Deep surrogates for finance: With an application to option pricing

• 成果类型: Article
• 署名作者: Chen, Hui; Didisheim, Antoine; Scheidegger, Simon
• 署名单位: Massachusetts Institute of Technology (MIT); National Bureau of Economic Research; University of Lausanne; University of London; London School Economics & Political Science; University of Melbourne
• 刊物名称: JOURNAL OF FINANCIAL ECONOMICS
• ISSN/ISSBN: 0304-405X
• DOI: 10.1016/j.jfineco.2025.104222
• 发表日期: 2026-03
• 页码: 104222
• 关键词: Surrogate; Deep neural network; universal approximation; inference; Tail risk index; parameter instability; illiquidity; stochastic volatility; gaussian-processes; generalized-method; TRANSFORM ANALYSIS; NETWORKS OVERCOME; neural-networks; approximation; models; dimensionality; RISK

摘要

We introduce deep surrogates - high-precision approximations of structural models based on deep neural networks, which speed up model evaluation and estimation by orders of magnitude and allow for various compute-intensive applications that were previously infeasible. As an application, we build a deep surrogate for a high-dimensional workhorse option pricing model. The surrogate enables us to re-estimate the model at high frequency to construct an option-implied tail risk measure, which is highly predictive of future market crashes. It also helps us systematically examine the model's out-of-sample performance, which reveals the tradeoffs between structural and reduced-form approaches for option pricing. Moreover, we construct a measure for the degree of parameter instability and connect it to option market illiquidity in the data. Finally, we use the surrogate to construct conditional distributions of option returns, which is useful for risk management and provides a new way to test the model.