from __future__ import annotations import json import logging from pathlib import Path from typing import Any import numpy as np import pandas as pd from sklearn.linear_model import HuberRegressor, LogisticRegression from sklearn.metrics import brier_score_loss, mean_absolute_error, roc_auc_score from sklearn.preprocessing import StandardScaler from trader_training.io_utils import read_json, read_parquet, run_root, sha256_json, write_json, write_parquet, write_text from trader_training.labels import _build_path_stats from trader_training.schemas import FEATURE_ORDER, FIT_SPLIT, LATEST_STRESS_SPLIT, TUNE_SPLIT, VALIDATION_LOCKED_SPLIT STATE_FEATURES = [ "position_side_sign", "time_in_position_minutes", "unrealized_pnl_bps", "mfe_since_entry_bps", "mae_since_entry_bps", "distance_to_stop_bps", "distance_to_target_bps", ] EVAL_SPLITS = (TUNE_SPLIT, VALIDATION_LOCKED_SPLIT, LATEST_STRESS_SPLIT) ALL_SPLITS = (FIT_SPLIT, TUNE_SPLIT, VALIDATION_LOCKED_SPLIT, LATEST_STRESS_SPLIT) def run_state_continue_experiment(args: Any) -> None: root = run_root(args) baseline_root = args.data_root / "trader-v4" / "runs" / args.baseline_run_id out_dir = root / "experiments" / "state_continue" ages = _parse_ages(args.ages_minutes) logging.info( "trader.training.state_continue_experiment_started runId=%s baselineRunId=%s ages=%s", args.run_id, args.baseline_run_id, ages, ) feature = _load_feature_frame(baseline_root) entry = _load_entry_labels(baseline_root) replay = _load_replay(baseline_root) plan = read_json(baseline_root / "label" / "price_plan_context.json") stop_bps = float(plan["stopDistanceBps"]) target_bps = float(plan["targetDistanceBps"]) cost_bps = float(plan["costBps"]) state_frame = _build_state_frame(feature, entry, replay, ages, stop_bps, target_bps, cost_bps) if args.max_rows_per_split: state_frame = _cap_rows_per_split(state_frame, int(args.max_rows_per_split)) dataset_hash = write_parquet(out_dir / "state_continue_train.parquet", state_frame) logging.info( "trader.training.state_continue_dataset_written runId=%s rowCount=%s splitCounts=%s path=%s", args.run_id, len(state_frame), state_frame["split_id"].value_counts().to_dict(), out_dir / "state_continue_train.parquet", ) source_manifest = _source_manifest(args, baseline_root, ages, stop_bps, target_bps, cost_bps, state_frame, dataset_hash) write_json(out_dir / "experiment_manifest.json", source_manifest) write_json(out_dir / "position_state_feature_schema.json", _state_feature_schema()) order_hash = write_json(out_dir / "position_state_feature_order.json", STATE_FEATURES) write_json( out_dir / "position_state_source_manifest.json", { "entry_predicted_edge_bps": "NOT_USED_IN_THIS_DIAGNOSTIC", "entry_direction_prob": "NOT_USED_IN_THIS_DIAGNOSTIC", "out_of_fold_used": False, "frozen_model_output_used": False, "replay_decision_trace_used": False, "state_feature_order_hash": order_hash, "row_count": len(state_frame), "split_counts": state_frame["split_id"].value_counts().to_dict(), }, ) feature_sets = { "market_only": FEATURE_ORDER, "market_plus_state": [*FEATURE_ORDER, *STATE_FEATURES], } results: dict[str, Any] = {} prediction_frames: list[pd.DataFrame] = [] for side in ("LONG", "SHORT"): side_frame = state_frame[state_frame["position_side"].eq(side)].copy() for feature_set_name, feature_columns in feature_sets.items(): key = f"{side.lower()}_{feature_set_name}" result, predictions = _train_side_models(side_frame, side, feature_columns) results[key] = result predictions["side"] = side predictions["feature_set"] = feature_set_name prediction_frames.append(predictions) logging.info( "trader.training.state_continue_model_trained