Professional Documents
Culture Documents
net/publication/276346462
CITATIONS READS
30 744
2 authors, including:
Robert Bray
Northwestern University
18 PUBLICATIONS 190 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Robert Bray on 20 July 2016.
This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.
The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.
INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
MANUFACTURING & SERVICE
OPERATIONS MANAGEMENT
Vol. 17, No. 2, Spring 2015, pp. 208–220
ISSN 1523-4614 (print) ISSN 1526-5498 (online) http://dx.doi.org/10.1287/msom.2014.0513
© 2015 INFORMS
Haim Mendelson
Graduate School of Business, Stanford University, Stanford, California 94305, haim@stanford.edu
T he bullwhip effect and production smoothing appear antithetical because their empirical tests oppose one
another: production variability exceeding sales variability for bullwhip, and vice versa for smoothing. But
this is a false dichotomy. We distinguish between the phenomena with a new production smoothing measure,
which estimates how much more variable production would be absent production volatility costs. We apply our
metric to an automotive manufacturing sample comprising 162 car models and find 75% smooth production by
at least 5%, despite the fact that 99% exhibit the bullwhip effect. Indeed, we estimate both a strong bullwhip
(on average, production is 220% as variable as sales) and robust smoothing (on average, production would
be 22% more variable without deliberate stabilization). We find firms smooth both production variability and
production uncertainty. We measure production smoothing with a structural econometric production scheduling
model, based on the generalized order-up-to policy.
Keywords: production smoothing; bullwhip effect; demand signal processing; generalized order-up-to policy;
martingale model of forecast evolution
History: Received: November 15, 2013; accepted: October 31, 2014. Published online in Articles in Advance
February 19, 2015.
it disregarded production stability. To conduct this Jackson 1994, Oh and Özer 2013). The MMFE de-
counterfactual analysis, we use structural estimation scribes the conditional expectations of a covariance-
(Lucas 1976, Reiss and Wolak 2007, Keane 2010); i.e., stationary, discrete stochastic process, 8xt 9
− . Let Ɛs 4xt 5
we (i) describe the data generating process with a pro- denote a forecaster’s conditional expectation of xt at
duction scheduling model, (ii) estimate the model’s time s ≤ t. The forecaster observes xt in period t, so
primitives with a sample of production schedules, Ɛt 4xt 5 = xt . And the forecaster has perfect memory, so
and (iii) use these primitive estimates to simulate its forecasts follow a martingale:
Downloaded from informs.org by [165.124.165.129] on 18 August 2015, at 18:13 . For personal use only, all rights reserved.
flexibility. H −1
X
= el0
t−l = 4xt 50
l=0
2. Model
Underpinning our empirical production scheduling 2.2. Production Logistics
model is Graves et al. (1998) and Chen and Lee’s An auto manufacturer faces exogenous demand for a
(2009) GOUTP, “an elegant model of 0 0 0 production single car model. In each period, the firm starts pro-
and inventory planning” (Aviv 2007, p. 778). DSP, ducing a new batch of cars, taking months to finish
the translation of demand forecast revisions into pro- each batch. The firm sources inputs quickly and back-
duction forecast revisions, drives the model dynamics logs demand when it stocks out of finished goods:
(Lee et al. 1997). the supply chain is decoupled (Kahn 1987, Gavirneni
et al. 1999, Lee et al. 2000, Chen and Lee 2009). The
2.1. Martingale Model of Forecast Evolution firm has an H-period-long forecast horizon, rolling
We model the firm’s information structure with forecasts for the next H demands. Its retail price is
the martingale model of forecast evolution (MMFE) fixed because “automakers only modestly respond
(Hausman 1969, Graves et al. 1986, Heath and with changes in price when faced with a demand
Bray and Mendelson: Production Smoothing and the Bullwhip Effect
210 Manufacturing & Service Operations Management 17(2), pp. 208–220, © 2015 INFORMS
shock to a particular vehicle. Instead, demand shocks that maps 4n − 15-period-information-lead-time cost
are almost entirely absorbed by changes in sales and signals into production quantities. Accordingly, we
production” (Copeland and Hall 2011, p. 233). call A the DSP IRF matrix and Ac the CSP IRF matrix.
