Can Google Trends Accurately Forecast GDP Growth?

Summary

• The OECD has created a Weekly GDP Growth tracker using 263 Google Trends search topics. The tracker is highly correlated with Google Mobility and accurately forecasts GDP growth.
• The model is constructed using a neural network algorithm trained to attach weights to Google Trends topics based on how well they explain quarterly changes in GDP growth over a 15-year period.
• The approach’s key features are that it captures country-specific Google search behaviours, identifies which topics contribute most to GDP fluctuations, and allows for non-linear relationships.
• The tracker suggests that Q4 2020 quarterly growth will be negative in many European countries, while Chile, Argentina, Brazil, India and South Africa will have positive growth.

Introduction – A New OECD Weekly Growth Tracker

Several new types of high-frequency indicators have emerged in recent years. And the pandemic has only accelerated the push to explore their viability. A new OECD paper finds Google Trends data can accurately track weekly GDP growth. Using a machine learning algorithm (neural network), the author finds:

• The Google Trends model is better at forecasting output changes than an (autoregressive) model that just uses lags of YoY GDP growth.
• It captures a sizeable share of business cycle variations, including around the Global Financial Crisis and the euro area sovereign debt crisis.
• The Weekly Tracker captures on average 60% of the fall observed in Q2 2020 (Chart 1). Topics related to consumption items and economic anxiety are the main drivers of the fall.
• The tracker closely correlates with weekly movements in mobility. The Google Mobility Index has already been shown to track activity well.

Source: OECD, page 17

Data – Google Trends as a High-Frequency Growth Tracker

Growth indicators that policymakers commonly use fall into two categories. ‘Hard’ indicators are collected by national administrations or statistical agencies and suffer from publication delays ranging from one to three months. ‘Soft’ indicators are timelier but can become less informative about GDP during recessions. Examples include PMIs and confidence surveys (Table 1).

In recent years, ‘high-frequency’ indicators have surfaced. These include flight departures, restaurant bookings, mobility reports based on anonymised personal data from Google and Apple, air quality indices, news-based indicators such as the Economic Policy Uncertainty Index, electricity consumption, and credit card transactions (OECD, 2020).

Google Trends data is another such indicator and has garnered significant attention. It provides Search Volume Indices, which measure search intensity by location and period. Google has classified searches into 1,200 categories, many of which contain monthly observations from 2004 to 2020 for many countries. The Appendix covers some limitations when using Google Trends data.

Using Google’s categories and topics, the author constructs a time series panel dataset that includes 263 topics across 46 countries and 183 months. The categories cover a large number of economic sectors, with a strong focus on consumption because they represent a large share of GDP (Table 2). Information regarding labour markets, housing, and business services are included as they are typically closely linked to the business cycle. Economic anxiety terms capture crises, and industrial activities provide information on the supply side.

Methodology – A Machine Learning Approach

The Tracker uses a two-step model to ‘nowcast’ weekly GDP growth based on Google Trends. In the first step, a quarterly model of GDP growth is estimated using Google Trends search intensities at a quarterly frequency. The relationship between the two is fitted using neural networks. In essence, the algorithm is trained to assign weights/elasticities to different search categories (and so economic sectors) depending on how well they explain changes in GDP growth over time.

The benefits of this approach are that it allows for non-linear relationships between variables to exist. For example (credit OECD paper), the elasticity of the GDP growth to searches for unemployment benefits is lower (the slope is flatter) on the left and higher on the right when the search intensity is higher (Chart 2). This pattern suggests that searches for unemployment benefits are stronger predictors of activity around times when lay-offs increase and therefore become dominant with regards to hiring in explaining changes in employment.

Source: OECD, page 22

Another benefit is that the model can capture cross-country heterogeneity. This means that the elasticities assigned to each search term are specific to each country. For example, Google searches of unemployment benefits may be more correlated with lower GDP growth in the UK than in the US. This also allows the model to make country-specific GDP growth estimations.

In the second step, the estimated elasticities from the quarterly model are applied to the weekly Google Trends series (a crucial assumption is that the relationship is frequency-neutral). Intuitively, the tracker is an estimate of the YoY growth rate of weekly GDP. Quarterly estimates built from these weekly values are provided three-to-seven weeks before actual GDP data is released.

Results – An Good Overall Performance

One way to measure model performance is to run pseudo-real-time simulations. That is, emulating the conditions a forecaster would have faced at each time period. The model is found to perform well in this exercise. Specifically, the error measure (root mean squared error) is lower than other popular models used for forecasting GDP growth (e.g., AR models), especially during the COVID-19 pandemic (Chart 3). Also, for Q3 2020, GDP growth estimates based on the tracker are on average only 1ppt away from actual reported figures.

Source: OECD, page 15

The author also finds a close correlation between the weekly growth tracker and weekly movements in mobility. For example, the timing and relative magnitude of the evolutions of the Weekly Tracker and the Google Mobility Index around the rebound in May-July are very close (Chart 4).

Source: OECD, page 20

The paper expands on the inner workings of the algorithm using Shapley values. These values decompose the predictions made by any algorithm into variable contributions. Focusing on what drove the significant changes in March 2020, they find Google searches on consumption goods and economic anxiety generated the low GDP estimates.

Across the full 15-year period, the author identifies searches corresponding to ‘Unemployment’, ‘Investment’ and ‘Student Loan’ as the largest contributors to variations in GDP growth (Chart 5). Intuitively, a red colour corresponds to a high search intensity. For example, the contribution of the topic ‘Unemployment’ is highly negative when the search intensity is high, and around zero for lower search intensities.

Source: OECD, page 22

Results – Forecasts For 2021

The OECD Weekly Tracker provides early insights into the COVID-19 recovery. Based on information added to the Tracker up until the second week of November, the model shows which countries have the strongest Q4 growth momentum (Chart 6).

Source: OECD, page 31

The results show the difference between the average tracker value over the first two weeks of November and Q3 2020. It suggests that quarterly growth will be negative in many European countries where the stringency of lockdown measures has recently been tightened. Meanwhile, Chile, India, Brazil and Korea are predicted to record a level of GDP higher than 2019.

Bottom Line

Neural networks have had attracted little attention from macroeconomists because of the small size of macroeconomic data. Google Trends Search Volume Indices provide a way around this because they offer a large number of comparable variables across a large set of countries, and at a high frequency. It is, therefore, exciting to see that such data could help make sense of real-time macroeconomic developments.

Specifically, the OECD has put forward a rigorous framework in which to track GDP growth by looking at what people in different countries have been searching on Google over the last seven days. Although the model does not on average outperform models based on more standard variables once these are eventually released, it does offer relatively accurate short-term predictions. For Q4, it is an early indication that European countries are set to suffer from recent lockdowns.

Appendix

Google Trend Data

Google Trends Search Volume Indices come with drawbacks. A couple mentioned in the paper are:

• Queries based on ‘Keywords’ are language-specific and subject to ambiguity. For example, Google Trends series for the keyword ‘apple’ mixes up searches for the fruit and the company.
• This is overcome by using ‘Topics and Categories’. However, the exact content of topics and categories is opaque and the algorithm that allocates keywords can make arbitrary choices.
• Google Trends data is not seasonally adjusted. It also contains breaks between Jan. 2011 to Jan. 2016 caused by changes in the data collection process.
• Finally, as the Google Search user base has increased dramatically since 2004, the relative search intensities of most search categories decrease over time.

Citation

Woloszko, N. (2020), Tracking activity in real time with Google Trends, OECD Economics Department Working Papers, No. 1634, OECD Publishing, Paris, https://doi.org/10.1787/6b9c7518-en.