L.L. Bean, Inc.
Item Forecasting and Inventory Management
Teaching Note
Synopsis
L.L. Bean must make stocking decisions on thousands of items sold through its catalogs. In
many cases, orders must be placed with vendors twelve or more weeks before a catalog lands on a
customer's doorstep, and commitments cannot be changed thereafter. As a result, L.L. Bean suffers
annual losses of over $20 million due to stockouts or liquidations of excess inventory.
The case deals with the implementation of standard critical-fractile methodology as part of
the ongoing operations of a company. This case has no number crunching, but can serve as a good
vehicle for discussing the gain in stocking an incremental unit that is demanded, the loss in failing to
do so, the effects of substitution and complementarity on inventory decisions, and the practicalities of
probability assessment of demand distributions. Treating the ratio of actual demand to forecast
demand as indistinguishable across items within a class is an important way of relating past data to a
probability distribution of future demand.
Assignment Questions
1. How does L.L Bean use past demand data and a specific item forecast to decide
how many units of that item to stock?
2. What item costs and revenues are relevant to the decision of how many units of
that item to stock?
3. What information should Scott Sklar have available to help him arrive at a
demand forecast for a particular style of men's shirt that is a new catalog item?
4. How would you address Mark Fasold's concern that the number of items
purchased usually exceeds the number forecast?
5. What should L.L. Bean do to improve its forecasting process?
1
, 895-057 L.L. Bean, Inc.
Analysis
The following issues are critical to successful implementation of item forecasting under
uncertainty at L.L. Bean:
• Demand vs. sales data;
• Deriving probability distributions of demand from relative-frequency
distributions of A/F;
• Computing appropriate values of G and L;
• Buyers' point forecasts;
• Behavioral considerations.
These are addressed in what follows.
Demand vs. Sales Data
The case says that true demand is known: if a customer calls or writes demanding Product X,
that action is recorded as a demand. Well, not exactly. Until recently, a telephone order for an out-
of-stock item was not recorded as a demand. Even though the telemarketer to whom the request
comes must enter the code for the product in order to receive an on-line status report on its
availability, the MIS system did not capture the code unless an order is placed. Most of that is now
being fixed. But what about substitutes? For instance, a customer asks for a green turtleneck, which
is out of stock, and accepts a purple one instead. The MIS system could capture the order for the
green turtleneck as a demand, and the purple as a substitute, but it doesn't.
How serious is failure to record a demand for an out-of-stock item? If LLB simply keeps
track of the fact that the item stocked out, i.e., the demand data are censored, not truncated, then
under some circumstances the censoring may not be too serious. Suppose, for example, that the
economics dictated that the stocking level for all items should be at the .75 fractile of their demand
distributions. Then 25% of all items should stock out. If 25% do stock out, then we can assume that
actual demand for those items exceeded the .75 fractile of their distribution of demand. Then,
retrospectively, we will have A/F ratios up to the .75 fractile of the frequency distribution of A/F,
thus will know the .75 fractile of this distribution, and can apply it to stocking decisions for next year.
If, however, the overage and underage costs of different items dictate different critical ratios, as is the
case at LLB, censoring will impact the frequency distribution of A/F in a way that will impair its
usefulness for next year. If an item with a critical ratio of .6 stocked out, for example, we cannot use
the censored data to find the .75 fractile of the A/F distribution
Even in this case, all is not necessarily lost. If, after a certain number of weeks into the
season, before any stockouts occur, the "percent done" for an item (i.e., the percent of total demand
that is incurred through the week in question) is known with near certainty, the total demand can be
projected with reasonable accuracy.
More insidiously, what happens if a customer has a list of five items, learns that the first item
on the list is out of stock, and cancels the whole order? There's no way of capturing demand for the
other four items. Nor any easy way of knowing whether, and how often, this is occurring.
Deriving Probability Distributions of Demand
The theory of A/F is that if past point forecasts take into account the effects of those factors
that are known to affect demand, what is left is indistinguishable "noise". The distribution of this
"noise" is applicable to point forecasts of future demand. If the critical ratio for an item is .75, if the
.75 fractile of the distribution of A/F is 1.3, and if the point forecast for a particular item is F = 1,000
units, then 1,300 units should be ordered.
