Forecasting the Price of Oil

Ron Alquist Lutz Kilian Robert J. Vigfusson

Bank of Canada University of Michigan Federal Reserve Board
CEPR

Prepared for the

Handbook of Economic Forecasting
Graham Elliott and Allan Timmermann (eds.)

This presentation reflects the authors own views and should not be attributed to the Bank of
Canada, the Federal Reserve System, or the Board of Governors of the Federal Reserve System.
 Motivation
Potential to improve forecast accuracy of macroeconomic
aggregates and macroeconomic policy responses

Forecasts of the prices of oil and its derivatives like gasoline

or heating oil important for:
Modeling purchases of energy-intensive durables
Predicting carbon emissions and climate change
Designing regulatory policies (fuel standards, gasoline taxes)
Business decisions (e.g., airlines, auto manufacturers, utilities)
 Key Oil Price Series
Since 1948:
1. US Producer Price Index (PPI) for crude oil
2. West Texas Intermediate (WTI) price of crude oil (almost
identical with PPI series in pre-1974 era)

Since 1974:
1. Refiners acquisition cost (domestically produced crude oil)
2. Refiners acquisition cost (imported crude oil)
3. Refiners acquisition cost (composite)

No single oil price series is perfect for all purposes.

 Key Oil Price Series
Import price matters for standard theories of transmission
of oil price shocks and is best proxy for global price of oil.
Kilian (2009): Refiners acquisition cost (imports)

Retail energy price matters for theories of relative price

shocks.
Hamilton (2003): Crude oil PPI Retail gasoline price
Edelstein and Kilian (2009): Retail energy price Retail gasoline
price

US domestic price series subject to regulation until

early 1980s and unrepresentative of actual market price.
Mork (1989): Refiners acquisition cost (composite)
 Pre-1974 Nominal Price of Oil
Nominal Price of Oil: 1948.1-1973.12
4.5

Actual WTI
Random Draw from Fitted Model

U.S. Dollars/Barrel 4

3.5

2.5

2
1950 1955 1960 1965 1970
 The Nominal Price of Crude Oil
Nominal Price of Oil: 1948.1-1973.12 Nominal Price of Oil: 1974.1-2010.4
4.5 140

WTI WTI
RAC Domestic
120
RAC Imported
4

100
U.S. Dollars/Barrel

U.S. Dollars/Barrel
3.5
80

60
3

40

2.5

20

2 0
1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
 Percent Changes in the Real Price of Oil
Real Price of Oil: 1948.2-1973.12 Real Price of Oil: 1974.2-2010.4
50 50

WTI WTI
40 40 RAC Domestic
RAC Imported
30 30

20 20
Percent Change

Percent Change
10 10

0 0

-10 -10

-20 -20

-30 -30

-40 -40

-50 -50
1950 1955 1960 1965 1970 1975 1980 1985 1990 1995 2000 2005 2010
 Selective Survey of Topics in Chapter
1. Predictability in population

2. Forecasting the nominal price of oil

3. Forecasting the real price of oil

4. Joint forecasts of oil prices and US real GDP growth

5. Forecasting oil price volatility and quantifying oil price

risks

6. Conclusions: How to forecast the price of oil?

1. Predictability in Population
 Predictability in Population from Macroeconomic
Aggregates to Price of Oil
It has become more widely accepted that price of oil is
endogenous with respect to macroeconomic conditions.
Hamilton (1983): Fails to reject null of no Granger causality
from US macro aggregates to nominal oil prices pre-1973.

Thus, lagged macroeconomic aggregates should have

predictive power for price of oil in population.

Predictability in population is a precondition for out-of-

sample forecastability (Inoue and Kilian 2004).
 Predictors of Nominal Oil Prices
US price changes: CPI inflation, %M1 and %M2

Commodity prices: CRB Industrial Raw Materials Index,

CRB Metals Index

Other: 3-month T-bill rate, trade-weighted USD exchange

rate

Commodity currencies: AUD, CAD, NZD, SAR (Chen, Rogoff,

and Rossi 2010).
 Summary of Predictability Results
Strongest evidence of in-sample predictability for:
M1
CRB indices
Currencies of some industrial commodity exporters
(e.g., CAD)

Rejection of Granger non-causality at standard significance

levels for WTI and RAC
 Predictors of Real Oil Prices
Quarterly: US real GDP, world industrial production.

