Rain history did not improve ridge forecasts of rodent composition on two Portal control plots in 2001–2024

Does known rain history improve forecasts of rodent species shares after recent captures and season are known? We tested 21 species on two permanent control plots at Portal, Arizona. We used 162 usable quarters from 1981–2024 in PortalData 6.117.0. We fitted separate ridge models for each species and selected penalties with past data alone. Forecast targets were in 2001–2024, at horizons of 3, 6, and 12 months. We compared recent captures and season with the same predictors plus four quarters of rain history. Every training response was dated at or before the forecast issue date. The primary loss was squared Hellinger distance for species composition. County rain gave added skill of −4.5%, −1.0%, and −8.4%, on 73, 71, and 71 target quarters. Paired 95% intervals from eight-quarter time blocks were −9.2 to +0.4%, −7.3 to +4.4%, and −17.8 to +0.02%. Positive skill means improvement. The basic models reduced error against a no-change forecast by 34.3%, 42.5%, and 18.8%. Composition skill estimates also stayed below zero with shorter and longer training histories, on-site rain, and a separate rain penalty with a rain-off option. Capture-rate forecasts had wider uncertainty and allowed some useful rain gains. The result is a scoped null test of a small linear model family at one site. It does not show that rain has no ecological effect. The package preserves source data, hashes, a protocol, species predictions, figures, code, and reproduction instructions.

biodiversity · rodent communities · rainfall · ecological forecasting · temporal validation · Hellinger distance · Portal Project