Forecasting Oil Prices Across the Distribution: A Quantile VAR Approach

1. Problem Statement

Forecasting real oil prices is notoriously difficult due to the asset's heavy-tailed, asymmetric distribution and its susceptibility to structural breaks during crises (e.g., the 2014–2016 collapse, the 2020 negative prices, and the 2022 Ukraine spike).

• Limitations of Existing Models: Standard econometric approaches, including Bayesian Vector Autoregressions (BVAR) with stochastic volatility (SV), focus on the conditional mean. While effective for point forecasting, they provide limited insight into tail risks (downside crashes or upside spikes) because they assume symmetric error distributions or summarize uncertainty around a single mean.
• The Gap: Macro-financial predictors (like uncertainty indices) often exhibit weak correlation with the level of oil prices but strong correlation with the skewness of returns. Mean-based models fail to capture these distributional asymmetries, leading to poor risk assessment during market disruptions.

2. Methodology: The Quantile Bayesian VAR (QBVAR)

The authors develop a Quantile Bayesian Vector Autoregression (QBVAR) to model the entire conditional distribution of oil prices, allowing dynamics to vary across quantiles.

A. Model Structure

The model extends the factor-augmented BVAR (Korobilis, 2022) to the quantile domain. For a vector of endogenous variables $y_t$ and a specific quantile $q \in (0,1)$:
$$y_t = \Phi_q x_t + \Lambda_q f_{t,q} + v_{t,q}$$

• Quantile-Specific Parameters: Unlike standard VARs, the coefficient matrix $\Phi_q$, factor loadings $\Lambda_q$, and latent factors $f_{t,q}$ are specific to each quantile. This allows the model to capture different drivers for median dynamics versus tail behavior.
• Factor Structure: To handle high dimensionality and avoid the variable-ordering dependence inherent in Cholesky-decomposition-based QVARs, the model uses a common factor structure ($\Lambda_q f_{t,q}$) to capture cross-equation comovements.
• Error Distribution: The error term $v_{t,q}$ follows an Asymmetric Laplace Distribution (ALD), which serves as the "working likelihood" for quantile regression.

B. Bayesian Estimation

• Mixture Representation: The authors utilize the location-scale mixture representation of the ALD (Kozumi and Kobayashi, 2011) to transform the non-Gaussian likelihood into a conditionally Gaussian form. This facilitates efficient Gibbs sampling.
• Priors:
  • Horseshoe Priors: Applied to VAR coefficients to aggressively shrink irrelevant predictors toward zero while preserving important signals (crucial for high-lag specifications).
  • Hierarchical Shrinkage: Global and local shrinkage parameters adapt to the sparsity of the data.
• Forecasting: For multi-step-ahead forecasting, the authors adopt a standard VAR approach with Gaussian innovations after estimation. They argue that while the ALD is necessary for estimation, the Central Limit Theorem justifies Gaussian approximations for iterated multi-step forecasts, avoiding the explosive paths that can arise from simulating auxiliary variables.

C. Evaluation Metric

Performance is measured using the Quantile Score (QS), also known as the "pinball loss" or check function ($\rho_q$). This is a proper scoring rule that penalizes over-predictions and under-predictions asymmetrically, depending on the target quantile (e.g., heavily penalizing under-prediction of downside risk at $q=0.1$).

3. Data and Experimental Design

• Sample: Monthly data from January 1975 to February 2025 (602 observations).
• Variables:
  • Target: Real Refiner Acquisition Cost (RAC) and Brent crude prices.
  • Predictors: A comprehensive set of 12 variables covering uncertainty (JLN, JLNF12), financial conditions (EBP, GZ spread, crack spreads), global activity (GECON, Kilian Index), supply/inventory (OECD stocks, Rig Count), and geopolitical risk (GPR indices).
• Benchmarks:
  1. Random Walk (RW): The difficult-to-beat univariate benchmark.
  2. BVAR(12): Standard Bayesian VAR with factor structure and horseshoe priors.
  3. BVAR-SV(12): State-of-the-art density benchmark incorporating stochastic volatility.
• Evaluation Windows: Recursive out-of-sample testing across two periods: 2008–2025 (including major crises) and 2013–2025 (post-crisis).

