International Journal of Forecasting 21 (2005) 491 – 501

www.elsevier.com/locate/ijforecast

A monthly crude oil spot price forecasting model using

relative inventories
Michael Yea,T, John Zyrenb, Joanne Shoreb
a
Department of Economics, St. Mary’s College of Maryland, St. Mary’s City, MD 20686, USA
b
Petroleum Division, EI-42, Office of Oil and Gas, Energy Information Administration, U.S. Department of Energy, 1000 Independence Ave.,
SW, Washington, DC 20585, United States

Abstract

This paper presents a short-term forecasting model of monthly West Texas Intermediate crude oil spot prices using readily
available OECD industrial petroleum inventory levels. The model provides good in-sample and out-of-sample dynamic
forecasts for the post-Gulf War time period. In-sample and out-of-sample forecasts from the model are compared with those
derived from other models. The model is intended for the practicing forecaster and designed to be simple enough to implement
easily in a spreadsheet or other software package, with the variables easy to update. The simplicity and ease of updating make
this model attractive for investigating various scenarios to see the impacts that market changes can have on monthly crude oil
spot prices if inventories, production, imports, or demand change. Finally, the model structure can easily be updated periodically
should there be a fundamental market change or a shift in the normal level of inventories.
D 2005 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.

Keywords: Forecast modeling; Crude oil price; Petroleum inventory

1. Introduction available data. The objective of this model is to

provide a dynamic forecast of monthly West Texas
In this paper, we develop a simple and practical Intermediate (WTI) prices for the post Gulf War I time
model for forecasting monthly crude oil spot prices period using readily available data (WTI spot prices
under normal market circumstances1 using readily and OECD petroleum inventories) readily available
and that can be implemented in spreadsheet applica-
tions. Thus, the model is intuitively appealing and
T Corresponding author. Tel.: +1 240 895 4696; fax: +1 240 895 useful to industry and government decision-makers in
4450. forecasting prices and in investigating the impacts on
E-mail address: mhye@smcm.edu (M. Ye).
1 price of changing market fundamentals, such as
This model deals with the fundamental relationship between
inventories and prices and does not take into account transitory petroleum inventories, production, imports, and
geopolitical situations giving rise to various risk premiums, such as demand. Because of the few variables involved, the
the War premium or terror premium. results from the forecast model can be easily
0169-2070/$ - see front matter D 2005 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.
doi:10.1016/j.ijforecast.2005.01.001
492 M. Ye et al. / International Journal of Forecasting 21 (2005) 491–501

interpreted. The model is also practical from a inventory levels are a measure of the balance, or
maintenance standpoint in that it is based on two imbalance, between petroleum production and
readily available data series. demand, they reflect changing market pressures on
We focus on the concepts of normal and relative crude oil prices, and thus provide a good market
levels of petroleum market variables such as demand, barometer for crude oil price change in the short run.
field production, net imports, and inventory. This is Intuitively, relative inventory levels must be nega-
similar to the natural rate theory widely applied in tively related to prices, with distributed lagged effects.
macroeconomic modeling and empirical studies. We The observed behavior of WTI spot prices and total
decompose the observed level of a petroleum market OECD industrial inventories since Gulf War I supports
variable into two components: the normal level, this argument. The large swings in WTI spot prices
determined by historical seasonal movements and during the late 1990s are coupled with counter-swings
trends, which reflects the normal market demand and in OECD total inventories, particularly when meas-
operational requirements; and the relative level, the ured in relative terms. For example, in 1998 produc-
difference between the observed and normal levels, tion consistently exceeded demand; as a result,
which reflects short-run market fluctuations. Season- inventories grew to unusually high levels. Demand
ality exists in petroleum market variables such as growth slowed during this period in part due to warm
demand, field production, and net imports. Since winters in the Northern Hemisphere and the Asian
change in inventory equals the difference between financial crisis, while supply increased substantially as
demand and supply (the sum of field production and Iraqi crude oil came back into the export market in
net imports), petroleum inventory demonstrates sea- 1997 through the bOil-for-FoodQ Program. WTI spot
sonal movements as well. Long-term inventory trends prices fell to near $10 per barrel by the end of 1998 due
exist due to the trends in government inventories, to the excess of production over demand and the
increases in long-term product demand, and the larger resulting inventory build. The supply/demand balance
distribution and storage infrastructure needed to meet reversed in 1999 when OPEC cut back production to a
demand growth and product differentiation. It is the level well below demand, and the demand for crude oil
relative levels of demand, field production, net simultaneously increased as the Asian economies
imports, and inventory that respond to the unexpected recovered. With demand exceeding production during
short-run variations in the market. 1999, inventories were drawn down to help meet
Among these relative level variables, it is the demand. The excess inventories that had been built up
relative inventory that matters most in the short run fell rapidly to well below normal levels, and WTI spot
for the determination of crude oil price.2 On the prices correspondingly rose to over $30 per barrel by
demand side, many consumers have few options and early March 2000. More recently, we have observed
little discretion as to how much volume they can use that the loss of Venezuelan crude oil production during
in the short term. People must heat their homes in the the oil workers’ strike in that country resulted in a large
winter and drive to work regardless of price. This drop in U.S. crude oil inventories, while demand
implies that short-term demand elasticity is small. On changed very little, even though crude oil prices
the supply side, supply chains that move crude oil climbed about $5 per barrel.3
from its source of production to refineries, and then
move product from refineries to consumers, are long
and cannot be adjusted quickly when unexpected 2. A forecast model
shifts in supply or demand occur. Empirically, we
have found that the short-run demand elasticity is The basic idea underlying the forecast model is the
indeed much less than the short-run inventory natural rate theory. Theoretically, the normal inventory
elasticity in the United States and the production
elasticity is virtually zero. Since total petroleum 3
Energy Information Administration, data from Weekly Petro-
leum Status Report and discussions from EIA’s web summary of the
2
For example, see Ye, Zyren, and Shore (2002) and Ye, Zyren, weekly data, bThis Week in Petroleum,Q during January 2003. http://
and Shore (2003). www.eia.doe.gov/oil_as/petroleum/info_glance/petroleum.html.
 M. Ye et al. / International Journal of Forecasting 21 (2005) 491–501 493