Monthly: CFNAI, US industrial production, OECD+6

industrial production, global real activity index (Kilian
2009).

Where applicable, Granger causality tests conducted on

filtered series (e.g., US real GDP):
Linear
Hodrick-Prescott
First difference
 Summary of Predictability Results
Strongest evidence of in-sample predictability for linearly
detrended series:
World industrial production
OECD+6 industrial production
Real activity index

Rejection of Granger non-causality at standard significance

levels for real WTI and RAC
 Why are linearly detrended global real activity
measures good at predicting real price of oil?
US is not the world
Oil price determined in global market

GDP poor proxy for business-cycle driven fluctuations in oil

demand because of large share of services
Industrial production is better indicator

Well-documented long swings in industrial commodity

prices such as oil
2. Forecasting the Nominal Price of Oil
 Do Oil Futures Prices Help Predict the Spot Price?
No-change (benchmark model)
+| = = 1, 3, 6, 9, 12

Futures Price
()
+| = = 1, 3, 6, 9, 12

Futures Spread

+| = 1 + + ln( / ) , = 1, 3, 6, 9, 12

where and are recursive OLS estimates from

+ = + + +
 Forecast Accuracy of Futures Prices
Forecast Evaluation Period: 1991.1-2009.12

Monthly Forecasts
Small ( 6%) improvements in forecast accuracy
Not statistically significant

Daily Forecasts
Short-horizon (1-12 months)
Similar to monthly forecasts, except at 12-month horizon.
Long-horizon (2-7 years)
No improvements in forecast accuracy
 Alternative Monthly Forecasting Methods
Local trends and structural change
Recursive drift
+| = 1 + = 1, , 12

Rolling drift
()
+| = 1 + = 1, , 12

Random walk in growth

+| = 1 + = 1, , 12

Hotelling (1931)
/12
+| = 1 + , = 3, 6, 12
 Alternative Monthly Forecasting Methods
CRB commodity prices
+| = 1 +
= 1, 3, 6, 9, 12

+| = 1 + , = 1, 3, 6, 9, 12

where ,

Commodity currencies (Chen, Rogoff, and Rossi 2010)

+| = 1 + = 1, 3, 6, 9, 12

+| = 1 + , = 1, 3, 6, 9, 12

where , ,
 Summary of Forecasting Results
CRB Indices
Up to 3-month horizon: Large (9-25%) and statistically
significant improvements in forecast accuracy.

Commodity Currencies
Up to 3-month horizon: Small (7-13%) but statistically
significant improvements in forecast accuracy for AUD and
CAD.
 Survey Forecasts
Monthly oil price forecasts from Consensus Economics, Inc.

+| = , = 3, 12

Quarterly oil price forecasts from Energy Information

Administration (EIA)

+| = , = 3, 12

Monthly Michigan Survey of Consumers (MSC) forecasts of

the price of gasoline
,
+| = , = 60
 Survey Forecasts Relative to No-Change Forecast
=3 = 12 = 60
MSPE Success MSPE Success MSPE Success
Ratio Ratio Ratio Ratio Ratio Ratio

+| = , 1.519 0.447 0.944 0.539 - -

+| = , 0.918 0.417 0.973 0.562 - -
,
+| = , - - - - 0.765 0.9071

+| = 1 + , - - 1.047 0.5661 - -

+| = 1 + , - - 1.016 0.5791 0.855 0.8111
NOTES: Boldface indicates statistical significance at the 10% level.
1
No significance test possible due to lack of variation in success ratio.
 Household Expectations of U.S. Retail Gasoline Prices (Cents/Gallon)
1992.11-2010.1
600

500
Expected price 5 years from now
400
Current price
300

200

100

0
1994 1996 1998 2000 2002 2004 2006 2008 2010

500

400 Expected real price 5 years from now

300 Current price
200

100

0
1994 1996 1998 2000 2002 2004 2006 2008 2010

600

500
Expected price 5 years from now
400
Actual price 5 years from now
300

200

100

0
1994 1996 1998 2000 2002 2004 2006 2008 2010
 Survey Forecasts Relative to No-Change Forecast
=3 = 12 = 60
MSPE Success MSPE Success MSPE Success
Ratio Ratio Ratio Ratio Ratio Ratio

