Predicting Stock Prices using Transformers

Abstract

This thesis explores how a Transformer-based architecture can enhance stock market prediction accuracy, using Mean Squared Error (MSE) as the measure. It also investigates if focusing on the last 15 minutes of trading data provides a strong basis for predictions. In the first part, we compare the MSE of the Transformer-based model with that of a Long Short-Term Memory network (LSTM), and a Feedforward Neural Network (FNN). This comparison aims to show how well each model captures short-term market trends at the end of the trading day to predict the opening price for the next day. We evaluate each model based on their MSE values. In the second part, we apply the Transformer model and LSTM network to real stock market data to see if they can generate more profit than passive investing. Passive investing involves a buy-and-hold strategy with minimal market trading. We use back testing to compare the performance of each model in generating profit.