Quality control tools for probabilistic forecasts of football match results

In football match forecasting, much attention has been devoted to statistical approaches of modelling the outcomes either as Win, Draw & Loss (WDL) categories, or as score lines. The second step lies in the evaluation of such forecasts. At first, emphasis is about computing synthetic criteria such as MSE and more generally probability scoring rules as introduced and popularized by weather forecasters. We can go further by looking at elementary components of forecast quality such as Reliability, Resolution and Discrimination both numerically and graphically. This presentation briefly reviews this validation step both theoretically and practically in the light of standard predictions of outcomes for the UEFA Champions League (C1).

Win forecasts of the C1 group stage matches derived from a simple Poisson regression model and Bookmaker odds derived probabilities are well calibrated with good resolution and discrimination properties for both Home and Away outcomes. On the contrary, forecasts of draws with these procedures highly lack refinement, resolution, and discrimination.

This talk was given by Jean-Louis Foulley from the Université de Montpellier, affiliated with IMAG.