Community Scripts

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:37:57
# Model            : Gradient Boosting
# Feature Eng.     : EMA (9,21), RSI 14 + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # --- EMA 9 and EMA 21 (required) ---
    ema_9 = close.ewm(span=9, adjust=False).mean()
    ema_21 = close.ewm(span=21, adjust=False).mean()
    df["ema_9"] = ema_9
    df["ema_21"] = ema_21
    df["dm_ema_9"] = (close - ema_9) / ema_9
    df["dm_ema_21"] = (close - ema_21) / ema_21

    # EMA crossover signal: positive when fast > slow
    df["ema_cross"] = ema_9 - ema_21
    # Crossover direction change (sign flip)
    df["ema_cross_signal"] = np.sign(df["ema_cross"])
    df["ema_cross_prev"] = df["ema_cross_signal"].shift(1)
    df["ema_cross_flip"] = (df["ema_cross_signal"] != df["ema_cross_prev"]).astype(float)

    # --- RSI 14 (required) ---
    delta = close.diff()
    gain = delta.clip(lower=0)
    loss = -delta.clip(upper=0)
    avg_gain = gain.ewm(com=13, adjust=False).mean()
    avg_loss = loss.ewm(com=13, adjust=False).mean()
    rs = avg_gain / (avg_loss + 1e-10)
    rsi_14 = 100 - (100 / (1 + rs))
    df["rsi_14"] = rsi_14
    # RSI normalised to [-1, 1] range
    df["rsi_norm"] = (rsi_14 - 50) / 50
    # RSI overbought/oversold flags
    df["rsi_ob"] = np.where(rsi_14 > 70, 1.0, 0.0)
    df["rsi_os"] = np.where(rsi_14 < 30, 1.0, 0.0)

    # --- Additional momentum and volatility features ---

    # EMA 50 for trend context
    ema_50 = close.ewm(span=50, adjust=False).mean()
    df["ema_50"] = ema_50
    df["dm_ema_50"] = (close - ema_50) / ema_50

    # Price momentum: rate of change over multiple horizons
    df["roc_4"] = close.pct_change(4)
    df["roc_8"] = close.pct_change(8)
    df["roc_16"] = close.pct_change(16)

    # ATR (Average True Range) for volatility
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low - close.shift(1)).abs()
    ], axis=1).max(axis=1)
    atr_14 = tr.ewm(span=14, adjust=False).mean()
    df["atr_14"] = atr_14
    # Normalised ATR
    df["natr_14"] = atr_14 / close

    # Bollinger Bands (20-period, 2 std)
    bb_mid = close.rolling(20).mean()
    bb_std = close.rolling(20).std()
    bb_upper = bb_mid + 2 * bb_std
    bb_lower = bb_mid - 2 * bb_std
    df["bb_pct"] = (close - bb_lower) / (bb_upper - bb_lower + 1e-10)
    df["bb_width"] = (bb_upper - bb_lower) / (bb_mid + 1e-10)
    # BB squeeze: narrow bands signal potential breakout
    df["bb_squeeze"] = np.where(df["bb_width"] < df["bb_width"].rolling(50).mean(), 1.0, 0.0)

    # MACD-like: difference between two EMAs
    ema_12 = close.ewm(span=12, adjust=False).mean()
    ema_26 = close.ewm(span=26, adjust=False).mean()
    macd_line = ema_12 - ema_26
    macd_signal = macd_line.ewm(span=9, adjust=False).mean()
    df["macd_line"] = macd_line / (close + 1e-10)
    df["macd_signal"] = macd_signal / (close + 1e-10)
    df["macd_hist"] = (macd_line - macd_signal) / (close + 1e-10)
    df["macd_cross"] = np.sign(macd_line - macd_signal)

    # Stochastic oscillator (14-period)
    lowest_low = low.rolling(14).min()
    highest_high = high.rolling(14).max()
    stoch_k = 100 * (close - lowest_low) / (highest_high - lowest_low + 1e-10)
    stoch_d = stoch_k.rolling(3).mean()
    df["stoch_k"] = stoch_k
    df["stoch_d"] = stoch_d
    df["stoch_diff"] = stoch_k - stoch_d

    # Volume of price movement (candle body and shadows)
    df["body"] = (close - open_).abs() / (atr_14 + 1e-10)
    df["upper_shadow"] = (high - pd.concat([close, open_], axis=1).max(axis=1)) / (atr_14 + 1e-10)
    df["lower_shadow"] = (pd.concat([close, open_], axis=1).min(axis=1) - low) / (atr_14 + 1e-10)
    df["candle_dir"] = np.sign(close - open_)

    # Rolling volatility (realised vol over 20 bars)
    log_ret = np.log(close / close.shift(1))
    df["realvol_20"] = log_ret.rolling(20).std()

    # Close relative to recent high/low channel (20-bar)
    roll_high_20 = high.rolling(20).max()
    roll_low_20 = low.rolling(20).min()
    df["chan_pct_20"] = (close - roll_low_20) / (roll_high_20 - roll_low_20 + 1e-10)

    # Lagged RSI and EMA cross for temporal context
    df["rsi_14_lag1"] = rsi_14.shift(1)
    df["rsi_14_lag2"] = rsi_14.shift(2)
    df["ema_cross_lag1"] = df["ema_cross"].shift(1)
    df["ema_cross_lag2"] = df["ema_cross"].shift(2)
    df["macd_hist_lag1"] = df["macd_hist"].shift(1)

    # RSI momentum: change in RSI
    df["rsi_delta_1"] = rsi_14.diff(1)
    df["rsi_delta_4"] = rsi_14.diff(4)

    # EMA9 slope (normalised)
    df["ema9_slope"] = ema_9.diff(3) / (ema_9.shift(3) + 1e-10)
    df["ema21_slope"] = ema_21.diff(3) / (ema_21.shift(3) + 1e-10)

    # Interaction: RSI * EMA cross direction
    df["rsi_ema_cross_interact"] = df["rsi_norm"] * df["ema_cross_signal"]

    # Fill NaN from indicator warm-up
    df = df.bfill().ffill()

    return df

# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title": "EMA Cross 9/21 + RSI14 Gradient Boost Scalper",
        "model_type": "GradientBoostingClassifier",
        "model_params": {
            "n_estimators": 400,
            "max_depth": 4,
            "learning_rate": 0.04,
            "subsample": 0.8,
            "min_samples_leaf": 20,
            "min_samples_split": 40,
            "max_features": "sqrt",
            "n_iter_no_change": 30,
            "validation_fraction": 0.1,
            "tol": 1e-4,
        },
        "signal_threshold": 0.56,
        "direction": "both",
        "stop_loss": 0.005,
        "take_profit": 0.010,
        "cooldown": 0,
        "max_positions": 1,
        "on_opposite": "reverse",
        "session_filter": [7, 17],
        "min_atr": 0.0002,
        "trend_filter": "sma_50",
        "target_horizon": 4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe/Calmar) using a GradientBoostingClassifier "
            "with EMA 9/21 crossover as the primary signal source and RSI 14 as confirmation. "
            "Deep feature set includes MACD, Bollinger Bands, Stochastic, ATR normalisation, "
            "candle structure, lagged features and RSI/EMA interaction terms. "
            "Gradient boosting chosen for its ability to capture non-linear interactions between "
            "trend, momentum and volatility features without overfitting when regularised via "
            "subsample, max_features, and early stopping. Threshold 0.56 filters marginal signals. "
            "Session filter [7,17] focuses on London/NY overlap for highest EUR/USD liquidity. "
            "SL 0.5% / TP 1.0% gives 1:2 risk-reward aligned with scalper momentum targets. "
            "Reverse on opposite signal to stay in sync with fast EMA crossover momentum."
        ),
        "notes": (
            "EMA 9/21 cross captures short-term momentum shifts typical of active EUR/USD sessions. "
            "RSI 14 filters entries in extreme overbought/oversold conditions. "
            "NATR min_atr filter removes flat/low-vol periods. "
            "Trend filter (SMA 50) ensures longs only above and shorts only below the medium-term trend. "
            "n_iter_no_change=30 provides early stopping to prevent overfitting on the training split. "
            "400 estimators with depth 4 and lr 0.04 balance bias-variance tradeoff for intraday data."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-05 09:59:46
# Model            : XGBoost
# Feature Eng.     : SMA (20,50,200), BB (20,2.0), RSI 14, MACD (12,26,9), ATR 14 + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet"
START_DATE = "2025-04-23"
END_DATE   = "2026-04-23"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):

    # ── SMA 20, 50, 200 + distance from close ──────────────────────────────
    for p in [20, 50, 200]:
        sma = close.rolling(p).mean()
        df[f"sma_{p}"]    = sma
        df[f"dm_sma_{p}"] = (close - sma) / sma

    # ── Bollinger Bands (20, 2.0) ───────────────────────────────────────────
    bb_mid   = close.rolling(20).mean()
    bb_std   = close.rolling(20).std(ddof=0)
    bb_upper = bb_mid + 2.0 * bb_std
    bb_lower = bb_mid - 2.0 * bb_std
    bb_range = bb_upper - bb_lower
    df["bb_mid"]   = bb_mid
    df["bb_upper"] = bb_upper
    df["bb_lower"] = bb_lower
    df["bb_width"] = bb_range / bb_mid
    df["bb_pct"]   = (close - bb_lower) / bb_range.replace(0, np.nan)

    # ── RSI 14 ──────────────────────────────────────────────────────────────
    delta = close.diff()
    gain  = delta.clip(lower=0)
    loss  = (-delta).clip(lower=0)
    avg_gain = gain.ewm(com=13, min_periods=14, adjust=False).mean()
    avg_loss = loss.ewm(com=13, min_periods=14, adjust=False).mean()
    rs  = avg_gain / avg_loss.replace(0, np.nan)
    df["rsi_14"] = 100.0 - (100.0 / (1.0 + rs))

    # ── MACD (12, 26, 9) ────────────────────────────────────────────────────
    ema12 = close.ewm(span=12, adjust=False).mean()
    ema26 = close.ewm(span=26, adjust=False).mean()
    macd_line   = ema12 - ema26
    signal_line = macd_line.ewm(span=9, adjust=False).mean()
    df["macd_line"]   = macd_line
    df["macd_signal"] = signal_line
    df["macd_hist"]   = macd_line - signal_line

    # ── ATR 14 + NATR ───────────────────────────────────────────────────────
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low  - close.shift(1)).abs()
    ], axis=1).max(axis=1)
    atr = tr.ewm(com=13, min_periods=14, adjust=False).mean()
    df["atr_14"] = atr
    df["natr"]   = atr / close

    # ── Additional derived features ─────────────────────────────────────────

    # Price momentum over multiple horizons
    for lag in [1, 4, 8, 16]:
        df[f"mom_{lag}"] = close.pct_change(lag)

    # Log return
    df["log_ret_1"] = np.log(close / close.shift(1))

    # Candle body and wick ratios
    body        = (close - open_).abs()
    candle_rng  = (high - low).replace(0, np.nan)
    df["body_ratio"]      = body / candle_rng
    df["upper_wick_ratio"] = (high - pd.concat([close, open_], axis=1).max(axis=1)) / candle_rng
    df["lower_wick_ratio"] = (pd.concat([close, open_], axis=1).min(axis=1) - low) / candle_rng
    df["candle_dir"]       = np.where(close >= open_, 1.0, -1.0)

    # Volume / volatility proxy: rolling std of returns
    ret = close.pct_change()
    df["vol_8"]  = ret.rolling(8).std()
    df["vol_20"] = ret.rolling(20).std()

    # RSI-derived features
    df["rsi_dist_50"]  = df["rsi_14"] - 50.0
    df["rsi_overbought"] = np.where(df["rsi_14"] > 70, 1.0, 0.0)
    df["rsi_oversold"]   = np.where(df["rsi_14"] < 30, 1.0, 0.0)

    # MACD cross signal
    df["macd_cross"] = np.where(
        (df["macd_hist"] > 0) & (df["macd_hist"].shift(1) <= 0), 1.0,
        np.where(
            (df["macd_hist"] < 0) & (df["macd_hist"].shift(1) >= 0), -1.0,
            0.0
        )
    )

    # BB squeeze: narrow bands relative to recent average
    df["bb_squeeze"] = np.where(df["bb_width"] < df["bb_width"].rolling(50).mean(), 1.0, 0.0)

    # Price relative to SMA crossovers
    df["sma20_above_sma50"]  = np.where(df["sma_20"] > df["sma_50"],  1.0, -1.0)
    df["sma50_above_sma200"] = np.where(df["sma_50"] > df["sma_200"], 1.0, -1.0)

    # High/low channel breakout features
    df["high_20"]   = high.rolling(20).max()
    df["low_20"]    = low.rolling(20).min()
    df["chan_pos"]  = (close - df["low_20"]) / (df["high_20"] - df["low_20"]).replace(0, np.nan)

    # Lagged RSI and MACD hist
    df["rsi_14_lag1"] = df["rsi_14"].shift(1)
    df["rsi_14_lag4"] = df["rsi_14"].shift(4)
    df["macd_hist_lag1"] = df["macd_hist"].shift(1)

    # ATR trend: expanding vs contracting volatility
    df["atr_ratio"] = df["atr_14"] / df["atr_14"].rolling(50).mean()

    # Fill NaN from warm-up
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title": "EUR/USD XGBoost SMA+RSI+MACD+BB Trend Rider",
        "model_type": "XGBClassifier",
        "model_params": {
            "n_estimators":      400,
            "max_depth":         4,
            "learning_rate":     0.04,
            "subsample":         0.75,
            "colsample_bytree":  0.70,
            "min_child_weight":  5,
            "gamma":             0.2,
            "reg_alpha":         0.1,
            "reg_lambda":        1.5,
            "objective":         "binary:logistic",
            "tree_method":       "hist",
            "random_state":      42,
        },
        "signal_threshold": 0.54,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 20],
        "min_atr":          0.0002,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe/Calmar) on EUR/USD 15-min data. "
            "XGBoost with moderate depth and regularisation to avoid overfitting on "
            "noisy FX data. Conservative SL/TP ratio of 1:2 improves expectancy. "
            "Session filter keeps the model active during liquid London/NY overlap. "
            "Min ATR filter avoids low-volatility noise. SMA-50 trend filter aligns "
            "trades with the prevailing medium-term trend, reducing whipsaw."
        ),
        "notes": (
            "Feature set: SMA 20/50/200 distances, BB width/pct, RSI 14, MACD histogram, "
            "ATR/NATR, multi-horizon momentum, candle structure, volatility, channel position, "
            "lagged indicators, and cross-over binary signals. "
            "Hyperparameters tuned for bias-variance balance: shallow trees (depth 4), "
            "high n_estimators with low learning rate, stochastic sampling, and L1/L2 "
            "regularisation reduce overfitting. Threshold 0.54 slightly above default to "
            "filter marginal signals while maintaining trade frequency."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-05 10:27:10
# Model            : XGBoost
# Feature Eng.     : EMA (50,200), ATR 14 + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet"
START_DATE = "2025-04-23"
END_DATE   = "2026-04-23"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── EMA 50 and EMA 200 ──────────────────────────────────────────────────
    ema_50  = close.ewm(span=50,  adjust=False).mean()
    ema_200 = close.ewm(span=200, adjust=False).mean()

    df["ema_50"]    = ema_50
    df["ema_200"]   = ema_200
    df["dm_ema_50"]  = (close - ema_50)  / ema_50
    df["dm_ema_200"] = (close - ema_200) / ema_200

    # EMA cross signal: positive when fast > slow
    df["ema_cross"]      = ema_50 - ema_200
    df["ema_cross_norm"] = df["ema_cross"] / ema_200

    # Cross direction change (momentum of the spread)
    df["ema_cross_delta"] = df["ema_cross"].diff(1)
    df["ema_cross_accel"] = df["ema_cross_delta"].diff(1)

    # ── ATR 14 ──────────────────────────────────────────────────────────────
    prev_close = close.shift(1)
    tr = pd.concat([
        high - low,
        (high - prev_close).abs(),
        (low  - prev_close).abs()
    ], axis=1).max(axis=1)

    atr = tr.ewm(span=14, adjust=False).mean()
    df["atr"]  = atr
    df["natr"] = atr / close

    # ── Price momentum features ─────────────────────────────────────────────
    for lag in [1, 2, 4, 8, 16]:
        df[f"ret_{lag}"] = close.pct_change(lag)

    # ── Volatility regime ───────────────────────────────────────────────────
    df["atr_ratio"]    = atr / atr.rolling(50).mean()          # ATR vs its own MA
    df["natr_ma20"]    = df["natr"].rolling(20).mean()

    # ── RSI 14 ──────────────────────────────────────────────────────────────
    delta  = close.diff(1)
    gain   = delta.clip(lower=0)
    loss   = (-delta).clip(lower=0)
    avg_g  = gain.ewm(span=14, adjust=False).mean()
    avg_l  = loss.ewm(span=14, adjust=False).mean()
    rs     = avg_g / avg_l.replace(0, np.nan)
    df["rsi_14"] = 100 - (100 / (1 + rs))
    df["rsi_delta"] = df["rsi_14"].diff(1)

    # ── MACD (12/26/9) ──────────────────────────────────────────────────────
    ema_12   = close.ewm(span=12, adjust=False).mean()
    ema_26   = close.ewm(span=26, adjust=False).mean()
    macd     = ema_12 - ema_26
    signal   = macd.ewm(span=9, adjust=False).mean()
    df["macd"]          = macd
    df["macd_signal"]   = signal
    df["macd_hist"]     = macd - signal
    df["macd_hist_delta"] = df["macd_hist"].diff(1)

    # ── Bollinger Bands (20, 2σ) ─────────────────────────────────────────────
    bb_mid  = close.rolling(20).mean()
    bb_std  = close.rolling(20).std()
    bb_up   = bb_mid + 2 * bb_std
    bb_lo   = bb_mid - 2 * bb_std
    df["bb_pos"]   = (close - bb_lo) / (bb_up - bb_lo).replace(0, np.nan)
    df["bb_width"] = (bb_up - bb_lo) / bb_mid

    # ── Stochastic %K / %D (14, 3) ──────────────────────────────────────────
    low14  = low.rolling(14).min()
    high14 = high.rolling(14).max()
    stoch_k = 100 * (close - low14) / (high14 - low14).replace(0, np.nan)
    stoch_d = stoch_k.rolling(3).mean()
    df["stoch_k"] = stoch_k
    df["stoch_d"] = stoch_d
    df["stoch_kd"] = stoch_k - stoch_d

    # ── Volume / body / wick features ───────────────────────────────────────
    body    = (close - open_).abs()
    candle  = (high - low).replace(0, np.nan)
    df["body_ratio"]  = body / candle
    df["upper_wick"]  = (high - pd.concat([close, open_], axis=1).max(axis=1)) / candle
    df["lower_wick"]  = (pd.concat([close, open_], axis=1).min(axis=1) - low) / candle
    df["bull_candle"]  = np.where(close > open_, 1, 0)

    # ── Rolling z-score of close vs SMA 50 ──────────────────────────────────
    sma_50       = close.rolling(50).mean()
    sma_50_std   = close.rolling(50).std()
    df["zscore_50"] = (close - sma_50) / sma_50_std.replace(0, np.nan)

    # ── High/Low breakout flags ──────────────────────────────────────────────
    df["high_20_break"] = np.where(close > high.rolling(20).max().shift(1), 1, 0)
    df["low_20_break"]  = np.where(close < low.rolling(20).min().shift(1),  1, 0)

    # ── Time-of-day features (cyclical encoding) ─────────────────────────────
    if hasattr(df.index, 'hour'):
        hour = df.index.hour + df.index.minute / 60.0
        df["hour_sin"] = np.sin(2 * np.pi * hour / 24.0)
        df["hour_cos"] = np.cos(2 * np.pi * hour / 24.0)

    # ── Fill NaN from indicator warm-up ─────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "EMA Cross 50/200 + ATR Momentum (XGBoost)",
        "model_type":       "XGBClassifier",
        "model_params": {
            "n_estimators":       400,
            "max_depth":          4,
            "learning_rate":      0.04,
            "subsample":          0.75,
            "colsample_bytree":   0.70,
            "min_child_weight":   3,
            "gamma":              0.15,
            "reg_alpha":          0.10,
            "reg_lambda":         1.50,
            "objective":          "binary:logistic",
            "random_state":       42,
            "n_jobs":             -1,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 20],
        "min_atr":          0.0002,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe/Calmar) on EUR/USD 15-min data. "
            "EMA 50/200 cross provides the primary trend regime filter. "
            "ATR 14 gates entries by volatility (min_atr avoids dead-market noise). "
            "XGBoost chosen for its ability to capture non-linear feature interactions. "
            "Conservative depth=4 and regularisation (alpha/lambda) prevent overfitting "
            "on the relatively short 1-year window. 2:1 reward/risk (SL=0.5%, TP=1.0%) "
            "ensures positive expectancy even at modest hit-rates. Session filter 06-20 UTC "
            "keeps the strategy in liquid London/NY hours only."
        ),
        "notes": (
            "Feature set combines trend (EMA cross, z-score), momentum (RSI, MACD, returns), "
            "volatility (ATR ratio, BB width), and price structure (body/wick ratios, "
            "stochastic). Cyclical hour encoding captures intraday seasonality without "
            "introducing lookahead. bfill().ffill() handles EMA warm-up NaNs gracefully."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-07 01:38:25
# Model            : XGBoost
# Feature Eng.     : EMA crossover trend (9/21) with RSI 14 confirmation on EURUSD 15min + Auto-add features: ON
# Signal / Entry   : —
# Optimization     : —
# Risk Mgmt        : —
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# QUANTIFY ME — STRATEGY MODULE
# EMA Crossover + RSI Confirmation (XGBoost, Sharpe Optimization)
# ============================================================

import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet"
START_DATE = "2026-03-28"
END_DATE   = "2026-04-25"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7


# ============================================================
# SECTION 1 — FEATURE ENGINEERING
# ============================================================

def feature_engineering(df, close, open_, high, low):
    """
    Add EMA crossover trend features + RSI confirmation.
    
    Features:
      - EMA 9 and EMA 21 for trend direction
      - RSI 14 for momentum confirmation
      - EMA crossover signal
      - Price deviation from EMA 21
      - High/Low proximity ratios
      - Volume-based volatility (NATR)
    """
    
    # EMA 9 and EMA 21 for trend
    df['ema_9'] = close.ewm(span=9, adjust=False).mean()
    df['ema_21'] = close.ewm(span=21, adjust=False).mean()
    df['ema_crossover'] = np.where(df['ema_9'] > df['ema_21'], 1, -1)
    
    # EMA crossover signal (1 when 9 crosses above 21, -1 when crosses below)
    df['ema_cross_signal'] = df['ema_crossover'].diff().fillna(0)
    df['ema_cross_signal'] = np.where(df['ema_cross_signal'] != 0, df['ema_cross_signal'], 0)
    
    # Price deviation from EMA 21 (normalized)
    df['price_ema_deviation'] = (close - df['ema_21']) / df['ema_21']
    
    # RSI 14 for momentum confirmation
    delta = close.diff()
    gain = np.where(delta > 0, delta, 0)
    loss = np.where(delta < 0, -delta, 0)
    
    avg_gain = pd.Series(gain, index=close.index).ewm(span=14, adjust=False).mean()
    avg_loss = pd.Series(loss, index=close.index).ewm(span=14, adjust=False).mean()
    
    rs = avg_gain / (avg_loss + 1e-10)
    df['rsi_14'] = 100 - (100 / (1 + rs))
    
    # RSI signal: overbought/oversold
    df['rsi_overbought'] = np.where(df['rsi_14'] > 70, 1, 0)
    df['rsi_oversold'] = np.where(df['rsi_14'] < 30, 1, 0)
    
    # High/Low proximity (distance from recent extremes)
    df['high_20'] = high.rolling(window=20).max()
    df['low_20'] = low.rolling(window=20).min()
    df['price_position'] = (close - df['low_20']) / (df['high_20'] - df['low_20'] + 1e-10)
    
    # NATR (Normalized ATR) for volatility
    atr_period = 14
    tr1 = high - low
    tr2 = np.abs(high - close.shift(1))
    tr3 = np.abs(low - close.shift(1))
    tr = np.maximum(tr1, np.maximum(tr2, tr3))
    atr = pd.Series(tr, index=close.index).rolling(window=atr_period).mean()
    df['natr'] = (atr / close) * 100
    
    # EMA momentum (rate of change in EMA)
    df['ema_9_roc'] = df['ema_9'].pct_change(periods=3)
    df['ema_21_roc'] = df['ema_21'].pct_change(periods=3)
    
    # Close relative to open (intrabar direction)
    df['close_above_open'] = np.where(close > open_, 1, 0)
    
    # Volume-based features (if available; otherwise skip)
    if 'volume' in df.columns:
        df['volume_ma'] = df['volume'].rolling(window=20).mean()
        df['volume_ratio'] = df['volume'] / (df['volume_ma'] + 1e-10)
    else:
        df['volume_ratio'] = 1.0
    
    # Fill NaN from indicator warm-up
    df = df.bfill().ffill()
    
    return df


# ============================================================
# SECTION 2 — STRATEGY CONFIG
# ============================================================

def strategy_config():
    """
    XGBoost strategy optimized for Sharpe ratio on EMA/RSI signals.
    