runId=%s side=%s featureSet=%s tuneAuc=%s tuneMaeRatio=%s", args.run_id, side, feature_set_name, result.get(TUNE_SPLIT, {}).get("continue_auc"), result.get(TUNE_SPLIT, {}).get("edge_mae_vs_constant_ratio"), ) predictions = pd.concat(prediction_frames, ignore_index=True) if prediction_frames else pd.DataFrame() write_parquet(out_dir / "state_continue_predictions.parquet", predictions) write_json(out_dir / "state_continue_result.json", results) write_text(out_dir / "state_continue_experiment_report.md", _report(args, baseline_root, source_manifest, results)) logging.info("trader.training.state_continue_experiment_finished runId=%s report=%s", args.run_id, out_dir / "state_continue_experiment_report.md") def _parse_ages(raw: str) -> list[int]: ages = [int(item.strip()) for item in raw.split(",") if item.strip()] if not ages or any(age <= 0 for age in ages): raise ValueError(f"invalid ages-minutes: {raw}") return sorted(set(ages)) def _load_feature_frame(baseline_root: Path) -> pd.DataFrame: feature = read_parquet(baseline_root / "feature" / "feature_frame.parquet") required = {"sample_id", "symbol", "event_time", "open_time_ms", "split_id", "walk_forward_fold", "data_quality_flag", *FEATURE_ORDER} missing = sorted(required.difference(feature.columns)) if missing: raise ValueError(f"baseline feature frame missing columns: {missing}") feature = feature[feature["data_quality_flag"].isin(["OK", "PARTIAL_OPTIONAL"])].copy() feature = feature[feature["split_id"].isin(ALL_SPLITS)].copy() return feature def _load_entry_labels(baseline_root: Path) -> pd.DataFrame: entry = read_parquet(baseline_root / "label" / "entry_labels.parquet") required = {"sample_id", "symbol", "event_time", "side", "entry_target", "split_id", "walk_forward_fold"} missing = sorted(required.difference(entry.columns)) if missing: raise ValueError(f"baseline entry labels missing columns: {missing}") entry = entry[(entry["entry_target"] == 1) & (entry["side"].isin(["LONG", "SHORT"]))].copy() entry["entry_open_time_ms"] = pd.to_datetime(entry["event_time"], utc=True).astype("int64") // 1_000_000 return entry[["sample_id", "symbol", "event_time", "side", "entry_open_time_ms"]].copy() def _load_replay(baseline_root: Path) -> pd.DataFrame: split_manifest = read_json(baseline_root / "split" / "split_manifest.json") replay_path = Path(split_manifest["source_replay_path"]) replay = read_parquet(replay_path) required = {"symbol", "event_time", "open_time_ms", "high", "low", "close", "spread_bps"} missing = sorted(required.difference(replay.columns)) if missing: raise ValueError(f"source replay missing columns: {missing}") return replay.sort_values(["symbol", "open_time_ms"]).reset_index(drop=True) def _build_state_frame( feature: pd.DataFrame, entry: pd.DataFrame, replay: pd.DataFrame, ages: list[int], stop_bps: float, target_bps: float, cost_bps: float, ) -> pd.DataFrame: future_stats = _build_path_stats(replay, horizon=30, target_bps=target_bps, stop_bps=stop_bps) future_stats = future_stats.rename(columns={"open_time_ms": "current_open_time_ms"}) current_feature = feature.rename(columns={"sample_id": "current_sample_id", "event_time": "current_event_time", "open_time_ms": "current_open_time_ms"}) replay_state_source = _state_source_by_age(replay, ages) frames: list[pd.DataFrame] = [] for age in ages: candidates = entry.copy() candidates["time_in_position_minutes"] = age candidates["current_open_time_ms"] = candidates["entry_open_time_ms"] + age * 60_000 candidates = candidates.merge( replay_state_source[replay_state_source["time_in_position_minutes"].eq(age)], on=["symbol", "current_open_time_ms", "time_in_position_minutes"], how="inner", ) candidates = candidates.merge(current_feature, on=["symbol", "current_open_time_ms"], how="inner") candidates = candidates.merge( future_stats, left_on=["symbol", "current_open_time_ms", "side"], right_on=["symbol", "current_open_time_ms", "side"], how="inner", ) if candidates.empty: continue frames.append(_state_rows_for_age(candidates, stop_bps, target_bps, cost_bps)) logging.info("trader.training.state_continue_age_built ageMinutes=%s rowCount=%s", age, len(candidates)) if not frames: raise ValueError("state continue experiment produced no rows") out = pd.concat(frames, ignore_index=True) out = out.replace([np.inf, -np.inf], np.nan) required = [*FEATURE_ORDER, *STATE_FEATURES, "continue_target", "expected_continue_edge_bps"] out = out.dropna(subset=required).copy() return out def _state_source_by_age(replay: pd.DataFrame, ages: list[int]) -> pd.DataFrame: frames: list[pd.DataFrame] = [] for _, group in replay.groupby("symbol", sort=False, observed=False): group = group.sort_values("open_time_ms").copy() for age in ages: rolling_high = group["high"].rolling(age + 1, min_periods=age + 1).max() rolling_low = group["low"].rolling(age + 1, min_periods=age + 1).min() frame = pd.DataFrame( { "symbol": group["symbol"], "current_open_time_ms": group["open_time_ms"], "time_in_position_minutes": age, "entry_price": group["close"].shift(age), "current_price": group["close"], "high_since_entry": rolling_high, "low_since_entry": rolling_low, } ) frames.append(frame.dropna()) return pd.concat(frames, ignore_index=True) if frames else pd.DataFrame() def _state_rows_for_age(frame: pd.DataFrame, stop_bps: float, target_bps: float, cost_bps: float) -> pd.DataFrame: side_sign = np.where(frame["side"].eq("LONG"), 1.0, -1.0) entry_price = frame["entry_price"].astype(float) current_price = frame["current_price"].astype(float) high_since = frame["high_since_entry"].astype(float) low_since = frame["low_since_entry"].astype(float) long_mask = frame["side"].eq("LONG") unrealized = np.where(long_mask, (current_price / entry_price - 1.0) * 10000.0, (entry_price / current_price - 1.0) * 10000.0) - cost_bps mfe = np.where(long_mask, (high_since / entry_price - 1.0) * 10000.0, (entry_price / low_since - 1.0) * 10000.0) mae = np.where(long_mask, (entry_price / low_since - 1.0) * 10000.0, (high_since / entry_price - 1.0) * 10000.0) stop_price = np.where(long_mask, entry_price * (1.0 - stop_bps / 10000.0), entry_price * (1.0 + stop_bps / 10000.0)) target_price = np.where(long_mask, entry_price * (1.0 + target_bps / 10000.0), entry_price * (1.0 - target_bps / 10000.0)) distance_to_stop = np.where(long_mask, (current_price / stop_price - 1.0) * 10000.0, (stop_price / current_price - 1.0) * 10000.0) distance_to_target = np.where(long_mask, (target_price / current_price - 1.0) * 10000.0, (current_price / target_price - 1.0) * 10000.0) expected_edge = frame["future_return_bps"].astype(float) - cost_bps continue_target = ((expected_edge >= 2.0) & (frame["mae_bps"].astype(float) < stop_bps)).astype("int8") out = frame[ [ "current_sample_id", "symbol", "current_event_time", "current_open_time_ms", "side", "split_id", "walk_forward_fold", *FEATURE_ORDER, ] ].copy() out = out.rename( columns={ "current_sample_id": "sample_id", "current_event_time": "event_time", "current_open_time_ms": "open_time_ms", "side": "position_side", } ) out["position_side_sign"] = side_sign.astype("float32") out["time_in_position_minutes"] = frame["time_in_position_minutes"].astype("float32") out["unrealized_pnl_bps"] = unrealized.astype("float32") out["mfe_since_entry_bps"] = np.maximum(mfe, 0.0).astype("float32") out["mae_since_entry_bps"] = np.maximum(mae, 0.0).astype("float32") out["distance_to_stop_bps"] = distance_to_stop.astype("float32") out["distance_to_target_bps"] = distance_to_target.astype("float32") out["continue_target"] = continue_target out["expected_continue_edge_bps"] = expected_edge.astype("float32") return out def _cap_rows_per_split(frame: pd.DataFrame, max_rows_per_split: int) -> pd.DataFrame: capped = [] for split_id, part in frame.sort_values("event_time").groupby("split_id", sort=False, observed=False): if len(part) > max_rows_per_split: part = part.tail(max_rows_per_split).copy() capped.append(part) logging.info("trader.training.state_continue_split_capped splitId=%s rowCount=%s maxRows=%s", split_id, len(part), max_rows_per_split) return pd.concat(capped, ignore_index=True) def _train_side_models(frame: pd.DataFrame, side: str, feature_columns: list[str]) -> tuple[dict[str, Any], pd.DataFrame]: train = frame[frame["split_id"].eq(FIT_SPLIT)].copy() if train.empty: raise ValueError(f"state continue {side} has no fit_inner rows") scaler = StandardScaler() x_train = scaler.fit_transform(train[feature_columns].astype("float32")) y_train_cls = train["continue_target"].astype(int).to_numpy() y_train_reg = train["expected_continue_edge_bps"].astype(float).to_numpy() clf = LogisticRegression(max_iter=500) clf.fit(x_train, y_train_cls) reg = HuberRegressor(alpha=0.001, epsilon=1.35, max_iter=300) reg.fit(x_train, y_train_reg) metrics: dict[str, Any] = {} prediction_frames: list[pd.DataFrame] = [] for split_id in ALL_SPLITS: part = frame[frame["split_id"].eq(split_id)].copy() if part.empty: continue x = scaler.transform(part[feature_columns].astype("float32")) y_cls = part["continue_target"].astype(int).to_numpy() y_reg = part["expected_continue_edge_bps"].astype(float).to_numpy() proba = clf.predict_proba(x)[:, 1] pred_edge = reg.predict(x) metrics[split_id] = _split_metrics(y_train_cls, y_train_reg, y_cls, y_reg, proba, pred_edge) pred_frame = part[["sample_id", "symbol", "event_time", "split_id", "position_side", "continue_target", "expected_continue_edge_bps"]].copy() pred_frame["continue_prob"] = proba.astype("float32") pred_frame["predicted_continue_edge_bps"] = pred_edge.astype("float32") prediction_frames.append(pred_frame) metrics["row_count"] = int(len(frame)) metrics["feature_count"] = len(feature_columns) metrics["feature_hash"] = sha256_json(feature_columns) return metrics, pd.concat(prediction_frames, ignore_index=True) def _split_metrics( y_train_cls: np.ndarray, y_train_reg: np.ndarray, y_cls: np.ndarray, y_reg: np.ndarray, proba: np.ndarray, pred_edge: np.ndarray, ) -> dict[str, Any]: train_rate = float(np.mean(y_train_cls)) constant_proba = np.full(len(y_cls), train_rate) train_median = float(np.median(y_train_reg)) constant_edge = np.full(len(y_reg), train_median) out: dict[str, Any] = { "row_count": int(len(y_cls)), "positive_rate": float(np.mean(y_cls)), "brier": float(brier_score_loss(y_cls, proba)), "constant_brier": float(brier_score_loss(y_cls, constant_proba)), "edge_mae": float(mean_absolute_error(y_reg, pred_edge)), "edge_constant_mae": float(mean_absolute_error(y_reg, constant_edge)), } if len(np.unique(y_cls)) == 2: out["continue_auc"] = float(roc_auc_score(y_cls, proba)) out["brier_vs_constant_ratio"] = float(out["brier"] / out["constant_brier"]) if out["constant_brier"] > 0 else None out["edge_mae_vs_constant_ratio"] = float(out["edge_mae"] / out["edge_constant_mae"]) if out["edge_constant_mae"] > 0 else None return out def _source_manifest( args: Any, baseline_root: Path, ages: