Conglin Xu
Senior Editor-Economics
Oil price forecasts are of immediate interest and importance to many industries, central banks, private forecasters, and international organizations. However, forecasting oil prices out of sample is difficult, even at short horizons, as oil prices have evolved in very different pattern over time depending on the driving factors.
Not surprisingly, enormous efforts have been devoted by economists and industry analysts towards developing methods to forecast price levels. There is still no general consensus on which methods are more reliable. While computational approaches such as artificial neural networks have gained popularity in financial markets, forecasting approaches based on traditional econometrics are still more commonly used.
No-change forecasts are often used as a benchmark. The agnostic view is that the real price of oil follows a random walk such that the change in the real price of oil is unpredictable. Actually, in contrast to purely financial assets, oil prices are expected to be predictable to some extent, although they have increasingly behaved as financial asset prices over the past decade.
Oil futures prices are typically used by central banks and the International Monetary Fund as forecasts of the oil price in the physical market. It is well known that oil futures prices deviate from the market expectation of the price of oil due to a time-varying risk premium. This risk premium can be so high, making oil futures prices poor predictors of the price at many horizons. However, this does not rule out that oil futures prices have some predictive power at some horizons.
In regression-based structural models, oil price movements are modeled as a function of a collection of fundamental variables such as behavior among members of the Organization of Petroleum Exporting Countries, oil inventory level, oil consumption and production, and some nonoil economic variables. Some of the models explain the price movement fairly well. However, this does not necessarily imply that they provide good forecasting performance.
Time-series models, including exponential smoothing models and autoregressive models such as ARIMA and the ARCH/GARCH family, are often used to forecast oil price levels and volatility. These models are competitive at short horizon forecasts.
Vector autoregressive (VAR) models have been popular recently to forecast oil prices. According to Professor Lutz Kilian at University of Michigan, VAR oil price forecasting models usually work well only at horizons of at most 6 months and tend to work particularly well when there are large predictable movements in the data such as during 2008-09 or 2014-15. The VAR model does less well when the oil price is stable and more parsimonious forecasting methods may do just as well or better.
Notably, Christiane Baumeister and Kilian (2015) employed a four-variable structural VAR forecasting model for the real price of oil and showed that more than half of the decline in the price of oil was predictable in real time as of June 2014.
Other more complex forecasting models include dynamic stochastic general equilibrium (DSGE) models.
Forecast combination
A general problem with individual forecast methods is that their forecast accuracy might be highly time-dependent as some models might capture specific price patterns better than others. As claimed by many forecasters, no forecasting model works well at all horizons, and no individual forecast does well all the time.
A forecast combination, or pooling alternative oil price forecasts, could render projections more robust over time, in addition to offering gains in forecast accuracy.
In a very recent analysis, the European Central Bank reviewed nine different models for oil price forecasting and showed that different models perform well in specific periods and over specific horizons.
ECB recommended an equally weighted combination of four models: 1. Futures; 2. Risk-adjusted futures, which aims to correct the forecast error of futures by adjusting for a time-varying risk premium linked to US economic activity; 3. Bayesian VAR model; 4. DSGE model. Combining the models has offered substantial gains in forecast accuracy, both over time and across forecast horizons.