Inventory Management Under Demand Uncertainty
Inventory Management Under Demand Uncertainty
Özalp Özer
Date: 02/2004; Rev’d: 06/2011
0. Overview
In this unit, you will learn the essentials of inventory management under demand
uncertainty.
In Section 4, we introduce the concept of risk pooling over multiple stocking locations.
We illustrate the benefit and cost of consolidating two warehouses into a single
warehouse.
In Section 5, we provide a tool, an exchange curve, to trade off service level and dollar
investment in safety stocks.
By the end of this unit, you should be able to due the following.
1. Demand Uncertainty
In the previous unit, we treated demand as a known quantity and provided models
with trades-offs between set-up (or fixed ordering) cost and holding cost. In reality,
however, demand for products displays variation over time. Several factors such as
changing consumer preferences and economic trends create uncertainties in demand.
Firms try to estimate future demand patterns using forecasting methods, which are
based on historical data and expertise about the market. Even though firms are aware of
the concept of demand uncertainty, they often come up with a single point estimate of
future demand. The following statement from a marketing manager is quite common.
“We expect demand for our product to be 10,000 units during the next selling season.”
Most of us may have also heard that forecasts are always wrong. Does this mean
that firms should not even try to come up with a forecast? No. Firms do need to
estimate demand for planning purposes. However, enhancing forecast statements by
providing scenario analysis improves the quality of the forecast. Consider following
two statements.
Firm A: “We expect demand for our product to be on average 10,000 units.
However, if the weather is sunny during our sales season we expect to sell 20,000
units; whereas, if it is rainy we expect to sell nothing.”
Firm B: “We expect demand for our product to be on average 10,000 units.
However, if weather is sunny we expect to sell 11,000; whereas, if it is rainy we
expect to sell 9,000 units.”
Both of these statements suggest that the average demand will be 10,000 units.
However, the first statement states that not selling even a single product is a possibility.
It also states that Firm A has a potential to sell almost twice more than Firm B. Either
scenario may be preferable, depending on risk preferences. However, the decision of
how much inventory to stock for Company A is more challenging than that for
Company B.
If no variability exists, hence the demand has no uncertainty, then the manager can
order exactly the average demand and avoid completely the possibility of stocking out.
However, when the demand is uncertain, the manager may run out of stock even if she
orders the mean demand. As discussed in the previous unit, stockout results in shortage
cost due to lost sales and loss of goodwill. To cope with uncertainty in demand, firms
may need to produce more than the average demand. The aim of this unit is to provide
you with some inventory management tools to counter the adverse effects of demand
uncertainty.
25%
20%
15%
Probability
10%
5%
0%
40 50 70 90 100 110 130 150 160
Unit Sales
For this example, the first row in Table 1 reports possible realized demand at the
end of a selling season. The second row reports the probability of observing d units of
Often histograms may resemble a bell shaped curve. For those instances, a Normal
distribution is a good approximation to model demand uncertainty and to provide a
probabilistic forecast.
Likelihood
AVG d
FIGURE 2: Normal Distribution with Mean AVG and Standard Deviation STD.
Example: Suppose that demand follows a Normal distribution with mean 100 units
and standard deviation 20 units. What is the probability of observing demand less
than or equal to 133 units?
The above concepts suffice to study the effect of demand uncertainty on stocking
decisions and to show why and how firms should protect themselves against demand
uncertainty. The detail of forecasting is the job of a statistician or a forecasting
specialist. During the rest of this unit, we assume that the probabilistic forecast of
Several managers from various industries face such notoriously difficult decisions.
A newsvendor located at the corner of your street has to decide how many copies of a
particular issue of a magazine she has to stock. A garment retailer who sells fashion
items with a short selling season (Zara’s winter collection and Zekitriko’s swim suit
collection to name a couple) has to decide how many items she has to stock in her retail
store. A toy manufacturer has to decide how many items to manufacture in Asia and
ship to North American retailers who sell during the Christmas season. A grocery store
manager who has to sell some of his products on a given day is another example.
All these managers face two fundamental trade-offs. At the end of the selling
season, if the realized demand turns out to be less than the available inventory, then the
manager needs to salvage leftover inventory at a discounted price. For certain products,
she may have to pay a disposal fee for leftover inventory. On the other hand, if realized
demand exceeds the available inventory, then the manager loses customers and hence
opportunity to make more money. Subsequently she may build a reputation of not
having enough products on hand to satisfy fickle consumer demand. The following
shopping season, consumers may choose to visit other retailers.
The newsvendor has to decide how many copies to buy from the regional distributor
early in the morning. An obvious solution is to buy 100 copies, the average demand.
