Portfolio selection in non-stationary markets

We consider the task of portfolio selection as a time series prediction problem. At each time-step we obtain prices of a universe of assets and are required to allocate our wealth across them with the goal of maximizing it, based on the historic price returns. We assume these returns are realizations of a general non-stationary stochastic process, and only assume they do not change significantly over short time scales. We follow a statistical learning approach, in which we bound the generalization error of a non-stationary stochastic process, using analogues of uniform laws of large numbers for non-i.i.d. random variables. We use the learning bounds to formulate an optimization algorithm for portfolio selection, and present favorable numerical results with financial data.