Detection of Seasonality Patterns in Equity News Flow

In order to apply News Analytics efficiently in quantitative trading models, it is necessary to consider certain adjustments or normalizations of time series representing news flow. Such normalizations may be relevant as particular seasonality patterns characterize the data. To identify those times at which news flow is especially relevant to the market, it may be necessary to distinguish true bursts of positive or negative information from mere seasonal peaks in volume. Having prior knowledge of such seasonality effects, allows for proper adjustments before conducting data analysis, and thus prevents a wrong interpretation for instance of the impact of increased News Flow on volatility and trading volume predictions.

In the paper “Detection of Seasonality Patterns in Equity News Flow” (contact RavenPack for a copy), I conducted a study focusing on the seasonality patterns in equity news flow. This was done over different time horizons covering the period 2005 through 2008. The research was conducted on the average news flow in a set of pre-specified news aggregation windows ranging intra-day from one- to sixty-minutes; while a daily aggregation window was applied for the intra-week and intra-month analysis. Finally, a monthly aggregation window was used in the intra-year analysis.

The key findings of the seasonality study can be summarized as:

• Strong seasonality exits on the intra-day, intra-week, and intra-year horizons with median sector-correlations above 73%, 99%, and 98%, respectively.
• Sector-correlations, on an intra-month level, are somewhat lower with a median of 40%. Still, 25% of the cross-sectional correlations are above 62%.
• Applying a cross-sectional linear regression model, it is possible to explain 91.6% of the daily news flow variability using the identified seasonality patterns (including a year-effect and a holiday schedule).
  

Fig. 1 presents an example of the intra-day seasonality pattern based on a 60 minute aggregation window.

The following graphs present examples for each of the intra-week, intra-month, and intra-year seasonality patterns for news flow covering Sector 001 for the period January 2005 to December 2008:

Intra-Week

Intra-Month

(For the Intra-Month pattern, deviations from the zero-line indicates deviations from the intra-week expected pattern.)

Intra-Year

Average Intra-Year

How Does the Market React to News?

The May edition of Factorial! (“Breaking News: How to use news sentiment to pick stocks”), the flagship factor research series of Macquarie Research, looks at how News Sentiment may be used to predict stock price movements. The objective was to identify alpha opportunities not explained by other available information. In general, the report can be split into two main sections. The first section considers an event study that has the objective of validating the “predictiveness” of stock returns using news sentiment, while the second section introduces a multi-factor model with quant factors based on news sentiment. As both sections present very interesting results, I decided to split these into a series of posting. The first posting will be focusing on the event study analysis, which can be considered part of the initial validation process of evaluation the impact of news sentiment on market returns.

The event study was conducted on the constituents of the Russell 3000 index covering both large and small cap companies for the period January 2005 through March 2009 with sentiment data supplied by RavenPack. As part of the study, Macquarie focused purely on high relevance events (Relevance=100 out of 100). In total, just below 53,000 events were identified during the four year period with about 30% translating into short signals and 70% into long signals (as represented by negative and positive sentiment respectively). An event was defined as a news story with at least three of five classifiers holding positive (negative) sentiment, and with the remaining classifiers being neutral. As part of the event study, a pre- and post event window of sixty days were applied.

The key findings of the Macquarie event study can be summarized as:

• Investors are more likely to react adversely to negative news sentiment than to react favorably to positive news sentiment.
• Investors appear to be more aggressive and impulsive on sell decisions and more conservative or tentative on buy decisions.
• Negative news sentiment is a stronger leading indicator of future underperformance than positive sentiment is for future outperformance.
• Investor reaction to neutral news stories is slightly negative, although this may be clouded by the fact that a deep bear market makes up a significant proportion of the sample period. They find that in bear markets, investors’ reaction to the same piece of news is much more negative.
• Investors’ perception of a news event changes depending on market conditions
  

More specifically, Macquarie finds that stocks continue to drift in the direction of the news sentiment for up to three months after a news event. See below figure taken from the report.

Looking at the negative sentiment events in more the detail, Macquarie makes the following observations:

1. Stocks tend to be underperforming prior to a negative news announcement. Macquarie suggests two potential reasons: (1) this may be because companies are reluctant to release bad news until they absolutely have to, so typically the market has already anticipated the bad news before it is finally released. (2) could reflect the fact that news sentiment is somewhat auto-correlated, meaning bad news is more likely to be followed by more bad news, so stocks that experience one negative news event are likely to have already had previous negative events in the past.
2. When bad news is released, the price impact is immediate. The key question for low frequency investors is whether there is any post-event price drift, because this will determine the potential profit opportunity from reacting to bad news. The Macquarie results suggest that stocks continue to drift downwards for the next five days, which is a promising sign for investors.
3. Stocks tend to rebound after about five days. This rebound may be due to the fact that investors initially overreact to negative news. The old adage that investors sell first and ask questions later when faced with bad news seems to apply. The rebound could also be due to short covering as investors who shorted the stock lock-in profits.
4. But after the bounce, securities tend to reverse and underperform for one to three months after a significant negative news announcement. This finding is noteworthy because it suggests that news sentiment does have predictive power in forecasting future underperformance (although the key question here is whether news sentiment adds anything beyond traditional quant factors – will be addressed in the next blog posting).

