Domain Knowledge Preservation in Financial Machine Learning: Evidence from Autocallable Note Pricing

Abstract

Machine learning applications in finance commonly employ feature decorrelation techniques developed for generic statistical problems. We investigate whether this practice appropriately addresses the unique characteristics of financial data, where correlations often encode fundamental economic relationships rather than statistical noise. Using autocallable structured notes as a laboratory, we demonstrate that preserving natural financial correlations outperforms conventional orthogonalization approaches. Our analysis covers autocallable notes with quarterly coupon payments, dual barrier structure, and embedded down-and-in up-and-out put options, priced using analytical methods with automatic differentiation for Greeks’ computation. Across neural networks, gradient boosting, and hybrid architectures, basic financial features achieve superior performance compared to decorrelated alternatives, with RMSE improvements ranging from 43% to 191%. The component-wise analysis reveals complex interactions between autocall mechanisms and higher-order sensitivities, particularly affecting vanna and volga patterns near barrier levels. These findings provide empirical evidence that financial machine learning benefits from domain-specific feature engineering principles that preserve economic relationships. Across all tested architectures, basic features consistently outperformed orthogonalized alternatives, with the largest improvements observed in CatBoost.