Can news headlines be traded?

Abstract

Papers have documented a relation between news and financial market movements. We explore this observation on the firm level from a practical investment perspective with the aid of ML methods. Various implementations of NLP models are trained to produce numerical sentiment value from the news, from which the best model is the state-of-the-art ‘Finbert’ plus a SVM with rbf kernel. The model is used on the news of S&P 500 constituents retrieved from ‘Reuters Newswire’ between the 1st of December 2020 and the 31st of March 2022. Finally, sentiment is aggregated daily to create Long, Short and Long-Short portfolios with 100 and 200 companies. We find that the relationship between sentiment and return is stronger on the same day, with some value being retained the following day. Namely, the Long-Short portfolio achieves the best performance, displaying a significant positive alpha on Fama-French factors and a low r square. However, the profitability of the strategies does not hold when considering transaction costs of 10bp. A further analysis using the EWCT technique to limit turnover shows that some profitability can still be achieved but only for the Long portfolio, which beats the benchmark by a small margin. This fact is solidified by the different sentiment proxies, which also demonstrate the potential for some profitability in the Long-Short portfolio, highlighting news volume as an essential component of the sentiment. We also find that lower turnover limits in the EWCT strategy provide better returns, meaning that sentiment momentum has value.

Date of Award	17 Oct 2022
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Dan Tran (Supervisor)