The study has examined the performance of six different specifications of Recurrent Neural Networks designed to predict Value at Risk at the one and five percent level. The models have been tested on the OMX30 stock index, the SEK/EUR exchange rate and the Class A Berkshire-Hathaway stock using a GARCH expanding window as baseline model. The proposed Neural Networks show decent predictive performance, serving as an indication of the potential use of Recurrent Neural Networks’ predictive capabilities of VaR. In three cases out of six does a proposed network outperform the baseline GARCH. However, when comparing the proposed models’ performance with the baseline GARCH, it is evident that GARCH on average is more precise and consistent in its predictions. Furthermore, the results show that the Neural Networks’ performance is very sensitive to the hyperparameter tuning, and that finding a model specification that performs well on both in-sample and out-of-sample data is rather difficult, as well as finding a single specification that performs acceptably on several data sets. Given the narrow selection of hyperparameters tuned, the fact that one of the proposed neural network models managed to beat the high performing GARCH in three out of six cases suggests that the subject could benefit from further studies. Future studies are recommended to extend the scope of hyperparameter tuning.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:uu-530654 |
| Date | January 2024 |
| Creators | Friedman, Dan, Matell, Axel |
| Publisher | Uppsala universitet, Statistiska institutionen |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.0021 seconds