Correlation Engine

Methodology, including why you should stay skeptical

Correlation, never causation. An edge means two change series moved together, possibly at a lag. Government announcements usually respond to events, so even a suggestive lead–lag ordering routinely points backwards, and most co-movement is driven by some third thing.

Changes, not levels. Every series is differenced until it passes a stationarity test, then adjusted for day-of-week cycles. Correlating raw trending levels manufactures fake relationships; this pipeline refuses to do it.

Multiple comparisons. Testing 4,875 hypotheses guarantees chance hits; the strip above says how many. Benjamini–Hochberg correction bounds the expected false-discovery share at 5%. The lags of a single pair are not independent tests, so treat q-values as approximate - which is what the noise panel is for. The calibration number in the strip above makes that comparison daily and quantitative: today's surviving edges as a multiple of what the identical pipeline averages on phase-and-amplitude-matched noise, with every dependence (lags, shared metrics, autocorrelation, ties) priced in empirically. Until it sits well above 1×, the correct level of trust is none.

Common drivers. The likeliest reason two series move together is that both react to the same events. Each published edge is therefore re-tested with changes in “VIX volatility index” conditioned out (a partial Spearman ρ, computed only on days that series trades): “holds” means the edge keeps its size, “fades” means conditioning pushed it below the effect-size floor, and “weekends, not stress” means the edge was already below the floor on trading days, so the drop is about weekend rows, not the market. Verdicts compare the partial ρ against the raw ρ recomputed on the same days, never against the full-sample ρ, so a shrunken sample cannot masquerade as a common driver. In today's noise panel 87% of noise edges “held”, so treat holding as the default outcome, not as evidence. This is context, not a filter; the publication gate ignores it. Nor is it a causal cleanup: market stress is itself driven by news, so conditioning can erase genuinely stress-mediated relationships and can occasionally manufacture association out of nothing (collider bias).

Stability. A pattern that appears once is noise. Edges are published only after recurring, with the same sign, across most of the recent daily runs.

Audit trail. Every fetched observation and every daily result is committed to the repository, so the full history of what this tool saw and claimed is public in the commit log.