Are Google Trends data an independently useful tool in predicting stock returns? In their March 2014 paper entitled “Do Google Trend Data Contain More Predictability than Price Returns?”, Damien Challet and Ahmed Bel Hadj Ayed apply non-linear machine learning methods to measure whether Google Trends data outperform past returns in predicting future stock returns. They focus on avoiding bias derived from choice of keywords (choosing words with obvious retrospective, but dubious prospective, import) and test strategy parameter optimization. Since Google Trends data granularity is weekly, they employ a six-month calibration interval to predict weekly stock returns. They apply a 0.2% trading friction for all backtested trades. Using weekly returns and Google Trends data for stock tickers and firm names plus other simple, non-overfitted words for the S&P 100 stocks as available through late April 2013, they find that:
- Backtests yield an average net weekly portfolio-level profit of about 0.17% with standard deviation of weekly returns 1.3%, whether based on past returns, Google Trends data or a combination of both.
- Results suggest that the information content of Google Trends data is equivalent to that of past returns (but the latter are more easily and reliably available).
In summary, evidence suggests that a careful analysis of weekly Google Trends data is no more informative (but more difficult to acquire/process) than past stock returns in predicting future stock returns.
Findings are a challenge to those of past studies, such as the study summarized in “Google Trends Predict the Stock Market?”.
Cautions regarding findings include:
- As discussed in the paper, avoiding bias in retroactive keyword selection and addressing changes in Google Trends methodology over time are problematic.
- As noted in the paper, testing involves tool bias, since the methods/computing power applied are not available over much of the sample period.
- Methods used are beyond the reach of many investors (or costly if delegated).
