Exploit and Vanish
Although many investors rely on charts to time trades the actual evidence that this yields any kind of advantage is vanishingly small. It’s not zero, however, as we saw in Technical Analysis on Display, because there’s a faint suggestion that human pattern recognition is so finely tuned that it may be able to extract information from the noise that machines cannot.
Intriguingly, though, research on Google search data suggests that there are exploitable trends in information signalled by interest in specific stocks. Unfortunately merely publishing this fact almost certainly signals its demise as a useful prediction tool. Exploit it while you can, because if you don’t, someone else will, probably using a supercomputer.
The basic idea behind charts is that they signal investor sentiment: investor sentiment may or may not tell us something useful about the fundamentals behind stocks and markets, but if enough people are buying or selling then this may not make any difference in the short-term. The lack of hard evidence supporting this as an effective trading strategy doesn’t put off hardened advocates of the method, who usually have their own personal observations to fall back on.
Typically, of course, personal observation is a very inefficient method of research. It’s entirely possible that one person’s experiences may be atypical of the actual general situation: which is one of the main reasons we tend to rely on carefully designed research programmes and statistical evaluation, rather than individual beliefs.
There’s a second key reason to be wary of personal experience, particularly when it comes to pattern recognition type problems. We are incredibly good at perceiving patterns, but will often identify them when they don’t exist – particularly in so-called conditions of uncertainty, such as those that regularly pertain in financial markets.
These illusory patterns cause us to generate false hypotheses, which normally would fail because they’re wrong and the actual evidence doesn’t support them. Unfortunately, markets being the random places that they are, you can often find examples that do match the observed patterns and, as it turns out, we don’t need much reinforcement to maintain incorrect hypotheses in the face of otherwise overwhelming data. Indeed, there’s an argument that our false pattern matching algorithms are so powerful we should never rely on chart based data on its own, because it’ll all too often be wrong.
However, it’s too simplistic to argue that charts never contain useful data. A more nuanced view suggests that following the same trends and signals as everyone else will lead to diminishing returns as investors try to get ahead of the crowd and ultimately destroy any real information contained in charts. Which, as a corollary, also suggests that chartists will have a better chance of success if they look for unusual and unpopular types of data, and then don’t tell anyone about them.
Of course, the internet and social media have created a whole new source of data about trends and sentiment – some of which we’ve looked at in a range of articles, including Nowcasting With Google which examined the use of Google Trend data to perform various sorts of forecasting. The evidence there indicated that Google was much better and faster at predicting ongoing trends than other more traditional methods. In particular Google data allows nowcasting – predicting what’s happening now – and is much more useful for central bankers trying to figure out the instantaneous impact of policy measures and economic happenstance than the traditional method of “wait and see”.
Now Bodo Herzog has turned his attention to whether Google search data can predict stock price movements. In Asset Prices and Google’s Search Data he’s drawn out the differences between Google search volume and Google search clicks in order to retrospectively analyse the relationship between search data and bank stock prices over the period 2004 to 2010; a particularly interesting selection given the financial crisis that played out over that period.
The difference in these measures is important. Google Trends measures search volume, which is a relative measure comparing the number of searches to its long-term average. Google Insights measures instant search clicks. This matters, because lots of searches indicates interest, not sentiment.
In simple terms if the search volume is small (i.e. it’s in line with the long-term average) but the number of clicks is high you’re looking at a sustained period of high interest in a stock – and you’re probably looking at an equally sustained rise in the stock price. However, once the search volume leaps then you’re into bubble territory:
“As soon as the Google Trends variable (long-run search) is above the average and Google search clicks are further growing, the asset price moves into a bubble or follows a typical unsustainable herd behavior. This effect is demonstrated by the statistically significant negative impact on stock prices via Google Trends. In other words, high attention measured by instant Google clicks has a positive impact on the price. But in case of an asset price boom, which is measured as Google search above its long-run average, the Google Trends variable predicts a kind of the turning point.”
The research also provides evidence for the so-called price momentum effect identified by Jedgadeesh and Titman, which we looked at in Losing Momentum: Is It Time To Exploit Mean Reversion?, where the argument is that investors overreact to signals. The paper also provides evidence for this – anything that increases investor awareness, such as a new story or a regulatory announcement that’s widely publicised, will likely increase Google search volume. As Google search is correlated with stock prices and trading volume, then such news items will likely see short-term price changes.
These are early days for this sort of research and it remains to be seen as to whether it’s robust to exploitation. We’ve seen lots of evidence that anomalies which predict abnormal returns can get arbitraged away by smart traders, so it’s not beyond the realms of possibility that this will as well (see: Pricing Anomalies: Now You See Me, Now You Don't). Obviously, because we’re looking at virtual real-time data this is something that is hard to get ahead of, but sadly the power of high-frequency algorithmic driven trading is likely to be better able to exploit the data than any private investor.
Do No Evil
Of course, the organization best placed to exploit this is the one that owns the data. Trying to figure out how best to do no evil in this environment would tax the resources of a saint, but in the meantime dedicated chartists could do a lot worse than to start mining Google data for exploitable trends.
A word to the wise though. If you find something useful don’t tell anyone. The next thing you know someone will be writing about it and your advantage will be gone.