See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220636425

Determining the reorder point and order-up-to-level in a periodic review

system so as to achieve a desired fill rate and a desired average time between
replenishments

Article  in  Journal of the Operational Research Society · September 2009

DOI: 10.1057/palgrave.jors.2602655 · Source: DBLP

CITATIONS READS

26 20,305

3 authors:

Edward A. Silver Hussein Naseraldin

The University of Calgary ORT Braude College
186 PUBLICATIONS   9,984 CITATIONS    15 PUBLICATIONS   533 CITATIONS   

SEE PROFILE SEE PROFILE

Diane Bischak
The University of Calgary
42 PUBLICATIONS   786 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Behavioral Aspects in Operations View project

Inventory Management View project

All content following this page was uploaded by Diane Bischak on 02 June 2014.

The user has requested enhancement of the downloaded file.

Determining the Reorder Point and Order-Up-To-Level in a Periodic Review
System So As to Achieve a Desired Fill Rate and a Desired Average Time
Between Replenishments

Edward A. Silver
Haskayne School of Business
University of Calgary
2500 University Drive NW
Calgary, AB T2N 1N4
Canada

Hussein Naseraldin
Joseph L. Rotman School of Management
105 St. George Street
Toronto, ON M5S 3E6
Canada

Diane P. Bischak
Haskayne School of Business
University of Calgary
2500 University Drive NW
Calgary, AB T2N 1N4
Canada

October 2007
Abstract

In this paper we consider a periodic review, reorder point, order-up-to-level system, a

type commonly used in practice. Motivated by a specific practical context, we present a novel
approach to determining the reorder point and order-up-to-level (for a given review interval) so
as to target desired values of i) customer fill rate and ii) average time between consecutive
replenishments. Specifically, by using a diffusion model (producing normally distributed
demand) we convert a periodic review, constant lead time setting into one having continuous
review and a random lead time. The method is simple to implement and produces quite
reasonable results.

Keywords: Inventory control, Heuristics, Stochastic, Diffusion process, Supply chain

management

Suggested running head: Determining s and S in a periodic system

1. Introduction

This paper is concerned with an inventory control system commonly used in practice.

Specifically, the status of an item is examined at equi-spaced (review) intervals and, if the

inventory position (on-hand plus on-order minus backorders) is at or below the reorder point

(denoted by s), then a replenishment, that raises it to the order-up-to-level (denoted by S), is

initiated. The review interval (denoted by R) is often preset at a convenient value (e.g., day,

week), which is what will be assumed here.

The research leading to this paper was motivated by the practical context of a major

international producer and distributor of food products. In particular, it was deemed crucial to

determine s and S so as to approximately satisfy two practical constraints: i) a specified fill rate

(fraction of demand satisfied without backordering), i.e., a marketing requirement, and ii) a

specified average time between consecutive replenishments (e.g., two weeks), desirable from the

perspective of the supplier (production department).

It is inherently difficult to find proper values of the two control parameters, the reorder

point and order-up-to-level, primarily due to periodic review causing undershoots of the reorder

point before replenishments are triggered. This is illustrated in Figure 1, where replenishments

are placed at times 0 and 3R and a shortage occurs because the undershoot at time 3R plus the

demand during the lead time L exceeds the reorder point s. The probability distribution of the

undershoot is a complicated function of the distance S − s and the distribution of demand during

the review interval R. The complexity is not present in the simpler periodic review, order-up-to-

level system where a replenishment is initiated at each review instant (see, for example, Robb

and Silver, 1998, or Silver et al., 1998).

1
 The aforementioned practical context necessitated a relatively simple procedure for

determining appropriate values of s and S. An early software package, IBM’s IMPACT system

(IBM, 1971), provided a simple, but overly conservative, choice of s by assuming that the

inventory position is just above s at the review prior to the one at which a replenishment is

initiated, hence s must provide protection over an interval of length R + L . This approach had

been advocated even earlier (Brown, 1967).

In contrast, most of the literature presents rather complicated procedures. Moreover,

none of these explicitly deals with both of the above-mentioned constraints. Schneider (1978,

1981) and Tijms and Groenevelt (1984) used asymptotic results from renewal theory (Roberts,

1962) to approximate the undershoot distribution. For a different type of service measure

(fraction of demand being on backorder), Schneider and Ringuest (1990) developed power

approximations for s and S in the spirit of the original power approximation work of Erhardt

(1979). Bashyam and Fu (1998) advocated a simulation-based approach to minimize setup and

holding costs subject to meeting a prescribed fill rate. Other authors (e.g., Erhardt and Mosier,

1984, and Zheng and Federgruen, 1991) considered shortage costs rather than a service

constraint. More recently, Moors and Strijbosch (2002) developed an efficient descriptive

method for determining the fill rate for given values of s and S under the assumption of gamma

distributed demand. Unlike in the approach we propose, their method would have to be

combined with a search on s and S to achieve a desired value of the fill rate.

In the next section we provide an overview of the general approach we advocate for

determining appropriate values of s and S. Section 3 is then concerned with the details of the

model development and analysis. Simulation testing (in terms of how closely the two constraints

are met) is presented in Section 4. This is followed, in Section 5, by an adjustment procedure

2
whose use may be appropriate under certain conditions. Summary comments are provided in

Section 6 and some technical details are provided in the appendix.

2. Overview of the suggested approach

The key idea in the approach is to recognize that the reorder point is reached at a random

time between reviews. As shown in Figure 1, at the instant that the inventory position drops to

the reorder point there is a time, denoted by τ, remaining until the next review instant. (Under

periodic review, the actual value of τ would not be observed in practice.) Thus, we can think in

terms of a continuous review model with effective lead time τ + L , where τ is a random variable.

Under the assumption of stationary, independent, normally distributed demands in non-

overlapping time intervals, the behaviour of the inventory position can be modeled by a diffusion

process which, in turn, permits us to develop estimates of the first two moments of τ as a

function of the distance S − s and a measure of variability (coefficient of variation) of the

demand process. These moments are used as building blocks in choosing S and s so as to meet

the two constraints in the following manner.

