Nowcasting the Unemployment Rate in Turkey : Let's ask Google

We investigate whether Google search query data can improve nowcasting performance of the monthly nonagricultural unemployment rate for Turkey, where monthly unemployment rate is revealed with a lag of three months. To do so, we employ linear regression models and Bayesian Model Averaging Procedure in our analysis and use data from January 2005 to October 2011. We show that Google search query data is successful at nowcasting nonagricultural unemployment rate both in-sample and out-of-sample. When compared with an autoregressive benchmark model, where we allow only the lag values of the monthly unemployment rate, the best model contains principal components of Google search query data and it is 47.7% more accurate in-sample and 38.4% more accurate out-ofsample in terms of relative root mean square errors (RMSE). The best model that does not include any Google data is 34.1% more accurate insample and 29.4% more accurate out-ofsample. We also show via Harvey et al (1997) modification of the Diebold-Mariano test that models with Google search query data indeed perform statistically better than the autoregressive benchmark model.