Past rain did not reduce mean scaled error in linear forecasts of seven rodent species in two Portal plots in 2000–2025

We tested whether past rain improves forecasts of rodent captures at Portal, Arizona. We used two plots with no treatments and fixed seven species using data before 2000. We then predicted counts one, three, and six months ahead for observed months in 2000–2025. Each fit used only targets observed by its forecast origin. The baseline used season, time, and the current count. A second model also used rain in three past time windows. We scaled each species error by its own baseline error. The mean error reductions were −0.95%, −2.04%, and −1.91% on the log(1 + count) scale. The 95% intervals from year blocks were [−3.80, +0.28], [−4.95, −0.38], and [−6.05, +0.76]%. Thus added rain gave no mean scaled gain in this test. The desert pocket mouse had a 6.39% gain at six months. This gain was smaller with site rain data and a shorter training window. Pooling unscaled species errors gave a six-month gain of 1.30%. The result depends on species weights. These results apply to the tested linear models and complete observed months at one site. They do not measure the causal effect of rain on populations. The package includes input tables, source checks, code, forecasts, error scores, and figures.

rodent dynamics · rainfall · ecological forecasting · Portal Project · negative result · temporal validation · reproducibility