In the previous blog, we created a quantile regression model that allowed us to estimate, in-running, a home team’s victory probability, and to create in-running confidence intervals for the home team’s final margin.
We evaluated that model based on a variety of performance metrics calculated using a 50% holdout sample from the original data set, which included games spanning the 2008 to 2016 period.
But nothing really measures a model’s performance better than a completely fresh data set from a non-overlapping time period, and in this blog we’ll be running the same metrics, but for games spanning the 2017 to 2019 period (up to and including the first week of the 2019 Finals). That’s 616 games entirely unseen by the model.
Read More