ETS Weight Optimization: 938 Combinations Tested
Finding the Optimal ADX / Entropy / Hurst / dH/dt Balance for Grid Trading Signals
The Key Finding
We tested 938 weight combinations for the 4 components of our ETS Trend Strength Indicator against 39 known grid trading outcomes (13 assets × 3 rolling windows). The optimal weights improved our composite prediction score by 74%.
Before vs After
How the component weights changed based on data optimization:
Dominant predictor — nearly doubled
Slightly reduced
Still 2× more valuable than Hurst level
Dramatically reduced — minimal standalone value
The headline finding: ADX (trend strength) alone accounts for 60% of the signal. Whether a market is trending or ranging is overwhelmingly the most important factor for grid trading suitability. The Hurst exponent's rate of change (dH/dt at 10%) is twice as predictive as the level itself (5%).
Performance Comparison
| Metric | Original (3-component) | Previous (4-component) | Optimized |
|---|---|---|---|
| Composite Score | 0.217 | 0.176 | 0.307 |
| Ranking | #90 | #235 | #1 🏆 |
| Weights | 40/35/25/0 | 35/30/20/15 | 60/25/5/10 |
| Signal Accuracy | — | — | 64.1% |
Note: The previous 4-component heuristic weights (35/30/20/15) actually scored lower than the original 3-component weights — confirming that heuristic adjustments can go wrong. Data-driven optimization is essential.
Does the Hurst Derivative Actually Help?
We compared the best 3-component weights (no dH/dt) against the best 4-component weights (with dH/dt):
Yes — dH/dt helps. Moving 10% of Hurst's weight from the level to the rate of change improves the score. It's not whether the market IS mean-reverting — it's whether it's BECOMING more or less mean-reverting. The transition detection confirmed by the Regime Transition Study has measurable value.
Limitations
Small dataset (39 observations)
13 assets × 3 windows. Risk of overfitting. The general pattern (ADX-dominant) is robust but exact percentages may shift with more data.
In-sample optimization
Weights optimized on the same data used for evaluation. True out-of-sample validation requires a 4th window (~3 months away).
Signal thresholds not optimized
The 50%/75% cutoffs for GRID_TRADING/NEUTRAL/TREND_FOLLOWING are fixed. A future study could optimize those alongside weights.
Indicators at one point in time
Calculated at Year 2 start only — no intra-period recalculation in this test.
Conclusions
ADX is king. Trend strength accounts for 60% of grid trading signal quality. Simple, well-established, dominant.
Hurst derivative > Hurst level. The rate of change (10%) is twice as predictive as the absolute value (5%). Transition detection matters more than state classification.
Data beats intuition. Our heuristic weights ranked #235. The optimized weights rank #1. A 74% improvement from letting data decide.
Weights applied to production. TrendStrengthIndicator.ts and calculate-ets-standalone.cjs both updated. Discovery, Analysis, and Trading pages automatically benefit.
Related Research
Regime Transition Detection
The study that discovered dH/dt's value — leading to this optimization
Hopfield V3 Pattern Recognition
Neural network approach to regime detection — complementary research
Rating System Validation
ETS feeds into this — improved ETS means improved ratings
Three Layers Methodology
How ETS fits into the broader CoinRoc methodology
See the Optimized ETS Scores
The data-optimized weights are now live on Discovery and Analysis pages. ETS scores more accurately predict grid trading suitability.