SPECIAL FOCUS

Forecasting Industrial Commodity Prices:

An Assessment
COMMODITY MARKETS OUTLOOK | APRIL 2024 SPECIAL FOCUS 35

The Special Focus of this edition evaluates the performance of five well-known approaches to forecasting the prices of three key
industrial commodities—aluminum, copper, and crude oil—over the period 2015Q1 to 2022Q1. High short-term volatility
and significant longer-term movements in commodity prices—both features of commodity markets in recent years—present
major challenges for policymakers in commodity-exporting EMDEs. Such challenges are easier to meet the more accurately
price changes can be forecast. The evaluation reveals four main results. First, there is no “one-approach-beats-all” for
commodity price forecasting, as the forecast accuracy of approaches varies significantly across commodities and time horizons.
Second, macroeconometric models tend to be more accurate at longer horizons, partly because they can incorporate the effects
of structural changes on prices. Third, it is critical to complement forecasts by incorporating judgment (information that
cannot be accounted for by statistical approaches), especially when confronted by unusual or unprecedented events. Finally,
these results underscore the value of employing a range of approaches in forecasting commodity prices.

FIGURE 18 Commodity dependence and commodity

Introduction price volatility
Almost two-thirds of emerging market and Many EMDEs depend heavily on commodity exports, with oil exporters
developing economies (EMDEs) heavily rely on typically more reliant than metals exporters. These export proceeds, in
turn, provide a sizeable share of fiscal revenues. Price volatility is high, as
commodities for export, fiscal revenue, and is evident in the size and amplitude of cycles, with price booms more
economic activity (figure 18.A-B). Among pronounced than price slumps, on average. The speed of commodity price
commodity-exporting EMDEs, resource sectors, rises in booms is also much faster than that for declines in slumps,
especially for crude oil.
on average, account for nearly 40 percent of
exports of goods and non-factor services, 31 A. Share of EMDE exports B. Resource revenues
percent of goods exports, and 10 percent of value Share of EMDE exports Percent of fiscal revenues
added. In some commodity-importing EMDEs, 80
80
70
commodities account for a large share of imports 60
60
and, in the presence of subsidies, fiscal spending. 50

The substantial volatility of commodity prices 40

30
40

exacerbates revenue management challenges in 20 20

EMDEs (figure 18.C-D). Large and persistent 10
0 0
price shocks caused by commodity price volatility Oil Aluminum Copper Oil Aluminum Copper

weaken fiscal and external positions and lower

economic growth in commodity-exporting C. Amplitude of booms and slumps D. Average speed of booms and
slumps
EMDEs (IMF 2015, Richaud et al. 2019).
Percent Booms Slumps Percent Booms Slumps
8
This Special Focus evaluates the performance of 200
150 6
five well-known approaches to forecasting the
100 4
prices of three key industrial commodities: 50 2
aluminum, copper, and crude oil (figure 19.A). 0
0
Copper
Oil

Aluminum

These three commodities account for about half of

Copper
Oil

Aluminum

global commodity exports. The evaluation focuses

on four model-based approaches (bivariate
regression models; Bayesian vector autoregression Sources: International Monetary Fund; UNComtrade (database); UNU-Wider (database); World Bank.
A. Figure shows the median share of exports accounted for by oil, copper, and aluminum for EMDEs
models; a macroeconometric model; and a that are commodity exporters. Oil includes 20 EMDEs, copper 6, and aluminum 4. Blue bars show
medians and orange whiskers show interquartile ranges.
machine-learning model) and forecasts from B. Unweighted average of resource revenues as a share of fiscal revenues for EMDE commodity
Consensus Economics. exporters of oil (25 countries), copper (4 countries), and aluminum (3 countries). Countries relying on
the export of multiple commodities are included in the averages for each commodity. Orange
whiskers indicate the range between the minimum and maximum values.
C. Data from January 1970 to October 2021. Amplitude measures the average real price change (in
Note: This Special Focus was prepared by Francisco Arroyo- percentage terms) from trough to peak for booms and from peak to trough for slumps.
Marioli, Jeetendra Khadan, Valerie Mercer-Blackman, Franziska D. Data from January 1970 to October 2021. Slump refers to the average monthly amplitude (in this
Ohnsorge, and Takefumi Yamazaki. The discussion is drawn from case, amplitude divided by the duration).
Arroyo-Marioli et al. (2023). Research assistance was provided by
Muneeb Ahmad Naseem, Vasiliki Papagianni, Lorez Qehaja, and
Kaltrina Temaj.
36 SPECIAL FOCUS COMMODITY MARKETS OUTLOOK | APRIL 2024

