Annual change gave no clear next-year level forecast gain at two Great Lakes for 1960-2024 issues

We tested whether the latest annual mean level change improves next-year mean level forecasts for Superior and Michigan-Huron. We used monthly coordinated levels, version v2025, for 1918-2025. Each of 65 issue years used only completed earlier targets. Adding annual change to a current-level model changed mean squared error by -0.004% skill for Superior and 11.68% skill for Michigan-Huron. Conditional 95% five-year block intervals were -9.37% to 8.38% and -9.74% to 27.53%. Thus neither lake gave clear added gain. In an exploratory check, adding December level to the current annual level cut error by 40.08% and 59.84%. After that control, annual change increased error by 1.81% and 4.48%, with intervals across zero. This suggests that the state lost by annual averaging can affect apparent change-based forecast value. It does not identify a physical cause. Revised data, assumed year-end availability, short records, simple models and conditional intervals limit the result. Code, aggregate scores, source hashes, a numerical audit and a figure are included. Source records and individual forecasts can be regenerated outside the package.

Great Lakes · annual forecasts · temporal aggregation · lake levels · chronological validation · null result