4. Key Results

A. Median Forecasts (Point Prediction)

• The QBVAR improves median forecasts by 2–5% relative to standard BVARs.
• This is a significant gain given the difficulty of outperforming well-specified mean-based models, suggesting that quantile-specific dynamics contain information relevant even for central tendency predictions.

B. Tail Forecasting Asymmetries

The paper identifies a fundamental asymmetry in oil price predictability:

• Left Tail (Downside Risk, $q=0.1$):
  • Performance: The QBVAR significantly outperforms all benchmarks, achieving 10–25% improvements in quantile scores.
  • Drivers: Uncertainty measures (JLN) and financial condition variables (GZ spread) are the strongest predictors.
  • Mechanism: These variables capture precautionary demand and financial stress that build gradually, providing early warning of price collapses. The gains intensify during crisis episodes (e.g., 2014–2016 collapse, 2020 pandemic).
• Right Tail (Upside Risk, $q=0.9$):
  • Performance: The QBVAR generally underperforms the BVAR-SV benchmark. Stochastic volatility models dominate here.
  • Reasoning: Price spikes are often driven by sudden, unpredictable supply shocks (e.g., wars, OPEC cuts) that are better captured by flexible variance dynamics than by linear quantile coefficients.
  • Exception: Geopolitical risk indices show some ability to predict upside risk, but generally, upside forecasting remains difficult.

C. Forecast Combinations

Recognizing the complementary strengths of the models, the authors test forecast combinations:

• Strategy: Combining QBVAR (strong on left tail) and BVAR-SV (strong on right tail).
• Result: Simple fixed-weight combinations (e.g., 40–50% QBVAR) successfully recover the right-tail losses of the pure QBVAR while preserving the massive left-tail gains.
• Optimization: Adaptive weighting strategies (based on recent performance) further enhance left-tail accuracy, confirming that the models are complementary rather than competing.

D. Performance During Major Events

• Price Declines: During the 2008 Financial Crisis, 2014–2016 collapse, and 2020 pandemic, the QBVAR consistently delivers the largest accuracy gains for downside risk, often outperforming RW and BVAR-SV by wide margins (up to 70% improvement in specific horizons).
• Price Surges: During the Gulf War, 2007–2008 boom, and 2021–2022 surge, results are mixed. The QBVAR performs well in the Gulf War and 2007–2008 but struggles in the post-COVID/Ukraine period, reinforcing the difficulty of forecasting supply-driven spikes.

5. Significance and Contributions

1. Methodological Innovation: The paper introduces a factor-based Bayesian QVAR that is invariant to variable ordering (unlike Cholesky-based QVARs) and utilizes hierarchical shrinkage priors to handle high-dimensional oil forecasting data.
2. Empirical Insight: It provides robust evidence that the determinants of oil prices are asymmetric. Financial stress and uncertainty are powerful predictors of crashes (left tail) but less so for spikes (right tail).
3. Practical Application: The study demonstrates that for risk management (e.g., Value-at-Risk, stress testing), modeling the full conditional distribution is superior to modeling only the mean.
   • Policy/Investment: Traditional models systematically underestimate downside risk during periods of rising financial stress. The QBVAR offers earlier warnings of potential price collapses.
   • Hybrid Approach: The optimal strategy for practitioners is a forecast combination that leverages the QBVAR for downside risk assessment and stochastic volatility models for upside risk, yielding the most robust distributional forecasts.

In conclusion, the paper argues that ignoring distributional heterogeneity leads to suboptimal risk assessment. By explicitly modeling quantile-specific dynamics, the QBVAR offers a superior framework for understanding and forecasting the asymmetric risks inherent in global oil markets.