The benefits of employing a menu of approaches Bayesian vector autoregressive models

to forecast commodity prices, rather than
attempting to identify a single “best” approach, The Bayesian vector autoregressive (BVAR)
have been discussed by earlier studies (for model, a multivariate VAR model, estimates the
example, Baumeister and Kilian 2015). relationships among two or more variables. It
Specifically, compared to a single approach, a differs from the standard multivariate VAR model
variety of approaches can enhance the reliability of in that the model parameters are treated as
forecasts by taking into account a broader view of random variables—with prior probabilities—
possible price outcomes. A previous edition of the rather than fixed values. A BVAR model with sign
Commodity Markets Outlook (World Bank 2023) restrictions is employed here to forecast
featured a Special Focus that presented a brief commodity prices following techniques developed
literature review on various approaches for by Kilian and Murphy (2014). The dependent
forecasting industrial commodity prices. It found variables in the model include industrial
that the five approaches examined here have commodity prices, log differences in metal
performed relatively well for selected commodities. production, and the global GDP growth rate. The
Building on that review, this Special Focus presents Bayesian estimation approach allows for structural
new empirical evidence by comprehensively identifications such as elasticity and sign
evaluating these approaches. restrictions and prior beliefs about future
economic events. The estimated impulse response
functions that satisfy the sign restrictions are used
Forecasting Approaches to forecast industrial commodity prices.
Bivariate regression models Macroeconometric model
Bivariate regression models are useful for For this exercise, we utilized the Oxford
capturing the basic relationships between Economics Global Economic Model (OEM). The
macroeconomic variables and commodity prices. OEM is a large-scale, cross-country, semi-
They are used to forecast industrial commodity structural projection model well suited to the
prices by simply regressing the change of a analysis of alternative projections for the global
commodity price on a “past” value of an economy (Oxford Economics 2019). It includes
explanatory variable. Six explanatory variables are 81 countries, 6 regional blocs, and the Eurozone.
employed in regressions: the Commodities Most have data available quarterly. Behavioral
Research Bureau Raw Industrial Commodity equations governing domestic economic activity,
Index; U.S. M1 growth; U.S. Treasury Bill 10- monetary and fiscal policy, global trade, and
year interest rates; China’s manufacturing commodity prices are used. The model combines
purchasing managers’ index (PMI); and global short-run momentum factors and long-term
composite and manufacturing PMIs. The first demand and supply fundamentals. In the short
three variables are often found to be useful run, shocks to demand drive business cycles
predictors of crude oil prices (Alquist, Kilian, and that can be influenced by fiscal and monetary
Vigfusson 2013). The last three reflect the policies. Over the long run, output is determined
importance of the outsized share of China in by supply-side factors such as investment, labor
global metal markets and the importance of global force participation, and productivity. The
activity in driving industrial commodity prices. resulting dynamics of short-run fluctuations and
This approach yields six forecasted prices for the long-run trends yield quarterly commodity price
three commodities and for each forecast horizon, forecasts.
reflecting the number of independent variables.
Machine-learning model
The final forecast is the average of all estimations
with statistically significant coefficients produced The machine-learning model combines algorithms
by bivariate regression models. and econometric methods to learn patterns and
COMMODITY MARKETS OUTLOOK | APRIL 2024 SPECIAL FOCUS 37

