Late-summer ENSO gave no clear dry-winter forecast gain in seven Arizona divisions in 1981-2026

We tested whether July-September Nino 3.4 improved probability forecasts of dry November-March seasons in seven Arizona climate divisions. We used fixed NOAA nClimDiv and ERSSTv5 snapshots. Each forecast used earlier winters only, starting with 30 training winters. The test has 46 winters, 1981-2026, and 322 forecasts. Dry means below the training one-third quantile. A logistic SST model reduced the equal-division mean Brier score by 0.01890, from 0.23114 to 0.21224. Brier skill was 8.18%, with a conditional 95% five-year block interval of -5.81% to 21.52%. Three- and seven-year blocks, quadratic SST, recent training and a lower dry threshold also left the mean gain uncertain. Skill was 24.62% in 1981-2003 and -6.08% in 2004-2026. That split is descriptive and does not prove a change in the climate relation. One division had a positive marginal interval, without a multiple-test correction. The result gives a transparent probability forecast benchmark. It does not show no ENSO effect, measure operational forecast skill, or predict water supply. Revised data, 46 test winters, simple models, changing training thresholds, and uncertainty conditional on fitted forecasts limit inference. The package gives exact source subsets, full source hashes, code, forecasts, all primary block draws, a figure, a separate numerical audit and an offline repeat.

Arizona · ENSO · dry winter · probability forecast · Brier score · chronological test · nClimDiv · ERSSTv5 · null result