We forecast recession probabilities for the United States, Germany and Japan. The predictions are based on the widely-used probit approach, but the dynamics of regressors are endogenized using a VAR. The combined model is called a ‘ProbVAR’. At any point in time, the ProbVAR allows to generate conditional recession probabilities for any sequence of forecast horizons. At the same time, the ProbVAR is as easy to implement as traditional probit regressions. The slope of the yield curve turns out to be a successful predictor, but forecasts can be markedly improved by adding other financial variables such as the short-term interest rate, stock returns or corporate bond spreads. The forecasting performance is very good for the United States: for the out-of-sample exercise (1995 to 2009), the best ProbVAR specification correctly identifies the ex-post classification of recessions and non-recessions 95% of the time for the one-quarter forecast horizon and 87% of the time for the four-quarter horizon. Moreover, the ProbVAR turns out to significantly improve upon survey forecasts. Relative to the good performance reached for the United States, the ProbVAR forecasts are slightly worse for Germany, but considerably inferior for Japan.