Abstrakt
We use machine learning methods in portfolio optimization problems. Most portfolio optimization problems require selection of one or more parameters and we create machine learning model to predict the optimal values of such parameter with respect to out-of-sample performance.
In this paper we use mean-CVaR portfolio optimization model and xgboost machine learning model. Extensive simulations were performed to create the dataset with the optimal choice of the desired parameter.
We explore the dependencies of the optimal choice of minimal in-sample mean on input data, like number of stocks or number of scenarios. Predictor importance and prediction evaluation is presented, showing that the model gives reasonable predictions for parameter that is otherwise very hard to select.