| Summary: | 碩士 === 國立清華大學 === 計量財務金融學系 === 106 === Mean-Variance Model, the basic model for portfolio construction, is very sensitive to the return we estimated. To overcome this disadvantage, Black-Litterman Model do not construct portfolio only according to the historical returns. It not only takes underlying returns under market equilibrium, or returns by CAPM under consideration, but also takes “Investor’s Views” under consideration. This helps investors to put their views into the model. In academic, we’ll use other return estimation models to make the views, such as MCMC, machine learning. Notice that the method of constructing views with MCMC were used by a robo-advisor in China, while the method with machine learning are still under research.
In this thesis, I’d like to make use of machine learning methods to get views for Black-Litterman Model. Not only basic machine learning methods, such as Lasso Regression, Random Forest, will be used, but also the method of “stacking” will be used. Moreover, I’ll add the volatility estimated by Stochastic Volatility Model to the factors used by machine learning, hoping it can both give a more accurate estimation, and help us to construct a more efficient portfolio. In this thesis, I’ll make comparisons between Black-Litterman Model with views constructed by different methods and based on different factors, getting empirical results using US ETF as underlying. Items to compare are accuracy of estimation, performance of portfolio during hole periods and in some specific periods.
|
|---|
