Update ml_engine/sniper_engine.py

ml_engine/sniper_engine.py

@@ -1,10 +1,10 @@
 # ============================================================
 # 🎯 ml_engine/sniper_engine.py 
-# (V2.2 - GEM-Architect: Full Visibility & OKX Patch)
+# (V2.3 - GEM-Architect: Dynamic Window Fix)
 # ============================================================
-# - Fixed: Logs now ALWAYS show ML and OB scores (Pass or Fail).
-# - Fixed: Data requirements relaxed to prevent 'No Data' on minor gaps.
-# - Integrity: OKX/Binance Unpacking Logic Preserved.
+# - Fixed: Rolling window error (min_periods > window).
+# - Logic: Auto-adjusts min_periods to fit requested window size.
+# - Integrity: Full ML & OB functionality restored.
 # ============================================================
 
 import os
@@ -17,18 +17,21 @@ import traceback
 from typing import List, Dict, Any, Optional
 
 N_SPLITS = 5
-# تم تخفيف الشرط هنا للسماح للموديل بالعمل حتى لو البيانات ناقصة قليلاً
-# الموديل سيتعامل مع الـ NaN بشكل طبيعي
-LOOKBACK_WINDOW = 200 
+# السماح ببيانات أقل للمرونة (بدلاً من 500)
+LOOKBACK_WINDOW = 150 
 
 # ============================================================
 # 🔧 1. Feature Engineering (Standard + Liquidity Proxies)
 # ============================================================
 
 def _z_score_rolling(x, w=500): 
-    # Use min_periods to allow calculation even with less data
-    r = x.rolling(w, min_periods=100).mean()
-    s = x.rolling(w, min_periods=100).std().replace(0, np.nan)
+    # ✅ FIX: Ensure min_periods never exceeds window size (w)
+    # If w=50 (like in vol_zscore), min_periods becomes 50.
+    # If w=500, min_periods becomes 100 (allowing partial calculation).
+    effective_min = min(w, 100)
+    
+    r = x.rolling(w, min_periods=effective_min).mean()
+    s = x.rolling(w, min_periods=effective_min).std().replace(0, np.nan)
     z = (x - r) / s
     return z.fillna(0)
 
@@ -45,7 +48,8 @@ def _add_liquidity_proxies(df):
     df_proxy['amihud'] = (df_proxy['ret'].abs() / df_proxy['dollar_vol'].replace(0, np.nan)).fillna(np.inf)
     
     dp = df_proxy['close'].diff()
-    roll_cov = dp.rolling(64, min_periods=30).cov(dp.shift(1))
+    # Reduced min_periods for reliability
+    roll_cov = dp.rolling(64, min_periods=20).cov(dp.shift(1))
     df_proxy['roll_spread'] = (2 * np.sqrt(np.maximum(0, -roll_cov))).bfill()
 
     sign = np.sign(df_proxy['close'].diff()).fillna(0)
@@ -100,6 +104,8 @@ def _add_standard_features(df):
         df_feat['ema_21_dist'] = 0
     
     df_feat['atr'] = ta.atr(df_feat['high'], df_feat['low'], df_feat['close'], length=100)
+    
+    # This was causing the error (w=50 vs min=100). Now fixed in helper.
     df_feat['vol_zscore_50'] = _z_score_rolling(df_feat['volume'], w=50) 
     
     df_feat['candle_range'] = df_feat['high'] - df_feat['low']
@@ -135,7 +141,7 @@ class SniperEngine:
         
         self._wall_cache = {}
 
-        print("🎯 [SniperEngine V2.2] High-Visibility Logic Loaded.")
+        print("🎯 [SniperEngine V2.3] Dynamic Window Logic Loaded.")
 
     def configure_settings(self, 
                            threshold: float, 
@@ -216,7 +222,6 @@ class SniperEngine:
             for i in range(depth):
                 weight = 1.0 / (1.0 + (self.ob_depth_decay * i))
                 
-                # ✅ Safe Indexing
                 bid_vol = float(bids[i][1])
                 ask_vol = float(asks[i][1])
                 
@@ -230,7 +235,6 @@ class SniperEngine:
             max_valid_wall = 0.0
             limit_price = best_ask * (1 + self.max_wall_dist)
             
-            # ✅ Safe Unpacking Loop
             for item in asks[:depth]:
                 p = float(item[0])
                 v = float(item[1])
@@ -287,7 +291,6 @@ class SniperEngine:
         ml_reason = "No Data"
         
         # ✅ Relaxed Check: Allow partial data (min 100) instead of strict 500
-        # LightGBM handles NaNs, better to run with partial data than show 'No Data'
         if len(ohlcv_1m_data) >= 100 and self.models:
             try:
                 df = pd.DataFrame(ohlcv_1m_data, columns=['timestamp', 'open', 'high', 'low', 'close', 'volume'])
@@ -303,19 +306,15 @@ class SniperEngine:
                 print(f"❌ [Sniper] ML Error: {e}")
                 ml_reason = "ML Err"
 
-        # --- Smart Order Book Analysis ---
         ob_res = {'score': 0.5, 'imbalance': 0.5, 'veto': False, 'reason': 'No OB'}
         if order_book_data:
             ob_res = self._score_order_book(order_book_data, symbol=symbol)
 
-        # ✅ FIXED: Prepare Score Strings for Visibility in ALL cases
         ob_str = f"OB:{ob_res['score']:.2f}"
         
-        # --- Final Hybrid Score ---
         if ob_res.get('veto', False):
             final_score = 0.0
             signal = 'WAIT'
-            # ✅ Added ml_reason to veto log
             reason_str = f"⛔ {ob_res['reason']} | {ml_reason} | {ob_str}"
         else:
             final_score = (ml_score * self.weight_ml) + (ob_res['score'] * self.weight_ob)
@@ -325,7 +324,6 @@ class SniperEngine:
                 reason_str = f"✅ GO: {final_score:.2f} | {ml_reason} | {ob_str}"
             else:
                 signal = 'WAIT'
-                # ✅ ADDED: Missing Info Restored here
                 reason_str = f"📉 Low Score: {final_score:.2f} | {ml_reason} | {ob_str}"
 
         return {