From e1a42a901152fa7ad300aed15303f151b7c67dd1 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=D0=9A=D1=83=D1=80=D0=BD=D0=B0=D1=82=20=D0=90=D0=BD=D0=B4?=
 =?UTF-8?q?=D1=80=D0=B5=D0=B9?= <sevenhill@live.ru>
Date: Tue, 14 Jul 2026 23:36:04 +0300
Subject: [PATCH] fix: align pooled symbol features at training

---
 tests/test_trade_objective_training.py    | 36 +++++++++++++++++++++++
 tools/train_torch_recurrent_forecaster.py |  4 +++
 2 files changed, 40 insertions(+)

diff --git a/tests/test_trade_objective_training.py b/tests/test_trade_objective_training.py
index 44206a6..1e674ed 100644
--- a/tests/test_trade_objective_training.py
+++ b/tests/test_trade_objective_training.py
@@ -12,6 +12,7 @@ from tools.train_torch_recurrent_forecaster import (
     RecurrentReturnModel,
     _barrier_outcome,
     _export_head_state,
+    _prepare_data,
 )
 
 
@@ -84,3 +85,38 @@ def test_multitask_head_export_matches_runtime_inference() -> None:
     actual = _torch_head_outputs(context[0].tolist(), entry, hidden_size=4)
 
     assert actual == pytest.approx(expected, abs=2e-6)
+
+
+def test_pooled_training_populates_symbol_identity_feature() -> None:
+    candles = [
+        _candle(
+            index,
+            open_=100.0 + index * 0.01,
+            high=100.2 + index * 0.01,
+            low=99.8 + index * 0.01,
+            close=100.0 + index * 0.01,
+        )
+        for index in range(180)
+    ]
+
+    prepared = _prepare_data(
+        symbol="BTCUSDT",
+        candles=candles,
+        feature_names=["return_1", "symbol_is_BTCUSDT", "symbol_is_ETHUSDT"],
+        lookback=8,
+        target_horizons=[3],
+        decision_horizon=3,
+        round_trip_cost=0.002,
+        stop_loss_percent=0.04,
+        take_profit_percent=0.035,
+        market_candles={"BTCUSDT": candles},
+        trend_candles=candles,
+        validation_window=24,
+        holdout_window=32,
+        clip=8.0,
+        device=torch.device("cpu"),
+    )
+
+    assert prepared is not None
+    assert torch.all(prepared.train_x[:, :, 1] == 1.0)
+    assert torch.all(prepared.train_x[:, :, 2] == 0.0)
diff --git a/tools/train_torch_recurrent_forecaster.py b/tools/train_torch_recurrent_forecaster.py
index 0c21e8b..2000bfa 100644
--- a/tools/train_torch_recurrent_forecaster.py
+++ b/tools/train_torch_recurrent_forecaster.py
@@ -361,6 +361,7 @@ def _train_pooled_symbols(
         prepared_by_symbol: dict[str, PreparedData] = {}
         for symbol in symbols:
             prepared = _prepare_data(
+                symbol=symbol,
                 candles=market_candles[symbol],
                 feature_names=feature_names,
                 lookback=lookback,
@@ -654,6 +655,7 @@ def _train_symbol(
     for lookback in lookbacks:
         _progress(f"{symbol}: preparing lookback={lookback}")
         prepared = _prepare_data(
+            symbol=symbol,
             candles=candles,
             feature_names=feature_names,
             lookback=lookback,
@@ -777,6 +779,7 @@ def _train_symbol(
 
 def _prepare_data(
     *,
+    symbol: str,
     candles: list[Candle],
     feature_names: list[str],
     lookback: int,
@@ -796,6 +799,7 @@ def _prepare_data(
     feature_rows = _feature_matrix(
         candles,
         feature_names,
+        symbol=symbol,
         market_candles=market_candles,
         trend_candles=trend_candles,
     )