Predictability Regime Index

Classifies whether the current market rewards momentum/trend-following or mean-reversion, by fusing four statistical predictability sensors — Variance Ratio, return autocorrelation, a Hurst exponent approximation, and a fractal efficiency ratio — onto one signed −100..+100 axis. It is a context/regime read, not a trade trigger.

Features

Variance Ratio across three horizons (fast/medium/slow), VR > 1 = momentum tendency, VR < 1 = mean-reversion tendency
Return autocorrelation across the same three windows (positive = momentum, negative = reversion)
Hurst exponent approximation via range/stdev scaling (H > 0.5 = persistent, H < 0.5 = anti-persistent)
Fractal efficiency ratio (directional path vs. noisy path)
Composite −100..+100 Predictability Regime Index (PRI) with EMA smoothing and a signal line
Confidence read combining distance from the random-walk centerline with sensor agreement
Momentum / Reversion / Noise / Mixed regime classification with background zones
Regime-shift markers on threshold crosses, alerts, and a light-theme dashboard

Scoring

Each sensor produces a −1..+1 score (positive = momentum-favoring, negative = reversion-favoring):

Sensor	Weight	Reads
Variance Ratio	0.38	ratio of k-period to 1-period return variance vs. the random-walk expectation
Autocorrelation	0.24	correlation of consecutive log returns
Hurst approximation	0.23	rescaled-range proxy for path persistence
Fractal efficiency	0.15	net displacement vs. total path length

The weighted sum is scaled to −100..+100 and EMA-smoothed into the PRI; a further EMA of the PRI produces the Signal line.

Regime classification

Momentum: PRI > 25 and Confidence > 35
Reversion: PRI < -25 and Confidence > 35
Noise: |PRI| ≤ 20 or Confidence < 25
Mixed: everything else

Confidence = |PRI| / 65 × agreement × 100, where agreement falls as the Variance Ratio, autocorrelation, and Hurst scores disagree with each other — a high PRI reading with sensors in conflict is down-weighted.

Edit this page on GitHub