Prediction-Based Portfolio Optimization Model using Neural Networks Fabio D. Freitas a,∗ , Alberto F. De Souza b , Ailson R. de Almeida a Secretaria da Receita Federal do Brasil – RFB, Programa de Pós Graduação em Engenharia Elétrica – UFES, Pietrangelo de Biase, 56 sala 308, 29.010-190 – Vitoria ES, Brazil. b Programa de Pós Graduação em Informática – UFES, Av. Fernando Ferrari, s/n, 29075-910 – Vitória, ES, Brazil. Abstract This work presents a new prediction-based portfolio optimization model that can capture short-term invest- ment opportunities. We used neural network predictors to predict stocks’ returns and derived a risk measure, based on the prediction errors, that have the same statistical foundation of the mean-variance model. The efficient diversification effects holds thanks to the selection of predictors with low and complementary pairwise error profiles. We employed a large set of experiments with real data from the Brazilian stock market to examine our portfolio optimization model, which included the evaluation of the Normality of the prediction errors. Our results showed that it is possible to obtain Normal prediction errors with non-Normal time series of stock returns, and that the prediction-based portfolio optimization model took advantage of short term opportunities, outperforming the mean-variance model and beating the market index. Key words: Neural Networks; Time Series Prediction; Portfolio Optimization. 1. Introduction Investment selection is a central problem in financial theory and practice and it is primarily concerned with the future performance of investments, mainly their expected returns. When investments are exposed to uncertainties, the investment selection framework must include a quantitative measure of the uncertainty of obtaining the expected return, i.e. a quantitative measure of risk. ∗ Corresponding author. Tel/Fax +552732228201. Email addresses: freitas@computer.org (Fabio D. Freitas), alberto@lcad.inf.ufes.br (Alberto F. De Souza), ailson@ele.ufes.br (Ailson R. de Almeida). Preprint submitted to Neurocomputing 8 June 2008