Forecasting Demand – Discrete Choice Models

Anton J. Kleywegt, Ph.D.

School of Industrial and Systems Engineering
Review of Previous Lesson

• We use causal models when we have data not

just about the dependent variable of interest,
but also about various factors that we think
are related with the dependent variable
• Linear regression models are simple, easy to
calibrate models that include these factors
• Factors are encoded into explanatory variables
that capture the effects of the factors on the
dependent variable using chosen basis
functions
• Categorical factors, encoded using binary
variables (indicator variables or dummy
variables) are especially useful in models
• Model parameters are estimated from data by
minimizing the sum of squared errors, or sum
of absolute errors, or similar error measures
Learning Outcomes
• Become familiar with the reasons for using discrete
choice models
• Develop skill in formulating and interpreting discrete
choice models
Forecasting Methods
• Classification of forecasting methods according to Armstrong, J.S. Long-range
Forecasting, Second Edition, 1985
Discrete Choice Models
• Suppose that the dependent variable Y of interest is discrete
• Examples
• Y = 0 if the democratic candidate wins the presidential election, and Y = 1 if the
republican candidate wins the presidential election
• Y = a if a customer chooses to buy product a, Y = b if the customer chooses to buy
product b, and Y = 0 if the customer chooses not to buy either product
• Y = c if a customer cancels an order and replaces it with a new order (changes the
order), Y = d if the customer cancels an order and asks for a refund, Y = e if the
customer cancels an order, does not replace it with a new order, and does not ask
for a refund, and Y = 0 if the customer does not cancel the order
Discrete Choice Models
• Suppose you have data on the dependent variable Y of interest, as well as different
factors Z1, Z2, …, Zm that you think affect Y
• For example, Y = a if a customer chooses to buy product a, Y = b if the customer
chooses to buy product b, and Y = 0 if the customer chooses not to buy either
product, Z1 denotes the price of product a during the week, and Z2 denotes the
price of product b during the same week
• You want to propose (based on intuition, experience, and/or trial-and-error) a relation
between the dependent variable Y of interest and chosen functions, called basis
functions or feature functions, of the factors Z1, Z2, …, Zm
• Call the basis functions
Discrete Choice Models
• Suppose one considers a linear model, as follows:
Discrete Choice Models
• Proposed linear model:

• Problem with linear model: The dependent variables take values 0 and 1 only, whereas
the linear expressions can take on any real value
Discrete Choice Models
• Partial fix for problem with linear model:

• Now the dependent variables are allowed to take on any values between 0 and 1
Discrete Choice Models
• Partially fixed linear model:

• Problem with the partially fixed linear model: The dependent variables take values
between 0 and 1 only, whereas the linear expressions can take values less than 0 or
more than 1
Discrete Choice Models
• Further fix:

• Note that this is not a linear model anymore – it is a nonlinear model, and it has lost the
nice properties of linear models
Discrete Choice Models
• There are many other nonlinear models to consider
• We can use any function G that takes values between 0 and 1 only:
Discrete Choice Models
• Popular, and simple idea: Take any function H that takes positive values only:
Discrete Choice Models
• Same idea, more generally stated:
• Set N of decision makers (potential customers)
• Each customer n has a set Sn of alternatives to choose from, including the null
alternative (for example, not to buy anything)
• There are factor values Z1,…,Zm that you think affect the dependent variables
(choice probabilities), for example Z1 denotes the price of product a during the
week, and Z2 denotes the price of product b during the same week
• For each customer n, and each alternative i in Sn, a vector of basis functions of the
chosen factors are formulated:
Discrete Choice Models
• Same idea, more generally stated (continued):
• Take any function H that takes positive values only
Discrete Choice Models
• One of the most popular choices for H:
Discrete Choice Models

• Example
• Each customer chooses product a or product b or neither (alternative 0)
• Thus, Sn = {a,b,0} for each customer n
• We want to forecast
Discrete Choice Models

• Example (continued)
• Chosen factors and basis functions:
Discrete Choice Models
• Example (continued)
• The resulting Multinomial Logit (MNL) model is
Discrete Choice Models
• Example (continued)
• Binary logit model: Binary logit model

0.9

0.8

Probability of purchase
0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 10 20 30 40 50 60 70 80 90 100
Price
Summary
• Modern demand models describe how customers
choose among available alternatives
• Linear regression models cannot model choices
among discrete alternatives
• There are various nonlinear regression models that we
can use to model customer choices
• One of the most popular and simple discrete choice
models is the multinomial logit model
• Parameter estimation for multinomial logit models is
relatively easy, and there are many software packages
that can calibrate parameters for multinomial logit
models, given good data