Two exogenous processes drive the firm’s inventory Equation (5) provides market-clearing constraints:
dynamics: The 0 Ac = 0 constraint makes the CSP IRFs integrate
H −1
to zero, which makes aggregate workloads insensitive
to cost shocks, and the 0 A = 0 constraint makes the
X
dt = + el0
t−l and (1)
Downloaded from informs.org by [165.124.165.129] on 18 August 2015, at 18:13 . For personal use only, all rights reserved.
l=1 i=0
incurs baseline production cost Ct1 = ct pt , inventory
overage-underage cost Ct2 = it2 , production capac- Figure 1 depicts Proposition 1’s optimal DSP IRF
ity overage-underage cost Ct3 = pt2 , and produc- matrix under various parameter values. A plot’s nth
tion input overage-underage cost Ct4 = hl=1 l 4pt −
P
curve depicts the n − 1th column of A41 1 5, the
Ɛt−l 4pt 552 1 for h < H − 1. For example, suppose pro- IRF that determines how demand signals with n-
ducing a car requires (i) line capacity, which the firm period information lead times translate into produc-
builds well in advance, (ii) labor, which the firm tion quantities:
schedules one month ahead, and (iii) raw materials, • When = 0 and and are small, A mirrors an
which the firm orders two months ahead. In this identity matrix and production mirrors demand. In
case, parameterizes the line capacity cost, which is this case, the firm acts like a cross-dock facility with
convex in production pt , 1 parameterizes the labor
stable inventories.
cost, which is convex in one-month forecast error
• When is large, the firm disperses mass
pt − Ɛt−1 4pt 5, and 2 parameterizes the raw material
throughout A’s columns, which spreads demand
cost, which is convex in two-month forecast error
shocks across the production horizon. This is signal
pt − Ɛt−2 4pt 5.
pooling: rather than an individual demand signal, each
The firm’s expected operating cost is Ɛ4Ct1 + Ct2 +
production quantity responds to a weighted sum of
Ct + Ct4 5 = Ɛ4ct pt 5 + 4it 5 + 4pt 5 + hl=1 l 4pt −
3
P
all demand signals. Signal pooling attenuates produc-
Ɛt−l pt 50 The firm chooses the IRF matrices that mini-
tion variability.
mize this expected cost, subject to the market-clearing
constraints of (5): • When l is large, the firm diverts mass away
from A’s first l rows, which attenuates production
minc Ɛ4ct pt 5 + 4it 5 + 4pt 5 schedule changes made with less than l periods
A1 A
h notice. This is signal delaying: the firm postpones its
response to demand shocks to stabilize short-run pro-
X
+ l 4pt − Ɛt−l pt 51 (7)
l=1 duction schedules. Signal delaying attenuates produc-
s.t. 0 A = 0 and 0 Ac = 00 tion uncertainty.
The following section develops an algorithm to
Relative to the marginal cost of inventory variability, reverse engineer , , and from the measured
parameter ≥ 0 denotes the marginal cost of pro- degree of signal pooling and signal delaying.
duction variability and l ≥ 0 denotes the marginal
cost of “lead-l production uncertainty,” the mean
square error of the l-period-ahead production fore- 3. Identification and Estimation
cast, 4pt − Ɛt−l pt 5. In other words, parameter- We now develop our theoretical model into an empir-
p
izes the firm’s aversion to overall production volatil- ical model. We treat
t as a dependent variable,
t as
ity and l parameterizes its aversion to production an independent variable, and et = Ac
tc as statistical
volatility that resolves in the last l periods. Aviv (2007, error. Section 3.1 presents our basic identification con-
p. 780) calls hl=1 l 4pt − Ɛt−l pt 5 in expression (7) an
P
ditions and estimators, and §3.2 refines these results.
