Before the match the model relies on historical data: each team's winrate on every map in the pool over 3 months, current form and head-to-head. Once the veto happens, the prediction is recalculated for the real map pool.