Volatility compression predicts persistence, not expansion

Verdict: REJECTED · Asset/TF: BTCUSDT 15m / 1h / 4h (+ 5m supplementary) · Sample: 2019-09 → 2026-08 (7y); 5m last 2y only

Hypothesis (preregistered, not Sharpe-mined): after unusually compressed volatility, a large move becomes more likely; direction might appear only when compression is combined with the preceding market state.

Three questions were separated in advance:

1. Does compression predict future magnitude?
2. Does compression + state add directional information?
3. Can that survive costs as a tradeable rule?

Q1 failed. We stopped. No Q2, no Q3, no ETH/SOL, no RLXBT strategy search.

Fundamental Effect

Four compression definitions, causal rolling percentiles (500-bar), compressed = pct < 10, events = start of consecutive compressed runs. Horizons 1/3/6/12/24. Chronological 60/20/20. Large-move threshold frozen on discovery only.

Across 60 validation gate cells (4 families × 3 TFs × 5 horizons), median event mean-ratio of |future log return| (compressed / unconditional) was 0.775. Only 2/60 (3.3%) exceeded 1.10. Median P(large move) lift 0.511. Future RV and range ratios median ~0.78, 0% above 1.10.

On 15m/1h at H=6, all-bar percentile buckets are monotone the wrong way (Spearman +1.0): stronger compression → smaller subsequent absolute returns (0–5% bucket ≈ 0.61–0.69× the 40–60% bucket). That is volatility persistence, not squeeze-then-expand.

5m (last 2y, supplementary): same failure, median ratio 0.776.

Directional Information

Not tested. Preregistered stop: if Q1 fails, do not manufacture a directional story.

Ablation A/B/C/D, continuation vs mean-reversion, and C vs B were never run.

Trading Results

None. No strategy was built.

OOS / Stability / Regimes / Cross-Asset

Final OOS was not used to refine anything; Q1 already failed on validation. No parameter plateau search. No ETH/SOL. No WFA/MC.

Falsification

The narrative “compression → larger expansion” is contradicted by the numbers: compressed states are followed by quieter magnitude, not louder. A lucky single threshold did not appear; the whole family fails.

Look-Ahead Audit

• Rolling RV/ATR/range/BB ratios: past-only windows — YES
• Causal 500-bar percentile — YES
• Large-move decile from discovery only — YES
• Chronological splits, no shuffle — YES
• Full-dataset normalization — NO (not used)

Final Verdict

REJECTED

Compression, as defined here, does not predict subsequent volatility expansion on BTCUSDT. The measured association is persistence: quieter now, quieter next. Because the magnitude premise fails, direction and trading claims were not tested. This is useful negative research: the squeeze narrative should not be the default entry story for this feature family. Reproduce from Binance OHLCV and /workspace/rlxbt-server/data/compression_q1_results.json with the prereg in research-log-compression.md.