Add state Continue diagnostic controls
This commit is contained in:
Codex
@@ -0,0 +1,151 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import sys
|
||||
import tempfile
|
||||
import unittest
|
||||
from pathlib import Path
|
||||
|
||||
import numpy as np
|
||||
import pandas as pd
|
||||
|
||||
TRAINING_ROOT = Path(__file__).resolve().parents[1]
|
||||
if str(TRAINING_ROOT) not in sys.path:
|
||||
sys.path.insert(0, str(TRAINING_ROOT))
|
||||
|
||||
from trader_training.onnx_export import LinearHead, export_heads
|
||||
from trader_training.schemas import FEATURE_ORDER
|
||||
from trader_training.state_continue_experiment import STATE_FEATURES, _predict_frozen_linear_model, _state_rows_for_age, _train_side_models, _verdict
|
||||
|
||||
|
||||
class StateContinueExperimentTest(unittest.TestCase):
|
||||
def test_state_rows_include_required_position_and_frozen_entry_features(self) -> None:
|
||||
row = {
|
||||
"current_sample_id": "s0",
|
||||
"symbol": "BTC-USDT-PERP",
|
||||
"current_event_time": pd.Timestamp("2026-01-01T00:05:00Z"),
|
||||
"current_open_time_ms": 300_000,
|
||||
"side": "LONG",
|
||||
"split_id": "fit_inner",
|
||||
"walk_forward_fold": 0,
|
||||
"time_in_position_minutes": 5,
|
||||
"entry_price": 100.0,
|
||||
"current_price": 100.1,
|
||||
"high_since_entry": 100.2,
|
||||
"low_since_entry": 99.95,
|
||||
"future_return_bps": 12.0,
|
||||
"mae_bps": 3.0,
|
||||
"entry_predicted_edge_bps": 8.5,
|
||||
"entry_direction_prob": 0.64,
|
||||
"add_count": 0.0,
|
||||
"minutes_since_last_add": 9999.0,
|
||||
}
|
||||
for feature_name in FEATURE_ORDER:
|
||||
row[feature_name] = 0.0
|
||||
frame = pd.DataFrame([row])
|
||||
|
||||
out = _state_rows_for_age(frame, stop_bps=8.0, target_bps=12.0, cost_bps=6.5)
|
||||
|
||||
self.assertEqual(set(STATE_FEATURES), set(STATE_FEATURES).intersection(out.columns))
|
||||
self.assertAlmostEqual(5.5, float(out.iloc[0]["expected_continue_edge_bps"]))
|
||||
self.assertEqual(1, int(out.iloc[0]["continue_target"]))
|
||||
self.assertAlmostEqual(8.5, float(out.iloc[0]["entry_predicted_edge_bps"]))
|
||||
self.assertAlmostEqual(0.64, float(out.iloc[0]["entry_direction_prob"]), places=6)
|
||||
self.assertAlmostEqual(0.0, float(out.iloc[0]["add_count"]))
|
||||
self.assertAlmostEqual(9999.0, float(out.iloc[0]["minutes_since_last_add"]))
|
||||
|
||||
def test_frozen_linear_onnx_weights_are_read_without_row_by_row_runtime(self) -> None:
|
||||
with tempfile.TemporaryDirectory() as tmp:
|
||||
model_path = Path(tmp) / "direction.onnx"
|
||||
export_heads(
|
||||
model_path,
|
||||
[
|
||||
LinearHead(
|
||||
"direction",
|
||||
"softmax",
|
||||
np.zeros((len(FEATURE_ORDER), 3), dtype=np.float32),
|
||||
np.array([0.0, 1.0, 2.0], dtype=np.float32),
|
||||
),
|
||||
LinearHead(
|
||||
"long_expected_net_edge_bps",
|
||||
"identity",
|
||||
np.zeros((len(FEATURE_ORDER), 1), dtype=np.float32),
|
||||
np.array([7.25], dtype=np.float32),
|
||||
),
|
||||
],
|
||||
feature_count=len(FEATURE_ORDER),
|
||||
)
|
||||
