This paper measures the economic value of conditioning stock and bond return forecasts on recession dummies. Motivated by recent studies, I restrict stock returns to only be predictable in recessions and bond returns to only be predictable in expansions. These restrictions generate notable increases in risk-adjusted performance in a standard setup with mean-variance utility and a simple bivariate regression model. I also let regression coefficients switch across recessions and expansions without restricting them. This extension performs particularly well when combined with forecast averaging. The risk-adjusted returns are robust over time and prevail using real-time recession signals. However, high accuracy of business cycle turning points is essential and seemingly small tweaks to the classification rule can have large economic implications. Goyal and Welch (2008) forcefully argue that simple regression models cannot forecast stock returns out of sample. Thornton and Valente (2012) reach similar conclusions for bonds. However, these studies do not condition on the business cycle. Henkel et al. (2011) let coefficients be regime-dependent and document that stock returns are predictable but only in economic recessions.1 Andreasen et al. (2016) find the mirror image for bond returns as they are only forecastable in expansions. Thus, return predictability seems to be highly dependent on the business cycle and, moreover, it is asymmetric across stocks and bonds. This motivates why I extend standard regression models to take macroeconomic conditions into account. Markov switching models are one of the most popular approaches to capturing state-dependence in returns. These models let the return distribution depend on unobservable realizations of a Markov chain without imposing an economic interpretation. Examples of papers taking this approach are Guidolin and Timmermann (2007), Guidolin and Hyde (2012), and Henkel et al. (2011). Henkel et al. (2011) find a strong connection between the inferred states of their return forecasting model and the business cycle. On one hand, Markov switching models provide flexibility by not imposing that the regimes are related to specific economic events. On the other hand, they are highly nonlinear and rely on potentially unstable numerical estimation methods which may hamper their usefulness in out-of-sample forecasting. If the underlying states are related to recessions, which are short-lived and infrequent, more robust techniques could be useful. I therefore instead use observable recession dummies which allows me to estimate the models using simple linear regressions. This choice is natural since most of the studies that identify differences in predictability across states of the economy have done so using recession dummies (see citations above). I do also show results for Markov switching models but find that, in contrast to the dummy switching strategies, they do not improve on constant coefficient models for real-time trading. Further, tying financial markets explicitly to recessions is in the spirit of the macro-finance asset pricing literature as summarized in Cochrane (2017).