Forecasting Digital Transaction Values in the Metaverse with a Comparison of ARIMA, Random Forest, and LSTM Models

The growth of the metaverse ecosystem has driven a significant increase in digital transaction activity, creating the need for accurate forecasting models to support business decision-making and regulatory planning. This study aims to compare the performance of three forecasting approaches, namely Long Short-Term Memory (LSTM), Random Forest, and Seasonal Autoregressive Integrated Moving Average (SARIMA), in predicting daily digital transaction values in the metaverse. The dataset comprises more than one hundred thousand transaction entries with temporal and user-behaviour variables. Experimental results show that Random Forest delivers the best performance in tracking fluctuating patterns, producing predictions that closely align with actual values. LSTM excels at capturing long-term trends, although it is less adaptive to extreme spikes, while SARIMA provides stable predictions with 95% confidence intervals that are useful for quantifying uncertainty. These findings indicate that model selection should align with analytical objectives, where Random Forest is better suited for dynamic short-term patterns, LSTM is effective for trend analysis, and SARIMA is valuable when uncertainty estimation is required. This study contributes to the forecasting literature in the digital economy and opens opportunities for hybrid model development to improve the accuracy of transaction predictions in the metaverse.