An extensive literature explains recession risks using a variety of financial and business cycle variables. The problem of selecting a parsimonious set of explanatory variables, which can differ between countries and prediction horizons, is naturally suited to machine-learning methods. The current paper compares models selected by conventional machine-learning methods with a customised algorithm, ‘Doombot’, which uses ‘brute force’ to test combinations of variables and imposes restrictions so that predictions are consistent with a coherent economic narrative. The same algorithms are applied to 20 OECD countries with an emphasis on out-of-sample testing using a rolling origin, including a window for the Global Financial Crisis. Despite the imposition of additional restrictions, Doombot is found to the best performing algorithm. Further testing confirms the imposition of judgmental constraints tends to improve rather than hinder out-of-sample performance. Moreover, these constraints provide a more coherent economic narrative and so mitigate the common ‘black box’ criticism of machine-learning methods.