How Top Investment Banks Use News Sentiment Data
With the emergence of alternative data, the alpha-landscape is currently getting seriously disrupted. The winners will be the firms that understand how to embrace and take advantage of all of the new opportunities arising from data abundance.
To achieve increased scalability and diversification, the industry trend is for portfolio managers to look at expanding into more global investment mandates. Unfortunately, applying alternative data on a global scale can be challenging because it is not easily available, and often lacks coverage. Furthermore, many datasets do not come with the depth of archive required to perform a proper backtest, or they may lack point-in-time sensitive scoring or tagging. Within the alternative data landscape, news sentiment is a real positive outlier.
To showcase how news sentiment can be used as a source for creating globally scalable and diversified alphas, the Data Science team at RavenPack recently created a simple sentiment signal, applying it on a global scale. The sentiment signals were used as an input for market-neutral strategies at country-, regional-, and global-level.
What were the main conclusions? News sentiment works everywhere with positive daily returns in 41 out of 49 countries, and with Information Ratios (IR) of 3.0 or higher in 3 out of 5 regions, including North America, Europe, and Asia Pacific. To achieve maximum diversification, the sentiment signals were combined into global portfolios leading to IRs as high as 4.8 for a one-day holding period, 2.6 for one week, and 0.8 for one month.
Other firms have also been looking into the use of news sentiment data. Across two reports, Empirical Research Partners tested the efficacy of sentiment applied to developed world markets with a special focus on fundamental investors. As part of their study, they considered eight different news sentiment data providers. Only two providers added meaningful value to their stock selection process, RavenPack being one of them. They argued that alternative data can be valuable to investors, however, a lot of the value lies in how it is being deployed. In particular, they found that media sentiment is different from price momentum, especially over the past month or quarter, which is something they take advantage of when improving the timing of their core multi-factor models. In fact, sentiment was found to be about twice as effective as a signal overlay compared to price momentum.
Interestingly, they also found that sentiment extracted only from news stories about earnings releases has been a better signal than tracking sell-side analysts’ earnings revision. Furthermore, they found, in line with previous studies, that media sentiment is particularly helpful in timing entry into value stocks.
In a recent report focused on CAPEX announcements vs. reported CAPEX, Citigroup's Equity Research team also took a special interest in use-cases for fundamental investors. As a source for announcement data, they considered RavenPack's corporate events that are automatically extracted from textual news and blogs using Natural Language Processing (NLP).
They found that announcement data provided a significant timing advantage to reported CAPEX, allowing investors to profit from the announcement of corporate reinvestment, as well as the subsequent price reversal following the reported numbers. The main conclusion: "Buy on the announcement, sell on the reported."
News sentiment has also proved valuable to fundamental investors beyond equities. In a recent report, J.P. Morgan proposed a framework for style timing in cross-asset risk premia, in which they applied various Machine Learning techniques to generate views on expected returns. After applying the Black-Litterman model, these views were translated into tactical risk premia portfolio tilts.
To capture high-frequency daily sentiment, J.P. Morgan utilized the RavenPack event taxonomy. Filtering on RavenPack’s fact-level score, they were able to focus on articles that contained sentiment on outlook and expectations that can be expected to be more forward-looking such as: business & consumer confidence, economic & inflation outlook, or corporate & sovereign credit sentiment. According to J.P. Morgan, RavenPack sentiment provided most value when modeling Credit/FX/Commodities Momentum, Equity/Commodities/Rates Value, Equity Quality, Equity/Rates Carry, and FX Volatility.
With the emergence of a multitude of longer-term use cases for fundamental investors, news sentiment is no longer used only by quant firms engaged in shorter-term trading. In the last couple of years, alternative data has seen a much wider adoption in the industry and new use-cases are constantly being developed as the marketplace matures.
For a wider catalogue of research papers to download, visit ravenpack.com.