frame = pd.DataFrame({"sample_id": ["s0", "s1"]})
|
||||
for feature_name in FEATURE_ORDER:
|
||||
frame[feature_name] = 0.0
|
||||
|
||||
out = _predict_frozen_linear_model(
|
||||
model_path,
|
||||
frame,
|
||||
{
|
||||
"direction": ("softmax", ("long_prob", "short_prob", "neutral_prob")),
|
||||
"long_expected_net_edge_bps": ("identity", ("long_edge",)),
|
||||
},
|
||||
)
|
||||
|
||||
self.assertEqual(["s0", "s1"], out["sample_id"].tolist())
|
||||
self.assertTrue(np.allclose(1.0, out[["long_prob", "short_prob", "neutral_prob"]].sum(axis=1)))
|
||||
self.assertLess(float(out.iloc[0]["long_prob"]), float(out.iloc[0]["neutral_prob"]))
|
||||
self.assertAlmostEqual(7.25, float(out.iloc[0]["long_edge"]), places=6)
|
||||
|
||||
def test_verdict_refuses_state_continue_when_edge_mae_is_not_good_enough(self) -> None:
|
||||
results = {}
|
||||
for side in ("long", "short"):
|
||||
results[f"{side}_market_only"] = {
|
||||
"validation_locked": {"continue_auc": 0.61, "edge_mae_vs_constant_ratio": 0.985},
|
||||
"latest_stress": {"continue_auc": 0.62, "edge_mae_vs_constant_ratio": 0.984},
|
||||
"regressor_converged": True,
|
||||
}
|
||||
results[f"{side}_market_plus_state"] = {
|
||||
"validation_locked": {"continue_auc": 0.63, "edge_mae_vs_constant_ratio": 0.979},
|
||||
"latest_stress": {"continue_auc": 0.64, "edge_mae_vs_constant_ratio": 0.978},
|
||||
"regressor_converged": True,
|
||||
}
|
||||
|
||||
verdict = _verdict(results)
|
||||
|
||||
self.assertEqual("NOT_READY_FOR_FORMAL_CHAIN", verdict["status"])
|
||||
self.assertTrue(any("above 0.97" in reason for reason in verdict["reasons"]))
|
||||
|
||||
def test_train_side_models_supports_ridge_regressor_diagnostic(self) -> None:
|
||||
rows = []
|
||||
for split_id in ("fit_inner", "tune_inner", "validation_locked", "latest_stress"):
|
||||
for index, target in enumerate((0, 1)):
|
||||
row = {
|
||||
"sample_id": f"{split_id}-{index}",
|
||||
"symbol": "BTC-USDT-PERP",
|
||||
"event_time": pd.Timestamp("2026-01-01T00:00:00Z") + pd.Timedelta(minutes=len(rows)),
|
||||
"split_id": split_id,
|
||||
"position_side": "LONG",
|
||||
"continue_target": target,
|
||||
"expected_continue_edge_bps": -3.0 if target == 0 else 6.0,
|
||||
}
|
||||
for feature_name in FEATURE_ORDER:
|
||||
row[feature_name] = float(index)
|
||||
for feature_name in STATE_FEATURES:
|
||||
row[feature_name] = float(index)
|
||||
rows.append(row)
|
||||
frame = pd.DataFrame(rows)
|
||||
|
||||
metrics, predictions = _train_side_models(
|
||||
frame,
|
||||
"LONG",
|
||||
[*FEATURE_ORDER, *STATE_FEATURES],
|
||||
regressor_kind="ridge",
|
||||
ridge_alpha=1.0,
|
||||
regression_target_clip_bps=5.0,
|
||||
)
|
||||
|
||||
self.assertEqual("ridge", metrics["regressor_kind"])
|
||||
self.assertEqual(5.0, metrics["regression_target_clip_bps"])
|
||||
self.assertTrue(metrics["regressor_converged"])
|
||||
self.assertEqual(8, len(predictions))
|
||||
self.assertIn("time_in_position_minutes", predictions.columns)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user