“adherence to production plans [metric], a measure
commonly used in the industry” to measure produc- 3.1. Basic Specification
tion uncertainty. We present three propositions: The first defines an
inverse mapping from optimal production policies to
2.4. Optimal Policy
model primitives, the second uses this inverse map-
The following proposition characterizes the optimal
ping to establish a set of identifying moment condi-
IRF matrices in terms of , , and = 61 1 0 0 0 1 h 7.
tions, and the third uses these identifying moment
Proposition 1. The firm sets A = A41 1 5 and conditions to create generalized method of moment
Ac = Ac 41 1 5, where (GMM) estimators of our model primitives.
−1 The following proposition establishes that A41 1 5
A41 1 5 = J + K K 0 4D0 C0 C D + I + L 5K is one-to-one—that each set of primitives has a unique
· K 0 D0 C0 C 4I − D J 5 − K 0 J 1 DSP signature.
Bray and Mendelson: Production Smoothing and the Bullwhip Effect
212 Manufacturing & Service Operations Management 17(2), pp. 208–220, © 2015 INFORMS
= 0.5 1 = 0.1
0.6
0.4
Downloaded from informs.org by [165.124.165.129] on 18 August 2015, at 18:13 . For personal use only, all rights reserved.
0.2
0.0
mnth element of DSP IRF matrix A(, , )
0.8
= 0.5 1 = 5
0.6
0.4
Column n
0.2 1
2
0.0
0.8 3
4
= 2 1 = 0.1
0.6 5
0.4
0.2
0.0
0.8
= 2 1 = 5
0.6
0.4
0.2
0.0
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
Row m
Proposition 2. When H > + 1, there exists in- tal variable inclusion restriction, ensuring the instru-
verse mapping A−1 4 · 5 that satisfies A−1 4A41 1 55 = ments have enough linearly independent variation
81 1 9. to characterize the demand signals (Cameron and
Trivedi 2005). And condition (9) is a classic instru-
We characterize A−1 in the appendix. The func-
mental variable exclusion restriction, ensuring the
tion enables us to empirically identify a firm’s prim-
instruments do not correlate with the error terms.
itives from its DSP IRFs. The identification argument With Proposition 3’s moment conditions, we define
is straightforward: production variability relative to estimators of , and in terms of matrices E =
inventory variability identifies , lead-l production p p
6
1 1 0 0 0 1
T 70 , E p = 6
1 1 0 0 0 1
T 70 , and Z = 61 1 0 0 0 1 T 70 :
uncertainty relative to inventory variability identi-
fies l , and the correlation between demand and Proposition 4. If (8) and (9) hold, then the following
lagged production identifies , since the firm tries to estimators are consistent:
match production starts in period t − with demand 81 ˆ 9
ˆ 1 ˆ = A−1 4Â51 where  = E p 0 ZZ 0 E4E 0 ZZ 0 E5−1 0
in period t.
Proposition 2 enables us to derive , and Note  is a GMM estimator corresponding to (9)’s
from A; Proposition 3, in turn, enables us to derive A moment conditions (Cameron and Trivedi 2005).
p
from
t and
t .
3.2. Refined Specification
Proposition 3. Matrix A is empirically identified in Proposition 4’s estimators have two drawbacks: (i) its
p
a sample of demand signals
t , production signals
t , and empirical requirements grow quickly with forecast
instrumental variables t , if horizon H , and (ii) it needlessly sacrifices H 2 degrees
of freedom by pre-estimating A. We now refine our
Rank6E4
t t0 57 = H and (8) identification conditions and estimators to address
E4et t0 5 = 00 (9) these shortcomings.
