We evaluate US Energy Information Agencies (EIA) forecasts of the world petroleum market, emphasising the importance of taking a multivariate perspective, considering asymmetric loss and allowing for time-variation. Forecasts for total demand, total supply, total stock withdrawals and the oil prices are biased, with biases that change over time and differ across variables. A loss function that takes into account asymmetry and interdependence can rationalise these biases. The implied asymmetric loss gives less weight to under-prediction of both demand and supply, while for oil prices, we document significant regime changes in the implied loss due to asymmetry. The EIA forecasts dominate a simple random walk benchmark when evaluated using symmetric and independent loss in the form of MSE statistical criteria. Yet, when allowing for asymmetry and interdependence that rationalize the EIA forecasts, the performance of the EIA forecasts worsens and is comparable to the random walk benchmark.
We develop a regime switching vector autoregression where artificial neural networks drive time variation in the coefficients of the conditional mean of the endogenous variables and the variance covariance matrix of the disturbances. The model is equipped with a stability constraint to ensure non-explosive dynamics. As such it is employable to account for changes in macroeconomic dynamics not only during typical business cycles, but also in a wide range of extreme events, like deep recessions and strong expansions. The methodology is put to test using aggregate data for the United States that include the abnormal realizations during the recent Covid-19 pandemic. The model delivers plausible and stable structural inference, and accurate out-of-sample forecasts. This performance compares favourably against a number of alternative methodologies recently proposed to deal with large outliers in macroeconomic data caused by pandemic.
Data | Code | Working paper version | Appendix
Structural vector autoregressions (SVARs) have been widely used for macroeconomic inferences since Sims (1980). A trade off between the computation efficiency and the complexity of the identification becomes a central issue. An increasingly popular way of identifying structural shocks is through a set of restrictions, which not only includes sign, but other qualitative identifications when shocks happen and in their aftermath. Computation using the traditional accept-reject approach can take a very long time since many draws can be rejected. An alternative way to identify only sign-restricted VARs using the Bayesian method, proposed by Baumeister & Hamilton (2015), does not run into this computational problem. However, this method limits identification restrictions to the impact matrix and, in practice, can lead to imprecise estimates of structural parameters. This paper proposes a parallel Metropolis--Hastings algorithm to identify and compute Bayesian SVARs, which extends Baumeister & Hamilton's (2015) method from sign to a set identification. Two specifications commonly used in crude oil market modelling are employed to illustrate that the new method offers improvements in terms of computational efficiency and a more precise estimates of crude oil demand elasticities. Furthermore, the novel method proposed in this paper gauges the uncertainty of restrictions on non-linear structural parameters, for example demand and supply elasticities. This paper provides empirical evidences that the uncertainty of restrictions leads to precise estimates of the non-linear parameters of interest.
Data | Code | Working paper version
This paper uses a Bayesian vector autoregression (BVAR) considering time-varying parameters and stochastic volatility modelling time variation in forecasting real crude oil prices. Two features are specific to this study. First, I minimise the one-step prediction Kullback–Leibler `distance' in the sample and then eliminate the outliers in the out-of-sample density prediction, allowing the highly parametric model to limit shrinkage, particularly in the long-term prediction range. Second, I extend the predictive assessment from standard statistical measures of the point and density forecasts to those that have a profitable opportunity in the futures market, as well as those that best predict the probability of real crude oil prices being extreme high or low. Over the 1992:01--2016:12 period, I find strong evidence supporting models using stochastic volatility for real crude oil price density forecasts as opposed to using a conventional VAR. A constant parameter VAR with stochastic volatility can provide well calibrated density forecasts and the highest probabilities for positive excess returns using crude oil futures. Time-varying parameters may contribute to density forecasts, but only when the majority of them are restricted to be time-invariant via a stochastic model specification selection prior and a linear opinion pool combining the 1- to 12-month VAR lag length choices.
Data | Code | Working paper version
Baumeister & Kilian (2015) combine forecasts from six empirical models to predict real oil prices. In this paper, we broadly reproduce their main economic findings, employing their preferred measures of the real oil price and other real-time variables. Mindful of the importance of Brent crude oil as a global price benchmark, we extend consideration to the North Sea based measure and update the evaluation sample to 2017:12. We model the oil price futures curve using a factor-based Nelson-Siegel specification estimated in real time to fill in missing values for oil price futures in the raw data. We find that the combined forecasts for Brent are as effective as for other oil price measures. The extended sample using the oil price measures adopted by Baumeister & Kilian (2015) yields similar results to those reported in their paper. And the futures-based model improves forecast accuracy at longer horizons. The real-time data set is available for download from https://www.niesr.ac.uk/real-time-forecast-combinations-oil-price.