Revised 27 September 2026. The June version of this page called its headline number the decade “90th percentile”. It was not. The script took the value at position ⌊0.9 n⌋ of each decade’s sorted values, and with ten or fewer years per decade that position is the last one: the decade maximum. So “120 kt in the 1950s → 185 kt in the 2010s” compares the single strongest storm of 1950–59 (1959) with that of 2010–19 (2015). The page also gave its satellite-era check as “rolling-15-year p90, r = 0.87”; that figure was a rolling standard deviation. And the June analysis ran on an earlier HURDAT2 download; the storm table was rebuilt on 19 September 2026 from hurdat2-nepac-1949-2025-091426.txt, which among small changes moves 2009 from 135 to 150 kt. The script now computes the maximum and a linearly interpolated 90th percentile side by side, runs every test on both, and reads the current table. The finding that the strong end is climbing survives, on both statistics. The claim that the distribution is widening does not: decade standard deviation and interquartile range do not trend.
Mean peak intensity within 500 km rises about +4.8 kt per decade (the phenology finding). The top of the distribution rises further in knots: across decades the maximum climbs about 9.7 kt per decade and the interpolated 90th percentile about 8.5, against 4.7 for the decade mean. Standard deviation and IQR do not trend. The strong end is pulling up; this is not a general spreading.
The top of the distribution climbed.
Each year contributes its peak sustained wind within 500 km of Saladita (75 years; 1977 and 1994 had no storm that close). Group by decade. The maximum is one storm. The 90th percentile is interpolated the standard way (numpy “linear”): with ten values it sits a tenth of the way from the second-strongest year to the strongest, so it is a coarse statistic too — but it is not the maximum.
SD is the population standard deviation. The 2020s cover six seasons. Three decade maxima reach Category 5 (≥137 kt): 2009 (150), 2015 (185), 2023 (145); none before 2009 does. The 185 kt of the 2010s is one storm in one year.
Both statistics, five of five.
* Rolling p-values are below 0.0001 but are not independent tests: each 15-year window shares 14 years with the next, and one strong year sits in up to 15 windows. Read the rolling rows as description. None of the decade-level tests clears the site’s cumulative Bonferroni threshold (α/22 = 0.0023). The maximum’s decade slope (9.7 ± 3.5 kt) and the mean’s (4.7 ± 1.0 kt) differ by less than two standard errors, so “faster than the mean” is a description of the point estimates, not a tested result.
Satellite era only — the top still climbs.
Pre-satellite records under-count weak storms, and earlier iterations of this site’s pattern hunt found that a “regime change” on the full record was a detection artifact. The same retest was applied here: 1978–2025, 47 storm years, 34 rolling 15-year windows.
Inside the satellite era the rolling maximum rises (r = 0.93), the rolling interpolated 90th percentile rises (r = 0.92), and so does the rolling standard deviation (r = 0.86). The June page quoted “r = 0.87” as the rolling p90; it was the rolling standard deviation. The overlap caveat applies with more force on 34 windows: 2009, 2015 and 2023 each sit in many of them. The upper-tail rise is not confined to the pre-satellite years, which is what the check was for. Whether spread is rising is unresolved: the rolling SD says yes, the non-overlapping decade SD and IQR say no.
The strongest storm of each decade near this coast has gained more knots than the average storm has. The mean tells the polite story. The decade’s worst storm tells the dangerous one — and it is one storm a decade.Synthesis · revised September 2026
The “extremes shift, means hold” pattern, read carefully.
The site reads a similar shape in other signals at Saladita. Rainfall: seasonal totals are flat while wet-run clustering rises and the longest wet spell shortens; the hardest-day trend is marginal (p = 0.058). Wave power: the annual-peak trend is positive but not significant (p = 0.27) at the nearest ERA5-Wave offshore node, and energetic days show no trend once a step in the model record around 1991–93 is allowed for. Hurricanes: over 1949–2025 the mean is also rising; what sets the storm result apart is that the decade maximum and 90th percentile gain more knots than the centre, while the spread does not measurably grow.
See the essay for how the three signals fit together after the September revisions.
Methodology & sources
Data. HURDAT2 East Pacific best-track database (NHC), file hurdat2-nepac-1949-2025-091426.txt, via the storm table built 19 September 2026. Per year, the peak sustained wind of any fix within 500 km of Saladita (17.8375°N, 101.7606°W); years with no fix inside 500 km (1977, 1994) are excluded. See _storm_climatology.js.
Methods. Years grouped by decade (8 decades, 2020s partial). Per decade: mean, median, population SD, IQR and 90th percentile (both by linear interpolation, numpy default), and maximum. Pearson (exact t p-value), Mann–Kendall and a 3,000-shuffle permutation null on the decade series of both the maximum and the 90th percentile; the same statistics in 63 rolling 15-year windows (Pearson, Mann–Kendall). Verdict “real” at three of five votes with p < 0.05. With about ten values per decade, any 90th percentile is coarse; that is why both statistics are shown.
Catalog robustness. Rolling-window maximum, 90th percentile and SD recomputed on 1978–2025 only and reported by name.
Reproducibility. python3 scripts/loop_iter8_intensity_distribution.py (revision 2). Artifact: _findings_intensity_widening.js, which records its HURDAT2 file as data_vintage. This finding came from iteration 8 of the site’s self-paced pattern hunt; iterations 1–7 produced one shipped finding (later walked back as a worked example) and five honest nulls. The June estimator error and its correction are described in the revision note above.