Forecasting electronic money trends in Indonesia using neural network models: A comparative analysis

Abstract

Forecasting electronic money transaction values is essential for effective financial planning and decision-making in various industries. This study evaluates the performance of three neural network models, which are Extreme Learning Machines (ELM), Multilayer Perceptron (MLP), and Neural Network Auto regression (NNETAR) for forecasting electronic money transaction values in Indonesia. The study fitted each model to electronic money transaction data, incorporating features like series modeling in differences, unilabiate lags, and output weight estimation techniques. The ELM utilized 24 hidden nodes and 20 repetitions, while the MLP used 5 hidden nodes and 20 repetitions, and NNETAR employed a 2-2-1 network architecture with 9 weights. Point forecasts were generated for future transaction values using each model. The results revealed variations in the point forecasts across the three models for each respective month, highlighting the diverse methodologies employed by ELM, MLP, and NNETAR in capturing underlying patterns within the data. For instance, the point forecasts for February 2024 ranged from 182,648 for the ELM to 170,525 for the MLP and 173,468 for NNETAR. Evaluation metrics, including Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Mean Absolute Scaled Error (MASE), were employed to assess the accuracy and reliability of the point forecasts. The results indicated that MLP consistently outperformed ELM and NNETAR across all evaluation metrics.

Keywords: Electronic money, forecasting, neural network, ELM, MLP, NNETAR
MSC2020: 91B84, 68T07, 62M10