Determining Lead-Lag Structure between Sentiment Index and Stock Price Returns

Abstract

This research contrasts and compares the state-of-the-art techniques of the two approaches within the domain of news sentiment analysis, as well as, investigates a novel document encoding representation of the 'TF-IDF momentum matrix'. The presented lexicon-based methodology is centred around Loughran & McDonald financial sentiment word lists and reaches 86.4% explained stock momentum variance, whereas the classification approach follows a thematic analysis pipeline implementing Latent Dirichlet Allocation and achieves that of 94.8%. As an additional element of model evaluation, the research implements Thermal Optimal Path method which relies on a dynamic programming approach for performance optimisation.