Abstract

Urban crime imposes substantial social and economic costs and motivates forecasting tools that can support proactive allocation of limited public sector resources. Yet in high-stakes governance settings, predictive performance alone is insufficient: decision makers also require transparency regarding why a model issues a particular warning. This study develops an interpretable, event-level crime forecasting model for Vilnius by adapting the Granger Network framework (Rotaru et al., 2022) to predict crime risk one week ahead at fine spatial resolution.

Using police-recorded violent and property crime events from 2019-2024, Vilnius is discretized into 300x300 m spatial tiles and modeled as a spatiotemporal system in which activity in one location can predict future risk in another. The Granger Network learns directed dependencies between tiles over time, producing weekly forecasts while remaining interpretable by design: the learned network structure can be inspected to understand predictive relationships across the city.

Model performance is evaluated on held-out spatial tiles (16% of the dataset), emphasizing generalization to locations not used for training. After hyperparameter tuning, predictive performance improves substantially over the baseline and becomes more consistent across tiles. On held-out tiles, mean AUC reaches 75.5% for property crime (a 9 percentage-point improvement) and 79.3% for violent crime (a 19 percentage-point improvement). The proportion of tiles with strong performance (AUC > 90%) is high, while apparent overfitting (AUC = 100%) is rare.

Overall, the results indicate that interpretable Granger Network forecasting can deliver operationally useful weekly crime risk prediction for Vilnius and provides a practical template for accountable public-sector analytics that balances performance with transparency.