Forecasting Births Using Google

Francesco C. Billari, University of Oxford
Francesco D'Amuri, Bank of Italy
Juri Marcucci, Bank of Italy

In this paper we propose a new leading indicator based on Google web-searches as a predictor of monthly birth rates. We then test its predictive power using US data. In a deep out-of sample comparison we show that popular time series specifications augmented with web-search-related data definitely improve their forecasting performance at forecast horizons of 6 to 24 months. The superior performance of these augmented models is confirmed by formal tests of equal forecast accuracy and superior predictive ability. Moreover, our results survive a falsification test and are confirmed also when a forecast horse race is conducted using different out-of-sample tests, and at the state rather than at the federal level. Conditioning on the same information set, the forecast error of our best model for predicting 2009 births is 35% lower than the Census bureau projections

  See paper

Presented in Session 155: Methods and Models in Fertility Research