Snowmelt and subsequent infiltration into potentially frozen soils are crucial hydrological processes in cold regions, with important implications for water availability and public safety. Robust operational forecasts are needed to anticipate these impacts. The objective of this study is to evaluate the representation of soil freezing in the Soil Vegetation and Snow (SVS) surface scheme used for hydrological forecasting at Environment and Climate Change Canada. Two approaches of contrasting complexity are considered. The first approach relies on a simple heat conduction method that was previously applied in the Canadian Prairies, whereas the second one combines a detailed multilayer snowpack scheme with a multilayer soil model solving heat and mass transfer. These two versions of SVS are evaluated at three sites in eastern Canada at latitudes of 47°N, 56°N, and 73°N. Local weather observations are used to drive the model at the site scale. Evaluation against in-situ measurements of soil temperature and water content reveal accurate performance of both approaches at the temperate site (47°N). At the subarctic and arctic tundra sites, the first approach fails to capture snow and frozen soil dynamics. Significant progress with the second approach is found only when properly representing (i) the properties of Arctic snow (vertical density profile and thermal insultation strongly affected by wind) and (ii) the effect of permafrost presence on soil moisture prior to freezing. The relative contribution of snow properties and soil water content to the ground thermal regime provides important insights to better represent frozen soil processes in cold regions.