level is determined by normal demand and normal Table 1

operating requirements, while the relative inventory Estimation used to create the normal inventory variable
level responds to market fluctuations. For practical Variable Coefficient t-statistic
purposes, we calculate the normal inventory level by a0 2581.334 106.029
appropriately de-seasonalizing and de-trending histor- T 0.204 1.367
February 42.313 1.375
ical data.4 The maintained hypothesis is that the
March 50.016 1.625
average inventory level over a sufficiently long period April 26.275 0.854
of time will be the desired level. Relative inventory May 26.937 0.858
level is defined as the deviation of actual inventories June 36.666 1.167
from a historically determined normal level. Invento- July 57.690 1.837
August 73.411 2.337
ries are seasonal, building in the summer months and
September 79.349 2.526
dropping during the winter because their build-ups October 74.710 2.378
and draw-downs reflect the seasonal imbalance November 62.240 1.981
between supply and demand as a result of the supply December 4.0458 0.131
of crude oil exhibiting less seasonal variation than
demand. Long-term trends may also exist, mainly due Monthly data for the industrial inventories of crude
to the impact which changes in government invento- oil and petroleum products for all OECD countries are
ries have on commercial inventories. Such trends will available from January 1988 to the present; however,
also be considered when defining the normal pattern. the International Energy Agency (IEA) changed its
Explicitly, the relative inventory level, denoted by data collection methodology in December 1990.5 We
RINt , is defined as limited the current study to the period from January
1992 to April 2003 to avoid the impact that Gulf War I
RINt ¼ INt  IN4t ð1Þ had on markets and to limit our analysis to a
where INt is the actual industrial OECD petroleum consistent data series. The ordinary least-square
inventory level in month t, and INt * is the normal estimation results in Table 1 show a statistically
level. Letting D k , k=2, 3, . . ., 12 be 11 seasonal significant seasonal pattern in inventory from Decem-
dummy variables and T be a linear trend, the normal ber 1990 to April 2003, in which WTI spot price is in
inventory level is calculated by nominal dollars per barrel and the inventory level is
measured in million barrels.
X
12
The statistical evidence for a long-term trend is
IN4t ¼ a0 þ b1 T þ bk D k ð2Þ
k¼2
marginal. However, with the recent world-wide
emphasis on larger strategic reserves for petroleum
where a 0, b 1, and b k , k=2, . . ., 12 are estimated and changes in inventory holding strategies, such as
coefficients from de-trending and de-seasonalizing the just-in-time replacement and product proliferation
observed total petroleum inventory. requirements, we include a trend for completeness in
case it will be needed in the near future for model-
4
For example, this is similar to the theoretical and empirical updating purposes.6 Normalized total OECD indus-
treatments of potential GDP. Theoretically, potential GDP is trial petroleum inventory is used as a surrogate for
determined by the labor force being at normal level, i.e., the desired world petroleum inventories.
unemployment rate at its natural rate, and the capital utilization rate
being at its normal rate. Empirically, potential GDP is the properly
de-trended historical GDP levels. In the same vein, monetary
movements are decomposed into an anticipated component and an
5
unanticipated component and the fiscal deficit is decomposed into IEA regularly revises the inventory data series. Therefore, newly
cyclical deficit (the normal deficit) and structural deficit (the relative downloaded series may not generate identical results to those
deficit). For related literature, see for example, Friedman (1968), presented in this paper. In this paper, we take petroleum inventories
Phelps (1970), Gordon and Hynes (1970), Lucas (1972, 1973), and as given, either as data or provided by another (forecasting) source.
Sargent (1976). In their study on net imports of crude oil, Husted All forecast results are conditional on this pre-existing data set.
6
and Kollintzas (1987) applied a similar idea and decomposed net Prediction results will change only negligibly when a trend is
imports to temporary and permanent components. not inserted into the normalizing equation.
494 M. Ye et al. / International Journal of Forecasting 21 (2005) 491–501

