Markov Analysis
Markov Analysis
IntroductionMarkov AnalysisA technique dealing with probabilities of
future occurrences with currently known probabilities
Numerous applications in◦ Business (e.g., market share analysis),
◦ Bad debt prediction
◦ University enrollment predictions
◦ Machine breakdown prediction
Markov Analysis
Matrix of Transition Probabilities Shows the likelihood that the system will
change from one time period to the next This is the Markov Process.
◦ It enables the prediction of future states or conditions.
States and State Probabilities
States are◦ used to identify all possible conditions of a process or
system
A system can exist in only one state at a time. Examples include:◦ Working and broken states of a machine
◦ Three shops in town, with a customer able to patronize one at a time
◦ Courses in a student schedule with the student able to occupy only one class at a time
Assumptions of Markov Analysis
1. A finite number of possible states
2. Probability of changing states remains the
same over time
3. Future state predictable from previous state
and the transition matrix
4. Size and states of the system remain the same
during analysis
5. States are collectively exhaustive
◦ All possible states have been identified
6. States are mutually exclusive
◦ Only one state at a time is possible
States and State Probabilities continued
1. Identify all states.
2. Determine the probability that the system is in this state.
◦ This information is placed into a vector of state probabilities.
p(i) = vector of state probabilities for period i
= (p1, p2, p3,…,pn)
where
n = number of states
p1, p2,…,pn = P (being in state 1, 2, …, state n)
Most of the time, problems deal with more than one item!
States and State Probabilities continued
Three Departmental Stores example: 100,000 customers monthly for the 3 grocery stores
◦ State 1 = store 1 = 40,000/100,000 = 40%
◦ State 2 = store 2 = 30,000/100,000 = 30%
◦ State 3 = store 3 = 30,000/100,000 = 30%
vector of state probabilities:
p(1) = (0.4,0.3,0.3)
where
p(1) = vector of state probabilities in period 1
p1 = 0.4 = P (of a person being in store 1)
p2 = 0.3 = P (of a person being in store 2)
p3 = 0.3 = P (of a person being in store 3)
States and State Probabilities continued
Three Departmental Stores example, continued: Probabilities in the vector of states for the stores
represent market share for the first period.
In period 1, the market shares are
◦ Store 1: 40%
◦ Store 2: 30%
◦ Store 3: 30%
But, every month, customers who frequent one store have a likelihood of visiting another store.
Customers from each store have different probabilities for visiting other stores.
States and State Probabilities continued
Three Grocery Store example, continued:
Store-specific customer probabilities for visiting a
store in the next month:
Store 1: Store 2:
Return to Store 1 = 80% Visit Store 1 = 10%
Visit Store 2 = 10% Return to Store 2 = 70%
Visit Store 3 = 10% Visit Store 3 = 20%
Store 3:
Visit Store 1 = 20%
Visit Store 2 = 20%
Return to Store 3 = 60%
States and State Probabilities continued
Three Grocery Store example, continued:
Combining the starting market share with the
customer probabilities for visiting a store next period
yields the market shares in the next period:
Initial
ShareP(1) P(2) P(3)
Store 1: 40% 80% 10% 10%
Next Period: 32% 4% 4%
Store 2: 30% 10% 70% 20%
Next Period: 3% 21% 6%
Store 3: 30% 20% 20% 60%
Next Period: 6% 6% 18%
New Shares: 41% 31% 28% = 100%
Matrix of Transition Probabilities
To calculate periodic changes, it is much more convenient to use
◦ a matrix of transition probabilities.
◦ a matrix of conditional probabilities of being in a future state given a current state.
Let Pij = conditional probability of being in state j in
the future given the current state of i, P (state j at time = 1 | state i at time = 0)
For example, P12 is the probability of being in state
2 in the future given the event was in state 1 in the prior period
Matrix of Transition Probabilities continued
Let P = matrix of transition probabilities
P11 P12 P13 ***** P1n
P21 P22 P23 ***** P2n
P =
Pm1 ****** Pmn
Important:
Each row must sum to 1.
