Empirical prediction intervals improve energy forecasting.

Area of Science:

• Energy economics
• Forecasting methodologies
• Statistical modeling

Background:

• Energy projections, such as the US Energy Information Administration's Annual Energy Outlook (AEO), are vital for decision-making.
• Past analyses reveal significant deviations between AEO projections and observed values, highlighting the need for robust uncertainty quantification.

Purpose of the Study:

• To evaluate the out-of-sample forecasting performance of empirical density forecasting methods for energy quantities.
• To assess the accuracy of uncertainty estimates in energy projections.
• To provide guidance on producing, evaluating, and ranking probabilistic forecasts.

Main Methods:

• Utilized the continuous ranked probability score (CRPS) to evaluate forecasting performance.
• Assessed a Gaussian density, estimated on historical forecasting errors.
• Proposed a log transformation for price forecast errors and a modified nonparametric empirical density method.

Main Results:

• A Gaussian density, estimated on past errors, yielded accurate uncertainty estimates for various AEO energy quantities.
• This Gaussian approach outperformed the scenario projections offered within the AEO.
• Probabilistic uncertainties were quantified for 18 core quantities in the AEO 2016 projections.

Conclusions:

• Gaussian density estimation on past errors is a reliable method for quantifying uncertainty in energy projections.
• The proposed methods offer improvements for evaluating and communicating uncertainty in energy outlooks.
• Findings provide practical guidance for enhancing the reliability of future energy forecasts and decision-making.