# Shortfalls of using Machine Learning algorithms to predict counterfactual outcome of the treated

To see the effect of a treatment we would like to know the difference between a person in a world in which she received the treatment and one in which she does not. Since we never observe the counterfactual performance of the treatment, we resort to different assumptions (similar pre-treatment trends), methods (synthetic controls), etc. to find the effect of the treatment.

Since one of the main advantages of the ML algorithm is its ability to predict with high efficiency, I was wondering if it can be used to forecast the counterfactual outcome of the treated for the post-treatment period. In my mind, this is sort of the unconfoundedness assumption on steroids.

What may be the shortfalls of such an exercise?

• I am not quite sure what you mean by surrogate market. My thinking is in the line of training a model of $y_t$ on a wide range of covariates $X_t$ for $t<T$ where $T$ is the start of the treatment period. Use this trained model to forecast a $\hat{y_t}$ for $t \geq T$ using $X_t$ for these time periods. Then use $\hat{y_t}-y_t$ to estimate the effect of the treatment. – FightMilk Jan 7 at 21:24