Forecasting stock return using a recurrent neural network apply to a financial optimization problem

Abstract

This paper presents a methodological proposal for optimizing financial asset portfolios by incorporating the returns predictions instead of the historical returns to calculate an efficient frontier. We changed the return means methodology to forecast by the return with LSTM neural network. We performed several simulation exercises to evaluate the methodology with real data from the US stock market to examine our portfolio optimization model. To evaluate our results, we compared the mean-variance frontier efficiency with the neural network return model. We selected one optimal portfolio that offered the highest expected return for a defined level of risk and compare both models. We show how the neural network return model has a better performance for different periods of time, outperforming the mean-variance model at the same level.