Uncertainty and Explainability in Deep Rough Volatility: A Neural Information-Theoretic Posterior Approach

Uncertainty quantification matters for pricing: The paper demonstrates that using a single "best estimate" parameter for pricing exotic options can be "materially unreliable" in certain market conditions, whereas uncertainty-aware price intervals provide more dependable coverage across different contract types (forward-start, barrier, and realized-variance options).

Deep learning needs honest uncertainty estimates: While neural networks have accelerated financial model calibration, the authors show that neural point estimates alone miss important remaining uncertainty. Their framework captures both parameter uncertainty and model uncertainty, providing a more complete picture of what we don't know after observing market data.

Explainability should focus on information gain, not just predictions: Rather than explaining individual price predictions, the paper introduces "Hellinger-SHAP," which identifies which parts of market data (specific maturity-moneyness regions) actually inform our understanding of different model parameters—helping practitioners understand what market signals matter for different aspects of the model.

Simulation-based inference enables transparent AI in finance: By learning posterior distributions of model parameters rather than point estimates, the framework makes the inference process more transparent and auditable, allowing practitioners to see the full range of plausible parameter values consistent with observed market data.

Combining uncertainty and explainability improves responsible AI: The paper demonstrates that uncertainty quantification and explainability methods work together—understanding both what we're uncertain about and why provides a more trustworthy foundation for financial decision-making than either approach alone.