No formal sources unfortunately, it's similar to using a linear regression model trained on historical data for inference (with some cross-validation/penalizati

No formal sources unfortunately, it's similar to using a linear regression model trained on historical data for inference (with some cross-validation/penalization to make sure we're not just finding signals in the noise). We know for sure that there's a ton of bias (in the bias-variance sense), and that the predictions won't be nearly as good as a blackbox model, but we may glean some useful insights out of it. Same idea with decision trees, in the past I've used them to find what factors influence customer support scores on online tickets, and we found some surprising and useful insights that were used for process improvements.