Measuring randomness did not predict anything

Can we predict when technical trading strategies will perform best by measuring how much a market deviates from a random walk? I tested this by using formal statistical measures to “gate” or filter my trading activity. The hypothesis was that by identifying periods where price action is non-random, I could selectively enable trend-following strategies. I used two primary metrics: the Variance Ratio (VR), where a value of 1 indicates a random walk, and the Hurst Exponent, where 0.5 indicates a random walk. Unlike previous tests that looked at the “quality” of price bars, this approach analyzed the autocorrelation structure of return series itself.

Testing the predictive power

I first measured these metrics on a 100-day rolling window across four major FX pairs. The metrics themselves were stable, meaning the degree of “non-randomness” persisted over time (Spearman correlation of +0.30 to +0.32). However, they failed to predict future performance. The correlation between these metrics and the subsequent 21-day trend strength was essentially zero (-0.04). In other words, knowing that a market was deviating from a random walk today provided no reliable signal that a trend-following strategy would succeed tomorrow. Interestingly, the median Hurst exponent for daily FX data hovered between 0.43 and 0.47, suggesting a persistent but weak tendency toward mean reversion rather than trending behavior.

Strategy gating results

I applied these metrics as a filter for my existing trend-following strategies using both In-Sample (IS) and Out-of-Sample (OOS) data. I tested eight variations of these gates, including thresholds based on theoretical anchors and historical quantiles.

Cross-sectional analysis

Finally, I looked at whether I could pick the “best” currency pairs by comparing their relative randomness. While I could reliably identify which pairs were more non-random than others (IC +0.27 to +0.29), this did not help in picking future winners. The only signal that showed any statistical significance was an inverse one: high Variance Ratio values predicted a short-term reversal the following day. However, once trading costs were factored in, the strategy’s monthly return dropped from a gross profit to a net return of -0.003% (PF 1.00). In other words, the edge is real, but it is too small to overcome the cost of the spread.

Verdict

This approach is rejected for both time-series filtering and cross-sectional selection. This finding aligns perfectly with my previous tests on regime filters. With this result, I have now completed a comprehensive sweep of the field, from heuristic measures like ADX and efficiency ratios to formal statistical tests like VR and Hurst exponents. None of these methods successfully predicted when technical strategies would function. The tools remain in my codebase for data analysis, but they will not be used to gate live execution.

How this connects

This verification builds on earlier ones (what failed before and what I tried this time, comparisons between approaches).

• The double-efficiency result vanished under robustness checks
• My AI position sizer lost to a shuffled placebo, and six more…
• Hunting regime-proof logic: exactly one candidate passed
• Restricted to gold, the viral method became a real edge