| Abstract [eng] |
This project compares the performance of machine learning models such as LSTM and GRU with classical forecasting methods such as ARIMA and ARDL in predicting the OMX Vilnius return index. The project includes a comprehensive literature review on the structure, calculation and use of the OMX Vilnius return index in various academic sources. An extensive review of deep machine learning and classical forecasting methods. The most common models in recent studies are reviewed, their basic principles explained, and the parameters used are described. The study uses historical daily data for the OMX Vilnius return index 2004-2024 and 41 other macroeconomic variables. Granger causality and Pearson correlation analysis are used to select the most significant variables. Various parameters of the machine learning models are evaluated by means of grid search, from which the best ones are selected. The Naive, LSTM, GRU, ARIMA and ARDL models are applied to daily and monthly, univariate and multivariate data in order to assess the full performance of these models. A total of 12 different models are fitted, after which the accuracy of each one is evaluated and compared to the results estimated by the naive method. Finally, once the most accurate model has been found, it is used as the basis for a actual forecast for the next six months (30/04/2024 - 30/09/2024). In summary, this project provides a comprehensive analysis of the OMX Vilnius gross index, as well as of the various machine learning and classical models whose comparative advantage is evaluated in forecasting the OMX Vilnius gross index. |