+| = , 1.519 0.447 0.944 0.539 - -

+| = , 0.918 0.417 0.973 0.562 - -
,
+| = , - - - - 0.765 0.9071

+| = 1 + , - - 1.047 0.5661 - -

+| = 1 + , - - 1.016 0.5791 0.855 0.8111
NOTES: Boldface indicates statistical significance at the 10% level.
1
No significance test possible due to lack of variation in success ratio.
3. Forecasting the Real Price of Oil
 Forecasting Models of the Real Price of Oil
AR, ARMA, ARIMA

Kilian-Murphy (2010) VAR

1. Percent change in world oil production
2. Real activity index
3. Change in inventories
4. Real price of oil

Consider two specifications

Unrestricted VAR
Bayesian VAR (Giannone, Lenza, Primiceri 2010)
 Summary of Forecasting Results
AR(I)MA Models
Statistically significant improvements ( 17%) in forecast
accuracy at 1-3 month horizon.

VAR Models
UVAR: Statistically significant improvements in forecast
accuracy ( 19%) at 1-6 month horizon.

BVAR: Results similar to those from unrestricted VAR

Shrinkage improves forecast accuracy in more heavily
parameterized model with 24 lags.
 Forecasting Scenarios for Real Price of Oil based on Kilian-Murphy SVAR
Conditional Forecast Expressed Relative to Baseline Forecast
Forecast Adjustment Based on U.S. Oil Production Stimulus Scenario
100

50
Percent

-50

-100
2010 2011

Forecast Adjustment Based on 2007-08 World Recovery Scenario

100

50
Percent

-50

-100
2010 2011

Forecast Adjustment Based on Iran 1979 Speculation Scenario

100

50
Percent

-50

-100
2010 2011
4. Joint Forecasts of Oil Prices and US Real GDP
Growth
 Joint Forecasting Models
A key reason price of oil considered important is its
perceived predictive power for US real GDP.

Examining this predictive power requires joint forecasting

model of price of oil and domestic real activity.

Two Models:
VAR models
Nonlinear dynamic forecasting models
 Bivariate VAR Models
Unrestricted VAR
= 1 + 4=1 11, + 4
=1 12, + 1,
4 4
= 2 + =1 21, + =1 22, + 2,

where is the real price of oil; and is US real GDP.

Restricted VAR
Assume oil price is exogenous (12, = 0 )

Summary
Only small (1-8%) improvements in forecast accuracy at 3-8
quarter horizon relative to AR(4) benchmark.
 Autoregressive Forecasts of Cumulative Real GDP Growth based on the Real
Price of Oil
(a) Four Quarters Ahead
10
Percent at Annual Rates

Realizations
-5 AR Forecast
Linear VAR Forecast
-10
1992 1994 1996 1998 2000 2002 2004 2006 2008 2010

(b) One Quarter Ahead

10
Percent at Annual Rates

Realizations
-5 AR Forecast
Linear VAR Forecast
-10
1992 1994 1996 1998 2000 2002 2004 2006 2008 2010
 Possible Explanations for
Limited Success of Linear Forecasting Models
Reflects inability to forecast more accurately real price of oil
No. Can rule this out.
Underlying predictive relationship is weak
Predictive relationship between oil prices and domestic
macroeconomic aggregates is time-varying:
1. Variation in the expenditure share of energy (Edelstein and Kilian 2009)
2. Variation in the extent of price regulation (Ramey and Vine 2010)
3. Variation due to changes in composition of underlying oil demand and
oil supply shocks (Kilian 2009)

Linear forecasting models inherently misspecified

 Nonlinear Dynamic Forecasting Models
Relationship between oil prices and US real GDP may be
asymmetric in that only oil-price increases matter (Mork 1989;
Hamilton 1996, 2003).

