A linear October-December predictor raised mean January-September forecast error in seven California divisions in 1926-2025

We tested whether October-December precipitation improves linear forecasts of January-September precipitation in seven California climate divisions. We used a fixed NOAA nClimDiv snapshot and chronological fits with earlier complete water years only. The test covers 100 water years, 1926-2025, after 30 initial training years. The equal mean of division MSE skills was -1.20%, with a conditional 95% five-year block interval of -2.21% to -0.22%. Measured early precipitation plus the mean remainder reduced full water-year error by 32.11% against the pre-season annual mean. Adding a fitted remainder reduced it by only 31.23%. Thus, known accumulation supplied all of the fitted model full-year gain in this record. Mean early-to-full-year correlation was 0.542, within the range from shuffled remainder years. Mean early-to-remainder correlation was 0.047. Accounting can explain the large full-year correlation under the tested null. This does not prove independence. An after-January forecast gave a positive but uncertain gain, and recent-period primary intervals included zero. One retrospective data product, simple model form, shared regional variation, and conditional intervals limit inference. The study does not measure hydrological recovery or identify a physical cause. The package includes exact source rows, all 4,200 forecasts, uncertainty draws, code, figures, a separate numerical audit, and an exact offline repeat.

California · precipitation · seasonal forecast · water year · known accumulation · forecast baseline · chronological validation · nClimDiv · block bootstrap · reproducibility