Back in the days when pricing a CDO/Baskets/LCDS were hot and perhaps get you a quant job doing these type of research. Days when Lehman, Merrill, and Bear publish research papers about credit derivatives pricing with armies of quants behind the calibrating the results and trying to help us intuitively understand the concept while accepting the assumptions could be garage in and garage out. I had my contributions in this as well ending up as a standard 6 hour courses in financial engineering programs. Pricing CDO/Basket Products course by David Lu
This was another article co-authored with my classmate and professor. This was after the crisis where try to crack the pricing of a Loan Credit Default Swap Ong_Li_Lu_LCDS Journal of Credit Risk
Interesting “a-ha” moments in life may many times happened after a “signal” made by a person or event. I started to think about reinforcement learning in 2005 when I first started to think about robust portfolio selection. On a side note, Andrew Ng wrote that he started to think about optimal control in his Phd Thesis back in 1998. We will talk about relationships in optimal control and reinforcement learning in later posts.
So when did I actually thought about supervised learning or machine learning? It happened in 2009 when I actually interviewed for a position at large rival investment bank at time square. The interviewer asked me how would I handle illiquid CDS spread. I told him I’d map to a liquid CDS curve. How would you price the curve? I told him I would map it with industry sector, region, ratings, capital structure, recovery rate or to an index. He didn’t seemed satisfied and I didn’t get the job. After the interview, though it was a front office COO role on the description, I notice he was a quant. As much as I made the mistake of not looking him up on Linkedin beforehand, it had me thinking about what he wanted was my thoughts on how I am going to model this spread.
We can estimate illiquid CDS spreads, or in other words to estimate unobserved CDS spreads for companies with illiquid debt based on observed spreads and other characteristics of other, liquid companies, we can use supervised learning or classification methods.
Our inputs can be the X features for the company: ratings, industry sector, geographical sector,financial ratios, implied volatilities, and Moody’s EDF (Expected Default Frequency). Take these observable features and train for an output for the observed universe of liquid CDS. We can use Decision Trees, Support Vector Machine or Neural Networks to handle this problem. It turns out that support vector machines are pretty good at handling these type of problems. If the readers interested of these type of problems it is also used in internal ratings modeling where this can be applied in regulatory issues such as Basel which the rules wouldn’t allow you to use rating agencies ratings.