Note, however, that not having a copy when a customer wants one costs the
newsvendor
We refer to this cost as the cost of underage. On the other hand, having an unsold copy
at the end of the day costs the newsvendor
We refer to this cost as the cost of overage. Intuitively, the vendor should buy more
than the average demand because the cost of underage is much higher than the cost of
overage. This observation suggests that optimal number of papers to stock should
depend on the costs of overage and underage.
Question: Which of the two costs are reflected in your company’s profit and
loss statement? Do you think the way these costs are reported in the P&L
statement affect inventory related decisions?
The cost of underage represents the marginal benefit of having one more unit of
inventory in stock when needed. Assume that the newsvendor is already committed to
stocking Q = 100 units in the morning. During the day, if she has one more unit of
paper in stock, she may earn cu = 30 cents more. From the historical data, the
newsvendor estimates that demand during a given day exceeds Q = 100 units with
probability 1 − F (Q) = Prob(Demand > Q units) = 50%. She can sell this additional unit
only if demand exceeds Q units. Hence, the expected marginal benefit of stocking one
more unit is
cu * (1 − F (Q)) ,
which is equal to 30 * 50% = 15 cents. On the other hand, the cost of underage represents
the marginal cost of having one more unit of inventory in stock than is not needed. The
expected marginal cost of having an additional unit is
co * F (Q) ,
which is 8.5*50% = 4.25 cents (Why?). The optimal stocking level should strike a
balance between the marginal benefit and cost of having an additional unit. The
newsvendor can increase his stocking level one unit at a time until the marginal benefit
of doing so is less than the marginal cost. In this particular example, increasing the
stock level from 100 to 101 is worthwhile. The optimal order size Q* should satisfy the
following
cu * (1 − F (Q)) = co * F (Q) .
This ratio is the probability of satisfying all demand during the day if the
newsvendor stocks Q * units at the beginning of the day. For our example, this ratio is
30/ (8.5+30) = 0.78. Thus, the newsvendor should stock enough newspaper such that
the probability of no stock out is 0.78 * 100% = 78%.
Next the newsvendor must convert this desired probability of stockout into stocking
quantity. This conversion depends on the distribution of demand or the probabilistic
forecast. Based on the historical data and her judgment, the newsvendor estimates that
demand for the newspaper follows the probabilistic forecast provided in Table 1, from
which the optimal stock quantity is Q * = 110 units. Given this quantity, her expected
cost of underage and overage is given by
Notice that instead of a marginal cost benefit analysis, the newsvendor can also
calculate her expected cost under all possible stocking quantity scenarios that are
Q = 20,...,160 units. She can choose the quantity that minimizes the expected cost. This
solution gives us the same answer as the marginal cost/benefit analysis.
Assume that demand for the newspaper follows a Normal distribution with mean
100 and standard deviation 20. We use the Normal table to convert this critical ratio to
an inventory level. For 0.78, z=0.7721. Thus, Q* should lie 0.7721 standard deviations
from the mean demand 100, which is 100 + 0.7721*20 = 115. Note that, given a
demand distribution, we convert a desired probability of not stocking out (78%) to an
inventory level (115 units).
1. The newsvendor orders more than the average demand. The average demand is
100 units, but the optimal inventory decision is 110 units. This outcome is due to
two reasons. First, the cost of underage is larger than cost of overage. The
opportunity cost of not having one more unit is larger than the potential cost of
creating this opportunity. Second, when the probabilistic forecast is symmetric
around the average, the chance of observing a demand smaller than the average is
less than 50% (Why?). Intuitively, the newsvendor has a higher chance of making
more money than losing if she orders more than the average demand.
2. The marginal analysis suggests that allowing for the possibility of a stockout
could be a better choice than promising to meet 100% of demand. In other words,
stocking inventory to meet possibly less than 100% of realized demand, and hence
________________________________________________________________
Exercise: If your local grocery store manager runs out of stock very rarely, can
you say that you are paying for a big markup? Why?
Exercise: How does the standard deviation of probabilistic forecast affect the
stocking decision of the newsvendor? If the standard deviation increases from
20 to 40, how much inventory should the newsvendor stock?
________________________________________________________________
3. Consider the above newsvendor example with the probabilistic forecast given in
Table 1. What should the newsvendor do if the critical ratio turns out to be 95%?
For this scenario, we do not have any Q that corresponds to this critical ratio. In this
case, the newsvendor should pick a quantity that gets her closer to this ratio and
results in lower expected cost. In this example, she can calculate expected cost both
for when Q=150 and for when Q=130. Next, she can choose the stocking level that
results in the smallest expected cost.
