Superior data gave no clear gain in six-month linear forecasts for Michigan-Huron with 1980-2024 start dates

We tested whether Lake Superior levels improve six-month linear forecasts for Lake Michigan-Huron after use of its own state. We used coordinated monthly water levels, version v2025, and 540 forecast start dates in 1980-2024. Each same-month fit used 50 earlier annual pairs. The own-lake model used the current level and six- and twelve-month changes. Adding the same Superior inputs raised mean squared error by 4.11%. The conditional 95% five-year block interval for error reduction was -19.08% to +9.17%. RMSE rose from 14.75 to 15.05 cm. The reverse forecast had -8.09% skill, with an interval from -18.87% to +0.50%. Planned checks did not resolve a gain for Michigan-Huron. These results apply to the tested linear models. They do not show an absence of a physical link or rule out other forecast methods. Final data versions, 45 test years, overlapping forecasts, omitted water-budget terms, and conditional uncertainty limit inference. The package includes all monthly inputs, exact source hashes, code, all forecasts, scores, a figure, and independent arithmetic checks.

Great Lakes · water level · seasonal forecast · cross-lake prediction · null result · chronological validation