A comparatiue study of machine learning models in stock price prediction

Abstract

This thesis predicts one-day ahead adjusted closing price estimates of four different stocks listed at the Oslo Stock Exchange. The prediction uses a rolling window approach, utilizing 60 days of historical adjusted closing prices as input for each observation. The thesis conducts a comparative analysis, evaluating the prediction performance of five different models: Ordinary Least Squares, Support Vector Regression, Random Forest, Extreme Gradient Boosting, and Long Short-Term Memory networks.

The findings of the comparative analysis indicate that the Long Short-Term Memory (LSTM) network consistently outperforms the other models for all the stocks considered. These findings align with previous research that highlights the effectiveness of LSTM models in stock market prediction. While the thesis does not provide a definitive conclusion on market efficiency or inefficiency, the predictive performance of the LSTM model suggests the presence of potential inefficiencies in the market. Additionally, this thesis identifies key hyperparameters used to avoid overfitting as well as optimizing predictive performance for the LSTM model.

Overall, this research contributes to the existing literature by investigating the effectiveness of the models on a different market, the Norwegian stock market, as well as noting the importance of model specific hyperparameter tuning.