Some insights regarding the optimal reorder period in periodic review inventory systems Edward A. Silver Haskayne School of Business The University of Calgary 2500 University Dr. NW Calgary, Alberta CANADA T2N 1N4 Phone: (403) 220-6996 Fax: (403) 282-0095 e-mail: [email protected]David J. Robb 1 Department of Information Systems and Operations Management The University of Auckland Private Bag 92019 Auckland NEW ZEALAND Phone: (64)(9)373-7599 ext.85990 Fax: (64)(9)373-7430 E-mail: [email protected]Abstract 2 The Periodic Review Inventory system is not only pervasive, but has an extensive literature dealing with various aspects, from its theoretical underpinnings through to its performance. However, the behaviour of the best review period with respect to basic inventory parameters such as demand and supply variability appears to be poorly understood. This analysis demonstrates and explains the somewhat counterintuitive results of how the best review period, a key decision parameter, changes as various parameters are modified. We also show that the cost function may be non-convex in the review period. Both Normal and Gamma distributions for lead-time demand are considered. The findings should also be of use to managers seeking improvement in inventory system performance. Key Words Inventory, Periodic Review, Stochastic Leadtimes, Normal Distribution, Gamma Distribution 1 Corresponding author 2 Acknowledgements. The research underlying this paper was supported by the Natural Sciences and Engineering Research Council of Canada under Grant A1485. Part of the research was conducted while the second author was on Research and Study Leave at the University of Calgary.
25
Embed
Some insights regarding the optimal reorder period in periodic review inventory systems
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Some insights regarding the optimal reorder period in periodic review inventory systems
Edward A. Silver Haskayne School of Business The University of Calgary 2500 University Dr. NW Calgary, Alberta CANADA T2N 1N4 Phone: (403) 220-6996 Fax: (403) 282-0095 e-mail: [email protected] David J. Robb1 Department of Information Systems and Operations Management The University of Auckland Private Bag 92019 Auckland NEW ZEALAND Phone: (64)(9)373-7599 ext.85990 Fax: (64)(9)373-7430 E-mail: [email protected]
Abstract2 The Periodic Review Inventory system is not only pervasive, but has an extensive literature
dealing with various aspects, from its theoretical underpinnings through to its performance.
However, the behaviour of the best review period with respect to basic inventory parameters
such as demand and supply variability appears to be poorly understood. This analysis
demonstrates and explains the somewhat counterintuitive results of how the best review
period, a key decision parameter, changes as various parameters are modified. We also show
that the cost function may be non-convex in the review period. Both Normal and Gamma
distributions for lead-time demand are considered. The findings should also be of use to
managers seeking improvement in inventory system performance.
Key Words
Inventory, Periodic Review, Stochastic Leadtimes, Normal Distribution, Gamma Distribution
1 Corresponding author 2 Acknowledgements. The research underlying this paper was supported by the Natural Sciences and
Engineering Research Council of Canada under Grant A1485. Part of the research was conducted
while the second author was on Research and Study Leave at the University of Calgary.
2
1. Introduction Treatment of inventory systems in which replenishments are conducted on a periodic basis
has been extensive, reflecting the ubiquitous nature of this method for controlling inventories.
Numerous complexities have been dealt with, including emergency replenishments (Bylka
2005), variable purchasing costs (Gavirneni 2004), and serially-correlated and inventory-
level-dependent demand (Urban 2005). However, despite the attention, the interaction of the
best review period (R*) with the basic parameters such as demand variability and leadtime
variability is not well understood. For example, many academics and practitioners, asked
what happens to the optimal review period when demand uncertainty increases, respond that
R* must always decrease. Others, perhaps more mathematically sophisticated, believe it
should always increase. However, the actual answer is “it depends” - on other parameters.
In this paper we demonstrate and explain these phenomena – initially observed in work
relating to date-terms trade credit (Robb and Silver 2004).
In our search for an explanation of this behaviour we discovered that while the expected total
relevant cost function is infinite at R=0 and R=∞, in between the tradeoff in the choice of R is
complex. The common argument that bigger R gives lower safety stock holding and shortage
costs doesn’t always hold true – the function may be non-convex. However, even when the
function is convex, demand and supply uncertainty may have opposite effects on the choice of
the best value of R.