First, we recognize that the expected time between consecutive replenishments in the (R,

s, S) system is the average size of a replenishment divided by the demand rate. But the average

size of a replenishment is S − s plus the average size of the undershoot (see Figure 1). The

latter, in turn, is easily developed from the expected value of τ. Thus we can select S − s so as

to target the desired expected time between replenishments.

Second, the moments of τ also lead to expressions for the mean and variance of total

demand over the effective lead time τ + L . Assuming the total demand is approximately

3
normally distributed, we can then determine a value of s appropriate to meet the fill rate

constraint.

3. Model development and analysis

In this section we first lay out the assumptions and a summary of the notation to be used.

Then we move into the details of the development and analysis of the model.

3.1 Assumptions and notation

The assumptions, most of which have already been mentioned, include the following:

i) The inventory position is reviewed every R units of time, where R is prespecified, not

controllable. (For convenience in the development we will set R=1, which means that the

review interval is redefined as unit time.)

ii) There is a constant replenishment lead time (L) from when a replenishment is triggered

(at a review instant) until it is available in stock.

iii) Demands in disjoint intervals of time are independent, stationary, normally distributed

variables.

iv) There is complete backordering of any demand during a stockout situation.

v) The service measure used is the fill rate, the fraction of demand to be routinely met from

stock.

vi) A target average time between consecutive replenishments is specified rather than

explicitly incorporating setup and carrying costs.

For reference purposes the following is a summary of most of the notation to be used (all

time variables are defined in units of R):

4
P2 – desired (fractional) value of the fill rate

n – desired (integer) average number of review intervals between consecutive

replenishments

s – reorder point

S – order-up-to-level

Q – size of a replenishment

L – replenishment lead time

µ – average demand in a unit time interval

σ – standard deviation of demand in a unit time interval

CV = σ/µ – coefficient of variation of demand in a unit time interval

τ – a random variable representing the time from when the inventory position hits s until

the next review instant

fτ (τ 0 ) – probability density function of τ

E(τ ) – the mean value of τ

Var(τ ) – the variance of τ

X – the total demand in τ + L

σ X – the standard deviation of X

k – safety factor

m = ( S − s) / µ

u – the first passage time for the inventory position to drop from S to s

fu (u0 ) – probability density function of u

5
3.2 E (τ ) and Var (τ ) for a given value of S − s

As indicated in Figure 2 (which reflects setting R = 1 ), τ is the time from the instant that

the inventory position first reaches s until the moment of the next review. The behaviour of

cumulative normally distributed demand (hence the change in the inventory position away from

the order-up-to-level, S) can be modeled as a continuous time diffusion process (Miltenburg and

Silver, 1984). Moreover, the probability density function of the first passage time, u, for the

inventory position to drop from S to s is given by (Cox and Miller, 1965)

S −s ⎡ ( S − s − µ u0 ) 2 ⎤
fu (u0 ) = exp ⎢ − ⎥ 0 < u0
σ 2π u03 ⎣ 2σ 2u0 ⎦

Setting

S −s
m= (1)
µ

and

σ
CV = (2)
µ

we obtain

m ⎡ ( m − u0 ) 2 ⎤
fu (u0 ) = exp ⎢ − 2 ⎥ 0 < u0 (3)
CV 2π u03 ⎣ 2(CV ) u0 ⎦

As an aside, Burgin (1969) developed an analytic expression for the expected value of u.

Now, multiple values of u can produce the same value of τ. Specifically τ = τ 0 will

result from any of u = 1 − τ 0 , 2 − τ 0 , etc. Hence, the density function of τ is given by

∞
fτ (τ 0 ) = ∑ fu (i − τ 0 )
i =1

Using (3) we have

6
 m ∞
1 ⎡ (m + τ 0 − i) 2 ⎤
fτ (τ 0 ) = ∑ exp ⎢ − ⎥ 0 <τ0 <1 (4)
CV 2π ⎣ 2(CV ) (i − τ 0 ) ⎦
2
i =1 (i − τ 0 )3

Unfortunately, it is not possible to analytically develop expressions for the moments of τ.

However, note that the density function, hence the moments, depend on only two parameters,

namely m and CV. For given values of these parameters τ0 can be discretized on a fine grid and

accurate estimates of E(τ ) and Var(τ ) can be found by numerical integration. An illustrative

case of the density function is shown in Figure 3. Initially we had hoped that τ would have a

uniform distribution, hence E(τ ) and Var(τ ) could be analytically determined. As illustrated in

Figure 3, this was not the case, particularly for combinations of low values of n and CV where

the first passage time distribution is not very spread out.

3.3 Determining S − s to target the desired average time between replenishments

For a given value of τ the expected size of the undershoot

E(undershoot | τ 0 ) = µτ 0

Hence E(undershoot) = µ E(τ ) . The average order size

E(Q) = S − s + E(undershoot) = S − s + µ E(τ )

The average time between replenishments

E(Q)
E(t ) =
µ

Using (1),

E(t ) = m + E(τ )

But we want to target an integer n for E(t). Thus we require

m + E(τ ) = n (5)

7
Now E(τ ) is a function of both m and CV. Thus, for a given value of CV we have to find the m

that satisfies (5). This was done using Mathematica where for each value of m, as mentioned

earlier, numerical integration is needed to estimate E(τ ) . In all of our experimentation the

resulting m is reasonably close to n- 0.5. The top part of Figure 4 shows E(τ ) as functions of

CV for n = 3 and 6. For very small values of CV the first passage time distribution approaches a

spike at m and for a specific n value equation (5) is satisfied by m = n − 0.5 and E(τ ) = 0.5 (with

Var(τ ) being very small). As CV increases, the u distribution, which is skewed to the right,

spreads out more, leading to E(τ ) increasing above 0.5. Eventually, as CV tends towards 0.5

(the largest value considered because larger values would be inappropriate for the assumption of

normally distributed demand), the u distribution flattens out so much that fτ (τ 0 ) tends towards a

uniform distribution, hence E(τ ) tends back to 0.5. Rather than providing detailed tables of

E(τ ) vs. CV, for several practical values of n we have instead used Stata (StataCorp. 2003) to fit

fractional polynomial functions of degree 3 (with very high R2 values). The results of fitting

functions to 51 values of E(τ ) , resulting from equi-spaced values of CV between 0.1 and 0.5, are

shown in Table 1. Note that the powers of CV have been adjusted to the average CV value (0.3)

in the approximation expressions; see the Stata Base Reference Manual (vol. 1, pp. 402-3) for

further information.