estimate relationships in data. It then makes FIGURE 19 Directional accuracy of commodity price
projections based on the patterns and estimated forecasts
relations without imposing any theoretical prior. While industrial commodity prices exhibit volatility that makes forecasting
Following Zhang et al. (2015), a hybrid machine- challenging over the short and longer terms, most evaluated forecasting
approaches accurately predicted the direction of price changes for the
learning model is employed here that comprises an three industrial commodities. BVAR and Consensus Forecasts had lower
empirical mode decomposition, and a generalized accuracy in predicting the direction of price changes for copper over
autoregressive conditional heteroskedasticity longer horizons. BVAR exhibited less directionally accuracy than other
approaches for oil. Bivariate regressions consistently demonstrated high
model. The hybrid model approach used in this directional accuracy across the three commodities.
exercise separates commodity price series into
different nonlinear and time-varying components. A. Industrial commodity prices B. Bayesian vector autoregression
Commodity price forecasts are constructed by models (BVAR)

adding forecasts of these components. Index, 100 = 2000 Percent 3-12 months 15-24 months
600 Aluminium Copper Oil 90
80
Forecasts from Consensus Economics
400 70
60
Consensus forecasts are published by Consensus 200 50
Economics (CE), a service that surveys several 40

Copper
Oil

Aluminum
forecasters for their projections of future output 0

2000
2002
2004
2006
2008
2010
2012
2014
2016
2018
2020
2022
growth, commodity prices, current account
balance, and other major macro variables. CE
forecasts of commodity prices are drawn from C. Bivariate regressions D. Consensus Economics forecasts

“Energy & Metals Consensus Forecasts” reports, Percent

100
Percent 3-12 months 15-24 months
90
which are based on a monthly survey of up to 40 90 80
leading private-sector commodity forecasters 80
70

covering 50 individual commodities. These 70

60

forecasts are a simple compilation, which means 50

60 40
they do not consider the systematic consistency of
Copper
Oil

Aluminum

Copper
Oil

Aluminum
methodologies used by different forecasters.

Data E. Macroeconometric model F. Machine learning approach

Percent 3-12 months 15-24 months Percent 3-12 months 15-24 months
100
The data frequency and sample periods used for 100
90
90
estimation vary across the approaches due to data 80 80

availability. Bivariate models are based on monthly 70 70

60 60
data for the period 2004Q1-2022Q1. For the 50 50

BVAR models, quarterly averages of monthly data 40 40

Copper
Oil

Aluminum
Copper
Oil

Aluminum

between 1995Q1-2022Q1 are used. The quarterly

GDP growth rates are drawn from Haver
Analytics, with forecasts based on the World
Source: World Bank.
Bank’s June 2022 Global Economic Prospects report Note: Charts show the percent of quarters for which each approach correctly predicts the
direction of price changes. Bars and diamonds represent averages of correct directional
(World Bank 2022). Production data for predictions for 3-12 months and 15-24 months forecast horizon, respectively.
aluminum and copper are drawn from the World C. Bivariate regressions are not evaluated beyond the 12-month horizon.

Bureau of Metal Statistics, and for oil from the

International Energy Agency. The machine
learning approach is estimated using monthly data
from 1995Q1 to 2022Q1.
38 SPECIAL FOCUS COMMODITY MARKETS OUTLOOK | APRIL 2024

TABLE 2 Forecast bias

Forecast bias does not differ significantly across approaches for most forecast horizons and commodities. Some approaches lead to
relatively larger biases, but these are not statistically significant. Bivariate regressions tend to overpredict copper prices, and BVAR produce
a larger bias for oil. CE forecasts produce smaller biases for all commodities.

Aluminum Copper Oil

Commodities
(US$/mt) (US$/mt) (US$/bbl)
Approaches/Horizons 3-12 months 15-24 months 3-12 months 15-24 months 3-12 months 15-24 months
CE forecasts -3.3 16.2 -108.1 -12.9 -1.6 1.5
Bivariate regressions 91.3 - 660.0 - 4.9 -
BVAR -20.2 -8.0 -98.5 -170.2 4.9 17.0
Macroeconometric model -0.3 2.9 -89.2 21.0 2.2 1.0
Machine learning 18.0 -2.5 -34.3 57.0 3.2 2.4
Actual prices (average,
1,948 6,493 56.6
2015Q1-2022Q1)

Source: World Bank.