Section 2 presents the model assumptions, notation, model, and decision rules. In Section 3
we describe a full factorial experiment utilised in our research. Section 4 demonstrates the
behaviour of the cost expression with respect to the review period, and Section 5 shows how
the optimal review period varies with incremental changes in basic parameters. We consider
both Normal and Gamma distributions to model leadtime demand.
2. Model Development In this section we describe the model environment, develop and discuss the behaviour of the
cost expression, and explain the methodology for selecting optimal decision parameters.
2.1 Model Environment We deal with the case of a single item with unit variable cost, c (see Table 1 for notation).
The inventory control system used is one that involves Periodic Review, and an Order-up-to
Level (R,S). Every R days the inventory is reviewed, and an order is placed (with a set-up
cost of A) to raise the inventory position to S. This order is available for consumption L (a
environmental conditions is ρ=100, 500, and 2000. For these values one may calculate
representative values of the probability of no stockout in a replenishment cycle (P1), e.g.,
setting R=w (i.e., without consideration of uncertainty), low, median, and high values
ofρwP −= 11 are ⎟
⎠⎞
⎜⎝⎛ −−−
200011,
500101,
100301 = (70%, 98%, 99.95%), respectively, i.e.,
reflecting a wide range of operating conditions.
For leadtimes, we consider means (µL) of 0.5, 10, and 30 days. Leadtime variability,
vL=σL/µL, ranges from extremes of almost deterministic, 0.05, up to 0.5, with a mid-point of
0.25. These figures are based on our observations of both domestic and international
leadtimes. For example, observations of vL for some two thousand shipments in the New
Zealand building industry average 0.30, with relatively little dependence on µL (i.e., σL tended
to increase linearly with µL).
Demand variability, vD=σD/µD (or, more properly, the coefficient of variation of the error in
the demand forecast) ranges from 0.1 to 10 with a mid-point of 2.5. These values may appear
high, but the time period considered is one day. The median coefficient of variation of
national monthly sales for more than 6000 products with positive annual sales in a major New
Zealand building products distributor is 1.58. If there were no autocorrelation in demand, this
would indicate a median value of vD of 8.6 (i.e., 1.58*√30). With positive autocorrelation the
actual values of vD would be somewhat lower. We conservatively use 2.5 as the mid-point.
Other empirical studies render similar ranges, e.g., Nahmias & Smith (1994) cite figures
which would convert to the range 0.55 to 7.07.
4. Behaviour of the Cost Expression as a Function of R In this section we demonstrate how and where the Expected Total Relevant Cost expression
may be non-convex for the case of Normal leadtime demand (X). We denote the first two
terms of (4) by R
wRE22
2
+= , and the last by 222)(*)(* LLDLr vvRkGR
kU µµρ++⎟
⎠⎞
⎜⎝⎛ += . E,
reflecting the cost curve for the deterministic EOQ case, is clearly convex in R. However, it
is not difficult to find cases where U, reflecting the incorporation of uncertainty, is non-
convex in R. Indeed we have proved that for k≥0 (the region of interest) U is either
decreasing in R, or has at most one local maximum (note that U is infinite at R=0 and it tends
to zero as R→∞). Moreover, cases exist in which the total expression NETRC=E+U is also
non-convex. Examples of this behaviour are illustrated in Figure 1 (with a local minimum at
R=0.248 and a local maximum at R=46.1) and Figure 2 (with two local minima, at R=5.42
and R=62.1, and a local maximum at R=34.9).
8
The full factorial experiment described in Section 3 was conducted. For each experimental
condition, R* was evaluated using a fine grid, from a low value of 0.06 to a high value of 115
days (all solutions were found within this range). The actual grid used was
0.06(0.002)1(0.01)5(0.02)20(0.05)115, i.e., involving 3521 evaluation points, with finer
resolution at lower values where the cost function generally changes more rapidly. Table 3
provides a partial listing of the results, for parameters in their centroidal position (row 1) and
the 25 extreme positions (e.g., row 2 has all the parameters at their low values and in row 3
only ρ is a its high value). Note that the last column, labelled “Type” will be explained later
in this section.
Among the 35 (243) experimental conditions NETRC has a local maximum in 9 instances.
