Paired bootstrap

Also written: bootstrap, paired_bootstrap

Resample whole customers with replacement many times, score both models on each resample, and look at the spread of their difference. Pairing removes shared luck, and an interval that includes zero means the models are effectively tied.