A Validated Volatility-Volume-Gap Classifier for Regime Identification in MNQ Intraday Data

Battle Count: 73

Highly relevant to quantitative trading research despite the negative directional results. The paper rigorously demonstrates the distinction between identifying a behavioral phenomenon (which it does successfully) and exploiting it directionally (which it cannot under current constraints). The VVG classifier identifies a real, statistically significant day-type with 77.6% peak-reversal rate and 25.6 bps next-day return spread. The methodology (expanding-window thresholds, walk-forward validation, year-stability gates, friction assumptions) represents best practices for intraday strategy research. The failure analysis (sample size, regime instability, variance-to-edge ratio) provides valuable lessons for any quant developing conditional trading signals. The classifier is preserved as a research asset for potential use in volatility modeling or as a regime filter for other signals.

The classifier itself is straightforward: compute three features (first-30-min return, overnight gap, first-bar volume z-score), compare each against expanding-window 66.7th percentile thresholds, and require all three to be exceeded simultaneously. The OLS regression filter in Strategy 3 adds minimal complexity. The main implementation challenges are: (1) proper expanding-window threshold computation without lookahead bias, (2) handling the 60-day warm-up period, (3) managing the small sample size in strategy testing, and (4) implementing the walk-forward validation protocol with year-stability gates. No ML model training is required.

The classifier construction logic is fully specified in Appendix A with explicit formulas for all three features, threshold computation, and the intersection rule. Dataset parameters (instrument, bar resolution, session definition, friction assumptions) are clearly stated. However, no code repository or data link is provided. The expanding-window methodology and 60-day warm-up are described but the exact data source/vendor is not named. The companion paper (arXiv:2605.04004) shares the same dataset and validation protocol, which aids reproducibility. The N=125 vs N=40 discrepancy in peak reversal analysis (multi-session observations vs unique days) could cause confusion.