Superior levels cut six-month Michigan-Huron forecast error by 7.0% for issue months in 1960-2024

We tested whether the observed level of Lake Superior improves six-month forecasts of Michigan-Huron beyond its own level and recent change. We used coordinated monthly means for 1918-2025, version v2025. Each of 780 issue months in 1960-2024 used only data through that month. Adding Superior level reduced mean squared error by 7.02%, with a conditional 95% five-year block interval of 2.43% to 11.01%. RMSE fell from 0.14473 to 0.13956 m. Adding Michigan-Huron level to a Superior forecast increased error by 2.53%. The main Michigan-Huron gain was positive at three and twelve months, in both test halves, and after removal of target years 2013-2020. An added own-lake twelve-month change retained a 6.50% gain. A rolling forty-year fit gave a 7.26% point gain but an interval that included zero. These results give a reproducible added-information benchmark. They do not identify water transport, a climate cause, or operational forecast skill. Final revised data, simple linear models, dependent errors and uncertainty conditional on fitted forecasts limit the result. The package gives the input panel, exact source hashes, all forecasts, bootstrap draws, code, a figure and a separate numerical audit.

Great Lakes · water level · seasonal forecast · chronological validation · Lake Superior · Michigan-Huron