“Simple correlation doesn’t work. Human judgment doesn’t scale.”
Hal Varian, Chief Economist at Google, begins his talk by asking the audience what day of the week people assume there will be the most searches for the term ‘hangover.’ The Google Insights for Search chart shows a regular ebb and flow, with searches peaking on Sundays, and an additional peak on January first.
Then he breaks up the searches by state, showing that New York has the highest search rate for the word ‘hangover.’ Then he correlates searches for ‘vodka’ with searches for ‘hangover,’ showing that peaks in vodka searches precede peaks in searches for hangover.
What is the practical application?
Fit the best model you can using the data you have. Catching the turning points is the most interesting part.
Example 1: Unemployment
Initial claims to unemployment, unemployment rate, and recession. Unemployment claims tend to peak about 6 months before a recession. Use Google Correlate to feed in a data series, and find the query most highly correlated with that series. There are some statistics here that explain how using the query can help predict one week ahead, which could mean predicting the end of the recession.
EG 2: Destination Planning
Each month, the Hong Kong tourism board releases the data on tourism. Using search terms like “visit Hong Kong” and segmenting searches by country, it becomes possible to predict tourism numbers as much as six weeks ahead of time.
Problems:
- Simple correlation doesn’t work.
- Human judgment doesn’t scale.
Seasonality, for example, can mean that correlation is not causation.
Example 3: Retail sales both adjusted and not adjusted for seasons
When it’s not adjusted for seasons, there are clear peaks for the Christmas season. When it’s adjusted, there is a clear valley in 2010 when there was a recession. Once adjusted, it is possible to predict both adjusted and non-adjusted peaks in sales. Coupon and rebate queries rise during the recession. When you go to Google Correlate, the sales queries are often about malls.
Example 4: University of Michigan Consumer Sentiment Survey
Best predictor of consumer confidence is pension and retirement queries – In a booming economy, there are many searches for pension and retirement. During economic hardship, people search business news, economics, hybrid and alternative vehicles. The cheaper the price of gasoline, the better consumers feel. Using these queries, you can predict the data about a month in advance.
Try this at home!