These instances are (vD=10, µL=0.5, vL=(0.05, 0.25, 0.5), w=(1,15), ρ=100) and (vD=10,
µL=10, vL=(0.05, 0.25, 0.5), w=1, ρ=100). Note that none of these are included in the partial
experiments of Table 3. Local maxima are thus only observed when vD is high and ρ is low,
and never when µL is high or w is high. We have found no such cases of local maxima in
NETRC for Gamma X, although U itself may be non-convex.
Substituting ρRR =' and ρ
µµ
ρ/' 2
22
⎟⎟⎠
⎞⎜⎜⎝
⎛+==
D
LLL v
vCC into (9), one obtains
( )( ) ''''
)'( 1 CRRpfRv
ERNETRC rNrND ++= −
≥ (11)
With considerable separate experimentation we have found that ( )( )''
'' 1 RpfR
CRrNrN−
≥+
has a local maximum only when C′<0.085, i.e., when C′ is small enough. Without a
local maximum in U it is less likely that E+U is non-convex. Thus, small µL and vL
and large vD would tend to cause non-convexity. Large vD, which appears as a
multiplier in U, would tend to make U more important relative to E. The smaller w is,
the more important the non-convex U is relative to E, hence the more likely it is that
E+U is non-convex. In summary, non-convexity in NETRC(R) is more likely for low
µL, low vL, high vD, and low w (these results are validated by the experiments). The
behaviour with ρ is trickier as it appears in two places in U. However, we have been
able to show analytically that ρd
dUtends to be positive the lower ρ is, which again
agrees with experimental results regarding non-convexity.
9
Understandably, the non-convexity only appears at relatively high levels of uncertainty in X.
However, non-convexity can occur even with very little variability in the leadtime. This
suggests that one has to be careful in using solution methods based on convexity, e.g., Eynan
& Kropp (2004). Also, Rao (2003) as proven convexity for the case of deterministic
leadtime. Moreover, a recent paper shows convexity in the case of lost sales and no set-up
costs, when order crossover is allowed and inventory is charged at time of order arrival rather
than order placement (Janakiraman and Roundy 2004)
To provide more insight into the non-convexity we introduce the term Ra, the value of R that
minimises U (note: w minimises E). We may categorise cases as one of four Types (see the
first two rows of Table 4). Type 1 (the most common in our experiments) has Ra<R*<w.
Type 2, the second most common, has w<R* but Ra is undefined (infinite). Type 3 has
w<R*<Ra and Type 4 has Ra<w<R*. Note that is not possible for R* to exceed both w and Ra
(a related observation is that if w<Ra, then R* must be between w and Ra). Note also that
Ra<ρ/2, since one can prove analytically that 0' 5.0'
<=RdR
dU, i.e., the U function is always
decreasing as R′→0.5. Hence, if there is an earlier local minimum, there must be a local
maximum between the local minimum and R′=0.5.
In Table 5 and in Figures 1 through 4 we present four cases, one from each of the types, from
the 35 conditions. Figure 1 shows (one of the 6 examples of) a Type 1 case with a local
maximum. Similarly, Figure 2 demonstrates (one of 3 examples of) a Type 2 case in which a
local maximum occurs. Figures 3 and 4 represent the two remaining classes, viz., Types 3
and 4. While U has a local maximum (at R=16.84) in Figures 3 and 4, the NETRC has no
such feature. Indeed, within the range of parameters investigated we have not found any local
maxima for Type 3 or Type 4 cases.
5. Behaviour of the Optimal Review Period (R*) with respect to basic parameters
5.1 Methodology for evaluating the impact of parameter changes While one can exploit the results of the previous section to indicate how R* varies given large
changes in the basic parameters (e.g., using Table 3), the impact of marginal changes upon R*
are just as important, particularly for managers engaged in continuous improvement. In
particular we would like to know how R* changes with minor perturbation in the parameters
of interest, viz., µL, vL, and vD. Ideally one would like analytical expressions
The last term is clearly negative. Similarly, in the region of interest (k≥0), the term in the
square brackets is negative, since (i) it is negative at k=0, (ii) the derivative of the term with
respect to k is ( ) 0≥≥ kprN , and (iii) the limit as k→∞ = 0.
Thus 0),(2
<∂∂
∂
L
L
RRNETRC
µµ which proves that R* increases as µL increases.
