Transformer networks are the backbone of several recent techniques of trajectory forecasting. But how the simple transformer networks do perform the forecasting? Do they model multimodal distributions? Obviously yes, but this work show how the single modalities represent different types of trajectories. A joint work by Luca Franco, Leonardo Placidi, Francesco Giuliari, Irtiza Hasan, myself and Fabio Galasso. Enjoy <<<here>>> just published on Pattern Recognition journal