Precipitation forecasts are of large societal value in the tropics. Here, we compare 1–5 day ensemble predictions from the European Centre for Medium-Range Weather Forecasts (ECMWF, 2009–2017) and the Meteorological Service of Canada (MSC, 2009–2016) over 30°S–30°N with an extended probabilistic climatology based on the Tropical Rainfall Measuring Mission 3B42 gridded dataset. Both models predict rainfall occurrence better than the reference only over about half of all land points, with a better performance by MSC. After applying the postprocessing technique Ensemble Model Output Statistics, this fraction increases to 87% (ECMWF) and 82% (MSC). For rainfall amount there is skill in many tropical areas (about 60% of land points), which can be increased by postprocessing to 97% (ECMWF) and 88% (MSC). Forecasts for extremes (>20 mm) are only marginally worse than those of occurrence but do not improve as much through postprocessing, particularly over dry areas. Forecast performance is generally best over arid Australia and worst over oceanic deserts, the Andes and Himalayas, as well as over tropical Africa, where models misrepresent the high degree of convective organization, such that even postprocessed forecasts are hardly better than climatology. Skill of 5-day accumulated forecasts often exceeds that of shorter ranges, as timing errors matter less. An increase in resolution and major model update in 2010 has significantly improved ECMWF predictions. Especially over tropical Africa new techniques such as convection permitting models or combined statistical-dynamical forecasts may be needed to generate skill beyond the climatological reference.