Upcoming Presentation on News Analytics & Finance

Next week, I will be presenting in London at the CARISMA workshop on "News Analytics Applied to Trading, Fund Management and Risk Control." The workshop will be held on February 1st, 2010 at 09h15 (local time) at the MWB, 25 Canada Square, Canary Wharf, London, E14 5LQ.

My presentation is entitled “The Role of News in Financial Markets" where I will provide an overview of how news analytics may be implemented into ones trading or investment environment considering aspects such as sentiment and the impact news may have in formulating and constructing trading strategies.

I also plan to show some of my latest research and review a few interesting studies recently published on news sentiment analysis.

The News Analytics workshop is a precursor to the 5th Annual CARISMA Conference on “The Interface of Behavioural Finance and Quantitative Finance” being held on February 2nd through February 3rd, 2010 also at the MWB.

You can find more information about the CARISMA News Analytics Workshop and Conference at: http://www.optirisk-systems.com/events/carisma2010.asp.

Capturing the Seasonal and Trend Components of News Sentiment Indices

Equity news flow is characterized by strong seasonal patterns that includes quarterly bursts of information due to earnings releases. To address this issue, I indirectly normalize by news flow and focus on the Sentiment Ratio (count of positive divided by the count of negative news items) when constructing Sentiment Indices. This seems to be a good approach under the assumption that the seasonal bursts of information impact both the positive and negative news counts symmetrically. Unfortunately, this does not seem to be the case.

Looking at Sentiment Indices constructed based on a one month trailing aggregation window reveals what seems to be a seasonal pattern in the Sentiment Ratio itself. This is not particularly surprising in that earning seasons are likely to contain more company-initiated news stories than on average. Generally, such stories tend to have more of a positive spin than the ones initiated by more objective third-parties.

In Figure 1, I have included two graphs that present both the "raw" and the seasonal-adjusted Sentiment Indices based on something called the Berlin Procedure. This approach helps decompose a time series into five primary components including trend, cyclical, seasonal, calendar, and irregular components. After the decomposition, you generally have three tools to assess the sentiment environment. Firstly, the seasonally-adjusted series that more clearly reveals the trend-cycle movements. Secondly, the estimates of the trend-cycle components of the time series; and thirdly, the original "raw" Sentiment Index. As expected, the seasonal-adjusted index becomes smoother since it removes the strong seasonal components.

Figure 1: A raw and seasonal-adjusted Sentiment Index (Jan 2005 through Nov 2009) based on a one month trailing news aggregation window. The seasonal-adjusted Sentiment Index (using the Berlin Procedure) considers an adjustment of the Sentiment Ratio due to an asymmetric impact of the count of positive and negative news items caused by company initiated news stories.

Before we perform a seasonal adjustment, we have to ask ourselves whether this is really necessary given a trading strategy. After all, these bursts of mainly positive news are "real" and could in fact affect the markets; hence we may simply want to use the "raw" index. That being said, it can probably be expected that market participants are well-aware of this Sentiment Bias as captured by the seasonal component, and thus to some extend discount such company initiated information. Taking this into consideration, it may be desirable to consider both indices simultaneously before making ones trading decisions. Doing so, it is possible to arrive at an estimate of the Sentiment Bias and thereby form a set of hypotheses on how such information should be considered in a trading model.

Comments and ideas are as always more than welcome.

What Does News Tell Us About the Current Sentiment of the US Equity Markets?

In a previous blog posting, I presented a methodology for how to construct a news sentiment index, that being either on a sector, industry or a broader equity index (in fact, any portfolio of stocks can be used as a basis). In the below figure, I have included a Market News Sentiment Index using a 90 day trailing news aggregation window, where values above 50 represent positive sentiment, values below 50 negative sentiment, and values at (or close to) 50 neutral sentiment. The figure includes data from January 2005 through November 2009.

