A statistically validated OOA-light GBM framework for crime forecasting

Crime forecasting is a critical research area that enables law enforcement agencies to anticipate trends and implement proactive prevention strategies. Traditional statistical and machine learning models, while useful, often struggle with the challenges of nonlinear, multivariate relationships, high-dimensional socio-economic data, and manual hyperparameter tuning, which lead to issues of overfitting, underfitting, and reduced forecasting reliability. To address these limitations, the present work proposes a novel orangutan optimization–driven light gradient boosting machine (OOA-Light GBM) framework that integrates socio-economic and crime-type time-series data into a unified predictive model. The workflow incorporates robust pre-processing, Light GBM for efficient nonlinear modelling, and the orangutan optimization algorithm (OOA) for automatic hyperparameter tuning through balanced exploration and exploitation. Experimental evaluation demonstrates that OOA-Light GBM significantly outperforms existing approaches, achieving 96.2% accuracy with the lowest root mean square error (RMSE) (70.3), mean absolute deviation (MAD) (60.2), and mean absolute percentage error (MAPE) (4.18%), while paired sample t-tests confirm that forecasted and actual values are statistically indistinguishable. These findings establish OOA-Light GBM as a mathematically robust and reliable framework for crime trend prediction, offering substantial improvements over prior methods and supporting data-driven policy planning.