But, the columns do NOT necessarily sum to 1.
****
****
Row Sum
1
1
1
Matrix of Transition Probabilities continued
Three Grocery Stores, revisited
The previously identified transitional probabilities for each of the stores can now be put into a matrix:
0.8 0.1 0.1
P = 0.1 0.7 0.2
0.2 0.2 0.6
Row 1 interpretation:
0.8 = P11 = P (in state 1 after being in state 1)
0.1 = P12 = P (in state 2 after being in state 1)
0.1 = P13 = P (in state 3 after being in state 1)
Predicting Future Market Shares
Grocery Store example
A purpose of Markov analysis is to predict the future
Given the
1. vector of state probabilities and
2. matrix of transitional probabilities.
It is easy to find the state probabilities in the future.
This type of analysis allows the computation of the probability that a person will be at one of the grocery stores in the future.
Since this probability is equal to market share, it is possible to determine the future market shares of the grocery store.
Predicting Future Market Shares continued
Grocery Store example
When the current period is 0, finding the state
probabilities for the next period (1) can be found
using:
p(1) = p(0)P
Generally, in any period n, the state probabilities for
period n+1 can be computed as:
p(n+1) = p(n)P
Predicting Future States continued
[
]
)1(
6.2.2.2.7.1.1.1.8.
(0)
6.2.2.2.7.1.1.1.8.
[ ].4 .3 .3 (0) = state probabilitiesπ
=
úúú
û
ù
êêê
ë
é=
úúú
û
ù
êêê
ë
é=
=
p
p(1)p
p
P
P
[ ].4 .3 .3(1) =p
(.4*.8 + .3*.1 + .3*.2), (.4*.1 + .3*.7 + .3*.2),
(.4*.1 + .3*.2 + .3*.6) (1)p = [ 0.41 0.31 0.28]
Predicting Future Market Shares continued
In general,
p(n) = p(0)Pn
Therefore, the state probabilities n periods in the future can be obtained from the
current state probabilities and the matrix of transition probabilities.
Another Example of Markov Analysis: The Machine Operations
States and State Probabilities
For example: If dealing with only 1 machine, given the
fact that it is currently functioning correctly.
The vector of states can then be shown.
p(1) = (1,0)
where
p(1) = vector of states for the machine in period 1
p1 = 1 = P (being in state 1) = P (machine working)
p2 = 0 = P (being in state 2) = P (machine broken)
Markov Analysis of Machine Operations
where
P11 = 0.8 = probability of machine working this period if
working last period
P12 = 0.2 = probability of machine not working this period if
working last
P21 = 0.1 = probability of machine working
this period if not working last
P22 = 0.9 = probability machine not working this period if
not working last
0.1 0.90.8 0.2P =
Markov Analysis of Machine Operations continued
What is the probability the machine will be working next month?
p(1) = p(0)P
= (1,0) 0.1 0.90.8 0.2
= [(1)(0.8)+(0)(0.1), (1)(0.2)+(0)(0.9)]= (0.8, 0.2)
Thus, if the machine works this month, then there is • an 80% chance that it will be working next month and • a 20% chance it will be broken.
Markov Analysis of Machine Operations continued
What is the probability the machine will be working in two months?
p(2) = p(1)P
= (0.8, 0.2)
0.1 0.90.8 0.2
= [(0.8)(0.8)+(0.2)(0.1), (0.8)(0.2)+(0.2)(0.9)]= (0.66, 0.34)Thus, if the machine works next month, then in two months there is • a 66% chance that it will be working and • a 34% chance it will be broken.
Equilibrium State and Absorbing State (Only Understanding of Concepts Required/No Maths Required)
Equilibrium ConditionsEquilibrium state probabilities are the long-run
average probabilities for being in each state.
Equilibrium conditions exist if state probabilities do not change after a large number of periods.
At equilibrium, state probabilities for the next period equal the state probabilities for current period.
Equilibrium Conditions continued
One way to compute the equilibrium share of the market is to use Markov analysis for a large number of periods and see if the future amounts approach stable values.
