File size: 18,003 Bytes
031a2d6 90420df 5f9784b 031a2d6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 | """
Phase 8 β Explainability Engine (SHAP + ElasticNet Coefficients)
Consumes outputs from Phase 6 (trained models + preprocessors + test data).
No model retraining. No feature engineering.
Produces:
data/explainability/
shap_global_xgb_{asset}.csv - mean |SHAP| per feature (XGBoost)
shap_global_en_{asset}.csv - mean |SHAP| per feature (ElasticNet)
shap_values_xgb_{asset}.csv - full SHAP matrix, test set (XGBoost)
shap_values_en_{asset}.csv - full SHAP matrix, test set (ElasticNet)
elastic_net_coefficients_{asset}.csv - actual EN coefficients from pkl model
feature_importance_comparison.csv - unified table: XGB gain | SHAP | EN coeff
latest_prediction_explanation.json - per-asset local explanation, most recent row
explainability_report.json - full summary
"""
from __future__ import annotations
import json
import logging
import time
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import joblib
import numpy as np
import pandas as pd
import shap
logging.basicConfig(level=logging.INFO, format="%(asctime)s | %(levelname)s | %(message)s")
logger = logging.getLogger(__name__)
ASSETS = ["spx", "ndx", "gold", "btc"]
MONTHS_PER_YEAR = 12
class _Timer:
def __init__(self) -> None:
self._t0 = time.time()
def log(self, tag: str, msg: str) -> None:
elapsed = time.time() - self._t0
m, s = divmod(int(elapsed), 60)
h, m = divmod(m, 60)
logger.info(f"[{h:02d}:{m:02d}:{s:02d}] [{tag}] {msg}")
class ExplainabilityEngine:
"""
Phase 8: SHAP-based explainability + ElasticNet coefficient analysis.
Loads Phase 6 pkl models + preprocessors, computes SHAP values on the
out-of-sample test set, and extracts EN coefficients from the model objects.
"""
def __init__(self, backend_root: str | Path) -> None:
self.root = Path(backend_root)
self.timer = _Timer()
self.models_dir = self.root / "models" / "phase6"
self.cache_dir = self.models_dir / "training_cache"
self.output_dir = self.root / "data" / "explainability"
self.output_dir.mkdir(parents=True, exist_ok=True)
# ββ Public API ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def run(self) -> dict:
self.timer.log("START", "Phase 8 Explainability started")
report: dict = {"assets": {}, "cross_asset_summary": {}}
comparison_rows: List[dict] = []
for asset in ASSETS:
self.timer.log(asset.upper(), f"Processing {asset.upper()}")
result = self._process_asset(asset)
report["assets"][asset] = result
comparison_rows.extend(result.pop("_comparison_rows"))
# Cross-asset feature importance comparison
comparison_df = pd.DataFrame(comparison_rows)
comparison_df.to_csv(self.output_dir / "feature_importance_comparison.csv", index=False)
self.timer.log("COMPARE", f"Feature importance comparison saved ({len(comparison_df)} rows)")
# Cross-asset summary: which features matter most on average
report["cross_asset_summary"] = self._cross_asset_summary(comparison_df)
# Write full report
with open(self.output_dir / "explainability_report.json", "w", encoding="utf-8") as f:
json.dump(report, f, indent=2, default=str)
self.timer.log("END", "Phase 8 completed successfully")
return report
# ββ Per-asset pipeline ββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _process_asset(self, asset: str) -> dict:
# Load artifacts
feature_cols = self._load_feature_columns(asset)
test_df = self._load_test_snapshot(asset)
X_test_raw, y_test = self._split_xy(test_df, asset)
# ββ ElasticNet βββββββββββββββββββββββββββββββββββββββββββββββββββββββ
en_model, en_pre = self._load_model(asset, "elastic_net")
X_test_en = en_pre.transform(X_test_raw)
transformed_cols = self._get_transformed_columns(asset)