    Hyperparameters tuned for:
      - Fast learning (learning_rate=0.08)
      - Shallow trees (max_depth=4) to avoid overfitting on 15min data
      - Moderate boosting (n_estimators=250) for good generalization
      - Regularization (subsample=0.85, colsample_bytree=0.8)
      - Balanced class weights via scale_pos_weight
    
    Signal threshold 0.55 chosen to be moderately selective while maintaining
    good trade frequency on the EMA crossover setup.
    """
    
    return {
        # Model specification
        "model_type": "XGBClassifier",
        "model_params": {
            "n_estimators": 250,
            "max_depth": 4,
            "learning_rate": 0.08,
            "subsample": 0.85,
            "colsample_bytree": 0.8,
            "min_child_weight": 1,
            "gamma": 0.5,
            "reg_alpha": 0.1,
            "reg_lambda": 1.0,
            "random_state": 42,
            "verbosity": 0,
        },
        
        # Entry signal
        "signal_threshold": 0.55,
        
        # Position management
        "direction": "both",
        "max_positions": 1,
        "on_opposite": "reverse",
        "cooldown": 0,
        
        # Risk management
        "stop_loss": 0.008,
        "take_profit": 0.015,
        
        # Filters
        "session_filter": None,
        "min_atr": None,
        "trend_filter": None,
        
        # Target
        "target_horizon": 4,
        
        # Metadata
        "title": "EMA Crossover + RSI Confirmation (XGBoost)",
        "objective": "Maximize Sharpe ratio with EMA 9/21 trend + RSI 14 momentum confirmation",
        "notes": (
            "Strategy uses fast EMA (9) crossover above/below slow EMA (21) "
            "as primary trend signal, confirmed by RSI 14 momentum. "
            "XGBoost learns non-linear interactions between these features. "
            "Moderate SL/TP (0.8%/1.5%) and bidirectional trading for scalping efficiency. "
            "Optimized for EURUSD 15min with 70/30 train/test split."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-25 02:27:36
# Model            : XGBoost
# Feature Eng.     : go long when EMA(9) crosses above EMA(21), exit when it crosses back below + Auto-add features: ON
# Signal / Entry   : —
# Optimization     : —
# Risk Mgmt        : —
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

# ── Inlined strategy_utils ──
"""
strategy_utils.py — Standard utility functions for generated strategies.
Claude imports these instead of writing boilerplate from scratch.
This ensures consistent behavior across all generated strategies.
"""
import numpy as np
import pandas as pd
from sklearn.preprocessing import LabelEncoder


# Max backtest window per timeframe. A finer timeframe over a longer window
# blows up the results dict / parquet load / Modal train time (the 2026-05-12
# OOM was a 1-min × multi-year sweep) — and a 1-min strategy gains nothing from
# 2 years of 1-min bars. Enforced HERE because every training path (UI / API /
# Modal) funnels through run_strategy → load_ohlc. Env-overridable so a future
# "max plan" / dedicated-server tier can lift it.
_TF_MAX_DAYS = {
    "1min":  30,
    "5min":  90,
    "15min": 365,
    "1h":    730,
}


def _fetch_ohlc_from_internal(symbol: str, tf: str, start: str, end: str):
    """Phase 3.2: fetch parquet bytes from Server A's /internal/ohlc endpoint
    instead of reading a local file. Used inside Modal containers / Mac worker
    pool (Phase 3.4) so every train sees the same source of truth as the chart.

    Returns: pd.DataFrame (parquet decoded), or raises on any failure so the
    caller can fall back / surface a clear error in the job.
    """
    import hashlib as _hashlib, hmac as _hmac, io as _io, os as _os
    import urllib.request as _ur, urllib.parse as _urp

    base = (_os.environ.get("QM_INTERNAL_OHLC_BASE") or "").rstrip("/")
    secret = (_os.environ.get("INTERNAL_WS_SECRET") or "").strip()
    if not base:
        raise RuntimeError("QM_INTERNAL_OHLC_BASE not set")
    if not secret:
        raise RuntimeError("INTERNAL_WS_SECRET not set")

    msg = f"{symbol}|{tf}|{start}|{end}".encode("utf-8")
    sig = _hmac.new(secret.encode("utf-8"), msg, _hashlib.sha256).hexdigest()

    qs = _urp.urlencode({
        "symbol": symbol, "tf": tf,
        "start": start, "end": end, "sig": sig,
    })
    url = f"{base}/internal/ohlc?{qs}"

    req = _ur.Request(url, headers={"User-Agent": "qm-worker/1.0"})
    with _ur.urlopen(req, timeout=30) as resp:
        if resp.status != 200:
            raise RuntimeError(f"/internal/ohlc returned {resp.status}")
        payload = resp.read()
    print(f"[load_ohlc:internal] {symbol} {tf} fetched {len(payload)} bytes", flush=True)
    return pd.read_parquet(_io.BytesIO(payload))


def _parse_symbol_tf_from_path(data_path: str):
    """Pull SYMBOL + TF out of a path like .../EURUSD_1min.parquet."""
    import os as _os, re as _re
    base = _os.path.basename(str(data_path))
    m = _re.match(r"^([A-Z]{6})_(\d+min|\d+h)\.parquet$", base)
    if not m:
        return None, None
    return m.group(1), m.group(2)


def load_ohlc(data_path, start_date="", end_date=""):
    """Load OHLC parquet, sort index, filter dates. Always returns consistent format.

    The lower bound is clamped per timeframe (see _TF_MAX_DAYS) — a request for
    more history than the cap silently starts later.

    Phase 3.2: when env QM_USE_INTERNAL_OHLC=="1", fetch over HTTP from
    Server A's /internal/ohlc endpoint instead of pd.read_parquet on a local
    file (which on Modal is a stale Volume snapshot). The endpoint applies the
    same day-cap, so the local cap-check below is a defensive no-op in that
    path. Flag defaults to "0" → unchanged behavior.

    Returns: (df, close, open_, high, low)
    """
    import os as _os, re as _re

    _use_internal = _os.environ.get("QM_USE_INTERNAL_OHLC", "0") == "1"
    if _use_internal:
        _sym, _tf = _parse_symbol_tf_from_path(data_path)
        if not _sym or not _tf:
            raise RuntimeError(
                f"QM_USE_INTERNAL_OHLC=1 but DATA_PATH basename does not match "
                f"SYMBOL_TF.parquet: {data_path}"
            )
        df = _fetch_ohlc_from_internal(_sym, _tf, start_date or "", end_date or "")
    else:
        df = pd.read_parquet(data_path)
    df.index = pd.to_datetime(df.index)
    df = df.sort_index()

    # Per-timeframe window cap (timeframe inferred from the parquet filename).
    _m = _re.search(r"_(\d+min|\d+h)\.parquet$", _os.path.basename(str(data_path)))
    _tf = _m.group(1) if _m else None
    _max_days = _TF_MAX_DAYS.get(_tf)
    if _max_days and _max_days > 0 and len(df):
        _env_override = _os.environ.get(f"QM_MAX_DAYS_{_tf.upper()}")
        if _env_override and _env_override.isdigit():
            _max_days = int(_env_override)
        try:
            _eff_end = pd.Timestamp(end_date) if end_date else df.index.max()
            _eff_end = min(_eff_end, df.index.max())
            _floor = _eff_end - pd.Timedelta(days=_max_days)
            _req_start = pd.Timestamp(start_date) if start_date else df.index.min()
            if _req_start < _floor:
                print(f"[load_ohlc] {_tf} backtest window capped to {_max_days}d: "
                      f"start {_req_start.date()} -> {_floor.date()}", flush=True)
                start_date = _floor
        except Exception as _e:
            print(f"[load_ohlc] window-cap check skipped ({_e})", flush=True)

    if start_date:
        df = df[df.index >= start_date]
    if end_date:
        df = df[df.index <= end_date]

    return df, df["close"], df["open"], df["high"], df["low"]


def make_target(close, horizon=4):
    """Create target: direction N bars ahead. Default 4 bars = 1 hour on 15-min data.

    Returns: target (pd.Series of -1, 0, 1)
    """
    return np.sign(close.shift(-horizon) - close)


def split_data(df, target, feature_cols, train_split=0.7, validation_date=""):
    """Train/test split. Handles both ratio and date-based splits.
    Drops NaN from target before splitting. Encodes labels to [0,1,2].

    Returns: dict with keys:
        X_train, X_test, y_train, y_test,
        y_train_enc, y_test_enc, enc,
        close_train, close_test,
        split_idx, split_dt, n_train, n_test
    """
    # Drop NaN from target
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]

    close = df["close"]

    # Build feature matrix
    X = df[feature_cols].copy()
    X = X.bfill().ffill()
    X = X.replace([np.inf, -np.inf], np.nan).fillna(0.0)

    # Split
    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)

    split_idx = max(1, min(split_idx, len(df) - 1))

    X_train = X.iloc[:split_idx]
    X_test = X.iloc[split_idx:]
    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]

    split_dt = str(df.index[split_idx])

    # Label encoding — always fit on [-1, 0, 1]
    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    return {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }


def compute_overlays(close, df_index):
    """Compute BB and MA overlays on full dataset. Always consistent.

    Returns: (bb_dict, ma_dict)
    """
    bb_mid = close.rolling(20).mean()
    bb_std = close.rolling(20).std()
    bb_upper = bb_mid + 2 * bb_std
    bb_lower = bb_mid - 2 * bb_std

    ma50 = close.rolling(50).mean()
    ma100 = close.rolling(100).mean()
    ma200 = close.rolling(200).mean()

    def _safe(s):
        s = s.reindex(df_index).bfill().ffill()
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in s.values]

    bb = {"upper": _safe(bb_upper), "mid": _safe(bb_mid), "lower": _safe(bb_lower)}
    ma = {"ma50": _safe(ma50), "ma100": _safe(ma100), "ma200": _safe(ma200)}

    return bb, ma


def run_backtest(signal, close, capital=10000, cost=2e-5):
    """Run backtest with transaction costs.
    Uses price-based trade returns (same as webapp _compute_trades).
    Signal 0 = hold (keep current position), not close.

    Returns: dict with equity, trade_returns, long_returns, short_returns, bar_returns
    """
    sig_arr = signal.values
    price_arr = close.values
    idx = signal.index
    n = len(price_arr)

    # Trade returns — price-based (matches webapp _compute_trades exactly)
    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    last_dir = None
    entry_price = None
    entry_bar = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != last_dir:
            # Direction change — close previous trade, open new
            if last_dir is not None and entry_price is not None and entry_price != 0:
                ret = float(last_dir * (c - entry_price) / entry_price - cost)
                trade_returns.append(ret)
                if last_dir == 1:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                trade_log.append({
                    "type": "Buy" if last_dir == 1 else "Sell",
                    "entry_time": str(idx[entry_bar]),
                    "exit_time": str(idx[i]),
                    "entry_price": round(entry_price, 5),
                    "exit_price": round(c, 5),
                    "pnl": round(last_dir * (c - entry_price), 5),
                    "pnl_pct": round(ret * 100, 3),
                    "exit_reason": "signal",
                })
            entry_price = c
            entry_bar = i
            last_dir = s

    # Close last open trade
    if last_dir is not None and entry_price is not None and n > 0 and entry_price != 0:
        c = price_arr[-1]
        ret = float(last_dir * (c - entry_price) / entry_price - cost)
        trade_returns.append(ret)
        if last_dir == 1:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        trade_log.append({
            "type": "Buy" if last_dir == 1 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[-1]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(c, 5),
            "pnl": round(last_dir * (c - entry_price), 5),
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": "end",
        })

    # Equity curve from trade returns
    cumret = 1.0
    equity_vals = np.full(n, float(capital))
    trade_idx = 0
    in_trade = False
    t_entry_price = None
    t_dir = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != t_dir:
            if t_dir is not None and t_entry_price is not None and t_entry_price != 0:
                t_ret = t_dir * (c - t_entry_price) / t_entry_price - cost
                cumret *= (1 + t_ret)
            t_entry_price = c
            t_dir = s
        equity_vals[i] = capital * cumret

    # Bar returns for Sharpe
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if price_arr[i - 1] != 0 and last_dir is not None:
            bar_returns[i] = sig_arr[i - 1] * (price_arr[i] - price_arr[i - 1]) / price_arr[i - 1] if sig_arr[i - 1] != 0 else 0.0

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


def compute_trade_stats(trades, capital=10000):
    """Single source of truth for trade statistics.
    Every display path reads from this — no recomputation anywhere.
    All values are rounded and JSON-safe (no inf/nan).
    """
    if not trades:
        return {"n": 0, "wins": 0, "losses": 0, "wr": 0, "avg": 0,
                "best": 0, "worst": 0, "ret": 0, "np": 0, "mdd": 0,
                "pf": 0, "rr": 0, "expect": 0}
    w = [r for r in trades if r > 0]
    l = [r for r in trades if r < 0]
    cumret = 1.0
    for r in trades:
        cumret *= (1 + r)
    net_p = capital * (cumret - 1)
    # Max drawdown
    eq = np.cumprod([1.0] + [1 + r for r in trades])
    peak = np.maximum.accumulate(eq)
    mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    # Profit Factor
    gross_w = sum(w) if w else 0
    gross_l = abs(sum(l)) if l else 0
    pf = gross_w / gross_l if gross_l > 0 else (9999.0 if gross_w > 0 else 0)
    # Risk:Reward
    avg_w = float(np.mean(w)) if w else 0
    avg_l = abs(float(np.mean(l))) if l else 0
    rr = avg_w / avg_l if avg_l > 0 else (9999.0 if avg_w > 0 else 0)
    # Expectancy
    expect = net_p / len(trades)
    return {
        "n": len(trades), "wins": len(w), "losses": len(l),
        "wr": round(len(w) / len(trades), 4),
        "avg": round(float(np.mean(trades)), 6),
        "best": round(max(w), 6) if w else 0,
        "worst": round(min(l), 6) if l else 0,
        "ret": round(cumret - 1, 6),
        "np": round(net_p, 2),
        "mdd": round(mdd, 6),
        "pf": round(pf, 2),
        "rr": round(rr, 2),
        "expect": round(expect, 2),
    }


def compute_metrics(bt_result, close_test, capital=10000):
    """Compute all standard metrics from backtest result.
    Uses trade-level compounding (same as webapp _trade_stats) for accuracy.

    Returns: dict with total_ret, bh_ret, sharpe_strat, sharpe_bh, mdd, n_trades
    """
    equity = bt_result["equity"]
    trade_returns = bt_result["trade_returns"]

    # Total return — trade-level compounding (matches webapp)
    if trade_returns:
        cumret = 1.0
        for r in trade_returns:
            cumret *= (1 + r)
        total_ret = cumret - 1
    else:
        total_ret = 0.0

    # Buy and hold
    bh_equity = capital * (close_test / close_test.iloc[0])
    bh_ret = (bh_equity.iloc[-1] - capital) / capital if capital != 0 else 0.0

    # Sharpe ratio — trade-level (matches webapp: sqrt(252*26) annualization)
    if len(trade_returns) >= 2 and float(np.std(trade_returns)) > 0:
        sharpe_strat = float(np.mean(trade_returns) / np.std(trade_returns) * np.sqrt(252 * 26))
    else:
        sharpe_strat = 0.0

    bh_rets = bh_equity.pct_change().dropna()
    if len(bh_rets) > 1 and bh_rets.std() != 0:
        sharpe_bh = float((bh_rets.mean() / bh_rets.std()) * np.sqrt(252 * 24 * 4))
    else:
        sharpe_bh = 0.0

    # Max drawdown — trade-level (matches webapp)
    if trade_returns:
        eq = np.cumprod([1.0] + [1 + r for r in trade_returns])
        peak = np.maximum.accumulate(eq)
        mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    else:
        mdd = 0.0

    return {
        "total_ret": float(total_ret),
        "bh_ret": float(bh_ret),
        "sharpe_strat": float(sharpe_strat) if not np.isnan(sharpe_strat) else 0.0,
        "sharpe_bh": float(sharpe_bh) if not np.isnan(sharpe_bh) else 0.0,
        "mdd": float(mdd),
        "n_trades": len(trade_returns),
    }


# Diagnostics line/histogram series (equity / drawdown / rolling_acc / conf_hist)
# only feed the small Diagnostics charts — they're never used by the price chart
# or scroll-back. On a 1-min model trained over the (2.2-capped) window these are
# still ~30k points each; downsample to a visually-identical resolution before the
# dict leaves the trainer so it doesn't carry that into Server-A RAM / Postgres.
_RESULTS_SERIES_MAX = 5000


def _downsample_idx(n, cap=_RESULTS_SERIES_MAX):
    """Evenly-spaced index list spanning [0, n-1] (first+last always kept), or
    None when no downsampling is needed (n <= cap)."""
    if n <= cap:
        return None
    return np.unique(np.linspace(0, n - 1, cap).astype(int)).tolist()


def _take(arr, idx):
    """Subset a list by an index list (idx may be None → return arr unchanged)."""
    if idx is None or not isinstance(arr, list):
        return arr
    return [arr[i] for i in idx]


# trade_log / train_trade_log are lists of per-trade dicts (display-only — the
# Trade Log tab). They scale with TRADE count, not bar count, so the bar-window
# cap (Phase 2.2) doesn't bound them — a degenerate near-every-bar model can put
# 10k+ trade dicts in the blob (>3 MB). Cap each (independently — a small-N model
# keeps every trade) to the most-recent N, recording `*_total` + `*_truncated`
# so the true count is still reported. Real strategies have far fewer than
# _TRADE_LOG_MAX trades, so this only ever bites pathological models.
_TRADE_LOG_MAX = 5000


def _cap_trade_log(tl):
    """Return (capped_list, original_len, was_truncated)."""
    if not isinstance(tl, list) or len(tl) <= _TRADE_LOG_MAX:
        return tl, (len(tl) if isinstance(tl, list) else 0), False
    return tl[-_TRADE_LOG_MAX:], len(tl), True


def build_return_dict(split_result, bt_result, metrics, model, feature_cols,
                      signal_full, p_pos_test, p_neg_test, custom_figs=None,
                      bt_train_result=None, pre_stats=None):
    """Assemble the complete return dict. Handles ALL serialization.
    Never returns Timestamps, numpy arrays, or non-JSON types.

    Returns: JSON-safe dict with all required keys
    """
    df = split_result["df"]
    close = split_result["close"]
    close_test = split_result["close_test"]
    X_test = split_result["X_test"]
    y_test = split_result["y_test"]
    equity = bt_result["equity"]
    bar_returns = bt_result["bar_returns"]

    # OHLC
    ohlc_dates = [str(x) for x in df.index.tolist()]

    def _safe_list(arr):
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in arr]

    # Overlays
    bb, ma = compute_overlays(close, df.index)

    # Buy and hold equity
    capital = equity.iloc[0] if len(equity) > 0 else 10000
    bh_equity = capital * (close_test / close_test.iloc[0])

    # Confusion matrix
    from sklearn.metrics import confusion_matrix
    pred_test = model.predict(X_test)
    y_test_arr = np.asarray(y_test)
    cm = confusion_matrix(y_test_arr, pred_test, labels=[-1, 0, 1])

    # Rolling accuracy
    sig_arr = signal_full.reindex(close_test.index).values
    correct = pd.Series((pred_test == y_test_arr).astype(float), index=X_test.index)
    active_test = pd.Series(sig_arr != 0, index=close_test.index) if len(sig_arr) == len(close_test) else pd.Series(True, index=close_test.index)
    correct_active = correct.where(active_test, other=np.nan)
    rolling_acc = correct_active.rolling(30, min_periods=1).mean()

    # Feature importance
    importances = model.feature_importances_
    fi_pairs = sorted(zip(feature_cols, importances), key=lambda x: x[1])[-15:]

    # Drawdown
    rolling_max = equity.cummax()
    drawdown = (equity - rolling_max) / rolling_max.replace(0, np.nan)
    drawdown = drawdown.fillna(0.0)

    # ── Downsample the Diagnostics-only series (see _downsample_idx) ──────────
    _eq_dates = [str(x) for x in close_test.index.tolist()]
    _eq_strat = _safe_list(equity.values)
    _eq_bh    = _safe_list(bh_equity.values)
    _eq_idx   = _downsample_idx(len(_eq_dates))
    _eq_dates, _eq_strat, _eq_bh = _take(_eq_dates, _eq_idx), _take(_eq_strat, _eq_idx), _take(_eq_bh, _eq_idx)

    _ra_dates = [str(x) for x in rolling_acc.index.tolist()]
    _ra_vals  = [float(x) if (not np.isnan(x) and not np.isinf(x)) else None for x in rolling_acc.values]
    _ra_idx   = _downsample_idx(len(_ra_dates))
    _ra_dates, _ra_vals = _take(_ra_dates, _ra_idx), _take(_ra_vals, _ra_idx)

    _dd_dates = [str(x) for x in drawdown.index.tolist()]
    _dd_vals  = _safe_list(drawdown.values)
    _dd_idx   = _downsample_idx(len(_dd_dates))
    _dd_dates, _dd_vals = _take(_dd_dates, _dd_idx), _take(_dd_vals, _dd_idx)

    _cp_pos = [float(x) for x in (p_pos_test.tolist() if hasattr(p_pos_test, 'tolist') else list(p_pos_test))]
    _cp_neg = [float(x) for x in (p_neg_test.tolist() if hasattr(p_neg_test, 'tolist') else list(p_neg_test))]
    _cp_pos = _take(_cp_pos, _downsample_idx(len(_cp_pos)))
    _cp_neg = _take(_cp_neg, _downsample_idx(len(_cp_neg)))

    # ── Trade logs — display-only (Trade Log tab); cap to most-recent N with a
    #    `_total` field so the true count is still reported (see _cap_trade_log).
    #    NB: ret_dist arrays are left FULL — a downstream path in callbacks.py
    #    recomputes n_trades/win-rate from len(ret_dist), so a sample would skew
    #    the displayed counts; they're small anyway and gzip handles them.
    _tl_test, _tl_test_n, _tl_test_tr = _cap_trade_log(bt_result.get("trade_log", []))
    _tl_tr,   _tl_tr_n,   _tl_tr_tr   = _cap_trade_log(bt_train_result.get("trade_log", []) if bt_train_result else [])

    return {
        "ohlc": {
            "dates": ohlc_dates,
            "open": _safe_list(df["open"].values),
            "high": _safe_list(df["high"].values),
            "low": _safe_list(df["low"].values),
            "close": _safe_list(df["close"].values),
        },
        "signals": {
            "dates": [str(x) for x in signal_full.index.tolist()],
            "values": [float(x) for x in signal_full.values],
        },
        "bb": bb,
        "ma": ma,
        "equity": {
            "dates": _eq_dates,
            "strategy": _eq_strat,
            "bh": _eq_bh,
        },
        "feature_importance": {
            "names": [p[0] for p in fi_pairs],
            "values": [float(p[1]) for p in fi_pairs],
        },
        "conf_matrix": cm.tolist(),
        "conf_hist": {
            "p_pos": _cp_pos,
            "p_neg": _cp_neg,
        },
        "rolling_acc": {
            "dates": _ra_dates,
            "values": _ra_vals,
        },
        "drawdown": {
            "dates": _dd_dates,
            "values": _dd_vals,
        },
        "ret_dist": [float(x) for x in bt_result["trade_returns"]],
        "ret_dist_long": [float(x) for x in bt_result["long_returns"]],
        "ret_dist_short": [float(x) for x in bt_result["short_returns"]],
        "train_ret_dist": [float(x) for x in bt_train_result["trade_returns"]] if bt_train_result else [],
        "train_ret_dist_long": [float(x) for x in bt_train_result["long_returns"]] if bt_train_result else [],
        "train_ret_dist_short": [float(x) for x in bt_train_result["short_returns"]] if bt_train_result else [],
        "trade_log": _tl_test,
        "train_trade_log": _tl_tr,
        "trade_log_total": _tl_test_n,
        "train_trade_log_total": _tl_tr_n,
        "trade_log_truncated": _tl_test_tr,
        "train_trade_log_truncated": _tl_tr_tr,
        **(pre_stats or {}),
        "metrics": metrics,
        "split_dt": split_result["split_dt"],
        "split_idx": int(split_result["split_idx"]),
        "n_train": int(split_result["n_train"]),
        "n_test": int(split_result["n_test"]),
        "feature_cols": list(feature_cols),
        "custom_figs": custom_figs or [],
    }


