RWI Konjunkturberichte

This paper introduces the short term forecasting model, which is used as a forecasting tool for the German GDP at the RWI. The model is based on a number of targeted monthly predictors selected from a large set of potential indicators. The selection is conducted by means of a soft thresholding algorithm, which ranks the whole set of potential indicators according to their marginal predictive power. Based on this order, we evaluate the past forecast precision of various subsets to identify our targeted predictors. In what follows, we set up a system of bridge equations, in which quarterly GDP growth is regressed on quarterly aggregates of the targeted predictors. The regression equations either consist of one indicator, one indicator plus lagged values of GDP, or a combination of two different indicators as explanatory variables. Estimated in sample, the regression coefficients enter the forecast equations. To tackle the ragged edge problem, the respective missing monthly indicator values are forecast by means of autoregressive model, augmented by seasonal information with regards to unusual weather and the scheduling of summer vacation. To pool the plethora of single forecasts, we calculate the mean of them, but check the robustness of mean forecasts with regards to using pooling schemes which account for models’ past forecast errors. We find that forecast errors are lowest using less than 30 from the 117 available indicators.

Link to the document