VPIN — Volume-synchronised Order-Flow Toxicity

2026-05-19 · status: dropped · 9.0s

Hypothesis: Bulk-volume-classification VPIN at 5-min buckets rank-correlates with |forward 1h return| at IC ≥ 0.05 walk-forward; top-quintile VPIN bars are disproportionately followed by large absolute returns.

Verdict: WEAK/DEAD — pooled IC -0.0063 below the 0.03 bar. BVC-derived VPIN on 1m bars doesn't carry enough taker-side signal to predict forward absolute returns. The real signal likely sits in raw taker buy/sell volume from Binance derivatives (richer than CDF-classified).

Key metrics

Approach

Bulk Volume Classification on 1m BTC bars (z = r / rolling-σ over 60m, p_buy = Φ(z)) → aggregate to 5min buckets → VPIN = rolling mean of |B−S|/V over last 50 buckets (~4h). Feature: VPIN at bucket end. Label: |fwd-60m log return|. Sampled every 12 buckets (~1h) to reduce autocorrelation.

Walk-forward: 12mo train / 3mo test, embargo = 1440 min. Bootstrap 95% CI on Spearman IC.

Pooled OOS

• Pooled Spearman IC: -0.0063 (n=55,819)

• Bootstrap 95% CI: [-0.0148, +0.0017]

• p-value: 0.136

Quintile table — |fwd-1h return| by VPIN quintile

Extreme-event hit rate

• VPIN > Q95 threshold: 0.3046

• |fwd-1h ret| > Q90 threshold: 89.1 bps

• Hits / extreme-VPIN bars: 3043 / 33,492 = 9.1% (baseline 10%)

• Lift over baseline: 0.91×

Per-window IC (21 windows)