| dc.description.abstract | Succeeding in accurately forecasting stock market indices is a sought-after capability for investors, traders, and policymakers. However, many financial researchers regard financial assets, including stock market indices, as unforecastable. Others have attempted to challenge these views, traditionally by fundamental or technical analysis of the indices’ historical data. Within the technical approaches, there has been a growing interest in machine learning (ML), a field within computer science. Within the ML methodologies, deep learning (DL), which consists of complex model designs inspired by brain neurons, has lately attracted the most attention. Several studies where DL is used for forecasting stock market index developments present impressive forecasting accuracies, but they seldom benchmark against models based on financial theories such as the random walk and efficient market hypotheses. This study proposes a framework to evaluate ML regression models against these financial theories. As benchmarks, the framework includes two random walk hypothesis-based models: the naive seasonal and naive drift. It is also examined if utilizing economic indicators increases forecasting performance. For experiments, the Oslo Stock Exchange, situated within the small, open, and oil-price-dependent economy of Norway, presents an interesting environment. The Oslo Stock Exchange all-share index is chosen as the target variable, and Norwegian economic data is gathered, resulting in a 72-feature dataset stretching back to 1988 with a daily frequency. Two state-of-the-art DL models are evaluated: long short-term memory networks and the temporal fusion transformer. The models are assessed by forecasting the target variable’s next-day closing valuation and then compared against the benchmark models. From the results, it is found that none of the DL models outcompete the random walk-based models. The most accurate DL implementation, the long short-term memory networks, has a 42,75 % higher error rate than the benchmarks. It is also found that including additional features beyond the target index’s price history in the training data leads to decreased performance. Several factors may cause these results. There may be too much noise or irrelevant information in the additional data, causing the models to overfit. Stock markets are dynamic, and since the dataset stretches back to the 1980s, it may learn the forecasting model patterns that are outdated in today’s market. | eng |