For crude oil prices, we use the nominal WTI crude in which subscript t is for the tth month; subscript i is
oil spot price, which is considered a world benchmark for ith month prior to the tth month; a, b i , c j , d, and e,
crude oil spot price. These daily spot prices were i=0, 1, 2, 3 and j=0, 1, 2, . . ., 5 referring to the 6
obtained from Reuters. Since OECD inventory data months from October 2001 to March 2002, are
are only available monthly, the daily crude oil spot coefficients to be estimated; Dj911 is a set of single
prices were aggregated to a monthly frequency. monthly variables to account for market disequili-
A number of factors were considered in the search brium following the September 11, 2001 terrorist
for the best forecast model. Firstly, we found, attacks in the United States; and LAPR99 is a level-
consistent with those findings in the literature, that shifting variable corresponding to the effect that
WTI price has a unit root while relative inventory may OPEC quota tightening had on the petroleum market
not.7 We decided not to use the first-order differences beginning in April 1999 (see Appendix B).
of WTI price and inventory because of the resulting The correlogram of squared residuals indicates a
diminished forecasting ability. Secondly, we found potential correlation at lag 3. An ARCH LM test (lag
that the sum of crude oil inventory and oil product 3) confirms the possibility of conditional variance at
inventories usually performed better than crude oil the 10 percent level of significance ( p=0.081).
inventory alone.8 Thirdly, we also found that inclusion Moreover, the correlogram of residuals indicates a
of a series of low stock indicators to capture the potential moving average correlation at lag 2. A
asymmetric market behavior will marginally reduce Breusch–Godfrey serial correlation LM test (lag 2)
in-sample forecast error, but will increase out-of- confirms this at the 5% confidence level ( p=0.011).
sample forecast error; thus, we decided not to include However, when the RSTK forecast equation includes
them. Fourthly, various lag lengths for the relative a MA(2) correction, the 1-month out-of-sample
inventory variables were also investigated. We forecast results are only marginally better, while the
included three lags plus the current period of relative 3-month out forecasts are slightly worse than the
inventories, based on the adjusted R 2, AIC, and SBC, uncorrected model of Eq. (3).
as well as in-sample and out-of-sample forecast Estimation results for the best-fit forecast equation
results. Details can be found in Appendix A. Finally, for the January 1992 to April 2003 time period are
we investigated the potential role of GDP in the shown in Table 2. The inventory variables for the most
model. We decided not to use it for two reasons. One part show the expected inverse relationship between
is that we have found no source of future prediction inventory size and price, where the current and lagged
for a monthly OECD GDP aggregate, which makes its coefficients on RINt are negative. The price effect for a
use in a simple forecast model impractical. We also
found that U.S. GDP alone does not provide good Table 2
forecast results in our model. The RSTK forecast equation
The best specification we find for the short-run Variable Coefficient Standard error t-statistic
forecast of monthly WTI spot price is the Relative a 6.450 1.126 5.726
Stock (RSTK) model RIN 0.009 0.004 2.236
RIN(-1) 0.010 0.006 1.731
X
3 X
5 RIN(-2) 0.014 0.006 2.518
WTIt ¼ a þ bi RINti þ cj Dj 911 RIN(-3) 0.010 0.004 2.123
i¼0 j¼0 DOCT01 3.613 1.346 2.685
DNOV01 4.040 1.376 2.936
þ dLAPR99 þ eWTIt1 þ et ð3Þ DDEC01 2.445 1.393 1.756
7 DJAN02 1.009 1.406 0.718
The existence of a unit root in the relative inventory series is
DFEB02 0.353 1.400 0.252
rejected at 0.95. Detailed unit root test results for all series are
DMAR02 2.678 1.361 1.967
available upon request. For unit root test results in the literature, see
LAPR99 2.921 0.550 5.314
for example, Arize (2000), Bentzen (1997), and Jones (1993).
8 WTI(-1) 0.658 0.059 11.104
However, at present, the OECD and petroleum trade literature
provides forecasts of future values only for combined industrial
Adj. R 2 Durbin-h AIC SBC
crude oil and petroleum products inventories. Thus, for all practical
0.940 1.525 3.468 3.746
purposes, this is the inventory variable that must be utilized.
 M. Ye et al. / International Journal of Forecasting 21 (2005) 491–501 495

