ml_config.py

"""
ml_config.py — All hyperparameters for the ML probability filter layer.
Edit here only; never hardcode values in other modules.
"""
from pathlib import Path

# ── PATHS ─────────────────────────────────────────────────────────────────────
ML_DIR = Path(__file__).parent / "ml_artifacts"
MODEL_PATH       = ML_DIR / "trade_filter.pkl"
THRESHOLD_PATH   = ML_DIR / "threshold.json"
FEATURE_IMP_PATH = ML_DIR / "feature_importance.csv"
LABEL_PATH       = ML_DIR / "label_stats.json"

# ── LABELING ──────────────────────────────────────────────────────────────────
# How many forward bars to check for target/stop hit.
# 1H timeframe → 24 bars = 1 trading day lookahead. Good balance of
# recency vs enough bars for a 1:2 RR to play out.
LABEL_FORWARD_BARS = 24

# Realistic costs: 0.06% taker fee each side + 0.04% slippage each side
TRADE_FEE_PCT     = 0.0006   # 0.06% taker fee per side
TRADE_SLIP_PCT    = 0.0004   # 0.04% slippage per side
ROUND_TRIP_COST   = (TRADE_FEE_PCT + TRADE_SLIP_PCT) * 2  # both sides

# ATR multipliers matching risk_engine.py
STOP_MULT  = 2.5
TARGET_RR  = 2.0   # target = stop_distance * TARGET_RR

# ── WALK-FORWARD ──────────────────────────────────────────────────────────────
WF_N_SPLITS      = 5      # number of walk-forward folds
WF_TRAIN_FRAC    = 0.70   # fraction of each fold used for training
WF_MIN_TRAIN_OBS = 500    # minimum training observations per fold

# ── MODEL HYPERPARAMETERS ─────────────────────────────────────────────────────
# These target LightGBM params; HistGradientBoostingClassifier maps them.
LGBM_PARAMS = dict(
    n_estimators      = 400,
    learning_rate     = 0.03,
    max_depth         = 5,         # shallow: reduces overfitting
    min_samples_leaf  = 40,        # minimum leaf size: ~1% of 4000 samples
    l2_regularization = 2.0,       # L2 ridge penalty
    max_features      = 0.70,      # feature bagging per split
    early_stopping_rounds = 30,
    validation_fraction   = 0.15,
    n_iter_no_change      = 30,
    random_state          = 42,
    verbose               = 0,
)

# ── THRESHOLD OPTIMIZATION ────────────────────────────────────────────────────
# Objective to maximize when searching for the optimal probability cutoff.
# Options: "sharpe", "expectancy", "f1", "precision_recall"
THRESHOLD_OBJECTIVE = "expectancy"

# Search grid for threshold sweep
THRESHOLD_MIN   = 0.35
THRESHOLD_MAX   = 0.80
THRESHOLD_STEPS = 91   # 0.35, 0.36, ..., 0.80

# ── INFERENCE ─────────────────────────────────────────────────────────────────
DEFAULT_PROB_THRESHOLD = 0.55    # conservative default before calibration

# ── FEATURE ENGINEERING ───────────────────────────────────────────────────────
# Raw features from the rule engine fed into the model.
# Order here defines column order in the feature matrix — DO NOT CHANGE
# without retraining.
FEATURE_COLUMNS = [
    # Trend / momentum
    "adx",
    "di_plus",
    "di_minus",
    "di_diff",          # engineered: di_plus - di_minus
    "di_ratio",         # engineered: di_plus / (di_plus + di_minus + 1e-9)
    # Volatility
    "atr_pct",
    "vol_ratio",
    "vol_compressed",
    "vol_expanding",
    "vol_expanding_from_base",
    # Volume / order flow
    "absorption",
    "failed_breakout",
    "recent_failed_count",
    "obv_slope_norm",
    "delta_sign",
    "spike",
    "climax",
    # Price context
    "dist_atr",
    "dist_atr_abs",     # engineered: abs(dist_atr)
    # Rule-engine scores (carry human priors into the model)
    "regime_confidence",
    "regime_score",
    "volume_score",
    "structure_score",
    "confidence_score",
    "total_score",
    # Interactions (multiplicative signal combinations)
    "adx_x_regime",     # engineered: adx * regime_score
    "vol_x_obv",        # engineered: vol_ratio * obv_slope_norm
    "score_x_conf",     # engineered: total_score * regime_confidence
]