Community Scripts

# ╔══════════════════════════════════════════════════════════════╗
# ║                  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        : 2026-05-06 01:50:17
# Model            : Gradient Boosting
# Feature Eng.     : BB (20,2.0), 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/USDCAD_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):
    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_period = 20
    bb_std    = 2.0
    bb_mid    = close.rolling(bb_period).mean()
    bb_std_s  = close.rolling(bb_period).std(ddof=0)
    bb_upper  = bb_mid + bb_std * bb_std_s
    bb_lower  = bb_mid - bb_std * bb_std_s

    df["bb_mid"]   = bb_mid
    df["bb_upper"] = bb_upper
    df["bb_lower"] = bb_lower
    df["bb_width"] = (bb_upper - bb_lower) / bb_mid
    df["bb_pct"]   = (close - bb_lower) / (bb_upper - bb_lower)

    # ── ATR (14) & Normalised ATR ────────────────────────────────────────────
    atr_period = 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=atr_period, adjust=False).mean()
    natr = atr / close

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

    # ── RSI (14) ─────────────────────────────────────────────────────────────
    rsi_period = 14
    delta      = close.diff()
    gain       = delta.clip(lower=0)
    loss       = (-delta).clip(lower=0)
    avg_gain   = gain.ewm(span=rsi_period, adjust=False).mean()
    avg_loss   = loss.ewm(span=rsi_period, adjust=False).mean()
    rs         = avg_gain / avg_loss.replace(0, np.nan)
    df["rsi"]  = 100 - (100 / (1 + 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
    macd_signal    = macd_line.ewm(span=9, adjust=False).mean()
    df["macd"]     = macd_line
    df["macd_sig"] = macd_signal
    df["macd_hist"]= macd_line - macd_signal

    # ── SMA filters (50, 200) ────────────────────────────────────────────────
    df["sma_20"]  = close.rolling(20).mean()
    df["sma_50"]  = close.rolling(50).mean()
    df["sma_200"] = close.rolling(200).mean()

    # Price relative to moving averages
    df["close_vs_sma20"]  = (close - df["sma_20"])  / df["sma_20"]
    df["close_vs_sma50"]  = (close - df["sma_50"])  / df["sma_50"]
    df["close_vs_sma200"] = (close - df["sma_200"]) / df["sma_200"]

    # ── Price momentum / returns ─────────────────────────────────────────────
    df["ret_1"]  = close.pct_change(1)
    df["ret_4"]  = close.pct_change(4)
    df["ret_8"]  = close.pct_change(8)
    df["ret_16"] = close.pct_change(16)
    df["ret_32"] = close.pct_change(32)

    # ── Candle body & 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["close_dir"]   = np.sign(close - open_)

    # ── Volatility regime ────────────────────────────────────────────────────
    df["vol_ratio"]  = natr / natr.rolling(50).mean()         # ATR vs its own average
    df["bb_squeeze"] = np.where(df["bb_width"] < df["bb_width"].rolling(50).quantile(0.25), 1.0, 0.0)

    # ── 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

    # ── Rate-of-change ───────────────────────────────────────────────────────
    df["roc_10"] = (close - close.shift(10)) / close.shift(10)

    # ── Rolling z-score of close (20-bar) ────────────────────────────────────
    roll_mean = close.rolling(20).mean()
    roll_std  = close.rolling(20).std(ddof=0).replace(0, np.nan)
    df["zscore_20"] = (close - roll_mean) / roll_std

    # ── Volume-related (if volume column exists) ─────────────────────────────
    if "volume" in df.columns and df["volume"].sum() > 0:
        vol_ma          = df["volume"].rolling(20).mean().replace(0, np.nan)
        df["vol_ratio_20"] = df["volume"] / vol_ma

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

    return df

# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title": "USD/CAD BB + ATR Gradient Boosting Mean-Rev",
        "model_type": "GradientBoostingClassifier",
        "model_params": {
            "n_estimators":     400,
            "max_depth":        4,
            "learning_rate":    0.04,
            "subsample":        0.8,
            "min_samples_leaf": 20,
            "max_features":     "sqrt",
            "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":   [7, 20],
        "min_atr":          0.0002,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe/Calmar) on USD/CAD 15-min data. "
            "GradientBoostingClassifier chosen for strong generalisation on noisy FX "
            "price data; moderate depth (4) and learning rate (0.04) with early stopping "
            "prevent overfitting. Features: Bollinger Bands (mean-reversion signal via "
            "bb_pct and bb_width), ATR/NATR (volatility filter), RSI, MACD, Stochastic, "
            "z-score, momentum returns, and candle-body ratios. 2:1 R:R (SL 0.5%, TP 1.0%) "
            "with session filter (07-20 UTC) to avoid illiquid overnight hours."
        ),
        "notes": (
            "session_filter [7,20] captures London + NY overlap on USD/CAD. "
            "min_atr 0.0002 avoids flat/choppy markets. on_opposite=reverse ensures "
            "the model flips direction quickly when sentiment changes. "
            "target_horizon=4 bars (1 hour) aligns with typical intraday FX moves."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  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        : 2026-05-06 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        : 2026-05-06 02:13:41
# 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/USDCAD_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 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, 200]:
        df[f"sma_{p}_slope"] = df[f"sma_{p}"].diff(5) / df[f"sma_{p}"].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 momentum / rate of change ──────────────────────────────────
    for p in [4, 8, 16, 32]:
        df[f"roc_{p}"] = close.pct_change(p)

    # ── RSI (manual, no external libs) ───────────────────────────────────
    def calc_rsi(series, period=14):
        delta = series.diff()
        gain  = delta.clip(lower=0)
        loss  = (-delta).clip(lower=0)
        avg_gain = gain.ewm(alpha=1.0 / period, min_periods=period, adjust=False).mean()
        avg_loss = loss.ewm(alpha=1.0 / period, min_periods=period, adjust=False).mean()
        rs  = avg_gain / avg_loss.replace(0, np.nan)
        rsi = 100 - (100 / (1 + rs))
        return rsi

    for p in [9, 14, 21]:
        df[f"rsi_{p}"] = calc_rsi(close, p)
        df[f"rsi_{p}_norm"] = (df[f"rsi_{p}"] - 50) / 50   # centre around 0

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

    # ── Bollinger Bands ───────────────────────────────────────────────────
    for p in [20, 50]:
        mid  = close.rolling(p).mean()
        std  = close.rolling(p).std()
        df[f"bb_upper_{p}"] = mid + 2 * std
        df[f"bb_lower_{p}"] = mid - 2 * std
        denom = (df[f"bb_upper_{p}"] - df[f"bb_lower_{p}"]).replace(0, np.nan)
        df[f"bb_pct_{p}"]   = (close - df[f"bb_lower_{p}"]) / denom
        df[f"bb_width_{p}"] = denom / mid

    # ── ATR (manual) ─────────────────────────────────────────────────────
    def calc_atr(h, l, c, period=14):
        prev_c = c.shift(1)
        tr = pd.concat([
            h - l,
            (h - prev_c).abs(),
            (l - prev_c).abs()
        ], axis=1).max(axis=1)
        return tr.ewm(alpha=1.0 / period, min_periods=period, adjust=False).mean()

    for p in [7, 14]:
        atr = calc_atr(high, low, close, p)
        df[f"atr_{p}"]  = atr
        df[f"natr_{p}"] = atr / close   # normalised ATR

