โ€บNavigation
Candidate 1 of 5 ยท Tier A ยท A1 Pairwise ยท Gate 0.5 ยท 21.2 sec
GATE 0.5 PASS GATE 3 PARTIAL 2-3ร— TYPE-I INFLATION

๐ŸŒ Aggregated HSIC โ€” kernel independence with bandwidth aggregation

Verdict: PARTIAL alignment. The method detects non-linear dependence as advertised, but its built-in significance machinery over-rejects on financial data by a factor of 2-3. Usable as a secondary check; not a drop-in production replacement.

Hub trail: /CLAUDE.md โ†’ findings โ†’ audits โ†’ audit hub โ†’ aggregated-hsic/

What this method does, in plain English

It maps every data point into a high-dimensional "kernel space" โ€” a coordinate system where curved relationships look more like straight lines โ€” then checks whether the two variables stay correlated in that space. It does this at four different "zoom levels" (kernel bandwidths) and combines the verdicts, so you don't have to commit to one zoom level. The number it outputs is a p-value: small means "these variables really are dependent"; large means "looks like they could be independent."

Headline numbers

Spike compute
21.2 sec
8 pairs ร— N=1000 ร— 4 bandwidths ร— 200 perms
Implementation
~150 LOC
Pure numpy โ€” no packaged FOSS exists
Type-I inflation
2.0-2.7ร—
Empirical 10-13% vs nominal 5%
Pairs flagged
8 / 8
All audit pairs at nominal ฮฑ=0.05

What we actually measured

Minimum p-value across 4 kernel bandwidths, per audit pair. Pairs where Spearman called them "independent" are highlighted.

Pair Prior audit Spearman ฯ HSICAgg p_min Flag at ฮฑ=0.05?
ofi โ†” turnover_imbalanceREDUNDANT-PERFECT+1.0000.0050YES
vwap โ†” closeREDUNDANT-PERFECT+0.9670.0050YES
vwap โ†” openREDUNDANT-PERFECT+0.9930.0050YES
buy_volume โ†” sell_volumeREDUNDANT-HIGH+0.7180.0050YES
ofi โ†” aggression_ratioREDUNDANT-MODERATE+0.7930.0050YES
vwap_close_deviation โ†” price_impactREDUNDANT-MODERATEโˆ’0.0200.0199YES
kyle_lambda_proxy โ†” ofiORTHOGONALโˆ’0.1700.0050YES
kyle_lambda_proxy โ†” buy_volumeORTHOGONAL+0.0490.0050YES

Does it do what the paper says?

Paper claim How we tested Result
HSIC = 0 iff X โŠฅ Y (for universal kernels)Synthetic independent Gaussians at N=500p_min = 0.49 โœ“
Catches non-linear dependence (Y = Xยฒ)Synthetic Y = Xยฒ + noisep_min at floor 0.005 โœ“
Aggregation removes bandwidth burdenPer-bandwidth p-values on pair 6: 0.025, 0.020, 0.060, 0.090Aggregation doing work โœ“
Significance correctly calibrated at ฮฑIAAFT surrogate Type-I rate on pairs 7+810% / 13% vs nominal 5% โœ—

What "Type-I error inflation" means

If a statistical test is "calibrated at ฮฑ=0.05", it should incorrectly call independent things "dependent" only 5% of the time. We measured what it actually does on financial data and found it does that 10-13% of the time โ€” so it's twice as trigger-happy as advertised. The fix is to use a stricter threshold (ฮฑ=0.01 or 0.02) to get a true 5% false-positive rate. The reason this happens: the original papers assumed the data is "independent and identically distributed", but financial returns are bumpy and autocorrelated. That breaks the calibration math.

Does it serve our orthogonality-measurement goal?

As nonlinear detector: YES

Flagged every pair Chatterjee flagged. Conceptually it measures the right thing. Useful for cross-method corroboration.

As production threshold: NO

Raw p-values over-reject by 2-3ร—. Production use needs hand-tightened ฮฑ to maintain intended Type-I rate.

Operational cost concern

~140ร— slower than Chatterjee for comparable verdicts. Panel-wide broadening would be 4 hours vs 41 sec.

Provenance: committed in 5d44f419 ยท "Gate 0.5 spike for Aggregated HSIC + validation methodology in FRAMEWORK"