Consider, for example, the distributor of Coke. The distributor faces random
demand for the product and receives the supply from the manufacturer. Delivery often
received after a fixed leadtime. For example, when the distributor places an order, he
receives the orders two weeks later. Demand is uncertain. Hence, the distributor may
need to hold inventory even if no fixed cost is charged for ordering the product. Three
main reasons for this are as follows.
Next we provide inventory models that help the distributor to better manage
inventory. She has to decide (i) how often the inventory status should be determined (ii)
when to place a replenishment order and (iii) how many cases of Coke to order.
For an ongoing inventory problem, the manager carries unsold inventory over to the
next time unit and incurs a holding cost. On the other hand, if demand for a product
exceeds available stock in the store, then the manager faces a stockout situation and
incurs a shortage cost.
The severity of the shortage cost depends on how the unmet demand is treated. Two
extreme alternatives are the following. All unmet demand can be backlogged; that is,
this demand will be satisfied through orders arriving in the future. Alternatively, all
unmet demand can be lost; that is, this demand disappears. Reality often lies
somewhere between these two extremes; that is, part of unmet demand is backlogged
while the rest is lost. The inventory policies discussed in this unit can safely be applied
to the full backlogging case. Analysis of a lost sale case is beyond the scope of this
unit. For a high service system, the model discussed here can be used as an
approximation for the lost sales problem. By service, we mean the percentage of
demand satisfied through on-hand inventory. Next, we define four important inventory
related concepts.
On-hand inventory is the physical inventory available in a store that the store can
use to satisfy demand directly. This quantity can never be negative.
Inventory on-order is the inventory that has been ordered but not yet received.
Safety stock is the inventory held in addition to average demand to hedge against
costly stockouts due to demand uncertainty. In other words, safety stock is the average
Question: Does the manager need to carry safety stock when demand is known
with certainty? Why?
Consider a product for which the replenishment leadtime is zero and the fixed cost of
ordering is negligible. Before placing an order, the manager can wait until on-hand
inventory is depleted. By doing so, she avoids carrying unnecessary inventory while
satisfying all demand given that replenishment orders arrive immediately. Hence, zero
leadtime with zero fixed ordering cost make holding safety stock unnecessary.
Question: Can you think of a product for which the leadtime is zero?
Answer: Water supply.
For an ongoing inventory problem, the key issue is to protect the system against
uncertainty in demand over the replenishment leadtime. Let
The average demand during the leadtime is the product of the average demand per
period and the leadtime.
DLT = L * AVG
The standard deviation of leadtime demand is the product of the standard deviation of
demand per time unit and the square root of the leadtime.
σ LT = L * STD .
If the average demand is 40 cases per week, the standard deviation is 10 cases per week
and replenishment takes 4 weeks, then the average leadtime demand is 160 cases and
the standard deviation of leadtime demand is 20 units. (Remember the rule: do not add
standard deviations! Add the variances and take the square root to obtain the total
standard deviation.) The standard deviation of leadtime demand captures the demand
uncertainty over the leadtime.
As in the newsvendor problem, if the demand per unit time follows a Normal
distribution with AVG and STD then the leadtime demand is normally distributed with
mean L*AVG and standard deviation L * STD . In fact, researchers have shown that for
long replenishment leadtime problems, using a Normal distribution to model demand
uncertainty works very well. In other words, if the manager faces a long replenishment
As we will see, anything management can do to reduce demand uncertainty over the
leadtime improves inventory performance. For example, helping suppliers to improve
operations may reduce demand variability during the replenishment leadtime. Such
improvements lead to better inventory performance as long as management is already
using a sound inventory management method.
The above discussion assumes that the leadtime is known with certainty. However,
leadtimes may be uncertain as well. For example, transportation time may depend on
weather conditions. If successive replenishment orders do not cross each other before
reaching to the store, and successive leadtimes do not depend on each other, then we
use the following formula to calculate the standard deviation of leadtime demand.
where AVGL is the average lead time and STD L is the standard deviation of the
leadtime.
Recall from our discussion in the previous unit that when demand is known,
determining the order size is equivalent to determining the order interval. When
demand is uncertain, however, this equality does not hold. The manager has to decide
how often to review the inventory status and place replenishment orders. We either fix
the order size or fix the order interval. This distinction gives rise to the two main
inventory review systems. The first is the continuous review (fixed-order-quantity)
system and the second is the periodic review (fixed-time-period) system.
In a continuous review system, the manager knows the inventory status at all times.