# ════════════════════════════════════════════════════════════════════════════
# STRATEGY FRAMEWORK v2 — Config-driven architecture
# Claude writes feature_engineering() + strategy_config(). Framework does rest.
# ════════════════════════════════════════════════════════════════════════════

import importlib

_MODEL_REGISTRY = {
    "XGBClassifier":              ("xgboost", "XGBClassifier"),
    "RandomForestClassifier":     ("sklearn.ensemble", "RandomForestClassifier"),
    "GradientBoostingClassifier": ("sklearn.ensemble", "GradientBoostingClassifier"),
    "LogisticRegression":         ("sklearn.linear_model", "LogisticRegression"),
    "ExtraTreesClassifier":       ("sklearn.ensemble", "ExtraTreesClassifier"),
    "AdaBoostClassifier":         ("sklearn.ensemble", "AdaBoostClassifier"),
}


def _build_model_from_config(config, X_train, y_train_enc):
    """Build, fit, and wrap a model from strategy_config dict."""
    model_type = config.get("model_type", "RandomForestClassifier")
    model_params = dict(config.get("model_params", {}))

    if model_type not in _MODEL_REGISTRY:
        raise ValueError(f"Unknown model_type '{model_type}'. Valid: {list(_MODEL_REGISTRY.keys())}")

    module_path, class_name = _MODEL_REGISTRY[model_type]
    mod = importlib.import_module(module_path)
    cls = getattr(mod, class_name)

    # XGBoost defaults
    if class_name == "XGBClassifier":
        model_params.setdefault("use_label_encoder", False)
        model_params.setdefault("eval_metric", "mlogloss")
        model_params.setdefault("tree_method", "hist")
        # Determinism > speed (2026-05-25). XGBoost hist with n_jobs=-1 is
        # NON-reproducible even with random_state set — the parallel histogram
        # gradient-sum order varies across threads, so the SAME code + data
        # gives a slightly different model (and backtest) every run. Forcing
        # single-thread makes training bit-reproducible so: (a) a user who
        # copies a strategy and reruns it gets identical numbers, (b) the
        # community "Live" score matches a redeploy, (c) "same code, different
        # result" support reports go away. Cost: single-threaded XGB (a few
        # seconds slower on large windows; hist is fast so it's minor). FORCED
        # (not setdefault) so the guarantee can't be silently broken by a
        # strategy passing n_jobs. Exact reproducibility holds within the
        # platform (pinned versions / same Modal image); a user's own machine
        # with different xgboost/numpy/CPU can still differ in low-order bits.
        model_params["n_jobs"] = 1

    # Common defaults
    model_params.setdefault("random_state", 42)

    from model_wrapper import ModelWrapper

    clf = cls(**model_params)
    clf.fit(X_train, y_train_enc)

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    return ModelWrapper(clf, original_classes=enc.classes_, n_features=X_train.shape[1])


def _generate_signals(model, X, threshold):
    """Framework-owned signal generation. Deterministic threshold logic."""
    proba = model.predict_proba(X)
    classes = list(model.classes_)

    idx_pos = classes.index(1) if 1 in classes else None
    idx_neg = classes.index(-1) if -1 in classes else None

    p_pos = proba[:, idx_pos] if idx_pos is not None else np.zeros(len(X))
    p_neg = proba[:, idx_neg] if idx_neg is not None else np.zeros(len(X))

    signal_vals = np.zeros(len(X))
    signal_vals = np.where(p_pos >= threshold, 1.0, signal_vals)
    signal_vals = np.where(p_neg >= threshold, -1.0, signal_vals)
    # Both exceed: pick stronger
    both = (p_pos >= threshold) & (p_neg >= threshold)
    signal_vals[both] = np.where(p_pos[both] >= p_neg[both], 1.0, -1.0)

    return pd.Series(signal_vals, index=X.index), p_pos, p_neg


# ── Filter functions (all no-ops when config value is None) ──────────────

def _apply_direction_filter(signal, direction):
    """Zero out signals that don't match allowed direction."""
    if direction is None or direction == "both":
        return signal
    s = signal.copy()
    if direction == "long":
        s[s < 0] = 0.0
    elif direction == "short":
        s[s > 0] = 0.0
    return s


def _apply_session_filter(signal, index, session_hours):
    """Zero out signals outside session hours [start, end] UTC."""
    if session_hours is None:
        return signal
    s = signal.copy()
    start_h, end_h = session_hours[0], session_hours[1]
    hours = index.hour
    if start_h <= end_h:
        mask = (hours >= start_h) & (hours < end_h)
    else:  # wrap around midnight, e.g. [22, 6]
        mask = (hours >= start_h) | (hours < end_h)
    s[~mask] = 0.0
    return s


def _apply_atr_filter(signal, close, high, low, min_atr):
    """Zero out signals when NATR(14) is below threshold."""
    if min_atr is None:
        return signal
    hl = high - low
    hc = (high - close.shift(1)).abs()
    lc = (low - close.shift(1)).abs()
    tr = pd.concat([hl, hc, lc], axis=1).max(axis=1)
    atr14 = tr.ewm(com=13, adjust=False).mean()
    natr = atr14 / close.replace(0, np.nan)
    s = signal.copy()
    s[natr < min_atr] = 0.0
    return s


def _apply_trend_filter(signal, close, trend_filter):
    """Only allow signals aligned with trend. e.g. 'sma_50': longs above SMA, shorts below."""
    if trend_filter is None:
        return signal
    # Parse: "sma_50" → SMA with period 50
    parts = trend_filter.lower().replace("-", "_").split("_")
    if len(parts) >= 2 and parts[0] in ("sma", "ema"):
        period = int(parts[1])
    else:
        return signal  # unknown filter, skip
    if parts[0] == "sma":
        trend_line = close.rolling(period).mean()
    else:
        trend_line = close.ewm(span=period, adjust=False).mean()
    s = signal.copy()
    # Longs only above trend, shorts only below
    s[(s > 0) & (close < trend_line)] = 0.0
    s[(s < 0) & (close > trend_line)] = 0.0
    return s


# ── run_backtest_v2: framework-owned SL/TP/cooldown/position management ──

def run_backtest_v2(signal, close, high, low, config, capital=10000, cost=2e-5):
    """Backtest with SL/TP/cooldown/direction handling built into the engine.
    Unlike run_backtest (v1), this function handles position exits internally.

    Returns: same dict shape as run_backtest()
    """
    stop_loss = config.get("stop_loss")
    take_profit = config.get("take_profit")
    cooldown = config.get("cooldown", 0)
    on_opposite = config.get("on_opposite", "reverse")

    sig_arr = signal.values
    close_arr = close.values
    high_arr = high.values
    low_arr = low.values
    idx = signal.index
    n = len(close_arr)

    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    equity_vals = np.full(n, float(capital))
    cumret = 1.0

    position = 0.0       # current direction: 1.0, -1.0, or 0.0 (flat)
    entry_price = None
    entry_bar = None      # index into arrays for entry time
    cooldown_remaining = 0

    def _log_trade(exit_bar, exit_px, ret, reason):
        trade_log.append({
            "type": "Buy" if position == 1.0 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[exit_bar]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(exit_px, 5),
            "pnl": round(position * (exit_px - entry_price), 5),
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": reason,
        })

    for i in range(n):
        c = close_arr[i]
        h = high_arr[i]
        lo = low_arr[i]
        s = sig_arr[i]

        # 1. Check SL/TP if in trade
        if position != 0.0 and entry_price is not None:
            hit_sl = False
            hit_tp = False
            exit_price = None

            if position == 1.0:  # long
                if stop_loss is not None and lo <= entry_price * (1 - stop_loss):
                    hit_sl = True
                    exit_price = entry_price * (1 - stop_loss)
                elif take_profit is not None and h >= entry_price * (1 + take_profit):
                    hit_tp = True
                    exit_price = entry_price * (1 + take_profit)
            else:  # short
                if stop_loss is not None and h >= entry_price * (1 + stop_loss):
                    hit_sl = True
                    exit_price = entry_price * (1 + stop_loss)
                elif take_profit is not None and lo <= entry_price * (1 - take_profit):
                    hit_tp = True
                    exit_price = entry_price * (1 - take_profit)

            if hit_sl or hit_tp:
                ret = float(position * (exit_price - entry_price) / entry_price - cost)
                trade_returns.append(ret)
                if position == 1.0:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                _log_trade(i, exit_price, ret, "SL" if hit_sl else "TP")
                cumret *= (1 + ret)
                position = 0.0
                entry_price = None
                entry_bar = None
                cooldown_remaining = cooldown
                equity_vals[i] = capital * cumret
                continue

        # 2. Cooldown
        if cooldown_remaining > 0:
            cooldown_remaining -= 1
            equity_vals[i] = capital * cumret
            continue

        # 3. Signal processing
        if s != 0.0:
            if position == 0.0:
                # Open new trade
                position = s
                entry_price = c
                entry_bar = i
            elif s != position:
                # Opposite signal
                if on_opposite == "reverse":
                    # Close current + open opposite
                    ret = float(position * (c - entry_price) / entry_price - cost)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "signal")
                    cumret *= (1 + ret)
                    position = s
                    entry_price = c
                    entry_bar = i
                else:  # close_only
                    # Close current, go flat
                    ret = float(position * (c - entry_price) / entry_price - cost)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "close_only")
                    cumret *= (1 + ret)
                    position = 0.0
                    entry_price = None
                    entry_bar = None
                    cooldown_remaining = cooldown

        equity_vals[i] = capital * cumret

    # Close last open trade at final close
    if position != 0.0 and entry_price is not None and n > 0 and entry_price != 0:
        c = close_arr[-1]
        ret = float(position * (c - entry_price) / entry_price - cost)
        trade_returns.append(ret)
        if position == 1.0:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        _log_trade(n - 1, c, ret, "end")
        cumret *= (1 + ret)
        equity_vals[-1] = capital * cumret

    # Bar returns for Sharpe (approximate)
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if close_arr[i - 1] != 0 and sig_arr[i - 1] != 0:
            bar_returns[i] = sig_arr[i - 1] * (close_arr[i] - close_arr[i - 1]) / close_arr[i - 1]

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


# ── run_strategy: the v2 orchestrator ────────────────────────────────────

def run_strategy(feature_fn, config_fn, data_path, start_date="", end_date="",
                 validation_date="", train_split=0.7, register_model_fn=None):
    """Config-driven strategy execution. Claude writes feature_fn + config_fn,
    framework does everything else.

    Returns: results dict (same format as webapp expects)
    """
    config = config_fn()

    # Auto-correct SL/TP if Claude passed percentage instead of decimal
    for _key in ("stop_loss", "take_profit"):
        _val = config.get(_key)
        if _val is not None and _val > 0.1:  # >10% is almost certainly a percentage
            config[_key] = _val / 100.0
            print(f"[strategy] Auto-corrected {_key}: {_val} -> {config[_key]} (was percentage, converted to decimal)")

    # 1. Load data
    df, close, open_, high, low = load_ohlc(data_path, start_date, end_date)

    # 2. Feature engineering (Claude's function)
    df = feature_fn(df, close, open_, high, low)
    close = df["close"]
    open_ = df["open"]
    high = df["high"]
    low = df["low"]

    # 3. Warm-up detection: drop rows where features have NaN BEFORE any fill
    feature_cols = [c for c in df.columns if c not in ("open", "high", "low", "close")]
    raw_nans = df[feature_cols].isna().any(axis=1)
    valid_rows = ~raw_nans
    if valid_rows.any():
        first_valid = valid_rows.idxmax()
        if raw_nans.loc[:first_valid].any():
            df = df.loc[first_valid:].copy()
            close = df["close"]
            open_ = df["open"]
            high = df["high"]
            low = df["low"]

    # 4. Target
    horizon = config.get("target_horizon", 4)
    target = make_target(close, horizon=horizon)

    # 5. Split (ffill only within each partition — no bfill leak)
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]
    close = df["close"]
    high = df["high"]
    low = df["low"]

    X = df[feature_cols].copy()
    X = X.replace([np.inf, -np.inf], np.nan)

    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)
    split_idx = max(1, min(split_idx, len(df) - 1))

    # ffill within train and test separately (no leak)
    X_train = X.iloc[:split_idx].ffill().fillna(0.0)
    X_test = X.iloc[split_idx:].ffill().fillna(0.0)
    X = pd.concat([X_train, X_test])

    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]
    high_test = high.iloc[split_idx:]
    low_test = low.iloc[split_idx:]

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    split_dt = str(df.index[split_idx])

    sp = {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }

    # 6. Build model from config
    model = _build_model_from_config(config, X_train, y_train_enc)

    # 7. Generate signals
    threshold = config.get("signal_threshold", 0.55)
    signal_train, p_pos_train, p_neg_train = _generate_signals(model, X_train, threshold)
    signal_test, p_pos_test, p_neg_test = _generate_signals(model, X_test, threshold)

    # 8. Apply filters (order: direction → session → ATR → trend)
    direction = config.get("direction", "both")
    signal_test = _apply_direction_filter(signal_test, direction)
    signal_train = _apply_direction_filter(signal_train, direction)

    session_filter = config.get("session_filter")
    signal_test = _apply_session_filter(signal_test, signal_test.index, session_filter)
    signal_train = _apply_session_filter(signal_train, signal_train.index, session_filter)

    min_atr = config.get("min_atr")
    if min_atr is not None:
        signal_test = _apply_atr_filter(signal_test, close_test, high_test, low_test, min_atr)

    trend_filter = config.get("trend_filter")
    if trend_filter is not None:
        signal_test = _apply_trend_filter(signal_test, close_test, trend_filter)

    signal_full = pd.concat([signal_train, signal_test])

    # 9. Backtest with SL/TP/cooldown (test + train)
    high_train = high.iloc[:split_idx]
    low_train = low.iloc[:split_idx]
    has_risk = (config.get("stop_loss") is not None or
                config.get("take_profit") is not None or
                config.get("cooldown", 0) > 0 or
                config.get("on_opposite", "reverse") != "reverse")

    if has_risk:
        bt = run_backtest_v2(signal_test, close_test, high_test, low_test, config, capital=10000)
        bt_train = run_backtest_v2(signal_train, close_train, high_train, low_train, config, capital=10000)
    else:
        bt = run_backtest(signal_test, close_test, capital=10000)
        bt_train = run_backtest(signal_train, close_train, capital=10000)

    # 10. Metrics
    metrics = compute_metrics(bt, close_test, capital=10000)

    # 11. Pre-compute all trade stats (single source of truth)
    pre_stats = {
        "train_stats": compute_trade_stats(bt_train.get("trade_returns", []), capital=10000),
        "test_stats":  compute_trade_stats(bt.get("trade_returns", []), capital=10000),
        "long_stats":  compute_trade_stats(bt.get("long_returns", []), capital=10000),
        "short_stats": compute_trade_stats(bt.get("short_returns", []), capital=10000),
    }

    # 12. Register model
    if register_model_fn is not None:
        register_model_fn(model)

    # 13. Build return dict
    return build_return_dict(sp, bt, metrics, model, feature_cols,
                             signal_full, p_pos_test, p_neg_test, custom_figs=[],
                             bt_train_result=bt_train, pre_stats=pre_stats)

# ── End strategy_utils ──


DATA_PATH = '/root/Desktop/QuantifyMe/data/ohlc/GBPUSD_15min.parquet'
START_DATE = '2026-04-15'
END_DATE = '2026-05-25'
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── EMA crossover core signals ──────────────────────────────────────────
    ema9  = close.ewm(span=9,  adjust=False).mean()
    ema21 = close.ewm(span=21, adjust=False).mean()
    ema50 = close.ewm(span=50, adjust=False).mean()
    ema200 = close.ewm(span=200, adjust=False).mean()

    df["ema9"]  = ema9
    df["ema21"] = ema21
    df["ema50"] = ema50
    df["ema200"] = ema200

    # Raw spread and normalised spread
    df["ema_diff"]       = ema9 - ema21
    df["ema_diff_norm"]  = (ema9 - ema21) / close

    # Cross signal: +1 when ema9 > ema21, -1 otherwise
    df["ema_cross_sign"] = np.where(ema9 > ema21, 1.0, -1.0)

    # Momentum of the spread (rate of change of spread)
    df["ema_diff_roc1"]  = df["ema_diff"].diff(1)
    df["ema_diff_roc3"]  = df["ema_diff"].diff(3)

    # Distance of price from ema50 and ema200 (normalised)
    df["dist_ema50"]  = (close - ema50)  / close
    df["dist_ema200"] = (close - ema200) / close

    # ── RSI (14) ────────────────────────────────────────────────────────────
    delta  = close.diff()
    gain   = delta.clip(lower=0.0)
    loss   = (-delta).clip(lower=0.0)
    avg_g  = gain.ewm(com=13, adjust=False).mean()
    avg_l  = loss.ewm(com=13, adjust=False).mean()
    rs     = avg_g / avg_l.replace(0.0, np.nan)
    rsi14  = 100.0 - 100.0 / (1.0 + rs)
    df["rsi14"] = rsi14

    # RSI normalised and centred
    df["rsi14_norm"] = (rsi14 - 50.0) / 50.0

    # ── MACD ────────────────────────────────────────────────────────────────
    ema12      = close.ewm(span=12, adjust=False).mean()
    ema26      = close.ewm(span=26, adjust=False).mean()
    macd_line  = ema12 - ema26
    signal_ln  = macd_line.ewm(span=9, adjust=False).mean()
    macd_hist  = macd_line - signal_ln

    df["macd_line"]    = macd_line / close
    df["macd_signal"]  = signal_ln / close
    df["macd_hist"]    = macd_hist / close
    df["macd_cross"]   = np.where(macd_line > signal_ln, 1.0, -1.0)

    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_mid   = close.rolling(20).mean()
    bb_std   = close.rolling(20).std(ddof=0)
    bb_upper = bb_mid + 2.0 * bb_std
    bb_lower = bb_mid - 2.0 * bb_std
    bb_width = (bb_upper - bb_lower) / bb_mid.replace(0.0, np.nan)
    bb_pct   = (close - bb_lower) / (bb_upper - bb_lower).replace(0.0, np.nan)

    df["bb_width"] = bb_width
    df["bb_pct"]   = bb_pct

    # ── ATR (14) ─────────────────────────────────────────────────────────────
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low  - close.shift(1)).abs()
    ], axis=1).max(axis=1)
    atr14 = tr.ewm(com=13, adjust=False).mean()
    df["atr14"]      = atr14
    df["natr14"]     = atr14 / close          # normalised ATR (volatility proxy)

    # ── Stochastic %K / %D (14, 3) ──────────────────────────────────────────
    low14  = low.rolling(14).min()
    high14 = high.rolling(14).max()
    stoch_k = 100.0 * (close - low14) / (high14 - low14).replace(0.0, np.nan)
    stoch_d = stoch_k.rolling(3).mean()
    df["stoch_k"]      = stoch_k / 100.0
    df["stoch_d"]      = stoch_d / 100.0
    df["stoch_diff"]   = (stoch_k - stoch_d) / 100.0

    # ── Rate of Change ───────────────────────────────────────────────────────
    df["roc1"]  = close.pct_change(1)
    df["roc4"]  = close.pct_change(4)
    df["roc8"]  = close.pct_change(8)
    df["roc16"] = close.pct_change(16)

    # ── Candle features ──────────────────────────────────────────────────────
    body       = (close - open_).abs()
    candle_rng = (high - low).replace(0.0, np.nan)
    df["body_ratio"]   = body / candle_rng
    df["upper_wick"]   = (high - pd.concat([close, open_], axis=1).max(axis=1)) / candle_rng
    df["lower_wick"]   = (pd.concat([close, open_], axis=1).min(axis=1) - low) / candle_rng
    df["candle_dir"]   = np.where(close >= open_, 1.0, -1.0)

    # ── Volume-like proxy: range relative to rolling average ────────────────
    df["range_ratio"]  = candle_rng / candle_rng.rolling(20).mean()

    # ── Lagged EMA diff features ─────────────────────────────────────────────
    for lag in [1, 2, 3, 4]:
        df[f"ema_diff_lag{lag}"] = df["ema_diff_norm"].shift(lag)

    # ── Lagged RSI ───────────────────────────────────────────────────────────
    for lag in [1, 2, 4]:
        df[f"rsi14_lag{lag}"] = df["rsi14_norm"].shift(lag)

    # ── Rolling volatility (std of returns) ──────────────────────────────────
    ret = close.pct_change()
    df["vol_8"]  = ret.rolling(8).std()
    df["vol_16"] = ret.rolling(16).std()
    df["vol_32"] = ret.rolling(32).std()

    # ── Trend strength: ADX-like (simplified) ────────────────────────────────
    plus_dm  = (high.diff()).clip(lower=0.0)
    minus_dm = (-low.diff()).clip(lower=0.0)
    overlap  = pd.concat([plus_dm, minus_dm], axis=1).min(axis=1)
    plus_dm  = plus_dm - overlap
    minus_dm = minus_dm - overlap
    smooth_tr  = tr.ewm(com=13, adjust=False).mean()
    plus_di    = 100.0 * plus_dm.ewm(com=13, adjust=False).mean()  / smooth_tr.replace(0.0, np.nan)
    minus_di   = 100.0 * minus_dm.ewm(com=13, adjust=False).mean() / smooth_tr.replace(0.0, np.nan)
    di_sum     = (plus_di + minus_di).replace(0.0, np.nan)
    adx        = ((plus_di - minus_di).abs() / di_sum * 100.0).ewm(com=13, adjust=False).mean()
    df["adx"]       = adx / 100.0
    df["plus_di"]   = plus_di / 100.0
    df["minus_di"]  = minus_di / 100.0

    # ── Session hour (UTC) ───────────────────────────────────────────────────
    if hasattr(df.index, "hour"):
        df["hour_sin"] = np.sin(2.0 * np.pi * df.index.hour / 24.0)
        df["hour_cos"] = np.cos(2.0 * np.pi * df.index.hour / 24.0)
    else:
        df["hour_sin"] = 0.0
        df["hour_cos"] = 1.0

    # ── Fill any NaN from warm-up periods ────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "EMA 9/21 Crossover + MACD Momentum (XGBoost)",
        "model_type":       "XGBClassifier",
        "model_params":     {
            "n_estimators":     400,
            "max_depth":        4,
            "learning_rate":    0.04,
            "subsample":        0.80,
            "colsample_bytree": 0.75,
            "min_child_weight": 3,
            "gamma":            0.10,
            "reg_alpha":        0.05,
            "reg_lambda":       1.50,
            "objective":        "binary:logistic",
            "tree_method":      "hist",
            "random_state":     42,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.0030,
        "take_profit":      0.0060,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 20],
        "min_atr":          None,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective":        (
            "Maximize Sharpe ratio on EUR/USD 15-min data. "
            "Core signal: EMA(9) vs EMA(21) crossover enriched with MACD, RSI, "
            "Bollinger %B, Stochastic, ATR, ADX, candle structure and rolling "
            "volatility. XGBoost with moderate depth (4) and strong regularisation "
            "(gamma, alpha, lambda) prevents overfitting on ~6 weeks of intraday data. "
            "A 0.55 probability threshold filters low-confidence signals. "
            "A 2:1 TP:SL ratio (30 bp SL / 60 bp TP) improves the reward-risk "
            "balance. Session filter [6,20] UTC keeps the model away from the thin "
            "Asian pre-open. trend_filter sma_50 aligns entries with the prevailing "
            "short-term trend to reduce chop. Cooldown=0 and reverse-on-opposite "
            "allow continuous participation in trending EMA crossover moves."
        ),
        "notes":            (
            "round-trip cost 2e-5 is accounted for by the framework. "
            "target_horizon=4 bars (1 hour ahead) suits EMA crossover which "
            "generates medium-frequency signals rather than tick-level scalps. "
            "All features are normalised or expressed as ratios to minimise "
            "scale sensitivity for the logistic-objective XGBoost."
        ),
    }

