Deep Learning-Based Stock Price Prediction for Norwegian Companies: A Comparative Studytive Study

Abstract

This thesis discusses the applicability of advanced machine learning (specifically GBDT models) and deep learning models in financial forecasting, especially in the setting of stock price prediction on the TITLON dataset. With this representation of the Oslo Stock Exchange data, which has complexities and volatilities intrinsic to financial markets, prediction of the stock prices is quite challenging. That would mean developing and evaluating robust models that are able to capture intricate patterns in data. The research is done at a holistic level: the performance of many models of machine learning is estimated, including Gradient Boosting Decision Trees like XGBoost and CatBoost, alongside some deep learning state-of-the-art architectures like Self-Normalizing Networks, Grownnet, DCNv2, AutoInt, MLP, ResNet, and FT-Transformer. These models will be contrasted for their capacity to reduce the error in predictions; two metrics we use to assess the performance of models are the root mean squared error and the mean absolute error. Another important novelty of this research is the application of a feature extraction with CNN for improving the model toward its precision in accuracy. Such results, before and after applying the CNN feature extraction, are rigorously compared to show that most of the models—deep learning architectures like SNN, DCNv2, AutoInt, and MLP—are substantially improved by this method. These models depicted a large reduction in RMSE and MAE, which goes in line with improved predictive accuracy and generalization ability to unseen data (the test set). Models such as ResNet and FT-Transformer demonstrated smaller improvements after feature extraction, proving that model-specific tuning could be required to some extent. GBDT models had massive performance improvements, particularly on the MAE, hence proving their robustness and competitiveness toward financial forecasting tasks when matched with efficient features. This thesis therefore concludes that although CNN-based feature extraction has the capacity to increase a model's predictive performance significantly, the choice of the appropriate model architecture and its hyperparameter tuning remain a prerequisite for optimal results. The findings enrich the literature on deep learning and machine learning techniques applied to financial forecasting, including valuable insights for future studies and practical implementations within the financial industry.

This thesis discusses the applicability of advanced machine learning (specifically GBDT models) and deep learning models in financial forecasting, especially in the setting of stock price prediction on the TITLON dataset. With this representation of the Oslo Stock Exchange data, which has complexities and volatilities intrinsic to financial markets, prediction of the stock prices is quite challenging. That would mean developing and evaluating robust models that are able to capture intricate patterns in data. The research is done at a holistic level: the performance of many models of machine learning is estimated, including Gradient Boosting Decision Trees like XGBoost and CatBoost, alongside some deep learning state-of-the-art architectures like Self-Normalizing Networks, Grownnet, DCNv2, AutoInt, MLP, ResNet, and FT-Transformer. These models will be contrasted for their capacity to reduce the error in predictions; two metrics we use to assess the performance of models are the root mean squared error and the mean absolute error. Another important novelty of this research is the application of a feature extraction with CNN for improving the model toward its precision in accuracy. Such results, before and after applying the CNN feature extraction, are rigorously compared to show that most of the models—deep learning architectures like SNN, DCNv2, AutoInt, and MLP—are substantially improved by this method. These models depicted a large reduction in RMSE and MAE, which goes in line with improved predictive accuracy and generalization ability to unseen data (the test set). Models such as ResNet and FT-Transformer demonstrated smaller improvements after feature extraction, proving that model-specific tuning could be required to some extent. GBDT models had massive performance improvements, particularly on the MAE, hence proving their robustness and competitiveness toward financial forecasting tasks when matched with efficient features. This thesis therefore concludes that although CNN-based feature extraction has the capacity to increase a model's predictive performance significantly, the choice of the appropriate model architecture and its hyperparameter tuning remain a prerequisite for optimal results. The findings enrich the literature on deep learning and machine learning techniques applied to financial forecasting, including valuable insights for future studies and practical implementations within the financial industry.