Principles of Manufacturing

Cross Dock Model

Stephen C. Graves and Sean P. Willems

May 30, 2018

In this note we describe and develop an inventory model for a cross dock operation. This model is very
useful for understanding the benefits that can come from postponement or delayed differentiation in a
supply chain. We will use an example to explain what we mean by a cross dock operation, and how it
can be a representative model for the tactics of postponement and delayed differentiation.

Consider the following example: A computer company procures monitors from Asia for its three
distribution centers in the U.S. At the start of each month, each distribution center places an order on
the Asian supplier. The lead time to each distribution center is 6 weeks: 4 weeks (equal to one month) in
ocean transit from Asia to the West coast port at Long Beach California, and then two weeks in rail-
transit from Long Beach to each distribution center.

Weekly demand for monitors at the distribution centers is assumed to be normally distributed with the
following parameters:

Mean demand per week Standard deviation per week

DC 1 100 25

DC 2 80 25

DC 3 140 25

As noted above, the company uses a periodic review control policy with a review period r = 4 weeks; the
replenishment lead time is L = 6 weeks. Each DC orders based on a periodic review, base stock policy.

B=( r + L ) µ + zσ r + L
We assume that the safety factor is set to z = 2, so that the cycle service level (CSL) is about 98%. From
our previous development of this model, we have that:

r  µr
On hand inventory: E [ I ] =B − µ  + L  = + zσ r + L
2  2

1
Pipeline inventory: E [ P ] = µ L

We can then make an estimate of the current inventory for each DC:

Mean demand Standard Cycle stock Safety stock pipeline

per week deviation per
week

DC 1 100 25 200 158 600

DC 2 80 25 160 158 480

DC 3 140 25 280 158 840

Total on hand = 1114; cycle stock = 640; safety stock = 474; total pipe = 1920

The company is considering establishing a cross dock operation at Long Beach to reduce its inventory
requirements. We assume that the three DCs are on the same order cycle, so that they order at the
same time each month. As a consequence, the supplier can consolidate their orders, and thus is able to
manufacture and ship them together. This provides an opportunity to adjust the allocations to each DC
at some point in the replenishment process. This is what the cross dock operation allows.
With a cross-dock operation at Long Beach, the company would receive the consolidated shipment in
the port, and could then re-allocate the supply of orders that go to each DC. The consolidated shipment
arrives at Long Beach four weeks after the orders were placed. Based on the demand during this four-
week period, the company would have an update on the inventory status at each of the DCs, and
possibly a revised forecast for future demand; it could then use this information to re-allocate the
inventory so as to get the best overall corporate performance. For instance, if DC1 had more recent
demand than expected, while DC3 had less, then the allocation to DC1 could be increased by diverting
(or re-assigning) stock destined for DC3 to DC1.
If we have a cross dock operation, as described above, then we denote the common lead time to the
cross dock as L1 , and the lead time from the cross dock to each DC as L2 where L1 + L2 =
L. We can
estimate the inventory for each DC with the following formulas:
µi r
Cycle stock: .
2
 L1
Safety stock: ≈ zσ i + L2 + r . (1)
N
As explanation of(1), we have risk pooling over the first (or common) lead time, which reduces the
exposure of each DC to variability. A DC that experiences high demand over the common lead time can
be allocated more inventory from DCs with low demand. The more DC’s the more opportunity that
there will be to ‘risk pool’ over this portion of the lead time.

Pipeline stock: µi ( L1 + L2 ) for each DC.

For the example, we then get:

safety
cycle stock pipeline
DC 1 100 25 200 135 600
DC 2 80 25 160 135 480
DC 3 140 25 280 135 840

640 406 1920

Total is 2966.
We see that the savings is just in safety stock but this can be substantial.
This model given by (1) can be very helpful as it provides a simple way to assess the benefits from the
cross dock. Furthermore, this model is applicable to general cases in which the final destination (or
determination) of a product can be postponed.
For instance we might have a factory that manufactures N products, with a lead time of L. But the first
part of the production process is common across all of the N products; and the identity of each product
does not get determined until midway in the overall production process. We can apply the cross dock
model (1) to this example where L1 is the common lead time, and L2 is the remaining lead time after

the identity of each product is set, and L1 + L2 =

L.
Another example of postponement is for a product that can be packaged in several different ways,
depending on the customer or country. In these cases a DC might receive the product unpackaged, and
then only package it within the DC based on the most recent demand. In this case L1 is the time to

replenish the DC, while L2 is any additional time at the DC to package or localize a product before

shipping to the customer. In the Appendix we show how to derive the safety stock expression (1) for the
cross dock model.
Appendix

The reference for this model is

Eppen, Gary and Linus Schrage. "Centralized ordering policies in a multi-warehouse system with lead
times and random demand." Multi-Level Production/Inventory Control Systems (1981): 51-67.

