Overview:
This summary presents key points from a presentation on Google data tools for social science research. The speaker, Hal Varian, highlights three tools: Google Correlate, Google Trends, and Google Consumer Surveys, emphasizing their uses and capabilities.

Google Correlate:
Shows queries most correlated with state-wise or weekly/monthly time series data entered by the user. Limited to US states for cross-sectional data but allows time series analysis in many countries.

Google Trends:
Displays an index of query activity for specific queries or query categories uploaded by the user. Useful for finding queries with the strongest correlation to a given time series.

Google Consumer Surveys:
Facilitates inexpensive and effective surveys of internet users. Easy to use and enables collection of data on a wide range of topics.

Audience Engagement:
A few individuals had experience with Google Correlate, a larger group had used Google Trends, and one person had utilized Google Consumer Surveys.

Google Correlate for House Sales Prediction:
To predict house sales over time, Google Trends data can be used in conjunction with economic data from the Federal Reserve Economic Database (FRED). By selecting a series of interest, such as New One-Family Houses Sold, and downloading the data from FRED, researchers can obtain monthly housing sales data. The CSV file containing the housing data can be uploaded to Google Correlate to find correlated search queries.

Analyzing the Correlated Queries:
Initially, researchers should examine the correlated queries to assess their relevance to the housing market. Some queries may exhibit spurious correlations due to the vast number of queries available. However, queries related to real estate, such as “80-20 mortgage appreciation rate” or “new home builder,” are more likely to be predictive of future house sales than one-time events like “Tahitian nani juice.” The correlation values provided by Google Correlate can help researchers identify the most relevant queries.

Bayesian Structural Time Series Model:
To improve the predictive accuracy of the model, a Bayesian Structural Time Series model is used. This model decomposes the time series into trend, seasonal, and regression components. The regression component utilizes the correlated queries as predictors for the housing sales series. The model also performs Bayesian variable selection to identify the most useful predictors and uses Bayesian model averaging to combine predictions from multiple models, reducing the reliance on a single model.

Advantages of Bayesian Model Averaging:
Bayesian model averaging improves predictive modeling by combining the outputs of multiple models. This approach is particularly beneficial in big data and machine learning contexts. While scientists often seek a single “correct” model, averaging models typically yields better predictions than relying on a single model.

Key Findings from the Bayesian Structural Time Series (BSTS) Analysis:
Appreciation rate emerged as the most prevalent predictor in 100% of the 5,000 BSTS models. IRS 1031 tax form for deferring capital gains on housing appeared in approximately 50% of the models. Century 21 real estate purchases and 80-20 mortgages also showed prominence as predictors. Real estate appraisals exhibited a negative correlation in the regression, indicating a complex relationship with housing prices.

In-Sample Fit and Model Performance:
With a limited number of predictors, BSTS achieved an impressive in-sample fit. The model captured seasonality, recessionary trends, and query frequency, leading to accurate predictions of housing starts. The in-sample fit, however, tends to overestimate the model’s performance out-of-sample.

One-Step-Ahead Out-of-Sample Forecasting:
A simple autoregressive regression of order one was employed for one-step-ahead out-of-sample forecasting. Incorporating query data into the forecasting model led to a 23% improvement in predictive accuracy. The mean absolute error obtained from this method indicates the model’s practical utility for short-term forecasting.

Importance of Now-Casting:
The model primarily engages in now-casting rather than forecasting, as it utilizes contemporaneous query data. This approach allows for real-time analysis of the correlation between housing-related queries and actual housing market activity. Central banks and other financial institutions find this now-casting capability particularly valuable for economic pulse-taking.

* In times of crisis, timely access to economic data is crucial. Traditional data collection methods, such as GDP compilation, can be slow and subject to revisions.
* Alternative data sources, such as MasterCard transaction data, offer near real-time insights into economic activity.
* These datasets can be used for “now-casting,” providing a more up-to-date picture of the economy.