On the next slide, the Markov analysis is repeated for 15 periods for the machine example.
By the 15th period, the share of time the machine spends working and broken is around 66% and 34%, respectively.
Machine Example: Periods to Reach Equilibrium
Period123456789101112131415
State 11.0 .8
.66 .562
.4934 .44538
.411766 .388236 .371765 .360235 .352165 .346515 .342560 .339792 .337854
0.0 .2
.34 .438
.5066 .55462
.588234 .611763 .628234 .639754 .647834 .653484 .657439 .660207 .662145
State 2
The Markov Process
(n) P (n+1)
Equilibrium Conditions
Matrix ofTransition
NewState
CurrentState
Equilibrium Equations
[ ]
[ ] [ ]
1
and 1
:or
Then:
P , (i) Assume:
)()1(
22
1212
11
2121
2221212 ,2121111
22212121211121
2221
121121
pp
pp
Therefore:PPPP
PPPP
pppp
Pii
-=-=
+=+=
++=
úûù
êëé==
=+
pppp
pppppp
pppppp
ppp
pp
Equilibrium Equations continued
It is always true thatp (next period) = p (this period) P
p (n+1) = p (n) Por
At Equilibrium:p (n+1) = p (n) = p (n) P*p (n) = p (n) P*
p = p P
Dropping the n term:
Equilibrium Equations continued
Machine Breakdown example
(p1, p2) = (p1, p2)
At Equilibrium:p = p P
Applying matrix multiplication:
0.1 0.90.8 0.2
(p1, p2) = [(p1)(0.8) + (p2)(0.1), (p1)(0.2) + (p2)(0.9)]
Multiplying through yields:p1 = 0.8 p1 + 0.1 p2p2 = 0.2 p1 + 0.9 p2
Equilibrium Equations continued
Machine Breakdown example
The state probabilities must sum to 1, therefore: S p’s = 1In this example, then:
p1 + p2 = 1
In a Markov analysis, there are always n state equilibrium equations and 1 equation of state probabilities summing to 1.
Equilibrium Equations continued
Machine Breakdown ExampleSummarizing the equilibrium equations:
p1 = 0.8 p1 + 0.1 p2p2 = 0.2 p1 + 0.9 p2
p1 + p2 = 1
Solving by simultaneous equations:p1 = 0.333333 p2 = 0.666667
Therefore, in the long-run, the machine will be functioning 33.33% of the time and broken down 66.67% of the time.
Absorbing StatesAny state that does not have a probability
of moving to another state is called an absorbing state.
If an entity is in an absorbing state now, the probability of being in an absorbing state in the future is 100%.
An example of such a process is accounts receivable.◦ Bills are either paid, delinquent, or written off as
bad debt.◦ Once paid or written off, the debt stays paid or
written off.
Absorbing StatesAccounts Receivable exampleThe possible states are
◦ Paid◦ Bad debt◦ Less than 1 month old debt◦ 1 to 3 months old debt
A transition matrix for this would look similar to:Paid Bad <1 1-3
Paid 1 0 0 0 Bad 0 1 0 0<1 0.6 0 0.2 0.2 1-3 0.4 0.1 0.3 0.2
Markov ProcessFundamental Matrix
P =0 1 0 01 0 0 0
0.6 0 0.2 0.20.4 0 0.3 0.2
I 0
A B
Partition the probability matrix into 4 quadrants to make 4 new sub-matrices:I, 0, A, and B
Markov ProcessFundamental Matrix continued
BA
0I PLet
Where I = Identity matrix,
and 0 = Null matrix
1 BIFThen
The FA indicates the probability that an amount in one of the non-absorbing states will end up in one of the absorbing states.
Once F is found, multiply by the A matrix: FA
Markov ProcessFundamental Matrix continued
Once the FA matrix is found, multiply by the M vector, which is the starting values for the non-absorbing states, MFA,
whereM = (M1, M2, M3, … Mn)
The resulting vector will indicate how many observations end up in the first non-absorbing state and the second non-absorbing state, respectively.