Function _score

python/openquant/feature_diagnostics.py:199–231

(
    y_true: Sequence[float],
    prob: Sequence[float],
    scoring: str,
    sample_weight: Sequence[float] | None,
)

        for i in range(dim):
            xtwy[i] += w * design[i] * yy
            for j in range(dim):
                xtwx[i][j] += w * design[i] * design[j]

    for i in range(dim):
        xtwx[i][i] += ridge

    beta = _solve_linear_system(xtwx, xtwy)
    return _LinearModel(coeffs=beta[1:], intercept=beta[0])


def _predict_proba(model: _LinearModel, x: Sequence[Sequence[float]]) -> list[float]:
    return [_sigmoid(model.intercept + _dot(row, model.coeffs)) for row in x]


def _score(
    y_true: Sequence[float],
    prob: Sequence[float],
    scoring: str,
    sample_weight: Sequence[float] | None,
) -> float:
    weights = [1.0] * len(y_true) if sample_weight is None else [float(v) for v in sample_weight]
    den = sum(weights)
    if den <= 0:
        return 0.0

    if scoring == "neg_log_loss":
        loss = 0.0
        for y, p, w in zip(y_true, prob, weights):
            p_clip = min(max(p, 1e-15), 1.0 - 1e-15)
            loss += w * (-(y * log(p_clip) + (1.0 - y) * log(1.0 - p_clip)))
        return -(loss / den)

    pred = [1.0 if p >= 0.5 else 0.0 for p in prob]

    if scoring == "accuracy":
        correct = sum(w for y, p, w in zip(y_true, pred, weights) if abs(y - p) < 1e-12)