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

Battle Count: 78

The paper is primarily relevant to risk management and derivatives valuation rather than direct trading strategies. However, it addresses a critical computational bottleneck in counterparty credit risk (CVA) for American-style options, which directly impacts pricing, hedging, and portfolio management decisions for derivatives desks. The scalability advantage in high-dimensional settings (50+ assets) is practically significant for institutional risk management. The work is more relevant to quantitative risk management and derivatives structuring than to algorithmic trading or market-making strategies.

The RLSM method is moderately complex to implement. It requires: (1) Monte Carlo path generation under BS/Heston dynamics, (2) randomized hidden-layer initialization, (3) backward induction with ridge regression at each exercise date, (4) forward evaluation for pricing and exposure, (5) portfolio aggregation with netting, and (6) CVA computation with hazard-rate curves. The core algorithm is simpler than full deep learning (no backpropagation), but hyperparameter tuning (hidden size, regularization, input scaling) adds practical complexity. The paper provides complete pseudocode and source code, reducing implementation barriers.

The paper provides a GitHub repository with complete source code (https://github.com/isidromoroso/thesis_final_code). All model parameters, Monte Carlo configurations, basis functions, and algorithm pseudocode are explicitly specified. The experimental setup closely follows Herrera et al. (2024), enabling direct comparison. However, some hyperparameter choices (hidden size, regularization) vary by dimension and product, requiring careful replication.