OPTIMIZATION OF BANK

LENDING DECISION
AN INTELLIGENT MODEL BASED ON GENETIC ALGORITHM
BY:
CHAKIT CHAWDA AND AKASH TIWARI
ABOUT THE DOMAIN:
• Credit Crunch: A scenario caused due to sustained period of careless and
inappropriate lending, resulting in losses for lending institutions and
investors in debt when the loans turn sour and the full extent of the bad
debts becomes known.
• Benefits of bank lending
decisions:
• Increasing profit
• Reduce possible risks
• Make managerial decisions
DEFINITIONS
 Default: Scenario where loans turn sour.
 Credit Rating: An estimate of the ability of a person or organization to fulfil their
financial commitments, based on previous dealings.
 Credit Limit: Maximum amount that can be lent to a customer based on his
income or/and occupation.
 Loan Size: Amount of loan requested by the customer.
 Loan Type: Purpose for which loan is requested (Mortgage, Personal, Auto etc)
 Expected Loan Loss: Proportion of the amount lent that is expected to be lost
determined by customer credit rating.
 Reserve Ratio: Bare minimum proportion of the total deposit amount to be
carried by the bank.
PROBLEM STATEMENT
• Choosing an optimum bank lending decisions that maximise the
bank profit in a credit crunch environment.
• Creation of a GA model that facilitates how banks would make an
efficient decision in case of cut back on lending supply(credit
crunch environment) and find out suitable customers.
• The GA model developed will be a constrained genetic algorithm
problem.
• In a credit crunch environment where all customers are eligible to get
the desired loan:
• This is a NP-hard optimization problem which can be solved using meta-
heuristic algorithms
• Evolutionary algorithm is a meta – heuristic.
 THE CONSTRAINT
• The bank obtains requests for loans from customers with different credit ratings for
various loan types, loan age and loan size.
• The bank has a maximum deposit of D, and has a reserve ratio of K for safety reasons,
meaning, the bank can lend a maximum of D*(1-K).
• The bank wants to maximise it’s benefit by making the right lending decision, while
meeting the constraint of the maximum lending of ’D*(1-K)‘ criteria.

Example:
D = $1,000,000
K = 0.15

The bank has a maximum deposit of $1,000,000

The bank can lend a total maximum of (1-0.15) * 1000000 = $ 850,000
and maintains a reserve of $150,000
TRADITIONAL TECHNIQUES USED:
• Linear and logistic regression
• Linear and quadratic programming
• Discriminant analysis
• Decision trees
DRAWBACKS
• Etc. OF THE TRADITIONAL
TECHNIQUES:
• Complexity and time consumption
• Objective and constraint functions of bank profit maximisation are nonlinear and sometimes non-
convex which can be costly to evaluate using traditional methods.

BENEFITS OF USING EVOLUTIONARY ALGORITHM

• Highly capable of extracting meaning from imprecise data.
• Capable of detecting trends that are too complex to be discovered by humans or conventional
techniques.
 THE GA
Customer Characteristics
LoanAge
Loan Age αα
MODEL
Loan
LoanAgeAge
Age α
Loan
Loan Age
Loan Size ααα L
LoanSize
Loan Size LL
Loan
Loan Size
Size L Lending Decision Chromosome
Loan Size
Loan
Loan Type LL φφ
Type
LoanType
Loan Type φφ
Loan
Loan Type
TypeRating
Credit Rating φφ
Credit
Credit Rating
CreditRating
Credit Rating
Credit Rating
Credit Limit
CreditLimit
Limit
Credit
CreditLimit
Limit GA operations: Mutation,
Credit
Credit Limit Crossover, Selection

Fitness Function Calculation

Bank Variables
Loan Interest Rate rL Intermediate Solutions
Expected Loan Loss λ
Deposit Rate rT
Feasibility Test
Reserve Ratio K
Transaction Cost T
Near - Optimum Lending Decision
1 0 0 1 1 0 0 1 1 1
 LENDING DECISION CHROMOSOME REPRESENTATION

Customer ID’s
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
1 0 0 1 0 0 1 0 0 0 0 1 0 1 1 1 0 0 0 0 0 0 0 1 0

