Technical Summary: Probabilistic Multivariate Time Series Forecasting with Diffusion Copulas

1. Problem Statement

Accurate financial risk assessment requires capturing two distinct but interrelated phenomena: individual asset volatility and the complex, asymmetric dependence structures that emerge during extreme market events. Existing approaches face significant limitations:

• Traditional Copula Models: Historically relied on fixed parametric families (e.g., Gaussian) which assume asymptotic tail independence. This contradicts empirical reality, where correlations often converge to unity during market crashes (contagion), leading to severe underestimation of tail risk.
• End-to-End Diffusion Models: While modern diffusion-based models offer powerful generative capabilities for multivariate forecasting, they often suffer from a "normality bias" when trained end-to-end. By optimizing a single joint objective, these models frequently compromise the marginal calibration of individual assets. Consequently, they struggle to sample consistent extremes across assets of varying magnitudes and tend to underestimate joint tail risk, classifying simultaneous market crashes as statistically impossible "Black Swans."

The core challenge is to decouple the learning of marginal distributions from their dependence structure to preserve marginal fidelity (especially heavy tails) while accurately modeling joint extreme dependencies.

2. Methodology: The Diffusion-Copula Framework

The authors propose a two-stage Diffusion-Copula framework that explicitly separates marginal modeling from dependence modeling, leveraging Sklar's theorem.

2.1 Predictive Marginal Modeling

The first stage models the predictive marginal distribution $p(y_t^i | X_t)$ for each asset $i$ independently.

• Architecture: A Mixture Density Network (MDN) is employed to approximate the conditional probability density.
• Components: The mixture consists of heterogeneous distributions: Normal, Laplace, and Student-t.
• Parameterization: A neural network $\Psi$ outputs dynamic parameters (mixing coefficients $\pi$, location $\mu$, scale $\sigma$, and degrees of freedom $\nu$) conditioned on a history window $X_t$ (14 lagged returns plus auxiliary features like rolling volatility).
• Objective: The network is trained to minimize Negative Log-Likelihood (NLL) with entropy regularization to ensure diverse component utilization. This approach captures heavy tails, multimodality, and skewness while maintaining precise calibration.

2.2 Dependence Modeling via Classification-Diffusion Copula (CDC)

The second stage models the joint dependence structure using a Classification-Diffusion Copula.

• Transformation: Marginal observations are transformed to the copula scale $u \in [0, 1]^d$ via the Probability Integral Transform (PIT) using the marginal CDFs $F_i$. These are then mapped to the Gaussian scale $z = \Phi^{-1}(u)$.
• Diffusion Process: The method defines a forward Ornstein-Uhlenbeck (OU) process that progressively destroys the dependence structure while preserving standard Gaussian marginals.
• Classification Mechanism: Instead of a standard score-matching objective, the CDC trains a neural network classifier $f_\theta$ to distinguish between:
  • Dependent states: Early diffusion time ($t=0$, the data distribution).
  • Independent states: Late diffusion time ($t=T_K$, the independent noise distribution).
• Density Recovery: The copula density $c(u)$ is recovered as the ratio of class probabilities:
  $$c(u) = \frac{P_\theta(t=0 | z)}{P_\theta(t=T_K | z)}$$
• Training Objective: A hybrid loss function combines a cross-entropy term (for density ratio estimation and calibration) and a denoising score matching term (for accurate gradients and stable sampling).

2.3 Sampling Procedure

Future trajectories are generated via a two-stage process:

1. Copula Sampling: Dependent samples are generated on the Gaussian scale using Langevin dynamics, starting from Gaussian noise and denoising back to $t=0$. These are transformed to the unit hypercube $[0, 1]^d$.
2. Inverse Marginal Transformation: The uniform samples are mapped back to asset returns using the inverse CDF of the trained MDN, conditioned on the history $X_t$. This ensures the final samples inherit the heavy-tailed properties of the marginal models.

3. Key Contributions

• Decoupled Framework: The paper introduces a hybrid architecture that explicitly separates marginal calibration from dependence modeling, addressing the "normality bias" of end-to-end diffusion models.
• Classification-Diffusion Copula: It adapts the CDC (Huk & Damoulas, 2026) for time series, utilizing a classifier-based approach to model complex, non-parametric dependence structures without the rigidity of traditional parametric copulas.
• Heavy-Tail Preservation: By employing MDNs with Student-t components for marginals, the framework guarantees heavy-tailed coverage, which is critical for financial risk management.
• Systemic Event Modeling: The approach successfully captures the "contagion" effect where correlations spike during crashes, a feature often missed by baseline models.

4. Experimental Results

The framework was evaluated on cryptocurrency markets (ADA, BNB, BTC, DOGE, ETH, LINK, LTC, SOL, XRP), comparing against state-of-the-art baselines: Conditional Score-based Diffusion (CSDI) and Transformer-modulated Diffusion (TMDM).

4.1 Marginal Calibration

• PIT and QQ Plots: The proposed model (CDC) adhered closely to the ideal diagonal line in Probability Integral Transformation (PIT) and Quantile-Quantile (QQ) plots. In contrast, CSDI and TMDM showed deviations near 0 and 1, indicating an underestimation of heavy tails.
• Metrics: CDC achieved the best Root Mean Squared Error (RMSE) and Tail Event Accuracy. It ranked second in Mean Absolute Error (MAE) and Continuous Ranked Probability Score (CRPS), demonstrating a superior balance between central tendency and tail risk calibration.

4.2 Joint Extreme Forecasting

• Correlation Stability: In extreme quantiles (deep tails), CDC maintained superior stability in correlation matrix estimation. Baseline models exhibited negative bias drift (overestimating correlation during crashes) or failed to capture heightened co-movement.
• Systemic Event Probability: While baselines assigned near-zero probability to simultaneous crashes involving multiple assets ($k \ge 4$), CDC maintained a significant probability buffer, correctly anticipating synchronized downturns.
• Black Swan vs. Expected Crashes:
  • Baselines: Classified high-magnitude simultaneous events as "Black Swans" (high surprise/Mahalanobis distance).
  • CDC: Classified these same events as "Expected Crashes" (low surprise), effectively integrating extreme volatility into its predictive distribution.
• Robustness: CDC achieved the lowest CRPS on days with simultaneous extreme events, confirming its robustness in modeling joint tail risks.

5. Significance and Claims

The paper claims that the Diffusion-Copula framework offers a significant advancement for financial risk management by resolving the trade-off between marginal calibration and joint coherence.

• Risk Management: The ability to model asymmetric contagion (correlation spikes during crashes) allows for the construction of "tail-robust" portfolios.
• Stress Testing: The generative nature of the diffusion process enables risk managers to perform synthetic stress testing, simulating plausible "black swan" scenarios essential for robust Expected Shortfall (ES) estimation and pricing complex multi-asset derivatives.
• Paradigm Shift: The work suggests that for financial time series, explicitly decoupling marginals from dependence is superior to end-to-end joint modeling, particularly when accurate tail risk assessment is the primary objective.

The authors acknowledge that the current framework is a two-stage estimator where marginal errors may propagate to the dependence structure, suggesting future work could explore model-free marginal approaches (e.g., Neural Spline Flows) or Flow-based models for higher-dimensional precision.