From our results, we cannot draw any strong conclusions about whether one of the underlying ­distributions tends to perform better than the others. However, the model using a Weibull attempt distribution appears to give a good balance between predictive accuracy and maximizing the average log probability.
These results match common sense - the Weibull distribution arises whenever we take the minimum out of a series of samples from an i.i.d. distribution with a finite lower bound \citep{FrchetSurLL}. At each time step, the attempt to beat the record reflects the minimum time out of all runners that year. Since runs cannot obtain a time less than 0, the conditions are met for a Weibull distribution.