    # ── Candle body / wick features ───────────────────────────────────────
    body       = (close - open_).abs()
    candle_rng = (high - low).replace(0, np.nan)
    df["body_ratio"]   = body / candle_rng
    df["upper_wick"]   = (high - np.maximum(close, open_)) / candle_rng
    df["lower_wick"]   = (np.minimum(close, open_) - low) / candle_rng
    df["candle_dir"]   = np.sign(close - open_)

    # ── Rolling volatility ────────────────────────────────────────────────
    log_ret = np.log(close / close.shift(1))
    for p in [8, 16, 32]:
        df[f"vol_{p}"] = log_ret.rolling(p).std()

    # ── Volume (if available) — graceful fallback ─────────────────────────
    if "volume" in df.columns and df["volume"].sum() > 0:
        vol_ma = df["volume"].rolling(20).mean()
        df["vol_ratio"] = df["volume"] / vol_ma.replace(0, np.nan)
    else:
        df["vol_ratio"] = 1.0

    # ── Lagged returns ────────────────────────────────────────────────────
    for lag in [1, 2, 3, 4, 8]:
        df[f"ret_lag_{lag}"] = log_ret.shift(lag)

    # ── Higher-timeframe SMA context (4-bar = 1h proxy) ──────────────────
    close_1h = close.rolling(4).mean()
    for p in [20, 50]:
        sma_1h = close_1h.rolling(p).mean()
        df[f"1h_dm_sma_{p}"] = (close_1h - sma_1h) / sma_1h

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

    return df

# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "USD/CAD SMA Trend + Momentum XGBoost Scalper",
        "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":            0.2,
            "reg_alpha":        0.1,
            "reg_lambda":       1.5,
            "objective":        "binary:logistic",
            "tree_method":      "hist",
            "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":   [7, 20],
        "min_atr":          0.0002,
        "trend_filter":     "sma_50",
        "target_horizon":   4,
        "objective": (
            "Maximise risk-adjusted return on USD/CAD 15-min bars. "
            "XGBoost with deep feature set (multi-period SMA distances and crossovers, "
            "RSI, MACD, Bollinger Bands, ATR, candle structure, lagged returns). "
            "Regularised tree ensemble (gamma, L1/L2, min_child_weight) prevents "
            "overfitting on the ~1-year window. 2:1 TP:SL ratio locks in positive "
            "expectancy; session filter restricts trading to liquid London/NY overlap."
        ),
        "notes": (
            "SMA-trio (20/50/200) distances are the primary trend-context features. "
            "MACD histogram momentum + RSI multi-period confirm entry timing. "
            "ATR normalisation makes volatility features scale-invariant. "
            "sma_50 trend filter ensures long trades only above 50-SMA and shorts below, "
            "aligning ML signals with dominant trend and improving Sharpe ratio."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  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 01:44:26
# Model            : XGBoost
# Feature Eng.     : BB (20,2.0), 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/AUDUSD_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):
    # ── 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
    df["bb_width"] = (bb_upper - bb_lower) / bb_mid
    # guard against zero range
    bb_range = bb_upper - bb_lower
    df["bb_pct"]   = np.where(bb_range != 0, (close - bb_lower) / bb_range, 0.5)

    # ── ATR (14) & Normalised ATR ────────────────────────────────────────────
    atr_period = 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=atr_period, min_periods=atr_period, adjust=False).mean()
    df["atr"]  = atr
    df["natr"] = np.where(close != 0, atr / close, 0.0)

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

    # ── 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

    # ── EMA trend features ───────────────────────────────────────────────────
    ema_20  = close.ewm(span=20, adjust=False).mean()
    ema_50  = close.ewm(span=50, adjust=False).mean()
    ema_200 = close.ewm(span=200, adjust=False).mean()
    df["ema_20"]        = ema_20
    df["ema_50"]        = ema_50
    df["ema_200"]       = ema_200
    df["close_vs_ema20"]  = (close - ema_20) / ema_20
    df["close_vs_ema50"]  = (close - ema_50) / ema_50
    df["ema20_vs_ema50"]  = (ema_20 - ema_50) / ema_50
    df["ema50_vs_ema200"] = (ema_50 - ema_200) / ema_200

    # ── Price momentum / rate-of-change ──────────────────────────────────────
    for n in [1, 4, 8, 16]:
        df[f"roc_{n}"] = close.pct_change(n)

    # ── Rolling volatility ────────────────────────────────────────────────────
    df["vol_10"]  = close.pct_change().rolling(10).std()
    df["vol_20"]  = close.pct_change().rolling(20).std()
    df["vol_ratio"] = np.where(df["vol_20"] != 0,
                               df["vol_10"] / df["vol_20"], 1.0)

    # ── Stochastic %K / %D (14, 3) ───────────────────────────────────────────
    low_14  = low.rolling(14).min()
    high_14 = high.rolling(14).max()
    stoch_range = high_14 - low_14
    stoch_k = np.where(stoch_range != 0,
                       100 * (close - low_14) / stoch_range, 50.0)
    df["stoch_k"] = stoch_k
    df["stoch_d"] = pd.Series(stoch_k, index=close.index).rolling(3).mean()

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

    # ── BB interaction features ───────────────────────────────────────────────
    df["price_above_bb_upper"] = np.where(close > bb_upper, 1, 0)
    df["price_below_bb_lower"] = np.where(close < bb_lower, 1, 0)
    df["bb_squeeze"]           = np.where(df["bb_width"] < df["bb_width"].rolling(50).quantile(0.20), 1, 0)

    # ── RSI regime bins (replacing pd.cut) ───────────────────────────────────
    df["rsi_oversold"]  = np.where(rsi < 30, 1, 0)
    df["rsi_overbought"] = np.where(rsi > 70, 1, 0)
    df["rsi_neutral"]   = np.where((rsi >= 30) & (rsi <= 70), 1, 0)

    # ── Volume proxy (if volume column exists) ───────────────────────────────
    if "volume" in df.columns:
        vol_ma = df["volume"].rolling(20).mean()
        df["volume_ratio"] = np.where(vol_ma != 0,
                                      df["volume"] / vol_ma, 1.0)

    # ── Lagged features (1-bar lag to avoid lookahead) ───────────────────────
    for feat in ["bb_pct", "rsi_14", "macd_hist", "natr", "stoch_k"]:
        df[f"{feat}_lag1"] = df[feat].shift(1)
        df[f"{feat}_lag2"] = df[feat].shift(2)

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

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "AUD/USD Bollinger + ATR Mean-Rev (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":              0.10,
            "reg_alpha":          0.10,
            "reg_lambda":         1.50,
            "objective":          "binary:logistic",
            "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":   [0, 23],
        "min_atr":          None,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe). "
            "XGBoost with moderate depth and heavy regularisation "
            "(gamma, alpha, lambda) prevents overfit on AUD/USD 15-min data. "
            "Bollinger Bands capture mean-reversion; ATR normalises volatility; "
            "RSI + MACD confirm momentum; 2:1 TP:SL ratio supports positive expectancy."
        ),
        "notes": (
            "Features: BB (20,2) width/pct, ATR-14/NATR, RSI-14, MACD histogram, "
            "EMA 20/50/200 spreads, Stochastic %K/%D, candle-body ratios, "
            "ROC at multiple horizons, volatility ratio, BB squeeze flag, "
            "lagged versions of key features. "
            "Threshold 0.56 filters marginal signals, improving precision. "
            "target_horizon=4 (1 hour) balances signal frequency vs. noise."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:32:18
# 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      : —
# ══════════════════════════════════════════════════════════════

