Forecasting Oil Prices Across the Distribution: A Quantile VAR Approach
Forecasting Oil Prices Across the Distribution: A Quantile VAR Approach
Authors: Hilde C. Bjornland, Nicolás Hardy, Dimitris Korobilis Date: 2026-04-14 Paper ID: openalex:2604.12927
Summary
This paper introduces a Quantile Bayesian Vector Autoregression (QBVAR) to model oil prices across the full conditional distribution. By allowing predictor effects to vary by quantile, the model captures asymmetries that traditional mean-focused approaches fail to account for. Evaluation using monthly data from 1975 to 2025 demonstrates significant gains in both median point prediction and left-tail risk assessment compared to standard Bayesian VARs. The results highlight the importance of quantile-specific dynamics for better understanding market uncertainty and risk during oil price volatility.
Key Contributions
- Introduces a Quantile Bayesian Vector Autoregression (QBVAR) to capture quantile-specific predictor dynamics for oil price forecasting.
- Demonstrates that QBVAR improves median point forecast accuracy by 2-5% over standard Bayesian VAR models.
- Shows that the QBVAR provides 10-25% improvement in left-tail risk forecasting, specifically during crisis periods, while highlighting the difficulty of right-tail prediction.
Open Questions & Future Work
Archivist Review
I reviewed the Quantile Bayesian VAR approach, which applies existing econometric techniques to oil price forecasting. I decided against creating a concept note for QBVAR as it is a standard extension of established statistical methodology. I approved the open question regarding right-tail forecasting, as it frames a non-trivial, persistent challenge in conditional distribution modeling that is broadly relevant to time-series forecasting research.
Approved Open Questions
- Right-Tail Forecast Asymmetry Bottleneck: This is a fundamental limitation in conditional distribution forecasting, highlighting the trade-off between modeling tail-specific dependencies and the stochastic nature of volatile upside events.
Rejected Candidates
- [concept] Quantile Bayesian Vector Autoregression (QBVAR) (
qbvar) - not_novel: Quantile variants of VAR models are standard econometric methodology and not a novel machine learning framework worth a standalone vault entry.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.12927
- paper_id
- 2604.12927
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
-
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-17T05:45:46Z