In the above, once the m is found that satisfies (5) for given values of n and CV, we then

use numerical integration to estimate Var(τ ) . Illustrative results are shown for two values of n

in the lower part of Figure 4. Again, very accurate fractional polynomial functions were fit

to Var(τ ) as a function of CV for selected values of n. The results are shown in Table 2.

8
3.4 Determining s to give a desired service level

As discussed in the previous subsection, for given values of n and CV we are able to

determine E(τ ) and Var(τ ) . Let X represent the total demand in the effective lead time τ + L .

As evident in Figure 1, a shortage will occur if X exceeds the reorder point s. We can model the

situation as a continuous review, reorder point system where the effective lead time τ + L is a

random variable with mean E(τ ) + L and variance Var(τ ) . Under such circumstances X has

moments (see Silver et al., 1998, p. 283)

E( X ) = [E(τ ) + L]µ (6)

and

Var( X ) = [E(τ ) + L]σ 2 + µ 2 Var(τ ) (7)

In the usual fashion, set

s = E( X ) + kσ X (8)

It is convenient to divide through by µ to obtain

s E( X ) σX
= +k
µ µ µ

Using (6) and (7), this reduces to

s
= E(τ ) + L + k [E(τ ) + L](CV ) 2 + Var(τ ) (9)
µ

If we assume that X is normally distributed (which we know is an approximation because τ is a

random variable; however, Tyworth and O’Neill, 1997, and Silver et al., 1998, pp. 272-3, argue

that as long as σ X / E( X ) ≤ 0.5 , there is little risk in making the normality assumption), then use

of (8) leads to the expected units short per replenishment cycle (see Silver et al., 1998)

EUSPRC  σ X Gu (k ) (10)

9
where
∞
Gu (k ) = ∫ ( z0 − k )φ ( z0 )dz0 (11)
k

is the unit normal loss function and φ ( z0 ) is the unit normal density function. There are tables,

spreadsheet functions (e.g., in Excel), and other accurate, rational approximations for the G

function. Note that there is a more precise version of (10) which turned out to not be required in

our numerical experiments, specifically

⎡ ⎛ nµ ⎞ ⎤
EUSPRC = σ X ⎢Gu (k ) − Gu ⎜ k + ⎟⎥ (12)
⎣ ⎝ σ X ⎠⎦

The target allowed (average) units short per replenishment cycle

AUSPRC = (1 − P2 )E(Q) (13)

From (10) and (13) and noting that E(Q) = nµ, we thus require

(1 − P2 )nµ
Gu (k ) = (14)
σX

Also, using (7), we can express

σX
= [E(τ ) + L](CV ) 2 + Var(τ ) (15)
µ

Hence, the choice of k in (14) depends only upon the values of four parameters, namely CV, n, L

and P2.

Finally, to obtain S we proceed as follows. Use of (1) and (5) gives

S −s
+ E(τ ) = n
µ

Hence

S s
= + n − E(τ ) (16)
µ µ

10
It is seen from (9), (15) and (16) that the key results are normalized with respect to µ, the average

demand per unit time.

Numerical illustration

Consider the following set of parameters values:

n = 4 (i.e., on average, a replenishment every 4 periods)

CV = 0.3

L = 2 periods

P2 = 0.9 (desired fill rate)

From the fractional polynomial approximations in Tables 1 and 2, E(τ ) = 0.5033 and Var(τ ) =

0.0839. Then from (15)

σX
= (2.5033(0.3) 2 + 0.0839 = 0.5561
µ

Next, (14) leads to

Gu (k ) = (0.1)4 / 0.5561 = 0.7194

Using a table lookup or Excel function (see Appendices B and C of Silver et al., 1998)

k = − 0.5309

Subsequently (9) gives

s
= 2.5033 + (− 0.5309)(0.5561) = 2.208
µ

Finally, (15) results in

S
= 5.705
µ

So, for example, if µ = 100, then s = 220.8 and S = 570.5.

11
 Our approach has involved approximations including the assumption of a normal

distribution of X. Thus, it is important to simulate the actual performance associated with using

the prescribed values of s and S. This is the topic of the next section.

4. Simulation testing of the approach

The simulation replicates what would happen in actual application of (R, s, S) control.

Specifically, the inventory position is only updated at each review instant (recall that we have set

R = 1 ). For the update

( inventory position at a review instant ) =

( inventory position after a possible order at previous review ) – (17)
( total demand in the review interval ) .

The total demand is randomly generated from a normal distribution with parameters µ and σ.

Note that in the simulation the times at which the inventory position hits s (hence the τ’s) are not

observed.

If the updated inventory position (denoted by I) is less than or equal to s, then an order

(of size Q = S – I) is placed. Otherwise nothing is done until at least the next review. Any order

placed arrives a constant time L later. Rather than keeping track of any backorders just before

replenishments arrive, we have used a far more efficient method of estimating the actual fill rate

achieved. It uses the “virtual measures” approach of variance reduction first suggested by Ignall

and Carter (1975). Specifically, let I i (≤ s ) be the inventory position when the ith order (of size

Qi) is placed. Then (for reference purposes see (12)) the expected units short at the end of the

lead time conditional on I i can be computed analytically as

12
 ⎡ ⎛ I − Lµ ⎞ ⎛ I i + Qi − Lµ ⎞ ⎤
EUS( I i ) = Lσ ⎢Gu ⎜ i ⎟ − Gu ⎜ ⎟⎥
⎣ ⎝ Lσ ⎠ ⎝ Lσ ⎠⎦