# ============================================================
# AUDUSD 15-min XGBoost Strategy
# SMA + RSI + MACD + Bollinger Bands + ATR Feature Set
# Optimized for Risk-Adjusted Return
# ============================================================

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

DATA_PATH = "/root/Desktop/QuantifyMe/data/ohlc/AUDUSD_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, 2.0) ───────────────────────────────────────────
    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
    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"]   = (close - bb_lower) / bb_range

    # ── 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"] = 100.0 - (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 + Normalised ATR ─────────────────────────────────────────────
    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).mean()
    df["atr_14"] = atr
    df["natr"]   = atr / close

    # ── Price momentum / rate-of-change ────────────────────────────────────
    for n in [1, 4, 8, 16]:
        df[f"roc_{n}"] = close.pct_change(n)

    # ── Candle body & 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["candle_dir"]  = np.where(close >= open_, 1.0, -1.0)

    # ── Volume (if present) ─────────────────────────────────────────────────
    if "volume" in df.columns:
        vol_ma = df["volume"].rolling(20).mean()
        df["vol_ratio"] = df["volume"] / vol_ma.replace(0, np.nan)

    # ── Lagged RSI & MACD histogram ─────────────────────────────────────────
    for lag in [1, 2, 3]:
        df[f"rsi_14_lag{lag}"]   = df["rsi_14"].shift(lag)
        df[f"macd_hist_lag{lag}"] = df["macd_hist"].shift(lag)

    # ── RSI overbought / oversold zones ─────────────────────────────────────
    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_up"] = np.where((df["rsi_14"] > 50) & (df["rsi_14"] <= 70), 1.0, 0.0)
    df["rsi_mid_dn"] = np.where((df["rsi_14"] >= 30) & (df["rsi_14"] < 50), 1.0, 0.0)

    # ── MACD cross signals ───────────────────────────────────────────────────
    df["macd_cross_up"] = np.where(
        (df["macd_line"] > df["macd_signal"]) &
        (df["macd_line"].shift(1) <= df["macd_signal"].shift(1)),
        1.0, 0.0
    )
    df["macd_cross_dn"] = np.where(
        (df["macd_line"] < df["macd_signal"]) &
        (df["macd_line"].shift(1) >= df["macd_signal"].shift(1)),
        1.0, 0.0
    )

    # ── Price position relative to SMA alignment ────────────────────────────
    df["trend_aligned_bull"] = np.where(
        (close > df["sma_20"]) & (df["sma_20"] > df["sma_50"]) & (df["sma_50"] > df["sma_200"]),
        1.0, 0.0
    )
    df["trend_aligned_bear"] = np.where(
        (close < df["sma_20"]) & (df["sma_20"] < df["sma_50"]) & (df["sma_50"] < df["sma_200"]),
        1.0, 0.0
    )

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

    # ── Rolling close statistics ─────────────────────────────────────────────
    df["close_zscore_20"] = (close - close.rolling(20).mean()) / close.rolling(20).std(ddof=0)
    df["close_zscore_50"] = (close - close.rolling(50).mean()) / close.rolling(50).std(ddof=0)

    # ── Volatility regime ────────────────────────────────────────────────────
    natr_ma = df["natr"].rolling(20).mean()
    df["vol_regime_high"] = np.where(df["natr"] > natr_ma, 1.0, 0.0)

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

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "AUD/USD XGBoost SMA+RSI+MACD+BB Momentum",
        "model_type":       "XGBClassifier",
        "model_params":     {
            "n_estimators":       500,
            "max_depth":          4,
            "learning_rate":      0.04,
            "subsample":          0.75,
            "colsample_bytree":   0.70,
            "min_child_weight":   5,
            "gamma":              0.1,
            "reg_alpha":          0.1,
            "reg_lambda":         1.5,
            "scale_pos_weight":   1.0,
            "objective":          "binary:logistic",
            "tree_method":        "hist",
            "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":   [0, 23],
        "min_atr":          None,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective":        (
            "Maximize risk-adjusted return (Sharpe/Calmar) on AUD/USD 15-min. "
            "XGBoost with depth-4 trees and conservative regularization (reg_lambda=1.5, "
            "min_child_weight=5) to reduce overfitting on FX data. "
            "2:1 RR (SL=0.5%, TP=1.0%) ensures positive expectancy with ~40%+ win rate. "
            "Subsample + colsample add stochastic diversity. 500 estimators with lr=0.04 "
            "balances bias-variance. Threshold 0.55 filters marginal signals."
        ),
        "notes":            (
            "Features: SMA(20/50/200) with distances, Bollinger Bands width+pct, RSI-14 "
            "with zone flags, MACD histogram + crosses, ATR-14 + NATR, momentum ROC(1/4/8/16), "
            "candle body/wick ratios, trend alignment flags, BB squeeze, z-scores, vol regime. "
            "Reverse on opposite signal to capture trend reversals. Session filter disabled "
            "to capture AUD/USD Asian + London + NY sessions. Target horizon = 4 bars (1 hour)."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 01:35:33
# Model            : Gradient Boosting
# Feature Eng.     : RSI 14, MACD (12,26,9) + 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/NZDUSD_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):

    # ── RSI 14 ──────────────────────────────────────────────────────────────
    delta = close.diff()
    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)
    df["rsi_14"] = 100 - (100 / (1 + rs))

    # RSI derived signals
    df["rsi_ob"]      = np.where(df["rsi_14"] > 70, 1, 0)   # overbought flag
    df["rsi_os"]      = np.where(df["rsi_14"] < 30, 1, 0)   # oversold flag
    df["rsi_mid"]     = df["rsi_14"] - 50                    # centred
    df["rsi_slope"]   = df["rsi_14"].diff(3)                 # momentum of RSI
    df["rsi_accel"]   = df["rsi_slope"].diff(2)              # acceleration

    # RSI regime: above/below 50
    df["rsi_bull"]    = np.where(df["rsi_14"] > 50, 1, -1)

    # ── 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()
    macd_hist   = macd_line - signal_line

    df["macd_line"]   = macd_line
    df["macd_signal"] = signal_line
    df["macd_hist"]   = macd_hist

    # MACD derived
    df["macd_cross"]     = np.where(macd_line > signal_line, 1, -1)
    df["macd_hist_sign"] = np.where(macd_hist > 0, 1, -1)
    df["macd_hist_chg"]  = macd_hist.diff(1)          # histogram change
    df["macd_hist_accel"]= df["macd_hist_chg"].diff(1) # second derivative
    df["macd_zero_cross"]= np.where(macd_line > 0, 1, -1)

    # ── 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(alpha=1/14, min_periods=14, adjust=False).mean()
    df["atr_14"]   = atr14
    df["natr_14"]  = atr14 / close          # normalised ATR
    df["atr_ratio"]= atr14 / atr14.rolling(50).mean()  # current vs recent vol

