Increasing the explanatory power of investor sentiment analysis for commodities in online media

Abstract

Online media are an important source for investor sentiment on commodities. Although there is empirical evidence for a relationship between investor sentiment from news and commodity returns, the impact of classifier design on the explanatory power of sentiment for returns has received little attention. We evaluate the explanatory power of nine classifier designs and find that (1) a positive relationship holds between more opinionated online media sentiment and commodity returns, (2) weighting dictionary terms by machine learning increases explanatory power by up to 25%, and (3) the commonly used dictionary of Loughran and McDonald is detrimental for commodity sentiment analysis.