Note: Macroeconometric model refers to the Oxford Economics Model. The table shows the average forecast bias, defined as the difference between the actual price and the predicted price,
for the 3-12 month and 15-24 month forecast horizons. The forecast evaluation period is 2015Q1-2022Q1. Forecasts for the macroeconometric model are only available at semi-annual
frequency; forecasts for all other approaches are available at a quarterly frequency. Bivariate regressions are not evaluated beyond the 12-month horizon.

Criteria for evaluating statistically significant only for horizons up to one

year ahead, other horizons are excluded. Historical
forecast performance OEM forecasts are only available semi-annually;
hence, the forecast accuracy tests are adjusted for
The five forecasting approaches are evaluated in the fewer degrees of freedom.
terms of three well-known statistical criteria:
Results
• Directional accuracy. The directional accuracy
simply assesses the likelihood of forecasts and Directional accuracy of forecasts
actual prices moving in the same direction.
The directions of price changes are often
• Forecast bias. The forecast bias, defined as the accurately predicted by most models for most
mean of the difference between the actual and commodities (figure 19.B-F). However, there are
forecasted prices (mean forecast error), few exceptions, notably lower directional accuracy
evaluates whether forecasts systematically of the BVAR compared to other approaches,
over- or under-predict their realized values. particularly for oil. For copper, the BVAR and
Consensus Forecasts correctly predict the direction
• Forecast accuracy. The accuracy of model of price changes less frequently. Bivariate
forecasting is evaluated by the Diebold and regressions tend to produce directionally accurate
Mariano (DM) test (Diebold and Mariano forecasts at shorter horizons.
1995). The DM test checks whether a
particular model is more accurate than Forecast bias
another by assessing the statistical significance
of the difference in forecast errors between Forecast bias does not differ significantly across
model pairs. The DM test is implemented for models for most forecast horizons and
each model against all other approaches. commodities. While there are some exceptions,
they are not statistically significant. Bivariate
The forecast evaluation covers price forecasts for regressions lead to forecasts with a higher bias for
each of the three industrial commodities, ranging aluminum and copper for horizons up to one year.
from one to eight quarters ahead, for the period BVARs produce forecasts with a greater bias for oil
2015Q1-2022Q1. As bivariate regressions are beyond the one-year horizon (table 2).
COMMODITY MARKETS OUTLOOK | APRIL 2024 SPECIAL FOCUS 39

TABLE 3 Model accuracy

Forecast accuracy differs significantly across models at short horizons. Beyond one year, CE forecasts and the macroeconometric approach
produce more accurate forecasts.

Horizon 3-12 months 15-24 months

Approaches/Commodities Aluminum Copper* Oil** Aluminum Copper Oil
CE forecasts
Bivariate regressions
BVAR
Macroeconometric model
Machine learning

Source: World Bank.

Note: Macroeconometric model refers to the Oxford Economics Model. The table identifies the approaches that have the highest incidence of success in forecasting commodity prices based
on the Diebold-Mariano test for 3-12 months and 15-24 months forecast horizons. Green cells represent the best approach and blue cells highlight the second best approach. * indicates that
Consensus Forecasts and the machine learning approach had similar forecast accuracy compared to other models for horizons up to one year. ** indicates that the accuracy of bivariate
regressions, Consensus Forecasts, and the machine learning approaches had similar forecast accuracy compared to other models for horizons up to one year.

