Transitioning from Econometrics to ML (predicitions)

[Vivian Dilmo]

Hi all,


I was trained as an econometrician. Nowadays, I find myself more and more required to do predictive modeling (train-test. split, feature selection etc.).

I have two things that I still need clarification on.
1. Where do the p-values go? It is very strange for me to report results that do not come with CIs.
2. I don't know when to choose Econometrics setup and to use the ML setup. Are there situation you prefer to use one over the other? Is there value in doing both? How do you compare results?

I would love to hear your thoughts on this matter.

Thanks,
Viv

[Jared Greathouse]

There's no difference between the two (in the sense that machine learning folks who study econometrics ask similar questions to us). This paper extends the synthetic control method to a more familiar ML setup by using singular value thresholding to denoise one's donor matrix and then predict the counterfactual in the post-intervention period. It is a mix of standard econometrics and ML.

in other words, they're not necessarily mutually exclusive. They can go together. Whether it makes sense to use elements from different traditions depends on your question and goals.

oh, and don't report p values. They don't tell me anything about if your design or model (in that order) is good or bad, and they're misunderstood and misused.

because I'm concerned with causal inference which is oftentimes based on prediction, I find myself playing with ML a lot. You may wanna read this book (don't worry it's totally free)