v2.2: signal model vuelve a BTC/ETH/SOL (probado rentable) + SMC features

model_signals.py

@@ -303,40 +303,31 @@ def main():
 
     if args.multi:
         import glob
-        pattern = os.path.join(DATA_DIR, f"labeled_*_{args.timeframe}.parquet")
-        files = sorted(glob.glob(pattern))
+        # Entrenar SOLO con monedas base (probado rentable a threshold 0.7)
+        # AI coins tienen comportamiento muy diferente - usar Strategy Runner SMC
+        BASE_TRAIN = ["BTCUSDT", "ETHUSDT", "SOLUSDT"]
         dfs = []
-        MIN_ROWS = 1500  # Monedas con menos historia = ruido
-        BASE_SYMBOLS = {"BTCUSDT", "ETHUSDT", "SOLUSDT", "LINKUSDT", "FETUSDT",
-                        "GRTUSDT", "THETAUSDT", "NMRUSDT", "RLCUSDT", "INJUSDT"}  # Baja volatilidad
-        for path in files:
-            df_sym = pd.read_parquet(path)
-            sym = os.path.basename(path).replace("labeled_", "").replace(f"_{args.timeframe}.parquet", "")
-            if len(df_sym) < MIN_ROWS:
-                logger.info("Skipped %s: only %d rows (min %d)", sym, len(df_sym), MIN_ROWS)
-                continue
-            logger.info("Loaded %s: %d rows", sym, len(df_sym))
-            # Threshold adaptativo: 2% para estables, 5% para AI volatiles
-            thresh = args.threshold if sym in BASE_SYMBOLS else args.threshold * 2.5
-            df_sym = create_signal_targets(df_sym, horizon=args.horizon,
-                                           buy_threshold=thresh, sell_threshold=-thresh)
-            dfs.append(df_sym)
+        for sym in BASE_TRAIN:
+            path = os.path.join(DATA_DIR, f"labeled_{sym}_{args.timeframe}.parquet")
+            if os.path.exists(path):
+                df_sym = pd.read_parquet(path)
+                logger.info("Loaded %s: %d rows", sym, len(df_sym))
+                dfs.append(df_sym)
         if not dfs:
-            logger.error("No labeled files found")
+            logger.error("No base symbol data found")
             return
         df = pd.concat(dfs, axis=0).sort_index()
-        logger.info("Combined: %d rows (%d symbols)", len(df), len(dfs))
+        logger.info("Combined: %d rows (%d base symbols)", len(df), len(dfs))
     else:
         path = os.path.join(DATA_DIR, f"labeled_{args.symbol}_{args.timeframe}.parquet")
         if not os.path.exists(path):
             return
         df = pd.read_parquet(path)
 
-    # Crear targets (solo si no se hizo en el loop multi)
-    if not args.multi:
-        df = create_signal_targets(df, horizon=args.horizon,
-                                    buy_threshold=args.threshold,
-                                    sell_threshold=-args.threshold)
+    # Crear targets de senal
+    df = create_signal_targets(df, horizon=args.horizon,
+                                buy_threshold=args.threshold,
+                                sell_threshold=-args.threshold)
     feature_cols = select_signal_features(df)
     result = train_signal_model(df, feature_cols)