# ── Framework v2: auto-generated wrapper ──
def train_and_backtest():
    _vd = VALIDATION_DATE if 'VALIDATION_DATE' in globals() else ''
    _ts = TRAIN_SPLIT if 'TRAIN_SPLIT' in globals() else 0.7
    return run_strategy(
        feature_engineering, strategy_config,
        DATA_PATH, START_DATE, END_DATE,
        _vd, _ts,
        register_model_fn=register_model
    )

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-05 10:44:36
# Model            : XGBoost
# Feature Eng.     : SMA (20,50,200) + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet"
START_DATE = "2025-04-23"
END_DATE   = "2026-04-23"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── SMA core features (required) ──────────────────────────────────────────
    for p in [20, 50, 200]:
        sma = close.rolling(p).mean()
        df[f"sma_{p}"]    = sma
        df[f"dm_sma_{p}"] = (close - sma) / sma

    # ── SMA cross-ratios ──────────────────────────────────────────────────────
    df["sma_20_50_ratio"]  = df["sma_20"]  / df["sma_50"]
    df["sma_50_200_ratio"] = df["sma_50"]  / df["sma_200"]
    df["sma_20_200_ratio"] = df["sma_20"]  / df["sma_200"]

    # ── SMA slope (rate of change of SMA over N bars) ─────────────────────────
    for p in [20, 50, 200]:
        df[f"sma_{p}_slope5"] = df[f"sma_{p}"].diff(5) / df[f"sma_{p}"].shift(5)

    # ── Price momentum / returns ───────────────────────────────────────────────
    for lag in [1, 2, 3, 4, 8, 16]:
        df[f"ret_{lag}"] = close.pct_change(lag)

    # ── Volatility (rolling std of returns) ───────────────────────────────────
    ret1 = close.pct_change(1)
    for w in [8, 20, 50]:
        df[f"vol_std_{w}"] = ret1.rolling(w).std()

    # ── ATR (manual) ──────────────────────────────────────────────────────────
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low  - close.shift(1)).abs()
    ], axis=1).max(axis=1)
    for w in [8, 14, 20]:
        atr = tr.rolling(w).mean()
        df[f"atr_{w}"]  = atr
        df[f"natr_{w}"] = atr / close

    # ── RSI (manual) ──────────────────────────────────────────────────────────
    for period in [7, 14, 21]:
        delta = close.diff(1)
        gain  = delta.clip(lower=0).rolling(period).mean()
        loss  = (-delta.clip(upper=0)).rolling(period).mean()
        rs    = gain / loss.replace(0, np.nan)
        df[f"rsi_{period}"] = 100 - (100 / (1 + rs))

    # ── MACD (manual) ─────────────────────────────────────────────────────────
    ema12 = close.ewm(span=12, adjust=False).mean()
    ema26 = close.ewm(span=26, adjust=False).mean()
    macd_line   = ema12 - ema26
    signal_line = macd_line.ewm(span=9, adjust=False).mean()
    df["macd_line"]   = macd_line
    df["macd_signal"] = signal_line
    df["macd_hist"]   = macd_line - signal_line
    df["macd_norm"]   = macd_line / close

    # ── Bollinger Bands (manual) ──────────────────────────────────────────────
    for w in [20]:
        mid  = close.rolling(w).mean()
        std  = close.rolling(w).std()
        upper = mid + 2 * std
        lower = mid - 2 * std
        bw    = (upper - lower) / mid
        pct_b = (close - lower) / (upper - lower).replace(0, np.nan)
        df[f"bb_upper_{w}"] = upper
        df[f"bb_lower_{w}"] = lower
        df[f"bb_width_{w}"] = bw
        df[f"bb_pct_{w}"]   = pct_b

    # ── Stochastic oscillator (manual) ────────────────────────────────────────
    for k_period in [14]:
        lo_k = low.rolling(k_period).min()
        hi_k = high.rolling(k_period).max()
        stoch_k = 100 * (close - lo_k) / (hi_k - lo_k).replace(0, np.nan)
        stoch_d = stoch_k.rolling(3).mean()
        df[f"stoch_k_{k_period}"] = stoch_k
        df[f"stoch_d_{k_period}"] = stoch_d

    # ── CCI (manual) ──────────────────────────────────────────────────────────
    for w in [14, 20]:
        tp     = (high + low + close) / 3
        tp_ma  = tp.rolling(w).mean()
        tp_md  = tp.rolling(w).apply(lambda x: np.mean(np.abs(x - x.mean())), raw=True)
        df[f"cci_{w}"] = (tp - tp_ma) / (0.015 * tp_md.replace(0, np.nan))

    # ── Williams %R (manual) ──────────────────────────────────────────────────
    for w in [14]:
        hi_w = high.rolling(w).max()
        lo_w = low.rolling(w).min()
        df[f"willr_{w}"] = -100 * (hi_w - close) / (hi_w - lo_w).replace(0, np.nan)

    # ── Donchian channel position ──────────────────────────────────────────────
    for w in [20, 50]:
        hi_d = high.rolling(w).max()
        lo_d = low.rolling(w).min()
        df[f"donch_pos_{w}"] = (close - lo_d) / (hi_d - lo_d).replace(0, np.nan)

    # ── Rolling high / low distances ──────────────────────────────────────────
    for w in [8, 20]:
        df[f"dist_hi_{w}"] = (high.rolling(w).max() - close) / close
        df[f"dist_lo_{w}"] = (close - low.rolling(w).min())  / close

    # ── Bar body / wick features ───────────────────────────────────────────────
    body      = (close - open_).abs()
    bar_range = (high - low).replace(0, np.nan)
    df["body_ratio"]      = body / bar_range
    df["upper_wick_ratio"] = (high - close.clip(lower=open_)) / bar_range
    df["lower_wick_ratio"] = (close.clip(upper=open_) - low) / bar_range
    df["bar_direction"]    = np.where(close >= open_, 1.0, -1.0)

    # ── Trend regime flags (binary) ────────────────────────────────────────────
    df["above_sma20"]  = np.where(close > df["sma_20"],  1.0, 0.0)
    df["above_sma50"]  = np.where(close > df["sma_50"],  1.0, 0.0)
    df["above_sma200"] = np.where(close > df["sma_200"], 1.0, 0.0)
    df["sma20_above50"] = np.where(df["sma_20"] > df["sma_50"], 1.0, 0.0)
    df["sma50_above200"] = np.where(df["sma_50"] > df["sma_200"], 1.0, 0.0)

    # ── Lagged returns for autoregressive signal ───────────────────────────────
    for lag in [1, 2, 3, 4, 5]:
        df[f"close_lag_{lag}"] = close.shift(lag)
        df[f"ret_lag_{lag}"]   = ret1.shift(lag)

    # ── Rolling correlation: price vs SMA distance ─────────────────────────────
    for w in [20]:
        df[f"autocorr_ret_{w}"] = ret1.rolling(w).apply(
            lambda x: pd.Series(x).autocorr(lag=1) if len(x) > 1 else 0.0, raw=False
        )

    # ── Volume-proxy: bar range z-score ───────────────────────────────────────
    for w in [20]:
        rng_mean = bar_range.rolling(w).mean()
        rng_std  = bar_range.rolling(w).std().replace(0, np.nan)
        df[f"range_zscore_{w}"] = (bar_range - rng_mean) / rng_std

    # ── Fill NaN from warm-up ─────────────────────────────────────────────────
    df = df.bfill().ffill()
    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "EUR/USD SMA Trend + Multi-Indicator XGBoost",
        "model_type":       "XGBClassifier",
        "model_params": {
            "n_estimators":       500,
            "max_depth":          4,
            "learning_rate":      0.03,
            "subsample":          0.8,
            "colsample_bytree":   0.7,
            "colsample_bylevel":  0.8,
            "min_child_weight":   5,
            "gamma":              0.1,
            "reg_alpha":          0.1,
            "reg_lambda":         2.0,
            "scale_pos_weight":   1,
            "objective":          "binary:logistic",
            "tree_method":        "hist",
            "random_state":       42,
            "n_jobs":             -1,
        },
        "signal_threshold": 0.56,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.01,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 18],
        "min_atr":          0.0002,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective":        (
            "Maximize risk-adjusted return (Sharpe/Calmar) via XGBoost with deep "
            "SMA-based trend features (20/50/200), momentum, volatility, RSI, MACD, "
            "Bollinger Bands, Stochastics, CCI, Donchian channels, and bar microstructure. "
            "Regularised tree ensemble (L1+L2, subsampling) prevents overfit on 15-min FX data. "
            "2:1 reward-to-risk with 0.5% SL / 1.0% TP targets consistent positive expectancy. "
            "Session filter [6,18] UTC focuses on liquid London+NY overlap. "
            "Trend filter sma_50 suppresses counter-trend noise."
        ),
        "notes": (
            "n_estimators=500 with low learning_rate=0.03 gives stable generalisation. "
            "max_depth=4 limits tree complexity to avoid overfit on 15-min EURUSD. "
            "min_child_weight=5 and gamma=0.1 add conservative splitting constraints. "
            "reg_lambda=2.0 strong L2 regularisation for stable leaf weights. "
            "colsample_bytree=0.7 adds feature bagging diversity. "
            "target_horizon=4 (1 hour ahead) balances signal frequency and predictability. "
            "signal_threshold=0.56 filters marginal predictions, improving precision."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 01:27:43
# Model            : XGBoost
# Feature Eng.     : SMA (20,50,200) + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/GBPUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── Required SMAs and distance metrics ──────────────────────────────────
    for p in [20, 50, 200]:
        sma = close.rolling(p).mean()
        df[f"sma_{p}"] = sma
        df[f"dm_sma_{p}"] = (close - sma) / sma

    # ── SMA slope (momentum of the moving average itself) ───────────────────
    for p in [20, 50]:
        sma = close.rolling(p).mean()
        df[f"sma_{p}_slope"] = sma.diff(5) / sma.shift(5)

    # ── SMA cross signals ────────────────────────────────────────────────────
    sma20 = close.rolling(20).mean()
    sma50 = close.rolling(50).mean()
    sma200 = close.rolling(200).mean()
    df["sma20_50_cross"] = (sma20 - sma50) / sma50
    df["sma50_200_cross"] = (sma50 - sma200) / sma200
    df["sma20_200_cross"] = (sma20 - sma200) / sma200

    # ── Price momentum over multiple horizons ────────────────────────────────
    for lag in [1, 3, 6, 12, 24, 48]:
        df[f"ret_{lag}"] = close.pct_change(lag)

    # ── Volatility: rolling standard deviation of returns ───────────────────
    ret1 = close.pct_change(1)
    for w in [10, 20, 40]:
        df[f"vol_{w}"] = ret1.rolling(w).std()

    # ── ATR (Average True Range, normalised) ─────────────────────────────────
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low  - close.shift(1)).abs()
    ], axis=1).max(axis=1)
    for w in [14, 28]:
        atr = tr.ewm(span=w, adjust=False).mean()
        df[f"natr_{w}"] = atr / close

    # ── Bollinger Bands (20-period, 2σ) ──────────────────────────────────────
    bb_mid  = close.rolling(20).mean()
    bb_std  = close.rolling(20).std()
    bb_up   = bb_mid + 2 * bb_std
    bb_lo   = bb_mid - 2 * bb_std
    bb_width = (bb_up - bb_lo) / bb_mid
    df["bb_pct_b"]  = (close - bb_lo) / (bb_up - bb_lo + 1e-12)
    df["bb_width"]  = bb_width
    df["bb_squeeze"]= np.where(bb_width < bb_width.rolling(50).mean(), 1.0, 0.0)

    # ── Keltner Channel (for squeeze confirmation) ───────────────────────────
    kc_mid  = close.ewm(span=20, adjust=False).mean()
    kc_atr  = tr.ewm(span=20, adjust=False).mean()
    kc_up   = kc_mid + 1.5 * kc_atr
    kc_lo   = kc_mid - 1.5 * kc_atr
    df["kc_pct"] = (close - kc_lo) / (kc_up - kc_lo + 1e-12)

    # ── RSI (Wilder) ─────────────────────────────────────────────────────────
    def wilder_rsi(src, period):
        delta = src.diff(1)
        gain  = delta.clip(lower=0)
        loss  = (-delta).clip(lower=0)
        avg_g = gain.ewm(alpha=1/period, adjust=False).mean()
        avg_l = loss.ewm(alpha=1/period, adjust=False).mean()
        rs    = avg_g / (avg_l + 1e-12)
        return 100 - 100 / (1 + rs)

    rsi14 = wilder_rsi(close, 14)
    rsi6  = wilder_rsi(close, 6)
    rsi28 = wilder_rsi(close, 28)
    df["rsi14"] = rsi14 / 100.0
    df["rsi6"]  = rsi6  / 100.0
    df["rsi28"] = rsi28 / 100.0
    df["rsi14_slope"] = rsi14.diff(3) / 100.0

    # RSI divergence proxy: price new high/low but RSI doesn't confirm
    price_high_12 = close.rolling(12).max()
    price_low_12  = close.rolling(12).min()
    rsi_high_12   = rsi14.rolling(12).max()
    rsi_low_12    = rsi14.rolling(12).min()
    df["rsi_bear_div"] = np.where(
        (close >= price_high_12 * 0.999) & (rsi14 < rsi_high_12 * 0.97), 1.0, 0.0)
    df["rsi_bull_div"] = np.where(
        (close <= price_low_12  * 1.001) & (rsi14 > rsi_low_12  * 1.03), 1.0, 0.0)

    # ── MACD ─────────────────────────────────────────────────────────────────
    ema12  = close.ewm(span=12, adjust=False).mean()
    ema26  = close.ewm(span=26, adjust=False).mean()
    macd   = ema12 - ema26
    signal = macd.ewm(span=9, adjust=False).mean()
    hist   = macd - signal
    df["macd_norm"]    = macd   / close
    df["macd_sig_norm"]= signal / close
    df["macd_hist_norm"]= hist  / close
    df["macd_hist_slope"] = hist.diff(2) / close

    # ── Stochastic Oscillator ─────────────────────────────────────────────────
    for k_period in [14, 5]:
        lo_k = low.rolling(k_period).min()
        hi_k = high.rolling(k_period).max()
        stoch_k = (close - lo_k) / (hi_k - lo_k + 1e-12) * 100
        stoch_d = stoch_k.rolling(3).mean()
        df[f"stoch_k_{k_period}"] = stoch_k / 100.0
        df[f"stoch_d_{k_period}"] = stoch_d / 100.0
        df[f"stoch_kd_{k_period}"] = (stoch_k - stoch_d) / 100.0

    # ── Williams %R ───────────────────────────────────────────────────────────
    hi14 = high.rolling(14).max()
    lo14 = low.rolling(14).min()
    df["williams_r"] = (hi14 - close) / (hi14 - lo14 + 1e-12)

    # ── CCI (Commodity Channel Index) ────────────────────────────────────────
    tp = (high + low + close) / 3.0
    tp_sma = tp.rolling(20).mean()
    tp_mad  = tp.rolling(20).apply(lambda x: np.mean(np.abs(x - x.mean())), raw=True)
    df["cci"] = (tp - tp_sma) / (0.015 * tp_mad + 1e-12) / 100.0

    # ── Volume-like proxy: candle body and wick ratios ────────────────────────
    candle_range = (high - low).replace(0, np.nan)
    df["body_ratio"]  = (close - open_).abs() / candle_range
    df["upper_wick"]  = (high - pd.concat([close, open_], axis=1).max(axis=1)) / candle_range
    df["lower_wick"]  = (pd.concat([close, open_], axis=1).min(axis=1) - low) / candle_range
    df["bull_candle"]  = np.where(close > open_, 1.0, 0.0)

    # ── Mean reversion signal: z-score of close vs SMA20 ────────────────────
    df["zscore_20"] = (close - sma20) / (close.rolling(20).std() + 1e-12)
    df["zscore_50"] = (close - sma50) / (close.rolling(50).std() + 1e-12)

    # ── Trend strength: ADX proxy ─────────────────────────────────────────────
    plus_dm  = (high.diff(1)).clip(lower=0)
    minus_dm = (-low.diff(1)).clip(lower=0)
    overlap  = pd.concat([plus_dm, minus_dm], axis=1).min(axis=1)
    plus_dm  = plus_dm - overlap
    minus_dm = minus_dm - overlap
    atr14 = tr.ewm(span=14, adjust=False).mean()
    plus_di  = 100 * plus_dm.ewm(span=14, adjust=False).mean()  / (atr14 + 1e-12)
    minus_di = 100 * minus_dm.ewm(span=14, adjust=False).mean() / (atr14 + 1e-12)
    dx = (plus_di - minus_di).abs() / (plus_di + minus_di + 1e-12) * 100
    adx = dx.ewm(span=14, adjust=False).mean()
    df["adx"]      = adx / 100.0
    df["plus_di"]  = plus_di / 100.0
    df["minus_di"] = minus_di / 100.0
    df["di_diff"]  = (plus_di - minus_di) / 100.0

    # ── Regime detection: above/below long-term SMA ──────────────────────────
    df["bull_regime"]  = np.where(close > sma200, 1.0, 0.0)
    df["mid_regime"]   = np.where(close > sma50,  1.0, 0.0)

    # ── Lag features (auto-regressive) ───────────────────────────────────────
    for col, lags in [("rsi14", [1, 2, 4]), ("macd_hist_norm", [1, 2]), ("bb_pct_b", [1, 2])]:
        for lag in lags:
            df[f"{col}_lag{lag}"] = df[col].shift(lag)

    # ── Time-of-day features ─────────────────────────────────────────────────
    if hasattr(df.index, "hour"):
        df["hour_sin"] = np.sin(2 * np.pi * df.index.hour / 24.0)
        df["hour_cos"] = np.cos(2 * np.pi * df.index.hour / 24.0)
        df["dow_sin"]  = np.sin(2 * np.pi * df.index.dayofweek / 5.0)
        df["dow_cos"]  = np.cos(2 * np.pi * df.index.dayofweek / 5.0)

    # ── Fill NaN from warm-up ────────────────────────────────────────────────
    df = df.bfill().ffill()
    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "GBP/USD SMA Trend + Multi-Indicator XGBoost Classifier",
        "model_type":       "XGBClassifier",
        "model_params": {
            "n_estimators":     600,
            "max_depth":        4,
            "learning_rate":    0.03,
            "subsample":        0.75,
            "colsample_bytree": 0.70,
            "min_child_weight": 5,
            "gamma":            0.10,
            "reg_alpha":        0.10,
            "reg_lambda":       1.50,
            "objective":        "binary:logistic",
            "tree_method":      "hist",
            "random_state":     42,
            "n_jobs":           -1,
        },
        "signal_threshold": 0.56,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 18],
        "min_atr":          0.0003,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return on GBP/USD 15-min bars. "
            "Strategy combines required SMA (20/50/200) distance and cross features "
            "with ADX trend strength, RSI divergence, Bollinger squeeze, Keltner, "
            "MACD histogram slope, Stochastic, CCI, Williams %R, and candle-structure "
            "ratios. XGBoost with strong regularisation and subsampling prevents "
            "overfitting on the relatively short 1-year window. "
            "Session filter 06-18 UTC keeps execution in liquid London/NY hours; "
            "0.5% SL and 1.0% TP yield 1:2 R:R; sma_50 trend filter aligns trades "
            "with intermediate momentum to improve win rate and Sharpe."
        ),
        "notes": (
            "Differs from prior RSI/MACD/BB/Stoch attempts by: (1) foregrounding "
            "SMA cross and distance features as primary trend signals; (2) adding "
            "ADX-based regime and DI differential; (3) including RSI divergence "
            "proxy flags; (4) z-score mean-reversion features; (5) candle body/wick "
            "structure ratios as micro-structure proxies; (6) time-of-day cyclical "
            "encoding; (7) heavier regularisation (gamma, alpha, lambda) and higher "
            "min_child_weight to reduce variance on the thin dataset."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 01:47:56
# Model            : Gradient Boosting
# Feature Eng.     : SMA (20,50,200) + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/GBPUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── SMA features (required) ──────────────────────────────────────────────
    for period in [20, 50, 200]:
        sma = close.rolling(period).mean()
        df[f"sma_{period}"] = sma
        df[f"dm_sma_{period}"] = (close - sma) / sma

    # ── SMA crossover signals ────────────────────────────────────────────────
    sma_20 = close.rolling(20).mean()
    sma_50 = close.rolling(50).mean()
    sma_200 = close.rolling(200).mean()

    df["sma_20_50_cross"] = np.where(sma_20 > sma_50, 1.0, -1.0)
    df["sma_20_200_cross"] = np.where(sma_20 > sma_200, 1.0, -1.0)
    df["sma_50_200_cross"] = np.where(sma_50 > sma_200, 1.0, -1.0)

    # ── Price momentum features ──────────────────────────────────────────────
    for lag in [1, 2, 4, 8, 16]:
        df[f"ret_{lag}"] = close.pct_change(lag)

    # ── Volatility: rolling std of returns ──────────────────────────────────
    ret_1 = close.pct_change(1)
    for window in [8, 20, 50]:
        df[f"vol_{window}"] = ret_1.rolling(window).std()

    # ── ATR (Average True Range) ─────────────────────────────────────────────
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low  - close.shift(1)).abs()
    ], axis=1).max(axis=1)

    for atr_period in [14, 50]:
        atr = tr.rolling(atr_period).mean()
        df[f"atr_{atr_period}"] = atr
        df[f"natr_{atr_period}"] = atr / close

    # ── RSI ──────────────────────────────────────────────────────────────────
    for rsi_period in [14, 28]:
        delta = close.diff()
        gain  = delta.clip(lower=0).rolling(rsi_period).mean()
        loss  = (-delta.clip(upper=0)).rolling(rsi_period).mean()
        rs    = gain / (loss + 1e-10)
        df[f"rsi_{rsi_period}"] = 100 - (100 / (1 + rs))

    # ── MACD ─────────────────────────────────────────────────────────────────
    ema_12 = close.ewm(span=12, adjust=False).mean()
    ema_26 = close.ewm(span=26, adjust=False).mean()
    macd_line   = ema_12 - ema_26
    signal_line = macd_line.ewm(span=9, adjust=False).mean()
    df["macd"]           = macd_line
    df["macd_signal"]    = signal_line
    df["macd_hist"]      = macd_line - signal_line
    df["macd_hist_norm"] = (macd_line - signal_line) / (close + 1e-10)

    # ── Bollinger Bands ───────────────────────────────────────────────────────
    for bb_period in [20, 50]:
        bb_mid  = close.rolling(bb_period).mean()
        bb_std  = close.rolling(bb_period).std()
        bb_upper = bb_mid + 2.0 * bb_std
        bb_lower = bb_mid - 2.0 * bb_std
        bb_width = (bb_upper - bb_lower) / (bb_mid + 1e-10)
        bb_pos   = (close - bb_lower) / (bb_upper - bb_lower + 1e-10)
        df[f"bb_width_{bb_period}"] = bb_width
        df[f"bb_pos_{bb_period}"]   = bb_pos

    # ── Stochastic Oscillator ────────────────────────────────────────────────
    for stoch_period in [14, 28]:
        lowest_low   = low.rolling(stoch_period).min()
        highest_high = high.rolling(stoch_period).max()
        stoch_k = (close - lowest_low) / (highest_high - lowest_low + 1e-10) * 100
        stoch_d = stoch_k.rolling(3).mean()
        df[f"stoch_k_{stoch_period}"] = stoch_k
        df[f"stoch_d_{stoch_period}"] = stoch_d

