The role of covariates in predicting stock markets

Abstract

This study aims to quantify the role of sentiment as a covariate in influencing stock market regime changes. Based on quarterly snapshots of economic conditions from McKinsey for 2010 to 2020, the study makes use of FinBert, a transformer-based pre-trained model which was fine-tuned on large amounts of financial texts, to extract sentiment. This sentiment distribution is then used as a predictor, along with other commonly used macroeconomic variables, CPI and GDP, to predict market state changes in a logistic regression model. Different models consisting of different lags and interaction terms are experimented with to obtain the best fit for studying the underlying relationship. Findings suggest that, contrary to standard expectations, sentiment has a negative effect on the probability of the market being in a ‘bull’ state. The study also benchmarks FinBert against VADER, a lexicon-based sentiment analysis tool, and the logit model against a random forest classifier, to examine and justify the choice of model. This work contributes to the existing literature by exploring the use of transformer-based pre-trained NLP model over commonly used lexicon-based methods, broader macroeconomy financial text over firm-specific text and the ease of interpretation via a logit model, shining light on possible ways sentiment analysis can be utilised for stock market prediction