Neurala nätverk inom volatilitetsprognoser: En jämförande analys på GARCH och neurala nätverk på svenska, polska och slovenska aktieindex
Year: 2024
Type: dissertation
Abstract: In recent years, the interest in machine learning techniques has exploded with widespread applications in many fields. In this study, we examine the effectiveness of neural networks in the context of financial market volatility forecasting and how they perform relative to econometric volatility forecasting models. To do this, we apply these models on the Swedish index OMXSPI, Polish index WIG and Slovenian index SBITOP. We perform volatility fore- casting using a wide range of GARCH family models and neural networks with an LSTM architecture and predict the forward 14-day volatility of daily returns using data from the 31st of July 2017 to the 31st of July 2023, with the last 342 days used for validation and testing. The in-sample performance and parameter selection of the econometric models were evaluated using Akaike information criterion and Bayesian information criterion metrics and the out-of-sample evaluation was done using root-mean-square percentage error (RMSPE) and root-mean-square error (RMSE). In addition, the GARCH model was implemented us- ing both a sliding and expanding window, where the expanding window was superior in the context of RMSPE and RMSE on the validation data. The experiments revealed that LSTM neural networks trained on trading data were con- sistently more effective in volatility forecasting based on their RMSPE and RMSE, and can therefore serve as a competitive method for volatility forecasting compared to GARCH-family econometric forecasting models.
Author: Linné, Axel
Language: English
Subfield: Finance
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7112461715