By Yimou Li, Zachary Simon, and David Turkington.
Published in the Journal of Financial Data Science, Winter 2022.
We put the stock-selection skill of machine learning models to the test, with an intense focus on making sure their selections are both investable and interpretable - and therefore, believable.
Imagine a line that shows remarkably stable investment performance outpacing the historical returns of nearly every mutual fund and known quantitative strategy. In a nutshell, this is the typical pitch for investment models based on machine learning. There are plenty of reasons to be skeptical and to keep machine learning on the sidelines of actual investment decisions. We argue that if complex models generate investable and interpretable results, they can be used with confidence alongside good, human judgement. We calibrated random forest, boosted trees and neural networks to predict stocks based on well-known factors and regime variables, and applied a new technique called the Model Fingerprint to show the logic behind each model’s stock picks. In the end, the machines learned many of the same rules as their human creators, but occasionally they landed on a less-obvious relationship that made us pause and think.