Loan not lent to customer 8 Loan lent to customer 16

Boolean Value

Decision Chromosome
Customer Characteristics
ID Name Loan Age(α) Loan Size Credit Rating Loan Type(φ)
1 Kothari 10 Large A Mortgage
2 Dakhare 3 Medium BB Personal
3 Dubey 15 Large AA Auto
… … … … … …
… … … … … …
MODEL PARAMETER DEFINITIONS:
ϑ= Loan Revenue
n= Number of customers
= Loan interest rate
λ= Expected Loan Loss
L= Loan Amount
μ= Loan Cost
δ= Unallocated overheadCost fraction of the loan
ϖ = Total Transaction cost
= Customer Transaction Rate
T = Available deposit amount – lent amount
K = Required Reserve Ratio
D = Financial institutions deposit
β = Demand Deposit Cost
= Weighted average of all different deposit rates based on deposit type = 0.9% (assumption)
CUSTOMER CHARACTERISTICS CALCULATIONS
Loan Age Category Loan Size Category
Category (α) Value Category (L) Value
1 1 ≤ α ≤  3 Micro $ 0 ≤ L ≤ $ 13,000
2 3 < α ≤  5 Small $ 13,001 < L ≤ $ 50,000
3 5 < α ≤  10 Medium $ 50,001< L ≤ $ 100,000
4 10 < α ≤  20 Large $ 100,001 < L ≤ $ 250,000

Loan Interest Rate assignment Customer Credit Rating – Expected Loan loss proportion
φ α rL Value Credit Rating λ Value
1 NA
AAA 0.0002 ≤ λ ≤ 0.0003
2 NA
Mortgage (M) 3 NA AA 0.0003 < λ ≤ 0.001
4 0.021 ≤ rL ≤ 0.028 A 0.001 < λ ≤ 0.0024
1 0.0599 ≤ rL ≤ 0.0601 BBB 0.0024 < λ ≤ 0.0058
Personal (P) 2 0.0601 < rL ≤ 0.0604 BB 0.0058 < λ ≤ 0.0119
3 0.0604 < rL ≤ 0.0609

1 0.0339 ≤ rL ≤ 0.03349

Auto (A) 2 0.0349 < rL ≤ 0.0379

3 0.0379 < rL ≤ 0.0399

 DEVELOPING THE FITNESS FUNCTION
Fitness function = function of (loan revenue(ϑ) , loan cost(μ) , Total transaction cost(ϖ), cost of demand
deposit(β ))
Fitness function = F(ϑ ,ϖ ,β ,μ)

Loan revenue (ϑ): Total Transaction Cost(ϖ):

ϖ=
ϑ=
Cost of Demand Deposit(β):
Loan Cost (μ): β= *D
μ=

Fitness Function

F = ϑ +ϖ -β -μ
GENETIC ALGORITHM PARAMETERS
Objective Determine a bank lending decision that maximizes
the bank profit.

Determine a bank lending decision that minimizes

the crediting cost.

Raw Fitness Fx=ϑ +ϖ -β -μ

Population Size (n) = 60
GA Generations (ℜ) = 60
Crossover ratio = 0.8
Mutation ratio = 0.006
Parameters Basic selection method is spanning the weighted
roulette
wheel selection six times.
Stopping Criteria The evolution continues until:
Fx(t) = ℜ where t is the time interval
DEMONSTRATING THE ALGORITHM
Assume 10 lending decisions to be made. We have the details of each lending decision, i.e. loan size, loan
type, loan age, expected loan loss, customer credit rating.

1. Initializing a feasible population of size 6.

2. Perform Roulette wheel selection to select copies of the chromosomes. This will be our parent set.
3. Perform single point crossover.
4. Perform the swap mutation operation.
5. Check for feasibility of the off-springs. Remove the infeasible solutions.
6. Merge the parent and off-spring set and choose 6 best chromosomes (based on fitness) to be carried
over to the next generation.
7. Stop if termination criterion is satisfied, else go to step 3.
CUSTOMER CHARACTERISTICS (FOR THE DEMONSTRATION)
Customer ID 1 2 3 4 5 6 7 8 9 10
Loan Size 10 25 4 11 18 3 17 15 9 10
Interest 0.021 0.022 0.021 0.027 0.025 0.026 0.023 0.021 0.028 0.022
Rating AAA BB A AA BBB AAA BB AAA A A
Loss (λ) 0.0002 0.0058 0.0001 0.0003 0.0024 0.0002 0.0058 0.0002 0.001 0.001