Forecast accuracy Additional considerations

The forecast accuracy of models differs Adding modeler priors. None of the forecasting
significantly across commodities and forecast approaches here have pre-designed scenarios or
horizons. All models are compared against each priors. However, in reality, it is often of interest to
other for all horizons using the DM test. Table 3 condition the forecasts using different scenarios
highlights the two forecasting approaches that (for example, different trajectories for the world
perform better for each commodity for forecast economy, a shock to the supply of oil, or changes
horizons 3-12 months (short-term) and 15-24 in policies). The OEM is particularly useful for
months ahead (medium-term). scenario exercises that consider changes in policy
• For aluminum prices, bivariate regression variables, global growth, inflation, and structural
models and the machine learning approach variables. Another purpose of conditional forecasts
are the most accurate for forecast horizons up is to incorporate information from higher
to one year. In contrast, the OEM and the frequency data or judgment into the model
machine learning approach performed better (Karlsson 2013). This underscores the main
than other approaches over the medium-term advantage of the BVAR model, which allows the
horizons. forecaster to simulate scenarios or test priors while
maintaining the statistical properties of the model.
• For copper prices, the OEM approach is more The forecaster can then make inferences about the
accurate than other approaches for all forecast posterior forecast, conditional on the prior. The
horizons. CE forecasts and machine learning version of the BVAR model used in this exercise
approaches had similar forecast accuracy leads to a larger oil price forecast bias than other
compared to other approaches for forecast models. However, in practice, the forecaster could
horizons up to one year. subsequently adjust the parameters to reflect
changes in priors to arrive at more informed
• For oil prices, three approaches—bivariate forecasts.
regressions, CE forecasts, and the machine
learning approach—produce forecasts with Adjusting to shocks and incorporating
similar accuracy compared to other judgment. Commodity markets are subject to
approaches for forecast horizons up to one a wide range of shocks affecting demand and
year. CE forecasts and the OEM are the two supply dynamics, ultimately leading to sharp price
most accurate approaches at forecast horizons movements. For example, the COVID-19
beyond one year. pandemic triggered a global recession in 2020, led
40 SPECIAL FOCUS COMMODITY MARKETS OUTLOOK | APRIL 2024

FIGURE 20 Forecasts and realizations: 2015Q1-2022Q1 and after the pandemic, as the model had been
The approaches often lead to better forecast accuracy during the pre-
adapted to accommodate the pandemic shock
pandemic period. This is in part driven by the disruption caused by the (figure 20.D). CE forecasts also performed
pandemic to the standard relationships between global macroeconomic relatively better as they reflected the aggregate
conditions and commodity prices. Among the five approaches, the expert-
and modeler-centric approaches (CE forecasts and the macroeconometric
views of many forecasters who could adjust their
model) produced more accurate forecasts in the post-pandemic period, in projections to account for the pandemic shock.
part because they were better able to incorporate this information. Though the pandemic period was highly unusual,
the behavior of different approaches serves as a
A. Aluminum prices and range of B. Copper prices and range of
forecast results from 5 approaches forecast results from 5 approaches
good illustration of the need to incorporate
$/mt $/mt
judgment and other factors outside the model
Range of forecast values, 12-month Range of forecast values, 12-month
3,500
Aluminum price 12,000 Copper price specification when forecasting.
3,000
10,000
2,500
8,000 Utilizing multiple approaches. Each forecasting
2,000
6,000
approach has unique strengths that should be
1,500

1,000
considered in practice. For instance, the BVAR
4,000
and OEM approaches excel in scenario analysis,
2015Q1
2015Q3
2016Q1
2016Q3
2017Q1
2017Q3
2018Q1
2018Q3
2019Q1
2019Q3
2020Q1
2020Q3
2021Q1
2021Q3
2022Q1

2015Q1
2015Q3
2016Q1
2016Q3
2017Q1
2017Q3
2018Q1
2018Q3
2019Q1
2019Q3
2020Q1
2020Q3
2021Q1
2021Q3
2022Q1
while machine learning methods demonstrate
proficiency in uncovering complex patterns in
C. Crude oil prices and range of D. Crude oil prices and range of time series data that traditional statistical models
forecast results from 2 approaches forecast results from 3 approaches
may overlook. Bivariate regressions are particularly
$/bbl Range of expert/modeler-centric $/bbl Range of price forecasts of statistical
100 forecasts, 12-month
Crude oil price
100 approaches, 12-month valued for their simplicity and ability to identify
Crude oil price
80 80 the most influential explanatory variables.
60 60
Meanwhile, CE forecasts serve as a valuable
40
sentiment indicator, offering a robust
40
benchmarking alternative with the additional
20 20
advantage of being timely and accessible.
2015Q1
2015Q3
2016Q1
2016Q3
2017Q1
2017Q3
2018Q1
2018Q3
2019Q1
2019Q3
2020Q1
2020Q3
2021Q1
2021Q3
2022Q1

2015Q1
2015Q3
2016Q1
2016Q3
2017Q1
2017Q3
2018Q1
2018Q3
2019Q1
2019Q3
2020Q1
2020Q3
2021Q1
2021Q3
2022Q1

Considering other industrial commodities.

