Stock Price Prediction Based on Temporal Fusion Transformer

Stock price prediction has been an important financial problem which receives increasing attention in the past decades. Existing literature focusing on stock markets forecasting considers the high volatility of stock prices caused by multiple factors. During past years, a number of deep neural networks have been gaining much attention, achieving more accurate results compared to the traditional linear and non-linear approaches. However, most of these neural networks only take time-series features into consideration, ignoring static metadata potentially affecting stock markets. This paper utilizes the state-of-art Temporal Fusion Transformer (TFT) for stock price prediction compared with Support Vector Regression (SVR) and Long Short-Term Memory (LSTM). The comparison for each model is evaluated based on two metrics: Mean Square Error (MSE) and Symmetric Mean Absolute Percentage Error (SMAPE). The results document that TFT achieves the lowest errors.