"""
Free-tier walk-forward accuracy seed.

Generates a small set of historical forecast_snapshots with honest statistical
p50 forecasts and true closes so /accuracy is non-empty without paid APIs.

Usage (from backend/ with venv):
  python scripts/seed_accuracy_backtest.py --symbols AAPL,MSFT,GOOGL --horizons 1,5,20 --windows 30

Keeps Yahoo + Turso writes tiny (default ~ a few hundred rows).
"""

from __future__ import annotations

import argparse
import asyncio
import os
import sys
import time
from pathlib import Path

sys.path.insert(0, str(Path(__file__).resolve().parent.parent))

from dotenv import load_dotenv

load_dotenv()

from app.services.accuracy_service import close_on_or_before, compute_errors
from app.services.data_service import StockDataService
from app.services.database_service import DatabaseService
from app.services.timesfm_service import TimesFMService


async def seed(
    symbols: list[str],
    horizons: list[int],
    windows: int,
    delay: float,
) -> dict:
    db = DatabaseService(url=os.getenv("TURSO_URL"), token=os.getenv("TURSO_TOKEN"))
    data = StockDataService()
    model = TimesFMService()
    # Free-tier seed: skip heavy model load; statistical path is honest and labeled
    model.available = False
    model.model = None
    written = 0
    evaluated = 0
    try:
        await db.initialize()
        for symbol in symbols:
            hist = await data.get_stock_data(symbol, period="2y")
            if hist is None or len(hist) < 80:
                print(f"[skip] {symbol}: insufficient history")
                continue
            closes = [float(x) for x in hist["Close"].tolist()]
            # Walk back from end: use windows ending before last bar
            max_h = max(horizons)
            end_idx = len(closes) - 1
            start_idx = max(60, end_idx - windows - max_h)
            for i in range(start_idx, end_idx - max_h):
                context = closes[: i + 1]
                entry = context[-1]
                # Forecast date ≈ bar timestamp
                ts = hist.index[i]
                forecast_date = int(ts.timestamp()) // 86400 * 86400
                for h in horizons:
                    target_i = i + h
                    if target_i >= len(closes):
                        continue
                    pred = await model.predict(context, horizon=h)
                    p50 = float(pred["quantiles"]["p50"][-1])
                    # Insert snapshot as if made on forecast_date
                    await db._execute(
                        """
                        INSERT OR IGNORE INTO forecast_snapshots
                        (symbol, country, exchange, forecast_date, horizon_days, target_date,
                         p50_price, entry_price, created_at)
                        VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
                        """,
                        [
                            symbol.upper(),
                            "us",
                            "",
                            forecast_date,
                            h,
                            forecast_date + h * 86400,
                            p50,
                            entry,
                            forecast_date,
                        ],
                    )
                    written += 1
                    # Evaluate immediately with known actual
                    actual = float(closes[target_i])
                    pct_err, direction_ok = compute_errors(entry, p50, actual)
                    await db._execute(
                        """
                        UPDATE forecast_snapshots
                        SET actual_price = ?, percentage_error = ?, direction_correct = ?,
                            evaluated_at = ?
                        WHERE symbol = ? AND horizon_days = ? AND forecast_date = ?
                          AND evaluated_at IS NULL
                        """,
                        [
                            actual,
                            pct_err,
                            direction_ok,
                            int(time.time()),
                            symbol.upper(),
                            h,
                            forecast_date,
                        ],
                    )
                    evaluated += 1
            print(f"[ok] {symbol}")
            await asyncio.sleep(delay)
    finally:
        await db.close()
    return {"written": written, "evaluated": evaluated}


def main():
    p = argparse.ArgumentParser()
    p.add_argument("--symbols", default="AAPL,MSFT,GOOGL,AMZN,META,NVDA,JPM,XOM,JNJ,WMT")
    p.add_argument("--horizons", default="1,5,20,60")
    p.add_argument("--windows", type=int, default=40, help="Lookback windows per symbol (keep small)")
    p.add_argument("--delay", type=float, default=0.3)
    args = p.parse_args()
    symbols = [s.strip().upper() for s in args.symbols.split(",") if s.strip()]
    horizons = [int(x) for x in args.horizons.split(",") if x.strip()]
    # Cap to protect free Turso write quota
    windows = min(args.windows, 60)
    result = asyncio.run(seed(symbols, horizons, windows, args.delay))
    print(result)


if __name__ == "__main__":
    main()