Using Facebook Likes and Google Trends data to forecast tourism

This post is guest authored by Ulrich Gunter, Irem Önder, Stefan Gindl, all from MODUL University Vienna, and edited by the EViews team.  (Note: all images on this post are for illustrative purposes only; are not taken from the published article and do not represent the exact analysis performed for the article). 

A recent article, "Exploring the predictive ability of LIKES of posts on the Facebook pages of four major city DMOs in Austria" in the scholarly journal Tourism Economics investigates the predictive ability of Facebook “likes” and Google Trends data on tourist arrivals in four major Austrian cities.  The use of online “big data” to perform short term forecasts or nowcasts is becoming increasingly important across all branches of economic study, but is particularly powerful in tourism economics.


A quick graph of Google Trends data for the Austrian city of Salzburg compared with tourist arrivals to the same city shows an obvious correlation:

The article used a number of EViews’ automatic and manual forecasting techniques introduced in recent versions to take advantage of this predictive power.

A brief outline of the steps taken to perform this analysis is as follows:

• Monthly tourist arrivals for the four cities of Graz, Innsbruck, Salzburg and Vienna are obtained from the TourMIS database.
• Daily Facebook likes on each city’s official Facebook pages are obtained using Facebook’s Graph API.
• Monthly Google Trends data for each city is obtained from the Google Trends website.
• Once data was obtained it is imported into EViews, using different pages for the different frequencies.
• Seasonal adjustment, unit root tests (with automatic lag-selection) and frequency conversion of daily data to monthly aggregates are all performed in EViews prior to estimation and forecasting.

• Perform univariate automatic model selection on the arrivals data using automatic ARIMA estimation and automatic ETS smoothing.

• ADL models regressing tourist arrivals against monthly aggregated Facebook likes or Google Trends, or both, are estimated.  Lag lengths are automatically selected.

• MIDAS regressions of monthly arrivals against daily Facebook Likes and monthly Google Trends are estimated.
• Using the EViews programming language, all the above estimation techniques are automated and used to perform recursive forecasts with horizons of 1, 2, 3, 6, 12 and 24 months.
• Finally, the EViews forecast evaluation tool is used to figure out the best-performing forecast models per city and forecast horizon (in terms of RMSE, MAE, and MAPE). The forecast encompassing test is also utilized.

The results from this analysis are mixed - for two of the cities, the univariate automatic forecasting methods perform best.  For the third city, the ADL model is best, and for the fourth city, the MIDAS approach is best.