Uber – experimental data – quantile regression

Uber – observational data

EconML Microsoft Research – https://econml.azurewebsites.net/spec/flowchart.html CONFIDENCE INTERVALS/MODEL SELECTION
The MetaLearner and DRLearner estimators offer the choice of any ML estimation model in all stages and allows for model selection via cross validation. This enhances flexibility, but because the sample data is used to choose among models it is impossible to calculate honest analytic confidence intervals. Moreover, most ML estimation approaches introduce bias for regularization purposes, so as to optimally balance bias and variance. Hence, confidence intervals based on such biased estimates will be invalid. For these models it is still possible to construct bootstrap confidence intervals, but this process is slow, may not be accurate in small samples and these intervals only capture the variance but not the bias of the model.

『相対化する知性』より