Reason: IAAFT generates phase-randomised surrogate series. The project policy is to use only real ClickHouse-sourced BTCUSDT data for all empirical analysis. Even though IAAFT inputs real data, the null calibration step relies on synthetic shuffled versions, which is not approved for this research.
Downstream impact: Spike 2 (Chatterjee ฮพโ STRONG verdict) and Spikes 4 / 5 / 6 (multi-symbol, broadening, bigblack-replication) all relied on IAAFT calibration for their null hypothesis. Without IAAFT, those verdicts have no statistical floor separating real signal from autocorrelation artifact, and the "11.8ร null max" + "REPLICATED" + "STRONG panel-wide" claims need to be downgraded or re-derived from a different null calibration approach.
The content below is preserved for the record โ it documents what was attempted and what the technique looks like โ but no decision should be made on the basis of these results.
Verdict: STRONG alignment. IAAFT is the standard-of-evidence enabler โ not a measurement metric itself. It generates synthetic data with the same bumpy autocorrelation as the original but no real cross-relationship to a partner column. That lets us prove a finding is REAL (signal exceeds the autocorrelation null) versus ARTIFACT (signal could be explained by autocorrelation alone).
Hub trail: audit hub โ iaaft-calibration/
What IAAFT does, in plain English
Imagine your data is a song. IAAFT generates a "new song" that has the same tempo and rhythm (linear autocorrelation) and the same note frequencies (amplitude distribution) as the original โ but the melody is scrambled. If you then compare your real data against many such surrogate "remixes" and find that the real-data dependence is way stronger than anything the surrogates can produce, you have evidence the dependence is genuine and not just a side-effect of the music's rhythm. Originally invented in physics (1996) for EEG signals, now widely used in finance for exactly this purpose.
The risk we were resolving
After Chatterjee and HSICAgg both flagged pair 8 as "highly dependent", we had two interpretations: (A) Both metrics correctly catching something Spearman is blind to. (B) Both metrics fooled by the fact that financial returns are autocorrelated within a window โ a shared blind spot. The "two metrics agree" argument is worthless if they could both be fooled by the same thing. IAAFT was the only way to tell A from B.
| Pair | Prior | Observed ฮพ | Null median | Null max | Ratio obs/null-max | Verdict |
|---|---|---|---|---|---|---|
| ofi โ turnover_imbalance | PERFECT | 0.997 | 0.001 | 0.048 | 20.8ร | REAL |
| vwap โ close | PERFECT | 0.900 | 0.001 | 0.072 | 12.5ร | REAL |
| vwap โ open | PERFECT | 0.950 | 0.002 | 0.055 | 17.3ร | REAL |
| buy_volume โ sell_volume | HIGH | 0.727 | โ0.000 | 0.053 | 13.7ร | REAL |
| ofi โ aggression_ratio | MODERATE | 0.761 | 0.000 | 0.059 | 12.9ร | REAL |
| vwap_close_deviation โ price_impact | MODERATE | 0.544 | 0.000 | 0.053 | 10.3ร | REAL |
| kyle_lambda_proxy โ ofi | ORTHOGONAL | 0.716 | 0.001 | 0.074 | 9.7ร | REAL |
| kyle_lambda_proxy โ buy_volume | ORTHOGONAL | 0.613 | โ0.002 | 0.052 | 11.8ร | REAL |
We also ran IAAFT calibration on HSICAgg for the 2 orthogonal-prior pairs. Result was unexpected:
| Pair | Empirical Type-I | Nominal ฮฑ | Inflation |
|---|---|---|---|
| kyle_lambda_proxy โ ofi | 10.0% | 5.0% | 2.0ร |
| kyle_lambda_proxy โ buy_volume | 13.3% | 5.0% | 2.7ร |
IAAFT is not itself an orthogonality metric. It's the calibration tool that converts "this metric says ฮพ = 0.613" into "this metric says ฮพ = 0.613 AND that's 12ร stronger than autocorrelation alone could produce, so it's real signal."
Every future metric candidate will require IAAFT calibration before its verdicts can be admissible. Now part of the validation methodology (METRIC-EVALUATION-PROTOCOL.md ยง9).
Provenance: committed in ddb47596 ยท "IAAFT surrogate calibration validates pair-8 finding as REAL"