feat(quant): configurable exit params TP/SL/trail + OOS trail experiments

aios_core/quant/ml_predictor.py

@@ -28,7 +28,14 @@
 
 LOG_TAG = "[QuantMLPredictor]"
 
-DEFAULT_FEATURES = ["open", "high", "low", "close", "volume", "ret1", "ema12", "ema26", "rsi", "vol_ma"]
+# Scale-free features (v2 model, trained by scripts/quant_ml_eval_train.py).
+# Identical scale across assets (BTC vs PEPE), which the old raw-price feature
+# set lacked; order must match the training script exactly.
+DEFAULT_FEATURES = [
+    "ret1", "ret3", "ret6", "ret12", "ret24",
+    "rsi", "bb_pos", "macd_norm", "ema_gap",
+    "vol_ratio", "vol_z", "bar_range_pct", "hl_pos",
+]
 
 
 class QuantMLPredictor:
@@ -42,8 +49,9 @@ def __init__(self, models_dir: Optional[Path] = None):
 
     # ------------------------------------------------------------ loading ----
     def _load_model(self) -> None:
-        """Загрузить CatBoost-модель если она есть."""
+        """Загрузить CatBoost-модель; v2 (scale-free features) приоритетнее."""
         candidates = [
+            self.models_dir / "catboost_price_dir_v2.cbm",
             self.models_dir / "catboost_price_dir.cbm",
             self.models_dir / "catboost_price_dir.pkl",
         ]
@@ -69,24 +77,44 @@ def available(self) -> bool:
 
     # -------------------------------------------------------- prediction ----
     def _features_from_csv(self, csv_path: Path) -> Optional[list]:
-        """Построить вектор признаков из последних строк CSV."""
+        """Построить вектор признаков из последних строк CSV.
+
+        Формулы должны совпадать 1:1 с scripts/quant_ml_eval_train.py
+        (_compute_features) — модель v2 обучена на этих же признаках.
+        """
         try:
             df = pd.read_csv(csv_path)
         except Exception as e:
             print(f"{LOG_TAG} [WARN] CSV {csv_path}: {e}")
             return None
-        if len(df) < 26:
+        if len(df) < 40:
             return None
         df = df.sort_values("timestamp_ms")
         df["close"] = pd.to_numeric(df["close"], errors="coerce")
         df["volume"] = pd.to_numeric(df["volume"], errors="coerce")
         g = df[["open", "high", "low", "close", "volume"]].copy()
         g["ret1"] = g["close"].pct_change()
+        g["ret3"] = g["close"].pct_change(3)
+        g["ret6"] = g["close"].pct_change(6)
+        g["ret12"] = g["close"].pct_change(12)
+        g["ret24"] = g["close"].pct_change(24)
+        chg = g["close"].pct_change()
+        up = chg.clip(lower=0).rolling(14).mean()
+        down = (-chg.clip(upper=0)).rolling(14).mean()
+        g["rsi"] = 100.0 - 100.0 / (1.0 + up / down.replace(0, 1e-9))
+        bb_mid = g["close"].rolling(20).mean()
+        bb_std = g["close"].rolling(20).std()
+        g["bb_pos"] = ((g["close"] - bb_mid + 2 * bb_std) / (4 * bb_std).replace(0, 1e-9)).clip(0, 1)
         g["ema12"] = g["close"].ewm(span=12).mean()
         g["ema26"] = g["close"].ewm(span=26).mean()
-        g["rsi"] = 100 - 100 / (1 + g["close"].pct_change().rolling(14).mean() /
-                                g["close"].pct_change().rolling(14).std().replace(0, 1e-9))
-        g["vol_ma"] = g["volume"].rolling(20).mean()
+        g["macd_norm"] = (g["ema12"] - g["ema26"]) / g["close"]
+        g["ema_gap"] = (g["ema12"] - g["ema26"]) / g["close"]
+        vol_mean = g["volume"].rolling(20).mean()
+        vol_std = g["volume"].rolling(20).std()
+        g["vol_ratio"] = g["volume"] / vol_mean.replace(0, 1e-9)
+        g["vol_z"] = (g["volume"] - vol_mean) / vol_std.replace(0, 1e-9)
+        g["bar_range_pct"] = (g["high"] - g["low"]) / g["close"]
+        g["hl_pos"] = (g["close"] - g["low"]) / (g["high"] - g["low"]).replace(0, 1e-9)
         last = g.dropna().iloc[-1]
         try:
             return [float(last[c]) for c in DEFAULT_FEATURES]
@@ -130,11 +158,17 @@ def predict_all(self, symbols: Optional[list[str]] = None) -> list[dict]:
         """Прогноз по всем активам, для которых есть данные."""
         if not self.available:
             return [{"ok": False, "error": "Модель не обучена. Запустите Colab-ноутбук Quant ML Training."}]
+        # Мёртвые/переименованные тикеры (MATIC->POL, RNDR->RENDER): старые
+        # папки данных остаются, но в сигналы не включаются.
+        dead = {"MATIC", "RNDR"}
         if symbols is None:
             symbols = sorted(d.name for d in QUANT_DIR.iterdir()
-                             if d.is_dir() and not d.name.startswith("_") and d.name not in ("export", "models", "uniswap_v3"))
+                             if d.is_dir() and not d.name.startswith("_")
+                             and d.name not in ("export", "models", "uniswap_v3", *dead))
         out = []
         for s in symbols:
+            if s in dead:
+                continue
             r = self.predict_symbol(s)
             if r.get("ok"):
                 out.append(r)

