An Extreme-Value Model for Predicting the Results of Horse Races

Results of horse races in 1979-80 are analysed to see if the extreme-value distribution can model the times to run horse races: horses with the same bookmakers' odds of winning are classed as one group; the distribution of times for each group is extreme-value with location parameter β depending on the win odds; and the scale parameter θ is common to all groups. The model predicts that the win probability p for the group is p = exp ((β0 - β)/θ) for some constant β0, and this is borne out by the data. Only the tail of the empirical distribution functions is consistent with the model; however, it is essentially this tail which determines the win probabilities, so if the aim is to do just that the model will serve a useful purpose.