A Predictive Benchmark for Animal Movement Forecasting

Analyzing animal movement data typically involves modeling how movement decisions relate to environmental conditions, for example by fitting step-selection functions using conditional logistic regression. An unresolved question is how well these models forecast future movement beyond the data on which they are trained. This issue is fundamental for understanding movement across spatial and temporal scales and is also highly relevant for conservation planning, reserve design, and mitigation of human–wildlife conflict. Recent work suggests that fitted step-selection functions can be used to forecast multi-scale movement through simulation, but empirical evaluation of these approaches remains limited. Further, to date there are no established metrics or evaluation procedures for evaluating the forecasting ability of animal movement models at multiple scales. 

To address these limitations, we introduce a new large-scale benchmark for evaluating the forecasting performance of animal movement models. The benchmark integrates three core components. First, it includes a large, curated collection of public movement datasets spanning diverse taxa, geographic regions, spatial and temporal resolutions, and movement syndromes. Second, each dataset is paired with environmental covariates that are broadly predictive across taxa and structured as geospatial rasters to support simulation from fitted models. Third, the benchmark provides a comprehensive suite of metrics for quantifying forecasting accuracy and uncertainty across multiple spatial and temporal scales. Together, these elements enable standardized, reproducible, and comparative assessment of movement forecasting algorithms.

Using this benchmark, we present empirical results across a range of traditional step-selection approaches as well as emerging machine-learning-based algorithms for movement forecasting. Performance is evaluated consistently across species, regions, and scales, allowing direct comparison of methods that have previously been assessed in isolation. The results highlight both the strengths and the limitations of current modeling strategies and reveal substantial variation in predictive skill across contexts. By providing shared data, covariates, and evaluation tools, the benchmark establishes a foundation for more rigorous and transparent progress in movement forecasting.