Research map / Time series / Causal and generative time series

MOVEBENCH: A Benchmark for Global-Scale Wildlife Movement Forecasting

NeurIPS 2026 · Time series › Causal and generative time series · approach: Probabilistic and conformal forecasting · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces MoveBench, a large-scale global wildlife GPS movement forecasting benchmark with 160 environmental covariates and an energy-score probabilistic evaluation, comparing four method families.

Problem

Wildlife movement forecasting lacks a standardized benchmark; existing trajectory benchmarks cover map-constrained human/vehicle movement and point-prediction metrics, which are ill-suited to stochastic animal movement influenced by environmental covariates.

Approach

A benchmark pairing 2.6M GPS locations from 800+ individuals across 110 species with 160 environmental raster covariates, evaluated with the multivariate energy score (a proper scoring rule usable with sample-based forecasts from any model, including deterministic ones).

Key results

• Most methods beat simple baselines on the temporal split but not on the individual split; deepSSF is best on temporal split (in-sample individuals) yet worse than CRW on individual split.
• AlphaEarth covariates outperform environmental covariates on average on MoveBench-Mini, and combining both does not help most methods.
• The 'don't move' baseline is the best under expected L2 error, motivating the energy score.

Surprising findings

• Deep learning methods do not consistently outperform a simple linear iSSF.
• Under expected L2 error, always predicting 'don't move' is optimal, showing point metrics reward hedging.

Limitations

• Data sourced from existing published studies, inheriting taxonomic and geographic sampling biases.
• Benchmark not intended for drawing general ecological conclusions without further analysis.
• Raster-based approach introduces significant engineering complexity.

Experimental setup

Datasets: Movebank Data Repository MoveBench (147 studies) ERA5-Land AlphaEarth Foundations embeddings

Benchmarks: MoveBench MoveBench-Mini

Metrics: energy score normalized energy score L2 error

Baselines: iSSF Linear iSSF MLP deepSSF MoveFormer don't move correlated random walk

Builds on: step selection function deepSSF MoveFormer energy score

Closest papers by meaning

• MammAlps: A Multi-view Video Behavior Monitoring Dataset of Wild Mammals in the Swiss Alps (CVPR 2025 · similarity 0.86)
• BuckTales: A multi-UAV dataset for multi-object tracking and re-identification of wild antelopes (NeurIPS 2024 · similarity 0.84)
• Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model (NeurIPS 2024 · similarity 0.83)
• Decision-aware Training of Spatiotemporal Forecasting Models to Select a Top-K Subset of Sites for Intervention (ICML 2025 · similarity 0.82)
• Trajectory World Models for Heterogeneous Environments (ICML 2025 · similarity 0.82)
• Animal-Bench: Benchmarking Multimodal Video Models for Animal-centric Video Understanding (NeurIPS 2024 · similarity 0.82)