Table 3 Table 4
NAIV model MALT model
Variable Coefficient Standard error t-statistic Variable Coefficient Standard error t-statistic
a 20.302 1.079 18.810 a 16.783 6.225 2.696
DOCT01 3.438 1.446 2.378 T 0.001 0.006 0.214
DNOV01 5.117 1.874 2.730 STK(-1) 0.005 0.002 2.134
DDEC01 5.191 2.034 2.552 ANN(-1) 0.003 0.002 1.938
DJAN02 4.755 2.031 2.341 DOCT01 3.544 1.447 2.449
DFEB02 4.005 1.860 2.154 DNOV01 2.796 1.453 1.925
DMAR02 1.004 1.442 0.696 DDEC01 1.691 1.457 1.160
LAPR99 4.002 1.348 2.969 DJAN02 1.113 1.470 0.757
AR(1) 0.953 0.035 27.378 DFEB02 0.048 1.473 0.032
AR(12) 0.090 0.037 2.392 DMAR02 2.796 1.468 1.905
LAPR99 2.084 0.603 3.458
Adj. R 2 DW AIC SBC WTI(-1) 0.757 0.044 17.044
0.918 1.891 3.760 3.974
Adj. R 2 Durbin-h AIC SBC
0.930 1.266 3.611 3.868
stock change is 0.0406 $/MMBbl, which is equiv-
alent to $1.22 for a 1 million barrel-per-day change
in production or demand.9 The anomalous positive predicted values in a spreadsheet model when these
lag 2 RIN coefficient is important for adding short-term data become available. Over time, forecast equation
structure to the dynamic forecast. The first few 9/11 coefficients may need to be re-estimated to improve
dummy variables are significant and negative to reflect accuracy. Normal levels may need to be re-estimated
the temporary effect that the terrorist incident had on to reflect changes in trend and infrastructures.
the U.S. economy and on petroleum product demand in
particular. The JAN02 and FEB02 variables are not
necessary and are included for convenience; their 3. Forecast evaluation
removal has little effect on the full period forecast
results. The MAR02 dummy is significant and To evaluate the relative inventory (RSTK) model
positive to reflect the return to economic normalcy developed in the last section, we compare in-sample
and the accompanying increased demand for petro- and out-of-sample forecast results with two alternative
leum products.10 The LAPR99 shift variable shows models. The first alternative model is a naV̈ve
the overall increase in prices resulting from OPEC autoregressive forecast (NAIV) model
tightening their control over world crude oil X
5
markets.11 Finally, WTI(-1) illustrates the autore- WTIt ¼ a þ bARð1Þ þ cj Dj 911 þ dLAPR99
gressive tendencies of petroleum prices. i¼0
The RSTK model is simple to implement, main- þ eARð12Þ þ et ð4Þ
tain, and update in a spreadsheet environment. The
forecaster may simply enter new observations and in which AR(1) and AR(12) are the 1st-order and 12th-
order autoregressive terms, respectively. Other varia-
9
The price effect (including the feedback) is calculated from the
model coefficients in the normal manner, i.e., (sum of RIN Table 5
coefficients)/(1coefficient of lagged WTI), or 0.013892/ Dynamic forecasts evaluation statistics
(10.657855). This calculated value of 0.041 is similar to the NaRve MALT RSTK
0.037 value for the price effect obtained by model simulation
studies. RMSE 4.150 1.747 1.538
10
Results without D911 dummy variables can be found in MAE 3.301 1.314 1.173
Appendix A. MAPE 16.428 6.242 5.562
11
Equivalently, Chow test results show that there is a break point Theil U 0.095 0.040 0.035
at April 1999 at more than 99.9% significance level. Further Bias P 0.001 0.002 0.000
statistical and market justifications for the level shift at April 1999 Variance P 0.641 0.107 0.046
can be found in Appendix B. Covariance P 0.359 0.892 0.953
496 M. Ye et al. / International Journal of Forecasting 21 (2005) 491–501

bles are as defined in Eq. (3) to ensure comparability.

