Companies that want to forecast future sales of their products can use internet search to predict trends, according to research by Georg von Graevenitz and colleagues, to be presented at the Royal Economic Society''s annual conference at the University of Bristol in April 2017.
The authors use data on searches for car models from Google Trends, plus lagged car sales data for Germany and the UK, to analyse how one influenced the other. Increased searches that demonstrated some intent to buy predicted a greater increase in sales for Germany than for the UK, but both were useful for forecasting.
The authors warn that increased numbers of searches for a brand name alone are not good predictors of sales, because they do not necessarily represent an intent to buy. On the other hand, this is a potentially powerful application of search data if used carefully, they add: ''As internet search also reveals variations in interest in many other things, the methods used in our study should have a wide range of applications.''
In 2009, the German government spent around €5 billion providing incentives for the replacement of older cars in an effort to keep the car industry afloat after the financial crisis of 2008. The first graph below shows the smoothed market shares of the main manufacturers selling cars in Germany before and after that year.
VW, Opel (GM), Ford and Renault did very well out of the subsidy while BMW and Mercedes clearly took a hit.
This temporary shift in demand for cars of specific manufacturers was preceded by a sudden and very strongly correlated increase in searches on Google for the names of the manufacturers and the name of the subsidy, ''Abwrackpraemie''.
The search intensity even reveals how interest initially diminished after the first €1.5 billion was used up and then grew again once the government announced an increase in the total level of the available funds. By September, the funds were used up and searches stopped.
We use this subsidy ''shock'' to analyse how data on searches from Google Trends can be used to predict sales of cars in the UK and Germany. We use monthly data for Germany, but were lucky to be able to obtain weekly data for the UK. This allowed us also to use weekly data that can be obtained from Google Trends on searches for car models.
The paper shows that predicting sales is possible with Google Trends data, but it requires some care. The data we have available are rather sparse (searches for models, searches for manufacturers, searches for the subsidy, lagged sales) and the dynamics of search and sales are tricky.
More importantly, Google Trends data contain information on searches motivated not just by the intention to buy but also by searches after a purchase and by fans of a product. We had to work out how to identify only those searches arising from the intent to buy.
The empirical results demonstrate that if no corrections for other types of search are made, the resulting estimates of the effect of search on sales can be significantly biased.
We find that increased searches predict a greater increase in sales for Germany than for the UK. Search increases of around three percentage points (about one standard deviation) preceded increases in sales four times larger in the German data. On the other hand, searches for the scrappage subsidy, which was lower in the UK, had an effect on sales that was twice as strong as in Germany.
Overall, the paper demonstrates how data on internet search can be used to predict product sales levels. As internet search also reveals variations in interest in many other things, the methods used in the paper should have a wide range of applications.
Predicting Sales with Google Trends Georg von Graevenitz, Christian Helmers, Valentine Millot and Oliver Turnbull
This article was originally published in Dr von Graevenitz''s blog ''Business Analytics, Management and Economics'' on 16 November 2016.