Predictive Uncertainty Quantification for Financial DNN Using Regular Vine Copula

Tuoyuan Cheng, Nixie Sapphira Lesmana, Saikiran Reddy Poreddy, Kan Chen

Abstract: Quantifying joint uncertainty in multivariate financial forecasts is critical for risk management, yet remains challenging due to heavy tails, skewness, and asymmetric cross-asset dependencies. Existing deep learning uncertainty quantification methods are often computationally expensive, inflexible, and struggle to capture complex non-Gaussian dependence structures. We introduce a post-hoc, latent-conditioned framework that augments any pretrained Deep Neural Network (DNN) with a regular-vine copula layer to produce full joint predictive distributions without retraining the base model. After the DNN learns point forecasts and produces latent representations, we fit a Vine Computational Graph (VCG) on multivariate targets with latent features specified as conditioning variables, enabling efficient multivariate conditional sampling via the inverse-Rosenblatt transform. This design captures complex non-Gaussian dependencies essential for realistic multivariate risk assessment. We evaluate our DNN-VCG hybrid framework on three real-world datasets: weekly CDS returns with LIBOR, daily equity-ETF returns, and 15-minute cryptocurrency returns, across univariate scoring rules (Winkler, pinball loss) and dependence-focused metrics (Kendall’s τ deviation, Chatterjee’s ξ deviation). The framework achieves sharp prediction intervals, accurate tail quantiles, and improved joint-dependence calibration compared with several state-of-the-art Financial Predictive Uncertainty Quantification (FinPUQ) methods. It is lightweight, architecture-agnostic, supports arbitrary marginal models and bivariate copula families, accommodates diverse financial forecasting tasks, offering a practical high-throughput solution for multivariate uncertainty quantification in finance.

External IDs:doi:10.1145/3768292.3771254