    # ── Volatility regime ───────────────────────────────────────────────────
    df["vol_high"] = np.where(df["natr_14"] > df["natr_14"].rolling(100).median(), 1, 0)

    # ── Price momentum ──────────────────────────────────────────────────────
    df["ret_1"]  = close.pct_change(1)
    df["ret_3"]  = close.pct_change(3)
    df["ret_8"]  = close.pct_change(8)
    df["ret_16"] = close.pct_change(16)

    # Scaled by ATR so the model sees normalised moves
    df["ret_1_atr"]  = df["ret_1"]  / (atr14 / close).replace(0, np.nan)
    df["ret_3_atr"]  = df["ret_3"]  / (atr14 / close).replace(0, np.nan)
    df["ret_8_atr"]  = df["ret_8"]  / (atr14 / close).replace(0, np.nan)

    # ── EMAs & trend structure ───────────────────────────────────────────────
    ema8  = close.ewm(span=8,  adjust=False).mean()
    ema21 = close.ewm(span=21, adjust=False).mean()
    ema50 = close.ewm(span=50, adjust=False).mean()
    ema100= close.ewm(span=100,adjust=False).mean()

    df["ema8_21_spread"] = (ema8 - ema21) / close
    df["ema21_50_spread"]= (ema21 - ema50) / close
    df["ema50_100_spread"]= (ema50 - ema100) / close

    df["price_vs_ema21"] = (close - ema21) / close
    df["price_vs_ema50"] = (close - ema50) / close

    df["trend_align"] = np.where(
        (ema8 > ema21) & (ema21 > ema50), 1,
        np.where((ema8 < ema21) & (ema21 < ema50), -1, 0)
    )

    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_mid  = close.rolling(20).mean()
    bb_std  = close.rolling(20).std(ddof=0)
    bb_up   = bb_mid + 2 * bb_std
    bb_lo   = bb_mid - 2 * bb_std
    bb_bw   = (bb_up - bb_lo) / bb_mid          # bandwidth
    bb_pct  = (close - bb_lo) / (bb_up - bb_lo) # %B

    df["bb_pct"]      = bb_pct
    df["bb_bw"]       = bb_bw
    df["bb_bw_ratio"] = bb_bw / bb_bw.rolling(50).mean()  # squeeze detector
    df["bb_upper_touch"] = np.where(close >= bb_up, 1, 0)
    df["bb_lower_touch"] = np.where(close <= bb_lo, 1, 0)

    # ── Stochastic %K %D (14, 3) ────────────────────────────────────────────
    lo14   = low.rolling(14).min()
    hi14   = high.rolling(14).max()
    stoch_k = 100 * (close - lo14) / (hi14 - lo14).replace(0, np.nan)
    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
    df["stoch_ob"]     = np.where(stoch_k > 80, 1, 0)
    df["stoch_os"]     = np.where(stoch_k < 20, 1, 0)

    # ── Candle structure ────────────────────────────────────────────────────
    body      = (close - open_).abs()
    candle_rng= (high - low).replace(0, np.nan)
    df["body_ratio"]   = body / candle_rng        # body vs full range
    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, -1)
    df["candle_dir_3"] = df["candle_dir"].rolling(3).sum()  # short-term bias

    # ── Volume-less momentum oscillator (Williams %R 14) ───────────────────
    df["williams_r"] = -100 * (hi14 - close) / (hi14 - lo14).replace(0, np.nan)

    # ── RSI x MACD composite signal ─────────────────────────────────────────
    df["rsi_macd_bull"] = np.where(
        (df["rsi_14"] > 50) & (macd_hist > 0), 1,
        np.where((df["rsi_14"] < 50) & (macd_hist < 0), -1, 0)
    )

    # ── Divergence proxy: price vs RSI direction (3-bar) ────────────────────
    price_dir3 = np.sign(close.diff(3))
    rsi_dir3   = np.sign(df["rsi_14"].diff(3))
    df["rsi_div"] = np.where(price_dir3 != rsi_dir3, 1, 0)

    # ── Mean-reversion signal: distance from 50-bar mean normalised by ATR ──
    sma50 = close.rolling(50).mean()
    df["zscore_50"] = (close - sma50) / (close.rolling(50).std(ddof=0).replace(0, np.nan))

    df["mean_rev_long"]  = np.where(df["zscore_50"] < -1.5, 1, 0)
    df["mean_rev_short"] = np.where(df["zscore_50"] >  1.5, 1, 0)

    # ── Interaction features ─────────────────────────────────────────────────
    df["rsi_bb_pct"]        = df["rsi_14"] * df["bb_pct"]
    df["macd_hist_rsi_mid"] = df["macd_hist"] * df["rsi_mid"]
    df["stoch_rsi"]         = df["stoch_k"] * df["rsi_14"] / 1e4  # normalised product

    # ── Lag features (avoid lookahead) ──────────────────────────────────────
    for lag in [1, 2, 4, 8]:
        df[f"rsi_lag{lag}"]       = df["rsi_14"].shift(lag)
        df[f"macd_hist_lag{lag}"] = df["macd_hist"].shift(lag)
        df[f"ret_lag{lag}"]       = df["ret_1"].shift(lag)

    # ── Hour-of-day & day-of-week cyclic encoding ───────────────────────────
    if hasattr(df.index, "hour"):
        hour = df.index.hour
        dow  = df.index.dayofweek
        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)

    # ── Final fill ───────────────────────────────────────────────────────────
    df = df.bfill().ffill()

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "NZD/USD RSI-MACD Gradient Boost Risk-Adjusted",
        "model_type":       "GradientBoostingClassifier",
        "model_params":     {
            "n_estimators":     600,
            "max_depth":        4,
            "learning_rate":    0.03,
            "subsample":        0.75,
            "max_features":     "sqrt",
            "min_samples_leaf": 20,
            "min_samples_split":40,
            "warm_start":       False,
        },
        "signal_threshold": 0.56,
        "direction":        "both",
        "stop_loss":        0.005,
        "take_profit":      0.010,
        "cooldown":         0,
        "max_positions":    1,
        "on_opposite":      "reverse",
        "session_filter":   [21, 21],
        "min_atr":          0.0002,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe/Calmar) using a deep "
            "GradientBoostingClassifier with many slow-learning trees and "
            "aggressive regularisation (min_samples_leaf=20, subsample=0.75). "
            "Feature set deliberately differs from prior RSI+BB+Stoch attempts "
            "by adding: ATR-normalised returns, z-score mean-reversion signals, "
            "RSI divergence proxy, Williams %R, candle structure ratios, cyclic "
            "time encoding, and interaction/lag features to give the model richer "
            "multi-timeframe context. SL=0.5%/TP=1% gives 1:2 RR aligned with "
            "maximising Sharpe."
        ),
        "notes": (
            "Prior PF=1.35 / ret=+18.36% used standard RSI+MACD+BB+Stoch without "
            "ATR normalisation or divergence detection. This version adds z-score "
            "mean-reversion context, candle structure, and temporal encoding to "
            "reduce false positives. session_filter=[21,21] is intentionally "
            "narrow — set to None if you want 24h coverage. min_atr=0.0002 "
            "avoids dead-market signals."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 01:54:56
# Model            : Gradient Boosting
# Feature Eng.     : BB (20,2.0), RSI 14, Stochastic (14,3) + 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/NZDUSD_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):
    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_period = 20
    bb_std    = 2.0
    bb_mid    = close.rolling(bb_period).mean()
    bb_std_   = close.rolling(bb_period).std(ddof=0)
    bb_upper  = bb_mid + bb_std * bb_std_
    bb_lower  = bb_mid - bb_std * 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
    df["bb_pct"]   = (close - bb_lower) / (bb_upper - bb_lower)

