Recent level variation gave no clear gain in six-month intervals for two Great Lakes at 1990-2025 origins

We tested whether recent monthly level variation improves six-month 90% intervals for Superior and Michigan-Huron. Both rules used the same local linear forecasts. Annual calibration used only earlier completed forecast errors. Each lake had 426 origins in 1990-2025. Splitting errors at the median of recent 12-month variation gave an equal-lake interval score reduction of -0.95%. The conditional 95% five-year block limits were -2.87% to +1.18%. Thus the primary test gave no clear gain. Total coverage stayed near 91%, while split-rule low-state coverage fell below 90% and high-state coverage rose above 93%. All sensitivity checks had negative point gains. Two lakes, revised data, dependent errors and one split rule limit inference. The test does not identify physical causes or evaluate operational forecasts. Code, numeric inputs, all forecast intervals, resampling draws, an audit, source hashes and a figure are included.

Great Lakes · prediction intervals · water levels · recent variation · chronological test · null result