← Moments Separation Team 1 Execution

Method

GMM on 6 months of data for 50 coins using:

  • Rolling std
  • Volatility ratio (std/mean)

GMM provides regime labels → validate with silhouette score → extract logic via tree classifier.

Data Split

Period Usage
2024-01 → 2024-06 Training
2024-06 → 2024-11 Testing

Sparse coins (e.g., TAOUSDT from 2024-03) use available data. Minimum: 7000 rows for train period.

Experiment Results

Run 1: Multiple Features

feature_cols = ['returns_std_5', 'returns_std_60', 'volume_std_60',
               'returns_std_15', 'returns_std_30', 'volume_std_5',
               'volume_std_15', 'volume_std_30']
Metric 2 Clusters 3 Clusters
Mean Silhouette 0.479 0.275
Best Silhouette 0.582 0.424
Worst Silhouette 0.383 0.148

Run 2: Single Feature

feature_cols = ['volume_std_30']
Metric Value
Mean Silhouette 0.731
Best Silhouette 0.804
Worst Silhouette 0.642

Strong separation achieved with single feature.

Validation

Tree classifier extracts boundaries. For coins without enough training data, use fallback rule:

mean(hist_std) + 0.5 * std  →  ~75% accuracy

Feature Testing (H1)

Model Spearman Mean Spearman Std
LGBM_Basic 0.0435 0.0171
LGBM_Basic_Regimes 0.0412 0.0194

Next: Separate data by regime in feature selector, run on same coins, compare selected features.