In a similar manner one can show the partial derivative with respect to the reorder period and
the leadtime variability,
⎟⎟⎠
⎞⎜⎜⎝
⎛
⎪⎭
⎪⎬⎫
⎪⎩
⎪⎨⎧
⎥⎦
⎤⎢⎣
⎡⎟⎟⎠
⎞⎜⎜⎝
⎛⎟⎠⎞
⎜⎝⎛ +−⎟⎟
⎠
⎞⎜⎜⎝
⎛+
+=
∂∂∂ −
≥−
≥ ρρρρ RpfCRRpRCR
dvdC
CRR
vvR
vRNETRCrNxNrN
L
D
L
L 11
232
2
23)(
)(2
),( (A5)
is also negative in the region of interest, since 02
2
2
>=D
LL
L vv
dvdC µ , and again the curly brackets
term is negative.
The second partial derivative with respect to R and vD is not as simple, as the term vD also
appears as a multiplier of the large bracketed term, with the latter involving vD through its
appearance in C.
Noting that 02
3
22
<−
=D
LL
D vv
dvdC µ , one obtains the following:
15
⎥⎦
⎤⎢⎣
⎡⎟⎟⎠
⎞⎜⎜⎝
⎛+
+⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜
⎝
⎛
⎥⎦
⎤⎢⎣
⎡⎟⎟⎠
⎞⎜⎜⎝
⎛+
+−⎟⎟
⎠
⎞⎜⎜⎝
⎛
+
+=
∂∂∂
−≥
−≥
−≥ ρ
µρρρ
ρµ RpfCR
Rv
vRpfCR
CRRpRCRR
R
vRvRNETRC
rNxND
LLrNxNrN
L
D
D
12
211
2
2
221
)(
),(
(A6)
Substitutingρ
µµ
ρρL
LandCCRR === ',',' , (A6) may be re-expressed as
( ) ( )[ ]⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜
⎝
⎛
+
++−+
+
+=
∂∂∂ −
≥−
≥ '''
')'(21'
23
'')'()(),( 1
'2'
1'2
22
RpfCR
CRRRpRR
CRRR
vRvRNETRC
rNxN
LL
rNLL
D
Dµµ
µρµ
(A7)
16
References
Burgin, T. A. 1975. The Gamma Distribution and Inventory Control. Operational Research Quarterly 26 (3):507-525.
Bylka, S. 2005. Turnpike policies for periodic review inventory model with emergency orders. International Journal of Production Economics 93-94:357-374.
Chopra, S, G Reinhardt, and M. Dada. 2004. The Effect of Lead Time Uncertainty on Safety Stocks. Decision Sciences 35 (1):1-24.
Eynan, A, and DH Kropp. 2004. Effective and Simple EOQ-Like Solutions for Stochastic Demand Periodic Review Systems.
Eynan, A., and D. H. Kropp. 1998. Periodic review and joint replenishment in stochastic demand environments. IIE Transactions 30 (11):1025-1033.
Fortuin, L. 1980. Five popular probability density functions: A comparison in the field of stock-control models. Journal of the Operational Research Society 31 (10):937-942.
Gavirneni, S. 2004. Periodic review inventory control with fluctuating purchasing costs. Operations Research Letters 32:374-379.
Janakiraman, G., and R. O. Roundy. 2004. Lost-Sales Problems with Stochastic Lead Times: Convexity Results for Base-Stock Policies. Operations Research 52 (5):795-803.
Kapuscinski, R, RQ Zhang, P Carbonneau, R Moore, and B Reeves. 2004. Inventory Decisions in Dell's Supply Chain. Interfaces 34 (3):191-205.
Nahmias, S., and S. A. Smith. 1994. Optimizing Inventory Levels in a Two-Echelon Retailer System with Partial Lost Sales. Management Science 40 (5):582-596.
Rao, US. 2003. Properties of the Periodic Review (R,T) Inventory Control Policy for Stationary,Stochastic Demand. Manufacturing & Service Operations Management 5 (1):37-53.
Robb, DJ, and EA Silver. 2004. Inventory Management under Date-terms Supplier Trade Credit with Stochastic Demand and Leadtime. In Haskayne School of Business Working Paper: The University of Calgary.
Shore, H. 2004. A general solution for the newsboy model with random order size and possibly a cutoff transaction size. Journal of the Operational Research Society 55 (11):1218-1228.
Silver, E. A., D. F. Pyke, and R. Peterson. 1998. Inventory Management and Production Planning and Scheduling. 3rd ed. New York: John Wiley & Sons.
