Predicting Accounts Receivable of the Social Security Administration for Employment Using LSTM Algorithm

This study explores the use of Long Short-Term Memory (LSTM) networks for predicting outstanding contributions from employers to the BPJS Ketenagakerjaan, Indonesia’s social security agency. The research aims to address the challenges BPJS faces due to delayed or unpaid contributions, which impact the institution's operational stability and financial health. The LSTM model, a deep learning technique well-suited for time-series prediction, was applied to historical data from BPJS Ketenagakerjaan to predict overdue contributions across three different training-validation splits: 70:30, 80:20, and 90:10. The results demonstrate that the 80:20 split achieved the highest validation accuracy of 84.71%, offering the optimal balance between training data and model generalization. The model's ability to predict overdue contributions with high accuracy could significantly improve BPJS's receivables management, allowing for more proactive financial planning and risk mitigation. The study also highlights the integration of an attention mechanism within the LSTM model, enhancing its predictive capabilities by focusing on the most relevant historical data. This research contributes to the field of predictive analytics in public sector financial management, showcasing the potential of machine learning in enhancing the efficiency and effectiveness of social security programs.