CUSUM event sampling + triple-barrier labeling (pipeline pilot)

2026-06-06 · status: inconclusive · 6.0s

Hypothesis: The CUSUM event sampler + vol-scaled triple-barrier labeler form a causal, leakage-free research pipeline: CUSUM downsamples noisy bars to meaningful directional moves, triple-barrier gives execution-consistent labels, and the +1 base rate is stable walk-forward (no drift/leakage). Infrastructure check, not an edge claim.

Verdict: INFRASTRUCTURE OK — pipeline runs causally end-to-end. CUSUM cut 55,826 bars to 19,944 events (64.3% downsampling); triple-barrier labels split 49/0/51 (+1/0/−1). +1 base rate ranges 0.44–0.53 across 21 walk-forward windows (spread 0.10) — stable, no drift/leakage. Ready to plug contextual features against the triple-barrier label. No edge claimed here.

Key metrics

Approach

BTC 1h bars. CUSUM threshold = 1.0 × trailing realised vol (std of log returns, 24-bar window). Triple barriers at ±1.5/1.5 vol units with a 24-bar vertical limit; barrier touches checked intrabar via high/low (same-bar double touch → stop, pessimistic). Average-uniqueness weights computed from overlapping [t0, t1] spans. All causal.

Sampling

• Bars: 55,826 → CUSUM events: 19,944 (64.3% downsampling)

• Labeled events: 19,944 · median holding to barrier: 2 bars · mean uniqueness: 0.601 (1.0 = no overlap)

Label distribution

• +1 (upper first): 9,789 (49.1%)

• −1 (lower first): 10,154 (50.9%)

• 0 (flat at expiry): 1 (0.0%)

• Barrier hit: up=9,715 · dn=10,083 · time=146

Walk-forward +1 base-rate stability (21 windows)