Nowcasting Growth using Google Trends Data: A Bayesian Structural Time Series Model

Main points

• Extend upon a popular frontier Bayesian nowcasting method to help identify variables that drive the cyclical component in aggregate economic time-series
• Proposed high dimensional variable selection uses a Bayesian decision theoretic perspective: maximise fit with a preference for sparsity
• We apply this method to nowcast US quarterly GDP growth based on traditional monthly macroeconomic data and a large set of Google Trends (GT) indices
• We find that GT improve nowcasts before traditional macroeconomic data get published; GT terms with high inclusion probability have good economic interpretation, reflecting leading signals of economic anxiety and wealth effects
• The proposed model outperforms competing models in the data, including post-Covid period, as well as in simulation

Commonly employed nowcasting models are good at generating short run predictions of the target variable, usually an aggregate economic time-series, such as quarterly real GDP growth in response to consumption, production, investment, and price data which are typically available at a higher frequency than GDP. In this paper, however, we are interested in augmenting the standard list of higher frequency macro data with a large set of Google Trend (GT) search term indices. GT measure the popularity of a search term or basket of search terms typed into Google. Google Trends have been used in a large swathe of recent research articles to measure the demand of supply of goods, as well as recession fears. We are the first to investigate whether GT provide information about aggregate economic behaviour above and beyond traditional macro data in real time. Our goal on the modelling front, is to efficiently make use of the different data sources as well as communicate to the policymaker clearly which model components affect nowcasts the most.

To extract the most value of the data, we make use the Bayesian structural time series model (BSTS). The BSTS decomposes a time series into high-frequency movements, captured by a MIDAS style mixed frequency regression and latent, slow moving trends as well as seasonal effects. In the context of nowcasting GDP, this framework can therefore be used to model the cyclical component of GDP in response to the macro data as well as long-run changes in economic conditions via unobserved components. Since such a model becomes quickly high-dimensional and hard to interpret, we bring forward two main contributions 1) we extend frontier Bayesian shrinkage priors and variable selection tools to the BSTS to communicate the importance of both the macro and search term data for the cyclical component, and 2) we reformulate the unobserved component structure to allow variable selection over long-run trends. Both innovations are geared toward making the model form data driven, as well communicate to the policymaker which parts of the model have an impact on nowcasts.

Our application to US quarterly GDP growth shows that Google’s search terms provide significant gains in nowcast precision prior to the arrival of traditional macro information, particularly during the pandemic. Our model shows that only a sparse subset of Google Trends related to signals of economic anxiety and wealth effects are important in the model. We additionally find that there is clear support of a shifting long-run trend in the data and that the proposed model clearly outperforms similar models in the literature. We back up the good empirical performance with a simulation exercise that shows that our model is especially potent in dense DGPs which is typically expected in a macro environment.