Predicting exchange rates with sentiment indicators:An empirical evaluation using text mining and multilayer perceptrons

Crone, Sven F. and Koeppel, Christian (2014) Predicting exchange rates with sentiment indicators:An empirical evaluation using text mining and multilayer perceptrons. In: 2014 IEEE Conference on Computational Intelligence for Financial Engineering and Economics, CIFEr Proceedings. IEEE/IAFE Conference on Computational Intelligence for Financial Engineering, Proceedings (CIFEr) . Institute of Electrical and Electronics Engineers Inc., GBR, pp. 114-121. ISBN 9781479923809

Abstract

Recent innovations in text mining facilitate the use of novel data for sentiment analysis related to financial markets, and promise new approaches to the field of behavioural finance. Traditionally, text mining has allowed a near-real time analysis of available news feeds. The recent dissemination of web 2.0 has seen a drastic increase of user participation, providing comments on websites, social networks and blogs, creating a novel source of rich and personal sentiment data potentially of value to behavioural finance. This study explores the efficacy of using novel sentiment indicators from MarketPsych, which analyses social media in addition to newsfeeds to quantify various levels of individual's emotions, as a predictor for financial time series returns of the Australian Dollar (AUD) - US Dollar (USD) exchange rate. As one of the first studies evaluating both news and social media sentiment indicators as explanatory variables for linear and nonlinear regression algorithms, our study aims to make an original contribution to behavioural finance, combining technical and behavioural aspects of model building. An empirical out-of-sample evaluation with multiple input structures compares multivariate linear regression models (MLR) with multilayer perceptron (MLP) neural networks for descriptive modelling. The results indicate that sentiment indicators are explanatory for market movements of exchange rate returns, with nonlinear MLPs showing superior accuracy over linear regression models with a directional out-of-sample accuracy of 60.26% using cross validation.