Πλοήγηση ανά Συγγραφέα "Bletsogiannis, Nikolaos"
Τώρα δείχνει 1 - 1 από 1
- Αποτελέσματα ανά σελίδα
- Επιλογές ταξινόμησης
Τεκμήριο Improved portfolio optimisation with machine learning techniques(03/01/2021) Bletsogiannis, Nikolaos; Μπλετσογιάννης, Νικόλαος; Athens University of Economics and Business, Department of International and European Economic Studies; Tzavalis, Elias; Arvanitis, Stylianos; Dendramis, YiannisThe standard approach for risky asset allocation is the mean-variance framework (Markowitz, 1952). This straightforward approach, of using the sample estimates the inputs, imposes a great amount of error. In this thesis, we implemented different methodologies in order to reduce the estimation error. In Chapter 2 we analyze approaches for a more accurate covariance matrix estimation. These are, the 1-factor model (Sharpe, 1963), shrinkage estimators (Ledoit & Wolf, 2001) and the sparsity principle (Torri, et al., 2018). Also, a brief reference to multivariate GARCH models has been made (Silvennoinen & Teräsvirta, 2008). In Chapter 3 we analyse machine learning models which are used to estimate the conditional expectation of returns. We start with the simple linear regression and continue with shrinkage methods like (Ridge, Lasso and Elastic Net), factor-based approaches like (PCA and PLS). Also, extensive analysis has been made to random forests and neural networks which provide a very accurate expected return estimation (Gu, et al., 2019). Most of the methodologies described above are implemented in real data by estimating the mean-variance (MV) and minimum variance (GMVP) portfolios. The main findings are that shrinkage and sparse precision matrices generate better out of sample returns than the sample minimum variance portfolio in Large N framework. For the mean-variance portfolio, the random forest model generates promising out of sample results but doesn’t outperform the benchmark 1/N approach. In addition, we generated simulated data which then are used to compare the estimation error the sample inverse covariance estimation has with the sparse estimate. The key result is that sparsity can ameliorate the results and provide a much more accurate estimate. Also, in situations where the number of assets is larger than the time series of returns, thus the sample covariance is singular, the sparce approach ends up in a fairly accurate result.