Table 3 shows the estimates for the naV̈ve model’s full
period (January 1992 to April 2003).12
The other comparison model is the Modified
Alternative (MALT) model, specified as follows:13
X
5
WTIt ¼ a þ bWTIt1 þ cj Dj 911
i¼0
þ dLAPR99 þ eSTKt1 þ f ANNt1
þ gT þ et ð5Þ
in which the subscript t is for the tth month, STK is Fig. 1. Comparison of dynamic forecasts and actual WTI price.
the total OECD industrial inventory, ANN is the 12-
month difference of STK, i.e., ANN t =STK t  indicating that there is little systematic forecast
STKt12, and T is a linear trend variable. Table 4 error.
shows the full period (January 1992 to April 2003) Fig. 1 shows the in-sample dynamic forecasts for
estimation results for the MALT model.14 the entire sample period produced by the models, i.e.,
In-sample forecast comparisons for the three a multi-step dynamic forecast using the actual value of
models can be found in Table 5, which summarizes the exogenous variables, but using model-predicted
the in-sample forecast root mean square (RMSE), values for the forecasting endogenous variable. As
mean absolute deviation (MAE), mean absolute can be seen, the RSTK creates a much better in-
percent error (MAPE) statistics, the Theil inequality sample forecast than the other two models.
coefficient, and the bias, variance and covariance The process of generating out-of-sample fore-
proportions for the three models in the estimation casts begins by fitting the models for the January
sample period.15 By all of these statistics, RSTK 1992 to December 1999 time period. A 6-month
model has the best in-sample forecasting ability. out-of-sample forecast was then generated (for the
The scale-sensitive statistics (RMSE, MAE, and months January 2000 to June 2000). Then one
MAPE) show that the NAIV model has the largest observation is added, and the models refitted
forecast errors while the RSTK model has the through January 2000, and another 6-month out-
smallest; the scale insensitive Theil U statistic of-sample forecast is generated (for February 2000
confirms this. Decomposition of the Theil U into through July 2000). This process is repeated until
bias, variance, and covariance shares shows that March 2003, when only a 1-month-ahead forecast
none of the models have significant forecast bias, can be generated (for April 2003). Thus, there are
but that the NAIV model replicates price variability 40 one-month-ahead forecasts, 39 two-month-ahead
poorly while the RSTK model is best. The RSTK forecasts, 38 three-month-ahead forecasts, and 35
model shows a very high covariance proportion, six-month-ahead forecasts available for the model
evaluation.16
12
April 1999 dummy variable is also statistically justified for the As can be seen from Table 6, the RSTK model
same reason stated in footnote 11. performs better in dynamic forecasts. The RMSE
13
The ALT model has been used internally at EIA to help evaluate
crude oil price forecasts. The original ALT model specification does
16
not include the LAPR99 and the six 9/11 dummy variables. Because Due to the particular model formulations used, two subtleties
the ALT model forecasts are improved when the dummy variables need be mentioned. The first is caused by the 9/11 dummy
are included (details are available upon request), these variables variables, which are only added to the model as the estimation
were added to the model to facilitate comparisons. period passes over them. The second is due to the nature of the
14
April 1999 dummy variable is also statistically justified for the relative inventories: the normal inventories were recalculated for
same reason stated in footnote 11. each estimation period to include only those observations available
15
Formulae for RMSE, MAE, MAPE, and Theil U can be found in for the estimation. For each model re-estimation, estimates of out-
Diebold (2001, Chapter 12) and reference to the bias, variance and of-sample normal inventories were then made beginning after the
covariance proportion in Pindyck and Rubinfeld (1991, Chapter 11). last in-sample period.
 M. Ye et al. / International Journal of Forecasting 21 (2005) 491–501 497

Table 6
Out-of-sample dynamic forecast statistics
1-month-ahead 2-month-ahead 3-month-ahead 6-month-ahead
NAIV MALT RSTK NAIV MALT RSTK NAIV MALT RSTK NAIV MALT RSTK
RMSE 2.64 2.25 2.13 3.56 3.00 2.78 3.82 3.07 2.91 5.36 3.20 3.06
MAE 2.21 1.85 1.75 2.90 2.37 2.22 3.28 2.49 2.42 4.31 2.68 2.42
Range 10.91 8.63 7.59 14.38 12.86 10.83 14.09 12.66 12.13 12.73 11.53 11.52
Mean error 0.32 0.01 0.09 1.04 0.14 0.08 1.83 0.31 0.06 4.06 0.78 0.02

and MAE increase for all models as the forecast inventories as the only independent variable.19 This
period lengthens. The absolute value of the mean practical model is needed because we found that
is largest for NAIV and smallest for the RSTK existing simple models lack a theoretical basis and
model in all forecast periods (except one-month- more complex models (in the general equilibrium
ahead). In general, the range tends to become framework or addressing specific theoretical and
larger as the forecast period lengthens, but for all econometric issues) provide insight into fundamen-
periods, the NAIV model has the largest range and tal understanding of the economics behind the
the RSTK has the smallest. market behavior, but require too much expertise
Figs. 2a and b compare the dynamic out-of- and specific data to be easily implemented and thus
sample 1-month-ahead and 3-month-ahead forecast are not practically useful. In addition to providing
errors for the three models for the period from good forecasting results, a desirable feature of the
January 2000 to January 2003.17 These figures show model presented in the paper is that it can readily
that the RSTK model produces the best out-of-sample be implemented in a spreadsheet or other software
forecast results and that the NAIV model has the package, and can be easily changed or updated. An
worst. In both periods, all three models had the largest important feature of any forecasting model is that a
negative errors at price troughs and the largest forecast of the independent variables be readily
positive errors when prices peak. The RSTK model available. On a routine basis, the U.S. Energy
produced worse forecasts than the MALT model for Information Administration forecasts monthly
the 6-month 9/11 period, where dummy variables OECD combined crude oil and oil products
were used in the forecasting equation. Additionally, inventory levels in conjunction with its Short-Term
for the 3-month out forecasts, the MALT model Energy Outlook (STEO).20 EIA is currently using
sometimes had slightly better results in a price trough, the RSTK model, among several other models, to
whereas the RSTK model had better performance in investigate future price impacts of demand or
price peaks.18 production changes and to estimate price effects
of market disruptions by using monthly inventories
derived from a 2-year forecast of OECD supply-
4. Conclusion demand balances published in EIA’s STEO.
A number of areas merit further exploration for
This paper presents and evaluates a simple and potential model improvement. First, it would be
practical dynamic forecast model using petroleum interesting to see how inventory levels may relate to
crude oil market volatility. For example, does higher

