Module 4: More
Realistic Models of
Demand
Prof. Chelsea C. White III
[Link]@[Link]
Introduction
We have been assuming we know 𝑃 𝑑 𝑡 + 1 =
𝑑, ∀ 𝑑.
Unlikely-
1. Seasonal effects
2. Underlying economic conditions vary
3. Product desirability varies
4. May only have access to sales (censored demand)
5. Regulatory or approval changes
Accurate estimates of demand difficult, especially if
lead times are long
Build-to-order?
⇒ desirability of build-to-order, which has
downsides:
1. Difficult to take economies of scale into
consideration
2. Can affect product design (postponement)
3. Requires agile supply chain
Probability-based Models
Two levels of understanding (for probability-based models of
demand)
1. What is the functional form of the cumulative distribution
function?
2. What are the parameter values of a given CDF?
REMINDER: Let 𝐷 be a random variable (e.g., 𝐷 = demand
over the next week).
Then, the CDF of 𝐷 is
𝑃 𝐷≤𝛼 ∀ 𝛼.
For discrete random variables
𝑃 𝐷 = 𝑑 ∀ 𝑑,
where 𝑃 𝐷 = 𝑑 ≥ 0 ∀ 𝑑,
𝑃 𝐷 = 𝑑 = 1.
𝑑
Example
i. Can we assume the random variable is, say, normally
distributed? 𝑁(𝜇, 𝜎)
ii. If so, what’s 𝜇 and 𝜎?
What to do if 𝑃(𝑑) isn’t known and/or i.i.d. isn’t a
reasonable assumption?
What do we know?
i. Assumptions about aspects of what affects
demand; e.g.,
a) How the general economy is trending
b) How the general economy impacts demand
ii. Data
iii. Some combination
Consider
A very simple (primitive) model of data-driven learning
(Bayesian):
Assume, besides the demand process {𝑑 𝑡 , 𝑡 = 1,2, … } ,
there is:
i. Modulation process {𝜇 𝑡 , 𝑡 = 0,1, … }
ii. Information process {𝑧 𝑡 , 𝑡 = 1,2, … }
What do they model?
1. Modulation process models underlying processes
(e.g., state of the economy).
2. Information process provides information about the
modulation process in addition to the demand
process (interest rates, unemployment rate, etc.).
Consider
Assume (at least initially) we know:
𝑃 𝑑 𝑡 + 1 ,𝑧 𝑡 + 1 ,𝜇 𝑡 + 1 𝜇 𝑡 , (∗)
a probability we have no control over (at least
initially).
Note, ∗ =
𝑃 𝑑 𝑡 + 1 ,𝑧 𝑡 + 1 𝜇 𝑡 + 1 ,𝜇 𝑡 × 𝑃(𝜇 𝑡 + 1 ∣ 𝜇 𝑡 )
𝑃(𝜇 𝑡 + 1 ∣ 𝜇 𝑡 ): the dynamics of, say, the general
economy. May be reasonably well understood for our
purposes – certainly well studied
𝑃 𝑑 𝑡 + 1 ,𝑧 𝑡 + 1 𝜇 𝑡 + 1 ,𝜇 𝑡 = (∗∗)
Consider
Likely, ∗∗ = 𝑃 𝑑 𝑡 + 1 𝜇 𝑡 𝑃 𝑧 𝑡 + 1 𝜇 𝑡
1. 𝑃 𝑧 𝑡 + 1 𝜇 𝑡 may be well understood/ studied.
2. 𝑃 𝑑 𝑡 + 1 𝜇 𝑡 is a product-specific. For a new
product, may not be understood at all. Assuming we know
this probability may be the weak link of the assumption
we know (∗).
Comments:
Suggests a need for certain parts of (∗) to be Bayesian-based
and other parts to perhaps be non-parametric, totally data-
based.
Consider
What does the literature say about this model?
Two general extensions away from 𝑃(𝑑) being known and
i.i.d.:
1. Where {𝜇(𝑡)} is not observed by {𝑧 𝑡 }; i.e., where
𝑃 𝑧 𝑡 + 1 𝜇 𝑡 = 𝑃(𝑧 𝑡 + 1 )
The unobserved case. Thus, the modulation process is
observed only through demand.
2. Where {𝜇 𝑡 } is completely observed by {𝑧 𝑡 }; i.e.,
where 𝑃 𝑧 𝑡 + 1 𝜇 𝑡 = 1 if 𝑧 𝑡 + 1 = 𝜇(𝑡) (usually
assuming 𝑃 𝜇 𝑡 + 1 𝜇 𝑡 = I, the identity matrix).
Thus, (∗) offer a unifying approach to both general
extensions.
State
How to model? Assume we know:
1. 𝑠 𝑡 + 1 = 𝑓(𝑦 𝑡 , 𝑑 𝑡 + 1 )
2. 𝑃(𝑑 𝑡 + 1 , 𝑧 𝑡 + 1 , 𝜇 𝑡 + 1 ∣ 𝜇 𝑡 )
What is now the “state”?
More than just {𝑠 𝑡 }; also includes {𝜇 𝑡 }.
But {𝜇 𝑡 } is in general partially observed
Information pattern: In selecting 𝑎(𝑡), the DM knows:
ℐ 𝑡 = {𝑠 𝑡 , 𝑠 𝑡 − 1 , … , 𝑎 𝑡 − 1 , 𝑎 𝑡 − 2 , … ,
𝑑 𝑡 , 𝑑 𝑡 − 1 … , 𝑧 𝑡 , 𝑧 𝑡 − 1 , … , 𝑥 0 }, Perfect memory!
where 𝑥 0 = {𝑥𝑖 0 }, 𝑥𝑖 0 = 𝑃(𝜇 0 = 𝑖).
A special case of the partially observed MDP (POMDP) –
POMDP
HISTORY OF THE POMDP:
What if “state” is only partially observed (can’t accurately count
inventory; can’t w.p.1 determine a patient’s cause of symptoms)
⇒ a generalization of the standard (state is completely observed)
MDP
Good news:
i. Robust model
ii. A sufficient statistic exists
iii. Some computationally useful structure
Bad news:
i. Still a computational challenge
ii. Many structural results for the MDP don’t extend to the
POMDP (but we will find some).
Sufficient Statistic
The sufficient statistic:
𝑠 𝑡 ,𝜇 𝑡 → 𝑠 𝑡 ,𝑥 𝑡 ,
where:
Belief Function
𝑥 𝑡 = {𝑥𝑖 𝑡 }
𝑥𝑖 𝑡 = 𝑃 𝜇 𝑡 = 𝑖 ℐ 𝑡 .
Nomenclature: Let
𝑃𝑖𝑗 𝑑, 𝑧 = 𝑃(𝑑 𝑡 + 1 = 𝑑, 𝑧 𝑡 + 1 = 𝑧, 𝜇 𝑡 + 1 = 𝑗 ∣
𝜇 𝑡 = 𝑖).