    # ── RSI (14) ─────────────────────────────────────────────────────────────
    rsi_period = 14
    delta      = close.diff()
    gain       = delta.clip(lower=0)
    loss       = -delta.clip(upper=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         = avg_gain / avg_loss.replace(0, np.nan)
    df["rsi"]  = 100 - (100 / (1 + rs))

    # ── Stochastic Oscillator (K=14, D=3) ────────────────────────────────────
    stoch_k_period = 14
    stoch_d_period = 3
    lowest_low   = low.rolling(stoch_k_period).min()
    highest_high = high.rolling(stoch_k_period).max()
    stoch_range  = (highest_high - lowest_low).replace(0, np.nan)
    df["stoch_k"] = 100 * (close - lowest_low) / stoch_range
    df["stoch_d"] = df["stoch_k"].rolling(stoch_d_period).mean()
    df["stoch_kd_diff"] = df["stoch_k"] - df["stoch_d"]

    # ── ATR (14) — for normalised volatility / min_atr filter ────────────────
    atr_period = 14
    tr = pd.concat([
        high - low,
        (high - close.shift(1)).abs(),
        (low  - close.shift(1)).abs()
    ], axis=1).max(axis=1)
    atr = tr.ewm(com=atr_period - 1, min_periods=atr_period).mean()
    df["atr"]  = atr
    df["natr"] = atr / close          # normalised ATR used by min_atr filter

    # ── Price momentum / rate-of-change ──────────────────────────────────────
    df["roc_4"]  = close.pct_change(4)    # 1-hour momentum on 15-min bars
    df["roc_8"]  = close.pct_change(8)    # 2-hour momentum
    df["roc_16"] = close.pct_change(16)   # 4-hour momentum

    # ── EMA trend context ─────────────────────────────────────────────────────
    df["ema_20"]  = close.ewm(span=20,  adjust=False).mean()
    df["ema_50"]  = close.ewm(span=50,  adjust=False).mean()
    df["ema_100"] = close.ewm(span=100, adjust=False).mean()
    df["sma_50"]  = close.rolling(50).mean()   # used by trend_filter

    df["ema_cross_20_50"]  = df["ema_20"]  - df["ema_50"]
    df["ema_cross_50_100"] = df["ema_50"]  - df["ema_100"]
    df["close_vs_ema20"]   = (close - df["ema_20"]) / df["ema_20"]

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

    # ── Volume-proxy: realised range rolling stats ────────────────────────────
    df["hl_range"]      = high - low
    df["hl_range_ma8"]  = df["hl_range"].rolling(8).mean()
    df["hl_range_ratio"]= df["hl_range"] / df["hl_range_ma8"]

    # ── RSI derived signals ───────────────────────────────────────────────────
    df["rsi_overbought"]  = np.where(df["rsi"] > 70, 1.0, 0.0)
    df["rsi_oversold"]    = np.where(df["rsi"] < 30, 1.0, 0.0)
    df["rsi_momentum"]    = df["rsi"].diff(4)

    # ── Stochastic derived signals ────────────────────────────────────────────
    df["stoch_overbought"] = np.where(df["stoch_k"] > 80, 1.0, 0.0)
    df["stoch_oversold"]   = np.where(df["stoch_k"] < 20, 1.0, 0.0)

    # ── BB squeeze: width vs rolling mean of width ────────────────────────────
    df["bb_width_ma20"]  = df["bb_width"].rolling(20).mean()
    df["bb_squeeze"]     = np.where(df["bb_width"] < df["bb_width_ma20"], 1.0, 0.0)

    # ── Interaction features ──────────────────────────────────────────────────
    df["rsi_bb_pct"]      = df["rsi"] * df["bb_pct"]
    df["stoch_k_bb_pct"]  = df["stoch_k"] * df["bb_pct"]
    df["rsi_stoch_diff"]  = df["rsi"] - df["stoch_k"]

    # ── Lagged features (avoids look-ahead) ──────────────────────────────────
    for lag in [1, 2, 3, 4]:
        df[f"rsi_lag{lag}"]     = df["rsi"].shift(lag)
        df[f"stoch_k_lag{lag}"] = df["stoch_k"].shift(lag)
        df[f"bb_pct_lag{lag}"]  = df["bb_pct"].shift(lag)
        df[f"roc4_lag{lag}"]    = df["roc_4"].shift(lag)

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


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "NZD/USD Stoch+BB+RSI Gradient Boosting Mean-Revert",
        "model_type":       "GradientBoostingClassifier",
        "model_params":     {
            "n_estimators":     400,
            "max_depth":        4,
            "learning_rate":    0.04,
            "subsample":        0.8,
            "min_samples_leaf": 20,
            "max_features":     "sqrt",
            "validation_fraction": 0.1,
            "n_iter_no_change": 30,
            "tol":              1e-4,
            "random_state":     42,
        },
        "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, 20],
        "min_atr":          0.0002,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximise risk-adjusted return (Sharpe / Calmar) on NZD/USD 15-min. "
            "GradientBoostingClassifier selected for its strong generalisation on "
            "structured tabular data with noisy financial features. n_estimators=400 "
            "with early stopping (n_iter_no_change=30) prevents overfitting. "
            "max_depth=4 keeps trees shallow to reduce variance. subsample=0.8 + "
            "max_features=sqrt add stochasticity for robustness. SL 0.5% / TP 1.0% "
            "gives a minimum 2:1 reward-risk ratio. Session filter 07-20 UTC covers "
            "Sydney open through NY overlap, maximising NZD/USD liquidity. "
            "Reverse on opposite signal keeps the model continuously positioned in "
            "the highest-confidence direction. min_atr filter avoids flat/illiquid "
            "periods where the model edges degrade."
        ),
        "notes": (
            "Features: Bollinger Bands (20,2) width & %B, RSI(14), Stochastic K/D "
            "(14,3), ATR(14)/NATR, EMA cross (20/50/100), SMA50 trend context, "
            "price ROC (4/8/16 bars), candlestick body/wick ratios, HL range "
            "normalisation, BB squeeze flag, RSI/Stoch overbought-oversold flags, "
            "interaction terms (RSI*%B, StochK*%B), and 4 lags each of RSI, StochK, "
            "%B and ROC4. Threshold 0.56 slightly above 0.50 to filter marginal "
            "signals without sacrificing too many trades."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:36:58
# Model            : XGBoost
# Feature Eng.     : BB (20,2.0), 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/USDCAD_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):
    # ── 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
    df["bb_width"] = (bb_upper - bb_lower) / bb_mid
    df["bb_pct"]   = (close - bb_lower) / (bb_upper - bb_lower)

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

    atr  = tr.ewm(alpha=1.0 / atr_period, min_periods=atr_period, adjust=False).mean()
    natr = atr / close

