Update backtest_engine.py

backtest_engine.py

@@ -104,8 +104,8 @@ def _transform_window_for_pattern(df_window):
 
 def calculate_sniper_features_exact(df):
     """
-    Sniper Features Calculation - TRUE MATCH with 'sniper_engine.py'.
-    Includes Liquidity Proxies & Microstructure Features.
+    Sniper Features Calculation - With AUTO-NORMALIZATION Patch.
+    Converts price-dependent features to percentages to fix Scale Mismatch.
     """
     # 1. Standard Features
     d = df.copy()
@@ -117,7 +117,7 @@ def calculate_sniper_features_exact(df):
     d['return_5m'] = c.pct_change(5).fillna(0)
     d['return_15m'] = c.pct_change(15).fillna(0)
 
-    # Technicals (Lower Case matches model expectations)
+    # Technicals
     d['rsi_14'] = ta.rsi(c, length=14).fillna(50)
     
     ema_9 = ta.ema(c, length=9).fillna(c)
@@ -126,10 +126,12 @@ def calculate_sniper_features_exact(df):
     d['ema_9_slope'] = ((ema_9 - ema_9.shift(1)) / ema_9.shift(1).replace(0, np.nan)).fillna(0)
     d['ema_21_dist'] = ((c - ema_21) / ema_21.replace(0, np.nan)).fillna(0)
     
-    # ATR (Length 100 as per sniper_engine.py usually)
-    d['atr'] = ta.atr(h, l, c, length=100).fillna(0)
+    # --- GEM-ARCHITECT PATCH: Normalize Price-Dependent Features ---
+    # ATR (Normalized by Close)
+    atr_raw = ta.atr(h, l, c, length=100).fillna(0)
+    d['atr'] = (atr_raw / c).fillna(0)  # Now it's a percentage (e.g., 0.002 instead of 0.25)
     
-    # Volume Z-Score
+    # Volume Z-Score (Already normalized by z-score logic)
     def _z_roll(x, w=50):
         r = x.rolling(w).mean()
         s = x.rolling(w).std().replace(0, np.nan)
@@ -137,43 +139,46 @@ def calculate_sniper_features_exact(df):
     
     d['vol_zscore_50'] = _z_roll(v, 50)
     
-    # Candle Geometry
-    rng = (h - l).replace(0, 1e-9)
-    d['candle_range'] = h - l
-    d['close_pos_in_range'] = ((c - l) / rng).fillna(0.5)
+    # Candle Geometry (Normalized)
+    rng = (h - l)
+    d['candle_range'] = (rng / c).fillna(0) # Normalized
+    d['close_pos_in_range'] = ((c - l) / rng.replace(0, 1e-9)).fillna(0.5)
 
-    # 2. Liquidity Proxies (The Missing 16 Features!)
-    # Amihud Illiquidity
+    # 2. Liquidity Proxies
+    # Amihud (Ratio of %Ret to $Vol -> inherently somewhat scale-invariant but sensitive)
     d['dollar_vol'] = c * v
     d['amihud'] = (d['return_1m'].abs() / d['dollar_vol'].replace(0, np.nan)).fillna(0)
     
-    # Roll Spread (Volatility Proxy)
+    # Roll Spread (Normalized)
     dp = c.diff()
     roll_cov = dp.rolling(64).cov(dp.shift(1)).fillna(0)
-    d['roll_spread'] = (2 * np.sqrt(np.maximum(0, -roll_cov)))
+    roll_spread_raw = (2 * np.sqrt(np.maximum(0, -roll_cov)))
+    d['roll_spread'] = (roll_spread_raw / c).fillna(0) # Normalized
     
-    # OFI (Order Flow Imbalance Proxy)
+    # OFI (Volume based -> Normalized by Z-Score later or used raw if model expects raw volume flow)
+    # Usually OFI is used as a signal direction, kept raw or z-scored. 
+    # Let's keep raw here as it's volume-based, not price-based.
     sign = np.sign(c.diff()).fillna(0)
     d['signed_vol'] = sign * v
     d['ofi'] = d['signed_vol'].rolling(30).sum().fillna(0)
     
-    # VPIN Proxy
+    # VPIN (Ratio -> Unitless -> Safe)
     buy_vol = (sign > 0) * v
     sell_vol = (sign < 0) * v
     imb = (buy_vol.rolling(60).sum() - sell_vol.rolling(60).sum()).abs()
     tot = v.rolling(60).sum().replace(0, np.nan)
     d['vpin'] = (imb / tot).fillna(0)
     
-    # Garman-Klass Volatility
+    # Garman-Klass (Log returns -> Unitless -> Safe)
     d['rv_gk'] = ((np.log(h / l)**2) / 2) - ((2 * np.log(2) - 1) * (np.log(c / o)**2))
     d['rv_gk'] = d['rv_gk'].fillna(0)
     
-    # VWAP Deviation
+    # VWAP Deviation (Normalized)
     vwap_win = 20
     vwap = (d['dollar_vol'].rolling(vwap_win).sum() / v.rolling(vwap_win).sum().replace(0, np.nan)).fillna(c)
-    d['vwap_dev'] = (c - vwap).fillna(0)
+    d['vwap_dev'] = ((c - vwap) / c).fillna(0) # Normalized
     
-    # Liquidity Score (Composite)
+    # Liquidity Score (Composite - Z-scores handle scaling automatically)
     d['L_score'] = (
         _z_roll(v, 500) + 
         _z_roll(1 / (d['amihud'] + 1e-12), 500) +