Stock Price Prediction Using the ETSFormer Model Case Study: PTBA

Abstract

The capital market in Indonesia is currently experiencing very rapid development. This growth is significantly evidenced by the increasing number of investors, especially from the millennial and Gen Z demographics. However, this growing investor base also faces a major challenge: high stock price volatility. These fluctuations are triggered by various factors, ranging from domestic economic policies and global geopolitical conditions to rapidly changing market sentiment. This research aims to build a stock price prediction model for PT Bukit Asam Tbk (PTBA) using the ETSFormer architecture, a modern Transformer-based method designed for time-series data. The historical stock price data used in this study covers a five-year period from 2020 to 2025. To ensure optimal model performance, the best model was identified using the Grid Search technique to find the most effective combination of hyperparameters. The results of this study determined that the best model was achieved with the hyperparameters model dimension = 16, batch size = 16, and a learning rate = 0.01, which yielded a validation loss of 0.0074. In the evaluation phase, this model demonstrated solid performance with a MAPE score of 3.28%, an MAE of 86.76, and an RMSE of 117.2. Although the resulting model is quite good at reading long-term trend directions, observations indicate limitations in capturing short-term price volatility. This implies that the model is more suitable for strategic trend analysis than for predicting daily fluctuations.