Quarter-specific error pools gave no clear gain in six-month lake-level intervals at 1990-2025 origins

We tested whether calendar-quarter error pools improve six-month 90% lake-level prediction intervals for Superior and Michigan-Huron. Both interval rules used the same local linear forecasts. Calibration used only earlier retrospective forecast errors. Each lake had 426 test origins in 1990-2025. Quarter pools gave an equal-lake mean interval score reduction of -0.87%. The conditional 95% five-year block interval was -3.52% to +1.78%. Thus the primary test gave no clear gain. Pooled coverage was 91.31% at Superior and 91.08% at Michigan-Huron. Seasonal total coverage differed by quarter, but smaller seasonal error pools did not clearly improve interval scores. Month-specific and twenty-year calibration checks also had negative point gains. This is a retrospective benchmark, not an operational comparison or a test of physical cause. Two lakes, revised observations, dependent errors and conditional uncertainty limit inference. Code, all numeric inputs, forecasts, intervals, block draws, source hashes, checks and a figure are included.

Great Lakes · prediction intervals · seasonal uncertainty · lake levels · chronological evaluation · null result