Bayesian Dynamic Modeling of Realized Volatility in Financial Asset Price Forecasting

Technical Summary: Bayesian Dynamic Modeling of Realized Volatility in Financial Asset Price Forecasting

Problem Statement
Financial forecasting requires efficient, sequential methods that can adapt to short-term volatility fluctuations while integrating high-frequency data. Traditional approaches face a trade-off: classical Stochastic Volatility (SV) models using latent log-variance processes are flexible but computationally intensive, often requiring Markov chain Monte Carlo (MCMC) or particle filtering with significant tuning. Conversely, Realized Volatility (RV) models forecast empirical volatility measures directly but often operate outside a joint price–volatility state-space. Hybrid models (e.g., Realized SV, Realized GARCH) combine these elements but typically rely on simulation-based inference or static likelihood optimization, lacking the analytic tractability required for rapid, online monitoring and scaling across many assets. The paper addresses the need for an analytically tractable, computationally fast framework that jointly models asset prices and realized volatility to improve short-term forecasting and risk management without the burdens of simulation-based methods.

Methodology
The authors introduce a new class of Realized Volatility Dynamic Linear Models (RV-DLMs). This framework extends traditional Bayesian discount Dynamic Linear Models (DLMs) for asset prices by integrating a dynamic gamma process model for realized variance.

1. Model Structure:

   • Price Component: The model retains the standard SV-DLM structure for log-prices ($y_t$), utilizing a latent state vector $\theta_t$ and a precision parameter $\phi_t$ (inverse variance). The evolution of $\phi_t$ follows a discount learning process controlled by a discount factor $\beta$.
   • Realized Variance Component: A new observation equation models the realized variance proxy ($z_t$) as $z_t = \eta_t / \phi_t$, where $\eta_t$ follows a Gamma distribution. This establishes a conditional Gamma relationship between the realized measure and the latent precision.
   • Bivariate Integration: The joint model for $(y_t, z_t)$ is constructed such that the marginal predictive distribution for $z_t$ is a scaled-F distribution, and the conditional distribution for $y_t$ given $z_t$ is a Student-t distribution. This preserves the conjugate normal-gamma structure of the underlying DLM.

2. Volatility Leverage and Feedback (RVL-DLM):

   • The framework allows for the inclusion of realized volatility proxies as predictors in the price equation. Specifically, the regression vector $F_t$ can include contemporaneous realized standard deviation ($x_t = \sqrt{z_t}$) and lagged values ($x_{t-1}$).
   • This formulation captures reduced-form leverage and feedback effects. The contemporaneous coefficient represents the immediate association between high-frequency volatility and price changes, while the lagged coefficient captures short-horizon effects. The model learns these coefficients sequentially from the data rather than imposing structural asymmetry terms.

3. Inference and Computation:

   • The model maintains full analytic tractability. Sequential filtering, forecasting, marginal likelihood evaluation, and retrospective smoothing are performed via closed-form updates.
   • One-step-ahead forecasting is achieved through direct compositional simulation: drawing $z_t$ from the scaled-F predictive distribution and then drawing $y_t$ from the conditional Student-t distribution.
   • The approach avoids MCMC and particle filters, offering negligible extra computational cost compared to standard DLMs.

Key Contributions

• Novel Model Class: The introduction of RV-DLMs and RVL-DLMs, which successfully overlay a dynamic gamma process for realized volatility onto standard Bayesian SV-DLMs.
• Analytic Tractability: The derivation of fully conjugate Bayesian analyses for the bivariate system, yielding scaled-F and Student-t predictive distributions without simulation-based inference.
• Integration of High-Frequency Data: A mechanism to incorporate intraday data (via OHLC-based realized variance proxies) directly into the sequential filtering of latent volatility states, improving the precision of volatility tracking.
• Reduced-Form Leverage Modeling: A flexible approach to modeling price-volatility dependence (leverage and feedback) through time-varying coefficients on realized volatility predictors, learned dynamically within the state-space framework.

Empirical Results
The authors applied the models to daily price data for 9 S&P 500 Select Sector SPDR ETFs and the S&P Index itself, covering the period from 2000 to 2025. They compared three models:

1. SV-DLM: A traditional univariate benchmark using only price data (with lagged RV as a fixed predictor).
2. RV-DLM: A bivariate model incorporating lagged realized volatility ($x_{t-1}$) as a predictor and using $z_t$ for direct volatility learning.
3. RVL-DLM: The most general model, including both lagged ($x_{t-1}$) and contemporaneous ($x_t$) realized volatility predictors.

• Predictive Performance: Cumulative log Bayes factors demonstrated that both RV-DLM and RVL-DLM variants consistently outperformed the traditional SV-DLM across all 10 series. The RVL-DLM, which includes the contemporaneous volatility predictor, provided further incremental gains over the RV-DLM.
• Coefficient Behavior: In the RVL-DLM, the posterior median for the contemporaneous coefficient ($x_t$) was generally negative, while the lagged coefficient ($x_{t-1}$) was generally positive. This aligns with high-frequency evidence of negative contemporaneous return-volatility asymmetry and positive lagged effects.
• Robustness: The models showed improved predictive density sharpness and reduced uncertainty, particularly during periods of market stress (e.g., the onset of the 2020 pandemic and the 2010 flash crash), where the realized volatility information provided critical updates to the latent state.

Significance and Claims
The paper claims that the RV-DLM framework offers a practical, scalable solution for financial forecasting that bridges the gap between the interpretability and speed of conjugate discount-learning filters and the empirical sharpness of realized volatility models.

• Efficiency: The method provides "computationally straightforward" analysis with "negligible extra computational cost," enabling scaling to higher dimensions for multivariate portfolio construction (decouple/recouple approaches).
• Practical Relevance: The models deliver contextual insights into the nature of volatility leverage and feedback effects, quantifying their impact on price forecasts without requiring complex structural assumptions.
• Forecasting Improvement: The primary gain is attributed to sharper predictive distributions and more efficient use of realized volatility information, rather than massive shifts in point forecasts.
• Future Scalability: The authors note that the closed-form structure naturally extends to multivariate settings using Wishart and multivariate-F distributions, positioning these models as building blocks for broader Bayesian risk evaluation and trading systems.

The study concludes that incorporating contemporaneous realized volatility into a fully conjugate sequential model is a robust strategy for improving short-term price forecasting and risk characterization in real-time systems.