As can be observed, the Market News Sentiment Index broke into negative territory well before the market high in October 2007 (in fact, it happened as early as June 2007), and thus could have acted as an early warning sign for following market downturn.

Since March 2009, sentiment has significantly improved entering a positive sentiment regime in May the same year. Using a 30 day news aggregation window, the regime shift would have been detected already in March 2009 indicating a sudden change in the "mood of the market". Since May 2009, the market sentiment has continued to improve with a drop in sentiment levels in August 2009 and September, which could be read out of the more "muted" market performance in September and October 2009. Recently, News Sentiment levels have significantly improved starting in October, with a following strong performance of the equity markets in November.

With the current News Sentiment levels at a record high, equity markets could well continue to increase in the periods ahead. Surely, it will be interesting to keep an eye on the development of the Market News Sentiment Index and look out for sudden changes (warning signs) in news sentiment.

Wishing you all a Happy Holiday and a prosperous 2010!

Forum on News Analytics

Recently I attended the Forum on News Analytics in London, organized by Carisma from Brunel University. In fact, Brian Sentance wrote a good summary of the event on his blog.

As part of the forum, a panel discussion was held with attendance from Northfield Information Services, Macquarie Equity Research, RavenPack, Thomson-Reuters, Semlab, and Capital IQ ClariFI.

What I mostly found interesting was that the different news analytics providers had chosen to approach the market from different angles. For instance, RavenPack and Reuters offered data feeds and "ready-made" analytics, while Semlab required the end-user to define their own set of rules. From my experience, trying to build the lingusitic models and NLP algorithms yourself is a daunting task and definitely a fulltime job. As a quant, I prefer to focus on the analysis of the news data.

Carisma, Northfield and Macquarie briefly discussed some of their research involving news analytics (all studies were based on RavenPack data). Generally, it seems that most of the practical research out there is based on RavenPack data.

The forum was well attended with people coming from both academia and industry including high and low frequency traders, risk managers, and people focusing on algorithmic execution. I had a chance to talk to people from both camps and there is no doubt that news analytics has surely caught people's attention, and is becoming a hot topic. Especially, it seemed as if people were very much interested in discussing the different techniques that one can use to extract sentiment, but also to use such information to construct triggering events and sentiment factors.

It was definitely an interesting event, and I am looking forward to attending Carisma’s next event "The Interface of Behavioral Finance and Quantitative Finance" and the pre-conference workshop "News Analytics Applied to Trading, Fund Management, and Risk Control" taking place in London early next year.

Sector Rotation Strategies Driven By News Sentiment Indices

Recently, I conducted a study on how to construct sector rotation strategies driven by news sentiment indices. The applied methodology was presented in a previous posting: " Construction of Market Sentiment Indices Using News Sentiment" from August 5th, 2009.

As part of the sector rotation study, I constructed a set of industry sentiment indices, and a market sentiment index as described in a previous posting. The strategy I tested included going long the Top 5 sentiment industries during positive market sentiment regimes (market sentiment index values must be above 50), and short the Bottom 3 sentiment industries during negative market sentiment regimes (market sentiment index values must be below 50).

Interestingly, I found value in every stage of the following four-step procedure:

Step 1: Market Return Momentum Strategy on DJIA

Go long the Dow Jones Industrial Average (DJIA) when the previous month's return has been positive and go short if the previous month's return has been negative (Black-line). Covering the period Feb. 2005 to Sept. 2009, the strategy yielded an annualized return of 11.25% with an Information ratio of 0.73.

Step 2: Industry Return Rotation Strategy with Market Return Overlay

Go long the Top 5 industries when the previous month's DJIA return has been positive and go short the Bottom 3 industries when the previous month's DJIA return has been negative - Momentum (Green-line). Covering the period Feb. 2005 to Sept. 2009, the strategy yielded an annualized return of 18.49% with an Information ratio of 0.95.

