Neural Network -GARCH-Copula Portfolio Optimization with Multifactor Data
Salonen, Johannes (2024-05-22)
Neural Network -GARCH-Copula Portfolio Optimization with Multifactor Data
(22.05.2024)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe2024060444524
https://urn.fi/URN:NBN:fi-fe2024060444524
Tiivistelmä
Return forecasting and portfolio selection have fascinated financial academics and practitioners alike for a long time. With the wake of artificial neural networks, and as importantly the computational capacity to take advantage of such models, financial academics and practitioners have turned their attention to more and more complex models to better understand and predict the behaviour of financial markets.
In literature, often one, seldom two, and rarely more categories of variables are utilized in return prediction. In this thesis, eight categories of variables are considered in return prediction, as data has become more available and immediate, and as such taking a multifactor approach was hypothesized to improve prediction accuracy. In literature when sophisticated neural network predictions have been considered, portfolio selection has often been simplified to equally weighted or similar approaches. In this thesis an error-GARCH-copula approach was utilized with the multifactor neural network to improve portfolio performance, as risk measures are just as important as returns in portfolio performance.
The combined methodology failed to consistently outperform market portfolios or portfolios based on simple linear regression predictions. The main issue with performance was seen to stem more from lack of informative predictions rather than the performance of the copula. Although as the return prediction relied heavily on dispersion measures, the NN-GARCH-copula portfolios also had worse risk measures than market portfolios. Based on these findings, it is suggested that input selection, more sophisticated architectures, and dynamic informative prediction intervals should be considered when using multifactor NN-copula approach.
In literature, often one, seldom two, and rarely more categories of variables are utilized in return prediction. In this thesis, eight categories of variables are considered in return prediction, as data has become more available and immediate, and as such taking a multifactor approach was hypothesized to improve prediction accuracy. In literature when sophisticated neural network predictions have been considered, portfolio selection has often been simplified to equally weighted or similar approaches. In this thesis an error-GARCH-copula approach was utilized with the multifactor neural network to improve portfolio performance, as risk measures are just as important as returns in portfolio performance.
The combined methodology failed to consistently outperform market portfolios or portfolios based on simple linear regression predictions. The main issue with performance was seen to stem more from lack of informative predictions rather than the performance of the copula. Although as the return prediction relied heavily on dispersion measures, the NN-GARCH-copula portfolios also had worse risk measures than market portfolios. Based on these findings, it is suggested that input selection, more sophisticated architectures, and dynamic informative prediction intervals should be considered when using multifactor NN-copula approach.