Lake Superior terms gave no clear added six-month level prediction skill for Michigan-Huron at 1970-2025 origins

We tested whether Lake Superior levels and recent changes improve six-month predictions of Michigan-Huron monthly levels beyond local history. We used the coordinated v2025 monthly release for 1918-2025. Each origin year used fits with training targets no later than the previous year. At 666 origins from January 1970 to June 2025, adding Superior terms reduced squared error by 5.93%. The conditional 95% five-year block interval was -0.59% to 11.55%. RMSE fell from 14.33 to 13.90 cm. Thus the primary test gave no clear added skill. A trend term gave a positive interval, but a 40-year training window gave only 1.26% point skill. The result depends on model and interval choices. Only two lakes, revised observations, simple linear models and conditional uncertainty limit inference. The test does not identify a physical cause or compare with operational forecasts. The package gives numeric inputs, raw source paths and hashes, predictions, code, saved block draws, numerical checks and a figure.

Great Lakes · Michigan-Huron · Lake Superior · six-month prediction · local persistence · chronological evaluation · null result