Randomized Neural Networks for estimation of exposure profiles and Credit Valuation Adjustment (CVA) for American Equity Options

The paper tests randomized neural networks (a fast-to-train neural network variant) for pricing American options and estimating their credit-related risk measures — exposure profiles and Credit Valuation Adjustment (CVA) — inside a Monte Carlo simulation. It compares this method against the standard Least-Squares Monte Carlo (LSM) approach under both Black-Scholes and Heston models, for single-asset and multi-asset portfolios. It finds the neural network matches LSM accuracy but scales more efficiently and cheaper in high-dimensional (many-asset) problems.

Why it matters: Risk and derivatives desks that must compute CVA and counterparty exposure for portfolios of American options could find this a more computationally efficient tool when many underlying assets are involved. The relevance is mainly to quantitative pricing/risk infrastructure rather than to trading signals or returns.

• Randomized neural networks matched the classical LSM benchmark for pricing, exposure, and CVA estimation.
• The key advantage is in high-dimensional, multi-asset problems, where it scales more efficiently at lower computational cost.
• Tested under both Black-Scholes and Heston dynamics with portfolio netting effects included.

⚠ This is a computational-methods study benchmarked against LSM in simulations, not a live-trading or market-return result, and applies mainly to specialized derivatives risk modeling.