Causal modeling is an umbrella term for a big selection of methods that allow us to model the results of our actions on the world.
Causal models differ from traditional machine learning models in plenty of ways.
Crucial distinction between them stems from the undeniable fact that the data contained in observational data used to coach traditional machine learning machinery is — normally — insufficient to consistently model the results of our actions.
The result?
Using traditional machine learning methods to model the outcomes of our actions leads — in principle — to biased decisions.
A superb example here is using a regression model trained on historical data for marketing mix modeling.
One other one?
Using XGBoost trained on historical observations to predict the probability of churn and sending a campaign if the expected probability of churn is bigger than some threshold.