17 19
Results for the complete set of dynamic forecast errors from 1- This paper can be viewed as a sequel to Ye, et al. (2002), in
month-ahead to 12-month-head are available upon request. which an early version of the model was presented without out-of-
18
The 2- and 3-month out-of-sample forecast errors for all three sample forecasts, evaluations, and comparisons with alternative
models demonstrated the (horizon1) MA property. However, the models.
20
1-month errors for the NAIV model were MA(1) at 5% significance Additionally, other organizations’ forecasts of OECD supply
level ( p=0.035) while those for the RSTK model were MA(1) at and demand, such as those projected by IEA, can be used to derive a
10% ( p=0.084). forecast of inventory levels.
498 M. Ye et al. / International Journal of Forecasting 21 (2005) 491–501

price volatility occur when relative inventories are low Table A.1
compared to normal? Regression summary statistics
A second possible extension would be the inves- Adj. R 2 AIC SBC
tigation of abnormal market behavior such as the Gulf Lags 0 to 1 0.94 3.49 3.73
War or periods of low or high inventories. It is not Lags 0 to 2 0.94 3.49 3.75
Lags 0 to 3 0.94 3.47 3.75
clear if market variable relationships are the same
Lags 0 to 4 0.94 3.48 3.78
during normal situations as during very tight or very Lags 0 to 5 0.94 3.47 3.79
loose markets. Inventory non-linearity needs to be Lag 1 0.94 3.51 3.73
carefully investigated to see if it can capture the Lags 1 to 2 0.94 3.51 3.75
omitted market psychological factors. No 9/11 dummies (Lags 0 to 3) 0.93 3.55 3.70
Finally, it is realized that crude oil and oil
product inventories may demonstrate statistically
significantly different features such as seasonality, available, one may propose, based on the outcome
trend, and variability. Therefore, it may be fruitful of investigation, the creation of alternative predicted
to conduct a full investigation on the potentially variables such as crude oil and oil products
separate roles of crude oil inventories and oil inventories separately.
products inventories. Even though there are pre-
dicted values for combined inventories currently
Acknowledgements

This work is partially sponsored by the Office of

Petroleum Reserves, U.S. Department of Energy. The
authors thank two anonymous referees for many
helpful comments.

Appendix A. Search the best RSTK model

specification

This appendix records our search for the best

specification among eight cases: seven cases with
various numbers of lagged OECD industrial relative
inventory levels and one case without the 9/11
variable.
Table A.1 gives the summary regression statistics
of these eight cases, in which AIC and SBC are the
Akaike Information Criterion and Schwarz Criterion,
respectively. Note that there are six 9/11 variables,
so that the SBC for the model without these
variables is lower than in the other models which
includes them. However, the higher AIC (and lower
adjusted R 2) for the 9/11-deficient model does
indicates a poorer fit.
Table A.2 shows the in-sample dynamic forecast21
error summary statistics for the eight specifications for

21
Fig. 2. (a) Comparisons of dynamic forecast errors—1-month- Static in-sample and out-of-sample forecast results are available
ahead; (b) comparisons of dynamic forecast errors—3-month-ahead. upon request.
 M. Ye et al. / International Journal of Forecasting 21 (2005) 491–501 499

Table A.2
In-sample dynamic forecast summary
RMSE MAE MAPE Theil Bias P Var P CoVar P
Lags 0 to 1 1.58 1.21 5.76 0.04 0.00 0.05 0.95
Lags 0 to 2 1.59 1.22 5.84 0.04 0.00 0.05 0.95
Lags 0 to 3 1.54 1.17 5.56 0.03 0.00 0.05 0.95
Lags 0 to 4 1.54 1.17 5.55 0.03 0.00 0.05 0.95
Lags 0 to 5 1.52 1.19 5.62 0.03 0.00 0.04 0.95
Lag 1 1.58 1.23 5.82 0.04 0.00 0.07 0.93
Lags 1 to 2 1.58 1.23 5.82 0.04 0.00 0.07 0.93
No 9/11 dummies (Lags 0 to 3) 1.86 1.38 6.55 0.04 0.00 0.06 0.94

