Google’s Chief Economist Mines Predictive Power of Search
With two-thirds of the global search market, Google logs an estimated 34,000 searches per second. That’s two million searches per minute, 121 million per hour, three billion per day, 88 billion per month and roughly one trillion a year, according to the latest ComScore data.
Tapping into this massive treasure trove of real-time data to literally take the pulse of the Web, and gain insight into human behavior and reactions, is surprisingly easy—and free.
Unlocking the search superpower’s data is only a click away at Google Trends, as the company’s chief economist, Hal Varian, demonstrated in December at an event co-hosted by the UC Davis Graduate School of Management, the Department of Economics, the Institute of Governmental Affairs and The Levine Family Fund.
Varian, an emeritus professor at UC Berkeley in business, economics and information management, is a leading econometrics expert and the author of two major economics textbooks that have been translated into 22 languages. He wrote a bestseller, Information Rules: A Strategic Guide to the Network Economy, and was a columnist for the New York Times from 2000–2007. He joined Google in 2002 and has been involved in auction design, econometric analysis, finance, corporate strategy and public policy.
Varian explained that Google Trends and its more advanced cousin, Google Insights for Search, are indexes of normalized keyword query search data dating back to 2004 that are categorized so they can be customized to the context. For example, filtering results for apple, the fruit, versus Apple, the computer firm.
Applying these free tools to searches related to automobile sales, unemployment and travel patterns, Varian led the capacity audience through an eye-opening tutorial on the predictive power of mining search data.
“You really can improve forecasting by including the query data,” he explained. “You can predict the present by looking at the current volume of queries.”
Using search results related to the labor market, Varian showed how he has constructed time series models that correlate closely with, and can quite accurately predict, the trends in unemployment insurance claims reported weekly by the U.S. Department of Labor.
Every Friday Bloomberg reports Wall Street analysts’ consensus predictions on initial unemployment claims for the following week. Varian’s model stacks up surprisingly well. He said the mean absolute error for Wall Street forecasters is about 3% versus 3.5% for his model, a hair worse.
“On the other hand, if you have 30 Wall Street analysts, that’s probably about $15 million a year (in salaries) at a minimum,” he said. “And our data is, of course, free and available to everybody. So maybe it is not quite as accurate, but it is very cost effective.”
Varian also shared how his team at Google analyzes search trends and internal data to fine tune its operations and future strategic directions. Their work has graduated from forecasting revenue growth, query growth and the need for new data centers, to bigger concerns.
“I like to answer the questions that management is going to ask next month,” he said.
“We try to bring data to bear on that problem because Google is a very data-driven company, and when you have the three top executives with Ph.D.s in computer science, they want to see quite of bit of quantitative analysis go into those decisions.”
Varian said Google has about 150 statisticians who are doing quantitative analyses on just about every product and aspect of the company. “These experiments are very carefully designed, treatment-controlled, so we can see the impact of a particular change,” he said. “So the system is evolving.”