Recently, the analysis of e-sport data has particularly attracted the attention of machine learning researchers. We contribute here by proposing, on the MOBA (Multiplayer Online Battle Arena) game DotA2 (Defense of the Ancients 2), an analysis of the relevance of geometric clues relating to the behavior of the players. The geometric clues we have chosen concern the trajectories of the players and the characteristics of the polygon described by the players of a team. More precisely, for the polygon clues, we computed, for each second of the game: the convex area, the diameter, the grouping capacity (average of the distances to the barycenter), the inertia (second order moment of the distances to the barycenter) and finally the distance to the opponent's fortress. During a regression on the number of alive heroes, the learning performance of an RNN (precisely a LSTM) fed by our simple clues allowed us to evaluate their relevance in describing this complex game.