Autoregressive integrated moving average-long short-term memory optimized hybrid model for cybercrime forecasting

Abstract

Cybercrime represents a growing global threat with adverse impacts on citizen security, the digital economy, and quality of life. In this context, an optimized hybrid model was developed that combines autoregressive integrated moving average (ARIMA) and long short-term memory (LSTM) for the monthly forecast of cybercrime complaints, applying the cross industry standard process for data mining (CRISP-DM) methodology and applying Python based data science techniques. The model combines the capabilities of the ARIMA statistical approach to capture linear components with the power of LSTM neural networks to address nonlinear temporal relationships. The architecture was trained on a set of 60,378 official records of complaints registered by the National Police of Peru between 2018 and 2023, achieving a mean absolute percentage error (MAPE) of 10.73%, which represents a significant improvement over the singular ARIMA and LSTM predictive models. Compared to previous studies in crime, health, and agriculture, this approach showed a greater ability to generalize over complex time series. It is concluded that the application of the proposed model is a relevant contribution for the police and other security agencies to anticipate crime trends and design preventive and effective strategies to combating cybercrime.