Know Today, Know Tomorrow: Ensemble Forecasting of Wildlife Sightings from Temporal Dynamics

Abstract

O_LIForecasting encounters between humans and large carnivores has largely relied on mechanistic models driven by causal factors such as food resources and weather. However, for short-term forecasting these approaches implicitly require unrealistically detailed real-time data on many covariates and an almost complete understanding of the underlying causal pathways. As a result, they offer little practical support for short-term, operational decision-making. C_LIO_LIWe developed a short-term forecasting system that predicts end-of-month cumulative bear sightings from the beginning of each month by exploiting temporal autocorrelation without mechanistic assumptions, using an ensemble of multiple components: (i) sequential estimation of daily sighting rates via a non-stationary Poisson process, (ii) seasonal baselines with ratio-based corrections from previous months, and (iii) rule-based transitions among components as daily sightings accumulate. C_LIO_LIApplied to Asiatic black bear (Ursus thibetanus) sighting records from two Japanese regions differing 18-fold in encounter frequency (maximum monthly counts: 83 vs. 1490) and with contrasting seasonal peaks, the ensemble achieved correlations of [≥]0.8 between predicted and observed month-end totals from day 1, increasing to [≥]0.98 by day 20 and substantially outperforming a null model that assumed no seasonal or interannual variation ({Delta}AIC: 477-652). C_LIO_LIAfter controlling for baseline spatial risk and for the region-wide daily bear-forecast level (temporal risk) provided by our ensemble, we detected strongly localized short-term recurrence in bear encounters: prior sightings increased encounter probability within 500 m for up to 3 days, with rapid decay in space and time. C_LIO_LISynthesis and applications. This observation-based ensemble demonstrates that temporal dynamics alone can approach the practical limits of short-term predictability of wildlife encounter rates, without relying on detailed environmental covariates or extensive new data collection. By quantifying both when (daily risk levels) and where (localized hotspots around recent sightings) encounters are most likely, the system offers wildlife agencies and residents an immediately implementable tool for issuing targeted warnings, adjusting outdoor activities, and reducing human injuries in regions experiencing increasing human-carnivore conflict. C_LI