American
International
University-
Bangladesh

Students & Supervisors

Abstract

Proper forecasting of crime is a necessity in planning of public safety, resource allocation, and prevention of violence. However, in low- and middle-income countries, the effectiveness of more complex predictive models is often limited by the sparsity of data, delays in reporting, and inconsistent trends in national statistical systems. This paper constructs a detailed forecasting model based on monthly crime data of the Bangladesh Police (2020-2025). We compare classical econometric models (ARIMA), deep learning networks (LSTM, ST-GCN), and machine learning ensembles, and assess their performance using various error measures. The findings indicate that the optimized tree-based ensembles are always better in low data regimes as compared to deep spatio-temporal models. Specifically, the suggested Adaptive Risk-Aware Ensemble (ARAE) dynamically re-weights base learners based on recent forecast errors, risk sensitivity of high-volatility periods. This method obtained the smallest mean absolute error (MAE = 1209), which was 14% smaller than standard ML ensembles, and over 40% smaller than deep learning baselines. The analysis on the importance of features shows the predictive importance of rolling averages and rate-based indicators, which underline the importance of domain-informed feature engineering. In general, the paper is able to present adaptive weighting-based Optimized ML ensembles to crime forecasting in resource-constrained set tings as a champion methodology that integrates both methodological and practical implications of proactive policing and policymaking.

Keywords

Publication Details

• Type of Publication:
• Conference Name: 4th International Conference on Data Analytics and Insights (ICDAI-2026)
• Date of Conference: 30/07/2026 - 30/07/2026
• Venue: Techno International New Town, Block- DG 1/1, Action Area 1, New Town, Kolkata-700156
• Organizer: Techno International New Town