INITIALISING FEASIBLE POPULATION

Gene 1 2 3 4 5 6 7 8 9 10
Ch1 0 1 1 0 0 1 0 0 1 0
Ch2 0 0 0 1 0 0 0 0 0 1
Ch3 0 0 0 0 1 1 0 1 0 1
Ch4 1 0 1 0 0 0 1 0 0 1
Ch5 1 0 1 0 0 1 0 1 1 0
Ch6 0 1 0 0 0 0 0 1 0 0
FEASIBILITY CALCULATION DEMONSTRATION
Maximum Bank Deposit (D) = $60
Reserve Ratio (K) = 0.15
Maximum Loan Amount that can be lent = (1-K)*D = 0.85*60 = $51
Þ Total Loan amount must not exceed $51
Demonstrating this for Chromosome 1:
Ch1 0 1 1 0 0 1 0 0 1 0

Loan Size 10 25 4 11 18 3 17 15 9 10

Amount 0 25 4 0 0 3 0 0 9 0

Total = 25 + 4 + 3 + 9 = 41 < 51
GENERATION 1 : THE
BEGINNING
MATING POOL

1 2 3 4 5 6 7 8 9 10 Fitness Fitness/avg. fitness no. of Mating pool

value copies

0 1 1 0 0 1 0 0 1 0 0.3244 0.96 1 0110010010

0 0 0 1 0 0 0 0 0 1 0.2427 0.72 0 0000110101

0 0 0 0 1 1 0 1 0 1 0.4702 1.39 2 0000110101

1 0 1 0 0 0 1 0 0 1 0.3094 0.91 1 1010001001

1 0 1 0 0 1 0 1 1 0 0.4394 1.29 2 1010010110

0 1 0 0 0 0 0 1 0 0 0.2470 0.73 0 1010010110

Average fitness = 0.3388

CROSSOVER AND SWAP MUTATION

MATING POOL RANDOM NUMBER OFFSPRING TWO RANDOM OFFSPRING

[1,10] NUMBERS [1,10]

0110010010 1 0110010010 2,6 0110010010

0000110101 0000110101 2,9 0010110111

0000110101 5 1010010101 1,9 0010010111

1010001001 0000101001 4,5 0001001001

1010010110 9 1010010110 3,8 1010010110

1010010110 1010010110 7,6 1010001110

 FEASIBLITY OFFSPRING
OFFSPRING TOTAL LOAN AVAILABLE FEASIBILITY FEASIBLE
AMOUNT OFFSPRINGS

0110010010 41 51 YES 0110010010

0010110111 59 51 NO _

0010010111 41 51 YES 0010010111

0001001001 38 51 YES 0001001001

1010010110 41 51 YES 1010010110

1010001110 55 51 NO _
 PARENT + FEASIBLE OFFSPRINGS
SR. NO CHROMOSOME FITNESS SORTED FITNESS CHROMOSOME NO.

1 0110010010 0.3244 0.4702 2

2 0000110101 0.4702 0.4702 3
3 0000110101 0.4702 0.4414 8
4 1010001001 0.3094 0.4394 5
5 1010010110 0.4394 0.4394 6
6 1010010110 0.4394 0.4394 10
7 0110010010 0.3244 0.3481 9
8 0010010111 0.4414 0.3244 7
9 0001001001 0.3481 0.3244 1
10 1010010110 0.4394 0.3094 4

The highlighted chromosomes are selected for the next generation.

 FITNESS VALUES THROUGH GENERATIONS
Generation = 0 Generation =1
Chromosome fitness Chromosome fitness
0110010010 0.3244 0000110101 0.4702
0001000001 0.2427 0000110101 0.4702
0000110101 0.4702 0010010111 0.4414
1010001001 0.3094 1010010110 0.4394
1010010110 0.4394 1010010110 0.4394
0100000100 0.2470 1010010110 0.4394

ITERATE FOR MORE GENERATION TO IMPROVE

THE FITNESS
GENERATION 1: THE CONCLUSION

USE THE NEWLY GENERATED CHROMOSOMES TO

IDENTIFY THE NEW MATING POOL FOLLOWED BY SAME
STEPS.
REFERENCE

• Genetic algorithm based model for optimizing bank lending decisions [Journal : Expert Systems with
Application/ ElSEVIER]
--------Noura Metawa, M. