Certification

Most backtests are overfit. cpz_quant.certification is the referee layer: the same open math that powers the CPZ Certification Standard on the CPZAI operating system, so any grade can be independently reproduced.

Probability of Backtest Overfitting (PBO)

Implements the combinatorially symmetric cross-validation (CSCV) method. Feed it the per-period returns of every configuration you tried, shape (T, N):

import numpy as np
from cpz_quant.certification import probability_of_backtest_overfitting

trials = np.column_stack([cfg1_returns, cfg2_returns, ...])  # all N trials
res = probability_of_backtest_overfitting(trials, n_splits=16)

res.pbo                       # P(best in-sample config is below-median OOS)
res.mean_logit                # < 0 leans overfit
res.performance_degradation   # slope of OOS vs IS performance

A PBO above ~0.5 means the selection process is more likely than not overfit.

Risk analytics

compute_risk_analytics(equity, benchmark=None, dates=None) produces the institutional due-diligence block from a daily equity curve: Sortino, Calmar, Ulcer index, tail ratio, CVaR, minimum track record length (MinTRL), up/down capture, and crisis behavior. Rust-accelerated when the optional extension is present, with an identical pure-NumPy fallback.

The certification grade

certify grades a strategy across six dimensions (robustness, out-of-sample, risk, factor independence, capacity/cost, live track record) with hard minimum gates, including the Deflated Sharpe Ratio, which discounts the Sharpe ratio for the number of trials that produced it:

from cpz_quant.certification import certify, CertMetrics, CertRigor

result = certify(
    backtest_verified=True,
    metrics=CertMetrics(sharpe_ratio=1.4, max_drawdown=-0.18, total_trades=420),
    rigor=CertRigor(deflated_sharpe=0.9, num_trials=25, oos_sharpe=1.1),
)
print(result.grade, result.composite, [g for g in result.gates if not g.passed])

Strategies without a qualifying live record receive a Provisional grade; the live dimension activates once enough live trades and days accumulate.

Regime-conditional performance

regime_conditional_performance breaks performance down by market regime, exposing strategies that only work in one environment.

Factor attribution

certification.factor_attribution neutralises common factors and reports how much alpha is genuinely idiosyncratic.