main

ai-engine/models/calibration.py

@@ -91,22 +91,26 @@ class Calibrator:
     def __init__(self):
         self.calibrators: Dict[str, IsotonicRegression] = {}
         self.metrics: Dict[str, CalibrationMetrics] = {}
+        # Less aggressive shrinkage — only meaningful overconfident bands are pulled.
+        # Default raised from ~0.85-0.90 to 0.95+ since the orchestrator and config
+        # already apply market-level multipliers; double-shrinkage was the root cause
+        # of 24-35pt avg calibrated-vs-raw drops in production traces.
         self.heuristic_fallback: Dict[str, float] = {
-            "ms": 0.90,
-            "ms_home": 0.90,
-            "ms_home_heavy_fav": 0.95,
-            "ms_home_fav": 0.90,
-            "ms_home_balanced": 0.85,
-            "ms_home_underdog": 0.80,
-            "ms_draw": 0.90,
-            "ms_away": 0.90,
-            "ou15": 0.90,
-            "ou25": 0.90,
-            "ou35": 0.90,
-            "btts": 0.90,
-            "ht_ft": 0.85,
-            "dc": 0.93,
-            "ht": 0.85,
+            "ms": 0.96,
+            "ms_home": 0.96,
+            "ms_home_heavy_fav": 0.98,
+            "ms_home_fav": 0.96,
+            "ms_home_balanced": 0.94,
+            "ms_home_underdog": 0.92,
+            "ms_draw": 0.94,
+            "ms_away": 0.96,
+            "ou15": 0.96,
+            "ou25": 0.96,
+            "ou35": 0.94,
+            "btts": 0.96,
+            "ht_ft": 0.92,
+            "dc": 0.97,
+            "ht": 0.92,
         }
         self._load_calibrators()
         
@@ -139,21 +143,32 @@ class Calibrator:
                 except Exception as e:
                     print(f"[Calibrator] Warning: Failed to load metrics for {market}: {e}")
 
+    # Below this sample count, blend isotonic with raw_prob to dampen overfit jumps.
+    # Above this count, trust isotonic fully.
+    TRUSTED_SAMPLE_FLOOR = 30
+    TRUSTED_SAMPLE_CEILING = 200
+    # Hard cap on how far calibration can move probability in either direction.
+    MAX_DELTA = 0.20
+
     def calibrate(self, market_type: str, raw_prob: float, odds_val: Optional[float] = None) -> float:
         """
-        Calibrate a raw probability using Isotonic Regression.
-        
+        Calibrate a raw probability using Isotonic Regression with safeguards.
+
         Args:
             market_type (str): 'ms_home', 'ou25', 'btts', 'ht_ft', etc.
             raw_prob (float): The raw probability from XGBoost (0.0 - 1.0)
             odds_val (float, optional): The pre-match odds, used for context-aware bucket mapping
-            
+
         Returns:
             float: Calibrated probability (0.0 - 1.0)
+
+        Safeguards:
+          * Low-sample trained models are blended with raw_prob to dampen overfit.
+          * MAX_DELTA caps the per-call adjustment (prevents 40pp swings).
         """
         # Normalize market type
         market_key = market_type.lower().replace("-", "_")
-        
+
         # Route to bucket if ms_home and odds provided
         if market_key == "ms_home" and odds_val is not None and odds_val > 1.0:
             if odds_val <= 1.40:
@@ -164,20 +179,42 @@ class Calibrator:
                 bucket_key = "ms_home_balanced"
             else:
                 bucket_key = "ms_home_underdog"
-                
+
             if bucket_key in self.calibrators:
                 market_key = bucket_key
-        
-        # If we have a trained Isotonic Regression model, use it
+
+        # If we have a trained Isotonic Regression model, use it (with safeguards)
         if market_key in self.calibrators:
             try:
-                calibrated = self.calibrators[market_key].predict([raw_prob])[0]
-                # Ensure output is valid probability
-                return float(np.clip(calibrated, 0.01, 0.99))
+                iso_pred = float(self.calibrators[market_key].predict([raw_prob])[0])
+
+                # Sample-count weighted blend with raw probability.
+                # Sparse models barely move probability; mature models dominate.
+                metrics = self.metrics.get(market_key)
+                n_samples = metrics.sample_count if metrics else 0
+                if n_samples >= self.TRUSTED_SAMPLE_CEILING:
+                    iso_weight = 1.0
+                elif n_samples <= self.TRUSTED_SAMPLE_FLOOR:
+                    # Very sparse: at least 30% trust to surface the signal
+                    iso_weight = max(0.30, n_samples / self.TRUSTED_SAMPLE_CEILING)
+                else:
+                    # Linearly ramp 30% → 100% between floor and ceiling
+                    span = self.TRUSTED_SAMPLE_CEILING - self.TRUSTED_SAMPLE_FLOOR
+                    iso_weight = 0.30 + 0.70 * (n_samples - self.TRUSTED_SAMPLE_FLOOR) / span
+                blended = iso_weight * iso_pred + (1.0 - iso_weight) * raw_prob
+
+                # Cap delta to avoid huge swings on noisy calibrators
+                delta = blended - raw_prob
+                if delta > self.MAX_DELTA:
+                    blended = raw_prob + self.MAX_DELTA
+                elif delta < -self.MAX_DELTA:
+                    blended = raw_prob - self.MAX_DELTA
+
+                return float(np.clip(blended, 0.01, 0.99))
             except Exception as e:
                 print(f"[Calibrator] Warning: Isotonic failed for {market_key}: {e}")
                 # Fall through to heuristic
-        
+
         # Fallback to heuristic calibration
         return self._heuristic_calibrate(market_key, raw_prob)