aios_core/quant/rl_signal_bridge.py

@@ -20,11 +20,23 @@
 from typing import Optional
 
 REPO_ROOT = Path(__file__).resolve().parents[2]
-MODEL_FILE = REPO_ROOT / "data" / "quant" / "models" / "ppo_v8.pt"  # лучшая LSTM-PPO (300 эп, +75.23%)
+MODEL_FILE = REPO_ROOT / "data" / "quant" / "models" / "ppo_v9.pt"  # LSTM-PPO v9 (300 эп, sum_rl +96% vs BH −114%)
 OUT_FILE = REPO_ROOT / "data" / "quant" / "rl_signals.json"
 
 LOG_TAG = "[RLSignalBridge]"
 
+# 32 актива среды обучения (quant_train_ppo.py, MultiAssetEnv):
+# self.names = sorted(assets.keys()) — onehot-индекс актива в этом порядке.
+# Используется как fallback; предпочтительно читать assets из чекпоинта
+# модели (ppo_v9.pt сохраняет {"policy", "assets"}), т.к. v8 обучалась с
+# MATIC, а v9 — с POL (разные индексы в onehot).
+ASSET_ORDER_DEFAULT = sorted([
+    "BTC", "ETH", "BNB", "SOL", "XRP", "ADA", "DOGE", "AVAX", "LINK", "DOT",
+    "POL", "LTC", "TRX", "ATOM", "UNI", "ETC", "FIL", "APT", "NEAR", "ARB",
+    "OP", "SUI", "TIA", "SEI", "TON", "INJ", "KAS", "FET", "WIF", "BONK",
+    "PEPE", "SHIB",
+])
+
 
 class RLSignalBridge:
     """Консультирующий доступ к RL-сигналам (read-only)."""
@@ -33,6 +45,7 @@ def __init__(self, model_file: Optional[Path] = None):
         self.model_file = Path(model_file or MODEL_FILE)
         self._policy = None
         self._model_available = self.model_file.exists()
+        self.asset_names = ASSET_ORDER_DEFAULT
 
     # ---- модель ----
     def _load_policy(self):
@@ -83,6 +96,7 @@ def forward(self, x):
 
             ckpt = torch.load(self.model_file, map_location="cpu")
             sd = ckpt.get("policy", ckpt)
+            self.asset_names = sorted(ckpt.get("assets") or ASSET_ORDER_DEFAULT)
             # определяем архитектуру по наличию lstm-слоя
             if "lstm.weight_ih_l0" in sd:
                 w = sd["fc_pre.weight"]
@@ -124,7 +138,7 @@ def _fetch_binance(self, sym: str, interval: str = "1h", limit: int = 120) -> li
             d = json.loads(r.read())
         return [[int(x[0]), float(x[1]), float(x[2]), float(x[3]), float(x[4]), float(x[5])] for x in d]
 
-    def _make_obs(self, rows: list, window: int = 10) -> Optional[list]:
+    def _make_obs(self, rows: list, asset_name: Optional[str] = None, window: int = 10) -> Optional[list]:
         import numpy as np
         if len(rows) < window + 1:
             return None
@@ -140,10 +154,14 @@ def mom(period):
                 m[i] = closes[i] / closes[i - period] - 1.0
             return m
         mom5 = mom(5); mom12 = mom(12)
-        vol_chg = [0.0] * n
-        for i in range(1, n):
-            vol_chg[i] = vols[i] / vols[i - 1] - 1.0 if vols[i-1] else 0.0
-        # волатильность
+        # vol_ratio: volume / rolling(10)-mean — признак #3 среды обучения v8.
+        # (Раньше здесь был vol_chg — несоответствие обучению: модель видела
+        #  другой 3-й статический признак.)
+        vol_ratio = [0.0] * n
+        for i in range(n):
+            lo = max(0, i - 9)
+            vol_ratio[i] = vols[i] / ((sum(vols[lo:i + 1]) / (i - lo + 1)) + 1e-9)
+        # vol_norm: rolling(10) std возвратов / среднее по ряду
         vol_arr = [0.01]*n
         for i in range(1, n):
             w = returns[max(0,i-10):i]
@@ -153,37 +171,45 @@ def mom(period):
         last = n - 1
         base = np.concatenate([
             rets_w,
-            [mom5[last], mom12[last], vol_chg[last], vol_arr[last]/vmean]
+            [mom5[last], mom12[last], vol_ratio[last], vol_arr[last]/vmean]
         ]).astype(np.float32)
         exp = self._obs_dim
         if base.shape[0] == exp:
             return base
         # мультиактив-модель: base (window+4) + onehot (n_assets)
         if exp > base.shape[0]:
             n_assets = exp - base.shape[0]
+            asset_names = getattr(self, "asset_names", ASSET_ORDER_DEFAULT)
+            if asset_name is None or asset_name not in asset_names:
+                # Актив не из обучающего универсума — честно сообщаем «нет сигнала»
+                # вместо того, чтобы подставлять чужой onehot-индекс.
+                return None
             onehot = np.zeros(n_assets, dtype=np.float32)
-            # предполагаем BTC как индекс 0 (первый актив)
-            onehot[0] = 1.0
+            onehot[asset_names.index(asset_name)] = 1.0
             obs = np.concatenate([base, onehot]).astype(np.float32)
             if obs.shape[0] == exp:
                 return obs
         return None
 