Table A.3
Out-of-sample dynamic forecast summary
Average error RMSE MAE Range
1M 3M 6M 1M 3M 6M 1M 3M 6M 1M 3M 6M
Lags 0 to 1 0.09 0.07 0.01 2.16 2.94 3.06 1.78 2.45 2.44 7.78 12.17 11.40
Lags 0 to 2 0.09 0.08 0.00 2.14 2.93 3.05 1.78 2.43 2.42 7.81 12.10 11.42
Lags 0 to 3 0.09 0.06 0.02 2.13 2.91 3.06 1.75 2.42 2.42 7.59 12.13 11.52
Lags 0 to 4 0.07 0.04 0.08 2.14 2.97 3.13 1.77 2.48 2.48 7.73 12.24 12.19
Lags 0 to 5 0.11 0.14 0.02 2.13 2.92 3.12 1.78 2.49 2.49 7.68 12.00 11.77
Lag 1 0.09 0.09 0.02 2.22 3.04 3.18 1.86 2.52 2.54 7.84 12.67 11.87
Lags 1 to 2 0.09 0.10 0.03 2.21 3.02 3.16 1.86 2.51 2.52 7.83 12.62 11.78
No 9/11 dummies (Lags 0 to 3) 0.31 1.10 2.04 2.20 3.52 4.50 1.90 2.92 3.83 9.37 13.09 14.79

the period from January 1992 to April 2003. The Table B.1
model without the 9/11 variables has decidedly poorer Model estimation summary statistics for LAPR99 shift variable
performance. Examination of the summary statistics Adj. R 2 AIC SBC
shows that lags 3, 4, and 5 appear to have the best NaRve model w/o LAPR99 0.92 3.79 3.98
performance. NaRve model w/ LAPR99 0.92 3.76 3.97
Table A.3 shows the out-of-sample dynamic MALT model w/o LAPR99 0.92 3.69 3.92
MALT model w/ LAPR99 0.93 3.61 3.87
forecast error summary statistics for the eight speci- RSTK model w/o LAPR99 0.93 3.66 3.92
fications for 1 month ahead (1M), 3 months ahead RSTK model w/ LAPR99 0.94 3.47 3.75
(3M), and 6 months ahead (6M). w/: with; w/o: without.

Table B.2
In-sample dynamic forecast statistics for LAPR99 comparisona
RMSE MAE MAPE Theil Bias Var CoVar
P P P
NaRve w/o LAPR99 5.38 4.48 22.06 0.12 0.00 0.60 0.40
NaRve w/ LAPR99 4.15 3.30 16.43 0.09 0.00 0.64 0.36
MALT w/o LAPR99 2.39 1.98 9.81 0.05 0.00 0.19 0.80
MALT w/ LAPR99 1.75 1.31 6.24 0.04 0.00 0.11 0.89
RSTK w/o LAPR99 3.94 3.40 16.09 0.09 0.00 0.32 0.67
RSTK w/ LAPR99 1.54 1.17 5.56 0.03 0.00 0.05 0.95
w/: with; w/o: without.
a
Fig. B.1. WTI spot price and Saudi crude oil production. See Footnote 15 for definitions.
500 M. Ye et al. / International Journal of Forecasting 21 (2005) 491–501