Step 3: Industry Return Rotation Strategy with Market Sentiment Overlay

Go long the Top 5 industries when the previous month exhibited positive market sentiment and go short the Bottom 3 industries when the previous month exhibited negative market sentiment (Red-line). Covering the period Feb. 2005 to Sept. 2009, the strategy yielded an annualized return of 27.23% with an Information ratio of 1.37.

Step 4: Industry Sentiment Rotation strategy with Market Sentiment Overlay

Go long the Top 5 industries when the previous month exhibited positive market sentiment and go short the Bottom 3 industries when the previous month exhibited negative market sentiment (Blue-line). Covering the period Feb. 2005 to Sept. 2009, the strategy yielded an annualized return of 29.63% with an Information ratio of 1.80.

The key findings of my study can be summarized as:

• Negative sentiment seems to be a strong leading indicator of future underperformance, while positive sentiment is not as clear a leading indicator of future outperformance.
• Based on an industry sentiment ranking, the Top industries seemed to reach their cumulative return high six months later than the Bottom industries (April 2008 versus September 2007).
• Tracking the sentiment of the Top and Bottom industries could be a valuable input into the creation of a dynamic leverage factor moving more aggressively into ones trading signals in extreme sentiment regimes.
• An industry sentiment rotation strategy with a market sentiment overlay would not only have outperformed an industry return rotation strategy during the 5 year period, but would have done so with a more attractive Information ratio (see above Figure).
  

While more research can be done to refine the methodology and to translate the sentiment index values into actual trading or investment signals, the above results highlights some potentially interesting and profitable relationships between news sentiment and market returns. The full results of my findings are documented in the paper Sector Rotation Strategies Driven By News Sentiment Indices (November, 2009).


Source: RavenPack

Extracting News Sentiment Using an Expert Consensus Methodology

One way to calculate news sentiment in finance is to use something RavenPack calls Expert Consensus. The idea is you take a group of say 10 financial analysts, give them the same set of news stories to read, and ask them to rate each story as positive, negative, or neutral relative to a stock.

The sets involve both news flashes and articles and includes about 10,000 stories picked from various publishers and segments of time. Because of the large volume, it can take about 200 man hours per analyst to complete their classification. In addition to sentiment ratings, analysts are asked to rate the impact of the story on a scale of (1-5) and provide individual classifications should there be multiple companies covered in the article. Analysts are asked to think in terms of stock price impact over different trading windows. Is the news likely to impact the price of the stock positively or negatively in an hour? By the end of the trading day? Over the Long term?

Once the training sets are complete, RavenPack looks for patterns across the classifications provided by the analysts to determine which types of news are consistently classified as positive or negative, thereby looking for consensus. So if 9 out of 10 analysts agree that bankruptcy news is negative with a high impact score, the algorithm would be trained using this data. The sentiment algorithm learns from the classifications provided by the group consensus and uses this information as its training set to analyze news stories automatically moving forward.

What’s neat about this technique is that it tries to emulate expert group behavior and can perform this task consistently, continuously and in a fraction of a second after news gets published. Normally, it would take minutes for a financial analyst or trader to digest each piece of news, let alone act on it under some form of consensus.

Market Reaction To Sentiment-Based Events

In a previous posting, I described how it is possible to identify and make trading decisions based on a range of events such as earnings or earnings guidance announcements, executive appointments, analysts ratings, etc.; this with news sentiment determining the expected direction of the market. When using sentiment to decide on the expected direction of certain triggering events based on RavenPack data, a number of approaches are available. For example:

