"How" models in finance
"How" models in finance
Lin William Cong's cool papers
Howdy!
This is my first real post. I’ve been trying to figure out what I should write, and I decided to just pump this out in the forty minute train ride this evening.
This is a free newsletter and will continue to be, I just want to get some thoughts on paper and explore things that interest me.
I did my PhD in finance, so naturally my first post is a finance thing. It’s pretty rough but at this point I want to try to get into a writing rhythm. Let me know what you liked, didn’t like, etc. I’m interested in lots of stuff so I’m happy to cover whatever.
Cameron
I recently had the opportunity to attend a talk by Lin William Cong, a finance professor at Cornell. Dr. Cong has a lot of very “futuristic” papers — things like crypto, reinforcement learning, as well as some more standard papers on corporate finance and asset pricing.
I like Lin’s papers because they are often very much focused on “how” problems rather than “why” problems. Financial economics as a literature is often primarily focused on why things are the way they are. Why do asset prices tend to move together? Why do firms issue equity when they do, and why do they pay dividends?
I would categorize Lin’s contemporary work as belonging to the “how” camp. It takes standard tools in machine learning and tries to apply them with a mostly-principled approach from a finance researcher. For example, in Lin’s reinforcement learning paper with coauthors, they model reinforcement learning in portfolio selection. They end up using several transformer/attention-style neural networks as well as data from firm balance sheets.
Lots of the reception to this paper in private conversation has been something along the lines of “why do we care”? To some extent I am sympathetic to this — reinforcement learning is strange, and not how the world currently works, and it doesn’t actually help us to understand about why financial markets work the way they do.
However, this is (in my view) extremely short sighted. Financial academics and practitioners have been diverging fairly steadily for decades now. Certainly, practitioners use factor models and talk all the time about “alpha” or whatever, but the big funds are no longer shackled to the same kinds of “why” problems that face massive financial institutions. Renaissance Technologies may not give a shit about why smaller stocks outperform larger stocks or whatever — they just want to be right.
And this is kind of what Lin tends to work on, at least recently. There are some issues I have with the reinforcement learning paper, for sure, but I think it’s important that academics start paying slightly more attention to how these models work and what we can learn from them. Lin & coauthors have a relatively standard neural network that I imagine most big financial firms can train relatively easy. Lin’s paper is one of very few papers using neural networks on asset prices and related data. I find it to be comical that, with all of financial academia’s understanding of financial markets, they still haven’t quite dug into the more advanced tools we’ve seen lately. Ce’st la vie.
Lin’s paper this time was “Uncommon Factors for Bayesian Asset Clusters”, which you can find at SSRN. I personally loved this paper. The primary motivation seemed to be to do interpretable analysis of “risk factors”, which in finance is a fancy term for “things that explain asset prices”. We’ve had a long discussion as an academic community over many decades what factors are real, which ones are fake, etc. The general consensus as I understand it now is that most are generally fake, and the set of factors that are meaningful is fairly small (on the order of 5-10).
In this paper, Lin & coauthors take a very simple analytic tool, decision trees. Decision trees are a very simple tool for partitioning your data into smaller blobs. This has been done before and isn’t particularly novel in and of itself.
The cool component here is that Lin & the gang add some very interesting Bayesian stuff to their model1. They say — okay. The decision tree will sort assets into different buckets. Then, within these buckets, let’s assume that each cluster responds differently to each factor, and let’s further assume that there’s a non-trivial likelihood that any of the factors are meaningless. They end up using a handful of conjugate priors (little magic tricks for Bayesians) and a Gibbs sampler.
They end up finding that lots of factors are common (the market return, for example, which we have a very good theoretical background to believe). Importantly, however, they end up finding that some factors matter differently for different clusters. There’s even a “junk” cluster full of assets that have historically been difficult to price.
The paper’s still definitely a work in progress, though, and I’m interested in seeing what happens with it.