2026-07-062026-07-062026-05-02https://repositorio.ifal.edu.br/handle/123456789/3107Spatiotemporal forecasting of violent property crimes is a relevant problem for the planning of preventive actions and for the allocation of public security resources. This work presents a methodological replication, in the context of the cities of Maceió and Arapiraca, of the fine-grained crime forecasting pipeline proposed by Albors Zumel, Tizzoni, and Campedelli (2025), based on a ConvLSTM architecture. From georeferenced records of violent property crimes provided by the Military Police of Alagoas, covering the period from 2012 to 2022, the work organizes the data into regular grids with cells of approximately 0.2 km2, aggregates occurrences in 12-hour windows, binarizes the cell-level signal, and trains a ConvLSTM with explicit imbalance treatment via weighted BCEWithLogitsLoss. Evaluation is performed by sweeping the classification threshold, reporting precision, recall, and F1 both in the traditional variant (cell-by-cell) and with spatial tolerance (Chebyshev neighborhood ≤ 1). The results show that, in both cities, the model achieves performance in the spatial variant significantly higher than in the traditional variant (spatial F1 of approximately 0.24 in Maceió and 0.21 in Arapiraca, against values below 0.05 in the traditional variant), evidencing that the network learns to correctly locate risk neighborhoods even when missing the exact cell. The work also discusses the main limitations of the approach, particularly the extreme sparsity of the data, and proposes directions for future developments, including the incorporation of sociodemographic and human mobility channels and comparison against baseline models.ptAttribution 3.0 BrazilPredição criminalConvLSTMCrimes contra o patrimônioSegurança pública – MaceióSegurança pública – ArapiracaCrime forecastingDeep learningPublic securityTrabalho de Conclusão de CursoCIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO::SISTEMAS DE INFORMACAO