Define
t = IH−H
t as the first H elements of
t ,
p p p
Instrumental variables t enable us to determine the
t = IH −H
t as the first H elements of
t , A41 1 5 =
0
causal effect of independent variables
t on depen- IH −H A41 1 5IH−H as the top-left H × H submatrix
p p
dent variables
t . Condition (8) is a classic instrumen- of A, and et =
t −A41 1 5
t as a vector of statistical
Bray and Mendelson: Production Smoothing and the Bullwhip Effect
Manufacturing & Service Operations Management 17(2), pp. 208–220, © 2015 INFORMS 213
5.1. Cost Parameters Figure 2 depicts the medians (with vertical dashed
Table 2 presents ˆ and ˆ l , which respectively esti- lines) and probability density functions (PDFs) of ˆl =
ˆ + hi=l+1 ˆ i , our estimates of the marginal cost of
P
mate the production variability and lead-l production
uncertainty marginal costs, relative to the inventory the variance of production signals with l-period infor-
variability marginal cost. We find aversions to both mation lead times. First, the American firms are par-
production variability and uncertainty, with signifi- ticularly averse to production schedule revisions—
cantly positive mean and median ˆ and . ˆ Overall, the median American ˆl is more than five times the
95% of our sample exhibits some aversion to produc- median non-American ˆl , for l = 0, 1, 2, and 3+.
ˆ ˆ 1 , ˆ 2 , or ˆ 3 .
tion instability, with positive , This finding points to Detroit’s sluggishness. Second,
Bray and Mendelson: Production Smoothing and the Bullwhip Effect
Manufacturing & Service Operations Management 17(2), pp. 208–220, © 2015 INFORMS 215
Figure 2 Production Volatility Cost Estimate Medians and PDFs by Information Lead Time
American Non-American
l=0 l=1
0.6
0.4 0.4
Downloaded from informs.org by [165.124.165.129] on 18 August 2015, at 18:13 . For personal use only, all rights reserved.
0.2 0.2
Density
0.0 0.0
l=2 l = 3+
3
0.9
2
0.6
1
0.3
0.0 0
0 1 2 3 4 5 6 0 1 2 3 4 5 6
∧
γl
surprising production fluctuations cost more than have at least three car models in our sample, Toyota—
predictable ones: the mean ˆ1 is statistically larger the inventor of the Toyota Production System—has
than the mean ˆ2 , which is statistically larger than the shortest production lead time, at 0.11 months; Toy-
the mean ˆ3 . For example, the median American ota produces statistically faster than the average firm.
manufacturer deems last-minute production schedule
changes to be three times as costly as inventory fluctu-
ations, but deems three-month-out production sched- 5.3. DSP Matrices
ule changes to be only half as costly as inventory We estimate matrices A, è = Ɛ4
t
t 0 5, and èe = Ɛ4et et 0 5
e
fluctuations. with  = A41 ˆ 5,
ˆ 1 ˆ è̂ = T −1 E 0 E, and è̂ =
0 0
T −1 4E p − ÂE 0 54E p − ÂE 0 50 0 Figure 3 plots these esti-
5.2. Lead Times mates, normalizing 4è̂5 to one. The  estimates
Table 3 tabulates the discrete PDFs of our production demonstrate substantial production policy hetero-
lead-time estimates, ˆ (we cap ˆ to two months since geneity: our data take advantage of our model’s flex-
producing a car should take less than 60 days). Most ibility. The mostly positive ˆ give  their dispersion;
ˆ are zero months, which supports the Lieberman the mostly positive ˆ give  their asymmetry; and
et al. (1995, p. 9) finding that production lead times
the mostly zero ˆ give  their diagonal ridge. The è̂
only “sometimes exceed one month.” At 0.52 months,
estimates confirm that demand signals are positively
the average Asian lead time is statistically smaller
correlated and more volatile with shorter information
than the average non-Asian lead time, at 0.80 months. e
This finding highlights the Asian firms’ proclivity for lead times. The large è̂ estimates indicate that the
JIT manufacturing. Indeed, of the 13 companies that error terms are influential, which limits our R2 values:
on average, el0 Aˆ
t account for 38%, 40%, 41%, 36%,
p
Table 3 Lead-Time Estimate PDFs
and 36% of el0
ˆ t , for l = 0, 1, 2, 3, and 4.
