Twitter and market efficiency in energy markets: Evidence using LDA clustered topic extraction

Abstract

We use an extended sample of tweets relating to energy markets in order to examine and quantify the existence of market efficiency. The tweets are used as a proxy for publicly available information and we examine the degree to which this information determines market movements on the next trading day for nine energy market indices. We mine the topics of increasing and decreasing days using latent Dirichlet allocation and find that the topics of tweets in increasing and decreasing days differ. We validate our approach by feeding the extracted topics into three classifier machines and find that the classifiers provide forecasts on market movements with accuracy 57.83% (39.02%) in bull (bear) markets. Our findings support the presence of semi-strong efficiency, since we find evidence of price movements not reflecting public information, while the asymmetry of forecast accuracy over increasing and decreasing markets suggests a different rate of information propagation across market regimes. Our findings can provide useful input to valuation models linked to market efficiency.