Over a large number (N) of replenishments the estimated average number of units short per

replenishment is

N
1
EUS =
N
∑ EUS( I )
i =1
i

Also, I i + Qi = S . Thus

EUS 1 N
⎛ I i − Lµ ⎞ ⎛ S − Lµ ⎞
=
Lσ N
∑G
i =1
u ⎜
⎝ Lσ ⎠
⎟ − Gu ⎜
⎝ Lσ ⎠
⎟

where the last term on the right side is a constant. Consequently, we only have to compute

⎛ I − Lµ ⎞
Gu ⎜ i ⎟ for each replenishment i.
⎝ Lσ ⎠

Now an estimate of the average order size is given by

N
1
E(Q) =
N
∑Q
i =1
i

and the expected fill rate is

EUS
FR = 1 −
E(Q)

The other performance measure we need is the average time between replenishments

µ
AT =
E(Q)

Besides point estimates of each of FR and AT, confidence intervals were developed based upon

the replication-deletion method for estimating a steady-state mean (see, e.g., Law and Kelton,

2000).

13
4.1 Selection of parameter values

As discussed earlier, after normalizing with respect to µ, there are four remaining,

independent, parameters, namely CV, n, L and P2. As shown in Table 3, each of these was set at

three reasonable values to produce a total of 81 experiments. A few associated comments are in

order.

As mentioned earlier, in the practical context that motivated this work, integer values of n

were desired. The lowest value of interest is 2 (in that n = 1 would imply ( R, S ) control, which

is much easier to analyze as the protection period is always exactly R + L). Because demand is

normally distributed we used an upper value of 0.5 for the CV. For higher values of CV a

different distribution (e.g., gamma) would likely be more appropriate. The L values were

selected so that the smallest was less than the review interval and the largest appreciably larger

than the interval. Finally, the three fill rates span the range typically targeted in practice.

4.2 Results

The results are displayed in Table 4. Although the details are not shown, all half-widths

of 95% confidence intervals were less than 0.007 (0.12%) for the estimates of the average time

between orders and less than 0.0004 (0.05%) for the estimates of the fill rates.

The major finding, indicated by the low percent deviations from the two targets (n and

P2), is that the approach works very well overall. Deviations from n as high as 5+% occur in the

average time (AT) between replenishments. However, such deviations are relatively unimportant

in practice because, even if the average time between replenishments was right at the targeted

value, such as 2 periods, there would still be occasions where the observed times would differ,

14
e.g., be 1 or 3 periods, simply due to the random nature of the demand, i.e., one could never

assure that the time between consecutive replenishments was a constant.

All of the entries in the last column of the table are negative, indicating that the fill rate

achieved is always somewhat below the target. The largest deviation of 3.0334% occurs in Case

3 in which n = 2 and CV = 0.5. (Case 3 also produces one of the highest deviations of AT from

n.) The combination of n = 2 and CV = 0.5 tends to produce a significant number of instances of

small τ values (in that the distribution of first passage time from S to s has a significant mass

below, but close to, 2). This occurs for all cases where CV = 0.5, but to a lesser degree as n, L

and P2 increase.

There are two possible causes of the inaccuracy. The more obvious reason is the

inaccuracy in assuming that X (the total demand in τ + L ) has a normal distribution. By

randomly generating values of τ (through values of the first passage time u – see equation (3))

and then for each such τ randomly generating a value of X from a normal distribution with mean

µ (τ + L) and variance σ 2 (τ + L) we developed empirical distributions of X for different

combinations of parameters, comparing each with a normal distribution having the same

moments. As expected, the empirical distribution became closer to the normal as the fixed

component L increased. This helps explain why the fill rate deviation decreases as L increases.

The second cause of the inaccuracy of our method is more subtle. If the reorder point (s)

is first reached at a distance τ from the next review instant, we have used an associated effective

lead time of length τ + L . However, recalling the diffusion representation of the demand process,

it is conceivable that by the time of the next review the inventory position could have come back

up above s due to negative overall demand in the interval τ. As will be shown in the next

section, this is far more likely to happen for low values of τ. In such a case the position will very

15
quickly drop back to the reorder point, resulting in a relatively large value of τ (occurring in the

next R interval). Thus, low values of τ will tend to be replaced by large values. Consequently,

our approach underestimates both E(τ ) and Var(τ ) , resulting in a fill rate somewhat lower than

the target as well as an average order size (which has a component proportional to E(τ ) ) larger

than targeted, hence leading to an average time between replenishments higher than n.

As mentioned earlier, it is precisely when CV takes on its highest value of 0.5 that there is

a significant probability of small τ values occurring. In the next section we propose a modified

approach for adjusting the estimates of E(τ ) and Var(τ ) under such circumstances.

5. Adjusting the estimates of E(τ ) and Var(τ )

As discussed in the previous section, our method, based on the diffusion model

representation of the normal distribution (which permits negative demands, hence the inventory

position rising), ignores the fact that when the reorder point is reached between review instants

there is a non-zero probability that the inventory position will be back above the reorder point at

the next review instant. The chance of this happening for a given value of τ is simply the

probability that total demand in τ is negative, i.e., Φ (− µτ / σ τ ) or Φ (− τ / CV ) , which

quickly becomes negligible as τ increases away from 0 ( Φ (⋅) is the cumulative distribution

function of a unit normal variable). Thus, in order for the effect to be important we need

significant probability mass associated with low values of τ in its unmodified (original)

distribution. Knowing that E(τ )  0.5 , this implies that we require considerable spread in that

distribution.

To obtain tractable adjustments in E(τ ) and Var(τ ) for the above effect we make two

simplifying assumptions, namely

16
i) the unmodified (original) distribution of τ is uniform, and

ii) when the inventory position manages to go back above the reorder point by the next

review instant, it quickly falls back to the reorder point, resulting in a τ value assumed to

be 1.

In actual fact the inventory position at the next review will be at some distance above s, which

won’t necessarily lead to a τ value of 1; in fact, an exact modeling approach, which would

become intractable, would recognize that for a given distance above s, we’re back to a diffusion

process starting at a specific distance above the reorder point, and so on. Regarding the first

assumption, the original distribution of τ tends to uniformity as both n and CV increase (through

the first passage time distribution becoming very spread out).

