kalibrierung.png

def fig_calibration(y_test, proba_raw, proba_cal) -> None:
    """Calibration curve: class_weight='balanced' distorts predict_proba(), and
    CalibratedClassifierCV fixes it. Shown side by side so the effect is
    visible directly, not just asserted in prose. See Module 25 for the full
    calibration-in-the-large/slope/DCA workflow that builds on this."""
    brier_raw = brier_score_loss(y_test, proba_raw)
    brier_cal = brier_score_loss(y_test, proba_cal)
    frac_raw, mean_raw = calibration_curve(y_test, proba_raw, n_bins=5, strategy="quantile")
    frac_cal, mean_cal = calibration_curve(y_test, proba_cal, n_bins=5, strategy="quantile")

    fig, ax = plt.subplots(figsize=(6.5, 6))
    ax.plot([0, 1], [0, 1], color=SECONDARY, linestyle="--", linewidth=1.0,
            label="Perfekte Kalibrierung")
    ax.plot(mean_raw, frac_raw, "o-", color=EVENT, linewidth=2.0, markersize=7,
            label=f"class_weight='balanced', unkalibriert\n(Brier = {brier_raw:.3f}, Ø vorhergesagt {proba_raw.mean():.0%})")
    ax.plot(mean_cal, frac_cal, "o-", color=PRIMARY, linewidth=2.0, markersize=7,
            label=f"nach CalibratedClassifierCV\n(Brier = {brier_cal:.3f}, Ø vorhergesagt {proba_cal.mean():.0%})")
    ax.set_xlabel("Vorhergesagte Wahrscheinlichkeit")
    ax.set_ylabel("Beobachtete Ereignisrate")
    ax.set_title("Kalibrierungskurve: 'balanced' verzerrt, Rekalibrierung korrigiert\n"
                 f"(30-Tage-Mortalität, Testdaten, beobachtet {y_test.mean():.0%})")
    ax.legend(loc="upper left", fontsize=9.5)
    ax.set_xlim(0, 1)
    ax.set_ylim(0, 1)
    save(fig, ASSETS / "kalibrierung.png")