So I ran a backtest of a prototype stat arb system on QC. Link with code and results is here:
https://www.quantconnect.cloud/backtest/3f856cd9dcc28e4efc07dc818cddc08e
I was pretty happy with the initial result – but the performance falls apart with different parameters, like a longer backtest period. I guess the period in the above run was just a favorable market regime.
The code is a mixture of my own work and work done by an AI agent co-developer. I handled a lot of the basics, like universe selection and neutralization logic. I asked the AI to implement ideas from Zura Kakushadze’s paper here:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2478345
I have a feeling the AI didn’t do as deep of a job as is possible. I think I will have to thoroughly read the paper and adapt the ideas into code on my own. (ChatGPT pretty much agrees haha)
The paper is about adding layers of processing to signal neutralization. This is important because it is a possible solution to a problem I run into when trading off of neutralized alpha alone.
I’ve recreated many of the alphas from Zura Kakushadze’s “101 Formulaic Alphas” from a data wrangling perspective, but they simply don’t work with only neutralization to take positions with.
For reference, the 101 alphas paper is here:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2701346
So all of this was inspired by the time I spent with WorldQuant as a contract Research Consultant. But it seems WorldQuant’s training materials left some things out. Probably by design to hinder competition. Would make sense to me if that was the case anyway. Kakushadze’s work on portfolio optimization also clues me in on this.
So here I am – thinking about optimization matters. I definitely have to brush up on linear algebra though.