Demand is recorded as it occurs. Due to the advances in information technology,
continuous review systems are easy to implement. The computerized checkout systems
in supermarkets, for example, instantly record all transmissions and update the
inventory level. Note, however, that knowing the inventory level at all times does not
prevent a stockout situation. (Why?) Continuous review is also sometimes known as
transaction reporting because all demand is recorded as it occurs.
In a periodic review system, the manager reviews inventory only at discrete points
in time. The manager counts inventory on-hand and places replenishment orders at
specific time intervals, such as at the end of each day, week or month. When necessary,
during these review periods, management places an order to bring the inventory
position back up to a desired level. A familiar example of a periodic review system is a
vending machine. The vendor company periodically checks and fills the vending
Question: Which review system would your supplier wish that your company is
using?
Answer: The supplier may prefer determined order intervals of the periodic
review system. With a continuous review system, he needs to be ready to ship
anytime.
Inventory Level
Inventory Position
R+Q
Inventory On-hand
Time
FIGURE 3: Inventory position and inventory on-hand over time under an (R,Q) policy
We need to determine the optimal values for two parameters, which are R and Q.
The EOQ formula provides a reasonable approximation to set the order quantity. We
set the reorder point to
R = DLT + ss
where ss stands for safety stock. Hence, the reorder point is the sum of two parts. The
first is the average demand during the replenishment leadtime, which equals L*AVG.
This part ensures that enough inventory is available to satisfy average demand until the
next order arrives. The second is the safety stock, which is necessary to protect the
system against demand uncertainty during the replenishment leadtime, that is, to protect
the system against larger-than-average demand during the replenishment leadtime. The
safety stock level is set to ensure that demand during the leadtime is less than or equal
to R during at least α of the order cycles. For example, if customer demand were
satisfied during 90 order cycles out of 100, then α is equal to 90%. In other words, R
satisfies the following equality
Similar to the newsvendor problem, the manager needs to know the leadtime
demand distribution to decide on the safety stock. Assume that leadtime demand
follows the probabilistic forecast provided in Table 1. If the service level is 97% then
safety stock is given by
ss = z * σ LT ,
For a given R and Q, we calculate the average total annual cost as:
D Q
TC ( R, Q ) = ( ) K + ( + ss ) h .
Q 2
The first is the average set up cost and the second is the average holding cost. Note that
the average holding cost includes the cost of holding safety stock. Larger safety stock
results in higher holding costs. Cost of holding inventory is due to cycle stock, which is
defined in Unit 7, as well as safety stock. Recall that when demand is known, inventory
holding cost is due to the cycle stock only.
________________________________________________________________
Answer: Since demand per unit time is Normal, leadtime demand also follows a
Normal distribution with
To calculate the desired level of safety stock, we obtain from the standard
Normal table the constant z that corresponds to 97.5%. It is 1.96. Hence,
The order-point, order-up-to policy is similar to an (R,Q) policy. We use an (s,S) policy
when the quantity demanded can be larger than one unit. For example, an arriving
customer may demand 15 units. For such systems, we set s=R and S=s+Q. The
manager places a replenishment order whenever the inventory position falls below the
order point s. Note that the order size under an (s,S) policy is variable. Due to a larger
demand, the inventory position may fall significantly below s. In this case, the order
quantity will be larger than Q units.
Question: Can you think of an example from your company for which base
stock policy could be a good candidate?
Under a (T,S) policy, the manager reviews the inventory every T units of time. If the
inventory position is less than S, she orders enough units to bring the inventory position
up to S. In many cases, the review period T is given. For example, a firm may already
have a fixed review cycle due to physical constraints or a regular delivery schedule.
Alternatively, the manager can set T to balance holding cost and ordering cost using the
EOQ formula; that is, T=Q*/AVG where Q* is the solution to the EOQ formula and
AVG is the average demand. Note that, under this policy, the system replenishes
inventory every T+L periods of time, where L is the replenishment leadtime. Similar to
the logic we described in Section 3.4, the manager has to protect the system against
demand uncertainty during the leadtime plus a review period. The average demand for
T+L periods is
DTL = (T + L) * AVG
σ TL = T + L * STD .
ss = zσ TL ,
where z is the constant corresponding to the desired service level. We set the order-up-
to level to
S = DTL + ss .
______________________________________________________________
Exercise: The owner of the vending machines at your company visits and
replenishes the machines every day even if vending machines are full. He
estimates that demand for Pepsi during any given day follows a Normal
distribution with average demand 200 units per week and standard deviation 24
units per week. He also wishes to have a 97.5% service rate. To achieve this
The VP of the Supply Chain, Emily Chung, is considering consolidating the two
warehouses into one warehouse, which will satisfy demand from both France and Italy.
