Guan recently cited some personal correspondence where I discuss the Anne Hathaway anomaly. Essentially, a Huffington Post blogger noted that when Anne Hathaway is mentioned in the news, Berkshire Hathaway's shares went up in value (n=6). He argues that this is because automated trading systems which consume the text of news and output trades are responding to false positives. This is an exceedingly small sample, but I think the underlying argument is plausible.

As I wrote to Guan, I don't really know what hedge funds using automated trading systems are doing in practice. This kind of software is proprietary for very good reasons. I can tell you how I would go about building an automated trading system (ATS), and this might give you some insight into how it works.

First, subscribe to a news source. ATSs consume a real-time stream from Reuters or some other provider of news.

Second, build an stock classifier. This is model which uses the text in headlines and articles to determine which stocks are discussed in the article. Typically this is something like naive Bayes, but perhaps with some natural language parsing or heuristics built in. The output is probabilistic, something like "we are 60% sure this article is about Berkshire Hathaway." Obviously the presence of the word "Hathaway" is going to be an important feaure in the model. And since "Hathaway" is a rare word, it would have comparatively high weight in most models of text.

Third, extract some features from the articles. This could be something like sentiment or it could be learned inductively from a subsample of "good events." For instance, we might find that certain phrases (n-grams) are present in articles that precede large returns for a certain stock (in the training set). The presence of this phrase would become an important feature. One might imagine articles about Ms. Hathaway often involve a lot of positive language, which a system could mistake for favorable news.

Fourth, search for trading rules which are profitable (and low risk/cost) on a training set. Rules include things like thresholds for how certain you need to be that an article is about a stock and the weights on the features you've extracted. As an example, a rule could be "if the article is about stock X with p > .75, and feature Y is present, buy Z shares and hold for 5 days with a certain stop-loss." There is really an endless amount of complexity that could be used here, but more complex rules are less likely to perform well out of sample.

Fifth, test the rules on a test data set which is later in time that the training set. The ones that perform well according to some risk/return profile go into production (and are constantly monitored).

---

I'm simplifying a lot here (particularly about how one might determine what is "new" news), but this is the gist. You can see how both the classifier and the rule might have false positives. Except for added transaction costs, false positives in trading systems may not really be risky at all (you are simply going long on a stock), so the system may be tuned to accomodate a fair number of them.

ATSs like this are essentially capitalizing on some non-instant, but fairly deterministic response of investors to news. Often, opportunities like this are termed "anomalies," i.e. deviations from market efficiency. The interesting bit is that they can easily lead to their own anomalies. To continue the example, it is possible one could profitably short Berkshire Hathaway when there is good news about Anne Hathaway--an anomaly created from ATSs trading based on another anomaly. As I wrote to Guan, I suspect we'll see more of these as the ATS arms race escalates. If you think about it recursively, it really emphasizes that these funds are playing a fixed-sum game. To the cleverest go the spoils.