    # ── Rate of Change (ROC) ──────────────────────────────────────────────────
    for roc_period in [5, 10, 20]:
        df[f"roc_{roc_period}"] = close.pct_change(roc_period)

    # ── Candle body and shadow features ──────────────────────────────────────
    body        = (close - open_).abs()
    candle_range = (high - low).abs()
    df["body_ratio"]       = body / (candle_range + 1e-10)
    df["upper_shadow"]     = (high - pd.concat([close, open_], axis=1).max(axis=1)) / (candle_range + 1e-10)
    df["lower_shadow"]     = (pd.concat([close, open_], axis=1).min(axis=1) - low) / (candle_range + 1e-10)
    df["bullish_candle"]   = np.where(close > open_, 1.0, -1.0)

    # ── Volume-proxy: candle range as volatility proxy ────────────────────────
    df["range_norm"] = candle_range / (close + 1e-10)
    df["range_ma_ratio"] = candle_range / (candle_range.rolling(20).mean() + 1e-10)

    # ── Lag features for return predictors ───────────────────────────────────
    for col_lag in ["rsi_14", "macd_hist", "bb_pos_20"]:
        for lag in [1, 2, 3]:
            df[f"{col_lag}_lag{lag}"] = df[col_lag].shift(lag)

    # ── Distance of close from recent high/low ────────────────────────────────
    for lookback in [10, 20, 50]:
        roll_high = high.rolling(lookback).max()
        roll_low  = low.rolling(lookback).min()
        df[f"dist_high_{lookback}"] = (close - roll_high) / (roll_high + 1e-10)
        df[f"dist_low_{lookback}"]  = (close - roll_low)  / (roll_low  + 1e-10)

    # ── Trend strength: ADX proxy ─────────────────────────────────────────────
    adx_period = 14
    tr_adx  = tr.copy()
    plus_dm  = pd.Series(np.where((high.diff() > 0) & (high.diff() > -low.diff()), high.diff(), 0.0), index=close.index)
    minus_dm = pd.Series(np.where((-low.diff() > 0) & (-low.diff() > high.diff()), -low.diff(), 0.0), index=close.index)
    atr_adx  = tr_adx.rolling(adx_period).mean()
    plus_di  = 100 * plus_dm.rolling(adx_period).mean()  / (atr_adx + 1e-10)
    minus_di = 100 * minus_dm.rolling(adx_period).mean() / (atr_adx + 1e-10)
    dx       = (100 * (plus_di - minus_di).abs() / (plus_di + minus_di + 1e-10))
    df["adx"]       = dx.rolling(adx_period).mean()
    df["plus_di"]   = plus_di
    df["minus_di"]  = minus_di

    # ── Fill NaN from indicator warm-up ──────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "GBP/USD SMA Trend Gradient Boosting Risk-Adj",
        "model_type":       "GradientBoostingClassifier",
        "model_params":     {
            "n_estimators":     400,
            "max_depth":        4,
            "learning_rate":    0.04,
            "subsample":        0.75,
            "min_samples_leaf": 20,
            "max_features":     "sqrt",
            "n_iter_no_change": 30,
            "validation_fraction": 0.1,
            "tol":              1e-4,
            "random_state":     42,
        },
        "signal_threshold": 0.57,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 18],
        "min_atr":          0.0002,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe/Calmar) on GBP/USD 15-min data. "
            "GradientBoostingClassifier chosen for its strong bias-variance tradeoff "
            "on medium-sized tabular datasets without needing GPU. "
            "Hyperparameters: moderate depth=4 prevents overfitting, learning_rate=0.04 "
            "with 400 estimators balances convergence vs generalisation, subsample=0.75 "
            "adds stochasticity to reduce variance, min_samples_leaf=20 enforces statistical "
            "significance at each leaf. Early stopping via n_iter_no_change guards against "
            "overfit on the training fold. Signal threshold 0.57 filters marginal signals "
            "to improve precision. SL=0.5%, TP=1.0% gives 1:2 RR. Session filter 6-18 UTC "
            "covers London+NY overlap — highest GBP/USD liquidity and tighter spreads. "
            "sma_50 trend filter ensures we only trade in the direction of medium-term trend, "
            "reducing whipsaw losses. target_horizon=4 bars (1 hour) gives the model enough "
            "time for moves to develop while staying relevant for intraday trading."
        ),
        "notes": (
            "Features: SMA 20/50/200 with distance metrics (core requirement), RSI 14/28, "
            "MACD, Bollinger Bands 20/50, Stochastic 14/28, ATR 14/50, NATR, ROC, ADX, "
            "candle body/shadow ratios, lagged RSI/MACD/BB features, distance from rolling "
            "high/low, SMA crossover signals, multi-lag return features. "
            "All features are backward-looking only (no lookahead bias). "
            "on_opposite=reverse for fast trend-following entries without missing reversals."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:56:13
# Model            : XGBoost
# Feature Eng.     : SMA (20,50,200) + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── SMA core features (required) ──────────────────────────────────────
    for period in [20, 50, 200]:
        sma = close.rolling(period).mean()
        df[f"sma_{period}"]    = sma
        df[f"dm_sma_{period}"] = (close - sma) / sma

    # ── SMA slope (momentum of the moving average itself) ─────────────────
    for period in [20, 50, 200]:
        df[f"sma_{period}_slope"] = df[f"sma_{period}"].diff(5) / df[f"sma_{period}"].shift(5)

    # ── SMA crossover signals ─────────────────────────────────────────────
    df["sma_20_50_cross"]  = df["sma_20"] - df["sma_50"]
    df["sma_50_200_cross"] = df["sma_50"] - df["sma_200"]
    df["sma_20_200_cross"] = df["sma_20"] - df["sma_200"]

    # ── Price vs SMA alignment score (how many SMAs price is above) ───────
    above_20  = np.where(close > df["sma_20"],  1, -1)
    above_50  = np.where(close > df["sma_50"],  1, -1)
    above_200 = np.where(close > df["sma_200"], 1, -1)
    df["sma_alignment"] = (above_20 + above_50 + above_200).astype(float)

    # ── Returns at multiple horizons ──────────────────────────────────────
    for lag in [1, 2, 4, 8, 16]:
        df[f"ret_{lag}"] = close.pct_change(lag)

    # ── Log returns ───────────────────────────────────────────────────────
    df["log_ret_1"] = np.log(close / close.shift(1))
    df["log_ret_4"] = np.log(close / close.shift(4))

    # ── ATR (14-period) ───────────────────────────────────────────────────
    tr1 = high - low
    tr2 = (high - close.shift(1)).abs()
    tr3 = (low  - close.shift(1)).abs()
    true_range = pd.concat([tr1, tr2, tr3], axis=1).max(axis=1)
    atr_14 = true_range.rolling(14).mean()
    df["atr_14"]  = atr_14
    df["natr_14"] = atr_14 / close

    # ── ATR ratio (current TR vs average — volatility burst) ──────────────
    df["atr_ratio"] = true_range / atr_14

    # ── Bollinger Bands (20, 2) ───────────────────────────────────────────
    bb_mid  = close.rolling(20).mean()
    bb_std  = close.rolling(20).std()
    bb_upper = bb_mid + 2.0 * bb_std
    bb_lower = bb_mid - 2.0 * bb_std
    bb_width = (bb_upper - bb_lower) / bb_mid
    df["bb_pct_b"]  = (close - bb_lower) / (bb_upper - bb_lower + 1e-12)
    df["bb_width"]  = bb_width
    df["bb_zscore"] = (close - bb_mid) / (bb_std + 1e-12)

    # ── RSI (14) ──────────────────────────────────────────────────────────
    delta  = close.diff()
    gain   = delta.clip(lower=0).rolling(14).mean()
    loss   = (-delta.clip(upper=0)).rolling(14).mean()
    rs     = gain / (loss + 1e-12)
    df["rsi_14"] = 100 - (100 / (1 + rs))

    # ── RSI momentum ─────────────────────────────────────────────────────
    df["rsi_14_diff"] = df["rsi_14"].diff(2)

    # ── MACD (12, 26, 9) ─────────────────────────────────────────────────
    ema12      = close.ewm(span=12, adjust=False).mean()
    ema26      = close.ewm(span=26, adjust=False).mean()
    macd_line  = ema12 - ema26
    signal_line = macd_line.ewm(span=9, adjust=False).mean()
    df["macd_line"]   = macd_line
    df["macd_signal"] = signal_line
    df["macd_hist"]   = macd_line - signal_line
    df["macd_hist_diff"] = df["macd_hist"].diff(1)

    # ── Stochastic %K / %D (14, 3) ───────────────────────────────────────
    low_14  = low.rolling(14).min()
    high_14 = high.rolling(14).max()
    stoch_k = 100 * (close - low_14) / (high_14 - low_14 + 1e-12)
    stoch_d = stoch_k.rolling(3).mean()
    df["stoch_k"] = stoch_k
    df["stoch_d"] = stoch_d
    df["stoch_kd_diff"] = stoch_k - stoch_d

    # ── Williams %R (14) ─────────────────────────────────────────────────
    df["williams_r"] = -100 * (high_14 - close) / (high_14 - low_14 + 1e-12)

    # ── CCI (20) ─────────────────────────────────────────────────────────
    typical_price = (high + low + close) / 3
    tp_sma  = typical_price.rolling(20).mean()
    tp_mad  = typical_price.rolling(20).apply(lambda x: np.mean(np.abs(x - x.mean())), raw=True)
    df["cci_20"] = (typical_price - tp_sma) / (0.015 * tp_mad + 1e-12)

    # ── Rate of Change (10) ───────────────────────────────────────────────
    df["roc_10"] = (close - close.shift(10)) / (close.shift(10) + 1e-12) * 100

    # ── Volume / candle-body features (no volume data assumed) ────────────
    df["body_size"]   = (close - open_).abs() / (high - low + 1e-12)
    df["upper_wick"]  = (high - np.maximum(close, open_)) / (high - low + 1e-12)
    df["lower_wick"]  = (np.minimum(close, open_) - low)  / (high - low + 1e-12)
    df["is_bullish"]  = np.where(close > open_, 1.0, 0.0)

    # ── Rolling volatility (realised) ────────────────────────────────────
    df["vol_8"]  = df["log_ret_1"].rolling(8).std()
    df["vol_20"] = df["log_ret_1"].rolling(20).std()
    df["vol_ratio"] = df["vol_8"] / (df["vol_20"] + 1e-12)

    # ── High-low range relative to SMA ───────────────────────────────────
    df["hl_range_sma20"] = (high - low) / (df["sma_20"] + 1e-12)

    # ── Price position within recent n-bar range ──────────────────────────
    for n in [8, 20]:
        roll_low  = low.rolling(n).min()
        roll_high = high.rolling(n).max()
        df[f"price_pos_{n}"] = (close - roll_low) / (roll_high - roll_low + 1e-12)

    # ── Lagged returns as autoregressive features ─────────────────────────
    for lag in [1, 2, 3, 4, 8]:
        df[f"lag_ret_{lag}"] = df["ret_1"].shift(lag)

    # ── Distance from 20-bar high/low ────────────────────────────────────
    df["dist_high_20"] = (high.rolling(20).max() - close) / (close + 1e-12)
    df["dist_low_20"]  = (close - low.rolling(20).min())  / (close + 1e-12)

    # ── Fill NaN from warm-up ─────────────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "EUR/USD SMA Trend + Multi-Indicator XGBoost",
        "model_type":       "XGBClassifier",
        "model_params":     {
            "n_estimators":     500,
            "max_depth":        4,
            "learning_rate":    0.03,
            "subsample":        0.8,
            "colsample_bytree": 0.7,
            "min_child_weight": 3,
            "gamma":            0.1,
            "reg_alpha":        0.05,
            "reg_lambda":       1.0,
            "objective":        "binary:logistic",
            "random_state":     42,
            "n_jobs":           -1,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 18],
        "min_atr":          None,
        "trend_filter":     "sma_200",
        "target_horizon":   4,
        "objective":        (
            "Maximize risk-adjusted return (Sharpe/Calmar) on EUR/USD 15-min data. "
            "SMA triple-stack (20/50/200) provides trend context; supplementary "
            "momentum (RSI, MACD, Stochastic), volatility (ATR, BB), and "
            "mean-reversion (CCI, Williams %R) features give the XGBoost model a "
            "rich multi-regime signal set. XGBoost chosen for its ability to rank "
            "feature importance and handle non-linear interactions. Shallow trees "
            "(max_depth=4) with high n_estimators and low learning_rate reduce "
            "overfitting. 2:1 reward:risk (SL 0.5% / TP 1.0%) ensures positive "
            "expectancy even at moderate win rates. Session filter 06-18 UTC "
            "concentrates trades in liquid London/NY overlap."
        ),
        "notes":            (
            "trend_filter=sma_200 aligns trades with the dominant trend — longs "
            "only above the 200-SMA, shorts only below — acting as a regime gate "
            "to suppress counter-trend noise. signal_threshold=0.55 slightly above "
            "0.50 to reduce false positives without starving signal count. "
            "colsample_bytree=0.7 introduces randomisation across features to "
            "decorrelate trees and improve generalisation on forex data."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:35:43
# Model            : Gradient Boosting
# Feature Eng.     : SMA (20,50,200) + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── SMA core features (required) ──────────────────────────────────────
    for period in [20, 50, 200]:
        sma = close.rolling(period).mean()
        df[f"sma_{period}"] = sma
        df[f"dm_sma_{period}"] = (close - sma) / sma

    # ── SMA slope (momentum of the moving average itself) ──────────────────
    for period in [20, 50, 200]:
        df[f"sma_{period}_slope"] = df[f"sma_{period}"].diff(5) / df[f"sma_{period}"].shift(5)

    # ── SMA crossover signals ──────────────────────────────────────────────
    df["sma_20_50_cross"]  = df["sma_20"] - df["sma_50"]
    df["sma_50_200_cross"] = df["sma_50"] - df["sma_200"]
    df["sma_20_200_cross"] = df["sma_20"] - df["sma_200"]

    # Sign of crossover difference (trend direction)
    df["trend_20_50"]  = np.where(df["sma_20_50_cross"]  > 0, 1, -1)
    df["trend_50_200"] = np.where(df["sma_50_200_cross"] > 0, 1, -1)

    # ── Price momentum ────────────────────────────────────────────────────
    for lag in [1, 4, 8, 16, 32]:
        df[f"return_{lag}"] = close.pct_change(lag)

    # ── Volatility features ───────────────────────────────────────────────
    # True Range
    prev_close = close.shift(1)
    tr = pd.concat([
        high - low,
        (high - prev_close).abs(),
        (low  - prev_close).abs()
    ], axis=1).max(axis=1)

    for atr_period in [14, 50]:
        atr = tr.rolling(atr_period).mean()
        df[f"atr_{atr_period}"] = atr
        df[f"natr_{atr_period}"] = atr / close

    # Rolling realised volatility
    log_ret = np.log(close / close.shift(1))
    for vol_period in [20, 50]:
        df[f"realvol_{vol_period}"] = log_ret.rolling(vol_period).std()

    # ── Bollinger Bands (20, 2σ) ──────────────────────────────────────────
    bb_mid = close.rolling(20).mean()
    bb_std = close.rolling(20).std()
    bb_upper = bb_mid + 2 * bb_std
    bb_lower = bb_mid - 2 * bb_std
    df["bb_width"]    = (bb_upper - bb_lower) / bb_mid
    df["bb_pct"]      = (close - bb_lower) / (bb_upper - bb_lower + 1e-12)
    df["bb_position"] = (close - bb_mid) / (bb_std + 1e-12)

    # ── RSI (14) ──────────────────────────────────────────────────────────
    delta = close.diff()
    gain  = delta.clip(lower=0).rolling(14).mean()
    loss  = (-delta.clip(upper=0)).rolling(14).mean()
    rs    = gain / (loss + 1e-12)
    df["rsi_14"] = 100 - (100 / (1 + rs))

    # RSI normalised to [-1, 1]
    df["rsi_14_norm"] = (df["rsi_14"] - 50) / 50

    # ── MACD (12, 26, 9) ─────────────────────────────────────────────────
    ema12 = close.ewm(span=12, adjust=False).mean()
    ema26 = close.ewm(span=26, adjust=False).mean()
    macd_line   = ema12 - ema26
    signal_line = macd_line.ewm(span=9, adjust=False).mean()
    df["macd"]          = macd_line
    df["macd_signal"]   = signal_line
    df["macd_hist"]     = macd_line - signal_line
    df["macd_hist_chg"] = df["macd_hist"].diff()

    # Normalise MACD by price
    df["macd_norm"]   = df["macd"]      / close
    df["macd_hist_norm"] = df["macd_hist"] / close

    # ── Stochastic Oscillator (14, 3) ─────────────────────────────────────
    low14  = low.rolling(14).min()
    high14 = high.rolling(14).max()
    stoch_k = 100 * (close - low14) / (high14 - low14 + 1e-12)
    stoch_d = stoch_k.rolling(3).mean()
    df["stoch_k"]    = stoch_k
    df["stoch_d"]    = stoch_d
    df["stoch_diff"] = stoch_k - stoch_d

    # ── Rate-of-change ────────────────────────────────────────────────────
    for roc_period in [5, 10, 20]:
        df[f"roc_{roc_period}"] = (close - close.shift(roc_period)) / (close.shift(roc_period) + 1e-12)

    # ── Candle body and wick features ─────────────────────────────────────
    body      = (close - open_).abs()
    candle_rng = (high - low).replace(0, np.nan)
    df["body_ratio"]       = body / candle_rng
    df["upper_wick_ratio"] = (high - pd.concat([close, open_], axis=1).max(axis=1)) / candle_rng
    df["lower_wick_ratio"] = (pd.concat([close, open_], axis=1).min(axis=1) - low)  / candle_rng
    df["candle_direction"]  = np.where(close >= open_, 1, -1)

    # ── Time-of-day features (hour) ───────────────────────────────────────
    hour = df.index.hour
    df["hour_sin"] = np.sin(2 * np.pi * hour / 24)
    df["hour_cos"] = np.cos(2 * np.pi * hour / 24)

    # ── Lagged returns as features ────────────────────────────────────────
    for lag in [1, 2, 3, 4]:
        df[f"close_lag_{lag}"] = close.shift(lag)
        df[f"ret_lag_{lag}"]   = log_ret.shift(lag)

    # ── Volume proxy: range-based ─────────────────────────────────────────
    df["range_abs"]  = high - low
    df["range_norm"] = (high - low) / close

    # ── High-Low channel position ─────────────────────────────────────────
    for ch_period in [20, 50]:
        ch_high = high.rolling(ch_period).max()
        ch_low  = low.rolling(ch_period).min()
        df[f"channel_pos_{ch_period}"] = (close - ch_low) / (ch_high - ch_low + 1e-12)

    # ── Fill NaN from warm-up ─────────────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "EUR/USD SMA Trend + Multi-Indicator GBM Scalper",
        "model_type":       "GradientBoostingClassifier",
        "model_params":     {
            "n_estimators":     400,
            "max_depth":        4,
            "learning_rate":    0.04,
            "subsample":        0.75,
            "max_features":     "sqrt",
            "min_samples_leaf": 20,
            "min_samples_split": 40,
            "validation_fraction": 0.1,
            "n_iter_no_change": 30,
            "tol":              1e-4,
            "random_state":     42,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 18],
        "min_atr":          None,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective": (
            "Maximise risk-adjusted return (Sharpe / Calmar). "
            "GradientBoostingClassifier chosen for its strong performance on "
            "tabular financial data with moderate feature sets. "
            "Shallow trees (max_depth=4) with high n_estimators and a low "
            "learning_rate reduce overfitting. subsample=0.75 adds stochastic "
            "regularisation. Early stopping via n_iter_no_change prevents "
            "over-training on the validation split. "
            "SL=0.5%, TP=1.0% gives a 1:2 risk-reward ratio, "
            "targeting positive expectancy even at sub-60% accuracy. "
            "Session filter (06-18 UTC) restricts trading to liquid hours "
            "covering London and New York overlap for EUR/USD. "
            "SMA-50 trend filter ensures trades align with the medium-term "
            "trend, reducing counter-trend noise."
        ),
        "notes": (
            "Features include required SMA(20,50,200) distances and crossovers, "
            "RSI-14, MACD histogram, Bollinger Band position, Stochastic, ATR, "
            "realised volatility, rate-of-change, candle structure ratios, "
            "channel position, lagged returns, and cyclical time encoding. "
            "Target horizon of 4 bars (1 hour) on 15-min data balances "
            "signal frequency with meaningful directional moves."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:04:57
# Model            : XGBoost
# Feature Eng.     : SMA (20,50,200), BB (20,2.0), RSI 14, MACD (12,26,9), ATR 14 + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# EUR/USD Multi-Indicator Momentum + Mean-Reversion (XGBoost)
# ============================================================

# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):

    # ── SMA 20, 50, 200 + distance-from-close ──────────────────────────────
    for period in [20, 50, 200]:
        sma = close.rolling(period).mean()
        df[f"sma_{period}"]    = sma
        df[f"dm_sma_{period}"] = (close - sma) / sma

    # ── Bollinger Bands (20, 2) ─────────────────────────────────────────────
    bb_mid   = close.rolling(20).mean()
    bb_std   = close.rolling(20).std(ddof=0)
    bb_upper = bb_mid + 2.0 * bb_std
    bb_lower = bb_mid - 2.0 * bb_std
    df["bb_mid"]   = bb_mid
    df["bb_upper"] = bb_upper
    df["bb_lower"] = bb_lower
    df["bb_width"] = (bb_upper - bb_lower) / bb_mid
    bb_range       = bb_upper - bb_lower
    df["bb_pct"]   = np.where(bb_range != 0, (close - bb_lower) / bb_range, 0.5)

    # ── RSI 14 ──────────────────────────────────────────────────────────────
    delta  = close.diff()
    gain   = delta.clip(lower=0)
    loss   = (-delta).clip(lower=0)
    avg_gain = gain.ewm(com=13, min_periods=14, adjust=False).mean()
    avg_loss = loss.ewm(com=13, min_periods=14, adjust=False).mean()
    rs         = np.where(avg_loss != 0, avg_gain / avg_loss, 100.0)
    df["rsi_14"] = 100.0 - (100.0 / (1.0 + rs))

    # ── MACD (12, 26, 9) ────────────────────────────────────────────────────
    ema_12 = close.ewm(span=12, adjust=False).mean()
    ema_26 = close.ewm(span=26, adjust=False).mean()
    macd_line   = ema_12 - ema_26
    signal_line = macd_line.ewm(span=9, adjust=False).mean()
    df["macd_line"]   = macd_line
    df["macd_signal"] = signal_line
    df["macd_hist"]   = macd_line - signal_line

    # ── ATR 14 + NATR ───────────────────────────────────────────────────────
    prev_close = close.shift(1)
    tr = pd.concat([
        high - low,
        (high - prev_close).abs(),
        (low  - prev_close).abs()
    ], axis=1).max(axis=1)
    atr = tr.ewm(com=13, min_periods=14, adjust=False).mean()
    df["atr_14"] = atr
    df["natr"]   = np.where(close != 0, atr / close, 0.0)

    # ── Price momentum features ─────────────────────────────────────────────
    for lag in [1, 2, 4, 8]:
        df[f"ret_{lag}"] = close.pct_change(lag)

    # ── Candle body / wick features ─────────────────────────────────────────
    candle_range = (high - low).replace(0, np.nan)
    df["body_ratio"] = np.where(
        candle_range.notna(),
        (close - open_).abs() / candle_range,
        0.0
    )
    df["upper_wick"] = np.where(
        candle_range.notna(),
        (high - pd.concat([close, open_], axis=1).max(axis=1)) / candle_range,
        0.0
    )
    df["lower_wick"] = np.where(
        candle_range.notna(),
        (pd.concat([close, open_], axis=1).min(axis=1) - low) / candle_range,
        0.0
    )
    df["candle_dir"] = np.where(close >= open_, 1.0, -1.0)

    # ── Volume-proxy: normalised range ─────────────────────────────────────
    df["norm_range"] = (high - low) / close.rolling(20).mean()

    # ── RSI derived ─────────────────────────────────────────────────────────
    df["rsi_ob"] = np.where(df["rsi_14"] > 70, 1.0, 0.0)
    df["rsi_os"] = np.where(df["rsi_14"] < 30, 1.0, 0.0)
    df["rsi_mid_cross"] = np.where(df["rsi_14"] > 50, 1.0, -1.0)

