Variable Selection for Inflation : A Pseudo Out-of-sample Approach

In this paper, we analyze the forecasting properties of a wide variety of variables for Turkish inflation, and thereby pin down the ones producing robust forecasts periodically. Defining the lag structure of a variable in two different ways, we determine the non-leading forecasters and leading indicators of inflation. We employ a pseudo out-of-sample approach and compare the forecasting performance of each variable ex-post with the benchmark model. We measure forecast errors over forecast horizons instead of over time for each horizon. Results suggest that no single variable gives the best forecasts at all times, hence inflation is best forecast by different variables each period. This finding promotes the use of forecast combination strategies and/or multivariate model settings.