A machine learning algorithm for stock picking built on information based outliers

Abstract

Stock picking based on regularities in time series is one of the most studied topics in the financial industry. Various machine learning techniques have been employed for this task. We build a trading strategy algorithm that receives as input indicators defined through outliers in the time series of stocks (return, volume, volatility, bid-ask spread). The procedure identifies the most relevant subset of indicators for the prediction of stock returns by combining an heuristic search strategy, guided from the Information Gain Criterium, with the Naive-Bayes classification algorithm. We apply the methodology to two sets of stocks belonging respectively to the EUROSTOXX50 and the DOW JONES index. Performance is smoother than in the Buy&Hold strategy and yields a higher risk-adjusted return, in particular in a turbulent period. However, outperformance vanishes when transaction costs (5–15 basis points) are inserted. Asset return and return/volume serial correlation turn out to be the most relevant indicators to build the trading algorithm.