ˆ = 0 ˆ = 1 ˆ = 2
Column n
0.75
Downloaded from informs.org by [165.124.165.129] on 18 August 2015, at 18:13 . For personal use only, all rights reserved.
!D
0.50
0.25
0.00
mnth element of matrix
0.5
0.4 11
0.3
Σ
D
12
0.2
0.1 13
0.0
1.2
0.8
Σ
0.4
C
0.0
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
Row m
and sales vary over time, a cost-minimizing strat- it were indifferent to production stability. We simu-
egy that equates marginal costs across time periods late the stability-indifferent counterfactual by setting
will smooth production relative to sales” (Blinder and and to zero, since these parameters compel the
Maccini 1991, p. 78). The bullwhip effect, on the other firm to stabilize production:
hand, is the tendency for demand shocks to amplify,
like the crack of a whip, as they wend their way up PS
c= ˆ =0 4p 5
ˆ =ˆ 4p 5/
=ˆ t =0 t
a supply chain (Lee et al. 1997). Although distinct,
e
these phenomena have shared a common measure— 6A41 ˆ 5
ˆ 1 ˆ è̂A41
ˆ 1ˆ 5
ˆ 0 + è̂ 7
the ratio of production variability to sales variability = e 0 (12)
ˆ è̂A401 01 5
6A401 01 5 ˆ 0 + è̂ 7
(Cachon et al. 2007, Bray and Mendelson 2012, Shan
et al. 2013). Mapping these two concepts to a sin-
Equation (12)’s smoothing estimate relies on two sim-
gle measure has forced them to be antithetical: either
plifications. First, it supposes the et statistical errors
a firm exhibited the bullwhip effect, with produc-
don’t change with or . Second, it pertains to p =
tion more volatile than sales, or production smooth- p t
ing, with sales more volatile than production. We pt − Ɛt−H 4pt 5 rather than pt , because
ˆ t doesn’t capture
have therefore implicitly defined production smooth- the production fluctuations firms can anticipate H = 5
ing to be the anti-bullwhip. Empirically, this framing months early. (Removing Ɛt−H 4pt 5 sacrifices about a
has enabled the bullwhip effect to steamroll produc- tenth of pt ’s variation.)
tion smoothing; e.g., production is more variable than We measure the bullwhip effect with the traditional
demand in 160 of the 162 car models in our sample. variance amplification ratio:
We reconcile the false dichotomy between produc- e
ˆ 5
ˆ è̂A41 ˆ 5
ˆ 0 + è̂ 7
tion smoothing and the bullwhip effect with a new BW ˆ t 5 = 6A41
ˆ 5/4d
d = 4p ˆ 1 ˆ 1
0
production smoothing measure. Rather than bench- t
6è̂7
mark production variability to sales variability—
an apples-to-oranges comparison, because of the Note, to match our smoothing metric, we measure
bullwhip—we benchmark it to the production vari- the bullwhip in terms of p = pt − Ɛt−H 4pt 5 and d t =
t
ability in the hypothetical scenario in which firms dt − Ɛt−H 4dt 5, rather than pt and dt (the results are sim-
have no incentive to smooth. That is, we measure ilar either way).
a firm’s smoothing with the ratio of what its pro- With our new smoothing measure, we no longer
duction variability actually is to what it would be if must pit the bullwhip effect against production
Bray and Mendelson: Production Smoothing and the Bullwhip Effect
Manufacturing & Service Operations Management 17(2), pp. 208–220, © 2015 INFORMS 217
Figure 4 Sales, Production, and Stability-Indifferent Counterfactual one (on average, production is 240% as variable as
Production of the Dodge Neon demand), and 95% of our production smoothing esti-
5.0 mates fall short of one (on average, production is
Production Demand only 81% as variable as it would be without explicit
Detrended volume
2.5
stabilization). The median American manufacturer
smooths significantly more than the median non-
American manufacturer and thus has a significantly
Downloaded from informs.org by [165.124.165.129] on 18 August 2015, at 18:13 . For personal use only, all rights reserved.
