#!/usr/bin/env python3
"""Run the backtest with TRAINED models (loads saved model weights).

Usage:
    python scripts/run_backtest_trained.py --data data/parquet/BTCUSDT_2021-05-19.parquet
    python scripts/run_backtest_trained.py --data data/parquet/LUNAUSDT_2022-05-10.parquet
"""
import argparse
import logging
import sys
from pathlib import Path

ML_DIR = Path(__file__).resolve().parent.parent / "ml"
sys.path.insert(0, str(ML_DIR))
PROJECT_ROOT = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(PROJECT_ROOT))

from flash_crash_watchdog.cascade import DetectionCascade
from flash_crash_watchdog.data.historical_loader import load_parquet
from flash_crash_watchdog.eval.backtest import run_backtest
from flash_crash_watchdog.models.stage2_isolation_forest import Stage2IsolationForest
from flash_crash_watchdog.models.stage3_tcn import Stage3TCN, TCNConfig


def main() -> int:
    parser = argparse.ArgumentParser(description="Run backtest with trained models")
    parser.add_argument("--data", required=True, help="Parquet file of crash data")
    parser.add_argument("--config", default="configs/pipeline.yml")
    parser.add_argument("--models", default="models/", help="Directory with trained models")
    parser.add_argument("--output", default="results/backtest_trained.json")
    parser.add_argument("--max-ticks", type=int, default=0,
                        help="Max ticks to process (0 = all, for speed use 500000)")
    args = parser.parse_args()

    logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")

    # Load cascade
    cascade = DetectionCascade.from_config(args.config)

    # Load trained models if they exist
    models_dir = Path(args.models)
    stage2_path = models_dir / "stage2_isolation_forest.joblib"
    stage3_path = models_dir / "stage3_tcn.pt"

    if stage2_path.exists():
        logger.info("Loading trained Stage 2 from %s", stage2_path)
        cascade.s2.load(stage2_path)
    else:
        logger.warning("No trained Stage 2 found at %s — using untrained fallback", stage2_path)

    if stage3_path.exists():
        logger.info("Loading trained Stage 3 from %s", stage3_path)
        cascade.s3.load(stage3_path)
    else:
        logger.warning("No trained Stage 3 found at %s — using untrained fallback", stage3_path)

    # Load data
    df = load_parquet(args.data)
    if args.max_ticks > 0 and len(df) > args.max_ticks:
        logger.info("Sampling down to %d ticks (from %d) for speed", args.max_ticks, len(df))
        indices = range(0, len(df), len(df) // args.max_ticks)
        df = df.iloc[indices[:args.max_ticks]].copy()

    # Run backtest
    results = run_backtest(cascade, df)
    results.print_summary()

    output_path = Path(args.output)
    output_path.parent.mkdir(parents=True, exist_ok=True)
    results.save(output_path)
    return 0


if __name__ == "__main__":
    raise SystemExit(main())