Research map / Time series / Causal and generative time series

SPATIALEPIBENCH: Benchmarking Spatial Information and Epidemic Priors in Forecasting

NeurIPS 2026 · Time series › Causal and generative time series · approach: Robust forecasting and evaluation · 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 SpatialEpiBench, a standardized rolling-evaluation benchmark of 11 epidemic forecasting datasets that tests adjacency-informed spatiotemporal forecasters and four epidemic priors against naive and univariate baselines, finding most underperform last-value persistence.

Problem

Spatiotemporal epidemic forecasting lacks a standardized benchmark with realistic rolling retraining and outbreak-focused evaluation, so it is unclear whether spatial information and epidemic priors actually improve forecasts.

Approach

A hybrid benchmark that applies a unified rolling-origin retraining protocol across 11 epidemic datasets, evaluates 4 model-agnostic epidemic priors (TID, filtered loss, epi-aware regularization with SIR/NGM variants, EINN) as plug-ins to general spatiotemporal forecasters, and uses LRTrend-based outbreak stratification, plus an agentic TusoAI search to test whether geographic adjacency helps.

Key results

• Every benchmarked spatiotemporal method beats the naive persistence baseline less than 50% of the time, often far less.
• Best-patched methods lower RMSE by an average of 9% and raise win rates by 15% relative to unpatched versions, yet most still underperform naive.
• All methods strongly underpredict during outbreaks; during outbreak periods almost no method outperforms naive, with up to 12% higher error for adjacency-informed methods.

Surprising findings

• Most spatiotemporal epidemic forecasters underperform a simple last-value baseline from 1 day to 1 month ahead, even during outbreaks.
• Over 250 attempts at modeling geographic adjacency by TusoAI yielded gains only from non-spatial modeling; removing spatial tricks often did not hurt much.

Limitations

• Only point-forecast metrics reported; probabilistic metrics like CRPS and WIS omitted.
• Focuses on ILI and COVID-19 and only geographic adjacency as spatial input; mobility/richer networks not studied.
• Stronger univariate baselines could be added; only 3 univariate baselines evaluated.

Experimental setup

Datasets: ILINet NCHSdeaths CANpositivity CHNGinpatient CHNGoutpatient CPRadmissions DVcli HHShosp JHUCase CAcase AUcase

Benchmarks: SpatialEpiBench

Metrics: RMSE RMSE filtered win rate MSE MAE medAE medSE

Baselines: naive persistence ARIMA DLinear DCRNN AGCRN STGCN GraphWaveNet MTGNN GTS StemGNN STNorm EpiGNN

Builds on: EpiGNN Cola-GNN EARTH EINN TusoAI LRTrend

Closest papers by meaning

• Aligning Time-series by Local Trends: Applications in Public Health (AAAI 2025 · similarity 0.86)
• Smooth Sailing: Lipschitz-Driven Uncertainty Quantification for Spatial Associations (NeurIPS 2025 · similarity 0.85)
• GST-UNet: A Neural Framework for Spatiotemporal Causal Inference with Time-Varying Confounding (NeurIPS 2025 · similarity 0.84)
• ST-FiT: Inductive Spatial-Temporal Forecasting with Limited Training Data (AAAI 2025 · similarity 0.84)
• A Bayesian Spatial Model to Correct Under-Reporting in Urban Crowdsourcing (AAAI 2024 · similarity 0.84)
• EARTH: Epidemiology-Aware Neural ODE with Continuous Disease Transmission Graph (ICML 2025 · similarity 0.84)