Forecasting business-cycle turning points with (relatively large) linear systems in real time

The detection of business-cycle turning points is usually performed with non-linear discrete-regime models such as binary dependent variable (e.g., probit or logit) or Markov-switching methods. The probit model has the drawback that the continuous underlying target variable is discretized, with a considerable loss of information. The Markov-switching approach in general presupposes a non-linear data-generating process, and the numerical likelihood maximization becomes increasingly dif cult when more covariates are used. To avoid these problems we suggest to rst use standard linear systems (subset VARs with zero restrictions) to forecast the relevant underlying variable(s), and in a second step to derive the probability of a suitably de ned turning point from the forecast probability density function. This approach will never fail numerically. We also discuss and show how this approach can be used in real time in the presence of publication lags and to capture features of the data revision process, and we apply the method to German data; the event of the recent Great Recession is rst signalled in June 2008, several months before the of cial published data con rms it (but due to publication and recognition lags it is found after it already began in reality).