Partitioning precipitation into rain or snow is an important aspect of hydrologic and climatological modelling, affecting a wide variety of downstream processes. Conventional methods typically rely on surface temperature, either as a strict threshold or to inform a probability function which predicts a 0% chance of rain at one temperature to 100% chance of rain at another. However, recent studies have shown that variables such as wind, pressure, humidity, and atmospheric profiles of temperature can all have a significant effect on precipitation phase at the surface. This study utilized CloudSat-derived estimates of precipitation phase and associated environmental variables ground-truthed at ECCC stations across Canada to build an improved statistical parameterization of precipitation phase. Our results showed that using a random forest model with atmospheric profiles of wetbulb temperature in addition to surface wetbulb temperature, elevation, and wind resulted in a probability of detection of 97.8% across the -1 to 4°C temperature interval, compared to a probability of detection of below 80% across the same interval for conventional methods. Importantly, the random forest parameterization was able to generalize spatially, performing well on stations it had not been trained on. Binning the data by Sturm’s snow classes did not provide any further improvement in skill, indicating that a model trained on all available data best captures spatial variability in rain-snow partitioning across the study area.