#!/usr/bin/env python3 """ Полная тренировка MoE v12 для всех 14 MOEX тикеров. Запуск: python3 train_experts_all.py python3 train_experts_all.py --ticker SBER # один тикер Архитектура: MultiTimeframeMoE (16 LSTM экспертов + XGBoost, H1 + MTF контекст D1/W1) Flow: 85% train → OOS thresholds → retrain 100% → save models/saved/{ticker}_moe_v12.joblib """ import sys, os, time, json, warnings import numpy as np from datetime import datetime warnings.filterwarnings('ignore') sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) import config config.set_device('train') # CUDA for training from config import MOEX_TICKERS, SAVE_DIR, TARGET_CONFIG, DEVICE from models.moe import MultiTimeframeMoE, _prepare_df, _get_feature_cols, _build_dataset, _flat_mask from models.experts import ExpertEnsemble import xgboost as xgb from config import MODEL_CONFIG, TICKER_SHORT_WEIGHTS, DEFAULT_SHORT_WEIGHT LOG_FILE = f'{SAVE_DIR}/moe_training_log.json' def _get_scale_weights(y_full, ticker): """Вычисляет scale_pos_weight для XGBoost.""" n = len(y_full) pos = float(y_full.sum()) neg = n - pos scale = min(max(neg / max(pos, 1), 1.0), 5.0) if ticker in TICKER_SHORT_WEIGHTS: scale *= TICKER_SHORT_WEIGHTS[ticker] return scale def train_moe_for(ticker): """Тренирует MoE v12 для одного тикера. Возвращает dict с метриками.""" t0 = time.time() print(f' [{datetime.now().strftime("%H:%M")}] Загрузка данных...') # ── Step 1: Train on 85% → OOS ── moe = MultiTimeframeMoE(ticker) ok = moe.train(limit=None, verbose=True, retrain_on_full=False) if not ok: return {'status': 'error', 'message': 'train() вернул False'} oos_long_th = moe.optimal_thresholds.get('long', 0.55) oos_short_th = moe.optimal_thresholds.get('short', 0.55) oos_val_acc = getattr(moe, 'oos_val_acc', 0) oos_selected = moe.selected_experts print(f' OOS: val_acc={oos_val_acc:.1%}, L≥{oos_long_th:.2f}, S≥{oos_short_th:.2f}') if oos_selected: print(f' Отобрано экспертов: {len(oos_selected)}/16') # ── Step 2: Retrain LSTM + XGBoost на 100% ── print(f' Retrain на 100% данных...') df_full = _prepare_df(ticker, limit=None) if df_full is None or len(df_full) < 200: return {'status': 'error', 'message': f'Недостаточно данных: {len(df_full) if df_full is not None else 0}'} all_features = _get_feature_cols() available = [c for c in all_features if c in df_full.columns] norm_stats = moe.ensemble.get_norm_stats() ensemble_full = ExpertEnsemble(n_features=len(available)) print(f' LSTM эксперты ({len(ensemble_full.expert_names)}) на {len(df_full)} строк...') ensemble_full.train_all(df_full, available, verbose=True, epochs=None) ensemble_full.set_norm_stats(*norm_stats) print(f' Forward pass...') signals_full = ensemble_full.predict_all(df_full, available) X_full = _build_dataset(df_full, signals_full, ensemble_full, available, keep_experts=oos_selected) y_full = np.column_stack([df_full['outcome_long'].values, df_full['outcome_short'].values]) min_len = min(len(X_full), len(y_full)) X_full, y_full = X_full[-min_len:], y_full[-min_len:] # ── ФЛЕТ-ФИЛЬТР ── flat_mask = _flat_mask(df_full).values[-len(X_full):] n_flat = int(flat_mask.sum()) y_full[flat_mask] = 0.0 print(f' Флет-фильтр: {n_flat}/{len(flat_mask)} ({n_flat/len(flat_mask):.1%})') # Статистика по успешности исходных таргетов (до фильтрации) long_sr = df_full['outcome_long'].mean() short_sr = df_full['outcome_short'].mean() print(f' Исходные success rate: long={long_sr:.1%}, short={short_sr:.1%}') print(f' Размер датасета: {len(X_full)} samples') print(f' XGBoost на {len(X_full)} samples...') params_xgb = MODEL_CONFIG['xgb'].copy() params_xgb.update({ 'max_depth': 4, 'learning_rate': 0.03, 'n_estimators': 300, 'subsample': 0.7, 'colsample_bytree': 0.7, 'reg_alpha': 0.1, 'reg_lambda': 1.0, 'min_child_weight': 5, }) p = params_xgb.copy() p['scale_pos_weight'] = min(max(y_full[:, 0].sum() / max(len(y_full) - y_full[:, 0].sum(), 1), 1.0), 5.0) xgb_long = xgb.XGBClassifier(**p) xgb_long.fit(X_full, y_full[:, 0]) pos_short = y_full[:, 1].sum() neg_short = len(y_full) - pos_short scale_short = min(max(neg_short / max(pos_short, 1), 1.0), 5.0) scale_short *= TICKER_SHORT_WEIGHTS.get(ticker, DEFAULT_SHORT_WEIGHT) p = params_xgb.copy() p['scale_pos_weight'] = scale_short xgb_short = xgb.XGBClassifier(**p) xgb_short.fit(X_full, y_full[:, 1]) # ── Оценка in-sample ── train_preds_long = xgb_long.predict(X_full) train_preds_short = xgb_short.predict(X_full) train_acc_long = (train_preds_long == y_full[:, 0]).mean() train_acc_short = (train_preds_short == y_full[:, 1]).mean() train_pos_rate_long = train_preds_long.mean() train_pos_rate_short = train_preds_short.mean() # Процент flat-баров, которые модель классифицирует как 1 (должны быть 0) flat_preds_long = train_preds_long[flat_mask] flat_preds_short = train_preds_short[flat_mask] flat_false_long = flat_preds_long.mean() if len(flat_preds_long) > 0 else 0 flat_false_short = flat_preds_short.mean() if len(flat_preds_short) > 0 else 0 # ── Step 3: Сборка production-модели ── moe.ensemble = ensemble_full moe.rf_long = xgb_long moe.rf_short = xgb_short moe.rf_model = None moe.calibrators = None moe.optimal_thresholds = {'long': oos_long_th, 'short': oos_short_th} moe.selected_experts = oos_selected moe.oos_val_acc = oos_val_acc moe.val_acc = oos_val_acc moe.feature_cols = available moe.n_base_features = len(available) # ── Save ── save_path = os.path.join(SAVE_DIR, f'{ticker.lower()}_moe_v12.joblib') moe.save(save_path) elapsed = time.time() - t0 return { 'status': 'ok', 'ticker': ticker, 'rows': len(df_full), 'val_acc': oos_val_acc, 'long_th': oos_long_th, 'short_th': oos_short_th, 'selected_experts': len(oos_selected) if oos_selected else 16, 'elapsed_min': round(elapsed / 60, 1), 'path': save_path, # Детальная статистика 'flat_pct': round(n_flat / len(flat_mask) * 100, 1), 'long_sr': round(long_sr * 100, 1), 'short_sr': round(short_sr * 100, 1), 'train_acc_long': round(train_acc_long * 100, 1), 'train_acc_short': round(train_acc_short * 100, 1), 'train_pos_rate_long': round(train_pos_rate_long * 100, 1), 'train_pos_rate_short': round(train_pos_rate_short * 100, 1), 'flat_false_long': round(flat_false_long * 100, 1), 'flat_false_short': round(flat_false_short * 100, 1), } if __name__ == '__main__': import argparse parser = argparse.ArgumentParser(description='Train MoE v12 for all MOEX tickers') parser.add_argument('--ticker', type=str, default=None, help='Single ticker only') args = parser.parse_args() tickers = [args.ticker.upper()] if args.ticker else MOEX_TICKERS print(f'\n{"="*70}') print(f' MoE v12 TRAINING — {datetime.now().strftime("%d.