For counterparty credit risk and CVA calculations, information about the future probability of default is essential. Typically, market-implied probabilities are used, derived from Credit Default Swaps (CDS) traded in the market. However, the CDS market is illiquid and often a CDS is not traded for the counterparty in question. Using proxy methods, i.e. CDS information for similar businesses in similar geographical locations, does not accurately reflect the true default probabilities.
This white paper explores how default probabilities can be predicted with deep learning, using information such as accounting data, trading activities, local and global economic data, credit ratings, etc. A deep neural network is trained using historical information, using CPUs or GPUs, and can then be used to accurately predict default probabilities. Backtesting showed that this method gives significantly more accurate results than the traditional CDS method.