Cointegration
| ← Ideas | Team 1 Execution |
Hypothesis
Check if coin relationships (beta) change depending on market regime. Current pairs may not be cointegrated in all regimes, but others might be.
Regime Features (per symbol, stationary)
| Feature | Description |
|---|---|
r_t |
15m log-return |
\|r_t\| |
Absolute return |
vol_1d |
Rolling std over 1 day (96 bars) |
vol_1w |
Rolling std over 7 days |
z1d |
Rolling z-score over 1 day |
Normalize with StandardScaler per symbol; clip outliers.
Labeling Process
| Step | Description |
|---|---|
| 1. Fit HMM | Per-symbol GaussianHMM on feature matrix |
| 2. Select states | n_states ∈ {2,3,4} by val loglik/BIC |
| 3. Predict | state_t for each timestamp |
| 4. Build labels | Wide format: SYMBOL_hmm_state columns |
Cointegration Testing
Pass regime labels to test_baskets_cointegration_parallel():
- Filter rows by basket states before Johansen test
- Policy: ‘all’ states or specific include_states