Method Selection
Primary: Boruta (reference ↗) - effective in DRW, strong conceptual foundation
Alternative: FSA (paper ↗)
Implementation Process
| Step |
Description |
| Input |
Feature functions + selected coin names (default: all) |
| Validation |
Test feature relevance for 96 steps (24h of 15m intervals) |
Boruta Feature Selection
| Component |
Description |
| Method |
Random Forest compares real vs permuted “shadow” features |
| Output |
Features: Confirmed / Tentative / Rejected |
| Data |
60% temporal subsets (multiple stability checks) |
| Minimum |
≥100 clean samples required |
MI Permutation Tests
| Step |
Description |
| Baseline |
Compute MI between feature and target |
| Resampling |
Block bootstrap with adaptive blocks (min 16 periods = 4h) |
| Selection |
Keep features with p-value ≤ 0.05 |
Jaccard Stability Analysis
| Step |
Description |
| 1. Subsample |
10 different 60% temporal subsets |
| 2. Pipeline |
Run Boruta → MI permutation per subset |
| 3. Metric |
Jaccard similarity (intersection/union) across subsets |
| 4. Final |
Retain features appearing in ≥50% of runs |