A1253
Title: Are transformer models good for stock price forecasting?
Authors: Aditya Maheshwari - Indian Institute of Management Indore (India) [presenting]
Mahima Kumavat - Indian Institute of Management Indore (India)
Akshat Vats - Indian Institute of Management Indore (India)
Abstract: Accurate stock price prediction is one of the most challenging tasks in financial decision-making, due to the inherent volatility, nonlinearity, and noise inherent in financial time series. Traditional statistical models are interpretable but fail to capture complex and nonlinear temporal dependencies and long-range patterns. The aim is to present a comparative analysis of state-of-the-art forecasting models, including ARIMA, Long Short-Term Memory (LSTM), Autoformer, Temporal Fusion Transformer (TFT), Informer, PatchTST, and the recently proposed Neural Hierarchical Interpolation for Time Series (NHITS), for five-day-ahead stock price prediction. Using 14 years of historical data of 20 stocks listed on NSE, each model is trained under identical conditions. These models are also validated and evaluated. This ensures that the comparison is fair. The experimental analysis shows that transformer-based models are better at financial time series forecasting. These models are ahead of classical linear approaches in capturing nonlinear dependencies. The temporal fusion transformer (TFT) model achieves the lowest average root mean square error (RMSE) of just 1.41\%, showing its superior predictive accuracy. Additionally, PatchTST and NHITS exhibit strong robustness and accuracy, highlighting their effectiveness in modeling local and global patterns using patching and hierarchical interpolation strategies.