Source: World Bank.
Notes: Grey shaded area denotes trough of the COVID-19 pandemic (2020Q2-Q3). Data shows
While the results are not reported here, lead,
1-year-ahead rollover over a 12-month horizon forecast period. The range includes the forecast
outcomes of the five approaches examined (Bivariate regressions, BVAR, Consensus Forecasts,
nickel, tin, and zinc prices are also evaluated using
the machine learning approach, and the Oxford Economic Model). the same forecast evaluation exercise (Arroyo-
C.D. Measures average crude oil prices (unweighted average of Brent, West Texas Intermediate,
and Dubai benchmarks) in U.S. dollars per barrel. Marioli et al. 2023). These four industrial metals
C. Expert- and modeler-centric forecast approaches refer to Consensus Forecasts and the
Oxford Economic Model.
have recently seen increasing demand in part due
D. Statistical forecast approaches refer to the Bivariate regressions, BVAR, and the machine to their use in clean energy technologies, though
learning approach.
they are less systemically important to the global
economy compared to the three studied here. For
these four industrial commodities, most
to gyrations in commodity prices, and, for a brief approaches performed well, with the OEM
period, even upended the standard relationships forecast showing the lowest bias and forecast error
among macroeconomic variables. Forecasting for forecast horizons of 12 months or more.
performance of all approaches naturally suffered
during the pandemic. The purely statistical Using futures prices. Prices of futures are often
forecasting approaches (bivariate regressions, used for forecasting purposes by many
BVAR, and, to a lesser extent, machine learning) organizations (Nixon and Smith 2012; World
were unable to quickly adjust to extraneous factors Bank 2023). They are simple to utilize as they
outside of the model. Past relationships embedded reflect market expectations of future spot prices.
into the models during the training period Earlier studies report that futures prices are often
temporarily broke down (figure 20.A-C). In unbiased but inefficient forecasts (with large
contrast, forecasts based on the OEM displayed a forecast errors in either direction) compared to
smaller difference in relative performance before forecasts produced by other approaches, including
COMMODITY MARKETS OUTLOOK | APRIL 2024 SPECIAL FOCUS 41