Predicting Real Estate Price Declines
* Google Correlate can be used to identify search queries that are correlated with real estate price declines.
* The most predictive query was “short sale process,” indicating a correlation between mortgage defaults and price declines.
* A cross-sectional model using additional queries, such as “underwater mortgage” and “HARP 3.0,” was able to predict price declines with reasonable accuracy.

Identifying Factors Associated with Short Lifespans
* Google Correlate was used to identify search queries correlated with short lifespans in the United States.
* Queries related to health issues, such as blood pressure medication, and social issues, such as gun violence and religious beliefs, were among the most predictive.
* A regression model combining these queries was able to predict short lifespans with moderate accuracy.

Geographic Patterns in Search Queries
* The geographic distribution of search queries can provide insights into social and economic conditions.
* Queries related to short lifespans were concentrated in the southern United States, particularly in states with a history of poverty and social problems.
* This suggests that search query data can be used to identify areas that may be in need of targeted interventions.

Predicting Hangover Searches from Martini Searches:
A correlation between queries for “hangover” and “martini recipes” was observed with a one-week lag, making martini recipes predictive of hangovers. The pattern holds across countries and is robust to variations during holidays.

Unemployment Claims Prediction:
Google Trends data on searches for “filing for unemployment” closely matches the actual initial claims reported by state employment offices. Combining time-series modeling with Google Trends data improves the accuracy of unemployment claims forecasting.

Conclusions:
Google Trends offers valuable insights into consumer behavior, economic trends, and real-time events. Combining time-series forecasting with Google Trends data enhances predictive accuracy in a variety of domains.

Google Correlate:
Google Correlate reveals correlations between search terms and real-world events. The peaks in the graph represent seasonal workers being laid off after Christmas and auto industry-related layoffs in Michigan during August.

Predictive Papers:
Baker and Fradkin’s research examines the impact of unemployment benefits on job search intensity using Google job searches. Individuals with unemployment insurance search 30% less for jobs, and their search intensifies as the benefits near exhaustion.

Racism and Obama’s Votes:
Seth Stevens Davidovitz’s study estimates that Obama lost about 4 percentage points of popular votes due to racism. The study compared Democratic votes in 2004 and 2008, correlating vote losses with racist searches in certain areas.

Google Surveys:
Google Surveys offer a paywall-based survey mechanism to access premium content. Survey writers create and upload surveys, and users answer them to gain access to the content. Both publishers (who get half the payment) and survey writers benefit, and users get access to premium content.

Survey Quality:
Google Consumer Surveys provide results that align well with traditional random digit dialing surveys. The methodology has limitations, such as biased samples for internet-related questions. Despite the limitations, Google Surveys offer diverse user participation and valuable insights.

Correlating Survey Data with Google Trends:
Combining Google Consumer Surveys with Google Trends led to the development of “survey amplification.” The technique involves associating each survey response with the respondent’s city and building a predictive model using Google Trends categories.

Predictive Models for Survey Responses:
Predictive models are constructed using survey responses and characteristics of the geographic area from where the survey originated. These models utilize demographic predictors, such as age, income, and education, as well as query data from the geographic area.

Marketing and Political Targeting:
The predictive models can be applied to other geographic divisions, enabling the estimation of survey responses for various regions. This approach allows for effective marketing and political targeting by tailoring messages to resonate with the population in specific areas.

Identifying Resonant Messages:
Predictors from a survey on product assembly preferences in America included Chevrolet, firearms, country music, and trucks. Such insights help in identifying resonant messages for advertising campaigns in different regions.

Targeting Product Features:
A survey on mobile phone customization preferences revealed that different features resonated with different demographics. Urbanistas, young women interested in fashion and design, were receptive to messages about customizing the look of their phones.

Availability and Accessibility:
Encouragement to explore Google Correlate, Google Trends, and Google Consumer Surveys for further insights. Emphasis on the accessibility and potential of these tools for uncovering valuable insights.