Using the two assumptions, in the appendix we develop the following expression for

AE(τ ) and AVar(τ ) , the adjusted values of E(τ ) and Var(τ ) :

1 C2 3 4 ⎛ 1 3 ⎞ ⎛ 1 ⎞ ⎛ C 3C 3 ⎞ ⎛ 1 ⎞
AE(τ ) = + − C + ⎜ − C2 + C4 ⎟Φ ⎜ − ⎟ − ⎜ − ⎟φ ⎜ − ⎟ (18)
2 2 4 ⎝2 2 ⎠ ⎝ C⎠ ⎝ 2 2 ⎠ ⎝ C⎠

AVar(τ ) = AE(τ 2 ) − [ AE(τ )]

2
(19)

where

1 C 2 5C 6 ⎛ 2 ⎞ ⎛ 1 ⎞ ⎛ 2C 5C ⎞ ⎛ 1⎞
3
AE(τ 2 ) = + − + ⎜ − C 2 + 5C 6 ⎟ Φ ⎜ − ⎟ + ⎜ − + + 5C 5 ⎟ φ ⎜ − ⎟ (20)
3 2 2 ⎝3 ⎠ ⎝ C⎠ ⎝ 3 3 ⎠ ⎝ C⎠

and, as earlier, Φ and φ are the cumulative and density functions of a unit normal variable and,

for simplicity, we have used C instead of CV to represent the coefficient of variation.

17
Numerical illustration

We illustrate the above approach for the case where the unadjusted approach had the

largest percent deviation from the targeted fill rate, specifically case 3 in Table 4 with n = 2, CV

= 0.5, P2 = 0.8 and L = 0.5.

Using (18), (20) and (19) we find

AE(τ ) = 0.582571

and

AVar(τ ) = 0.092826

(The unadjusted E(τ ) and Var(τ ) values were 0.502672 and 0.086764.) The approach of Section

3.4, using AE(τ ) and AVar(τ ) rather than E(τ ) and Var(τ ) , then leads to

s S
= 0.8114 and = 2.2289
µ µ

In contrast, the unadjusted approach gave

s S
= 0.7209 and = 2.2183
µ µ

As expected, the increases in AE(τ ) and AVar(τ ) , relative to E(τ ) and Var(τ ) , have led to an

appreciably higher value of s. This, in turn, results in an actual fill rate (obtained through

simulation) of 0.8002, i.e., much closer to (deviation of only 0.250% from) the target value of

0.80 than the actual fill rate of 0.7757 achieved by the unadjusted method (see case 3 in Table 4).

6. Summary

In this paper we have considered an (R, s, S) inventory control system, a type commonly

used in practice. Motivated by a practical context, we have presented a novel approach to

determining s and S (for a given review interval R) so as to target desired values of i) customer

18
fill rate and ii) average time between consecutive replenishments. Specifically, by using a

diffusion model we converted a periodic review, constant lead time situation into one having

continuous review and a random lead time. The basic method is simple to implement and

produces quite reasonable results. If a practitioner felt that even nominal underachievement of

the desired fill rate was a serious concern, then the adjustment of the preceding section, again

easy to implement, could be used.

There are a number of possible extensions of this work, including:

i) Dealing with a variable lead time. Except for the possibility of crossing of orders (which

cannot happen in the context of the current paper because τ is, by definition, less than R),

this should be a straightforward extension.

ii) Using a demand distribution other than the normal, particularly to encompass situations

where the CV exceeds 0.5. One possibility would be to modify the normal so that it

cannot take on negative values as suggested by Strijbosch and Moors (2006). Another

would be to use the gamma distribution. A third option would be to consider a discrete

distribution, such as the Compound Poisson, so as to be able to model large, individual

demand transactions.

iii) Incorporating shortage costing rather than a service measure.

iv) Treating a different control system, specifically an (R, s, mQ) system where orders must

be placed in an integer multiple m of a prescribed, basic replenishment quantity. This

would be appealing where the supply comes from production, as in the practical situation

that motivated this research. In particular, the production department may prefer to

produce only in integer multiples of a convenient lot size (Q). An equivalent context

19
 would be where supply is only delivered as an integer number of non-unit-sized (Q)

packs.

Acknowledgements

The authors thank Daniel Costa, Supply Chain Director, Nestlé Italiana, for describing a

practical situation that motivated the topic studied. The research leading to this paper was

supported by the Natural Sciences and Engineering Research Council of Canada under grants

A1485 and 239153-05. It was carried out while the second author held a postdoctoral position at

the University of Calgary. Associated financial support from the Haskayne School of Business

is gratefully acknowledged.

References

Bashyam S and Fu MC (1998). Optimization of (s, S) inventory systems with random lead times
and a service level constraint. Mngt Sci 44: 243-256.

Brown, RG (1967). Decision Rules for Inventory Management. Holt, Rinehart and Winston:
New York, p. 238.

Burgin TA (1969). Time to sell a fixed quantity of stock depleted by distributed demand.
Operational Research Quarterly 20: 421-428.

Carter G and Ignall E (1975). Virtual measures: a variance reduction technique. Mngt Sci 21:
607-616.

Cox DR and Miller HD (1965). The Theory of Stochastic Processes. John Wiley & Sons: New
York.

Ehrhardt R (1979). The power approximation for computing (s, S) policies. Mngt Sci 25: 777-
786.

Ehrhardt R and Mosier C (1984). Revision of the power approximation for computing (s, S)
policies. Mngt Sci 30: 618-622.

IBM (1971) System/360 Inventory Control (360 A-MF-04X). Program description manual
GH20-0555-1. New York.

20
Law AM and Kelton WD (2000). Simulation Modeling and Analysis. Third Edn. McGraw-Hill:
Boston.

Miltenburg GJ and Silver EA (1984). The diffusion process and residual stock in periodic
review, coordinated control of families of items. Int J Prod Res 22: 629-646.

Moors JJA and Strijbosch LWG (2002). Exact fill rates for (R, s, S) inventory control with
gamma distributed demand. J Opl Res Soc 53: 1268-1274.

