admin/quant-trader-service
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Improve Trader V4 training pipeline

Browse Source 
Align entry labels with max future edge, tune direction labeling, and harden regression evaluation.

Add training diagnostics, price-plan search, feature screening, and nonlinear benchmark scripts.
This commit is contained in:
Codex
2026-06-27 19:57:29 +08:00
parent e58e4a5572
commit 9acb3460a1
27 changed files with 2059 additions and 341 deletions
Download Patch File Download Diff File Expand all files Collapse all files
training/tests/test_training_contract.py
+110 -4
View File
@@ -15,6 +15,7 @@ if str(TRAINING_ROOT) not in sys.path:
from trader_training.onnx_export import LinearHead, export_heads
from trader_training.io_utils import read_json, write_json
from trader_training.labels import ENTRY_LABEL_METHOD, _path_stats_for_group, build_entry_labels
from trader_training.promote import promote_artifact_bundle
from trader_training.replay import build_splits
from trader_training.schemas import FEATURE_ORDER, LATEST_STRESS_SPLIT, MODEL_OUTPUTS, OUTPUT_MAPPING, TRAINING_SPLITS, VALIDATION_LOCKED_SPLIT
@@ -22,10 +23,10 @@ from trader_training.schemas import FEATURE_ORDER, LATEST_STRESS_SPLIT, MODEL_OU
class TrainingContractTest(unittest.TestCase):
def test_feature_order_is_v4_contract_size(self) -> None:
self.assertEqual(39, len(FEATURE_ORDER))
self.assertEqual(54, len(FEATURE_ORDER))
self.assertEqual(len(FEATURE_ORDER), len(set(FEATURE_ORDER)))
self.assertEqual("ret_1m_bps", FEATURE_ORDER[0])
self.assertEqual("minutes_to_next_funding", FEATURE_ORDER[-1])
self.assertEqual("book_pressure_reversal_15m", FEATURE_ORDER[-1])
def test_output_mapping_matches_model_outputs(self) -> None:
for model_name, fields in MODEL_OUTPUTS.items():
@@ -67,6 +68,110 @@ class TrainingContractTest(unittest.TestCase):
self.assertEqual([VALIDATION_LOCKED_SPLIT, LATEST_STRESS_SPLIT], manifest["sealed_splits"])
self.assertEqual("FINAL_GATE_ONLY", manifest["latest_stress_policy"])
def test_path_stats_keeps_same_bar_target_stop_as_stop_first(self) -> None:
frame = pd.DataFrame(
{
"event_time": pd.date_range("2026-01-01", periods=6, freq="min", tz="UTC"),
"open_time_ms": np.arange(6, dtype=np.int64) * 60_000,
"symbol": "BTC-USDT-PERP",
"close": [100.0, 100.0, 100.0, 100.0, 100.0, 100.0],
"high": [100.0, 100.05, 100.20, 100.0, 100.0, 100.0],
"low": [100.0, 99.95, 99.70, 100.0, 100.0, 100.0],
"spread_bps": [1.0, 1.1, 1.2, 1.3, 1.4, 1.5],
}
)
stats = _path_stats_for_group(frame, "LONG", horizon=3, target_bps=10.0, stop_bps=8.0)
first = stats.loc[stats["open_time_ms"].eq(0)].iloc[0]
self.assertEqual(0, first["target_hit"])
self.assertEqual(1, first["stop_hit"])
self.assertEqual(1, first["ambiguous_hit"])
self.assertEqual(120_000, first["time_to_stop_ms"])
self.assertAlmostEqual(-8.0, first["gross_edge_bps"])
def test_entry_label_uses_max_future_edge_not_fixed_target_hit(self) -> None:
with tempfile.TemporaryDirectory() as tmp:
data_root = Path(tmp)
run_root = data_root / "trader-v4" / "runs" / "unit-entry"
feature_path = run_root / "feature" / "feature_frame.parquet"
replay_path = run_root / "replay" / "replay_1m.parquet"
plan_path = run_root / "label" / "price_plan_context.json"
config_path = data_root / "label_config.json"
feature_path.parent.mkdir(parents=True)
replay_path.parent.mkdir(parents=True)
times = pd.date_range("2026-01-01", periods=5, freq="min", tz="UTC")
pd.DataFrame(
{
"sample_id": ["s0", "s1"],
"symbol": "BTC-USDT-PERP",
"event_time": times[:2],
"open_time_ms": [0, 60_000],
"split_id": "fit_inner",
"walk_forward_fold": 0,
"data_quality_flag": "OK",
"spread_bps": 1.0,
"spread_rank_24h_pct": 0.1,
"realized_vol_15m_bps": 2.0,
}
).to_parquet(feature_path, index=False)
pd.DataFrame(
{
"event_time": times,
"open_time_ms": np.arange(5, dtype=np.int64) * 60_000,
"symbol": "BTC-USDT-PERP",
"open": [100.0, 100.0, 100.0, 100.0, 100.0],
"high": [100.0, 100.05, 100.19, 100.20, 100.0],
"low": [100.0, 99.99, 99.98, 99.97, 100.0],
"close": [100.0, 100.0, 100.0, 100.0, 100.0],
"spread_bps": 1.0,
}
).to_parquet(replay_path, index=False)
write_json(
config_path,
{
"entry": {
"max_hold_minutes": 3,
"target_bps": 50.0,
"stop_bps": 50.0,
"min_expected_net_edge_bps": 3.0,
}
},
)
write_json(
plan_path,
{
"pricePlanId": "unit-plan",
"pricePlanConfigHash": "unit-hash",
"targetDistanceBps": 50.0,
"stopDistanceBps": 50.0,
"maxHoldMinutes": 3,
"costBps": 6.5,
"entryLabelMethod": ENTRY_LABEL_METHOD,
},
)
build_entry_labels(
Namespace(
data_root=data_root,
run_id="unit-entry",
feature_path=feature_path,
replay_path=replay_path,
label_config_path=config_path,
cost_config_path=None,
price_plan_context_path=plan_path,
)
)
labels = pd.read_parquet(run_root / "label" / "entry_labels.parquet")
row = labels[labels["sample_id"].eq("s0") & labels["side"].eq("LONG")].iloc[0]
self.assertEqual(0, row["target_hit"])
self.assertEqual(1, row["entry_target"])
self.assertEqual(ENTRY_LABEL_METHOD, row["label_method"])
self.assertAlmostEqual(13.5, row["expected_net_edge_bps"], places=6)
self.assertAlmostEqual(row["mfe_bps"] - row["cost_bps"], row["max_achievable_net_edge_bps"], places=6)
def test_exported_onnx_accepts_java_feature_shape(self) -> None:
import onnxruntime as ort
@@ -78,13 +183,14 @@ class TrainingContractTest(unittest.TestCase):
LinearHead(
"direction",
"softmax",
np.zeros((39, 3), dtype=np.float32),
np.zeros((len(FEATURE_ORDER), 3), dtype=np.float32),
np.array([0.1, 0.2, 0.3], dtype=np.float32),
)
],
feature_count=len(FEATURE_ORDER),
)
session = ort.InferenceSession(str(path))
output = session.run(None, {"features": np.zeros((1, 39), dtype=np.float32)})[0]
output = session.run(None, {"features": np.zeros((1, len(FEATURE_ORDER)), dtype=np.float32)})[0]
self.assertEqual((1, 3), output.shape)
self.assertAlmostEqual(1.0, float(output.sum()), places=6)