feat: add orderbook shadow training pipeline

crypto_spot_bot/bot.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import asyncio
 import logging
+import math
 import sqlite3
 from datetime import datetime
 
@@ -14,7 +15,7 @@ from crypto_spot_bot.models import BotStatus, Signal, Ticker, utc_now
 from crypto_spot_bot.patterns import PatternAnalyzer
 from crypto_spot_bot.strategy import SpotStrategy, torch_model_readiness_reasons
 from crypto_spot_bot.storage import Storage
-from crypto_spot_bot.time_series import TimeSeriesForecaster
+from crypto_spot_bot.time_series import TimeSeriesForecaster, _barrier_outcome
 
 
 logger = logging.getLogger(__name__)
@@ -31,6 +32,7 @@ class CryptoSpotBot:
         pattern_analyzer: PatternAnalyzer,
         learner: TradeLearner,
         forecaster: TimeSeriesForecaster | None = None,
+        shadow_forecaster: TimeSeriesForecaster | None = None,
         llm_advisor=None,
     ):
         self.settings = settings
@@ -41,6 +43,7 @@ class CryptoSpotBot:
         self.pattern_analyzer = pattern_analyzer
         self.learner = learner
         self.forecaster = forecaster
+        self.shadow_forecaster = shadow_forecaster
         self.llm_advisor = llm_advisor
         self.running = False
         self.started_at: datetime | None = None
@@ -52,6 +55,8 @@ class CryptoSpotBot:
         self._last_reconciliation_at: datetime | None = None
         self._last_prune_at: datetime | None = None
         self._consecutive_loop_errors = 0
+        self._orderbook_feature_cache_key: tuple[tuple[str, int], ...] = ()
+        self._orderbook_feature_cache: dict[str, dict[int, dict[str, float]]] = {}
 
     async def start(self) -> None:
         if self.running:
@@ -416,6 +421,25 @@ class CryptoSpotBot:
         self.market.patterns = patterns
 
     def _update_forecasts(self) -> None:
+        cache_key = tuple(
+            (symbol, rows[-1].timestamp if rows else 0)
+            for symbol, rows in sorted(self.market.candles.items())
+        )
+        earliest_timestamp = min(
+            (rows[0].timestamp for rows in self.market.candles.values() if rows),
+            default=0,
+        )
+        if cache_key != self._orderbook_feature_cache_key:
+            orderbook_features, _manifest = self.storage.recent_aggregated_orderbook_features(
+                interval=self.settings.base_interval,
+                symbols=self.market.symbols,
+                after_timestamp_ms=earliest_timestamp,
+                min_samples_per_bucket=20,
+            )
+            self._orderbook_feature_cache = orderbook_features
+            self._orderbook_feature_cache_key = cache_key
+        else:
+            orderbook_features = self._orderbook_feature_cache
         if (
             self.forecaster is None
             or not self.settings.time_series_forecast_enabled
@@ -429,8 +453,98 @@ class CryptoSpotBot:
                 symbol=symbol,
                 market_candles=self.market.candles,
                 trend_candles=self.market.trend_candles.get(symbol, []),
+                orderbook_features=orderbook_features,
             ).as_dict()
         self.market.forecasts = forecasts
+        self._update_shadow_forecasts(orderbook_features)
+
+    def _update_shadow_forecasts(
+        self,
+        orderbook_features: dict[str, dict[int, dict[str, float]]],
+    ) -> None:
+        if self.shadow_forecaster is None:
+            self.market.shadow_forecasts = {}
+            return
+        model_sha256 = self.shadow_forecaster.artifact_sha256()
+        if not model_sha256:
+            self.market.shadow_forecasts = {}
+            return
+        forecasts: dict[str, dict] = {}
+        for symbol in self.market.symbols:
+            candles = self.market.candles.get(symbol, [])
+            forecast = self.shadow_forecaster.forecast(
+                candles,
+                symbol=symbol,
+                market_candles=self.market.candles,
+                trend_candles=self.market.trend_candles.get(symbol, []),
+                orderbook_features=orderbook_features,
+            ).as_dict()
+            forecast["shadow"] = True
+            forecast["model_sha256"] = model_sha256
+            forecasts[symbol] = forecast
+            self._record_and_settle_shadow(symbol, candles, forecast, model_sha256)
+        self.market.shadow_forecasts = forecasts
+
+    def _record_and_settle_shadow(
+        self,
+        symbol: str,
+        candles: list,
+        forecast: dict,
+        model_sha256: str,
+    ) -> None:
+        if candles and forecast.get("usable"):
+            probability = float(
+                forecast.get("probability_take_profit_first")
+                if forecast.get("probability_take_profit_first") is not None
+                else forecast.get("probability_up", 0.5)
+            )
+            expected = float(forecast.get("expected_return_percent", 0.0) or 0.0)
+            eligible = bool(
+                not forecast.get("block_entry")
+                and expected >= float(forecast.get("calibrated_min_edge_percent", 0.0) or 0.0)
+                and probability >= float(forecast.get("calibrated_min_probability_up", 0.5) or 0.5)
+            )
+            self.storage.insert_shadow_prediction(
+                model_sha256=model_sha256,
+                symbol=symbol,
+                forecast_timestamp_ms=candles[-1].timestamp,
+                horizon=max(1, int(forecast.get("horizon", 1) or 1)),
+                reference_price=float(candles[-1].close),
+                expected_return_percent=expected,
+                probability_up=probability,
+                eligible_signal=eligible,
+            )
+        if not candles:
+            return
+        indexes = {candle.timestamp: index for index, candle in enumerate(candles)}
+        round_trip_cost = 2.0 * (
+            float(self.settings.taker_fee_rate) + float(self.settings.slippage_rate)
+        )
+        for row in self.storage.pending_shadow_predictions(
+            model_sha256=model_sha256,
+            symbol=symbol,
+        ):
+            index = indexes.get(int(row.get("forecast_timestamp_ms", 0) or 0))
+            horizon = max(1, int(row.get("horizon", 1) or 1))
+            if index is None or index + horizon >= len(candles):
+                continue
+            outcome = _barrier_outcome(
+                candles,
+                end_index=index,
+                horizon=horizon,
+                stop_loss_percent=float(self.settings.stop_loss_percent),
+                take_profit_percent=float(self.settings.take_profit_percent),
+                round_trip_cost=round_trip_cost,
+            )
+            if outcome is None:
+                continue
+            actual_log_return, take_profit_first = outcome
+            actual_return_percent = (math.exp(actual_log_return) - 1.0) * 100.0
+            self.storage.settle_shadow_prediction(
+                int(row["id"]),
+                actual_return_percent=actual_return_percent,
+                take_profit_first=take_profit_first >= 0.5,
+            )
 
     def status(self) -> BotStatus:
         live_ready = self.settings.live_ready