Narrowing of the Daily Temperature Distribution Biases Threshold Day Estimates in the High Arctic: Evidence From Five Svalbard Stations, 1979 to 2025

ABSTRACT Arctic threshold day indices, which govern permafrost, snow cover and engineering design, are commonly obtained by displacing an observed daily temperature distribution by a projected mean change. That is exact only if warming is a pure translation. We test this on daily records from five Svalbard stations, 1979 to 2025. Between 1979‐2000 and the later period the 5th percentile warmed by 5.14°C against 1.22 at the 95th, and the daily standard deviation fell by 16.6%. The narrowing is significant at every station under a moving block bootstrap and insensitive to the split year, so counts obtained by translation are biased, warm day counts by 12.0–41.6 days per year. It is confined to the cold season, 5.12°C in winter against 0.34 in summer, where the interval contains zero at every station; mean warming therefore remains adequate for summer indices and not for winter ones. The two longest records, from 1922 and 1946, place the onset after about 1980. Of six models tested on withheld years, extrapolating the mean rather than freezing it cuts the threshold count error from 18.8 to 14.1 days per year, and linear quantile regression cuts it to 7.8, improving 16 of 20 cases against that control. That model is promising rather than recommended: its mean advantage is not established and it fails at the warmest and southernmost station. Gradient boosting, unable to extrapolate beyond its training range, is no better than the trend extrapolated shift.