Seasonal memory in Hudson Bay sea ice: autumn freeze-up predicts spring break-up, 1990–2026

Abstract Hudson Bay is one of the largest seasonally ice-covered seas in the Northern Hemisphere, and the timing of its ice season structures the regional climate, its marine ecosystems and the communities that depend on them. Between 1990 and 2026 the ice-free season lengthened by 6.8 days per decade. Most of that change is conventionally attributed to warming in the season in which it occurs, but here we show that the spring half of the signal cannot be explained that way: April–May air temperature over the bay has not risen, yet the date on which half the bay is open water has advanced by 2.8 days per decade. Using 37 ice seasons of daily 25-km satellite ice concentration, we develop a seasonal-memory model in which the preceding autumn's freeze-up date predicts the following spring's break-up independently of spring temperature. The relationship is strong (β = −0.50 days per day, p < 0.001, R² = 0.59) and survives first- differencing (β = −0.29, p = 0.025), instrumental-variable estimation (β = −0.58) and double machine learning (β = −0.40). The memory coefficient is negative in 89% of 1,428 grid cells and has grown from about −0.15 in 1990–2007 to −0.61 in 2008–2026, consistent with thinning ice. A two-predictor physical model outperforms gradient boosting with eleven predictors out of sample (R² = 0.48 versus 0.36), indicating that the relationship is additive and low-dimensional. Autumn freeze-up accounts for roughly 70% of the observed spring advance. Hudson Bay's ice-free season is therefore lengthening as a coupled two-season process, and autumn ice formation carries information about the following spring at a lead time of six months.