She reasons that such a consolidation would result in reduced overhead and better
coordination of operations with manufacturing partners located in China. Under the
current system it takes 4 weeks to ship products from China to the warehouses in
Europe. She estimates that this replenishment leadtime will remain the same under the
new system. She also claims that a higher-than-average demand at one location may be
offset by a lower-than-average demand at the other location. As a result, she expects to
carry less safety stock. She is also aware of the current distribution system’s advantage
over the single warehouse system. Under the current system, warehouses are located
closer to the end customer. This proximity enables HP to satisfy demand faster.
Nevertheless, she would like the new senior operations manager, Otis, to calculate
inventory related gains from such a consolidation. She states that under the current
system service level is 95%; that is, each warehouse carries enough safety stock for A
level items so that the probability of stocking out is 5%. Emily’s only requirement is
that under the new warehouse system the service level remain the same.
Otis starts his analysis by considering the inkjet printers. In France, annual demand
for inkjet printers follows a Normal distribution with mean 15,000 units and standard
deviation is 1,000 units. On the other hand, In Italy, annual demand for HP printers
follows a Normal distribution with mean 10,000 and standard deviation 900 units. Otis
calculates the constant z from the standard Normal distribution Table. For a service
level of 95%, z should be set equal to 1.65. Hence, the current safety stock at each
warehouse should be
Total safety stock for inkjet printers held in Europe is 905 units. Under the new
warehouse system, demand will be satisfied through a single warehouse. Total demand
follows a Normal distribution with mean 15,000+9,000 units. Six years after
Otis also finds out that the cost of holding inventory is $30 per inkjet printer per
unit time for both warehouses in Italy and France. Percentage savings in safety stock
30 * (905 − 641)
dollars for the inkjet printer is = 29% . He knows that HP’s total
30 * 905
inventory investment in France and Italy is approximately $6Million. If HP were to
achieve similar inventory reduction across all product categories, he concludes that the
consolidation has the potential to slash inventory related costs by about $1.74 Million.
________________________________________________________________
Exercise: After Otis runs all calculations for a sample of 100 products, he
reports final results to Emily. She expresses her appreciation about such a
thorough analysis. However, Emily points out that under this consolidation, HP
should be saving also from fixed ordering costs. She asks Otis to investigate
whether the order quantity for the inkjet printer remains the same after the
consolidation. She tells Otis that the fixed cost of ordering inkjet printers is
$500 per order, independent of the distribution strategy. Do you agree with
Emily? Why/Why not?
________________________________________________________________
The above analysis assumes that demand for HP products in Italy and in France are
independent. In other words, observing a higher-than-average demand at one location is
independent of observing a lower-than-average demand in the other location. However,
if demand at each location were positively correlated, then demand realizations would
be similar. Observing a higher than average demand at one location would signal a
similar pattern in the other location. Thus, the benefits from risk pooling diminish if the
demands in separate locations are positively correlated.
________________________________________________________________
HC ( z ) = h * z * σ LT .
To delineate an exchange curve, we increase the service level and calculate the holding
cost using the above equation. Figure 4 is an example of an “exchange curve”.
Cost of Holding Safety Stock
Using exchange curves, management can judge whether current service levels are
supported by sound stocking decisions. Consider a firm whose inventory investment
and service level for the same product corresponds to point A in Figure 4. What can be
said about this firm? The firm can improve its inventory management practices and
reduce inventory investments with the same level of service.
The above analysis is for a single item. A similar approach, however, can be used to
construct exchange curves for aggregate values. Such aggregate measures can be used
to compare a firm’s performance to the performance of its competitors.
Question: Can you think of other IT investments that may shift the exchange
curve downwards?
6. Conclusion
In this unit, we introduce you to some of the most important inventory management
tools for uncertain demand environments. One important use of inventory models and
tools is the insight we obtain from our analyses. The tools enable a better understanding
of the role of safety stock, the relationships between different parameters, and the
sensitivity of policy parameters to inaccuracies in data.
Now, we know how to set appropriate inventory levels, how to monitor them and
when to place replenishment orders. We can quantify and estimate the impact of not
having inventory available when needed. To avoid costly stock outs, we can choose the
best safety stock levels. Note, however, that all of our discussion so far addresses a
single firm with a manager who has a single objective maximizing firm’s stakeholders’
value by better managing inventories.
To stay competitive, firms need to improve the processes within the firm as well as
across the global supply chain. Often supply chains are not vertically integrated.
Different firms own different parts of the supply chain and these firms often have
selfish objectives. Coordination and synchronization between supply chain members is
often difficult, leading to problems such as the “bullwhip effect”. The following unit
will touch upon some of these issues.