Patel JK and Read CB (1982). Handbook of the Normal Distribution. Marcel Dekker: New
York, p. 23.

Robb DJ and Silver EA (1998). Inventory management with periodic ordering and minimum
order quantities. J Opl Res Soc 49: 1085-1094.

Roberts DM (1962). Approximations to Optimal Policies in a Dynamic Inventory Model. In:

Arrow K, Karlin S and Suppes P (eds.), Studies in Applied Probability and Management Science.
Stanford University Press: Stanford, pp. 207-229.

Schneider H (1978). Methods for determining the reorder point of an (s, S) ordering policy when
a service level is specified. J Opl Res Soc 29: 1181-1193.

Schneider H (1981). Effect of service-levels on order-points or order-levels in inventory models.

Int J Prod Res 19: 615-531.

Schneider H and Ringuest JL (1990). Power approximation for computing (s, S) policies using
service level. Mngt Sci 36: 822-834.

Silver EA, Pyke D and Peterson R (1998). Inventory Management and Production Planning and
Scheduling. Third Edn. John Wiley & Sons: New York.

Stata Corp. (2003). Stata Statistical Software: Release 8.0. Stata Corporation: College Station,
TX.

Strijbosch LWG and Moors JJA (2006). Modified normal distributions in (R, S)-inventory
control. Eur J Opl Res 172: 201-212.

Tijms HC and Groenevelt H (1984). Simple approximations for the reorder point in periodic and
continuous review (s, S) inventory systems with service level constraints. Eur J Oper Res 17:
175-190.

Tyworth JE and O’Neill L (1997). Robustness of the normal approximation of lead-time

demand in a distribution setting. Nav Res Logist 44: 165-186.

Zheng YS and Federgruen A (1991). Finding optimal (s, S) policies is about as simple as
evaluating a single policy. Oper Res 39: 654-665.

21
Appendix – Derivation of Adjusted Mean and Variance of τ

To simplify the presentation we use C, rather than CV, to represent the coefficient of

variation of demand in a unit time interval. As indicated in the main text the adjustment is based

on two assumptions:

i) the original distribution of τ is uniform

ii) when the inventory position manages to go back above the reorder point by the

next review, it quickly falls back to the reorder point and we assume that the

associated value of τ is effectively 1.

For a given value of the original τ, denoted by τ 0 , two things can happen. With

( )
probability Φ − τ 0 / C the inventory position gets back above the reorder point by the next

review resulting (by assumption ii) in an adjusted τ value of 1. With probability

( )
1 − Φ − τ 0 / C the original τ 0 value remains in effect. Thus, using the uniform distribution of

τ 0 (assumption i) the adjusted first two moments, AE(τ ) and AE(τ 2 ) , are

1 1

( ) ⎣ ( ⎦ )
AE(τ ) = ∫ Φ − τ 0 / C dτ 0 + ∫ τ 0 ⎡1 − Φ − τ 0 / C ⎤ dτ 0
0 0

and

1 1

( ) ⎣ ( )
AE(τ 2 ) = ∫ Φ − τ 0 / C dτ 0 + ∫ τ 0 2 ⎡1 − Φ − τ 0 / C ⎤ dτ 0
⎦
0 0

These can be simplified to

1
1
( )
AE(τ ) = + ∫ (1 − τ 0 ) Φ − τ 0 / C dτ 0
2 0

and

22
 1
1
( ) (
AE(τ 2 ) = + ∫ 1 − τ 0 2 Φ − τ 0 / C dτ 0
3 0
)
Substituting z0 = − τ 0 / C results in

0
1
(
AE (τ ) = − 2C 2 ∫ z0 − C 2 z03 Φ ( z0 )dz0
2
) (21)
−1/ C

and

0
1
AE τ( )2

3
(
= − 2C 2 ∫ z0 − C 4 z05 Φ ( z0 )dz0 ) (22)
−1/ C

In these two equations there are three integrations involving the Φ function, namely

0
J1 =
−1/C
∫ z0 Φ ( z0 ) dz0 (23)

0
J2 = ∫
−1/C
z03 Φ ( z0 ) dz0 (24)

and

0
J3 = ∫
−1/C
z05 Φ ( z0 ) dz0 (25)

Each of these can be integrated by parts to produce

0
1 ⎛ 1⎞ z02
2C 2 ⎝ C ⎠ −1/∫C 2
J1 = − Φ ⎜ − ⎟ − φ ( z0 ) dz0 (26)

0
1 ⎛ 1⎞ z04
4C 4 ⎝ C ⎠ −1/∫C 4
J2 = − Φ ⎜ − ⎟ − φ ( z0 ) dz0 (27)

and

0
1 ⎛ 1⎞ z06
6C 6 ⎝ C ⎠ −1/∫C 6
J3 = − Φ ⎜ − ⎟ − φ ( z0 ) dz0 (28)

23
We next make use of the following result for incomplete moments of the unit normal distribution

∞
from Patel and Read (1982). If I r (b) = ∫ z0 rφ ( z0 ) dz0 , then for r a positive, even integer
b

I r (b) = ⎡⎣b r −1 + (r − 1)b r −3 + " + (r − 1)(r − 3)" 5 ⋅ 3 ⋅ b ⎤⎦ φ (b)

(29)
+(r − 1)(r − 3)" 5 ⋅ 3 ⋅ [1 − Φ (b) ]

