Forecasting Inflation Using News Indices

This paper is dedicated to the use of news sentiment indices in machine learning models to forecast inflation on horizons from one to six months. On the one hand, news can shape public expectations and thus public behaviour, while on the other hand, it contains information that is difficult to account for using standard macro variables but can significantly enhance the accuracy of inflation forecasts. For this study, economic news from a nine-year period is collected from the RIA Novosti website, the sentiment of each news item is determined, then all news is divided into nine topic groups using the latent Dirichlet allocation method. The sentiment of each topic is determined based on the probability-weighted sentiment of the news related to that topic. The resulting topic time series are used in machine learning models along with standard macro variables. The most accurate of the models considered in the paper is the long short-term memory (LSTM) model using news indices. The news sentiment indices for the 'Sanctions', 'Gas Sector', and 'Economic Growth' topics make the greatest contribution to the shaping of the inflation forecast with the help of this model, and the major contributors among the standard macro variables are the salary level, production index, and the prices for Brent crude oil futures.