     # ---- предсказание ----
-    def predict_symbol(self, binance_symbol: str) -> Optional[dict]:
+    def predict_symbol(self, binance_symbol: str, asset_name: Optional[str] = None) -> Optional[dict]:
         policy = self._load_policy()
         if policy is None:
             return None
         try:
             rows = self._fetch_binance(binance_symbol)
-            obs = self._make_obs(rows)
+            obs = self._make_obs(rows, asset_name)
             if obs is None:
                 return None
             import torch
             with torch.no_grad():
                 o_t = torch.tensor(obs, dtype=torch.float32).unsqueeze(0)
                 mean, _, _ = policy(o_t)
                 act = mean[0][0].item()
+                # В обучении (kg_v8) действие clamp(-1,1) до конвертации в
+                # дискрету {0,1,2}; без clamp mean может выйти за [-1,1] и дать
+                # недопустимую позицию (например, -0.5).
+                act = max(-1.0, min(1.0, act))
                 pos = int((act + 1) / 2 * 2) / 2.0  # 0, 0.5, 1
             return {
                 "symbol": binance_symbol,
@@ -204,12 +230,12 @@ def run_all(self, symbols: Optional[dict] = None) -> dict:
             symbols = {
                 "BTC": "BTCUSDT", "ETH": "ETHUSDT", "SOL": "SOLUSDT", "BNB": "BNBUSDT",
                 "XRP": "XRPUSDT", "ADA": "ADAUSDT", "DOGE": "DOGEUSDT", "LINK": "LINKUSDT",
-                "DOT": "DOTUSDT", "MATIC": "MATICUSDT",
+                "DOT": "DOTUSDT", "POL": "POLUSDT",
             }
         signals = []
         for name, sym in symbols.items():
             try:
-                res = self.predict_symbol(sym)
+                res = self.predict_symbol(sym, asset_name=name)
                 if res:
                     res["asset"] = name
                     signals.append(res)

aios_core/quant_directional_policy.py

@@ -33,6 +33,9 @@ class DirectionalV2Config:
     half_spread_rate: float = 0.0005
     slippage_rate: float = 0.0005
     candle_seconds: int = 3_600
+    take_profit_pct: float = 0.02
+    stop_loss_pct: float = -0.01
+    trail_ratio: float = 0.988
 
     @classmethod
     def from_env(cls) -> DirectionalV2Config:
@@ -57,6 +60,9 @@ def from_env(cls) -> DirectionalV2Config:
             half_spread_rate=max(0.0, float(os.environ.get("AIOS_QUANT_HALF_SPREAD_RATE", "0.0005"))),
             slippage_rate=max(0.0, float(os.environ.get("AIOS_QUANT_SLIPPAGE_RATE", "0.0005"))),
             candle_seconds=max(60, int(os.environ.get("AIOS_QUANT_CANDLE_SECONDS", "3600"))),
+            take_profit_pct=max(0.0, float(os.environ.get("AIOS_QUANT_TAKE_PROFIT_PCT", "0.02"))),
+            stop_loss_pct=min(0.0, float(os.environ.get("AIOS_QUANT_STOP_LOSS_PCT", "-0.01"))),
+            trail_ratio=min(1.0, max(0.0, float(os.environ.get("AIOS_QUANT_TRAIL_RATIO", "0.988")))),
         )
 
     def entry_execution_price(self, mid_price: float) -> float:

aios_core/quant_directional_v2.py

@@ -189,11 +189,11 @@ def run_multi_exchange_cycle(engine) -> dict[str, Any]:
                     held_seconds = max(0.0, now - float(position.get("opened_at", now)))
 
                     reason = ""
-                    if net_pnl_pct >= 2.0:
+                    if net_pnl_pct >= config.take_profit_pct * 100.0:
                         reason = "take_profit"
-                    elif net_pnl_pct <= -1.0:
+                    elif net_pnl_pct <= config.stop_loss_pct * 100.0:
                         reason = "stop_loss"
-                    elif max_seen > entry_mid * 1.01 and mid_price <= max_seen * 0.988:
+                    elif max_seen > entry_mid * 1.01 and mid_price <= max_seen * config.trail_ratio:
                         reason = "trailing_stop"
                     elif bearish_exit_confirmed(config, analysis, held_seconds=held_seconds):
                         reason = "confirmed_bearish_exit"