Unrestricted multivariate nonlinear forecasting model

4 4
= 1 + =1 11, + =1 12, + 1,
4 4 4
= 2 + =1 21, + =1 22, + =1 + 2,

,+,3 ,+,1
where , , + is censored oil-price
,+,3
increase variable (e.g., = max{0, } and is the
maximum log real price of oil during the past 3 years).
 Summary of Nonlinear Forecasting Results
Can forecast US GDP growth using model with:
Exogenous oil prices (12, = 0 ) and

No feedback from lagged oil prices to GDP (22, = 0 )

Models that combine restrictions 12, = 0 and

22, = 0

3-year net nominal and real oil-price increases achieve

forecast-accuracy improvements for US real GDP growth at
4-quarter horizon.
 Nonlinear Forecasts of Cumulative Real GDP Growth: Forecasting
Success?
4-Quarter Ahead Recursive Forecasts of Cumulative Real GDP Growth
10
Percent (Annual Rates)

-5 Actual
Nonlinear Forecast Based on 3-Year Net Oil Price Increase
-10
1992 1994 1996 1998 2000 2002 2004 2006 2008 2010

4-Quarter Ahead Recursive MSPE Ratio Relative to AR(4) Benchmark

2

1.5
MSPE Ratio

0.5

0
1992 1994 1996 1998 2000 2002 2004 2006 2008
5. Forecasting Oil Price Volatility and
Quantifying Oil Price Risks
 12-Month Ahead Predictive Density of the Real WTI Price as of 2009.12
Based on No-Change Forecast
0.035

0.03

0.025

0.02

0.015

0.01

0.005

0
0 20 40 60 80 100 120 140 160 180 200
2009.12 Dollars/Barrel
 Alternative Measures of Nominal Oil Price Volatility
1-Month Implied Volatility
40

30
Percent

20

10

0
2002 2003 2004 2005 2006 2007 2008 2009 2010

Realized Volatility
40

30
Percent

20

10

0
2002 2003 2004 2005 2006 2007 2008 2009 2010

Recursive GARCH Volatility

40

30
Percent

20

10

0
2002 2003 2004 2005 2006 2007 2008 2009 2010
 Limitations of Volatility Measures
Models of delayed investment decisions require long-run,
real oil price volatility, not short-run nominal volatility.

Volatility is not risk.

Consumers are not concerned with real oil-price volatility
associated with price decreases.

Risks depend on the predictive distribution of the variable

of interest and on users loss function (Machina and
Rothschild 1987).

Need to distinguish between upside and downside risk.

 Oil Price Risks
Consider event of Rt h exceeding an upper threshold of (upside
risk) or falling below the lower threshold of (downside risk):


+ + , 0


+ + , 0
Examples:
1. = = 0 0 = Pr + <
0 = Pr + >
2. = = 1 1 = + |+ < Pr + <
1 = + |+ > Pr + > .
12-Month-Ahead Upside and Downside Risks in the Real WTI Price Based on
No-Change Forecast
Target Probabilities
1
UR(0)>80
DR(0)<45

0.5

-0.5

-1
2002 2003 2004 2005 2006 2007 2008 2009 2010

Probability Weighted Expected Excess/Shortfall

60
UR(1)>80
50
DR(1)<45
40

30

20

10

-10

-20

-30
2002 2003 2004 2005 2006 2007 2008 2009 2010
6. Conclusions: How to Forecast the Price of Oil?
 How to Forecast the Price of Oil?
Nominal price of oil
Futures prices generally poor predictors of spot prices
1-3 months: Adjust no-change forecast by recent price
change in industrial raw materials

6-48 months: No-change forecast

60 months: Adjust no-change forecast by expected
inflation

Real price of oil

1-6 months: Recursive VAR forecast
Beyond 6 months: No-change forecast
Background Slides
 Predictability from Nominal U.S. Aggregates to the Nominal Price of Oil
(p-values of the Wald test statistic for Granger Non-Causality)
Evaluation Period:
1973.1-2009.12 1975.2-2009.12

Monthly WTI WTI RAC RAC RAC Composite

Predictors: Oil Imports Domestic Oil
CPI 0.004 0.108 0.021 0.320 0.161

M1 0.181 0.039 0.010 0.000 0.000

M2 0.629 0.234 0.318 0.077 0.209

CRB Industrial Raw 0.000 0.000 0.000 0.000 0.000

Materials Index
CRB Metals Index 0.000 0.002 0.005 0.000 0.003

3-Month 0.409 0.712 0.880 0.799 0.896

T-Bill Rate
Trade-Weighted - 0.740 0.724 0.575 0.746
Exchange Rate
Predictability from Selected Bilateral Nominal Dollar Exchange Rates to the
Nominal Price of Oil
(p-values of the Wald test statistic for Granger Non-Causality)
Evaluation Period:
1973.1- 1975.2-2009.12
2009.12