• Using a single sentiment score allows for the simplest approach in determining whether a news item is positive or negative. For instance, a news item with a RavenPack sentiment score (0-100) above 50 is positive and below 50 is negative. However, previous studies indicate that taking a multi-classifier approach to sentiment adds value due to the diversification effect available amongst the different classifiers.
• Using average or weighted scores allows for a simple evaluation of multi-classifier sentiment, where high (low) values could potentially be considered reflecting greater positive (negative) sentiment. Alternatively, a simple positive or negative classification can be used depending on whether the score is above or below 50 and contextually relevant to a given entity (i.e. a company). The "average" approach was applied as part of constructing market sentiment indices (factors) in the study "Construction of Market Sentiment Indices Using News Sentiment".
• Using unique combinations of sentiment scores, which was the focus in the study: "Impact of News Sentiment on (intra-day) Abnormal Stock Returns". An advantage of such approach is that it is possible to evaluate how different combinations of sentiment classifications can impact the market potentially under different environments and for different event types. In contrast, a disadvantage is that one may end up with only a few occurrences of each "signal", thus requiring an evaluation of a larger universe of companies as part of the research and model construction phase.
• Using a voting (consensus) mechanism could potentially create stronger confidence in correctly classifying a news item as being overall positive or negative as it relies on several sentiment analysis techniques that have to be in agreement. This approach is more restrictive than the average sentiment approach, and therefore a large number of potentially interesting signals a filtered out resulting in less signals overall but of higher quality. This approach has been applied for instance in a study by Macquarie US Equity Research: "Breaking News: How to use news sentiment to pick stocks".
  

In what follows, I will highlight some of the results from an event study that considers about 100 different types of events, as covered by RavenPack. As part of this study, events are considered positive or negative using a voting mechanism with at least three out of the five RavenPack sentiment scores having to be greater (less) than 50 (out of 100), and the other two scores being neutral. The company relevance score must also be 100%.

Considering the constituents of the Russell 3000, the following charts present the results of selected events (the numbers in parentheses denote the total number of events of that type and the percent of all events in the sample). Furthermore, the charts presented show the mean and median market- and sector-adjusted returns for stocks around positive and negative events.


Earnings and guidance events

Overall, earnings and revenues related events are the most common in the database, which comes as no surprise since most companies make such information available on a quarterly horizon. Considering positive and negative revenue announcements, the key results can be summarized as:

• For positive revenue announcements (Figure 6), on average there tends to be a mild post-event upwards drift relative to the market. However, the median return actually gives a slight downwards drift.
• For negative revenue announcements (Figure 7), the post-event drift is much stronger regardless of whether the mean or median is used. This suggests that even after negative revenue announcements, there is still substantial alpha available from shorting this category of stocks.

Analyst Ratings

The second most common category of events is those generated by sell-side analyst activity. Figures 10 and 11 show the average stock price reaction to changes in ratings by sell-side analysts.

• On average, sell-side analysts tend to be contrarian in their ratings changes. In other words, they tend to upgrade stocks that have been underperforming, and downgrade stocks that have been outperforming. This is opposite to the conventional wisdom that sell-side analysts tend to follow recent price momentum.
• Post-upgrade and post-downgrade drift is less clear-cut. From the perspective of the median stock, it appears that both upgrades and downgrades lead to underperformance.
• Using average returns, post-event performance is relatively in-line with the market.
  

M&A activity events

Assessing the impact of M&A activity has always been challenging, because every takeover situation is a little different from the last. Figures 14 shows the case where a company announces it is acquiring another company, while Figure 15 shows what happens when a company is itself the target of a takeover. The key results can be summaries as:

• A few outlier stocks tend to have a big impact on returns to the strategy as reflected by the big difference between the mean and the median.
• On the day of the announcement, the price of the acquiree’s stock jumps up sharply on average. However, post-event, there is little drift on average.
• For the acquirer, there is no post-event drift when using average returns. However, median returns tend to indicate that stocks that look for acquisitions tend to underperform before and after the event.

Other events

Ad hoc events like bankruptcies (Figure 16) tend to occur infrequently, but when they do happen, the returns are often very large. Fraud (Figure 17) is another somewhat rare event. The key results can be summaries as:

• Following a bankruptcy announcement, the stock price plunges, but then tends to bounce back in the next five to ten trading days.
• On average, stocks that report fraud tend already to have been underperforming before the fraud is announced. After the event, these stocks continue to underperform.