en_coeffs_df = self._extract_en_coefficients(en_model, transformed_cols, asset)
shap_en_df, shap_en_global = self._compute_shap_linear(
en_model, en_pre, X_test_raw, transformed_cols, asset
)
# ββ XGBoost ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
xgb_model, xgb_pre = self._load_model(asset, "xgboost")
X_test_xgb = xgb_pre.transform(X_test_raw)
shap_xgb_df, shap_xgb_global = self._compute_shap_tree(
xgb_model, X_test_xgb, transformed_cols, asset
)
xgb_gain_df = self._load_xgb_gain_importance(asset)
# ββ Latest-prediction local explanation ββββββββββββββββββββββββββββββ
local_explanation = self._explain_latest_prediction(
asset, X_test_raw, y_test,
en_model, en_pre, shap_en_df,
xgb_model, xgb_pre, shap_xgb_df,
transformed_cols,
)
# ββ Comparison rows for cross-asset table ββββββββββββββββββββββββββββ
comparison_rows = self._build_comparison_rows(
asset, transformed_cols, shap_xgb_global, shap_en_global, en_coeffs_df, xgb_gain_df
)
self.timer.log(asset.upper(), "Done")
return {
"en_active_features": int((en_coeffs_df["coefficient"] != 0).sum()),
"en_total_features": len(en_coeffs_df),
"xgb_top_feature": shap_xgb_global.iloc[0]["feature"] if not shap_xgb_global.empty else None,
"en_top_feature": shap_en_global.iloc[0]["feature"] if not shap_en_global.empty and shap_en_global.iloc[0]["mean_abs_shap"] > 0 else "all_zero",
"latest_explanation": local_explanation,
"_comparison_rows": comparison_rows,
}
# ββ SHAP: XGBoost (TreeExplainer) ββββββββββββββββββββββββββββββββββββββββ
def _compute_shap_tree(
self,
model,
X_transformed: np.ndarray,
feature_names: List[str],
asset: str,
) -> Tuple[pd.DataFrame, pd.DataFrame]:
explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_transformed) # (n_test, n_features)
shap_df = pd.DataFrame(shap_values, columns=feature_names)
shap_df.to_csv(self.output_dir / f"shap_values_xgb_{asset}.csv", index=False)
global_df = pd.DataFrame({
"feature": feature_names,
"mean_abs_shap": np.abs(shap_values).mean(axis=0),
}).sort_values("mean_abs_shap", ascending=False).reset_index(drop=True)
global_df.to_csv(self.output_dir / f"shap_global_xgb_{asset}.csv", index=False)
self.timer.log(asset.upper(), f"XGBoost SHAP computed β top feature: {global_df.iloc[0]['feature']} ({global_df.iloc[0]['mean_abs_shap']:.5f})")
return shap_df, global_df
# ββ SHAP: ElasticNet (LinearExplainer) βββββββββββββββββββββββββββββββββββ
def _compute_shap_linear(
self,
model,
preprocessor,
X_raw: pd.DataFrame,
feature_names: List[str],
asset: str,
) -> Tuple[pd.DataFrame, pd.DataFrame]:
X_transformed = preprocessor.transform(X_raw)
# Background = training data mean (masker for linear models)
explainer = shap.LinearExplainer(model, X_transformed, feature_perturbation="interventional")
shap_values = explainer.shap_values(X_transformed) # (n_test, n_features)
shap_df = pd.DataFrame(shap_values, columns=feature_names)
shap_df.to_csv(self.output_dir / f"shap_values_en_{asset}.csv", index=False)
global_df = pd.DataFrame({
"feature": feature_names,
"mean_abs_shap": np.abs(shap_values).mean(axis=0),
}).sort_values("mean_abs_shap", ascending=False).reset_index(drop=True)
global_df.to_csv(self.output_dir / f"shap_global_en_{asset}.csv", index=False)
top = global_df.iloc[0]
self.timer.log(asset.upper(), f"ElasticNet SHAP computed β top feature: {top['feature']} ({top['mean_abs_shap']:.5f})")
return shap_df, global_df
# ββ ElasticNet coefficients βββββββββββββββββββββββββββββββββββββββββββββββ
def _extract_en_coefficients(
self,
model,
feature_names: List[str],
asset: str,
) -> pd.DataFrame:
"""
Extract actual coefficients from the ElasticNetCV pkl object.