Tyworth, J. E., and L. O'Neill. 1997. Robustness of the Normal Approximation of Lead-Time Demand in a Distribution Setting. Naval Research Logistics 44:165-186.
Urban, T. L. 2005. A periodic-review model with serially-correlated, inventory-level-dependent demand. International Journal of Production Economics 95 (3):287-295.
Table 1. Some Notation (note: $ may be replaced by any unit of currency) A $ Fixed cost component incurred with each replenishment B2 $/$ Fractional charge per unit short c $/unit Unit variable cost of an item
C Days 2
22
D
LLL v
vC
µµ +=
C′ - C′=C/ ρ D Units/day Demand per day, distributed with mean and standard deviation ),( DD σµ . k - Safety factor L Days Replenishment leadtime, distributed with mean and standard deviation ),( LL σµ . r $/$/day Inventory carrying charge R Days Reorder Period R′ - R/ρ Ra Days The local minimum of the normalised shortage and safety stock function (U) S Units Order-up-to Level, xx kS σµ += SS Units Safety Stock T Days The normalised set-up cost and cycle stock holding cost function U Days The normalised shortage cost and safety stock cost function vD - Demand variability, vD=σD/µD vL - Leadtime variability, vL=σL/µL
w Days The Economic Order Quantity expressed as a time period, also known as the Wilson number, crAw
Dµ2
=
X Units Demand during the “critical protection period”, R+L, distributed with mean and standard deviation ),( XX σµ . Coefficient of
variationx
xxv
µσ
=
ρ Days Shortage to Holding Cost ratio, ρ=B2/r (interpretation: number of days holding costs equivalent to the cost of running out of stock, B2) µL Days Leadtime mean µL′ - µL′/ ρ
18
Table 2. Properties of Statistical Distributions, adapted from (Fortuin 1980) Distribution Normal Gamma Probability Density Function fx(x)
)(2
1),,( 2)( 2
∞<<−∞
=−
−
x
exfxx
xxxxN
µ
πσσµ
)0,0,0(
)(),,(
1
>>≥Γ
=−−
γαα
γγαγαα
x
exxfx
xG
Regularised Probability Density Function fr(k)
π2)()(
22k
urNekfkf
−
== )(),(
1
αα
α
Γ=
−− k
rGekkf
Regularised Complementary Cumulative Density Function pr≥(k)
dttfkpk rNrN ∫∞
≥ = )()( dttfkpk rGrG ∫∞
≥ = ),(),( αα
Regularised Linear Loss Integral Gr(k)
)(.)(
)()()(
kpkkf
dttfktkG
rNrN
k rNrN
≥
∞
−=
−= ∫
)(),()(),1(
),()(),(
αααααααααα
αα
Γ+Γ+−++Γ
=
−= ∫∞
kkk
dttfktkGk rGrG
Where the (upper) Incomplete Gamma Function ∫∞ −−=Γx
t dttex 1),( αα and the Complete Gamma Function ∫∞ −−=Γ=Γ0
1)0,()( dtte t ααα
19
Table 3. Centroidal and Extreme (25) Results of (35) Factorial Experiment
Note: The Four Types are as follows: 1. Ra<R*<w. 2. w<R* but Ra is undefined (infinite). 3. w<R*<Ra . 4. Ra<w<R*.
Note: ∆vD(R*) is the partial derivative. Other abbreviations are Normal, Gamma, Local Maximum, Local Minimum, Analytic Value, and Numeric Approximation.
22
Figure 1. Normalised Expected Costs for Normal X, with high demand variability, but all other factors at their lowest levels.
Figure 2. Normalised Expected Costs for Normal X, with high demand variability and high average leadtime, but all other factors at their lowest levels.
vD=10, µL=10, vL=0.05, w=1, ρ=100w<R*=5.42, Ra undefined, ∆vD(R*)<0
0
20
40
60
80
100
0 50 100 150Reorder Period, R (days)
Ave
rage
NE
C (d
ays)
EUE+U
24
Figure 3. Normalised Expected Costs for Normal X, with high leadtime variability, but all other factors at their lowest levels.
Figure 4. Normalised Expected Costs for Normal X, with high leadtime variability and Wilson number, but all other variables at their lowest levels (note the multiplier on U for visibility. U is very small relative to E, and thus E+U is hard to distinguish from E).