Can We Predict When the Market Changes?
Testing Regime Transition Detection for Smarter Grid Trading
The Hypothesis
Our current indicators (ADX, Entropy, Hurst) correctly classify the current regime. But they're reactive — by the time ADX rises above 25, the grid has already been trading in the wrong conditions.
What if we could detect regime transitions BEFORE they fully happen? We tested this using five approaches — from simple indicator derivatives to Hidden Markov Models.
Six Strategies Tested
| ID | Strategy | Spacing Logic |
|---|---|---|
| A | Static 2.59% | Fixed, no regime awareness (baseline) |
| B | ETS Current | Uses ADX/Hurst/Entropy signal levels (already in CoinRoc) |
| C | Hurst Derivative | Rate of change of Hurst exponent (dH/dt) |
| D | Composite Score | Weighted dH/dt + dADX/dt + ATR breakout + dEntropy/dt |
| E | HMM Regime | 4-state Hidden Markov Model trained on Year 1 |
| F | Wider-Only + Composite | Composite detection but spacing floor = 2.59% |
Results — Best Strategy Distribution
Which strategy was best (by Calmar ratio) across 39 asset-window observations:
Baseline wins half the time
Already in CoinRoc — validates current system
Catches transitions ETS misses
Too complex for the signal it provides
Key Findings
Static 2.59% is genuinely well-calibrated. It won 51% of the time — confirming that our current production spacing is already strong.
ETS Current is the second-best approach (31%) — and it's already in the app. This validates that our ADX + Hurst + Entropy combination adds real value.
Hurst derivative catches transitions ETS misses — leading to its integration into the ETS calculation as a 4th component.
HMM adds complexity without consistent improvement. Simpler approaches outperformed the academic gold standard in this test.
Action Taken
Based on this study, we integrated the Hurst derivative (dH/dt) into the ETS Trend Strength Indicator as a 4th component. Then we optimized all four weights by testing 938 combinations against known outcomes.
See the ETS Weight Optimization Study for the data-driven weight calibration results.
Optimized ETS Weights:
Related Research
ETS Weight Optimization
938 combinations tested — ADX is 60% of the signal
Adaptive Grid v2.1
ATR-based spacing — where regime detection feeds in
Hopfield V3 Pattern Recognition
Exit timing is the most valuable AI feature
Three Layers Methodology
Selection + Optimization + Grid = +30.8pp improvement