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

    # ── Price momentum / returns ─────────────────────────────────────────────
    df["ret_1"]  = close.pct_change(1)
    df["ret_4"]  = close.pct_change(4)
    df["ret_8"]  = close.pct_change(8)
    df["ret_16"] = close.pct_change(16)

    # ── Distance from Bollinger mid / bands ──────────────────────────────────
    df["close_minus_mid"]   = (close - bb_mid)   / bb_mid
    df["close_minus_upper"] = (close - bb_upper) / bb_mid
    df["close_minus_lower"] = (close - bb_lower) / bb_mid

    # ── BB squeeze flag: width below rolling median ───────────────────────────
    bb_width_med = df["bb_width"].rolling(50).median()
    df["bb_squeeze"] = np.where(df["bb_width"] < bb_width_med, 1.0, 0.0)

    # ── BB mean-reversion z-score ────────────────────────────────────────────
    df["bb_z"] = (close - bb_mid) / (bb_sigma + 1e-12)

    # ── Candle body / wick features ──────────────────────────────────────────
    body       = (close - open_).abs()
    candle_rng = (high - low).replace(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["bull_candle"]  = np.where(close > open_, 1.0, 0.0)

    # ── RSI (14) built from scratch ──────────────────────────────────────────
    rsi_period = 14
    delta  = close.diff()
    gain   = delta.clip(lower=0)
    loss   = (-delta).clip(lower=0)
    avg_g  = gain.ewm(alpha=1.0 / rsi_period, min_periods=rsi_period, adjust=False).mean()
    avg_l  = loss.ewm(alpha=1.0 / rsi_period, min_periods=rsi_period, adjust=False).mean()
    rs     = avg_g / (avg_l + 1e-12)
    rsi    = 100.0 - (100.0 / (1.0 + rs))
    df["rsi_14"] = rsi

    # RSI deviation from neutral 50
    df["rsi_dev"] = (rsi - 50.0) / 50.0

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

    # ── Rolling volatility (realised over 20 bars) ───────────────────────────
    df["vol_20"] = df["ret_1"].rolling(20).std()

    # ── ATR z-score vs 50-bar rolling mean ───────────────────────────────────
    atr_mean     = atr.rolling(50).mean()
    atr_std      = atr.rolling(50).std(ddof=0)
    df["atr_z"]  = (atr - atr_mean) / (atr_std + 1e-12)

    # ── Volume-of-BB-touches over last 10 bars ───────────────────────────────
    near_upper = (close >= bb_upper * 0.998).astype(float)
    near_lower = (close <= bb_lower * 1.002).astype(float)
    df["touch_upper_10"] = near_upper.rolling(10).sum()
    df["touch_lower_10"] = near_lower.rolling(10).sum()

    # ── SMA 50 (trend filter helper) ─────────────────────────────────────────
    df["sma_50"]       = close.rolling(50).mean()
    df["close_vs_sma"] = (close - df["sma_50"]) / df["sma_50"]

    # ── EMA cross (9 / 21) ───────────────────────────────────────────────────
    ema9  = close.ewm(span=9,  adjust=False).mean()
    ema21 = close.ewm(span=21, adjust=False).mean()
    df["ema_cross"] = (ema9 - ema21) / close

    # ── Bar-of-day / session ─────────────────────────────────────────────────
    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

    # ── Lag features on bb_pct and rsi ───────────────────────────────────────
    for lag in [1, 2, 4]:
        df[f"bb_pct_lag{lag}"]  = df["bb_pct"].shift(lag)
        df[f"rsi_14_lag{lag}"]  = df["rsi_14"].shift(lag)
        df[f"macd_hist_lag{lag}"] = df["macd_hist"].shift(lag)

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

    return df


# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "USD/CAD BB Mean-Reversion + ATR 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":   [7, 20],
        "min_atr":          0.0002,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximise risk-adjusted return (Sharpe/Calmar) on USD/CAD 15-min "
            "using Bollinger Band mean-reversion signals augmented by ATR, RSI, "
            "MACD, and EMA-cross features fed into a regularised XGBoost classifier. "
            "SL=0.5% / TP=1.0% gives a 1:2 RR floor. Conservative depth (4) and "
            "strong L1/L2 regularisation prevent overfitting on a single year of data."
        ),
        "notes": (
            "BB squeeze flag and bb_z capture regime; atr_z filters noisy bars. "
            "Session filter 07-20 UTC covers London + NY overlap for tighter spreads. "
            "min_atr=0.0002 avoids dead-market whipsaws. Lag features on bb_pct and "
            "rsi_14 give the model short-term momentum context without look-ahead."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 03:05:53
# Model            : Gradient Boosting
# Feature Eng.     : RSI 14, MACD (12,26,9) + 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/AUDUSD_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):
    # ── RSI 14 ──────────────────────────────────────────────────────────────
    period = 14
    delta = close.diff()
    gain = delta.clip(lower=0)
    loss = -delta.clip(upper=0)
    avg_gain = gain.ewm(com=period - 1, min_periods=period).mean()
    avg_loss = loss.ewm(com=period - 1, min_periods=period).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()
    macd_hist = macd_line - signal_line
    df["macd_line"] = macd_line
    df["macd_signal"] = signal_line
    df["macd_hist"] = macd_hist

    # ── RSI derived features ─────────────────────────────────────────────────
    df["rsi_14_lag1"] = df["rsi_14"].shift(1)
    df["rsi_14_lag2"] = df["rsi_14"].shift(2)
    df["rsi_14_delta"] = df["rsi_14"] - df["rsi_14_lag1"]
    df["rsi_ob"] = np.where(df["rsi_14"] > 70, 1, 0)
    df["rsi_os"] = np.where(df["rsi_14"] < 30, 1, 0)

    # ── MACD derived features ────────────────────────────────────────────────
    df["macd_hist_lag1"] = macd_hist.shift(1)
    df["macd_hist_delta"] = macd_hist - macd_hist.shift(1)
    df["macd_cross_bull"] = np.where((macd_line > signal_line) & (macd_line.shift(1) <= signal_line.shift(1)), 1, 0)
    df["macd_cross_bear"] = np.where((macd_line < signal_line) & (macd_line.shift(1) >= signal_line.shift(1)), 1, 0)
    df["macd_above_zero"] = np.where(macd_line > 0, 1, 0)
    df["macd_hist_positive"] = np.where(macd_hist > 0, 1, 0)

    # ── 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_14 = tr.ewm(span=14, adjust=False).mean()
    df["atr_14"] = atr_14
    df["natr_14"] = atr_14 / close

    # ── Bollinger Bands (20, 2) ───────────────────────────────────────────────
    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_b"] = (close - bb_lower) / (bb_upper - bb_lower).replace(0, np.nan)
    df["bb_width"] = (bb_upper - bb_lower) / bb_mid.replace(0, np.nan)
    df["bb_squeeze"] = np.where(df["bb_width"] < df["bb_width"].rolling(50).mean(), 1, 0)

    # ── Price momentum features ───────────────────────────────────────────────
    df["ret_1"] = close.pct_change(1)
    df["ret_4"] = close.pct_change(4)
    df["ret_8"] = close.pct_change(8)
    df["ret_16"] = close.pct_change(16)

    # ── Rolling volatility ────────────────────────────────────────────────────
    df["vol_8"] = df["ret_1"].rolling(8).std()
    df["vol_20"] = df["ret_1"].rolling(20).std()