The saved CSV has all-zeros due to heavy regularization β the pkl has the truth.
"""
coef = np.asarray(model.coef_).flatten()
intercept = float(model.intercept_) if hasattr(model, "intercept_") else 0.0
df = pd.DataFrame({
"feature": feature_names,
"coefficient": coef,
"abs_coefficient": np.abs(coef),
"direction": np.where(coef > 0, "positive", np.where(coef < 0, "negative", "zero")),
}).sort_values("abs_coefficient", ascending=False).reset_index(drop=True)
df.attrs["intercept"] = intercept
# Add intercept row at end
intercept_row = pd.DataFrame([{
"feature": "__intercept__",
"coefficient": intercept,
"abs_coefficient": abs(intercept),
"direction": "positive" if intercept >= 0 else "negative",
}])
df = pd.concat([df, intercept_row], ignore_index=True)
df.to_csv(self.output_dir / f"elastic_net_coefficients_{asset}.csv", index=False)
n_active = int((np.abs(coef) > 0).sum())
self.timer.log(asset.upper(), f"EN coefficients extracted β {n_active}/{len(coef)} active, intercept={intercept:.6f}")
return df
# ββ Latest-prediction local explanation ββββββββββββββββββββββββββββββββββ
def _explain_latest_prediction(
self,
asset: str,
X_test_raw: pd.DataFrame,
y_test: pd.Series,
en_model, en_pre,
shap_en_df: pd.DataFrame,
xgb_model, xgb_pre,
shap_xgb_df: pd.DataFrame,
feature_names: List[str],
) -> dict:
"""
Explain the most recent (last row of test set) prediction for each model.
Returns top-5 drivers from each model.
"""
last_idx = -1 # most recent test observation
# Raw feature values
raw_values = X_test_raw.iloc[last_idx].to_dict()
# EN prediction + SHAP
X_last_en = en_pre.transform(X_test_raw.iloc[[last_idx]])
en_pred = float(en_model.predict(X_last_en)[0])
en_shap_last = shap_en_df.iloc[last_idx].to_dict()
en_top5 = sorted(en_shap_last.items(), key=lambda x: abs(x[1]), reverse=True)[:5]
# XGB prediction + SHAP
X_last_xgb = xgb_pre.transform(X_test_raw.iloc[[last_idx]])
xgb_pred = float(xgb_model.predict(X_last_xgb)[0])
xgb_shap_last = shap_xgb_df.iloc[last_idx].to_dict()
xgb_top5 = sorted(xgb_shap_last.items(), key=lambda x: abs(x[1]), reverse=True)[:5]
actual = float(y_test.iloc[last_idx])
explanation = {
"asset": asset,
"actual_return": round(actual, 6),
"elastic_net": {
"prediction": round(en_pred, 6),
"error": round(en_pred - actual, 6),
"top_drivers": [
{"feature": f, "shap_value": round(v, 6), "direction": "bullish" if v > 0 else "bearish"}
for f, v in en_top5
],
},
"xgboost": {
"prediction": round(xgb_pred, 6),
"error": round(xgb_pred - actual, 6),
"top_drivers": [
{"feature": f, "shap_value": round(v, 6), "direction": "bullish" if v > 0 else "bearish"}
for f, v in xgb_top5
],
},
"raw_feature_snapshot": {k: round(float(v), 6) if isinstance(v, (int, float, np.floating)) else str(v)
for k, v in raw_values.items()},
}
return explanation
# ββ Cross-asset summary βββββββββββββββββββββββββββββββββββββββββββββββββββ
def _cross_asset_summary(self, comparison_df: pd.DataFrame) -> dict:
"""Average XGB SHAP importance across all assets β universal drivers."""
