You are on page 1of 24

DECISION TREES: SOLUTION

Statistics for Business 1

Session Pre-requisites
Go through our earlier
videos
Bayes Rule
Decision Trees

In preparation

Please have your computer open, and follow along


Maximum benefit guaranteed!

Strategy

Payoff ($) Probabilit


y
-5
0.3
3
0.5
10
0.2

Playing the averages


If you played this game again and again, how much
would you make on average?

EMV

Payoff ($) Probabilit


y
-5
0.3
3
0.5
10
0.2
If we played this game a 100 times, we would
make 200 dollar

Playing the averages

If you played this game again and again, how much


would you make on average?

Average payoff = -5 x 0.3 + 3 x 0.5 + 10 x 0.2 = 2


dollars

EMV

Payoff ($) Probabilit


y
-5
0.3
3
0.5
Playing the averages
10 how much
0.2
If you played this game again and again,
would you make on average?

There may be streaks of losses/gains, but things


smooth out
EMV = -5 x 0.3 + 3 x 0.5 + 10 x 0.2 = 2 dollars
E.g. If we played this game n times, we would make (n
x EMV)

Simplify

EMV = p1 x r1 + p2 x r2 + p3 x r3

Simplify

EMV = p1 x r1 + p2 x r2 + p3 x r3
On a decision tree, p1, p2 & p3 are all conditional
probabilities!

Context

TrendyBank or TBank is a small retail bank in


Bangalore
Key revenue channel: Interest on loans
Venkat, President of TBank

TBank decides who to make out loans to, based on


high scores from a credit-rating service
Possible disconnect between scores and default rates

Borrowers

Level 1
Borrowers make under Rs.100,000 per month
Or less than Rs.1.2 million per annum in salary

Level 2
Borrowers make over Rs.100,000 per month
Or more than Rs.1.2 million per annum in salary

Full tree

Numbers in red are negative


Additional costs or losses

Modified tree

When credit scores are not considered, what


decision path(s) must the bank choose?

EMV

EMV = 0.753 x 8000 0.247 x 5000


Considering the earnings from multiple borrowers at
this node, the bank makes an average of $4789 per
borrower

Choice

We have decided to choose the path that


maximises the EMV. Alternatively, we could
adopt these strategies
Maximin - maximise the minimum payoff at the node
Maximax - maximise the maximum payoff at the node
Minimax regret minimise the maximum regret at
the node

EMV

Depending on the borrowers salary, there are two


paths
If the loan is approved, how much will the bank make
on average?

Folding back

Start from the rightmost nodes and move towards


the left
1. Calculate the EMV for the Paid node
2. Choose the path that maximises EMV at the Level
node

EMVs

Level 1

Level 2

Overall

Software

When the decision tree becomes large and


complex, software packages come to the rescue
PrecisionTree plugin for Microsoft Excel
http://www.palisade.com/precisiontree/

Sensitivity

The conclusions of a decision tree problem depend


on the parameters holding steady. Rarely the
case!
E.g. Interest rates are dictated by the reserve banks

Consider the following scenarios


Interest rates on Treasury bills fall from 5% to 4%
Interest rates on Treasury bills rise from 5% to 7%
The bank lowers its Loan interest rate from 8% to 7.5%
The bank hikes it Loan interest rate from 8% to 9%

Benchmarks

Expected Value of Sample Information


Suppose that the credit scores were available for free
EVSI = EMV of the decision tree with this (free) sample
information minus EMV without this information

Perfect information is information from a


magical test that will always predict the correct
behaviour of the borrower
EVPI = EMV with (free) perfect information minus EMV
without it

Thanks

Picture of the Team

Thanks for patiently going through this course


I look forward to your continued participation in the SFB
series!

You might also like