Verification of snowpack stability and avalanche danger is a prerequisite for the improvement of avalanche forecasting and the development of models. Although avalanche danger is based on snowpack stability, little is known about the variation of regional snowpack stability at a given danger level. Verification can be done by observation of avalanche occurrence and/or stability tests. To verify avalanche forecasts, and to get a more detailed picture of regional snowpack stability patterns at different danger levels, a large-scale field study has been performed. On four occasions during the winter of 2002 stability data were collected in the region of Davos. During each 1- to 3-day sampling period between 50 and 70 full snow profiles with rutschblock tests were recorded, primarily on shady slopes. At the same time the avalanche danger was estimated based oil observations in the field. For analysis the profiles were assigned to one of five stability classes: Very Poor, Poor, Fair, Good, Very Good. Relating the stability to the prevailing (verified) danger level showed distinct patterns of stability, At the danger level Low 90% of the profiles were rated as Good, or Very Good, whereas at the danger level Considerable, more than 50% showed Poor or Very Poor stability. The coefficient of variation was about 20% independent of the danger level. Significant differences in aspect and elevation existed. Some of the variation could be explained by differences in snow depth and snowpack consolidation (ram resistance). A preliminary analysis of failure layers showed that a persistent weak layer of large faceted crystals above a crust could be found in the majority of the profiles during certain periods. Despite a generally large variation in stability this weak layer was very widespread, and strongly influenced snow stability during the course of the winter, even 2 months after its formation. Due to the stability variation found in this study, verification of avalanche forecasts based on single stability tests cannot be recommended.