Findings · A worked example

A 1969 "regime change" we found and then disproved.

Five statistical methods agreed the EPAC hurricane season at Saladita shifted abruptly in 1969 — June 17 to May 31, a 17.5-day jump. A follow-up test designed to disambiguate said the jump was a satellite-catalog artifact, not climate. The chorus methodology caught its own false positive within two iterations. This page documents both rounds honestly.

HURDAT2 1949–2025 · Pettitt + CUSUM + MK + AIC + permutation · Iter 3 + 4 retest · 9 June 2026
Headline finding — corrected by iteration 4

Pettitt + CUSUM + MK + AIC + permutation flagged a 17.5-day jump at 1969. Re-run on the satellite-era subset only (1978–2025), 4 of 5 methods turn null. The 1969 jump was the pre-satellite catalog catching up with reality, not the atmosphere shifting.

The result

Five methods, one year.

The first day-of-year that any East Pacific tropical cyclone reaches within 1,500 km of Saladita has been tracked yearly since 1949. The mean across the full record is day 156 — early June. Linear regression shows that mean drifting earlier by about 2.45 days per decade. That fit was the existing finding.

Run five different change-point tests and a different shape emerges. The data is fit better by a step than by a slope.

Method
Statistic
p
Pettitt (non-parametric change-point)
K = 667
0.006
Pettitt (permutation null, 5,000 shuffles)
0.002
CUSUM permutation null
max-abs = 268
0.0002
Mann–Kendall trend
z = −2.68
0.007
AIC piecewise vs linear
ΔAIC = −7.68
piecewise preferred

All four p-values reach conventional significance. The independent ΔAIC test prefers a model with one change-point over a continuous linear trend. CUSUM alone clears a Bonferroni correction for 11 cumulative chorus tests run to date.

Before and after

17.5 days earlier.

Before 1969 (n = 21 years)

Mean first arrival

June 17
DOY 168.6
After 1969 (n = 56 years)

Mean first arrival

May 31
DOY 151.1

The two regimes are well-separated. Year-to-year variability is similar before and after; the shift is in the mean. Within each regime, no additional trend reaches significance.

Catalog caveat — load-bearing

1969 sits at the dawn of routine satellite hurricane surveillance. Operational geostationary coverage of the East Pacific basin was building through 1966–1980; pre-1969, ship and aircraft reconnaissance were the primary detection methods. Short-lived early-season tropical depressions that did not strengthen into named storms — and those that formed outside shipping lanes — would be systematically missed.

A jump in detection capability mimics a jump in real phenology. The first recorded arrival shifts earlier as the catalog catches more early-season weak storms, even if real storm timing did not change. The 17.5-day discrete shift cannot be confidently attributed to climate without further controls.

What the data shows: a real signal in the catalog. What we don't yet know: how much of it is the atmosphere and how much is the satellites watching it.

What this means for the prior finding

A regime jump and a slope.

The previously published linear-trend finding — about 2.45 days per decade earlier over 77 years — implies a total shift of around 19 days end-to-end. The piecewise model shows a 17.5-day shift at one moment in 1969. The total magnitudes match almost exactly. The data is compatible with both shapes; AIC prefers the piecewise model by 7.68 units, which is strong but not decisive.

The two views can coexist. A real underlying gradual climate shift, plus a one-time catalog improvement, would produce data that looks like a jump centered exactly when satellite tracking matured. So would either explanation alone.

The satellite-era retest — iteration 4

Same tests, satellite years only.

The same five-method ensemble, run only on 1978–2025 (n=48 years), where geostationary catalog completeness is uniform.

Method
Full 1949–2025
1978–2025 only
Pettitt p (permutation)
0.003
0.73
CUSUM p (permutation)
0.0002
0.70
Mann–Kendall p
0.007
0.92
ΔAIC (piecewise vs linear)
−7.68
−4.76
Methods voting "real"
5 / 5
1 / 5

Four of the five methods turn fully non-significant once the pre-satellite years are removed. AIC still marginally prefers a piecewise fit — but AIC will almost always prefer the extra parameter on noisy data, and the implied "change-point" within the satellite era lands at 1999 with a magnitude of −4 days, well within sampling noise. The hypothesis-testing methods all converge on null.

Conclusion: the 1969 jump was the catalog catching up to a real but gradual shift. The linear-trend finding remains the defensible characterization of how the EPAC season has moved across this coast.

The 1969 jump in the data was the satellites learning to see early-season storms — not the atmosphere learning to send them sooner.
Synthesis · June 2026
Why this page exists

A worked example.

The chorus methodology agrees a pattern is real only when multiple statistical methods cross-check each other. Iteration 3 of the pattern hunt produced 5/5 method agreement on a striking 17.5-day discrete shift. Iteration 4 — a follow-up designed specifically to disambiguate climate from catalog — turned the ensemble verdict to null.

The page stays up. False positives, caught and documented, are part of how science is supposed to look.

Methodology & sources

Data. HURDAT2 East Pacific best-track database (NHC), 1949–2025. First-arrival DOY = earliest day of year on which any tracked tropical cyclone enters 1,500 km of Saladita (17.5897°N, 101.4317°W). See _storm_climatology.js.

Methods. Pettitt 1979 non-parametric change-point statistic with normal-approximation p-value, plus a 5,000-shuffle permutation null. CUSUM (cumulative sum of deviations from the grand mean) with bootstrap null. Mann–Kendall trend test, normal-approximation z. AIC comparison of two-parameter linear regression vs. three-parameter piecewise-constant model (best break selected by minimum SSE). The first four are agreement votes; the fifth is an independent model-selection criterion.

Ensemble verdict. "Real" if ≥3 of 4 voting methods clear p < 0.05 AND ΔAIC favors the structural alternative. All 5 conditions satisfied here.

Reproducibility. python3 scripts/loop_iter3_regime_change.py. Artifact: _findings_epac_regime_change.js. This finding was generated by chorus-stack iteration 3 of the self-paced pattern hunt; iteration 1 and 2 returned null.

Related findings