Seasonal slopes raised six-month Michigan-Huron prediction error by 5.25% at 1970-2024 origins

We tested whether annual changes in response slopes improve six-month lake level predictions beyond monthly intercepts and common slopes. We used coordinated v2025 levels for 1918-2025. Fits used training targets before each test year. At 660 Michigan-Huron origins in 1970-2024, four annual harmonic interactions increased mean squared error by 5.25%. The conditional 95% five-year block interval for the increase was 1.44% to 9.68%. RMSE rose from 14.47 to 14.84 cm. Losses persisted with a trend, a 40-year training window and an own-lake twelve-month change. The secondary Superior test gave a 4.78% increase. This gives a reproducible comparison of local-history models. It does not show that physical responses are constant through the year. Two lakes, revised data, one horizon and conditional uncertainty limit inference. The package includes numeric inputs, code, source hashes, predictions, block draws, an independent audit and a figure.

Great Lakes · seasonal response · lake level prediction · harmonic interactions · chronological validation · negative result