2
Item Forecasting and Inventory Management
Teaching Note
Synopsis
L.L. Bean must make stocking decisions on thousands of items sold through its catalogs. In
many cases, orders must be placed with vendors twelve or more weeks before a catalog lands on a
customer's doorstep, and commitments cannot be changed thereafter. As a result, L.L. Bean suffers
annual losses of over $20 million due to stockouts or liquidations of excess inventory.
The case deals with the implementation of standard critical-fractile methodology as part of
the ongoing operations of a company. This case has no number crunching, but can serve as a good
vehicle for discussing the gain in stocking an incremental unit that is demanded, the loss in failing to
do so, the effects of substitution and complementarity on inventory decisions, and the practicalities of
probability assessment of demand distributions. Treating the ratio of actual demand to forecast
demand as indistinguishable across items within a class is an important way of relating past data to a
probability distribution of future demand.
Assignment Questions
1. How does L.L Bean use past demand data and a specific item forecast to decide
how many units of that item to stock?
2. What item costs and revenues are relevant to the decision of how many units of
that item to stock?
3. What information should Scott Sklar have available to help him arrive at a
demand forecast for a particular style of men's shirt that is a new catalog item?
4. How would you address Mark Fasold's concern that the number of items
purchased usually exceeds the number forecast?
5. What should L.L. Bean do to improve its forecasting process?
1
, 895-057 L.L. Bean, Inc.
Analysis
The following issues are critical to successful implementation of item forecasting under
uncertainty at L.L. Bean:
• Demand vs. sales data;
• Deriving probability distributions of demand from relative-frequency
distributions of A/F;
• Computing appropriate values of G and L;
• Buyers' point forecasts;
• Behavioral considerations.
These are addressed in what follows.
Demand vs. Sales Data
The case says that true demand is known: if a customer calls or writes demanding Product X,
that action is recorded as a demand. Well, not exactly. Until recently, a telephone order for an out-
of-stock item was not recorded as a demand. Even though the telemarketer to whom the request
comes must enter the code for the product in order to receive an on-line status report on its
availability, the MIS system did not capture the code unless an order is placed. Most of that is now
being fixed. But what about substitutes? For instance, a customer asks for a green turtleneck, which
is out of stock, and accepts a purple one instead. The MIS system could capture the order for the
green turtleneck as a demand, and the purple as a substitute, but it doesn't.
How serious is failure to record a demand for an out-of-stock item? If LLB simply keeps
track of the fact that the item stocked out, i.e., the demand data are censored, not truncated, then
under some circumstances the censoring may not be too serious. Suppose, for example, that the
economics dictated that the stocking level for all items should be at the .75 fractile of their demand
distributions. Then 25% of all items should stock out. If 25% do stock out, then we can assume that
actual demand for those items exceeded the .75 fractile of their distribution of demand. Then,
retrospectively, we will have A/F ratios up to the .75 fractile of the frequency distribution of A/F,
thus will know the .75 fractile of this distribution, and can apply it to stocking decisions for next year.
If, however, the overage and underage costs of different items dictate different critical ratios, as is the
case at LLB, censoring will impact the frequency distribution of A/F in a way that will impair its
usefulness for next year. If an item with a critical ratio of .6 stocked out, for example, we cannot use
the censored data to find the .75 fractile of the A/F distribution
Even in this case, all is not necessarily lost. If, after a certain number of weeks into the
season, before any stockouts occur, the "percent done" for an item (i.e., the percent of total demand
that is incurred through the week in question) is known with near certainty, the total demand can be
projected with reasonable accuracy.
More insidiously, what happens if a customer has a list of five items, learns that the first item
on the list is out of stock, and cancels the whole order? There's no way of capturing demand for the
other four items. Nor any easy way of knowing whether, and how often, this is occurring.
Deriving Probability Distributions of Demand
The theory of A/F is that if past point forecasts take into account the effects of those factors
that are known to affect demand, what is left is indistinguishable "noise". The distribution of this
"noise" is applicable to point forecasts of future demand. If the critical ratio for an item is .75, if the
.75 fractile of the distribution of A/F is 1.3, and if the point forecast for a particular item is F = 1,000
units, then 1,300 units should be ordered.
2