Looking at the positive sentiment events in more the detail, Macquarie makes the following observations:

1. Securities tend to outperform slightly prior to the release of a positive news event. This is similar to the reaction of securities around the time of negative news events, where securities underperform prior to the announcement date.
2. On the news release date, securities tend to outperform. The extent of the outperformance for positive new events is less than the corresponding underperformance for a negative event. This may suggest that investors are more likely to react adversely to negative news sentiment than react favorably to positive news sentiment. Macquarie highlights that they suspect the presence of a behavioral bias at the root of this asymmetry – as they saw a similar result when studying the market reaction to analyst earnings revisions.
3. There is no short-term reversal after positive news events. Unlike negative news events, investors do not appear to overreact to positive news. Instead, stocks with positive news continue to drift up slowly relative to the market. One explanation for this could be that investors are more willing to wait and take their time digesting positive news before buying in, whereas they just want to get out as quickly as possible when selling after bad news.
4. Securities tend to outperform only slightly one to three months after a positive news announcement. In other words, the day +1 to day +60 return for positive news events is much more muted than that for negative news events. This suggests that negative news sentiment may be a stronger leading indicator of future returns than positive news sentiment.

Impact of News Sentiment on (Intra-day) Abnormal Stock Return

Recently, I conducted a study on the impact of news sentiment on abnormal stock returns with the objective to identify certain event categories that could be traded profitably. The study was based on a portfolio of large cap stocks covering the Dow Jones Industrial Average and the Eurostoxx50 for the years 2005 through 2008. My research focused on intra-day trading, assuming a maximum holding period of five hours (or until market close). Based on a set of sentiment classifiers used to process textual news stories (data supplied by RavenPack), I constructed a set of 63 event categories characterized by unique combinations of sentiment classiffications. Of these, I found that at least 39 categories showed interesting results in terms of creating trading signals. Of all sentiment-based events, 87% showed positive average Information Ratios (IR) amounting to 90,000 events over the four year period with about 50% translating into short signals. On average, the most profitable short trading signals generated 0.70% per trade before transaction costs in the five hours following the event, this with Hit Ratios in the range of 60-65%, see figure below.

The key findings of the Intra-day Abnormal Return study can be summarized as:

• 39 of 63 event categories show positive average Information ratios measured over the five hour post-event window.
• 87% of all sentiment-based events show positive average Information ratios amounting to 90,000 events over the four year period.
• 50% of the identified events translate into short signals, while the remaning events are considered long signals.
• The Top 10 and Top 20 event categories generate about 550 and 6,500 signals over the four year period, respectively.
• 62% of the 39 event categories hold positive average Information ratios in at least three of the four years covered in the study.
• On average, the most profitable short trading signals yielded 0.70% per trade with Hit Ratios in the range of 60-65% (Fig. 2).
• On average, the most profitable long trading signals yielded 0.40% per trade with Hit Ratios in the range of 40-70% (Fig. 4).
  

The most successful long strategy yielded about 0.40% per trade, with a Hit Ratio in the range of 40-70%, see below figure. In total, the “Top 10” and “Top 20” event categories generated about 550 and 6,500 trading signals over the relevant period, respectively.

To consider the consistency of the trading signals, I conducted a year-by-year analysis and found that 62% of the interesting event categories held positive average Information Ratios in at least three of the four years covered in the study.

I think the above findings indicate that News Sentiment could potentially be applied profitably as a factor or an overlay to a trading model, and thus add value in the alpha-generation process at an intra-day level. The full results of my findings are documented in Hafez, Peter. Impact of News Sentiment on Abnormal Returns (May 15, 2009). In the near future, I plan to present a series of interesting results found by Macquarie Equity Research applying news sentiment in more traditional low frequency investment strategies.

What is News Analytics?

Let us first define what is meant by News Analytics. According to Wikipedia, News Analytics refer to metrics derived from textual news stories for the purpose of representing their qualitative nature in a quantitative manner. Measurements of any particular qualitative property are expressed as a specific quantity or unit. Many attributes in textual news stories including sentiment, relevance, and novelty are studied as quantitative properties. Expressing news stories as numbers permits the manipulation of everyday information in a mathematical and statistical way.

News analytics are used in financial modeling, particularly in quantitative and algorithmic trading. They are usually derived through automated text analysis and applied to digital texts using elements from natural language processing and machine learning such as latent semantic analysis, support vector machines, "bag of words" among other techniques.

News Analytics are delivered in a variety of formats, often as machine readable XML documents or .csv files. They include numerical values, tags, and other properties that tend to represent underlying news stories. For back-testing purposes, historical information is often delivered via flat files, while live data for production is processed and delivered in milliseconds through direct data feeds or APIs.

Being able to express news stories as numbers permits the manipulation of everyday information in a mathematical and statistical way that allows computers not only to make decisions once made only by humans, but to do so both faster and more efficiently. Since market participants are always looking for an edge, the speed of computer connections and the delivery of news analytics, measured in milliseconds, have become essential. News analytics not only allows market participants to capture alpha opportunities in the market, it can also be used to improve on risk management and trading execution.

An Introduction

Welcome to the Sentiment News Blog where I plan to post interesting information about the application of news sentiment analysis in both short- and long-term investment decision making.

In search for alpha, quantitative analysts are continuously looking to identify new data sources to maintain their edge in the alpha-generation process. As many of the popular quant-factors have become overcrowded (“quantcentration”), they have also become less effective as more people are chasing the same (limited) alpha opportunities. Asset Management companies have realized the necessity to invest in continued research into the development of less crowded proprietary factors. The availability of News Sentiment data opens up for new sources of potential alpha generation, but also raises the interesting question: “Is news sentiment predictive or mostly explanatory?”. Going forward, I will post some of the research findings that tries to answer this question either directly or indirectly, this amongst other interesting topics within this field.

Comments and ideas are more than welcome, and I hope that we will be able to engage in lively discussions as we start the journey down the News Analytics highway.

Peter Hafez