    # ── Trend alignment flags ───────────────────────────────────────────────
    df["above_sma_20"]  = np.where(close > df["sma_20"],  1.0, -1.0)
    df["above_sma_50"]  = np.where(close > df["sma_50"],  1.0, -1.0)
    df["above_sma_200"] = np.where(close > df["sma_200"], 1.0, -1.0)
    df["sma_20_50_cross"] = np.where(df["sma_20"] > df["sma_50"], 1.0, -1.0)

    # ── MACD cross flag ─────────────────────────────────────────────────────
    df["macd_cross"] = np.where(df["macd_line"] > df["macd_signal"], 1.0, -1.0)

    # ── Bollinger squeeze ───────────────────────────────────────────────────
    bb_width_ma = df["bb_width"].rolling(20).mean()
    df["bb_squeeze"] = np.where(df["bb_width"] < bb_width_ma, 1.0, 0.0)

    # ── Lagged RSI and MACD hist for regime detection ───────────────────────
    df["rsi_14_lag1"]   = df["rsi_14"].shift(1)
    df["macd_hist_lag1"] = df["macd_hist"].shift(1)
    df["rsi_slope"]     = df["rsi_14"] - df["rsi_14_lag1"]
    df["macd_hist_slope"] = df["macd_hist"] - df["macd_hist_lag1"]

    # ── Rolling volatility ratio ─────────────────────────────────────────────
    vol_short = close.pct_change().rolling(8).std()
    vol_long  = close.pct_change().rolling(32).std()
    df["vol_ratio"] = np.where(vol_long != 0, vol_short / vol_long, 1.0)

    # ── Hour-of-day (session proxy) ─────────────────────────────────────────
    if hasattr(df.index, "hour"):
        df["hour_sin"] = np.sin(2 * np.pi * df.index.hour / 24.0)
        df["hour_cos"] = np.cos(2 * np.pi * df.index.hour / 24.0)
    else:
        df["hour_sin"] = 0.0
        df["hour_cos"] = 1.0

    # ── Fill any NaN from warm-up ───────────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "EUR/USD SMA+RSI+MACD+BB Momentum XGBoost",
        "model_type":       "XGBClassifier",
        "model_params":     {
            "n_estimators":     600,
            "max_depth":        4,
            "learning_rate":    0.04,
            "subsample":        0.8,
            "colsample_bytree": 0.75,
            "min_child_weight": 3,
            "gamma":            0.1,
            "reg_alpha":        0.05,
            "reg_lambda":       1.5,
            "objective":        "binary:logistic",
            "random_state":     42,
            "n_jobs":           -1,
        },
        "signal_threshold": 0.54,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 20],
        "min_atr":          0.0002,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective":        (
            "Maximize risk-adjusted return (Sharpe/Calmar) on EUR/USD 15-min. "
            "XGBoost with moderate depth and strong regularisation to avoid "
            "overfitting on noisy FX data. 2:1 TP:SL ratio with trend filter "
            "on SMA-50 to avoid counter-trend noise. Session filter 06-20 UTC "
            "covers London + NY overlap for best liquidity."
        ),
        "notes": (
            "Features: SMA cross distances, RSI overbought/oversold flags, "
            "MACD histogram slope, Bollinger squeeze, ATR normalisation, "
            "candle body/wick ratios, momentum returns at 1/2/4/8 bars, "
            "session encoding via hour sin/cos. "
            "Threshold 0.54 keeps precision high while capturing enough trades. "
            "Reverse on opposite signal maximises capital utilisation."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:19:50
# Model            : XGBoost
# Feature Eng.     : EMA (50,200), ATR 14 + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── EMA 50 / 200 and distance features ──────────────────────────────
    ema_50  = close.ewm(span=50,  adjust=False).mean()
    ema_200 = close.ewm(span=200, adjust=False).mean()

    df["ema_50"]    = ema_50
    df["ema_200"]   = ema_200
    df["dm_ema_50"]  = (close - ema_50)  / ema_50
    df["dm_ema_200"] = (close - ema_200) / ema_200

    # EMA cross signal: +1 when ema_50 > ema_200, -1 otherwise
    df["ema_cross"] = np.where(ema_50 > ema_200, 1.0, -1.0)

    # Spread between the two EMAs, normalised by price
    df["ema_spread"] = (ema_50 - ema_200) / close

    # Rate of change of ema_spread (momentum of the cross)
    df["ema_spread_roc"] = df["ema_spread"].diff(4)

    # ── ATR 14 & NATR ───────────────────────────────────────────────────
    prev_close = close.shift(1)
    tr = pd.concat([
        high - low,
        (high - prev_close).abs(),
        (low  - prev_close).abs()
    ], axis=1).max(axis=1)

    atr  = tr.ewm(span=14, adjust=False).mean()
    natr = atr / close

    df["atr"]  = atr
    df["natr"] = natr

    # ── RSI 14 ───────────────────────────────────────────────────────────
    delta  = close.diff()
    gain   = delta.clip(lower=0)
    loss   = (-delta).clip(lower=0)
    avg_gain = gain.ewm(span=14, adjust=False).mean()
    avg_loss = loss.ewm(span=14, adjust=False).mean()
    rs  = avg_gain / (avg_loss + 1e-12)
    rsi = 100.0 - (100.0 / (1.0 + rs))
    df["rsi_14"] = rsi
    # Normalised RSI centred around 0
    df["rsi_norm"] = (rsi - 50.0) / 50.0

    # ── MACD (12/26/9) ───────────────────────────────────────────────────
    ema_12   = close.ewm(span=12, adjust=False).mean()
    ema_26   = close.ewm(span=26, adjust=False).mean()
    macd     = ema_12 - ema_26
    signal   = macd.ewm(span=9,  adjust=False).mean()
    df["macd"]        = macd / close
    df["macd_signal"] = signal / close
    df["macd_hist"]   = (macd - signal) / close

    # ── Bollinger Bands (20, 2σ) ─────────────────────────────────────────
    bb_mid  = close.rolling(20).mean()
    bb_std  = close.rolling(20).std(ddof=0)
    bb_up   = bb_mid + 2.0 * bb_std
    bb_lo   = bb_mid - 2.0 * bb_std
    df["bb_width"] = (bb_up - bb_lo) / (bb_mid + 1e-12)
    df["bb_pct"]   = (close - bb_lo)  / (bb_up - bb_lo + 1e-12)

    # ── Price momentum (log-returns at multiple horizons) ────────────────
    for lag in [1, 4, 8, 16]:
        df[f"logret_{lag}"] = np.log(close / close.shift(lag))

    # ── Volume-less volatility proxy: high-low range / ATR ───────────────
    df["hl_range_norm"] = (high - low) / (atr + 1e-12)

    # ── Candle body and shadow ratios ────────────────────────────────────
    body  = (close - open_).abs()
    range_ = (high - low + 1e-12)
    df["body_ratio"]       = body / range_
    df["upper_shadow_ratio"] = (high - pd.concat([close, open_], axis=1).max(axis=1)) / range_
    df["lower_shadow_ratio"] = (pd.concat([close, open_], axis=1).min(axis=1) - low) / range_

    # ── Trend strength: close relative to recent N-bar high/low ──────────
    for window in [20, 50]:
        roll_hi = high.rolling(window).max()
        roll_lo = low.rolling(window).min()
        denom   = (roll_hi - roll_lo + 1e-12)
        df[f"close_rank_{window}"] = (close - roll_lo) / denom

    # ── EMA 50 slope (rate of change) ────────────────────────────────────
    df["ema_50_slope"]  = ema_50.diff(4)  / (close + 1e-12)
    df["ema_200_slope"] = ema_200.diff(8) / (close + 1e-12)

    # ── Hour-of-day and day-of-week as cyclic features ───────────────────
    if hasattr(df.index, "hour"):
        hour = df.index.hour
        dow  = df.index.dayofweek
        df["hour_sin"] = np.sin(2.0 * np.pi * hour / 24.0)
        df["hour_cos"] = np.cos(2.0 * np.pi * hour / 24.0)
        df["dow_sin"]  = np.sin(2.0 * np.pi * dow  / 5.0)
        df["dow_cos"]  = np.cos(2.0 * np.pi * dow  / 5.0)

    # ── Fill NaNs from indicator warm-up ────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "EUR/USD EMA Cross + ATR Momentum (XGBoost)",
        "model_type":       "XGBClassifier",
        "model_params": {
            "n_estimators":        400,
            "max_depth":           4,
            "learning_rate":       0.04,
            "subsample":           0.75,
            "colsample_bytree":    0.75,
            "min_child_weight":    3,
            "gamma":               0.1,
            "reg_alpha":           0.05,
            "reg_lambda":          1.5,
            "objective":           "binary:logistic",
            "random_state":        42,
            "n_jobs":              -1,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 20],
        "min_atr":          0.0002,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe) by combining EMA crossover "
            "trend regime with ATR-normalised volatility, RSI, MACD, and Bollinger "
            "features. XGBoost hyperparameters tuned for bias-variance balance: "
            "moderate depth (4), aggressive shrinkage (lr=0.04), column/row "
            "subsampling, and L1/L2 regularisation. SL=0.5% / TP=1.0% targets a "
            "2:1 reward-risk ratio. Session filter [6,20] UTC focuses on the "
            "liquid London/NY overlap. min_atr filters dead markets."
        ),
        "notes": (
            "EMA 50/200 cross provides the macro trend regime; distance features "
            "capture how far price has stretched from trend. ATR/NATR quantifies "
            "volatility regime. RSI, MACD histogram, and BB %b add mean-reversion "
            "and momentum context. Candle body ratios encode micro-structure. "
            "Cyclic time features allow the model to learn intraday seasonality. "
            "target_horizon=4 bars (1 hour on 15-min data) balances trade frequency "
            "against predictability. on_opposite=reverse reduces idle time and "
            "captures trend continuation efficiently."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 01:39:22
# Model            : XGBoost
# Feature Eng.     : EMA (9,21), RSI 14 + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):
    # ── EMA 9 and EMA 21 ──────────────────────────────────────────────────────
    ema_9  = close.ewm(span=9,  adjust=False).mean()
    ema_21 = close.ewm(span=21, adjust=False).mean()

    df["ema_9"]    = ema_9
    df["ema_21"]   = ema_21
    df["dm_ema_9"]  = (close - ema_9)  / ema_9
    df["dm_ema_21"] = (close - ema_21) / ema_21

    # EMA cross signal: +1 when ema_9 > ema_21, -1 otherwise
    df["ema_cross"] = np.where(ema_9 > ema_21, 1.0, -1.0)

    # EMA cross momentum: difference normalised by ema_21
    df["ema_spread"] = (ema_9 - ema_21) / ema_21

    # Rate of change of EMA spread (1-bar and 3-bar)
    df["ema_spread_chg1"] = df["ema_spread"].diff(1)
    df["ema_spread_chg3"] = df["ema_spread"].diff(3)

    # ── RSI 14 ────────────────────────────────────────────────────────────────
    delta  = close.diff(1)
    gain   = delta.clip(lower=0)
    loss   = (-delta).clip(lower=0)
    avg_gain = gain.ewm(alpha=1/14, min_periods=14, adjust=False).mean()
    avg_loss = loss.ewm(alpha=1/14, min_periods=14, adjust=False).mean()
    rs     = avg_gain / avg_loss.replace(0, np.nan)
    rsi_14 = 100 - (100 / (1 + rs))
    df["rsi_14"] = rsi_14

    # RSI normalised to [-1, 1]
    df["rsi_norm"] = (rsi_14 - 50) / 50

    # RSI momentum
    df["rsi_chg1"] = rsi_14.diff(1)
    df["rsi_chg3"] = rsi_14.diff(3)

    # RSI zone flags (overbought / oversold)
    df["rsi_ob"] = np.where(rsi_14 > 70, 1.0, 0.0)
    df["rsi_os"] = np.where(rsi_14 < 30, 1.0, 0.0)

    # ── ATR 14 (for normalisation & volatility context) ───────────────────────
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low  - close.shift(1)).abs()
    ], axis=1).max(axis=1)
    atr_14 = tr.ewm(span=14, adjust=False).mean()
    df["atr_14"]  = atr_14
    df["natr_14"] = atr_14 / close  # normalised ATR

    # ── Price momentum features ───────────────────────────────────────────────
    for n in [1, 3, 5, 10, 20]:
        df[f"ret_{n}"] = close.pct_change(n)

    # ── Volatility: rolling std of returns ───────────────────────────────────
    ret1 = close.pct_change(1)
    df["vol_5"]  = ret1.rolling(5).std()
    df["vol_20"] = ret1.rolling(20).std()

    # ── Bollinger Band features (20, 2) ───────────────────────────────────────
    bb_mid  = close.rolling(20).mean()
    bb_std  = close.rolling(20).std()
    bb_up   = bb_mid + 2 * bb_std
    bb_lo   = bb_mid - 2 * bb_std
    df["bb_pct"]   = (close - bb_lo) / (bb_up - bb_lo).replace(0, np.nan)
    df["bb_width"] = (bb_up - bb_lo) / bb_mid

    # ── MACD-style: difference of EMA12 and EMA26 ────────────────────────────
    ema_12 = close.ewm(span=12, adjust=False).mean()
    ema_26 = close.ewm(span=26, adjust=False).mean()
    macd   = ema_12 - ema_26
    signal = macd.ewm(span=9, adjust=False).mean()
    df["macd"]          = macd / close
    df["macd_signal"]   = signal / close
    df["macd_hist"]     = (macd - signal) / close

    # ── High-Low channel position ─────────────────────────────────────────────
    hh20 = high.rolling(20).max()
    ll20 = low.rolling(20).min()
    df["hl_pos_20"] = (close - ll20) / (hh20 - ll20).replace(0, np.nan)

    hh5 = high.rolling(5).max()
    ll5 = low.rolling(5).min()
    df["hl_pos_5"]  = (close - ll5)  / (hh5  - ll5).replace(0, np.nan)

    # ── Bar body & shadow features ────────────────────────────────────────────
    body  = (close - open_).abs()
    range_ = (high - low).replace(0, np.nan)
    df["body_ratio"]   = body / range_
    df["upper_shadow"]  = (high - pd.concat([close, open_], axis=1).max(axis=1)) / range_
    df["lower_shadow"]  = (pd.concat([close, open_], axis=1).min(axis=1) - low)  / range_
    df["bull_bar"]     = np.where(close > open_, 1.0, 0.0)

    # ── Rolling correlation: EMA spread vs RSI (captures confluence) ──────────
    df["corr_spread_rsi"] = df["ema_spread"].rolling(10).corr(rsi_14)

    # ── Time-of-day features (hour & minute encoded cyclically) ───────────────
    if hasattr(df.index, "hour"):
        hour = df.index.hour
        df["hour_sin"] = np.sin(2 * np.pi * hour / 24)
        df["hour_cos"] = np.cos(2 * np.pi * hour / 24)
        dow = df.index.dayofweek
        df["dow_sin"]  = np.sin(2 * np.pi * dow / 5)
        df["dow_cos"]  = np.cos(2 * np.pi * dow / 5)

    # ── Fill NaN from indicator warm-up ──────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "EMA Cross (9/21) + RSI Confirmation — XGBoost",
        "model_type":       "XGBClassifier",
        "model_params":     {
            "n_estimators":     400,
            "max_depth":        4,
            "learning_rate":    0.04,
            "subsample":        0.75,
            "colsample_bytree": 0.70,
            "min_child_weight": 5,
            "gamma":            0.15,
            "reg_alpha":        0.10,
            "reg_lambda":       1.50,
            "objective":        "binary:logistic",
            "random_state":     42,
            "n_jobs":           -1,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [6, 20],
        "min_atr":          0.0002,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective":        (
            "Maximize risk-adjusted return (Sharpe / Calmar) on EUR/USD 15-min data. "
            "XGBoost with moderate depth (4) and heavy regularisation (gamma, alpha, lambda) "
            "to avoid overfitting on a 1-year window. "
            "EMA cross provides trend direction; RSI filters against overbought/oversold entries. "
            "Asymmetric TP/SL (2:1) boosts expectancy. Session filter restricts trading to "
            "London + NY overlap (06–20 UTC) where EUR/USD liquidity is highest. "
            "min_atr removes low-volatility bars where spreads erode edge."
        ),
        "notes":            (
            "Features: EMA 9/21 cross & spread, RSI 14, MACD histogram, Bollinger Band %B, "
            "ATR-normalised volatility, price momentum (1/3/5/10/20 bars), rolling vol, "
            "high-low channel position, candlestick body/shadow ratios, cyclical time encoding. "
            "Target horizon = 4 bars (1 hour ahead). Train/test split 70/30 (no leakage). "
            "Cooldown = 0 because on_opposite='reverse' keeps the model always positioned "
            "when high-confidence signals appear."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 03:08:55
# Model            : Gradient Boosting
# Feature Eng.     : SMA (20,50,200), BB (20,2.0), RSI 14, MACD (12,26,9), ATR 14 + Auto-add features: ON
# Signal / Entry   : Enter when model confidence > threshold; exit on opposite signal or SL/TP
# Optimization     : Maximize risk-adjusted return
# Risk Mgmt        : Stop loss 0.5%, Take profit 1.0%
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/EURUSD_15min.parquet"
START_DATE = "2025-04-24"
END_DATE   = "2026-04-24"
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):

    # ── SMA 20, 50, 200 + distance from close ─────────────────────────────
    for p in [20, 50, 200]:
        sma = close.rolling(p).mean()
        df[f"sma_{p}"]    = sma
        df[f"dm_sma_{p}"] = (close - sma) / sma

    # ── Bollinger Bands (20, 2) ────────────────────────────────────────────
    bb_mid   = close.rolling(20).mean()
    bb_std   = close.rolling(20).std(ddof=0)
    bb_upper = bb_mid + 2.0 * bb_std
    bb_lower = bb_mid - 2.0 * bb_std
    df["bb_width"] = (bb_upper - bb_lower) / bb_mid
    denom          = bb_upper - bb_lower
    df["bb_pct"]   = np.where(denom != 0, (close - bb_lower) / denom, 0.5)

    # ── RSI 14 ────────────────────────────────────────────────────────────
    delta    = close.diff()
    gain     = delta.clip(lower=0)
    loss     = (-delta).clip(lower=0)
    avg_gain = gain.ewm(com=13, min_periods=14).mean()
    avg_loss = loss.ewm(com=13, min_periods=14).mean()
    rs       = avg_gain / avg_loss.replace(0, np.nan)
    df["rsi_14"] = 100 - (100 / (1 + rs))

    # ── MACD (12, 26, 9) ──────────────────────────────────────────────────
    ema_fast        = close.ewm(span=12, adjust=False).mean()
    ema_slow        = close.ewm(span=26, adjust=False).mean()
    macd_line       = ema_fast - ema_slow
    signal_line     = macd_line.ewm(span=9, adjust=False).mean()
    df["macd_line"] = macd_line
    df["macd_sig"]  = signal_line
    df["macd_hist"] = macd_line - signal_line

    # ── ATR 14 + NATR ─────────────────────────────────────────────────────
    hl    = high - low
    hc    = (high - close.shift(1)).abs()
    lc    = (low  - close.shift(1)).abs()
    tr    = pd.concat([hl, hc, lc], axis=1).max(axis=1)
    atr   = tr.ewm(com=13, min_periods=14).mean()
    df["atr_14"]  = atr
    df["natr"]    = atr / close

    # ── Price momentum / rate-of-change ───────────────────────────────────
    for p in [4, 8, 16, 32]:
        df[f"roc_{p}"] = close.pct_change(p)

    # ── Candle body and wick features ─────────────────────────────────────
    body          = (close - open_).abs()
    candle_range  = (high - low).replace(0, np.nan)
    df["body_ratio"]   = body / candle_range
    df["upper_wick"]   = (high - pd.concat([close, open_], axis=1).max(axis=1)) / candle_range
    df["lower_wick"]   = (pd.concat([close, open_], axis=1).min(axis=1) - low)  / candle_range
    df["body_dir"]     = np.sign(close - open_)

    # ── Volume-normalised (uses candle range as proxy if no volume col) ───
    # Rolling z-score of close
    roll_mean = close.rolling(20).mean()
    roll_std  = close.rolling(20).std(ddof=0).replace(0, np.nan)
    df["close_zscore_20"] = (close - roll_mean) / roll_std

    # ── RSI divergence proxy ──────────────────────────────────────────────
    df["rsi_delta_4"]  = df["rsi_14"].diff(4)
    df["price_delta_4"] = close.pct_change(4)
    df["rsi_price_div"] = df["rsi_delta_4"] - (df["price_delta_4"] * 100)

    # ── MACD histogram slope ──────────────────────────────────────────────
    df["macd_hist_slope"] = df["macd_hist"].diff(2)

    # ── SMA crossover signals ─────────────────────────────────────────────
    df["sma20_vs_50"]   = np.where(df["sma_20"] > df["sma_50"],  1.0, -1.0)
    df["sma50_vs_200"]  = np.where(df["sma_50"] > df["sma_200"], 1.0, -1.0)

    # ── Volatility regime (ATR percentile proxy) ──────────────────────────
    atr_roll_min = atr.rolling(96).min()
    atr_roll_max = atr.rolling(96).max()
    atr_range    = (atr_roll_max - atr_roll_min).replace(0, np.nan)
    df["atr_pctile_96"] = (atr - atr_roll_min) / atr_range

    # ── Stochastic oscillator %K, %D ──────────────────────────────────────
    low_14  = low.rolling(14).min()
    high_14 = high.rolling(14).max()
    stoch_k = 100 * (close - low_14) / (high_14 - low_14).replace(0, np.nan)
    stoch_d = stoch_k.rolling(3).mean()
    df["stoch_k"] = stoch_k
    df["stoch_d"] = stoch_d
    df["stoch_diff"] = stoch_k - stoch_d

    # ── Rolling high/low breakout distance ────────────────────────────────
    df["dist_hi_20"] = (high.rolling(20).max() - close) / close
    df["dist_lo_20"] = (close - low.rolling(20).min())  / close

    # ── Hour-of-day and day-of-week cyclical features ─────────────────────
    hour = pd.Series(df.index.hour,        index=df.index, dtype=float)
    dow  = pd.Series(df.index.dayofweek,   index=df.index, dtype=float)
    df["hour_sin"] = np.sin(2 * np.pi * hour / 24)
    df["hour_cos"] = np.cos(2 * np.pi * hour / 24)
    df["dow_sin"]  = np.sin(2 * np.pi * dow  / 5)
    df["dow_cos"]  = np.cos(2 * np.pi * dow  / 5)

    # ── Fill NaN from indicator warm-up ───────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "EUR/USD Gradient Boost SMA+RSI+MACD Swing",
        "model_type":       "GradientBoostingClassifier",
        "model_params":     {
            "n_estimators":       400,
            "max_depth":          4,
            "learning_rate":      0.04,
            "subsample":          0.75,
            "min_samples_leaf":   20,
            "min_samples_split":  30,
            "max_features":       "sqrt",
            "validation_fraction": 0.1,
            "n_iter_no_change":   30,
            "tol":                1e-4,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [7, 18],
        "min_atr":          0.0002,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective":        (
            "Maximize risk-adjusted return (Sharpe / Calmar) on EUR/USD 15-min. "
            "GradientBoostingClassifier chosen for robustness to noisy FX features and "
            "good probability calibration. Shallow trees (depth 4), high n_estimators with "
            "early stopping prevent overfitting. Subsample=0.75 adds stochasticity. "
            "SL=0.5%, TP=1.0% gives 1:2 R:R. Session filter 07-18 UTC captures London+NY overlap. "
            "min_atr filters out flat/illiquid periods. sma_50 trend filter aligns trades with "
            "medium-term momentum. Threshold 0.55 balances precision vs recall."
        ),
        "notes":            (
            "Features: SMA(20,50,200) with distance ratios, BB(20,2) width+pct, RSI-14, "
            "MACD(12,26,9) line/signal/hist + slope, ATR-14 + NATR, ROC(4,8,16,32), "
            "candle body/wick ratios, close z-score, RSI-price divergence proxy, "
            "stochastic %K/%D, rolling high/low breakout distances, "
            "SMA crossover flags, ATR percentile regime, hour/DOW cyclical encodings. "
            "on_opposite=reverse means a strong counter-signal immediately flips the position, "
            "reducing idle time and capturing reversals within the London-NY session."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-25 02:29:29
# Model            : XGBoost
# Feature Eng.     : buy when price closes below the lower Bollinger Band(20,2) and RSI(14) < 35, exit at the middle band + Auto-add features: ON
# Signal / Entry   : —
# Optimization     : —
# Risk Mgmt        : —
# Risk Filter      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# Bollinger Band Mean-Reversion + RSI Filter (XGBoost, Sharpe)
# ============================================================

# SECTION 0 — IMPORTS & CONSTANTS
import numpy as np
import pandas as pd

# ── Inlined strategy_utils ──
"""
strategy_utils.py — Standard utility functions for generated strategies.
Claude imports these instead of writing boilerplate from scratch.
This ensures consistent behavior across all generated strategies.
"""
import numpy as np
import pandas as pd
from sklearn.preprocessing import LabelEncoder


# Max backtest window per timeframe. A finer timeframe over a longer window
# blows up the results dict / parquet load / Modal train time (the 2026-05-12
# OOM was a 1-min × multi-year sweep) — and a 1-min strategy gains nothing from
# 2 years of 1-min bars. Enforced HERE because every training path (UI / API /
# Modal) funnels through run_strategy → load_ohlc. Env-overridable so a future
# "max plan" / dedicated-server tier can lift it.
_TF_MAX_DAYS = {
    "1min":  30,
    "5min":  90,
    "15min": 365,
    "1h":    730,
}


def _fetch_ohlc_from_internal(symbol: str, tf: str, start: str, end: str):
    """Phase 3.2: fetch parquet bytes from Server A's /internal/ohlc endpoint
    instead of reading a local file. Used inside Modal containers / Mac worker
    pool (Phase 3.4) so every train sees the same source of truth as the chart.

