CDO calibration via Magnus Expansion and Deep Learning

TL;DR

This paper enhances CDO calibration accuracy by combining Magnus expansion for efficient SPDE solving with deep learning for initial default distance inference, achieving good market data fit and high computational efficiency.

Contribution

It introduces a novel combination of Magnus expansion and deep learning techniques for improved CDO calibration and SPDE solution efficiency.

Findings

Magnus expansion provides high-accuracy solutions for large basket SPDEs.

Deep learning significantly improves initial default distance inference from CDS quotes.

The proposed methods enable highly parallelized, GPU-accelerated calibration fitting.

Abstract

In this paper, we improve the performance of the large basket approximation developed by Reisinger et al. to calibrate Collateralized Debt Obligations (CDO) to iTraxx market data. The iTraxx tranches and index are computed using a basket of size $K= 125$. In the context of the large basket approximation, it is assumed that this is sufficiently large to approximate it by a limit SPDE describing the portfolio loss of a basket with size $K\rightarrow \infty$. For the resulting SPDE, we show four different numerical methods and demonstrate how the Magnus expansion can be applied to efficiently solve the large basket SPDE with high accuracy. Moreover, we will calibrate a structural model to the available market data. For this, it is important to efficiently infer the so-called initial distances to default from the Credit Default Swap (CDS) quotes of the constituents of the iTraxx for the large basket approximation. We will show how Deep Learning techniques can help us to improve the performance of this step significantly. We will see in the end a good fit to the market data and develop a highly parallelizable numerical scheme using GPU and multithreading techniques.