Implementation of Temporal Fusion Transformer (TFT) for Short-Term Sales Prediction of Telkomsel Data Packages in East Java

Abstract

The development of the cellular telecommunications industry has driven an increasing demand for fast, stable, and affordable data services. Accurate forecasting of data package sales is a significant challenge for telecommunications operators due to high demand fluctuations and the complexity of time series patterns. This study aims to implement a Temporal Fusion Transformer (TFT) model based on Seasonal-Trend Decomposition using Loess (STL) to predict short-term sales of Telkomsel data packages in East Java. The data used are sales transactions with hourly time resolution from January to June 2024, focusing on the five data packages with the highest transaction volume. The STL method is applied in the pre-processing stage to separate the trend, seasonal, and residual components, which are then used as additional features in the TFT modeling. Model performance is evaluated using Mean Absolute Error (MAE) and Quantile Risk (q-Risk). The results show that the TFT model is able to produce accurate predictions with an MAE value of 3.6941 and an average q-Risk of 0.0808. Furthermore, interpretability analysis revealed that historical sales variables, seasonal components, and calendar variables significantly contributed to the prediction results. These findings indicate that the STL-based TFT approach is effective for short-term sales forecasting and has the potential to support data-driven operational decision-making in the telecommunications sector.