Monthly WTI WTI RAC RAC RAC

Predictors: Oil Imports Domestic Oil Composite
Australia 0.038 0.066 0.073 0.017 0.044

Canada 0.004 0.003 0.002 0.006 0.002

New Zealand 0.128 0.291 0.309 0.045 0.169

South Africa 0.017 0.020 0.052 0.021 0.037

 Predictability from Selected Real Aggregates to the Real Price of Oil
(p-values of the Wald test statistic for Granger Non-Causality)
Evaluation Period:
1973.I-2009.IV 1975.II-2009.IV

Quarterly WTI WTI RAC RAC RAC

Predictors: Oil Imports Domestic Oil Composite
U.S. Real GDP
LT 0.353 0.852 0.676 0.397 0.561
HP 0.253 0.821 0.653 0.430 0.573
DIF 0.493 0.948 0.705 0.418 0.578
World Industrial
Production
LT 0.032 0.095 0.141 0.081 0.098
HP 0.511 0.766 0.800 0.665 0.704
DIF 0.544 0.722 0.772 0.668 0.691
 Predictability from Selected Real Aggregates to the Real Price of Oil
(p-values of the Wald test statistic for Granger Non-Causality)
Evaluation Period:
1973.1-2009.12 1976.2-2009.12
WTI WTI RAC RAC RAC Composite
Monthly Oil Imports Domestic Oil
Predictors: = 12 = 24 = 12 = 24 = 12 = 24 = 12 = 24 = 12 = 24
Chicago Fed
National Activity 0.823 0.951 0.735 0.952 0.881 0.998 0.707 0.979 0.784 0.995
Index (CFNAI)
U.S. Industrial
Production
LT 0.411 0.633 0.370 0.645 0.410 0.746 0.091 0.421 0.182 0.510
HP 0.327 0.689 0.357 0.784 0.415 0.878 0.110 0.549 0.194 0.668
DIF 0.533 0.859 0.458 0.866 0.473 0.909 0.114 0.490 0.222 0.699
OECD Industrial
Production1
LT 0.028 0.001 0.009 0.033 0.023 0.199 0.021 0.187 0.018 0.230
HP 0.195 0.034 0.072 0.278 0.138 0.714 0.121 0.530 0.114 0.706
DIF 0.474 0.060 0.130 0.353 0.182 0.741 0.174 0.604 0.209 0.757
Global Real Activity
Index 0.041 0.000 0.055 0.020 0.141 0.034 0.004 0.004 0.028 0.018
 Forecast Accuracy Relative to Monthly No-Change Forecast
Evaluation Period: January 1991- December 2009
() ()
1 + + ln( / )

MSPE Ratio Success Ratio MSPE Ratio Success Ratio

1 0.988 0.465 1.001 0.539
3 0.998 0.465 1.044 0.531
6 0.991 0.509 1.051 0.535
9 0.978 0.548 1.042 0.583
12 0.941 0.557 1.240 0.537

NOTES: Boldface indicates statistical significance at the 10% level.

Forecast Accuracy Relative to Daily No-Change Forecast
Evaluation Period: Since January 1986
()

MSPE Ratio Success Ratio
1 0.963 0.522
3 0.972 0.516
6 0.973 0.535
9 0.964 0.534
12 0.929 0.562

NOTES: There are 5968, 5926, 5861, 5744, and 5028 daily observations at
horizons of 1 through 12 months, respectively. Boldface indicates statistical
significance at Leamers (1978) critical value.
Forecast Accuracy Relative to Daily No-Change Forecast
(in years) Starting Sample MSPE Success
date size Ratio Ratio
2 11/20/90 3283 1.159 0.515

3 05/29/91 515 1.168 0.518

4 11/01/95 194 1.212 0.294

5 11/03/97 154 1.280 0.247

6 11/03/97 134 1.158 0.276

7 11/21/97 22 1.237 0.500

NOTES: Boldface indicates statistical significance using Leamers

(1978) critical value.
 Forecast Accuracy Relative to No-Change Forecast
Evaluation Period: January 1991- December 2009
, ,
1 + , 1 + ,
MSPE Ratio Success Ratio MSPE Ratio Success Ratio
1 0.913 0.583 1.031 0.579
3 0.782 0.601 0.750 0.601
6 1.055 0.583 1.219 0.623
9 1.076 0.553 1.304 0.575
12 1.035 0.548 1.278 0.539

NOTES: Boldface indicates statistical significance at the 10% level.