    Returns: pd.DataFrame (parquet decoded), or raises on any failure so the
    caller can fall back / surface a clear error in the job.
    """
    import hashlib as _hashlib, hmac as _hmac, io as _io, os as _os
    import urllib.request as _ur, urllib.parse as _urp

    base = (_os.environ.get("QM_INTERNAL_OHLC_BASE") or "").rstrip("/")
    secret = (_os.environ.get("INTERNAL_WS_SECRET") or "").strip()
    if not base:
        raise RuntimeError("QM_INTERNAL_OHLC_BASE not set")
    if not secret:
        raise RuntimeError("INTERNAL_WS_SECRET not set")

    msg = f"{symbol}|{tf}|{start}|{end}".encode("utf-8")
    sig = _hmac.new(secret.encode("utf-8"), msg, _hashlib.sha256).hexdigest()

    qs = _urp.urlencode({
        "symbol": symbol, "tf": tf,
        "start": start, "end": end, "sig": sig,
    })
    url = f"{base}/internal/ohlc?{qs}"

    req = _ur.Request(url, headers={"User-Agent": "qm-worker/1.0"})
    with _ur.urlopen(req, timeout=30) as resp:
        if resp.status != 200:
            raise RuntimeError(f"/internal/ohlc returned {resp.status}")
        payload = resp.read()
    print(f"[load_ohlc:internal] {symbol} {tf} fetched {len(payload)} bytes", flush=True)
    return pd.read_parquet(_io.BytesIO(payload))


def _parse_symbol_tf_from_path(data_path: str):
    """Pull SYMBOL + TF out of a path like .../EURUSD_1min.parquet."""
    import os as _os, re as _re
    base = _os.path.basename(str(data_path))
    m = _re.match(r"^([A-Z]{6})_(\d+min|\d+h)\.parquet$", base)
    if not m:
        return None, None
    return m.group(1), m.group(2)


def load_ohlc(data_path, start_date="", end_date=""):
    """Load OHLC parquet, sort index, filter dates. Always returns consistent format.

    The lower bound is clamped per timeframe (see _TF_MAX_DAYS) — a request for
    more history than the cap silently starts later.

    Phase 3.2: when env QM_USE_INTERNAL_OHLC=="1", fetch over HTTP from
    Server A's /internal/ohlc endpoint instead of pd.read_parquet on a local
    file (which on Modal is a stale Volume snapshot). The endpoint applies the
    same day-cap, so the local cap-check below is a defensive no-op in that
    path. Flag defaults to "0" → unchanged behavior.

    Returns: (df, close, open_, high, low)
    """
    import os as _os, re as _re

    _use_internal = _os.environ.get("QM_USE_INTERNAL_OHLC", "0") == "1"
    if _use_internal:
        _sym, _tf = _parse_symbol_tf_from_path(data_path)
        if not _sym or not _tf:
            raise RuntimeError(
                f"QM_USE_INTERNAL_OHLC=1 but DATA_PATH basename does not match "
                f"SYMBOL_TF.parquet: {data_path}"
            )
        df = _fetch_ohlc_from_internal(_sym, _tf, start_date or "", end_date or "")
    else:
        df = pd.read_parquet(data_path)
    df.index = pd.to_datetime(df.index)
    df = df.sort_index()

    # Per-timeframe window cap (timeframe inferred from the parquet filename).
    _m = _re.search(r"_(\d+min|\d+h)\.parquet$", _os.path.basename(str(data_path)))
    _tf = _m.group(1) if _m else None
    _max_days = _TF_MAX_DAYS.get(_tf)
    if _max_days and _max_days > 0 and len(df):
        _env_override = _os.environ.get(f"QM_MAX_DAYS_{_tf.upper()}")
        if _env_override and _env_override.isdigit():
            _max_days = int(_env_override)
        try:
            _eff_end = pd.Timestamp(end_date) if end_date else df.index.max()
            _eff_end = min(_eff_end, df.index.max())
            _floor = _eff_end - pd.Timedelta(days=_max_days)
            _req_start = pd.Timestamp(start_date) if start_date else df.index.min()
            if _req_start < _floor:
                print(f"[load_ohlc] {_tf} backtest window capped to {_max_days}d: "
                      f"start {_req_start.date()} -> {_floor.date()}", flush=True)
                start_date = _floor
        except Exception as _e:
            print(f"[load_ohlc] window-cap check skipped ({_e})", flush=True)

    if start_date:
        df = df[df.index >= start_date]
    if end_date:
        df = df[df.index <= end_date]

    return df, df["close"], df["open"], df["high"], df["low"]


def make_target(close, horizon=4):
    """Create target: direction N bars ahead. Default 4 bars = 1 hour on 15-min data.

    Returns: target (pd.Series of -1, 0, 1)
    """
    return np.sign(close.shift(-horizon) - close)


def split_data(df, target, feature_cols, train_split=0.7, validation_date=""):
    """Train/test split. Handles both ratio and date-based splits.
    Drops NaN from target before splitting. Encodes labels to [0,1,2].

    Returns: dict with keys:
        X_train, X_test, y_train, y_test,
        y_train_enc, y_test_enc, enc,
        close_train, close_test,
        split_idx, split_dt, n_train, n_test
    """
    # Drop NaN from target
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]

    close = df["close"]

    # Build feature matrix
    X = df[feature_cols].copy()
    X = X.bfill().ffill()
    X = X.replace([np.inf, -np.inf], np.nan).fillna(0.0)

    # Split
    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)

    split_idx = max(1, min(split_idx, len(df) - 1))

    X_train = X.iloc[:split_idx]
    X_test = X.iloc[split_idx:]
    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]

    split_dt = str(df.index[split_idx])

    # Label encoding — always fit on [-1, 0, 1]
    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    return {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }


def compute_overlays(close, df_index):
    """Compute BB and MA overlays on full dataset. Always consistent.

    Returns: (bb_dict, ma_dict)
    """
    bb_mid = close.rolling(20).mean()
    bb_std = close.rolling(20).std()
    bb_upper = bb_mid + 2 * bb_std
    bb_lower = bb_mid - 2 * bb_std

    ma50 = close.rolling(50).mean()
    ma100 = close.rolling(100).mean()
    ma200 = close.rolling(200).mean()

    def _safe(s):
        s = s.reindex(df_index).bfill().ffill()
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in s.values]

    bb = {"upper": _safe(bb_upper), "mid": _safe(bb_mid), "lower": _safe(bb_lower)}
    ma = {"ma50": _safe(ma50), "ma100": _safe(ma100), "ma200": _safe(ma200)}

    return bb, ma


def run_backtest(signal, close, capital=10000, cost=2e-5):
    """Run backtest with transaction costs.
    Uses price-based trade returns (same as webapp _compute_trades).
    Signal 0 = hold (keep current position), not close.

    Returns: dict with equity, trade_returns, long_returns, short_returns, bar_returns
    """
    sig_arr = signal.values
    price_arr = close.values
    idx = signal.index
    n = len(price_arr)

    # Trade returns — price-based (matches webapp _compute_trades exactly)
    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    last_dir = None
    entry_price = None
    entry_bar = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != last_dir:
            # Direction change — close previous trade, open new
            if last_dir is not None and entry_price is not None and entry_price != 0:
                ret = float(last_dir * (c - entry_price) / entry_price - cost)
                trade_returns.append(ret)
                if last_dir == 1:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                trade_log.append({
                    "type": "Buy" if last_dir == 1 else "Sell",
                    "entry_time": str(idx[entry_bar]),
                    "exit_time": str(idx[i]),
                    "entry_price": round(entry_price, 5),
                    "exit_price": round(c, 5),
                    "pnl": round(last_dir * (c - entry_price), 5),
                    "pnl_pct": round(ret * 100, 3),
                    "exit_reason": "signal",
                })
            entry_price = c
            entry_bar = i
            last_dir = s

    # Close last open trade
    if last_dir is not None and entry_price is not None and n > 0 and entry_price != 0:
        c = price_arr[-1]
        ret = float(last_dir * (c - entry_price) / entry_price - cost)
        trade_returns.append(ret)
        if last_dir == 1:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        trade_log.append({
            "type": "Buy" if last_dir == 1 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[-1]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(c, 5),
            "pnl": round(last_dir * (c - entry_price), 5),
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": "end",
        })

    # Equity curve from trade returns
    cumret = 1.0
    equity_vals = np.full(n, float(capital))
    trade_idx = 0
    in_trade = False
    t_entry_price = None
    t_dir = None

    for i in range(n):
        s = sig_arr[i]
        c = price_arr[i]
        if s != 0.0 and s != t_dir:
            if t_dir is not None and t_entry_price is not None and t_entry_price != 0:
                t_ret = t_dir * (c - t_entry_price) / t_entry_price - cost
                cumret *= (1 + t_ret)
            t_entry_price = c
            t_dir = s
        equity_vals[i] = capital * cumret

    # Bar returns for Sharpe
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if price_arr[i - 1] != 0 and last_dir is not None:
            bar_returns[i] = sig_arr[i - 1] * (price_arr[i] - price_arr[i - 1]) / price_arr[i - 1] if sig_arr[i - 1] != 0 else 0.0

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


def compute_trade_stats(trades, capital=10000):
    """Single source of truth for trade statistics.
    Every display path reads from this — no recomputation anywhere.
    All values are rounded and JSON-safe (no inf/nan).
    """
    if not trades:
        return {"n": 0, "wins": 0, "losses": 0, "wr": 0, "avg": 0,
                "best": 0, "worst": 0, "ret": 0, "np": 0, "mdd": 0,
                "pf": 0, "rr": 0, "expect": 0}
    w = [r for r in trades if r > 0]
    l = [r for r in trades if r < 0]
    cumret = 1.0
    for r in trades:
        cumret *= (1 + r)
    net_p = capital * (cumret - 1)
    # Max drawdown
    eq = np.cumprod([1.0] + [1 + r for r in trades])
    peak = np.maximum.accumulate(eq)
    mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    # Profit Factor
    gross_w = sum(w) if w else 0
    gross_l = abs(sum(l)) if l else 0
    pf = gross_w / gross_l if gross_l > 0 else (9999.0 if gross_w > 0 else 0)
    # Risk:Reward
    avg_w = float(np.mean(w)) if w else 0
    avg_l = abs(float(np.mean(l))) if l else 0
    rr = avg_w / avg_l if avg_l > 0 else (9999.0 if avg_w > 0 else 0)
    # Expectancy
    expect = net_p / len(trades)
    return {
        "n": len(trades), "wins": len(w), "losses": len(l),
        "wr": round(len(w) / len(trades), 4),
        "avg": round(float(np.mean(trades)), 6),
        "best": round(max(w), 6) if w else 0,
        "worst": round(min(l), 6) if l else 0,
        "ret": round(cumret - 1, 6),
        "np": round(net_p, 2),
        "mdd": round(mdd, 6),
        "pf": round(pf, 2),
        "rr": round(rr, 2),
        "expect": round(expect, 2),
    }


def compute_metrics(bt_result, close_test, capital=10000):
    """Compute all standard metrics from backtest result.
    Uses trade-level compounding (same as webapp _trade_stats) for accuracy.

    Returns: dict with total_ret, bh_ret, sharpe_strat, sharpe_bh, mdd, n_trades
    """
    equity = bt_result["equity"]
    trade_returns = bt_result["trade_returns"]

    # Total return — trade-level compounding (matches webapp)
    if trade_returns:
        cumret = 1.0
        for r in trade_returns:
            cumret *= (1 + r)
        total_ret = cumret - 1
    else:
        total_ret = 0.0

    # Buy and hold
    bh_equity = capital * (close_test / close_test.iloc[0])
    bh_ret = (bh_equity.iloc[-1] - capital) / capital if capital != 0 else 0.0

    # Sharpe ratio — trade-level (matches webapp: sqrt(252*26) annualization)
    if len(trade_returns) >= 2 and float(np.std(trade_returns)) > 0:
        sharpe_strat = float(np.mean(trade_returns) / np.std(trade_returns) * np.sqrt(252 * 26))
    else:
        sharpe_strat = 0.0

    bh_rets = bh_equity.pct_change().dropna()
    if len(bh_rets) > 1 and bh_rets.std() != 0:
        sharpe_bh = float((bh_rets.mean() / bh_rets.std()) * np.sqrt(252 * 24 * 4))
    else:
        sharpe_bh = 0.0

    # Max drawdown — trade-level (matches webapp)
    if trade_returns:
        eq = np.cumprod([1.0] + [1 + r for r in trade_returns])
        peak = np.maximum.accumulate(eq)
        mdd = float(((eq - peak) / peak).min()) if len(eq) > 1 else 0.0
    else:
        mdd = 0.0

    return {
        "total_ret": float(total_ret),
        "bh_ret": float(bh_ret),
        "sharpe_strat": float(sharpe_strat) if not np.isnan(sharpe_strat) else 0.0,
        "sharpe_bh": float(sharpe_bh) if not np.isnan(sharpe_bh) else 0.0,
        "mdd": float(mdd),
        "n_trades": len(trade_returns),
    }


# Diagnostics line/histogram series (equity / drawdown / rolling_acc / conf_hist)
# only feed the small Diagnostics charts — they're never used by the price chart
# or scroll-back. On a 1-min model trained over the (2.2-capped) window these are
# still ~30k points each; downsample to a visually-identical resolution before the
# dict leaves the trainer so it doesn't carry that into Server-A RAM / Postgres.
_RESULTS_SERIES_MAX = 5000


def _downsample_idx(n, cap=_RESULTS_SERIES_MAX):
    """Evenly-spaced index list spanning [0, n-1] (first+last always kept), or
    None when no downsampling is needed (n <= cap)."""
    if n <= cap:
        return None
    return np.unique(np.linspace(0, n - 1, cap).astype(int)).tolist()


def _take(arr, idx):
    """Subset a list by an index list (idx may be None → return arr unchanged)."""
    if idx is None or not isinstance(arr, list):
        return arr
    return [arr[i] for i in idx]


# trade_log / train_trade_log are lists of per-trade dicts (display-only — the
# Trade Log tab). They scale with TRADE count, not bar count, so the bar-window
# cap (Phase 2.2) doesn't bound them — a degenerate near-every-bar model can put
# 10k+ trade dicts in the blob (>3 MB). Cap each (independently — a small-N model
# keeps every trade) to the most-recent N, recording `*_total` + `*_truncated`
# so the true count is still reported. Real strategies have far fewer than
# _TRADE_LOG_MAX trades, so this only ever bites pathological models.
_TRADE_LOG_MAX = 5000


def _cap_trade_log(tl):
    """Return (capped_list, original_len, was_truncated)."""
    if not isinstance(tl, list) or len(tl) <= _TRADE_LOG_MAX:
        return tl, (len(tl) if isinstance(tl, list) else 0), False
    return tl[-_TRADE_LOG_MAX:], len(tl), True


def build_return_dict(split_result, bt_result, metrics, model, feature_cols,
                      signal_full, p_pos_test, p_neg_test, custom_figs=None,
                      bt_train_result=None, pre_stats=None):
    """Assemble the complete return dict. Handles ALL serialization.
    Never returns Timestamps, numpy arrays, or non-JSON types.

    Returns: JSON-safe dict with all required keys
    """
    df = split_result["df"]
    close = split_result["close"]
    close_test = split_result["close_test"]
    X_test = split_result["X_test"]
    y_test = split_result["y_test"]
    equity = bt_result["equity"]
    bar_returns = bt_result["bar_returns"]

    # OHLC
    ohlc_dates = [str(x) for x in df.index.tolist()]

    def _safe_list(arr):
        return [float(x) if (x is not None and not np.isnan(x) and not np.isinf(x)) else None
                for x in arr]

    # Overlays
    bb, ma = compute_overlays(close, df.index)

    # Buy and hold equity
    capital = equity.iloc[0] if len(equity) > 0 else 10000
    bh_equity = capital * (close_test / close_test.iloc[0])

    # Confusion matrix
    from sklearn.metrics import confusion_matrix
    pred_test = model.predict(X_test)
    y_test_arr = np.asarray(y_test)
    cm = confusion_matrix(y_test_arr, pred_test, labels=[-1, 0, 1])

    # Rolling accuracy
    sig_arr = signal_full.reindex(close_test.index).values
    correct = pd.Series((pred_test == y_test_arr).astype(float), index=X_test.index)
    active_test = pd.Series(sig_arr != 0, index=close_test.index) if len(sig_arr) == len(close_test) else pd.Series(True, index=close_test.index)
    correct_active = correct.where(active_test, other=np.nan)
    rolling_acc = correct_active.rolling(30, min_periods=1).mean()

    # Feature importance
    importances = model.feature_importances_
    fi_pairs = sorted(zip(feature_cols, importances), key=lambda x: x[1])[-15:]

    # Drawdown
    rolling_max = equity.cummax()
    drawdown = (equity - rolling_max) / rolling_max.replace(0, np.nan)
    drawdown = drawdown.fillna(0.0)

    # ── Downsample the Diagnostics-only series (see _downsample_idx) ──────────
    _eq_dates = [str(x) for x in close_test.index.tolist()]
    _eq_strat = _safe_list(equity.values)
    _eq_bh    = _safe_list(bh_equity.values)
    _eq_idx   = _downsample_idx(len(_eq_dates))
    _eq_dates, _eq_strat, _eq_bh = _take(_eq_dates, _eq_idx), _take(_eq_strat, _eq_idx), _take(_eq_bh, _eq_idx)

    _ra_dates = [str(x) for x in rolling_acc.index.tolist()]
    _ra_vals  = [float(x) if (not np.isnan(x) and not np.isinf(x)) else None for x in rolling_acc.values]
    _ra_idx   = _downsample_idx(len(_ra_dates))
    _ra_dates, _ra_vals = _take(_ra_dates, _ra_idx), _take(_ra_vals, _ra_idx)

    _dd_dates = [str(x) for x in drawdown.index.tolist()]
    _dd_vals  = _safe_list(drawdown.values)
    _dd_idx   = _downsample_idx(len(_dd_dates))
    _dd_dates, _dd_vals = _take(_dd_dates, _dd_idx), _take(_dd_vals, _dd_idx)

    _cp_pos = [float(x) for x in (p_pos_test.tolist() if hasattr(p_pos_test, 'tolist') else list(p_pos_test))]
    _cp_neg = [float(x) for x in (p_neg_test.tolist() if hasattr(p_neg_test, 'tolist') else list(p_neg_test))]
    _cp_pos = _take(_cp_pos, _downsample_idx(len(_cp_pos)))
    _cp_neg = _take(_cp_neg, _downsample_idx(len(_cp_neg)))

    # ── Trade logs — display-only (Trade Log tab); cap to most-recent N with a
    #    `_total` field so the true count is still reported (see _cap_trade_log).
    #    NB: ret_dist arrays are left FULL — a downstream path in callbacks.py
    #    recomputes n_trades/win-rate from len(ret_dist), so a sample would skew
    #    the displayed counts; they're small anyway and gzip handles them.
    _tl_test, _tl_test_n, _tl_test_tr = _cap_trade_log(bt_result.get("trade_log", []))
    _tl_tr,   _tl_tr_n,   _tl_tr_tr   = _cap_trade_log(bt_train_result.get("trade_log", []) if bt_train_result else [])

    return {
        "ohlc": {
            "dates": ohlc_dates,
            "open": _safe_list(df["open"].values),
            "high": _safe_list(df["high"].values),
            "low": _safe_list(df["low"].values),
            "close": _safe_list(df["close"].values),
        },
        "signals": {
            "dates": [str(x) for x in signal_full.index.tolist()],
            "values": [float(x) for x in signal_full.values],
        },
        "bb": bb,
        "ma": ma,
        "equity": {
            "dates": _eq_dates,
            "strategy": _eq_strat,
            "bh": _eq_bh,
        },
        "feature_importance": {
            "names": [p[0] for p in fi_pairs],
            "values": [float(p[1]) for p in fi_pairs],
        },
        "conf_matrix": cm.tolist(),
        "conf_hist": {
            "p_pos": _cp_pos,
            "p_neg": _cp_neg,
        },
        "rolling_acc": {
            "dates": _ra_dates,
            "values": _ra_vals,
        },
        "drawdown": {
            "dates": _dd_dates,
            "values": _dd_vals,
        },
        "ret_dist": [float(x) for x in bt_result["trade_returns"]],
        "ret_dist_long": [float(x) for x in bt_result["long_returns"]],
        "ret_dist_short": [float(x) for x in bt_result["short_returns"]],
        "train_ret_dist": [float(x) for x in bt_train_result["trade_returns"]] if bt_train_result else [],
        "train_ret_dist_long": [float(x) for x in bt_train_result["long_returns"]] if bt_train_result else [],
        "train_ret_dist_short": [float(x) for x in bt_train_result["short_returns"]] if bt_train_result else [],
        "trade_log": _tl_test,
        "train_trade_log": _tl_tr,
        "trade_log_total": _tl_test_n,
        "train_trade_log_total": _tl_tr_n,
        "trade_log_truncated": _tl_test_tr,
        "train_trade_log_truncated": _tl_tr_tr,
        **(pre_stats or {}),
        "metrics": metrics,
        "split_dt": split_result["split_dt"],
        "split_idx": int(split_result["split_idx"]),
        "n_train": int(split_result["n_train"]),
        "n_test": int(split_result["n_test"]),
        "feature_cols": list(feature_cols),
        "custom_figs": custom_figs or [],
    }


# ════════════════════════════════════════════════════════════════════════════
# STRATEGY FRAMEWORK v2 — Config-driven architecture
# Claude writes feature_engineering() + strategy_config(). Framework does rest.
# ════════════════════════════════════════════════════════════════════════════

import importlib

_MODEL_REGISTRY = {
    "XGBClassifier":              ("xgboost", "XGBClassifier"),
    "RandomForestClassifier":     ("sklearn.ensemble", "RandomForestClassifier"),
    "GradientBoostingClassifier": ("sklearn.ensemble", "GradientBoostingClassifier"),
    "LogisticRegression":         ("sklearn.linear_model", "LogisticRegression"),
    "ExtraTreesClassifier":       ("sklearn.ensemble", "ExtraTreesClassifier"),
    "AdaBoostClassifier":         ("sklearn.ensemble", "AdaBoostClassifier"),
}


def _build_model_from_config(config, X_train, y_train_enc):
    """Build, fit, and wrap a model from strategy_config dict."""
    model_type = config.get("model_type", "RandomForestClassifier")
    model_params = dict(config.get("model_params", {}))

    if model_type not in _MODEL_REGISTRY:
        raise ValueError(f"Unknown model_type '{model_type}'. Valid: {list(_MODEL_REGISTRY.keys())}")

    module_path, class_name = _MODEL_REGISTRY[model_type]
    mod = importlib.import_module(module_path)
    cls = getattr(mod, class_name)

