Borrowing data from other populations to forecast epidemic size

Forecasting infectious disease dynamics is important for ecological management and emerging disease preparedness, yet many systems lack the long-term datasets required to develop reliable predictions. Spatial replication may provide an alternative by allowing information to be shared among populations, but the extent to which this improves forecasting remains unclear. Here, I used epidemic and temperature data from 20 replicated semi-natural Daphnia-parasite pond populations monitored across four seasons (80 epidemics) to test whether information from multiple populations improves forecasts and whether environmental similarity identifies informative populations. I forecasted disease prevalence, infected host density, and healthy host density using a benchmark model, autoregressive integrated moving average (ARIMA), and time-series regression models. Models were trained using focal-population data alone, mean dynamics across other populations, or temperature-weighted mean dynamics from environmentally similar populations. Epidemics showed strong seasonal structure, with prevalence and infected host density generally peaking during the warmest period of the season. However, no forecasting approach consistently outperformed others across all variables. Temperature improved forecasts of prevalence and infected host density but not healthy host density. Multi-population training improved forecasts in some cases, particularly for ARIMA models, whereas temperature-weighted averaging provided no additional benefit over simple averaging. Regression models consistently produced the most accurate forecasts of healthy host density. These results show that borrowing information among populations can improve forecasts under specific conditions but is not universally beneficial, highlighting the importance of matching forecasting approaches and data sources to the ecological processes being predicted.