Telecommunication Stock Price Prediction Using Smooth Transition Autoregressive

Abstract

Stock price forecasting remains challenging due to nonlinear dynamics and regime-switching behavior caused by market competition, regulatory changes, technological development, and external shocks. This study evaluates the ability of Smooth Transition Autoregressive (STAR) models to capture nonlinear return dynamics. It compares their forecasting performance with the Autoregressive Integrated Moving Average (ARIMA) benchmark for Indonesian telecommunications stocks, namely PT Telkom Indonesia Tbk (TLKM) and PT XL Axiata Tbk (EXCL). Daily closing price data from 2 January 2018 to 19 September 2025 were transformed into stock returns and preprocessed using the Yeo–Johnson transformation to stabilize variance. Model estimation was conducted using in-sample data, while forecasting performance was evaluated on 372 out-sample observations using Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Directional Accuracy (DA). The best STAR specifications obtained were ESTAR(3,2) for TLKM and LSTAR(4,3) for EXCL. The results showed that ESTAR outperformed ARIMA for TLKM, achieving a MAPE of 8.73%, compared with 34.53% for ARIMA, indicating stronger nonlinear regime-switching behavior. In contrast, ARIMA achieved better forecasting accuracy for EXCL with a MAPE of 10.46%, compared with 14.22% for LSTAR, suggesting that linear autoregressive dynamics were sufficient for this series. These findings provide preliminary evidence that the predictive advantage of nonlinear STAR models depends on stock-specific statistical characteristics rather than model complexity alone. The study contributes empirical evidence on nonlinear stock return modeling in the Indonesian telecommunication sector while highlighting the need for broader validation frameworks and exogenous predictors in future forecasting research.