""" Statistical tests on VaR exceedances. - Kupiec (1995) proportion of failures: is the exceedance rate consistent with 1 - confidence? - Christoffersen (1998) independence: do exceedances cluster? - Conditional coverage: both at once. - Basel traffic light: zone from the binomial probability of observing at most that many exceedances. """ import numpy as np import pandas as pd from scipy.special import xlogy from scipy.stats import binom, chi2 from data.schema import Out # Basel traffic-light thresholds on the cumulative binomial # probability of the observed exceedance count. GREEN_LIMIT = 0.95 YELLOW_LIMIT = 0.9999 def _bernoulli_loglik(failures, successes, p): """ log L = x ln p + (n - x) ln(1 - p), with 0 ln 0 = 0. """ return xlogy(failures, p) + xlogy(successes, 1 - p) def kupiec_test(exceedances, observations, confidence): p = 1 - confidence x = exceedances n = observations lr = -2 * ( _bernoulli_loglik(x, n - x, p) - _bernoulli_loglik(x, n - x, x / n) ) return float(lr), float(chi2.sf(lr, df=1)) def christoffersen_independence_test(hits): """ Likelihood-ratio test of a first-order Markov chain against independent exceedances. """ hits = np.asarray(hits, dtype=int) previous, current = hits[:-1], hits[1:] n00 = np.sum((previous == 0) & (current == 0)) n01 = np.sum((previous == 0) & (current == 1)) n10 = np.sum((previous == 1) & (current == 0)) n11 = np.sum((previous == 1) & (current == 1)) pi01 = n01 / (n00 + n01) if n00 + n01 else 0.0 pi11 = n11 / (n10 + n11) if n10 + n11 else 0.0 pi = (n01 + n11) / len(current) lr = -2 * ( _bernoulli_loglik(n01 + n11, n00 + n10, pi) - _bernoulli_loglik(n01, n00, pi01) - _bernoulli_loglik(n11, n10, pi11) ) lr = max(float(lr), 0.0) return lr, float(chi2.sf(lr, df=1)) def traffic_light(exceedances, observations, confidence): probability = binom.cdf( exceedances, observations, 1 - confidence ) if probability < GREEN_LIMIT: return "Green" if probability < YELLOW_LIMIT: return "Yellow" return "Red" def backtest_statistics( forecasts: pd.DataFrame, significance: float, ) -> pd.DataFrame: """ One row of coverage statistics per confidence level. """ rows = [] for confidence, group in forecasts.groupby(Out.CONFIDENCE): hits = group[Out.EXCEEDANCE].to_numpy() n = len(hits) x = int(hits.sum()) kupiec_lr, kupiec_p = kupiec_test(x, n, confidence) ind_lr, ind_p = christoffersen_independence_test(hits) cc_lr = kupiec_lr + ind_lr cc_p = float(chi2.sf(cc_lr, df=2)) # Severity: realised loss over predicted ES on exceedance # days. Around 1 means ES is well calibrated. exceeded = group[group[Out.EXCEEDANCE]] es_ratio = ( float( exceeded[Out.REALIZED_LOSS].mean() / exceeded[Out.ES].mean() ) if x else np.nan ) if kupiec_p < significance: verdict = ( "Rejected: too many exceedances" if x > n * (1 - confidence) else "Rejected: too few exceedances (overly conservative)" ) elif ind_p < significance: verdict = "Rejected: exceedances cluster" elif cc_p < significance: verdict = "Rejected: conditional coverage" else: verdict = "Not rejected" rows.append( { Out.CONFIDENCE: confidence, Out.OBSERVATIONS: n, Out.EXPECTED_EXCEEDANCES: n * (1 - confidence), Out.EXCEEDANCES: x, Out.EXCEEDANCE_RATE: x / n, Out.KUPIEC_LR: kupiec_lr, Out.KUPIEC_P: kupiec_p, Out.INDEPENDENCE_LR: ind_lr, Out.INDEPENDENCE_P: ind_p, Out.CONDITIONAL_LR: cc_lr, Out.CONDITIONAL_P: cc_p, Out.ZONE: traffic_light(x, n, confidence), Out.ES_RATIO: es_ratio, Out.VERDICT: verdict, } ) return pd.DataFrame(rows)