ConformalKelly: Conformal Prediction Intervals as the Scale in Fractional Kelly Position Sizing

Battle Count: 79

Directly addresses position sizing and portfolio construction - core quantitative trading problems. Introduces a novel use of conformal prediction intervals as the volatility denominator in Kelly sizing. Provides extensive empirical evidence on design choices (slow vs. fast adaptation, per-asset vs. pooled calibration, winsorisation). The drawdown dial based on conformal miscoverage is a practical risk management tool. However, the lockbox results show the strategy does not beat passive benchmarks out-of-sample, limiting direct deployability. The design principles (slow conformal quantiles beat adaptive methods for sizing) are broadly applicable to any Kelly-based sizing system.

Core components are relatively simple: ridge regression forecasting, rolling conformal quantile computation, fractional Kelly sizing with caps. However, the full pipeline includes: horizon ensemble with conformal inverse-variance weighting, geometric anchor shrinkage, staggered overlapping books, drawdown dial with pooled miscoverage tracking, and strict causality enforcement. The agentic search methodology adds meta-complexity. The harness enforces specific cost models, leverage caps, and implementation lags. Reproduction requires the private repository with git-tagged configurations.

The paper provides extensive reproducibility details: immutable prepare.py harness, git-tagged configurations, pre-registration document, verbatim lockbox output, and a diagnostics script that replays from git tags. However, the project repository is private (not publicly archived), the pre-registration ordering is verifiable only through committed history available on request, and the data is a frozen Kaggle snapshot. The agentic search over ~200 configurations introduces multiple-comparisons concerns mitigated by the sealed lockbox. A disclosed z-constant inconsistency (1.2816 vs 1.1503) is recorded but not corrected.