Lab · ML Experiments

ML — Pattern Discovery

Inverted workflow: find conditional edges in BTC data first, build strategies second.
55 experiments

DVOL/Realised Vol Spread — BTC Variance Risk Premium

Promoted
2026-05-19 volatilityoptionsvrpderibit
Hypothesis
Variance Risk Premium VRP = DVOL_implied − RV_realised predicts forward 4h realized vol with negative Spearman IC ≥ 0.05 in magnitude walk-forward. High VRP → vol mean-reverts down; low VRP → vol spikes up.
Verdict
**PROMOTE** — VRP IC = -0.2772 (negative as predicted), negative in 16/16 windows, ΔRV spread Q5−Q1 = +34.0%. VRP is a robust mean-reversion signal for 4h vol.
pooled_n
1,878
n_windows
16
pooled_CI
[-0.3159, -0.2315]
pooled_IC
-0.2772
windows_neg
16
delta_rv_spread_q5_minus_q1_pct
+33.9909

DVOL/Realised Vol Spread — BTC Variance Risk Premium

2026-05-19 · status: promoted · 1.8s

Hypothesis: Variance Risk Premium VRP = DVOL_implied − RV_realised predicts forward 4h realized vol with negative Spearman IC ≥ 0.05 in magnitude walk-forward. High VRP → vol mean-reverts down; low VRP → vol spikes up.

Verdict: PROMOTE — VRP IC = -0.2772 (negative as predicted), negative in 16/16 windows, ΔRV spread Q5−Q1 = +34.0%. VRP is a robust mean-reversion signal for 4h vol.

Key metrics

metric value
pooled_IC -0.2772
pooled_CI [-0.3159, -0.2315]
pooled_n 1,878
windows_neg 16
n_windows 16
delta_rv_spread_q5_minus_q1_pct +33.9909

Approach

DVOL daily close → annualised IV (= DVOL/100). Forward-fill onto 4h BTC bars. Realised 4h annualised vol = std(1m log returns over 4h) · √525600. VRP = IV − RV. Label: realised 4h ann vol starting at t+1. Sampled every 6 bars (~24h) for IC bootstrap independence. Walk-forward 12mo/3mo, embargo 1440 min.

Expected sign: Spearman IC of VRP → fwd RV should be negative (high VRP = market overpricing vol = realised will come down).

Pooled OOS IC (VRP → fwd 4h RV ann)

  • Spearman IC: -0.2772 (n=1,878)

  • 95% CI: [-0.3159, -0.2315]

  • p-value: 1.8e-34

Quintile table — fwd 4h RV ann by VRP quintile

quantile count mean_pct std_pct
0 376 69.32 42.8
1 375 45.66 22.6
2 376 41.38 22.14
3 375 39.04 22.51
4 376 44.21 33.65

Quintile table — ΔRV (fwd − current) by VRP quintile

quantile count mean_delta_pct
0 376 -26.21
1 375 -4.22
2 376 1.06
3 375 4.61
4 376 7.78

Per-window IC (16 windows)

window n IC ci_low ci_high p_value
2022-03-25 → 2022-06-25 92 -0.5025 -0.6501 -0.3274 0
2022-06-25 → 2022-09-25 92 -0.4865 -0.6423 -0.3051 0
2022-09-25 → 2022-12-25 91 -0.4647 -0.6232 -0.2944 0
2022-12-25 → 2023-03-25 90 -0.5815 -0.7248 -0.3748 0
2023-03-25 → 2023-06-25 92 -0.2251 -0.3933 -0.0474 0.031
2023-06-25 → 2023-09-25 92 -0.4517 -0.6202 -0.26 0
2023-09-25 → 2023-12-25 91 -0.4107 -0.591 -0.2339 0.0001
2023-12-25 → 2024-03-25 91 -0.3492 -0.5303 -0.1672 0.0007
2024-03-25 → 2024-06-25 92 -0.3964 -0.5668 -0.2094 0.0001
2024-06-25 → 2024-09-25 92 -0.4962 -0.6424 -0.3561 0
2024-09-25 → 2024-12-25 91 -0.6388 -0.7512 -0.4708 0
2024-12-25 → 2025-03-25 90 -0.702 -0.7985 -0.571 0
2025-03-25 → 2025-06-25 92 -0.5304 -0.6652 -0.3373 0
2025-06-25 → 2025-09-25 92 -0.4946 -0.657 -0.292 0
2025-09-25 → 2025-12-25 91 -0.5906 -0.7331 -0.3774 0
2025-12-25 → 2026-03-25 90 -0.4768 -0.6369 -0.288 0

per-window IC

IV vs RV