0.0
smaller bullwhip—intuitively, Americans are better
suited to smoothing, since being less lean gives them
– 2.5 more stabilizing inventories.
So far, we have only considered production
1998 2000 2002 2004 2006
variability, but firms also attenuate production
Year uncertainty. We measure the smoothing of lead-l pro-
duction uncertainty by comparing its values in the
5.0 Production Counterfactual production
current and stability-indifferent scenarios:
Detrended volume
2.5 PS
cl = ˆ =ˆ 4p − Ɛt−l 4p 55
ˆ =0 4p − Ɛt−l 4p 55/
=0 t t =ˆ t t
Pl−1 0 e
0.0 ˆ è̂A401 01 5
e 6A401 01 5 ˆ 0 + è̂ 7ei
= P i=0 i e 0
l−1 0 ˆ 5
ˆ è̂A41 ˆ 5
ˆ 0 + è̂ 7ei
i=0 ei 6A41
ˆ 1 ˆ 1
– 2.5
We likewise define the lead-l bullwhip effect as the
amplification of lead-l uncertainty, from demand to
1998 2000 2002 2004 2006
production (Bray and Mendelson 2012):
Year
American Non-American
1.0 • • • ••• • • • • •• •• • ••• •
• ••
• • ••• •
• •• •• • • • • • •
•
•
• ••• • •• • • ••
• • • • • •
••• • •
• •
••
• • • •
• •• • • • ••
•• • •
•• • • • •
• • • •
0.8 •• •• •
• • • •
•• •
PS
••
• • •
• • • • •
• • •• • ••
• • •• • •
• • • • • • ••
• • •
••
••
0.6 •• •
• •
•• •
••• •
•
•
• •
•
1 2 3 4 5 1 2 3 4 5
BW
Bray and Mendelson: Production Smoothing and the Bullwhip Effect
218 Manufacturing & Service Operations Management 17(2), pp. 208–220, © 2015 INFORMS
Table 4 Quartiles of the Lead-l Production Smoothing and processing. We show that the transformation of de-
Bullwhip Effect Estimates mand into production empirically identifies manufac-
l =1 l =2 l =3 l =4 l =5 turing costs. Our model-cum-estimators give empiri-
cists a flexible means to invoke inventory theory:
BW
c
l
Q1 1074 1063 1061 1065 1063 managers could use our framework to conduct oper-
400065 400065 400055 400065 400065 ational counterfactuals, such as estimating the value
Q2 2051 2018 2012 2018 2016 of shortening production lead times or improving
Downloaded from informs.org by [165.124.165.129] on 18 August 2015, at 18:13 . For personal use only, all rights reserved.
X
4wh+1 − wh 5 + 1 − wi = 00
i=0
⇒ K 0 D0 C0 C D KAè+K 0 D0 C0 C 4D J −I 5è
+K 0 KAè+K 0 J è+K 0 L KAè+K 0 L J è = 0 These h + 1 linear equations implicitly define 4A5
and 4A5.
