Original verdict relied on IAAFT null calibration โ now prohibited by project policy. The finding needs to be re-derived from an approved null-calibration method before it can be used for decisions. Numbers below are preserved for the record.
Verdict: STRONG signal โ no, they were not. When we ran Chatterjee on every one of 171 microstructure feature pairs on BTCUSDT and applied BH-FDR multiplicity correction at q=0.10, 100% of pairs are flagged, and 69% (118 of 171) are Spearman-independent (|ฯ| < 0.30) yet Chatterjee-decisive. The 8-pair pattern is the panel's dominant pattern.
Hub trail: audit hub โ chatterjee-broadening/
Why this test mattered
We picked 8 feature pairs to test based on the prior audit's classifications. A skeptic could say: "Of course you found something โ you picked pairs that look interesting." This test eliminates that skepticism by running the same method on all 171 pairs we have access to, with multiple-comparison correction (BH-FDR at q=0.10) so we're not just picking up false positives from running 171 tests in parallel. If only the originally-chosen 8 pairs showed dependence, we'd have a cherry-pick problem. If many pairs show dependence, we've discovered a real panel-wide phenomenon.
| Bucket | Count | % of 171 |
|---|---|---|
| Total pairs | 171 | 100.0% |
| Spearman |ฯ| โฅ 0.95 (production "drop") | 11 | 6.4% |
| Chatterjee ฮพ_max โฅ 0.10 (empirical decisive) | 171 | 100.0% |
| Chatterjee BH-FDR significant at q=0.10 | 171 | 100.0% |
| Spearman "independent" (|ฯ|<0.30) AND Chatterjee-decisive | 118 | 69.0% |
These would be retained by the production Pearson filter (low |ฯ|) but show decisive non-linear dependence under Chatterjee + IAAFT. "Spearman says these are unrelated; Chatterjee says one is mostly a curved function of the other."
| X | Y | Spearman ฯ | Chatterjee ฮพ_max | IAAFT p | BH q |
|---|---|---|---|---|---|
| kyle_lambda_proxy | turnover_imbalance | โ0.192 | 0.748 | 0.0050 | 0.0050 |
| kyle_lambda_proxy | ofi | โ0.192 | 0.748 | 0.0050 | 0.0050 |
| aggression_ratio | price_impact | +0.076 | 0.733 | 0.0050 | 0.0050 |
| aggression_ratio | buy_volume | +0.178 | 0.726 | 0.0050 | 0.0050 |
| aggression_ratio | duration_us | โ0.120 | 0.722 | 0.0050 | 0.0050 |
| low | trade_intensity | โ0.292 | 0.706 | 0.0050 | 0.0050 |
| kyle_lambda_proxy | sell_volume | +0.205 | 0.703 | 0.0050 | 0.0050 |
| buy_volume | ofi | +0.233 | 0.684 | 0.0050 | 0.0050 |
| buy_volume | turnover_imbalance | +0.233 | 0.682 | 0.0050 | 0.0050 |
| duration_us | ofi | โ0.145 | 0.682 | 0.0050 | 0.0050 |
| duration_us | turnover_imbalance | โ0.145 | 0.680 | 0.0050 | 0.0050 |
| close | trade_intensity | โ0.286 | 0.676 | 0.0050 | 0.0050 |
| duration_us | kyle_lambda_proxy | +0.175 | 0.673 | 0.0050 | 0.0050 |
| ofi | price_impact | +0.080 | 0.666 | 0.0050 | 0.0050 |
| kyle_lambda_proxy | price_impact | โ0.134 | 0.664 | 0.0050 | 0.0050 |
| price_impact | turnover_imbalance | +0.080 | 0.664 | 0.0050 | 0.0050 |
| aggregation_density | duration_us | โ0.210 | 0.658 | 0.0050 | 0.0050 |
| aggression_ratio | kyle_lambda_proxy | โ0.084 | 0.657 | 0.0050 | 0.0050 |
| duration_us | high | +0.216 | 0.643 | 0.0050 | 0.0050 |
| high | vwap_close_deviation | โ0.123 | 0.640 | 0.0050 | 0.0050 |
Significance is not utility. Many dependencies are statistically significant only because N=1000. Feature selection requires combining dependence + relevance-to-target.
At 100 dbps, BTC bars are dominated by 1-2 trade events. Many features collapse onto the same trade stream. "100% flagged" could partly reflect substrate physics.
36 of 53 production features are INSUFFICIENT-CRYPTO-SPARSE at 100 dbps. Realistic surface is the 17-19 reliably-populated features.
Provenance: committed in f4235b13 ยท "multi-symbol robustness + Chatterjee broadening โ both validate"