Can Google location data predict GDP

The corona crisis has boosted interest in so-called nowcasting. As society changes in unforeseen ways, economists don’t want to wait for the publication of official statistics and look for data sources that provide earlier clues about the state of the economy. A new example of such data is the location data provided in Google’s mobility reports. This data reflects activity levels at locations like shops, supermarkets and workplaces.

In an article in Dutch newspaper NRC, ING Bank senior economist Bert Colijn said country-level mobility data shows a ‘very strong correlation’ with GDP (even though Google indicates the method to produce the data varies across countries). ING used this data to calculate a ‘lockdown index’. Despite the term, they suggest this index reflects domestic demand rather than strictness of the lockdown.

The article in NRC contains a scatter plot that shows the correlation between mobility data and GDP change. The chart above is a recreation of that chart. I flipped the axes and added data for a few countries that were not in the original chart (shown in orange). The chart shows that in countries like Great-Britain and Spain, activity levels were about 35-40% lower than normal and GDP decreased by between 20-25%, whereas in South Korea, activity was about 5% lower than normal and GDP decreased by about 5%.

The correlation for the countries in the original chart is .95. After adding the ‘orange’ countries, this decreases to .83, which is still a pretty strong correlation. (In case you suspect the original chart involved cherry-picking of countries that best fit the model, I should mention that some of the added country data wasn’t available yet when the original chart was published.)

So would it be possible to predict changes in GDP using location data? The dashed line in the chart above is a prediction based on the original (blue) countries. It predicts values for some of the added countries (Japan and Slovakia) pretty well, but others less so. However, the more relevant question is whether mobility data can predict future changes in GDP. That’s a difficult question to answer: as the impact of corona on the economy changes, the link between location data and GDP might change as well. So the prudent position would be that location data can be used as an indicator of possible changes in the economy.

Method

Google location data is available here. ING used data for the first half of 2020 and calculated a ‘lockdown index’ based on data it deemed most relevant, i.e. data for ‘shops, supermarkets and offices’. I thought it might take some tedious reverse-engineering to reconstruct exactly how they calculated this index, but it appears they simply took the average values of all data for the variables ‘retail and recreation’, ‘grocery and pharmacy’ and ‘workplaces’ (Google also provides data for ‘parks’, ‘transit stations’ and ‘residential’), for all available dates before 1 July 2020.

ING compares the location data to GDP data for H1 2020. They appear to have used OECD data, and added their own estimates for a few countries for which Q2 data wasn’t available yet at the time (e.g., the Netherlands). Using more recent OECD data, I added some countries not included in the NRC chart.