Are Deeper Neural Networks Always Better For Finance?

Introduction

We read about extremely deep and powerful neural networks everyday. Surely someone has to ask, does it work the same in Finance?

I am reviewing a highly cited, very famous paper on complexity on machine learning for equity return predictions. The core finding upends what most people assume about conventional understanding to model complexity: the best-performing neural network is a very shallow network with no more than 3 layers.

Everything you have heard about deep learning seem to runs backwards for return prediction. Why?