VARs, machine learning techniques, and these approaches. Commodity prices are driven by
univariate time series models. Despite their forces that may not be captured by backward-
relative underperformance, some studies find that looking statistical techniques. These techniques
futures prices do contain important predictive can be improved with reference to events or
information, and the financialization of information known to the modeler but not yet
commodity markets may have helped improve incorporated in the data.
their predictive power over time (Arroyo-Marioli
et al. 2023; Ellwanger and Snudden 2023). Importance of multiple approaches. A single
approach can sometimes produce large forecast
Conclusions errors. Moreover, forecast accuracy varies
significantly across approaches. These results
This Special Focus evaluates the performance of collectively emphasize the importance of
five widely used approaches to forecasting the employing a rich menu of approaches in
prices of industrial commodities. The evaluation forecasting commodity prices.
focuses on the prices of aluminum, copper, and
crude oil as these commodities account for almost For policymakers, these results underscore the
half of global commodity exports. It examines four uncertainty around commodity price forecasts and
model-based approaches (bivariate regressions; the need to develop contingency plans for
Bayesian vector autoregression models; a alternative outcomes, particularly for economies
macroeconometric model; and a machine learning heavily dependent on commodities for revenues.
technique) and CE forecasts. These approaches are The usefulness of forecasting is sometimes less
evaluated in terms of their performance with about predicting the future with accuracy and
respect to directional accuracy, forecast bias, and more about looking at how changes in certain
forecast accuracy over the period 2015Q1- assumptions might lead to different outcomes, as
2022Q1. The evaluation finds four major results. well as the risk associated with those outcomes. In
practice, it is crucial to use various models, each
No “one-approach-beats-all” for commodity with its strengths, coupled with an informed
price forecasting. Most approaches produce assessment of potential changes in commodity
directionally accurate forecasts at horizons of less markets.
than one year. Forecast bias does not differ
significantly across approaches for most forecast
horizons and commodities. However, the forecast
accuracy of approaches varies significantly across
References
commodities and time horizons. Since market Alquist, R., L. Kilian, and R. J. Vigfusson. 2013.
conditions often change, it is not possible for one “Forecasting the Price of Oil.” In Handbook of
approach to systematically outperform the others Economic Forecasting, edited by G. Elliott, C. W.
across different commodities, particularly over J. Granger, and A. Timmermann, 427-507.
shorter horizons. Amsterdam: Elsevier.
Arroyo-Marioli, F., J. Khadan, F. Ohnsorge, and T.
Macroeconometric models are better for longer-
Yamazaki. 2023. “Forecasting Industrial Commodity
term forecasts. These models tend to be more Prices: Literature Review and a Model Suite.” Policy
accurate at horizons of one year or more, partly Research Working Paper, No. 10611. Washington,
because they can incorporate the impact of D.C.: World Bank Group.
structural changes on prices. These models are also
useful for conducting scenario analyses and Baumeister, C., and L. Kilian. 2015. “Forecasting the
Real Price of Oil in A Changing World: A Forecast
considering forecasts conditional on certain
Combination Approach.” Journal of Business and
outcomes. Economic Statistics 33 (3): 338-51.
Need for incorporating judgment. It is critical to Diebold, F., and R. Mariano. 1995. “Comparing
complement mechanistic forecasts with judgment Predictive Accuracy.” Journal of Business and Economic
and information that cannot be accounted for by Statistics 13 (3): 253-63.
42 SPECIAL FOCUS COMMODITY MARKETS OUTLOOK | APRIL 2024

Ellwanger, R., and S. Snudden. 2023. “Futures Prices Oxford Economics. 2019. “The Oxford Global
are Useful Predictors of the Spot Price of Crude Oil.” Economic Model.” July. Oxford Economics, Oxford,
The Energy Journal 44 (4): 45-62. UK. www.oxfordeconomics.com/service/subscription-
services/industries/commodity-price-forecasts.
IMF (International Monetary Fund). 2015. “The
Commodities Roller Coaster” In Fiscal Monitor. Richaud, C., A. Galego, F. Ayivodji, S. Matta, and S.
Washington, DC: International Monetary Fund. Essl. 2019. “Fiscal Vulnerabilities in Commodity
Exporting Countries and the Role of Fiscal Policy.”
Karlsson, S. 2013. “Forecasting with Bayesian Vector
MTI Discussion Paper 15, World Bank, Washington,
Autoregression.” In Handbook of Economic Forecasting
DC.
Volume 2, Part B, edited by G. Elliott and A.
Timmermann. 791-897. Amsterdam: Elsevier.
World Bank. 2022. Global Economic Prospects.
Kilian, L., and D. P. Murphy. 2014. “The Role of Washington, DC: World Bank.
Inventories and Speculative Trading in the Global
Market for Crude Oil.” Journal of Applied Econometrics World Bank. 2023. “Forecasting Industrial
29 (3): 454-78. Commodity Prices.” In Commodity Markets Outlook.
April. Washington, DC: World Bank.
Nixon, D., and T. Smith. 2012. “What Can the Oil
Futures Curve Tell Us About the Outlook for Oil Zhang, J., Y.-J. Zhang, and L. Zhang. 2015. “A Novel
Prices?” Bank of England Quarterly Bulletin 52 (1): Hybrid Method for Crude Oil Price Forecasting.”
39-47. Energy Economics 49 (C): 649-59.