adaptive-wfo-epoch_skill

Overview

This skill implements adaptive epoch selection for Walk-Forward Optimization (WFO), picking per-fold training epochs by optimizing Walk-Forward Efficiency (WFE). It sweeps log-spaced epoch candidates, computes WFE = OOS_Sharpe / IS_Sharpe with data-driven IS thresholds, and selects Pareto-efficient epochs to reduce overfitting and carry priors between folds. The configuration derives all parameters from the search space and sample sizes rather than magic numbers.

How this skill works

For each WFO fold the skill generates a log-spaced set of epoch candidates and trains models at those epoch counts. It computes IS and OOS Sharpe, applies a signal-to-noise-derived IS_Sharpe threshold, and calculates WFE for valid candidates. The efficient frontier of WFE versus training cost is identified and an optimal epoch is chosen; that epoch can be warm-started or used as a prior for the next fold. All variance, granularity, and annualization factors are computed from first principles and the search bounds.

When to use it

• Selecting per-fold training epochs in temporal WFO pipelines
• Mitigating overfitting by measuring transfer from in-sample to out-of-sample
• Building an efficient frontier of epoch versus performance/cost trade-offs
• Carrying epoch priors or warm-starts across sequential WFO folds
• Evaluating robustness of model training regimes with principled thresholds

Best practices

• Generate epoch candidates with log spacing to cover early rapid-change and late diminishing-return regimes
• Use compute_is_sharpe_threshold(n_samples) to adapt IS thresholds to fold sample size
• Prefer carrying selected epoch as a prior with derived prior/observation variances to avoid hyperparameter churn
• Classify WFE with guideline thresholds (0.30, 0.50, 0.70) but adjust to domain risk tolerance
• Annualize Sharpe correctly via the provided market/time_unit factors before comparing across datasets

Example use cases

• Walk-forward evaluation of trading models where per-fold training length matters
• Comparing training epoch cost versus OOS robustness for resource-constrained pipelines
• Automating epoch selection in CI/CD model training for sequential production updates
• Warm-starting training in adjacent folds using a principled prior variance derived from the search range
• Computing alerting rules when WFE falls below investigate/reject thresholds across folds