Structural Limits of OHLCV-Based Intraday Momentum Signals in MNQ Futures: A Systematic Falsification Study

Battle Count: 60

This paper is directly relevant to quantitative trading as it systematically tests 14 common retail intraday momentum signals (ORB, gap fills, volume spikes, Asia session patterns, event-day trends, mean reversion) on a highly liquid futures instrument (MNQ) with realistic transaction costs. The findings establish an empirical gross edge ceiling of ~1-2 points per trade for single-bar OHLCV pattern signals at 5-minute resolution, consistent with competitive market equilibrium. The paper also demonstrates that regime-detection architectures (GMM-based) with longer hold periods can escape this ceiling, as evidenced by two positive controls. The validation protocol (five simultaneous criteria, walk-forward structure, permutation tests) provides a rigorous template for evaluating any intraday trading signal. The three distinct failure modes identified (friction ceiling, T-statistic insufficiency, year instability) offer actionable diagnostic categories for strategy developers.

The falsification framework itself (walk-forward validation, five-criteria gating, permutation tests) is moderately complex but well-documented. The signal implementations range from simple (ORB, gap fill, volume z-scores) to moderately complex (Kalman filter velocity estimation, GMM regime classification with multiple features, Markov transition probability estimation, Ornstein-Uhlenbeck parameterization). The positive control signals require GMM fitting, rolling z-score computation, Markov transition matrices, and ATR computation on 1-minute bars merged to 5-minute pipelines. The paper explicitly notes that several implementation details (ATR source resolution, GMM fit schedule, StandardScaler passes, contract concatenation method) caused replication failures, indicating non-trivial implementation subtleties. No code repository is provided.

The paper explicitly documents implementation details that caused replication failures in earlier versions: ATR computed on 1-minute bars merged to 5-minute pipeline, GMM fit schedule (static 2022 for RTH vs. per-fold refit for London), continuous contract construction (quarterly concatenation, no roll adjustment), and the second StandardScaler pass for London GMM. Appendix A provides full specifications. However, the underlying NinjaTrader data is not publicly distributed, and the 53+ parameter search combinations for the RTH signal introduce selection bias that cannot be fully corrected. The paper acknowledges these gaps and notes that external researchers encountered ambiguity on several points before the current version.