In this section we provide a simplified development of the safety stock expression for the cross dock
model, given by the formula(1).
The assumptions are:
• We have N inventory sites (e.g., the DCs), each with normally distributed demand with
parameters µ j , σ j for site j.

• We assume each site has the same demand standard deviation; that is σ=
j σ , ∀j. The
development extends directly for non-equal standard deviations, but is more complicated.
• Each site reorders with a periodic review, base stock policy with a common review period r.
• = L1 + L2 . The first
The lead time for replenishment to each site is deterministic, given by L

lead time L1 is the time for the system replenishment order to reach the cross-dock; and the

second lead time L2 is the time from the cross dock to each site. We assume here that this

second lead time is the same for all sites, but the model can be extended to allow site-specific
L2 .
• We assume that the N sites place a single consolidated order on the supplier at each review
period. When the consolidated order arrives at the cross dock, we assume that we can re-
allocate the order so as to equalize the probability of stock-out at each site at the end of this
replenishment cycle. This is a critical assumption that we will explain more as we develop the
model.
• For ease of presentation we assume, without loss of generality, that r > L so that there is at
most one order in process at any point in time. All results apply when r ≤ L but the explanation
is more involved due to the need to account for the time in which multiple orders are
outstanding.
• Any shortages are backordered.
We will use D j ( a , b ) to denote the demand at site j over the time interval from time a to time b. We
N
denote the consolidated (or system) demand as DS ( a , b ) = ∑ D ( a, b ).
j =1
j

We will use I j ( t ) to denote the on-hand inventory at site j at time t. We denote the system inventory
N
as I S ( t ) = ∑ I ( t ).
j =1
j

Let us suppose that time t is a review epoch (i.e., the time at which an order is placed). Each site makes
an initial order given by:

Qˆ j ( t=
) B j − I j (t ) . (2)

That is, the initial order is just the amount that would bring the inventory position 1 for the site back to
its base stock B j .

Given these initial orders from each site, then the consolidated order at time t is:

∑(B − I j (t )).
N N
Q (t )
=
=j 1 =j 1
∑ Q=
ˆ (t )
j j (3)

This consolidated order arrives at the cross dock at time t + L1 ; at that time it is split into orders

destined for each site, which we denote as Q j ( t ) . We can now adjust the initial orders, having

observed the demand D j ( t , t + L1 ) over the interval ( t , t + L1 ) at each site. We need to make these
N N
adjustments so as to allocate fully the consolidated order; 2 that is, =
=j 1 =j 1
∑
Q j (t ) ∑
= Qˆ j ( t ) Q ( t ) .

The intent of the adjustment is to send more inventory to the sites that experienced heavier demand,
and less inventory to the sites with lighter demand. Eppen and Schrage (in references cited above)
propose the following adjustment 3:

1
Since we assume r > L, at any review epoch the inventory position (on-hand plus on-order) will equal the on-hand
inventory, as there will be no inventory on-order at that time instant.
2
We assume that the cross-dock does not hold any inventory; thus, everything that is received must then be
‘cross-docked’ and shipped out.
3
There’s an important assumption here, in that we assume that these quantities are positive. This is a reasonable
assumption as long as the sites have similar expected demand rates, and the demand over the common lead time
is not too variable.
 D ( t , t + L1 )
Q j ( t )= Qˆ j ( t ) + D j ( t , t + L1 ) − S . (4)
N

Thus, for each site we adjust the initial order by adding in the demand experienced at the site, minus the
average demand across all of the sites. Hence, the adjustment increases the order amounts that go to
the sites with heavier demand, namely those sites that experience demand more than the average; and
the adjustment reduces the order amounts that go to the sites with lower demand.

N N
We also see that =
=j 1 =j 1
∑
Q j (t ) ∑
= Qˆ ( t )
j Q ( t ) , so the adjusted order quantities exactly consume the

consolidated order.
We can now use (4) along with (2), to characterize the on-hand inventory after each order is received,

namely the inventory at time t + L1 + L2 =t + L+ at each site.