%m.%Y %H:%M")}') print(f' Трекеры: {", ".join(tickers)}') print(f' Архитектура: 16 LSTM экспертов + XGBoost + MTF контекст') print(f' TARGET_CONFIG: SL={TARGET_CONFIG["atr_mult_sl"]}×ATR TP={TARGET_CONFIG["atr_mult_tp"]}×ATR') print(f' Device: {DEVICE}') print(f'{"="*70}\n') results_log = {} t_start = time.time() errors = 0 for i, ticker in enumerate(tickers, 1): print(f'\n[{i}/{len(tickers)}] {ticker} {"="*50}') entry = {'ticker': ticker, 'status': 'error'} try: result = train_moe_for(ticker) if result is None: print(f' ⏭ Пропущен') entry['status'] = 'skipped' else: entry = result print(f' ✓ {result["elapsed_min"]}мин — ' f'val_acc={result["val_acc"]:.1%}, ' f'L≥{result["long_th"]:.2f}, S≥{result["short_th"]:.2f}, ' f'экспертов={result["selected_experts"]}/16, ' f'флет={result.get("flat_pct", "?")}%') except Exception as e: import traceback errors += 1 entry['status'] = 'error' entry['error'] = str(e) print(f' ❌ Ошибка: {e}') traceback.print_exc() results_log[ticker] = entry total_time = time.time() - t_start print(f'\n{"="*70}') print(f' ТРЕНИРОВКА ЗАВЕРШЕНА') print(f' Время: {total_time/60:.1f} мин') print(f' Ошибок: {errors}') ok = [v for v in results_log.values() if v.get('status') == 'ok'] if ok: accs = [v['val_acc'] for v in ok] print(f' Успешно: {len(ok)}/{len(tickers)}') print(f' Средняя val_acc: {np.mean(accs):.1%}') print(f'{"="*70}') print() print(f' ДЕТАЛЬНАЯ СТАТИСТИКА ПО КАЖДОМУ ТИКЕРУ:') print(f' {"="*60}') print(f' {"Тикер":>6} {"Rows":>6} {"val_acc":>8} {"L≥thr":>6} {"S≥thr":>6} ' f'{"Эксп":>4} {"Флет%":>6} {"L_SR%":>6} {"S_SR%":>6} ' f'{"Acc_L":>6} {"Acc_S":>6} {"ТоргL":>6} {"ТоргS":>6} ' f'{"flong":>6} {"fshort":>6} {"мин":>5}') print(f' {"-"*100}') for v in ok: print(f' {v["ticker"]:>6} {v["rows"]:>6} {v["val_acc"]:>7.1%} ' f'{v["long_th"]:>5.2f} {v["short_th"]:>5.2f} ' f'{v["selected_experts"]:>4} ' f'{v.get("flat_pct",0):>5.1f}% ' f'{v.get("long_sr",0):>5.1f}% {v.get("short_sr",0):>5.1f}% ' f'{v.get("train_acc_long",0):>5.1f}% {v.get("train_acc_short",0):>5.1f}% ' f'{v.get("train_pos_rate_long",0):>5.1f}% {v.get("train_pos_rate_short",0):>5.1f}% ' f'{v.get("flat_false_long",0):>5.1f}% {v.get("flat_false_short",0):>5.1f}% ' f'{v["elapsed_min"]:>4.1f}') # Итоговые средние if ok: print(f' {"-"*100}') print(f' {"СРЕДНЕЕ":>6} {np.mean([v["rows"] for v in ok]):>6.0f} ' f'{np.mean([v["val_acc"] for v in ok]):>7.1%} ' f'{np.mean([v["long_th"] for v in ok]):>5.2f} {np.mean([v["short_th"] for v in ok]):>5.2f} ' f'{np.mean([v["selected_experts"] for v in ok]):>4.0f} ' f'{np.mean([v.get("flat_pct",0) for v in ok]):>5.1f}% ' f'{np.mean([v.get("long_sr",0) for v in ok]):>5.1f}% {np.mean([v.get("short_sr",0) for v in ok]):>5.1f}% ' f'{np.mean([v.get("train_acc_long",0) for v in ok]):>5.1f}% {np.mean([v.get("train_acc_short",0) for v in ok]):>5.1f}% ' f'{np.mean([v.get("train_pos_rate_long",0) for v in ok]):>5.1f}% {np.mean([v.get("train_pos_rate_short",0) for v in ok]):>5.1f}% ' f'{np.mean([v.get("flat_false_long",0) for v in ok]):>5.1f}% {np.mean([v.get("flat_false_short",0) for v in ok]):>5.1f}% ' f'{np.mean([v["elapsed_min"] for v in ok]):>4.1f}') print(f'{"="*70}') print(f' Лог: {LOG_FILE}') with open(LOG_FILE, 'w') as f: json.dump(results_log, f, indent=2, default=str)