Evaluation of Transformer-Generated Proxy Credit Default Swap Spreads

Abstract

Following the 2007-2008 financial crisis, accurately assessing Counterparty Credit Risk (CCR) has become very important in the financial industry, especially in the Over-The-Counter (OTC) derivatives market. Credit Valuation Adjustment (CVA) integrates CCR into the pricing of OTC derivatives like Credit Default Swaps (CDSs) by using precise Probability of Default (PD) estimations, typically derived from CDS spreads. When no liquid CDS spreads are available one uses proxy spreads instead. Tradi tional linear regression models for generating these, like the Nomura cross-sectional model, are commonly used in the industry but show problems in certain market environments. Therefore, this thesis evaluates the effectiveness of a Machine Learn ing (ML) model known as the Transformer for generating proxy CDS spreads and compares its outputs to the Nomura model. Using actual, liquid market data for Western European financial companies, the Transformer model was trained to generate proxy spreads for five credit rating cate gories: AA, A, BBB, BB, and B. Results indicate that the Transformer model signif icantly outperforms the Nomura model, particularly in higher-rated categories, by generating spreads that are better aligned with the liquid market spreads they aim to simulate. Despite the promising results, further tests and analyses are needed to confidently be able to declare the Transformer model’s superiority and potentially take it to production. This includes hyperparameter optimization, data diversifica tion, and model interpretability improvements.