I j ( t + L+ ) = I j ( t ) + Q j ( t ) − D j ( t , t + L )
DS ( t , t + L1 )
= I j ( t ) + ( B j − I j ( t ) ) + D j ( t , t + L1 ) − − D j ( t, t + L ) (5)
N
DS ( t , t + L1 )
=B j − − D j ( t + L1 , t + L )
N
Thus, at time t + L+ the on-hand inventory is the base stock level minus the average demand over the
common lead time from t to t + L1 and minus the site-specific demand over the lead time from the cross

dock to the site, the time interval ( t + L1 , t + L ) .

To determine how to set the base stock levels, we will proceed similarly as we did for the base stock
model. We recall from the development of the base stock model that the order that is placed at time t
needs to cover demand until the next order is received; the next order is placed at time t + r and will
then arrive at time t + r + L . Hence, the order placed at time t needs to cover the demand until time
t + r + L , and hence needs to cover the demand over the coverage interval of length r + L.
To determine how to set the base stock for the cross-dock model, we can use (5) to project the

inventory just prior to the receipt of the next order, namely the inventory at time t + r + L− :
 I j ( t + r + L−=
) I j ( t + L+ ) − D j ( t + L, t + r + L )
D ( t , t + L1 ) (6)
= Bj − S − D j ( t + L1 , t + r + L ) .
N
We can now use (6) to set the base stock level and to determine the expected inventory levels from this
cross-dock policy. In general we want to set the base stock so that with high probability we will not
stock out; that is, we want:

I j ( t + r + L− ) ≥ 0
DS ( t , t + L1 ) (7)
⇒ Bj ≥ + D j ( t + L1 , t + r + L ) .
N
The right-hand-side of (7) is a random variable, which is normally distributed by assumption, and has
moments given by:

 D ( t , t + L1 )  µ
E S + D j ( t + L1 , t + r + L=)  S L1 + µ j × ( r + L2 )
 N  N
 D ( t , t + L1 )  σ S2
Var  S + D j ( t + L1 , t + r + L= ) L + σ 2j × ( r + L2 )
2 1
(8)
 N  N

where µS is the system demand rate, and σ S2 = N σ 2 is the system demand variance, and σ 2 is the

demand variance for each site, assumed to be the same for all sites. With this latter assumption, we can
re-write the variance as:

 D ( t , t + L1 )
Var  S
 N


(
+ D j ( t + L1 , t + r + L )  = σ 2 × 1 + r + L2 .
L
N ) (9)

Since the bracketed term in (8) is normally distributed with these moments, we suggest setting the base
stock as:

 D ( t , t + L1 )   D ( t , t + L1 ) 
=Bj E  S − D j ( t + L1 , t + r + L )  + z Var  S − D j ( t + L1 , t + r + L )  .
 N   N  (10)
µS
L1 + µ j × ( r + L2 ) + zσ
L1
= + r + L2
N N
where z is a safety factor.
This formula (10) for the base stock is similar to that for the original base-stock model, in that it equals
an expected demand over the coverage interval r + L plus a safety stock term that is proportional to
the standard deviation of demand over the coverage interval. What is different is that in (10) the
 demand over the common lead time (time until the cross dock) is the average demand across the sites,
rather than the site-specific demand; the average demand has less variability due to risk pooling, and as
such the cross dock model will result in less safety stock (shown next).
Now we can use (10) to get an estimate of the on hand inventory at each site. We just need to plug it
into (5) and take an expectation to get the expected “high point” for the inventory, and then do the
same using (6) to get the expectation for the “low point” of the inventory. We do this next:

 D ( t , t + L1 ) 
E  I j ( t + L+ ) =B j − E  S − D j ( t + L1 , t + L ) 
 N 
µS µS
L1 + µ j × ( r + L2 ) + zσ
L1
= + r + L2 − L1 − µ j L2 (11)
N N N
L1
= µ j r + zσ + r + L2
N
And

 D ( t , t + L1 ) 
E  I j ( t + r + L− ) = B j − E  S − D j ( t + L1 , t + r + L ) 
 N 
µS µS
L1 + µ j × ( r + L2 ) + zσ L1 − µ j × ( r + L2 )
L1
= + r + L2 − . (12)
N N N
L1
= zσ + r + L2
N
We can now average these two expressions to estimate the average inventory for the cross dock model:
µ jr L1
E  I j = + zσ + r + L2 . (13)
2 N
Thus, we have a familiar expression with the average inventory being a cycle stock that depends on the
length of the review period, plus a safety stock term. The safety stock term is the same as stated earlier
in (1), and reflects the benefit from risk pooling over the common lead time 4.

4
In the note the safety stock expression (1) has a site-dependent standard deviation, whereas the development in
the appendix assumes that all sites have the same standard deviation. When the sites can have different standard
deviations, then (1) is an approximation.