The study of the impact of investor sentiment on stock returns has gained increasing momentum in the past few years. It has been widely accepted that public mood is correlated with financial markets. However, only a few studies discussed how the public mood would affect one of the fundamental problems of computational finance: portfolio management. In this study, we use public financial sentiment and historical prices collected from the New York Stock Exchange (NYSE) to train multiple machine learning models for automatic wealth allocation across a set of assets. Unlike previous studies which set as target variable the asset prices in the portfolio, the variable to predict here is represented by the best asset allocation strategy ex post. Experiments performed on five portfolios show that long short-term memory networks are superior to multilayer perceptron and random forests producing, in the period under analysis, an average increase in the revenue across the portfolios ranging between 5% (without financial mood) and 19% (with financial mood) compared to the equal-weighted portfolio. Results show that our all-in-one and end-to-end approach for automatic portfolio selection outperforms the equalweighted portfolio. Moreover, when using long short-term memory networks, the employment of sentiment data in addition to lagged data leads to greater returns for all the five portfolios under evaluation. Finally, we find that among the employed machine learning algorithms, long short-term memory networks are better suited for learning the impact of public mood on financial time series.

The recent dominance of machine learning-based natural language processing methods has fostered the culture of overemphasizing model accuracies rather than studying the reasons behind their errors. Interpretability, however, is a critical requirement for many downstream AI and NLP applications, e.g., in finance, healthcare, and autonomous driving. This study, instead of proposing any "new model", investigates the error patterns of some widely acknowledged sentiment analysis methods in the finance domain. We discover that (1) those methods belonging to the same clusters are prone to similar error patterns, and (2) there are six types of linguistic features that are pervasive in the common errors. These findings provide important clues and practical considerations for improving sentiment analysis models for financial applications.

No abstract

No abstract

No abstract