Glossary¶
This glossary defines key terms used throughout the temporalcv documentation, including the Knowledge Tier system used to tag confidence levels in claims.
Knowledge Tiers¶
- T1¶
Academically Validated—Claims with full academic citations and peer-reviewed support. Apply directly; these are established statistical methods.
Examples:
Diebold-Mariano test (Diebold & Mariano, 1995)
Harvey small-sample correction (Harvey et al., 1997)
HAC variance estimation (Newey & West, 1987)
Cross-fitting debiasing (Chernozhukov et al., 2018)
Action: Trust and apply. These methods have strong theoretical foundations.
- T2¶
Empirical Finding—Results validated through prior project work but lacking formal publication. May need adjustment for different domains.
Examples:
70th percentile threshold from prior forecasting work
20% improvement suspicion threshold
Webb weights for < 13 clusters
Action: Apply but monitor. Verify in your specific context.
- T3¶
Assumption Needing Justification—Domain-specific choices that require explicit justification for your data. Document why the assumption holds.
Examples:
13-week volatility window
AR(1) process assumption for residuals
Minimum 30 samples for DM test
Action: Question first. Run sensitivity analysis when possible.
Statistical Terms¶
- HAC¶
Heteroskedasticity and Autocorrelation Consistent variance estimation. Used in the DM test to account for serial correlation in forecast errors. Implementation uses Bartlett kernel with Andrews (1991) automatic bandwidth. See
temporalcv.compute_hac_variance().- DM test¶
Diebold-Mariano test for comparing predictive accuracy of two forecasts. Tests null hypothesis of equal predictive accuracy. See
temporalcv.dm_test(). T1- PT test¶
Pesaran-Timmermann test for directional forecast accuracy. Tests whether directional predictions are better than random. See
temporalcv.pt_test(). T1- MC-SS¶
Move-Conditional Skill Score. Measures forecast skill only during periods when the target variable moves beyond a threshold. Essential for high-persistence series where persistence baselines inflate apparent skill. See
temporalcv.compute_move_conditional_metrics(). T2
Cross-Validation Terms¶
- walk-forward CV¶
Time-series cross-validation where training data always precedes test data. Can use expanding (all prior data) or sliding (fixed window) approach. See
temporalcv.WalkForwardCV.- cross-fitting¶
Double ML-style K-fold where each observation receives out-of-sample predictions. Eliminates regularization bias in metric estimates. See
temporalcv.CrossFitCV. T1- gap¶
Number of samples skipped between training end and test start. Should be
>= horizonto prevent information leakage in forecasting.- leakage¶
Information from the test period improperly entering the training process. Examples: future values in features, full-series normalization, threshold computation using test data.
Validation Terms¶
- gate¶
A validation check that returns PASS, WARN, HALT, or SKIP status. Gates form the first line of defense against data leakage. See
temporalcv.gates.- HALT¶
Gate status indicating critical failure. Pipeline should stop and investigate before proceeding. Typically indicates data leakage.
- WARN¶
Gate status indicating potential issue. Pipeline may continue but results should be interpreted with caution.
- PASS¶
Gate status indicating check passed. No evidence of problems found.
- SKIP¶
Gate status indicating check was skipped due to insufficient data or inapplicable conditions.
Uncertainty Quantification¶
- conformal prediction¶
Distribution-free prediction intervals with finite-sample coverage guarantees. Uses calibration residuals to quantify uncertainty. See
temporalcv.SplitConformalPredictor. T1- wild bootstrap¶
Bootstrap method for clustered data that preserves within-cluster correlation structure. Uses random weight multipliers. See
temporalcv.wild_cluster_bootstrap(). T1- coverage¶
Proportion of test observations falling within prediction intervals. Target is typically 1 - alpha (e.g., 95% for alpha=0.05).
Regime Terms¶
- volatility regime¶
Classification of periods as HIGH or LOW volatility based on rolling standard deviation relative to median. T3 See
temporalcv.classify_volatility_regime().- direction regime¶
Classification of periods as UP, DOWN, or FLAT based on returns exceeding threshold. T3 See
temporalcv.classify_direction_regime().