Stock Price Prediction and Loss Risk Analysis of PT Sawit Sumbermas Sarana Tbk Using a Hybrid TCN-GAN Model

Abstract

The Crude Palm Oil (CPO) industry is a strategic sector for the Indonesian economy. Yet, stock prices of companies in this sector tend to be highly volatile due to global market dynamics and export policies, increasing investment risk. Conventional models, such as ARIMA, rely on linearity assumptions that limit their ability to capture nonlinear dynamics, while deep learning models, such as RNN, GRU, and LSTM, still suffer from vanishing-gradient problems. Therefore, this study proposes a hybrid Temporal Convolutional Network–Generative Adversarial Network (TCN-GAN) model for stock price prediction and investment risk analysis using the Value-at-Risk (VaR) method with Historical Simulation. The TCN-GAN combines TCN's ability to capture long-term temporal patterns with the adversarial mechanism of GAN to improve prediction accuracy. The data consist of daily closing prices of PT Sawit Sumbermas Sarana Tbk (SSMS.JK) from Yahoo Finance, covering January 1, 2020, to September 30, 2025. A sensitivity analysis on sliding window lengths of 10, 20, and 30 days was conducted to validate model robustness, with window 20 identified as optimal. The TCN-GAN model significantly outperforms the ARIMA baseline, which yielded a MAPE of 18.12% and RMSE of 368.68, by achieving a MAPE of 3.22% and RMSE of 84.23. The model was further used to predict stock prices for the next five periods, yielding an average of IDR 1,647.82. The VaR analysis at a 95% confidence level with a five-day holding period indicates a maximum potential loss of IDR 146,204.