    # ── EMA trend features ───────────────────────────────────────────────────
    ema_20 = close.ewm(span=20, adjust=False).mean()
    ema_50 = close.ewm(span=50, adjust=False).mean()
    df["ema_20"] = ema_20
    df["ema_50"] = ema_50
    df["price_vs_ema20"] = (close - ema_20) / ema_20
    df["price_vs_ema50"] = (close - ema_50) / ema_50
    df["ema20_vs_ema50"] = (ema_20 - ema_50) / ema_50

    # ── Candle body / 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["bull_candle"] = np.where(close > open_, 1, 0)

    # ── Rolling high/low position ─────────────────────────────────────────────
    roll_high_20 = high.rolling(20).max()
    roll_low_20 = low.rolling(20).min()
    roll_range_20 = (roll_high_20 - roll_low_20).replace(0, np.nan)
    df["price_position_20"] = (close - roll_low_20) / roll_range_20

    # ── RSI + MACD interaction ────────────────────────────────────────────────
    df["rsi_macd_product"] = df["rsi_14"] * macd_hist
    df["rsi_norm"] = (df["rsi_14"] - 50) / 50

    # ── Session / time features ───────────────────────────────────────────────
    if hasattr(close.index, "hour"):
        df["hour"] = close.index.hour
        df["session_london"] = np.where((close.index.hour >= 7) & (close.index.hour < 16), 1, 0)
        df["session_ny"] = np.where((close.index.hour >= 13) & (close.index.hour < 21), 1, 0)
        df["session_asia"] = np.where((close.index.hour >= 22) | (close.index.hour < 7), 1, 0)
    else:
        df["hour"] = 0
        df["session_london"] = 0
        df["session_ny"] = 0
        df["session_asia"] = 0

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

    return df

# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title": "AUD/USD RSI+MACD Gradient Boosting Scalper",
        "model_type": "GradientBoostingClassifier",
        "model_params": {
            "n_estimators": 400,
            "max_depth": 4,
            "learning_rate": 0.05,
            "subsample": 0.8,
            "min_samples_leaf": 20,
            "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": 0.01,
        "cooldown": 0,
        "max_positions": 1,
        "on_opposite": "reverse",
        "session_filter": None,
        "min_atr": None,
        "trend_filter": None,
        "target_horizon": 4,
        "objective": (
            "Maximize risk-adjusted return on AUD/USD 15-min data using GradientBoostingClassifier. "
            "RSI-14 captures momentum extremes and divergence conditions; MACD (12,26,9) provides "
            "trend direction and momentum shifts via crossovers and histogram slope. Additional "
            "features (BB, ATR, EMA trend, candle structure, session timing) enrich the feature "
            "space. GradientBoosting with shallow trees (depth=4), moderate learning rate (0.05), "
            "and early stopping via n_iter_no_change prevents overfitting. SL=0.5%, TP=1.0% gives "
            "a 1:2 risk/reward ratio, improving Sharpe and Calmar. Threshold=0.55 filters low-confidence "
            "signals to reduce noise. Reverse on opposite signal maximizes capital efficiency."
        ),
        "notes": (
            "n_estimators=400 with early stopping balances bias-variance. max_depth=4 keeps trees "
            "shallow to reduce overfitting on FX microstructure noise. subsample=0.8 adds stochastic "
            "gradient boosting regularization. min_samples_leaf=20 prevents fitting to outlier bars. "
            "max_features='sqrt' adds feature randomization similar to random forests. "
            "target_horizon=4 (1 hour) aligns with typical AUD/USD intraday swing durations."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 02:24:20
# Model            : Gradient Boosting
# Feature Eng.     : BB (20,2.0), RSI 14, Stochastic (14,3) + 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/AUDUSD_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):
    # ── Bollinger Bands (20, 2) ──────────────────────────────────────────────
    bb_period = 20
    bb_std    = 2.0
    bb_mid    = close.rolling(bb_period).mean()
    bb_std_   = close.rolling(bb_period).std(ddof=1)
    bb_upper  = bb_mid + bb_std * bb_std_
    bb_lower  = bb_mid - bb_std * 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
    df["bb_pct"]   = (close - bb_lower) / (bb_upper - bb_lower)

    # ── RSI (14) ─────────────────────────────────────────────────────────────
    rsi_period = 14
    delta      = close.diff()
    gain       = delta.clip(lower=0)
    loss       = -delta.clip(upper=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         = avg_gain / avg_loss.replace(0, np.nan)
    df["rsi"]  = 100.0 - (100.0 / (1.0 + rs))

    # ── Stochastic Oscillator (K=14, D=3) ────────────────────────────────────
    stoch_k = 14
    stoch_d = 3
    low_min  = low.rolling(stoch_k).min()
    high_max = high.rolling(stoch_k).max()
    k_raw    = 100.0 * (close - low_min) / (high_max - low_min).replace(0, np.nan)
    df["stoch_k"] = k_raw
    df["stoch_d"] = k_raw.rolling(stoch_d).mean()

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

    # ── Trend / Momentum features ─────────────────────────────────────────────
    df["sma_20"]  = close.rolling(20).mean()
    df["sma_50"]  = close.rolling(50).mean()
    df["sma_100"] = close.rolling(100).mean()

    df["price_vs_sma20"]  = (close - df["sma_20"])  / df["sma_20"]
    df["price_vs_sma50"]  = (close - df["sma_50"])  / df["sma_50"]
    df["price_vs_sma100"] = (close - df["sma_100"]) / df["sma_100"]
    df["sma20_vs_sma50"]  = (df["sma_20"] - df["sma_50"]) / df["sma_50"]

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

    # ── Rate-of-Change features ───────────────────────────────────────────────
    for p in [4, 8, 16]:
        df[f"roc_{p}"] = close.pct_change(p)

    # ── Volatility regime ────────────────────────────────────────────────────
    df["vol_8"]  = close.pct_change().rolling(8).std()
    df["vol_20"] = close.pct_change().rolling(20).std()
    df["vol_ratio"] = df["vol_8"] / df["vol_20"].replace(0, np.nan)

    # ── Candle body / shadow features ────────────────────────────────────────
    df["body"]        = (close - open_).abs()
    df["upper_shadow"] = high - pd.concat([close, open_], axis=1).max(axis=1)
    df["lower_shadow"] = pd.concat([close, open_], axis=1).min(axis=1) - low
    df["body_ratio"]  = df["body"] / (high - low).replace(0, np.nan)

    # ── RSI-derived features ──────────────────────────────────────────────────
    df["rsi_above_50"] = np.where(df["rsi"] > 50, 1, 0)
    df["rsi_overbought"] = np.where(df["rsi"] > 70, 1, 0)
    df["rsi_oversold"]   = np.where(df["rsi"] < 30, 1, 0)
    df["rsi_lag1"] = df["rsi"].shift(1)
    df["rsi_lag4"] = df["rsi"].shift(4)