 Recursive Forecasts of Real Price of Oil from AR and ARMA Models
U.S. Refiners Acquisition Cost for Imported Crude Oil
Evaluation period: 1991.12-2009.8
=1 =3 =6 =9 = 12
MSPE SR MSPE SR MSPE SR MSPE SR MSPE SR
AR(12) 0.849 0.599 0.921 0.552 0.969 0.522 1.034 0.441 1.022 0.517
AR(24) 0.898 0.576 0.978 0.557 1.008 0.565 1.056 0.446 1.058 0.453
AR(SIC) 0.826 0.613 0.936 0.557 1.015 0.488 1.039 0.515 1.007 0.532
AR(AIC) 0.842 0.613 0.940 0.562 0.983 0.483 1.013 0.500 0.989 0.527
ARMA(1,1) 0.837 0.580 0.932 0.514 0.982 0.493 1.006 0.510 0.992 0.527
ARI(11) 0.856 0.604 0.939 0.571 1.003 0.517 1.095 0.471 1.091 0.512
ARI(23) 0.898 0.561 0.978 0.538 1.009 0.546 1.068 0.500 1.068 0.508
ARI(SIC) 0.833 0.594 0.951 0.605 1.041 0.546 1.053 0.505 1.016 0.527
ARI(AIC) 0.849 0.604 0.958 0.605 1.008 0.556 1.042 0.500 1.015 0.527
ARIMA(0,1) 0.841 0.599 0.945 0.581 1.009 0.546 1.032 0.515 1.017 0.512
NOTES: Boldface means significance at the 10% level.
 Recursive Forecasts of Real Price of Oil from
Kilian-Murphy (2010) VAR Model
U.S. Refiners Acquisition Cost for Imported Crude Oil
Evaluation period: 1991.12-2009.8
UVAR BVAR
MSPE Ratio SR MSPE Ratio
12 1 0.814 0.561 0.800
3 0.834 0.567 0.876
6 0.940 0.546 0.967
9 1.047 0.564 1.052
12 0.985 0.632 1.004
24 1 0.961 0.561 0.801
3 1.081 0.614 0.883
6 1.298 0.604 0.993
9 1.476 0.583 1.095
12 1.415 0.647 1.059
NOTES: Model includes the four oil market variables used in Kilian and Murphy (2010).
BVAR based on Giannone, Lenza, and Primicieri (2010).
 MSPE Ratios of VAR and VARX Models Relative to AR(4) Benchmark
Cumulative U.S. Real GDP Growth Rates
Evaluation Period: 1990.Q1-2010.Q2
Real RAC Price of Imports Nominal RAC Price of Imports

Horizon Oil Price Oil Price Oil Price Oil Price

Endogenous Exogenous Endogenous Exogenous
1 1.10 1.10 1.11 1.11
2 1.04 1.04 1.05 1.05
3 0.99 0.98 1.00 0.99
4 0.97 0.96 0.98 0.97
5 0.96 0.95 0.96 0.96
6 0.95 0.94 0.95 0.95
7 0.92 0.92 0.92 0.92
8 0.92 0.92 0.92 0.92
MSPE Ratios for Alternative Specifications of Restricted Exogenous Model
Cumulative U.S. Real GDP Growth Rate
Oil Price Series 1990.Q1-2010.Q2 1990.Q1-2007.Q4
Horizon Horizon
h 1 h4 h 1 h4
Real RAC imports 0.91 0.85 1.11 1.71
RAC composite 1.16 0.99 1.49 2.05
RAC domestic 1.23 0.89 1.55 1.73
WTI 1.03 0.70 1.23 1.22
PPI 1.24 1.09 1.63 2.28

Nominal RAC imports 1.01 0.74 1.22 1.37

RAC composite 1.26 0.82 1.58 1.54
RAC domestic 1.23 0.80 1.50 1.40
WTI 0.92 0.66 1.02 1.08
PPI 1.23 0.88 1.59 1.78