Abstract. The Alpine region is strongly affected by torrential floods, sometimes leading to severe negative impacts on society, economy, and the environment. Understanding such natural hazards and their drivers is essential to mitigate related risks. In this article we propose a statistical method based on a mere discriminative index to objectively isolate the atmospheric variables associated with torrential events with long return periods. The method is applied to the Grenoble Metropolitan area in the Northern French Alps using a list of dates of damaging torrential events since 1851. We consider seven atmospheric variables that describe the nature of the air masses involved and the possible triggers of precipitation, using both 20CRv2c and ERA5 reanalyses. Two spatial scales are considered – a local scale (the Grenoble conurbation) and a regional scale (the French Alps) –, in order to study the spatial variability of the atmospheric signature. This analysis is done conditionally on the main types of generating atmospheric circulation derived from Lamb weather classes, namely the North-West, Southeast-Southwest and Barometric Swamp classes. The results show that a simple discriminative index – the so-called silhouette index – is able to isolate the variables associated with torrential events, by objectively determining which of them differ particularly from the climatology at the dates of torrential events. All atmospheric variables turn out to be less discriminant for torrential events before 1950 according to 20CRv2c – this is likely linked to 20CRv2c issues over the remote past. For the post-1950 period, in the North-West class – which is both the most frequent class generating torrential events and the best discriminated – humidity and particularly humidity transport (IVT) plays the greatest role. In the Southeast-Southwest class – the second most frequent class generating torrential events–, instability potential (CAPE) is mostly at play. In the Barometric Swamp class both humidity (PWAT) and instability (CAPE) are discriminant –and even more at the local scale–, showing more mixed situations generating torrential events in this class. The gain in prediction provided by the discriminant variables is substantial. Depending on the class, torrential events are 4 to 14 times more likely when the respective discriminant variables are extreme (typically above their 0.95-quantile). Although the results are likely to be region-dependent worldwide, the methodology used in this article is generic and could be used elsewhere to find the most discriminating atmospheric variables – provided a list of flooding dates is available. It is also remarkable that the same atmospheric variables with the same discriminative power are found whether we consider them at local or regional scale. This means that, although torrential events are triggered by very local precipitation, the atmospheric signature for such events is actually much wider. Thus, although the present study applies to a small region of the Northern French Alps, it is reasonable to presume that similar results would apply to other torrential catchments in the French Alps.