The forecasting power of the multi-language narrative of sell-side research: A machine learning evaluation

Abstract

This is probably the first ever analysis of sell-side daily economic research to use Natural Language Processing, and it shows that the narrative of such reports can be used to predict economic time series. The NLP indexes are based on Polish and English language reports released at the same time and exhibit predictive power for different sets of economic variables. VAR models with the NLP indexes generate smaller forecast errors than ARIMA. The wordscores scaling model uses Monetary Policy Council statements to generate scores and allows NLP indexes to be created with better forecasting power than the sentiment-based ones.