I’ve always wanted to build sports simulations with ML, and trajectory forecasting is fundamental to that. But when I looked into recent SOTA models, many achieved good per-agent prediction metrics yet struggled to produce coherent and plausible joint futures across agents (players + ball). So I built CausalTraj to address this, and it is now open-sourced.