gg

2026-04-21 16:53:56 +03:00
parent 1346924387
commit 2ccd6831eb
26 changed files with 430403 additions and 3 deletions
@@ -0,0 +1,58 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+import pandas as pd
+
+
+AI_ENGINE_DIR = Path(__file__).resolve().parents[1]
+DATA_DIR = AI_ENGINE_DIR / "data" / "v26_shadow"
+CONFIG_PATH = AI_ENGINE_DIR / "models" / "v26_shadow" / "market_profiles.json"
+REPORT_PATH = AI_ENGINE_DIR / "reports" / "training_v26_shadow.json"
+REPORT_PATH.parent.mkdir(parents=True, exist_ok=True)
+
+
+def _market_accuracy(frame: pd.DataFrame, target_col: str) -> float:
+    if target_col not in frame.columns or frame.empty:
+        return 0.0
+    counts = frame[target_col].value_counts(normalize=True)
+    if counts.empty:
+        return 0.0
+    return round(float(counts.max()), 4)
+
+
+def main() -> None:
+    train_csv = DATA_DIR / "train.csv"
+    validation_csv = DATA_DIR / "validation.csv"
+    if not train_csv.exists() or not validation_csv.exists():
+        raise SystemExit("Run extract_training_data_v26.py first")
+
+    train_df = pd.read_csv(train_csv)
+    validation_df = pd.read_csv(validation_csv)
+    config = json.loads(CONFIG_PATH.read_text(encoding="utf-8"))
+    report = {
+        "version": config.get("version"),
+        "calibration_version": config.get("calibration_version"),
+        "train_rows": int(len(train_df)),
+        "validation_rows": int(len(validation_df)),
+        "label_priors": {
+            "MS": _market_accuracy(validation_df, "label_ms"),
+            "OU25": _market_accuracy(validation_df, "label_ou25"),
+            "BTTS": _market_accuracy(validation_df, "label_btts"),
+            "HT": _market_accuracy(validation_df, "label_ht_result"),
+            "HTFT": _market_accuracy(validation_df, "label_ht_ft"),
+            "CARDS": _market_accuracy(validation_df, "label_cards_ou45"),
+        },
+        "artifact_path": str(CONFIG_PATH),
+        "notes": [
+            "v26.shadow runtime currently uses artifact-based calibration and ROI gating",
+            "market profile JSON remains the source of truth for runtime thresholds",
+        ],
+    }
+    REPORT_PATH.write_text(json.dumps(report, indent=2), encoding="utf-8")
+    print(f"[OK] Shadow training report written to {REPORT_PATH}")
+
+
+if __name__ == "__main__":
+    main()