Deep Learning Model Error Bounds in Credit Derivatives Valuation

A significant challenge in financial engineering is the rapid and accurate calculation of the value of credit derivatives; risk management scenarios often demand the valuation of millions of contracts. While deep learning models theoretically offer the desired speedup, they are often regarded as black-box solutions, lacking two critical requirements for acceptance by financial practitioners and regulators: interpretability and uncertainty measurement. Multilayer perceptron models demonstrate impressive performance against the industry-standard pricing models for credit default swap (CDS) index and CDS option contracts—with prediction errors generally well below 1% while delivering speedups of 418x and 235x, respectively. Leveraging differential machine learning, we propose a novel method for constructing intuitive, conservative first-order gradient–based error bars, thereby addressing both model interpretability and uncertainty. Error bars for CDS index and CDS option captured 75% and 89% of ground truth prices, respectively, with error bar widths less than or equal to 20% of the corresponding ground truth price in 80% and 86% of cases, respectively.