if "xgb_mean_abs_shap" not in comparison_df.columns:
return {}
avg = (
comparison_df.groupby("feature")["xgb_mean_abs_shap"]
.mean()
.sort_values(ascending=False)
.reset_index()
)
avg.columns = ["feature", "avg_xgb_shap_across_assets"]
avg.to_csv(self.output_dir / "cross_asset_shap_importance.csv", index=False)
return {
"top5_universal_drivers": avg.head(5).to_dict(orient="records"),
"note": "Average XGBoost SHAP importance across SPX, NDX, Gold",
}
# ββ Comparison table builder ββββββββββββββββββββββββββββββββββββββββββββββ
def _build_comparison_rows(
self,
asset: str,
feature_names: List[str],
shap_xgb_global: pd.DataFrame,
shap_en_global: pd.DataFrame,
en_coeffs_df: pd.DataFrame,
xgb_gain_df: pd.DataFrame,
) -> List[dict]:
# Build lookup dicts
xgb_shap_map = dict(zip(shap_xgb_global["feature"], shap_xgb_global["mean_abs_shap"]))
en_shap_map = dict(zip(shap_en_global["feature"], shap_en_global["mean_abs_shap"]))
en_coeff_map = {}
for _, row in en_coeffs_df.iterrows():
if row["feature"] != "__intercept__":
en_coeff_map[row["feature"]] = row["coefficient"]
xgb_gain_map = {}
if xgb_gain_df is not None:
xgb_gain_map = dict(zip(xgb_gain_df["feature"], xgb_gain_df["importance"]))
rows = []
for feat in feature_names:
rows.append({
"asset": asset,
"feature": feat,
"xgb_gain": round(xgb_gain_map.get(feat, 0.0), 6),
"xgb_mean_abs_shap": round(xgb_shap_map.get(feat, 0.0), 6),
"en_mean_abs_shap": round(en_shap_map.get(feat, 0.0), 6),
"en_coefficient": round(en_coeff_map.get(feat, 0.0), 6),
"en_direction": ("positive" if en_coeff_map.get(feat, 0.0) > 0
else "negative" if en_coeff_map.get(feat, 0.0) < 0
else "zero"),
})
return rows
# ββ Helpers βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _load_model(self, asset: str, model_type: str):
if model_type == "elastic_net":
model_path = self.models_dir / f"elastic_net_{asset}.pkl"
pre_path = self.models_dir / f"elastic_net_{asset}_preprocessor.pkl"
else:
model_path = self.models_dir / f"xgb_{asset}.pkl"
pre_path = self.models_dir / f"xgb_{asset}_preprocessor.pkl"
model = joblib.load(model_path)
pre = joblib.load(pre_path)
return model, pre
def _load_test_snapshot(self, asset: str) -> pd.DataFrame:
path = self.cache_dir / f"test_snapshot_{asset}.csv"
df = pd.read_csv(path)
df["date"] = pd.to_datetime(df["date"], format="mixed")
return df.sort_values("date").reset_index(drop=True)
def _load_feature_columns(self, asset: str) -> List[str]:
path = self.models_dir / f"elastic_net_{asset}_feature_columns.json"
with open(path, encoding="utf-8") as f:
data = json.load(f)
return data["raw_feature_columns"]
def _get_transformed_columns(self, asset: str) -> List[str]:
path = self.models_dir / f"elastic_net_{asset}_feature_columns.json"
with open(path, encoding="utf-8") as f:
data = json.load(f)
return data["transformed_feature_columns"]
def _split_xy(self, df: pd.DataFrame, asset: str) -> Tuple[pd.DataFrame, pd.Series]:
target_col = f"{asset}_target"
raw_cols = self._load_feature_columns(asset)
# Drop non-feature columns
drop_cols = [c for c in df.columns if c not in raw_cols and c != target_col]
X = df[raw_cols].copy()
y = df[target_col].copy()
return X, y
def _load_xgb_gain_importance(self, asset: str) -> Optional[pd.DataFrame]:
path = self.models_dir / "feature_importance" / f"xgb_{asset}_importance.csv"
if path.exists():
return pd.read_csv(path)
return None
|