−1
⇒ A = K 0 4D0 C0 C D +I +L 5K Proofs of Propositions 3–6. First, estimator  is consis-
· K 0 D0 C0 C 4I −D J 5−K 0 J −K 0 L J tent (note, condition (8) enables us to invert Ɛ4
t t0 5 Ɛ4t
t 0 5
and condition (9) enables us to drop Ɛ4et t0 5):
−1
⇒ A = K 0 4D0 C0 C D +I +L 5K
lim  = lim E p 0 ZZ 0 E4E 0 ZZ 0 E5−1
· K 0 D0 C0 C 4I −D J 5−K 0 J
T → T →
p −1
−1 = Ɛ4
t t0 5 Ɛ4t
t 0 5 Ɛ4
t t0 5 Ɛ4t
t 0 5
⇒ A = J +K K 0 4D0 C0 C D +I +L 5K
−1
= Ɛ 4A
t + et 5t0 Ɛ4t
t 0 5 Ɛ4
t t0 5 Ɛ4t
t 0 5
· K 0 D0 C0 C 4I −D J 5−K 0 J 0
−1
= A Ɛ4
t t0 5 Ɛ4t
t 0 5 Ɛ4
t t0 5 Ɛ4t
t 0 5
Differentiating the firm’s objective with respect to Ac like-
wise yields Ac . = A0
−1
Proof of Proposition 2. Let A 4A5 = 84A51 4A51 Since  converges to A, A−1 4Â5 must converge to 81 1 9,
4A590 First we derive 4A5 with three identities: which proves Propositions 3 and 4. Second, since Propo-
1. K 0 D0 C0 C D K + K 0 L K is symmetric and positive sition 2’s inverse mapping only references elements in A’s
definite. top-left H × H submatrix, function A41 1 5 is also one-to-
2. 4D0 C0 C 4I − D J 5 − L J 5e+1 = 4D0 C0 C 4I − D J 5 − one, which implies (11)’s moment conditions are consistent
L J 5e − e0 . at the true , , and values only:
3. 4D0 C0 C 4I − D J 5 − L J 5el = 4D0 C0 C 4I − D J 5 −
lim T −1 E p 0 − A41 ˆ 5E
ˆ 0 Z
L J 5el+1 , for l < . ˆ 1
T →
The first identity implies K4K 0 D0 C0 C D K + K 0 L K5−1 K 0
T
is symmetric and positive definite, which implies p ˆ 5
ˆ t t 0
= lim T −1
X
t − A41
ˆ 1
e00 K4K 0 D0 C0 C D K + K 0 L K5−1 K 0 e0 is nonzero. The second T →
t=1
identity yields
T
= lim T −1 ˆ 55
ˆ t + et t 0
X
4A41 1 5 − A41
ˆ 1
e00 Ae+1 e00 + K4K D0 C0 C D K + K 0 L K5−1
0
= J T →
t=1
· K 0 4D0 C0 C 4I − D J 5 − L J 5 e+1
ˆ 5
ˆ Ɛ4
t 0 5
= A41 1 5 − A41
ˆ 1 t
= e00 K4K 0 D0 C0 C D K + K 0 L K5−1 =0
0
·K 4D0 C0 C 4I − D J 5 − L J 5e+1 ˆ 9
ˆ = 81 1 9.
if and only if 81
ˆ 1
= e00 K4K 0 D0 C0 C D K + K 0 L K5−1 This proves Propositions 5 and 6.
Bureau of Labor Statistics (2013) Producer Price Index Industry Hausman WH (1969) Sequential decision problems: A model to
Data. Government Printing Office, Washington, DC. Accessed exploit existing forecasters. Management Sci. 16(2):93–111.
January 31, 2015, http://download.bls.gov/pub/time.series/ Heath DC, Jackson PL (1994) Modeling the evolution of demand
pc/pc.data.17.PrimaryMetal. forecasts ITH application to safety stock analysis in produc-
Cachon GP, Olivares M (2010) Drivers of finished-goods inventory tion/distribution systems. IIE Trans. 26(3):17–30.
in the U.S. automobile industry. Management Sci. 56(1):202–216. Holt CC, Modigliani F, Muth JF, Simon HA (1960) Planning Pro-
Cachon GP, Gallino S, Olivares M (2012a) Does inventory increase duction, Inventories, and Work Forces (Prentice-Hall, Englewood
sales? The billboard and scarcity effects in U.S. automo- Cliffs, NJ).
bile dealerships. Working paper, University of Pennsylvania,
Downloaded from informs.org by [165.124.165.129] on 18 August 2015, at 18:13 . For personal use only, all rights reserved.