Statistical methods

Resampling uses a stationary bootstrap, so the null distribution preserves the block structure of a return series. Multiplicity control follows Romano and Wolf's stepwise method, studentising before taking the maximum and recentring per configuration.

White's Reality Check and Hansen's Superior Predictive Ability run on the same matrix. Bonferroni is computed and weighs in 0 decisions: it is returned so the distance between a conservative bound and the corrected one stays visible.

The Deflated Sharpe compares the best configuration against the ceiling luck alone reaches across the trials that were run. The standard error carries skewness and kurtosis, and the ceiling uses the expected maximum of N trials with its Euler correction. The result returns the null distribution, the luck ceiling, and whether the observed value clears it.

The effective number of trials is derived from the p-value of the maximum rather than counted, and two estimators are published side by side because they disagree: a per-grid count depends on the order the configurations were swept in, and the stepwise one does not.

The Probabilistic and the Deflated Sharpe stay two separate calculations: the first asks whether a Sharpe clears a benchmark on the sample it was measured on, the second asks what the search that found it was worth. The overfitting probability is computed by CSCV over blocks of the configuration matrix.

Overfitting is tested on the configurations that ran ungated, and not on the survivors: asking the selection to grade itself returns the answer it was built to give.

A correction that returns a number on a sample too small to support it is worse than a refusal, because the number travels and the sample size does not.

Training is an anchored expanding walk-forward, and the embargo between the training block and the test block defaults to the label horizon. The report names the model that actually ran along the LightGBM to scikit-learn to ridge seam.

The ml family holds 38 nodes, 24 of them validated by executable checks. Machine-learning nodes are outside the export targets for a structural reason: their features are cross-sectional ranks computed across the universe of the run, which a single-symbol target cannot compute.

The backend charges no friction it was not given: zero declared cost means zero charged cost, and the report states it.

ADF, KPSS, Jarque-Bera and Kolmogorov-Smirnov run on returns; Engle-Granger and Johansen run on the equity curves. No threshold is wired to any of them: p-values and critical values are returned for a reader, and 0 configurations are excluded on their result.

KPSS with its parameters estimated on the same data is conservative, and that property is written beside the test itself.

Ljung-Box, the Hurst exponent, GARCH and EGARCH conditional volatility, and change-point detection by BOCPD or CUSUM are nodes of the catalogue, each covered by executable validation.

Returns and the equity curve are separate inputs, and each result states which of the two it was computed on: a Hurst exponent on returns and a Hurst exponent on an equity curve answer different questions, and a figure that omits its input cannot be read.

Six controls attempt to refute a GO verdict, and their own error rates are measured: 0.017 false positives, 0.000 false negatives, and a power rising from 0.17 to 1.00 between 0.03 and 0.27 sigma.

No language model takes part in a verdict. The assistant triggers the validation suite and reads its output; the calculation remains separate from the conversation.

The perimeter of a result, the direction of its known effects and the control available in the build are registered on one page, entry by entry.

Validity domain →