Table B.3 standard deviation of WTI price increased from

Out-of-sample dynamic forecast statistics for LAPR99 comparison $3.11 to $4.58 per barrel; and the coefficient of
RMSE MAE Range variation of Saudi production increased from 2.63% to
1M 3M 6M 1M 3M 6M 1M 3M 6M 6.74% and the coefficient of variation of WTI price
NaRve w/o 2.75 4.26 6.52 2.26 3.60 5.36 12.10 14.83 14.42 increased from 16.32% to 17.14%. Especially reveal-
LAPR99 ing is the correlation between Saudi production and
NaRve w/ 2.64 3.82 5.36 2.21 3.28 4.31 10.91 14.09 12.73 WTI price, which is 0.009 in the early period and
LAPR99
MALT w/o 2.31 3.30 3.77 1.89 2.76 3.23 9.76 13.71 12.65
0.717 in the later period.22 This difference is readily
LAPR99 visible in Fig. B.1, where no correlation is observed in
MALT w/ 2.25 3.07 3.20 1.85 2.49 2.68 8.63 12.66 11.53 the early 1990s but is obvious from the late 1990s to
LAPR99 the present.23
RSTK w/o 2.20 3.52 4.50 1.89 2.92 3.83 9.37 13.09 14.79 The impacts that the LAPR99 level shift has on the
LAPR99
RSTK w/ 2.13 2.91 3.06 1.75 2.42 2.42 7.59 12.13 11.52
models are shown in Tables B.1, B.2, and B.3, which
LAPR99 contain summary statistics for the three models with
w/: with; w/o: without. and without the level shift variable. In addition to
being statistically significant, the LAPR99 variable
adds explanatory power to all three models.
All tables in this appendix show that the specification
with 9/11 dummies and current and three lagged
OECD industrial relative inventories give the best References
results in all aspects.
Arize, A. C. (2000). U.S. petroleum consumption behavior and oil
price uncertainty: Tests of cointegration and parameter insta-
Appendix B. April 1999 level shift variable and bility. Atlantic Economic Journal, 28(4), 463 – 477.
Bentzen, J. (1997). The impact of petroleum taxes on residential oil
structural change demand in Denmark. Journal of Energy Finance and Develop-
ment, 2(1), 81 – 96.
In this appendix, we first show evidence for a Diebold, F. X. (2001). Elements of forecasting (2nd ed.). Cincinnati7
structural change in crude oil markets during April South-Western Thomson Learning.
1999. Then, we report the impacts of a level shift Friedman, M. (1968). The role of monetary policy. American
Economic Review, 58(1), 1 – 17.
variable, LAPR99, reflecting such structural change Gordon, D. F., & Hynes, A. (1970). On the theory of price dynamics.
on the forecast ability of the model. In E. S. Phelps, et al., (Eds.), Micro-economics of inflation and
OPEC became more active in the market in early employment theory (pp. 369 – 393). Norton7 New York.
1999, reining in supply to cause prices to rise Husted, S., & Kollintzas, T. (1987). Linear rational expectations
significantly. This change in behavior resulted in equilibrium laws of motion for selected U.S. raw material
imports. International Economics Review, 28(3), 651 – 670.
structural change in the relationship between invento- Jones, C. T. (1993). A single equation study of U.S. petroleum
ries and crude oil price in April 1999. The relationship consumption: The role of model specification. Southern
between Saudi production and WTI spot price is the Economic Journal, April, 687 – 700.
most telling statistical indication of an April 1999 Lucas Jr., R. (1972). Expectations and the neutrality of money.
regime change. As shown in Fig. B.1, the mean, Journal of Economic Theory, 4, 103 – 124.
Lucas Jr., R. (1973). Some international evidence on output-
standard deviation, and coefficient of variation of inflation tradeoffs. The American Economic Review, 63(3),
WTI and Saudi production all changed significantly 326 – 334.
from the first period (from December 1990 to March Phelps, E. S. (1970). Introductory chapter. In E. S. Phelps, et al.,
1999) to the second (April 1999 to April 2003): the (Eds.), Micro-economics of inflation and employment theory.
mean of Saudi production decreased from 8245 to Norton7 New York.
8058 thousand barrels per day, while mean WTI price 22
Detailed results are available upon request.
increased from $19.06 to $26.75 per barrel; the 23
The financial community recognizes this recent highly positive
standard deviation of Saudi production increased from correlation between WTI price and Saudi crude oil production. See,
216 to 544 thousand barrels per day while the for example, Rothman and Pfeifer (2003).
 M. Ye et al. / International Journal of Forecasting 21 (2005) 491–501 501

Pindyck, R. S., & Rubinfeld, D. L. (1991). Econometric models and John ZYREN joined the Energy Information Agency at the U.S.
economic forecasts (3rd ed.). Boston7 McGraw-Hill. Department of Energy as a senior industry economist after working
Rothman, M., & Pfeifer, S. (2003). They came, they met and they as a private consultant in the energy industry. His principal areas of
will come again. . . OPEC update: Energy commodities–oil, research are in the quantitative modeling of crude oil and petroleum
Merrill Lynch, 11 June 2003 (pp. 1 – 2). product price behavior.
Sargent, T. (1976). A classical macroeconometric model for the
United States. Journal of Political Economy, 84(2), 207 – 238. Joanne SHORE has spent over 25 years analyzing and modeling
Ye, M., Zyren, J., & Shore, J. (2002). Forecasting crude oil spot energy markets. She began her career in the nuclear industry, but
price using OECD petroleum inventory levels. International later joined Gulf Oil’s corporate planning department where she
Advances in Economic Research, 8(4), 324 – 334. performed studies in areas such as downstream profitability and the
Ye, M., Zyren, J., & Shore, J. (2003). Elasticity of demand for implications of economic, political, and social changes on the oil
relative petroleum inventory in the short run. Atlantic Economic business. Ms. Shore then spent 15 years in private consulting,
Journal, 31(1), 87 – 102. focusing on analyses of crude oil, gasoline, distillate, and propane
markets to explain price movements and to determine the impacts of
Biographies: Michael YE teaches economics at St. Mary’s College changing regulations. In 2000, she joined the Energy Information
of Maryland. He has experience as a consultant to a wide variety of Administration’s Petroleum Division as lead operations research
private and government organizations, including the U.S. Depart- analysis team leader.
ment of Defense and U.S. Department of Energy. He has published
many papers on modeling in microeconomics, operations research,
and statistics.