feat: add orderbook shadow training pipeline

tools/calibrate_torch_thresholds.py

@@ -26,6 +26,7 @@ from crypto_spot_bot.bybit import BybitClient
 from crypto_spot_bot.config import load_settings
 from crypto_spot_bot.indicators import add_indicators
 from crypto_spot_bot.models import Candle
+from crypto_spot_bot.orderbook_features import ORDERBOOK_FEATURES, load_orderbook_feature_map
 from crypto_spot_bot.time_series import (
     DEFAULT_TORCH_FEATURES,
     _barrier_outcome,
@@ -97,6 +98,14 @@ def main() -> None:
     context_symbols = sorted(set(symbols + _symbols(args.context_symbols, ())))
     horizon = args.horizon if args.horizon > 0 else settings.time_series_forecast_horizon
     round_trip_cost = _artifact_round_trip_cost(artifact, settings)
+    orderbook_features: dict[str, dict[int, dict[str, float]]] = {}
+    if args.orderbook_db:
+        orderbook_features, _orderbook_manifest = load_orderbook_feature_map(
+            args.orderbook_db,
+            interval=settings.base_interval,
+            symbols=symbols,
+            min_samples_per_bucket=args.orderbook_min_samples_per_bucket,
+        )
 
     market_candles: dict[str, list[Candle]] = {}
     for symbol in context_symbols:
@@ -125,6 +134,7 @@ def main() -> None:
             min_candles=max(30, settings.time_series_min_candles),
             calibration_window=args.calibration_window,
             batch_size=args.batch_size,
+            orderbook_features=orderbook_features,
         )
         records.extend(symbol_records)
         per_symbol_counts[symbol] = len(symbol_records)
@@ -304,6 +314,8 @@ def _parse_args() -> argparse.Namespace:
     parser.add_argument("--min-oos-folds-with-trades", type=int, default=2, help="Minimum walk-forward folds that must produce trades.")
     parser.add_argument("--min-oos-profit-factor", type=float, default=1.10, help="Minimum out-of-sample profit factor.")
     parser.add_argument("--min-benchmark-edge-percent", type=float, default=0.0, help="Required total-net percent advantage over the benchmark.")
+    parser.add_argument("--orderbook-db", default="", help="SQLite cache used by an artifact with L1 features.")
+    parser.add_argument("--orderbook-min-samples-per-bucket", type=int, default=20)
     return parser.parse_args()
 
 
@@ -352,6 +364,7 @@ def _forecast_records(
     min_candles: int,
     calibration_window: int,
     batch_size: int,
+    orderbook_features: dict[str, dict[int, dict[str, float]]] | None = None,
 ) -> list[ForecastRecord]:
     entry = _torch_recurrent_entry(symbol, artifact)
     model = _torch_recurrent_model_name(symbol, artifact)
@@ -364,6 +377,7 @@ def _forecast_records(
         symbol=symbol,
         market_candles=market_candles,
         trend_candles=trend_candles,
+        orderbook_features=orderbook_features,
     )
     closes = [float(candle.close) for candle in candles]
     decision_horizon = _calibration_horizon(entry, horizon, explicit=horizon_is_explicit)
@@ -376,6 +390,18 @@ def _forecast_records(
         start += 1
     if calibration_window > 0:
         start = max(start, end - calibration_window)
+    lookback = max(1, int(float(entry.get("lookback", 64))))
+    requires_orderbook = any(name in ORDERBOOK_FEATURES for name in feature_names)
+    symbol_orderbook = (orderbook_features or {}).get(symbol.upper(), {})
+    valid_indices = {
+        index
+        for index in range(start, max(start, end))
+        if not requires_orderbook
+        or all(
+            candles[position].timestamp in symbol_orderbook
+            for position in range(index - lookback + 1, index + 1)
+        )
+    }
     batched_records = _batch_forecast_records(
         symbol=symbol,
         candles=candles,
@@ -389,6 +415,7 @@ def _forecast_records(
         start=start,
         end=end,
         batch_size=batch_size,
+        valid_indices=valid_indices,
     )
     if batched_records is not None:
         return batched_records
@@ -398,6 +425,8 @@ def _forecast_records(
     # belong exclusively to the final quality gate and cannot influence replay.
     skill = _entry_validation_skill(entry)
     for index in range(start, max(start, end)):
+        if index not in valid_indices:
+            continue
         prediction = _torch_recurrent_predict(
             _log_returns(closes[: index + 1]),
             symbol,
@@ -476,6 +505,7 @@ def _batch_forecast_records(
     start: int,
     end: int,
     batch_size: int,
+    valid_indices: set[int] | None = None,
 ) -> list[ForecastRecord] | None:
     if torch is None or RecurrentReturnModel is None:
         return None
@@ -494,7 +524,9 @@ def _batch_forecast_records(
     indices = [
         index
         for index in range(start, max(start, end))
-        if index - lookback + 1 >= 0 and index + decision_horizon < len(closes)
+        if index - lookback + 1 >= 0
+        and index + decision_horizon < len(closes)
+        and (valid_indices is None or index in valid_indices)
     ]
     if not indices:
         return []