More than feeling lucky? Using Google Trends to predict the stock marketPosted by Glenn Thrope on Oct 18, 2012 | Tags: Google, Google trends, stock market | Comments Off on More than feeling lucky? Using Google Trends to predict the stock market
As Google has become the primary means of finding information on the Internet, the queries entered into the search engine have emerged as an increasingly interesting source of data. Google makes much of this data available through its website, http://www.google.com/trends.
This search data has been analyzed in all types of intriguing ways. Scientists developed a way of using this data to track the spread of influenza in different regions. Others used the data to draw fascinating conclusions about psychological differences between people in developed and developing nations – namely, those in developed nations are more interested in the future, whereas those in developing nations are more focused on the past. (Interestingly, Google search data appears to be a poor predictor of U.S. Congressional elections.)
Furthermore, this data has obvious uses for businesses. For example, a company like Airbnb might consider a state-by-state breakdown of searches for the term “sublet” as it decides where to deploy its service next. A company like Big Skinny might use a similar breakdown for “new wallet” queries as it decides where to next setup a sales booth.
Perhaps investors can also use this data. For the bushy-tailed MBA student, it doesn’t take long to start asking: Can Google search data predict the stock market? Many believe that stock prices are, at least in the short run, driven partly by the swoons and whims of public opinion. If this were the case, it would be reasonable to hypothesize that Google search queries could serve as a harbinger of these shifts in public mood.
In the rest of this post I address this question. Though my discussion is far from comprehensive, I provide simple examples and computations that might be developed and refined in a more robust study.
I selected three companies for analysis: General Motors, Danaher, and Apple. I chose these companies because they all have several years of data, are in different industries, and are mature. When dealing with younger, less stable companies, there is more noise and thus more difficulty establishing correlation between any two variables.
For each of these companies, I’ve graphed Google’s search volume for the name of that company – i.e. searches for “General Motors” – against that company’s stock price. Here are the results:
In each of these cases, there is a positive correlation between search volume and stock price. However, this correlation appears to be fairly weak for General Motors and Danaher. Even though we see a stronger correlation for Apple, it’s not clear whether the search volume is actually predictive. Maybe stock price is actually predicting search volume!
To dive a bit deeper, I computed coefficients of determination for the graphs above. This allows me to quantify the strength of the correlations. I also correlated search volume to stock price a month into the future, as well as to stock price a month into the past. These calculations help capture any predictive linkages between the data sets.
Here are the results of this analysis:
Looking at the first column, these numbers suggest that the search volume never explains more than 50% of the variation in the stock price, which it does for Apple. Furthermore, looking at the next two columns, when we try to use search volumes to predict stock prices – and visa versa – we are unable to explain a greater percentage of the variation. This suggests that, at least in one-month intervals, search volume alone does not predict stock prices for these companies.
There are obviously holes in this analysis. I selected only a few companies, and I arbitrarily picked one month as my interval of investigation. I also did not consider search keywords other than the names of the companies themselves. Nevertheless, the fact that no strong correlations jump out here suggests that using search data to predict stock prices is probably not a slam-dunk.
Note: All data & analysis used in this post are available here. I used GM data from 11/10 to 10/12, and Danaher and Apple data from 1/04 to 10/12.