import numpy as np
import pandas as pd
import pytest

from btcmodel import forward
from btcmodel.evaluate import HORIZONS, score
from btcmodel.models import MODELS

from .test_models import synthetic_prices

MODELS_UNDER_TEST = [MODELS["random_walk"], MODELS["powerlaw"]]


@pytest.fixture
def prices():
    return synthetic_prices(lambda day: 0.001 + 0 * day, end="2025-03-31", noise=0.03)


def test_snapshot_records_every_model_and_horizon(prices, tmp_path):
    history = prices.loc[:"2024-11-26"]
    path = forward.snapshot(history, MODELS_UNDER_TEST, tmp_path, pd.Timestamp("2024-11-27"))
    recorded = pd.read_csv(path)
    assert path.name == "2024-11-26.csv"
    assert len(recorded) == len(MODELS_UNDER_TEST) * len(HORIZONS)
    assert recorded.target.iloc[0] == "2024-12-26"


def test_snapshot_refuses_stale_data_and_duplicates(prices, tmp_path):
    history = prices.loc[:"2024-11-26"]
    with pytest.raises(ValueError, match="real time"):
        forward.snapshot(history, MODELS_UNDER_TEST, tmp_path, pd.Timestamp("2024-12-10"))
    forward.snapshot(history, MODELS_UNDER_TEST, tmp_path, pd.Timestamp("2024-11-27"))
    with pytest.raises(FileExistsError):
        forward.snapshot(history, MODELS_UNDER_TEST, tmp_path, pd.Timestamp("2024-11-27"))


def test_scores_only_due_horizons_and_match_the_backtest(prices, tmp_path):
    history = prices.loc[:"2024-11-26"]
    forward.snapshot(history, MODELS_UNDER_TEST, tmp_path, pd.Timestamp("2024-11-27"))

    assert forward.score_snapshots(prices.loc[:"2024-12-20"], tmp_path).empty
    assert forward.next_due(prices.loc[:"2024-12-20"], tmp_path) == pd.Timestamp("2024-12-26")

    scores = forward.score_snapshots(prices, tmp_path)
    assert sorted(scores.horizon.unique()) == [30, 91]

    model = MODELS["powerlaw"]
    horizons = np.array([30, 91])
    direct = score(
        "powerlaw",
        model.forecast(history, horizons),
        np.log(
            prices.close.loc[history.index[-1] + pd.to_timedelta(horizons, unit="D")]
        ).to_numpy(),
    )
    recorded = scores[scores.model == "powerlaw"].sort_values("horizon")
    np.testing.assert_allclose(recorded.crps, direct.crps, atol=1e-5)