    # XGBoost defaults
    if class_name == "XGBClassifier":
        model_params.setdefault("use_label_encoder", False)
        model_params.setdefault("eval_metric", "mlogloss")
        model_params.setdefault("tree_method", "hist")
        # Determinism > speed (2026-05-25). XGBoost hist with n_jobs=-1 is
        # NON-reproducible even with random_state set — the parallel histogram
        # gradient-sum order varies across threads, so the SAME code + data
        # gives a slightly different model (and backtest) every run. Forcing
        # single-thread makes training bit-reproducible so: (a) a user who
        # copies a strategy and reruns it gets identical numbers, (b) the
        # community "Live" score matches a redeploy, (c) "same code, different
        # result" support reports go away. Cost: single-threaded XGB (a few
        # seconds slower on large windows; hist is fast so it's minor). FORCED
        # (not setdefault) so the guarantee can't be silently broken by a
        # strategy passing n_jobs. Exact reproducibility holds within the
        # platform (pinned versions / same Modal image); a user's own machine
        # with different xgboost/numpy/CPU can still differ in low-order bits.
        model_params["n_jobs"] = 1

    # Common defaults
    model_params.setdefault("random_state", 42)

    from model_wrapper import ModelWrapper

    clf = cls(**model_params)
    clf.fit(X_train, y_train_enc)

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    return ModelWrapper(clf, original_classes=enc.classes_, n_features=X_train.shape[1])


def _generate_signals(model, X, threshold):
    """Framework-owned signal generation. Deterministic threshold logic."""
    proba = model.predict_proba(X)
    classes = list(model.classes_)

    idx_pos = classes.index(1) if 1 in classes else None
    idx_neg = classes.index(-1) if -1 in classes else None

    p_pos = proba[:, idx_pos] if idx_pos is not None else np.zeros(len(X))
    p_neg = proba[:, idx_neg] if idx_neg is not None else np.zeros(len(X))

    signal_vals = np.zeros(len(X))
    signal_vals = np.where(p_pos >= threshold, 1.0, signal_vals)
    signal_vals = np.where(p_neg >= threshold, -1.0, signal_vals)
    # Both exceed: pick stronger
    both = (p_pos >= threshold) & (p_neg >= threshold)
    signal_vals[both] = np.where(p_pos[both] >= p_neg[both], 1.0, -1.0)

    return pd.Series(signal_vals, index=X.index), p_pos, p_neg


# ── Filter functions (all no-ops when config value is None) ──────────────

def _apply_direction_filter(signal, direction):
    """Zero out signals that don't match allowed direction."""
    if direction is None or direction == "both":
        return signal
    s = signal.copy()
    if direction == "long":
        s[s < 0] = 0.0
    elif direction == "short":
        s[s > 0] = 0.0
    return s


def _apply_session_filter(signal, index, session_hours):
    """Zero out signals outside session hours [start, end] UTC."""
    if session_hours is None:
        return signal
    s = signal.copy()
    start_h, end_h = session_hours[0], session_hours[1]
    hours = index.hour
    if start_h <= end_h:
        mask = (hours >= start_h) & (hours < end_h)
    else:  # wrap around midnight, e.g. [22, 6]
        mask = (hours >= start_h) | (hours < end_h)
    s[~mask] = 0.0
    return s


def _apply_atr_filter(signal, close, high, low, min_atr):
    """Zero out signals when NATR(14) is below threshold."""
    if min_atr is None:
        return signal
    hl = high - low
    hc = (high - close.shift(1)).abs()
    lc = (low - close.shift(1)).abs()
    tr = pd.concat([hl, hc, lc], axis=1).max(axis=1)
    atr14 = tr.ewm(com=13, adjust=False).mean()
    natr = atr14 / close.replace(0, np.nan)
    s = signal.copy()
    s[natr < min_atr] = 0.0
    return s


def _apply_trend_filter(signal, close, trend_filter):
    """Only allow signals aligned with trend. e.g. 'sma_50': longs above SMA, shorts below."""
    if trend_filter is None:
        return signal
    # Parse: "sma_50" → SMA with period 50
    parts = trend_filter.lower().replace("-", "_").split("_")
    if len(parts) >= 2 and parts[0] in ("sma", "ema"):
        period = int(parts[1])
    else:
        return signal  # unknown filter, skip
    if parts[0] == "sma":
        trend_line = close.rolling(period).mean()
    else:
        trend_line = close.ewm(span=period, adjust=False).mean()
    s = signal.copy()
    # Longs only above trend, shorts only below
    s[(s > 0) & (close < trend_line)] = 0.0
    s[(s < 0) & (close > trend_line)] = 0.0
    return s


# ── run_backtest_v2: framework-owned SL/TP/cooldown/position management ──

def run_backtest_v2(signal, close, high, low, config, capital=10000, cost=2e-5):
    """Backtest with SL/TP/cooldown/direction handling built into the engine.
    Unlike run_backtest (v1), this function handles position exits internally.

    Returns: same dict shape as run_backtest()
    """
    stop_loss = config.get("stop_loss")
    take_profit = config.get("take_profit")
    cooldown = config.get("cooldown", 0)
    on_opposite = config.get("on_opposite", "reverse")

    sig_arr = signal.values
    close_arr = close.values
    high_arr = high.values
    low_arr = low.values
    idx = signal.index
    n = len(close_arr)

    trade_returns = []
    long_returns = []
    short_returns = []
    trade_log = []
    equity_vals = np.full(n, float(capital))
    cumret = 1.0

    position = 0.0       # current direction: 1.0, -1.0, or 0.0 (flat)
    entry_price = None
    entry_bar = None      # index into arrays for entry time
    cooldown_remaining = 0

    def _log_trade(exit_bar, exit_px, ret, reason):
        trade_log.append({
            "type": "Buy" if position == 1.0 else "Sell",
            "entry_time": str(idx[entry_bar]),
            "exit_time": str(idx[exit_bar]),
            "entry_price": round(entry_price, 5),
            "exit_price": round(exit_px, 5),
            "pnl": round(position * (exit_px - entry_price), 5),
            "pnl_pct": round(ret * 100, 3),
            "exit_reason": reason,
        })

    for i in range(n):
        c = close_arr[i]
        h = high_arr[i]
        lo = low_arr[i]
        s = sig_arr[i]

        # 1. Check SL/TP if in trade
        if position != 0.0 and entry_price is not None:
            hit_sl = False
            hit_tp = False
            exit_price = None

            if position == 1.0:  # long
                if stop_loss is not None and lo <= entry_price * (1 - stop_loss):
                    hit_sl = True
                    exit_price = entry_price * (1 - stop_loss)
                elif take_profit is not None and h >= entry_price * (1 + take_profit):
                    hit_tp = True
                    exit_price = entry_price * (1 + take_profit)
            else:  # short
                if stop_loss is not None and h >= entry_price * (1 + stop_loss):
                    hit_sl = True
                    exit_price = entry_price * (1 + stop_loss)
                elif take_profit is not None and lo <= entry_price * (1 - take_profit):
                    hit_tp = True
                    exit_price = entry_price * (1 - take_profit)

            if hit_sl or hit_tp:
                ret = float(position * (exit_price - entry_price) / entry_price - cost)
                trade_returns.append(ret)
                if position == 1.0:
                    long_returns.append(ret)
                else:
                    short_returns.append(ret)
                _log_trade(i, exit_price, ret, "SL" if hit_sl else "TP")
                cumret *= (1 + ret)
                position = 0.0
                entry_price = None
                entry_bar = None
                cooldown_remaining = cooldown
                equity_vals[i] = capital * cumret
                continue

        # 2. Cooldown
        if cooldown_remaining > 0:
            cooldown_remaining -= 1
            equity_vals[i] = capital * cumret
            continue

        # 3. Signal processing
        if s != 0.0:
            if position == 0.0:
                # Open new trade
                position = s
                entry_price = c
                entry_bar = i
            elif s != position:
                # Opposite signal
                if on_opposite == "reverse":
                    # Close current + open opposite
                    ret = float(position * (c - entry_price) / entry_price - cost)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "signal")
                    cumret *= (1 + ret)
                    position = s
                    entry_price = c
                    entry_bar = i
                else:  # close_only
                    # Close current, go flat
                    ret = float(position * (c - entry_price) / entry_price - cost)
                    trade_returns.append(ret)
                    if position == 1.0:
                        long_returns.append(ret)
                    else:
                        short_returns.append(ret)
                    _log_trade(i, c, ret, "close_only")
                    cumret *= (1 + ret)
                    position = 0.0
                    entry_price = None
                    entry_bar = None
                    cooldown_remaining = cooldown

        equity_vals[i] = capital * cumret

    # Close last open trade at final close
    if position != 0.0 and entry_price is not None and n > 0 and entry_price != 0:
        c = close_arr[-1]
        ret = float(position * (c - entry_price) / entry_price - cost)
        trade_returns.append(ret)
        if position == 1.0:
            long_returns.append(ret)
        else:
            short_returns.append(ret)
        _log_trade(n - 1, c, ret, "end")
        cumret *= (1 + ret)
        equity_vals[-1] = capital * cumret

    # Bar returns for Sharpe (approximate)
    bar_returns = np.zeros(n)
    for i in range(1, n):
        if close_arr[i - 1] != 0 and sig_arr[i - 1] != 0:
            bar_returns[i] = sig_arr[i - 1] * (close_arr[i] - close_arr[i - 1]) / close_arr[i - 1]

    return {
        "equity": pd.Series(equity_vals, index=close.index),
        "trade_returns": trade_returns,
        "long_returns": long_returns,
        "short_returns": short_returns,
        "bar_returns": bar_returns,
        "trade_log": trade_log,
    }


# ── run_strategy: the v2 orchestrator ────────────────────────────────────

def run_strategy(feature_fn, config_fn, data_path, start_date="", end_date="",
                 validation_date="", train_split=0.7, register_model_fn=None):
    """Config-driven strategy execution. Claude writes feature_fn + config_fn,
    framework does everything else.

    Returns: results dict (same format as webapp expects)
    """
    config = config_fn()

    # Auto-correct SL/TP if Claude passed percentage instead of decimal
    for _key in ("stop_loss", "take_profit"):
        _val = config.get(_key)
        if _val is not None and _val > 0.1:  # >10% is almost certainly a percentage
            config[_key] = _val / 100.0
            print(f"[strategy] Auto-corrected {_key}: {_val} -> {config[_key]} (was percentage, converted to decimal)")

    # 1. Load data
    df, close, open_, high, low = load_ohlc(data_path, start_date, end_date)

    # 2. Feature engineering (Claude's function)
    df = feature_fn(df, close, open_, high, low)
    close = df["close"]
    open_ = df["open"]
    high = df["high"]
    low = df["low"]

    # 3. Warm-up detection: drop rows where features have NaN BEFORE any fill
    feature_cols = [c for c in df.columns if c not in ("open", "high", "low", "close")]
    raw_nans = df[feature_cols].isna().any(axis=1)
    valid_rows = ~raw_nans
    if valid_rows.any():
        first_valid = valid_rows.idxmax()
        if raw_nans.loc[:first_valid].any():
            df = df.loc[first_valid:].copy()
            close = df["close"]
            open_ = df["open"]
            high = df["high"]
            low = df["low"]

    # 4. Target
    horizon = config.get("target_horizon", 4)
    target = make_target(close, horizon=horizon)

    # 5. Split (ffill only within each partition — no bfill leak)
    mask = target.notna()
    df = df[mask].copy()
    target = target[mask]
    close = df["close"]
    high = df["high"]
    low = df["low"]

    X = df[feature_cols].copy()
    X = X.replace([np.inf, -np.inf], np.nan)

    if validation_date:
        split_idx = len(df[df.index <= validation_date])
    else:
        split_idx = int(len(df) * train_split)
    split_idx = max(1, min(split_idx, len(df) - 1))

    # ffill within train and test separately (no leak)
    X_train = X.iloc[:split_idx].ffill().fillna(0.0)
    X_test = X.iloc[split_idx:].ffill().fillna(0.0)
    X = pd.concat([X_train, X_test])

    y_train = target.iloc[:split_idx]
    y_test = target.iloc[split_idx:]
    close_train = close.iloc[:split_idx]
    close_test = close.iloc[split_idx:]
    high_test = high.iloc[split_idx:]
    low_test = low.iloc[split_idx:]

    enc = LabelEncoder()
    enc.fit([-1, 0, 1])
    y_train_enc = enc.transform(y_train)
    y_test_enc = enc.transform(y_test)

    split_dt = str(df.index[split_idx])

    sp = {
        "df": df, "X_train": X_train, "X_test": X_test,
        "y_train": y_train, "y_test": y_test,
        "y_train_enc": y_train_enc, "y_test_enc": y_test_enc,
        "enc": enc,
        "close": close, "close_train": close_train, "close_test": close_test,
        "split_idx": split_idx, "split_dt": split_dt,
        "n_train": len(X_train), "n_test": len(X_test),
    }

    # 6. Build model from config
    model = _build_model_from_config(config, X_train, y_train_enc)

    # 7. Generate signals
    threshold = config.get("signal_threshold", 0.55)
    signal_train, p_pos_train, p_neg_train = _generate_signals(model, X_train, threshold)
    signal_test, p_pos_test, p_neg_test = _generate_signals(model, X_test, threshold)

    # 8. Apply filters (order: direction → session → ATR → trend)
    direction = config.get("direction", "both")
    signal_test = _apply_direction_filter(signal_test, direction)
    signal_train = _apply_direction_filter(signal_train, direction)

    session_filter = config.get("session_filter")
    signal_test = _apply_session_filter(signal_test, signal_test.index, session_filter)
    signal_train = _apply_session_filter(signal_train, signal_train.index, session_filter)

    min_atr = config.get("min_atr")
    if min_atr is not None:
        signal_test = _apply_atr_filter(signal_test, close_test, high_test, low_test, min_atr)

    trend_filter = config.get("trend_filter")
    if trend_filter is not None:
        signal_test = _apply_trend_filter(signal_test, close_test, trend_filter)

    signal_full = pd.concat([signal_train, signal_test])

    # 9. Backtest with SL/TP/cooldown (test + train)
    high_train = high.iloc[:split_idx]
    low_train = low.iloc[:split_idx]
    has_risk = (config.get("stop_loss") is not None or
                config.get("take_profit") is not None or
                config.get("cooldown", 0) > 0 or
                config.get("on_opposite", "reverse") != "reverse")

    if has_risk:
        bt = run_backtest_v2(signal_test, close_test, high_test, low_test, config, capital=10000)
        bt_train = run_backtest_v2(signal_train, close_train, high_train, low_train, config, capital=10000)
    else:
        bt = run_backtest(signal_test, close_test, capital=10000)
        bt_train = run_backtest(signal_train, close_train, capital=10000)

    # 10. Metrics
    metrics = compute_metrics(bt, close_test, capital=10000)

    # 11. Pre-compute all trade stats (single source of truth)
    pre_stats = {
        "train_stats": compute_trade_stats(bt_train.get("trade_returns", []), capital=10000),
        "test_stats":  compute_trade_stats(bt.get("trade_returns", []), capital=10000),
        "long_stats":  compute_trade_stats(bt.get("long_returns", []), capital=10000),
        "short_stats": compute_trade_stats(bt.get("short_returns", []), capital=10000),
    }

    # 12. Register model
    if register_model_fn is not None:
        register_model_fn(model)

    # 13. Build return dict
    return build_return_dict(sp, bt, metrics, model, feature_cols,
                             signal_full, p_pos_test, p_neg_test, custom_figs=[],
                             bt_train_result=bt_train, pre_stats=pre_stats)

# ── End strategy_utils ──


DATA_PATH = '/root/Desktop/QuantifyMe/data/ohlc/AUDUSD_15min.parquet'
START_DATE = '2026-04-15'
END_DATE = '2026-05-25'
VALIDATION_DATE = ""
TRAIN_SPLIT = 0.7

# SECTION 1 — FEATURE ENGINEERING
def feature_engineering(df, close, open_, high, low):

    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_period = 20
    bb_std    = 2.0
    bb_mid    = close.rolling(bb_period).mean()
    bb_sigma  = close.rolling(bb_period).std(ddof=0)
    bb_upper  = bb_mid + bb_std * bb_sigma
    bb_lower  = bb_mid - bb_std * bb_sigma

    df["bb_mid"]   = bb_mid
    df["bb_upper"] = bb_upper
    df["bb_lower"] = bb_lower

    # %B — position of close within the band (0 = lower, 1 = upper)
    bb_range       = bb_upper - bb_lower
    df["bb_pct_b"] = np.where(bb_range > 0, (close - bb_lower) / bb_range, 0.5)

    # Bandwidth — normalised band width (regime filter)
    df["bb_bandwidth"] = np.where(bb_mid > 0, bb_range / bb_mid, 0.0)

    # Distance from each band (signed, normalised by sigma)
    df["dist_lower"] = np.where(bb_sigma > 0, (close - bb_lower) / bb_sigma, 0.0)
    df["dist_upper"] = np.where(bb_sigma > 0, (bb_upper - close) / bb_sigma, 0.0)
    df["dist_mid"]   = np.where(bb_sigma > 0, (close - bb_mid)   / bb_sigma, 0.0)

    # Below lower band flag
    df["below_lower"] = np.where(close < bb_lower, 1, 0)
    # Above upper band flag
    df["above_upper"] = np.where(close > bb_upper, 1, 0)

    # ── RSI (14) ─────────────────────────────────────────────────────────────
    rsi_period = 14
    delta      = close.diff()
    gain       = delta.clip(lower=0)
    loss       = (-delta).clip(lower=0)
    avg_gain   = gain.ewm(com=rsi_period - 1, min_periods=rsi_period).mean()
    avg_loss   = loss.ewm(com=rsi_period - 1, min_periods=rsi_period).mean()
    rs         = np.where(avg_loss > 0, avg_gain / avg_loss, 100.0)
    rsi        = 100.0 - 100.0 / (1.0 + rs)
    df["rsi"]  = rsi

    # RSI-derived flags and distances
    df["rsi_oversold"]   = np.where(rsi < 35, 1, 0)
    df["rsi_overbought"] = np.where(rsi > 65, 1, 0)
    df["rsi_dist_35"]    = rsi - 35.0   # negative when oversold
    df["rsi_dist_65"]    = rsi - 65.0   # positive when overbought
    df["rsi_norm"]       = (rsi - 50.0) / 50.0  # centred, ±1 range

    # ── Core entry condition features ────────────────────────────────────────
    # Buy setup: close < lower BB AND RSI < 35
    df["long_setup"]  = np.where((close < bb_lower) & (rsi < 35), 1, 0)
    # Sell setup: close > upper BB AND RSI > 65
    df["short_setup"] = np.where((close > bb_upper) & (rsi > 65), 1, 0)

    # ── ATR (14) — volatility context ────────────────────────────────────────
    atr_period = 14
    hl         = high - low
    hc         = (high - close.shift(1)).abs()
    lc         = (low  - close.shift(1)).abs()
    tr         = pd.concat([hl, hc, lc], axis=1).max(axis=1)
    atr        = tr.ewm(com=atr_period - 1, min_periods=atr_period).mean()
    df["atr"]  = atr
    df["natr"] = np.where(close > 0, atr / close, 0.0)

    # ── Momentum / Rate-of-Change ─────────────────────────────────────────────
    for n in [1, 3, 5, 10]:
        df[f"roc_{n}"] = np.where(
            close.shift(n) > 0,
            (close - close.shift(n)) / close.shift(n),
            0.0
        )

    # ── EMA trend context (fast / slow) ──────────────────────────────────────
    ema_fast = close.ewm(span=9,  min_periods=9).mean()
    ema_slow = close.ewm(span=21, min_periods=21).mean()
    df["ema_fast"] = ema_fast
    df["ema_slow"] = ema_slow
    df["ema_diff"] = np.where(ema_slow > 0, (ema_fast - ema_slow) / ema_slow, 0.0)
    df["ema_bull"] = np.where(ema_fast > ema_slow, 1, 0)

    # SMA-50 trend filter helper (used by framework trend_filter)
    df["sma_50"] = close.rolling(50).mean()

    # ── Candle body & wick features ───────────────────────────────────────────
    body        = (close - open_).abs()
    candle_rng  = (high - low).replace(0, np.nan)
    df["body_ratio"]  = (body / candle_rng).fillna(0.0)
    df["upper_wick"]  = np.where(candle_rng.notna(), (high  - close.clip(lower=open_)) / candle_rng.fillna(1), 0.0)
    df["lower_wick"]  = np.where(candle_rng.notna(), (close.clip(upper=open_) - low)   / candle_rng.fillna(1), 0.0)
    df["bull_candle"]  = np.where(close > open_, 1, 0)

    # ── Volume-like proxy — true range z-score ────────────────────────────────
    tr_mean = tr.rolling(20).mean()
    tr_std  = tr.rolling(20).std(ddof=0).replace(0, np.nan)
    df["tr_zscore"] = ((tr - tr_mean) / tr_std).fillna(0.0)

    # ── Lagged RSI and %B (1, 2, 3 bars back) ────────────────────────────────
    for lag in [1, 2, 3]:
        df[f"rsi_lag{lag}"]    = df["rsi"].shift(lag)
        df[f"bb_pct_b_lag{lag}"] = df["bb_pct_b"].shift(lag)

    # ── RSI slope ────────────────────────────────────────────────────────────
    df["rsi_slope3"] = df["rsi"] - df["rsi"].shift(3)

    # ── Mean-reversion proximity: how far price is from middle band ───────────
    df["pct_to_mid"] = np.where(close > 0, (bb_mid - close) / close, 0.0)

    # ── Fill any NaNs from warm-up ────────────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "BB Mean-Reversion + RSI Oversold/Overbought (XGBoost)",
        "model_type":       "XGBClassifier",
        "model_params": {
            "n_estimators":      500,
            "max_depth":         4,
            "learning_rate":     0.03,
            "subsample":         0.75,
            "colsample_bytree":  0.70,
            "min_child_weight":  5,
            "gamma":             0.1,
            "reg_alpha":         0.05,
            "reg_lambda":        1.5,
            "objective":         "binary:logistic",
            "random_state":      42,
            "n_jobs":            -1,
        },
        "signal_threshold": 0.55,
        "direction":        "both",
        "stop_loss":        0.0010,
        "take_profit":      0.0020,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [7, 17],
        "min_atr":          0.00005,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximize Sharpe ratio by exploiting Bollinger Band mean-reversion "
            "with RSI confirmation. Entry conditions (close < lower BB, RSI < 35 "
            "for longs; close > upper BB, RSI > 65 for shorts) are encoded as "
            "features together with momentum, ATR volatility, candle structure, "
            "and lagged indicators. XGBoost with strong regularisation "
            "(reg_lambda=1.5, gamma=0.1, min_child_weight=5) and a low learning "
            "rate avoids overfitting on the 6-week window. Session filter "
            "[7,17] UTC targets liquid London/NY overlap, reducing noise. "
            "TP:SL ratio of 2:1 supports positive expected value even at "
            "moderate win rates, pushing Sharpe higher."
        ),
        "notes": (
            "Features: %B position, RSI (raw + flags + slope + lags), "
            "EMA cross, ATR/NATR, ROC(1/3/5/10), candle body/wick ratios, "
            "TR z-score, distance-to-midband, long/short setup flags. "
            "Round-trip cost ~2e-5 is implicitly absorbed by the 10-pip TP target. "
            "Cooldown=0 allows immediate re-entry after mean-reversion completes."
        ),
    }

# ── Framework v2: auto-generated wrapper ──
def train_and_backtest():
    _vd = VALIDATION_DATE if 'VALIDATION_DATE' in globals() else ''
    _ts = TRAIN_SPLIT if 'TRAIN_SPLIT' in globals() else 0.7
    return run_strategy(
        feature_engineering, strategy_config,
        DATA_PATH, START_DATE, END_DATE,
        _vd, _ts,
        register_model_fn=register_model
    )