    # ── Stochastic-derived features ───────────────────────────────────────────
    df["stoch_cross_up"]   = np.where((df["stoch_k"] > df["stoch_d"]) &
                                      (df["stoch_k"].shift(1) <= df["stoch_d"].shift(1)), 1, 0)
    df["stoch_cross_down"] = np.where((df["stoch_k"] < df["stoch_d"]) &
                                      (df["stoch_k"].shift(1) >= df["stoch_d"].shift(1)), 1, 0)
    df["stoch_oversold"]   = np.where(df["stoch_k"] < 20, 1, 0)
    df["stoch_overbought"] = np.where(df["stoch_k"] > 80, 1, 0)

    # ── BB-derived features ───────────────────────────────────────────────────
    df["bb_squeeze"] = np.where(df["bb_width"] < df["bb_width"].rolling(50).quantile(0.20), 1, 0)
    df["above_bb_upper"] = np.where(close > bb_upper, 1, 0)
    df["below_bb_lower"] = np.where(close < bb_lower, 1, 0)
    df["bb_pct_lag1"]    = df["bb_pct"].shift(1)
    df["bb_pct_lag4"]    = df["bb_pct"].shift(4)

    # ── Session hour (UTC) ────────────────────────────────────────────────────
    df["hour_utc"] = df.index.hour if hasattr(df.index, "hour") else 0

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

# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "AUD/USD Stochastic BB Mean-Reversion (GBM)",
        "model_type":       "GradientBoostingClassifier",
        "model_params":     {
            "n_estimators":      400,
            "max_depth":         4,
            "learning_rate":     0.04,
            "subsample":         0.80,
            "min_samples_leaf":  20,
            "max_features":      "sqrt",
            "n_iter_no_change":  30,
            "validation_fraction": 0.10,
            "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":   None,
        "min_atr":          None,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective": (
            "Maximize risk-adjusted return (Sharpe/Calmar) on AUD/USD 15-min. "
            "GradientBoostingClassifier with moderate depth and learning rate chosen "
            "to balance bias-variance. 2:1 reward-to-risk (SL=0.5%, TP=1.0%). "
            "Stochastic crossovers, BB mean-reversion, and RSI regime signals "
            "form the core feature set; MACD, volatility, and candle features add "
            "context. Early stopping (n_iter_no_change=30) prevents overfitting."
        ),
        "notes": (
            "Features: Bollinger Bands (20,2) width/pct, RSI(14) with lag/regime flags, "
            "Stochastic(14,3) K/D with crossover detection, ATR/NATR volatility, MACD "
            "histogram, short/medium SMAs, ROC(4/8/16), volatility ratio, candle body "
            "ratios, and UTC session hour. No session or trend filter to allow full "
            "mean-reversion opportunities across all sessions."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  STRATEGY REQUEST LOG                       ║
# ╚══════════════════════════════════════════════════════════════╝
# Generated        : 2026-05-06 01:40:19
# 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/AUDUSD_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 crossover signal: positive when fast > slow
    df["ema_cross"]       = ema_9 - ema_21
    df["ema_cross_prev"]  = df["ema_cross"].shift(1)
    # Binary: did a cross just occur?
    df["ema_cross_up"]    = np.where((df["ema_cross"] > 0) & (df["ema_cross_prev"] <= 0), 1, 0)
    df["ema_cross_down"]  = np.where((df["ema_cross"] < 0) & (df["ema_cross_prev"] >= 0), 1, 0)
    # Trend direction encoded as -1 / 1
    df["ema_trend"]       = np.where(ema_9 > ema_21, 1, -1)

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

    # RSI regime flags
    df["rsi_oversold"]   = np.where(rsi_14 < 30, 1, 0)
    df["rsi_overbought"] = np.where(rsi_14 > 70, 1, 0)
    df["rsi_mid"]        = rsi_14 - 50          # centred

    # RSI momentum (1-bar change in RSI)
    df["rsi_delta"]      = rsi_14.diff(1)
    df["rsi_delta2"]     = rsi_14.diff(3)

    # ── Additional momentum / volatility features ─────────────────────────
    # ATR-like normalised 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)
    atr_14 = tr.ewm(span=14, adjust=False).mean()
    df["atr_14"]  = atr_14
    df["natr_14"] = atr_14 / close          # normalised ATR

    # Rate-of-change over various horizons
    for n in [4, 8, 16]:
        df[f"roc_{n}"] = close.pct_change(n)

    # Bollinger Band width and %B (using 20-period SMA)
    sma_20   = close.rolling(20).mean()
    std_20   = close.rolling(20).std()
    bb_upper = sma_20 + 2 * std_20
    bb_lower = sma_20 - 2 * std_20
    df["bb_width"] = (bb_upper - bb_lower) / sma_20
    df["bb_pct"]   = (close - bb_lower) / (bb_upper - bb_lower).replace(0, np.nan)

    # Candle body and wick features
    df["body"]        = (close - open_) / close
    df["upper_wick"]  = (high - pd.concat([close, open_], axis=1).max(axis=1)) / close
    df["lower_wick"]  = (pd.concat([close, open_], axis=1).min(axis=1) - low)  / close

    # Volume-normalised momentum proxy: price range relative to ATR
    df["range_vs_atr"] = (high - low) / atr_14.replace(0, np.nan)

    # Lagged EMA cross signal
    df["ema_cross_lag1"] = df["ema_cross"].shift(1)
    df["ema_cross_lag2"] = df["ema_cross"].shift(2)

    # Combined signal: RSI and EMA cross alignment
    df["rsi_ema_bull"] = np.where((rsi_14 > 50) & (ema_9 > ema_21), 1, 0)
    df["rsi_ema_bear"] = np.where((rsi_14 < 50) & (ema_9 < ema_21), 1, 0)

    # Hour-of-day (cyclical encoding) — no lookahead
    hour = df.index.hour
    df["hour_sin"] = np.sin(2 * np.pi * hour / 24)
    df["hour_cos"] = np.cos(2 * np.pi * hour / 24)

    # Day-of-week (cyclical encoding)
    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 warm-up periods
    df = df.bfill().ffill()

    return df

# SECTION 2 — STRATEGY CONFIG
def strategy_config():
    return {
        "title":            "AUD/USD EMA Cross (9/21) + RSI14 XGBoost Scalper",
        "model_type":       "XGBClassifier",
        "model_params":     {
            "n_estimators":     400,
            "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",
            "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":   [0, 23],
        "min_atr":          None,
        "trend_filter":     None,
        "target_horizon":   4,
        "objective":        (
            "Maximise risk-adjusted return on AUD/USD 15-min bars. "
            "XGBoost chosen for its ability to capture non-linear interactions between "
            "the EMA-cross regime, RSI momentum, volatility (NATR/BB width), and time-of-day. "
            "Shallow trees (max_depth=4) with strong regularisation (reg_lambda=1.5, gamma=0.1) "
            "reduce overfitting on the limited 1-year window. "
            "2:1 R:R (SL=0.5%, TP=1.0%) improves Sharpe; reverse on opposite signal captures "
            "trend momentum without missing transitions."
        ),
        "notes":            (
            "Features: EMA-9/21 cross and distances, RSI-14 with regime flags and delta, "
            "ATR-14, NATR, Bollinger Band width/%B, 4/8/16-bar ROC, candle anatomy, "
            "time cyclical encodings. Threshold 0.54 filters marginal signals to raise precision. "
            "No session filter applied — AUD/USD has meaningful moves across Asian and London sessions."
        ),
    }

# ╔══════════════════════════════════════════════════════════════╗
# ║                  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
    )