Conclusion
Predicting records is of paramount importance for science - with use cases ranging from predicting temperature records to forecasting the development of new technologies.
In this article, we have developed a Bayesian account for predicting records. Our work is quite general and extends to all situations in which the underlying attempts follow an i.i.d. distribution.
We have shown that our approach is competitive with a previous frequentist method by \citep{tryfos_forecasting_1985} in terms of MSE for predicting records in 6 athletic events.
We have also investigated the predictive accuracy of different attempt distributions on data for 11 athletic events. While the conclusions aren't clear-cut, the evidence suggests that the Weibull distribution results in a better fit for the data in terms of a log-likelihood loss.
Using our method we have forecasted the records for these 11 athletic events for the period of 2022 to 2032. We hope other researchers will be able to use this as a basis for comparison.
We have released an open-source PyMC3 package accompanying this paper,
fmax. Researchers and practitioners can use this framework to model record distributions using their own data.