Our solution is a 2-stage gradient boosting decision tree model.
In the first stage, the model aggregates a set of sampled data to create features.
We trained two models in this stage: one predicts whether an event is a Higgs boson signal or not, while the other predicts TES and JES, which are part of the nuisance parameters.
In the second stage, the aggregated features generated by the two models are used to estimate the 68% confidence interval of mu.
We trained two models in this stage as well: one using quantile regression and another employing the regression-as-classification approach.
By merging the predictions from these models through post-processing, we were able to enhance the accuracy of our solution.