International Journal of Forecasting 30 (2014) 898–917 Contents lists available at ScienceDirect International Journal of Forecasting journal homepage: www.elsevier.com/locate/ijforecast Predicting recessions with a composite real-time dynamic probit model Christian R. Proaño a,∗ , Thomas Theobald b,1 a The New School for Social Research, New York, NY, United States b Macroeconomic Policy Institute (IMK), Düsseldorf, Germany article info Keywords: Dynamic binary choice models Out-of-sample forecasting Yield curve Real-time econometrics abstract In this paper we propose a composite indicator for real-time recession forecasting based on alternative dynamic probit models. For this purpose, we use a large set of monthly macroeconomic and financial leading indicators from the German and US economies. Alternative dynamic probit regressions are specified through automated general-to- specific and specific-to-general lag selection procedures on the basis of slightly different initial sets. The resulting recession probability forecasts are then combined in order to decrease the volatility of the forecast errors and increase their forecasting accuracy. This procedure features not only good in-sample forecast statistics, but also good out-of-sample performances, as is illustrated using a real-time evaluation exercise. © 2014 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved. 1. Introduction As is widely acknowledged, the timely and accurate prediction of turning points in the business cycle is one of the most policy-relevant aspects of macroeconomic fore- casting. However, this task is also one of the most chal- lenging. Not only are there many potential nonlinearities at the onset of a turning point in economic activity, there is also a significant level of uncertainty around macroeco- nomic data at the current edge, 2 in addition to the model uncertainty inherent in all applied work. With respect to mitigating the model uncertainty prob- lem, Bates and Granger (1969) were among the first to propose a combinatorial approach. They showed that the ∗ Corresponding author. E-mail addresses: christian.proano@gmail.com (C.R. Proaño), thomas-theobald@boeckler.de (T. Theobald). 1 Also Free University (FU) Berlin, Germany. 2 The current edge is defined as the last observation(s) of a certain vintage of macroeconomic data. These observations are usually subject to future data revisions. They are also called end-point data. inclusion of inferior ex-ante forecasts could increase the predictive power of the best ex-ante forecasts if the in- ferior forecasts contained some novel information. More recently, Timmermann (2006) also emphasized the use- fulness of forecast combination due to (1) diversification, (2) structural breaks, (3) misspecification of individual forecasts, and (4) systematic differences in the individual loss functions. In contrast, methods for reducing the uncertainty inherent in end-point data are less developed. Pesaran and Timmermann (2005) have stressed the urgent need to develop robust interactive systems of model specification and evaluation which are designed explicitly to work in real time, as ‘‘by setting out in advance a set of rules for observation windows and variable selection, estimation, and modification of the econometric model, automation provides a way to reduce the effects of data snooping and facilitates learning from the performance of a given model when applied to a historical data set’’ (Pesaran & Timmermann, 2005, p. 212). Binary response models have been used extensively in the literature for the prediction of business cycle turning http://dx.doi.org/10.1016/j.ijforecast.2014.02.007 0169-2070/© 2014 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.