A count model gave no clear flood range score gain over past-error scaling at 83 US gauges in 2022-2025

We tested whether a negative-binomial count model improves annual high tide flood day ranges beyond a published past-error scaling rule. We used a verified NOAA table and 307 station-years at 83 US gauges in 2022-2025. Each test year used only earlier rows to fit one common extra-variance parameter. The mean was the NOAA range midpoint with a 0.5-day floor. The count model reduced the equal-year mean 68%-target interval score by 0.30 day against scaling. Conditional 95% limits were -1.21 to +0.47 days. Scores fell in three of four years. Counts inside the range rose from 59.3% to 72.0%, while mean width rose from 6.81 to 8.48 days. Removing 2023 changed the mean score difference to +0.13 day. The test therefore gave no clear score gain over scaling. The result is a count-model benchmark, not a new distribution or a physical explanation. Four test years, shared coastal errors, fixed possibly biased means, revised observations and inherited threshold checks limit inference. The resample limits hold fits fixed and omit serial dependence. The package includes frozen inputs, source hashes, every forecast, code, an independent scalar audit, sensitivity tests and a static figure.

high tide flooding · count model · negative binomial · interval score · chronological test · null result · coastal planning