Structured Conditional Neural Solvers Improve Stochastic-Volatility Pricing Accuracy

01 ABSTRACT

To address the difficulty physics-informed neural networks face in preserving derivative accuracy across heterogeneous parameter regimes, the authors introduce BeTINN, which extends time-induced modulation via multi-directional time-parameter conditioning and a joint residual pathway. On the Heston stochastic-volatility PDE, BeTINN improves joint accuracy across normalized option values and key first- and second-order sensitivities relative to baselines. The authors view the results as supporting structured conditional adaptation and note that validation physics loss does not fully determine differential fidelity; these interpretations are the authors' views and require independent replication.

02 KEY FINDINGS

1. BeTINN extends time-induced modulation through multi-directional time-parameter conditioning and a joint residual pathway for parametric PDE solution families.
2. On the Heston stochastic-volatility PDE, BeTINN improves joint accuracy across option values and key sensitivity metrics relative to baselines.
3. Ablations characterize architectural component roles and indicate that validation physics loss is informative but does not fully determine differential fidelity.