Recognizing that

0 ∞ ∞

∫ z0 φ ( z0 ) dz0 = ∫ z0 φ ( z0 ) dz0 − ∫ z0 j φ ( z0 ) dz0

j j

−1/C −1/C 0

we use (29) in (26), (27) and (28) with r = 2, 4 and 6 respectively to obtain, after considerable

simplification,

1 ⎛1 1 ⎞ ⎛ 1⎞ 1 ⎛ 1⎞
J1 = − + ⎜ − 2 ⎟ Φ ⎜ − ⎟ + φ ⎜− ⎟ (30)
4 ⎝ 2 2C ⎠ ⎝ C ⎠ 2C ⎝ C ⎠

3 ⎛3 1 ⎞ ⎛ 1⎞ ⎛ 3 1 ⎞ ⎛ 1⎞
J2 = − + ⎜ − 4 ⎟ Φ ⎜ − ⎟ + ⎜ + 3 ⎟φ ⎜ − ⎟ (31)
8 ⎝ 4 4C ⎠ ⎝ C ⎠ ⎝ 4C 4C ⎠ ⎝ C ⎠

and

5 ⎛5 1 ⎞ ⎛ 1⎞ ⎛ 5 5 1 ⎞ ⎛ 1⎞
J3 = − + ⎜ − 6 ⎟ Φ ⎜ − ⎟ + ⎜ + 3 + 5 ⎟φ ⎜ − ⎟ (32)
4 ⎝ 2 6C ⎠ ⎝ C ⎠ ⎝ 2C 6C 6C ⎠ ⎝ C ⎠

Finally, use of (21) to (25) and (30) to (32) leads directly to the expressions of (18) and (20) of

the main text.

24
Figure 1 – Undershoots in a Periodic Review System

25
Figure 2 – Relation Between First Passage Time (u) and τ

26
Figure 3 – Illustrative Density Function of τ for m = 1.5, CV = 0.3

27
 Figure 4 – E(τ ) that Satisfies Equation (5) and Associated Var(τ ) for n = 3, 6

E(τ )

Var(τ )

28
 Table 1 – Fractional Polynomial Approximations of E(τ ) as a Function of CV for Selected
Values of n

n Adj. R2 Approximation (note C ª CV)

2 0.9863 .53608 + .44271(C – 3.333) + 1.7634(C –½ – 1.826) + 1.0508(C –½ ln C + 2.198)
–1

3 0.9835 .51211 + 1.8652(C – .3) – 1.1430(C2 – .09) + 3.1367(C2 ln C + .1084)

4 0.9740 .50325 + 2.1455(C½ – .5477) – .65943(C½ ln C +.6594) + .50973(C½ (ln C)2 – .794)
5 0.9726 .50079 – .13438(C –1 – 3.333) – .39946(C –½ – 1.826) – .28188(C –½ ln C + 2.198)
6 0.9784 .50004 – .00237(C –2 – 11.11) – .03307(C –1 – 3.333) – .02296 (C –1 ln C + 4.013)

29
 Table 2 – Fractional Polynomial Approximations of Var(τ ) as a Function of CV for Selected
Values of n

n Adj. R2 Approximation (note C ª CV)

2 0.9993 .07401 + .53380(C – .3) – .58217(C2 – .09)
3 0.9965 .08283 + .36794(C½ – .5477) – .35809(C½ ln C + .6594)
4 0.9933 .08387 – .12888(C –½ – 1.826) – .10939(ln C + 1.204)
5 0.9931 .08371 + .01876(C –1 – 3.333) + .00887(C –1 ln C + 4.013)
6 0.9936 .08352 – .00078(C –2 – 11.11) + .00503(C –1 – 3.333)

30
 Table 3 – Settings of Independent Parameters

Parameter Definition Values Selected

target average number of review (unit) intervals

n 2, 4, 6
between replenishments

coefficient of variation of (the normally distributed)

CV 0.1, 0.3, 0.5
demand in a unit interval

L lead time 0.5, 2, 4

P2 target fill rate 0.8, 0.9, 0.99

31
 Table 4 – Simulation Results

Average Time (AT) Between Orders Fill Rate (FR)

% Deviation from % Deviation from
Observed Observed
Case P2 L n CV Target, n Target, P2
1 0.8 0.5 2 0.1 1.9802 -0.9875% 0.7999 -0.0074%
2 0.8 0.5 2 0.3 2.0226 1.1313% 0.7897 -1.2911%
3 0.8 0.5 2 0.5 2.1023 5.1132% 0.7757 -3.0334%
4 0.8 0.5 4 0.1 3.9648 -0.8802% 0.7991 -0.1081%
5 0.8 0.5 4 0.3 4.0047 0.1187% 0.7921 -0.9935%
6 0.8 0.5 4 0.5 4.0788 1.9711% 0.7804 -2.4478%
7 0.8 0.5 6 0.1 5.9476 -0.8729% 0.7991 -0.1068%
8 0.8 0.5 6 0.3 5.9849 -0.2516% 0.7941 -0.7361%
9 0.8 0.5 6 0.5 6.0606 1.0094% 0.7848 -1.9035%
10 0.8 2 2 0.1 1.9803 -0.9859% 0.7999 -0.0094%
11 0.8 2 2 0.3 2.0227 1.1334% 0.7902 -1.2282%
12 0.8 2 2 0.5 2.1011 5.0562% 0.7786 -2.6789%
13 0.8 2 4 0.1 3.9642 -0.8941% 0.7991 -0.1145%
14 0.8 2 4 0.3 4.0054 0.1348% 0.7925 -0.9377%
15 0.8 2 4 0.5 4.0812 2.0311% 0.7833 -2.0818%
16 0.8 2 6 0.1 5.9470 -0.8836% 0.7991 -0.1121%
17 0.8 2 6 0.3 5.9873 -0.2125% 0.7941 -0.7417%
18 0.8 2 6 0.5 6.0628 1.0468% 0.7863 -1.7072%
19 0.8 4 2 0.1 1.9803 -0.9853% 0.7998 -0.0216%
20 0.8 4 2 0.3 2.0223 1.1128% 0.7909 -1.1379%
21 0.8 4 2 0.5 2.1012 5.0594% 0.7809 -2.3913%
22 0.8 4 4 0.1 3.9644 -0.8894% 0.7993 -0.0929%
23 0.8 4 4 0.3 4.0064 0.1592% 0.7931 -0.8633%
24 0.8 4 4 0.5 4.0786 1.9638% 0.7855 -1.8070%
25 0.8 4 6 0.1 5.9472 -0.8806% 0.7991 -0.1068%
26 0.8 4 6 0.3 5.9857 -0.2390% 0.7946 -0.6701%
27 0.8 4 6 0.5 6.0596 0.9939% 0.7875 -1.5668%
28 0.9 0.5 2 0.1 1.9803 -0.9844% 0.8999 -0.0064%
29 0.9 0.5 2 0.3 2.0227 1.1352% 0.8900 -1.1065%
30 0.9 0.5 2 0.5 2.1022 5.1110% 0.8794 -2.2843%
31 0.9 0.5 4 0.1 3.9641 -0.8965% 0.8993 -0.0723%
32 0.9 0.5 4 0.3 4.0074 0.1845% 0.8937 -0.7016%
33 0.9 0.5 4 0.5 4.0770 1.9260% 0.8848 -1.6908%
34 0.9 0.5 6 0.1 5.9476 -0.8734% 0.8993 -0.0800%
35 0.9 0.5 6 0.3 5.9847 -0.2552% 0.8948 -0.5826%
36 0.9 0.5 6 0.5 6.0607 1.0117% 0.8869 -1.4545%
37 0.9 2 2 0.1 1.9803 -0.9846% 0.8999 -0.0065%
38 0.9 2 2 0.3 2.0224 1.1191% 0.8917 -0.9201%
39 0.9 2 2 0.5 2.1024 5.1206% 0.8843 -1.7415%
40 0.9 2 4 0.1 3.9643 -0.8914% 0.8993 -0.0829%

