Wednesday, September 2, 2009

Identifying Events Using a Sentiment Approach

Posted by at
Event-driven strategies are often cited as a useful alternative alpha source for quantitative investors because returns to these strategies tend to be uncorrelated with traditional multifactor quant models. However, there have always been two major challenges with event-driven investing including the infrequent nature of such events and the lack of data to detect and capture events in a timely fashion.

In some of my previous research, see the posting Impact of News Sentiment on (Intra-day) Abnormal Stock Return from June 10 2009, I proposed constructing intra-day triggering events based on unique combinations of sentiment classifications after applying a relevance filtering. A similar study was conducted by Macquarie US Equity Research with a focus on longer-term investing, see the posting How Does The Market React To News? from June 17, 2009.

Even though it is possible to construct attractive trading and investment strategies based purely on sentiment, it may be desirable to further enhance the strategy by considering other aspects of a news story. For example, news stories can be further categorized into frequent events like earnings or earnings guidance announcements, executive appointments, analysts ratings, and less frequent (and unexpected) events like product recalls, lawsuits, layoffs, or bankruptcies. News filtering can be based on the type of event of a given story, since the expected market reaction to certain events could be very different. Using this information, it may also be desirable to decide on the traded or invested amount, thereby being more bullish on a story about a new business contract than on a story about a charity donation.

Figure 2 shows the frequency of the top 20 events reported by Russell 3000 companies as sampled from the RavenPack News Scores database. As expected, Revenues is the most common event, with over 60,000 instances from January 2005 to August 2009. This particular event category captures when a company announces sales or revenue figures. Roughly speaking, earnings related events are most common, followed by sell-side analyst upgrades or downgrades, and then by company restructuring type events (reorganizations, layoffs, executive resignations, etc.).

Interestingly, it can be observed from Figure 3 that the frequency of events evolves through time. Some events, like earnings announcements, are periodic and occur every quarter (creating a seasonal pattern), whereas others are cyclical and rise and fall with the businesses cycle. Layoffs are an unfortunate recent example.

In a recent publication by Macquarie US Equity Research a number of triggering events were constructed using an ‘event-driven’ news sentiment approach. The study considered the constituents of the Russell 3000 and was based on the RavenPack News Scores database, which in addition to measuring the sentiment of news articles also flags over 100 scheduled and unscheduled events typically reported by the media.

Using news sentiment to determine the expected direction of an event is a big advantage as the other alternative is to rely purely on the market reaction of similar stocks to the same event in the past. Figure 5 shows the average excess returns of stocks after positive sentiment events, negative sentiment events, and all events, for the Top 20 events reported in the news.

On average, Macquarie finds that excess returns, particularly in the month after the event, are well delineated by the sentiment of the event. In other words, after an event with positive sentiment, on average returns tend to be positive, whereas after a negative event, they tend to be negative. However, this is not a universal result – there are a number of events (e.g., buybacks) where the sentiment of the event appears to have no impact on next month’s returns. Overall, the results in Figure 5 suggest that using sentiment to determine whether an event is likely to have a positive or negative impact has merit.

In future postings I plan to dig a little deeper into the report looking in more detail at the market reaction to some of these events.