Title: Forecasting industrial production in Germany
Authors: Sercan Eraslan - Deutsche Bundesbank (Germany) [presenting]
Klemens Hauzenberger - Deutsche Bundesbank (Germany)
Abstract: A parsimonious forecast model is introduced for industrial production in Germany. Our framework is based on the combination of point and density forecasts obtained from a pool with a relatively low number of models capturing, in particular, the long-run dynamics of industrial production. The model pool consists of three model classes: vector AR and error correction as well as multicointegration models. While standard vector error correction models are based on the relation between industrial production and new orders received, the multicointegration framework allows for a second long-run relation in the model: the relation between industrial production and inventories and/or stock of orders. Moreover, we augment all model classes with survey-based indicators specifically targeted at the industrial sector. We consider both linear and threshold-type nonlinear specifications, and estimate the models with Bayesian and frequentist methods. At a final step we combine the point and density forecasts in various ways. Our first results based on a real-time forecast evaluation between 2006 and 2016 points out that the parsimonious forecast routine with around ten models can reduce the root mean squared forecast errors by up to 30\% over a 12-month forecast horizon against the random-walk benchmark.