Portfolio selection using feature selection approach

Abstract

In financial portfolio selection problem, there are always trade-offs between re- turn and risk. Security market provides huge amount of historical data with respect to tradings, such as closing price and volume. As a result, risk measurement and criterion to evaluate a portfolio can be diversified according to investors considerations. Inspired of feature selection method in machine learn- ing, we purposed a search framework based on forward selection and backward elimination to optimize portfolio selections, allowing investors to customize their risk measurement and criterion to select a ‘good’ portfolio. Several different criterion for portfolio selection were tested in this research including Sharpe Ratio, C-Sharpe and Information Ratio. The searching model was evaluated in Dow Jones Industrial Average 30 Index, nasdaq 100, shsz300 and S&P 500 with data in 1280 trading days. Compared with traditional Mean-Variance optimization methods, our method reduces the size of the problem from n * n to m * n, where n is number of securities that can be really large, m is the number we want to select within a portfolio, which is typically an integer around 10. According to the experiments, our method showed improved efficiency in large scale problems and gave flexibility on searching criterion, making it possible to test and solve portfolio selection problems under multiple different assessment metrics.