ML vs Statistical Analysis

I am an assistant professor in accounting and I am seeing more and more machine learning techniques are used in academic papers in our field. But I am confused why we still use simple linear regressions with R-Squared of 20-50%if neural networks can predict an outcome a lot better?

I appreciate it if you can shed some light on this.

Good question.

There are a few reasons I can think of, but the answer will really depend case by case since I don’t know the accounting field well.

  1. They don’t know that neural networks can be applied to improve predictions
  2. They don’t know how to apply neural networks and more advance ML techniques
  3. It’s the benchmark standard because everyone understands it immediately, everyone can compare results immediately, therefore it’s the best technique to get the point across
  4. Neural networks cannot predict outcomes better. A bigger, more complex model doesn’t always help if your problem doesn’t call for it; sometimes linear regression works just fine. For complex domains like NLP and image processing neural networks work well, but if you’re working with, say, only a small set of features then linear regression might be good enough.


1 Like

Thanks Nick for getting back on this. I think it is more #4.