list[int], stop_bps: float, target_bps: float, cost_bps: float, state_frame: pd.DataFrame, dataset_hash: str, ) -> dict[str, Any]: return { "experiment": "state_continue_diagnostic_v1", "run_id": args.run_id, "baseline_run_id": args.baseline_run_id, "baseline_root": str(baseline_root), "ages_minutes": ages, "target_bps": target_bps, "stop_bps": stop_bps, "cost_bps": cost_bps, "dataset_hash_sha256": dataset_hash, "row_count": int(len(state_frame)), "split_counts": state_frame["split_id"].value_counts().to_dict(), "side_counts": state_frame["position_side"].value_counts().to_dict(), "feature_inputs": { "market_feature_count": len(FEATURE_ORDER), "state_features": STATE_FEATURES, "state_feature_count": len(STATE_FEATURES), }, "leakage_policy": { "uses_future_entry_label_as_feature": False, "uses_same_round_model_prediction_as_feature": False, "entry_predicted_edge_bps": "not used", "entry_direction_prob": "not used", }, } def _state_feature_schema() -> list[dict[str, Any]]: return [ {"name": "position_side_sign", "unit": "-1/1", "source": "synthetic position state", "leakage_check": "known at current position time"}, {"name": "time_in_position_minutes", "unit": "minute", "source": "entry time to current time", "leakage_check": "known at current position time"}, {"name": "unrealized_pnl_bps", "unit": "bps", "source": "entry price and current close", "leakage_check": "uses <= current time price"}, {"name": "mfe_since_entry_bps", "unit": "bps", "source": "high since entry", "leakage_check": "uses only entry..current high"}, {"name": "mae_since_entry_bps", "unit": "bps", "source": "low/high since entry", "leakage_check": "uses only entry..current low/high"}, {"name": "distance_to_stop_bps", "unit": "bps", "source": "price plan and current close", "leakage_check": "uses fixed plan and current price"}, {"name": "distance_to_target_bps", "unit": "bps", "source": "price plan and current close", "leakage_check": "uses fixed plan and current price"}, ] def _report(args: Any, baseline_root: Path, manifest: dict[str, Any], results: dict[str, Any]) -> str: baseline = read_json(baseline_root / "model" / "model_train_manifest.json") continue_metrics = baseline["CONTINUE"]["metrics"] lines = [ "# State Continue Experiment Report", "", f"- run_id: `{args.run_id}`", f"- baseline_run_id: `{args.baseline_run_id}`", f"- row_count: `{manifest['row_count']}`", f"- ages_minutes: `{manifest['ages_minutes']}`", "", "## Baseline run-10 Continue", "", "| head | auc | mae_vs_constant |", "| --- | ---: | ---: |", f"| long_continue_prob | {continue_metrics['long_continue_prob'].get('auc')} | |", f"| short_continue_prob | {continue_metrics['short_continue_prob'].get('auc')} | |", f"| long_expected_continue_edge_bps | | {continue_metrics['long_expected_continue_edge_bps'].get('mae_vs_constant_ratio')} |", f"| short_expected_continue_edge_bps | | {continue_metrics['short_expected_continue_edge_bps'].get('mae_vs_constant_ratio')} |", "", "## Diagnostic Result", "", "| side | feature_set | split | rows | auc | brier_ratio | mae_ratio |", "| --- | --- | --- | ---: | ---: | ---: | ---: |", ] for key, item in results.items(): side, feature_set = key.split("_", 1) for split_id in EVAL_SPLITS: metric = item.get(split_id, {}) lines.append( f"| {side.upper()} | {feature_set} | {split_id} | {metric.get('row_count')} | {metric.get('continue_auc')} | {metric.get('brier_vs_constant_ratio')} | {metric.get('edge_mae_vs_constant_ratio')} |" ) lines.extend( [ "", "## Verdict Rule", "", "状态特征只有在 `market_plus_state` 同时好过 `market_only`，并且 validation_locked / latest_stress 没有反向变差时，才进入正式链路。", "", ] ) return "\n".join(lines)