Journal of Forecasting, J. Forecast. 32, 577–586 (2013) Published online 22 July 2013 in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/for.2260 Evaluation of Regime Switching Models for Real-Time Business Cycle Analysis of the Euro Area MONICA BILLIO, 1 LAURENT FERRARA, 2 DOMINIQUE GUÉGAN 3 AND GIAN LUIGI MAZZI 4 1 Dipartimento di Economia, Università Ca’ Foscari Venezia, Italy 2 EconomiX UMR 7235, Banque de France and University Paris Ouest, France 3 CES UMR 8174, University Paris 1, Panthéon-Sorbonne, France 4 Eurostat, Luxembourg ABSTRACT In this paper, we aim at assessing Markov switching and threshold models in their ability to identify turning points of economic cycles. By using vintage data updated on a monthly basis, we compare their ability to date ex post the occurrence of turning points, evaluate the stability over time of the signal emitted by the models and assess their ability to detect in real-time recession signals. We show that the competitive use of these models provides a more robust analysis and detection of turning points. To perform the complete analysis, we have built a historical vintage database for the euro area going back to 1970 for two monthly macroeconomic variables of major importance for short-term economic outlook, namely the industrial production index and the unemployment rate. Copyright © 2013 John Wiley & Sons, Ltd. KEY WORDS regime switching models; SETAR models; business cycle; turning points INTRODUCTION We recently witnessed the development of new modern tools in business cycle analysis, mainly based on nonlinear parametric modelling. Nonlinear models have the great advantage of being flexible enough to take into account certain stylized facts of the economic business cycle, such as asymmetries in the phases of the cycle. In this respect, emphasis has been placed on the class of nonlinear dynamic models that accommodate the possibility of regime changes: in particular, the Markov switching (MS) model popularized by Hamilton (1989) and its multivariate version (Diebold and Rudebusch, 1996; Kim and Nelson, 1999), which allows co-movements among countries and sectors to be taken into account. Besides the well-known MS approach, another parametric model allows for different regimes in business cycle analysis: the threshold autoregressive model (Tong, 1990), which can describe the asymmetry observed in economic variables (see also Tiao and Tsay, 1994; Proietti, 1998 or Ferrara and Guégan, 2006, for applications on business cycle analysis). The main difference between these two approaches concerns the use of a transition variable to describe switches: it is an observed exogenous or endogenous variable for SETAR modelling or a latent Markov chain for MS modelling. We resort to these two different classes of models because they underpin a complementary rather than an alternative approach, as the switching mechanism captured by them is not exactly the same. Thus, in this paper, we propose a new approach for business cycle analysis comparing the ability of these two models to date and detect in real time turning points on two different economic datasets: industrial production index and unemployment rate. These two series are of major importance for short-term economic outlook; they are usually considered to date economic cycles and are available for the euro area on a monthly basis. After a specification phase to determine the best modelling in each class, we increase successively the original dataset by moving through all the available vintage monthly releases. We specify again and re-estimate the models at each step in order to detect in real time the occurrence of turning points which can characterize changes in the economic phases. For each approach, we use a grid-search-like method that selects the model that on average is better able to timely detect turning points of a given reference chronology. Using two different classes of models we show, on the two datasets previously mentioned, that it is preferable to use them in a complementary way for dating and detecting recessions and expansions. Indeed, each model provides a part of the analysis which can be strengthened by the other one as soon as the key instrument to detect a change in the data is not the same: one explains changes endogenously and the other exogenously. The paper is organized as follows. The next section describes the two datasets and presents the calibration done for each series using SETAR and MS modellings. The third section examines the capability of the two models to detect turning points in real time. The fourth section concludes.  Correspondence to: Monica Billio, Dipartimento di Economia, Università Ca’ Foscari Venezia, 30121 Venezia, Italy. E-mail: billio@unive.it Copyright © 2013 John Wiley & Sons, Ltd