Using Clustering Neutralization to Improve Value Strategy

Note: This is a republication of a full write-up of the machine learning research to understand and improve systematic value strategy, which was originally shared in my 2019 Q4 letter [Link]

I’m also start trying Medium. It is published on Medium too. [Link]


This article is a collection of thoughts and experiments of my personal endeavor from a quantitative perspective to understand why “Value” as an investing style didn’t perform compared to other styles for an extended long period of time. It consists of mainly two parts: 1) review of others’ relevant researches & 2) an experiment of using Clustering to explore reasons for such phenomenon and potential ways to improve systematic value strategies.

“Value” means differently for different investors. It is important to note that “Value” in context of this write-up mainly refers to the “value” factor, e.g. Fama-French HML factor or variations of it. In other words, it is a systematic long-short strategy longing a group of the cheapest stocks and shorting a group of the most expensive stocks.

My main hypothesis is that GAAP/IFRS currently does not do a as well job in representing fairly companies’ value as historically. Thus “Value” may not work on broad stock universe yet may still work in certain types of stocks. To classify stocks into such “types”, I tried to use Clustering, an unsupervised machine learning technique, to test if it works.

You can also download the pdf here:  [Link]


Continue reading