32
 Average Time (AT) Between Orders Fill Rate (FR)
% Deviation from % Deviation from
Observed Observed
Case P2 L n CV Target, n Target, P2
41 0.9 2 4 0.3 4.0070 0.1746% 0.8937 -0.6998%
42 0.9 2 4 0.5 4.0820 2.0488% 0.8867 -1.4744%
43 0.9 2 6 0.1 5.9470 -0.8836% 0.8992 -0.0853%
44 0.9 2 6 0.3 5.9882 -0.1969% 0.8949 -0.5629%
45 0.9 2 6 0.5 6.0672 1.1203% 0.8886 -1.2690%
46 0.9 4 2 0.1 1.9803 -0.9857% 0.9000 -0.0029%
47 0.9 4 2 0.3 2.0222 1.1097% 0.8934 -0.7369%
48 0.9 4 2 0.5 2.1017 5.0863% 0.8872 -1.4212%
49 0.9 4 4 0.1 3.9642 -0.8953% 0.8993 -0.0737%
50 0.9 4 4 0.3 4.0062 0.1545% 0.8943 -0.6329%
51 0.9 4 4 0.5 4.0797 1.9923% 0.8892 -1.2038%
52 0.9 4 6 0.1 5.9467 -0.8883% 0.8993 -0.0761%
53 0.9 4 6 0.3 5.9873 -0.2120% 0.8952 -0.5331%
54 0.9 4 6 0.5 6.0654 1.0905% 0.8901 -1.1045%
55 0.99 0.5 2 0.1 1.9803 -0.9842% 0.9900 -0.0042%
56 0.99 0.5 2 0.3 2.0213 1.0664% 0.9865 -0.3555%
57 0.99 0.5 2 0.5 2.1023 5.1134% 0.9825 -0.7566%
58 0.99 0.5 4 0.1 3.9646 -0.8838% 0.9896 -0.0393%
59 0.99 0.5 4 0.3 4.0058 0.1462% 0.9865 -0.3505%
60 0.99 0.5 4 0.5 4.0831 2.0766% 0.9841 -0.5969%
61 0.99 0.5 6 0.1 5.9468 -0.8866% 0.9893 -0.0682%
62 0.99 0.5 6 0.3 5.9856 -0.2406% 0.9870 -0.2994%
63 0.99 0.5 6 0.5 6.0618 1.0294% 0.9851 -0.4934%
64 0.99 2 2 0.1 1.9803 -0.9836% 0.9900 -0.0033%
65 0.99 2 2 0.3 2.0222 1.1119% 0.9877 -0.2352%
66 0.99 2 2 0.5 2.1002 5.0096% 0.9859 -0.4176%
67 0.99 2 4 0.1 3.9645 -0.8870% 0.9896 -0.0405%
68 0.99 2 4 0.3 4.0051 0.1271% 0.9876 -0.2405%
69 0.99 2 4 0.5 4.0828 2.0704% 0.9865 -0.3585%
70 0.99 2 6 0.1 5.9472 -0.8800% 0.9895 -0.0531%
71 0.99 2 6 0.3 5.9888 -0.1860% 0.9880 -0.2031%
72 0.99 2 6 0.5 6.0620 1.0325% 0.9868 -0.3242%
73 0.99 4 2 0.1 1.9803 -0.9863% 0.9900 0.0000%
74 0.99 4 2 0.3 2.0229 1.1462% 0.9884 -0.1656%
75 0.99 4 2 0.5 2.1018 5.0914% 0.9872 -0.2806%
76 0.99 4 4 0.1 3.9642 -0.8957% 0.9897 -0.0331%
77 0.99 4 4 0.3 4.0052 0.1309% 0.9883 -0.1686%
78 0.99 4 4 0.5 4.0760 1.8988% 0.9874 -0.2581%
79 0.99 4 6 0.1 5.9474 -0.8759% 0.9896 -0.0417%
80 0.99 4 6 0.3 5.9862 -0.2295% 0.9885 -0.1527%
81 0.99 4 6 0.5 6.0636 1.0599% 0.9876 -0.2469%

33
 Figure Captions

Figure 1 – Undershoots in a Periodic Review System

Figure 2 – Relation Between First Passage Time (u) and τ

Figure 3 – Illustrative Density Function of τ for m = 1.5, CV = 0.3

Figure 4 – E(τ ) that Satisfies Equation (5) and Associated Var(τ ) for n = 3, 6

Table Headings

Table 1 – Fractional Polynomial Approximations of E(τ ) as a Function of CV for Selected

Values of n

Table 2 – Fractional Polynomial Approximations of Var(τ ) as a Function of CV for Selected

Values of n

Table 3 – Settings of Independent Parameters

Table 4 – Simulation Results

34

View publication stats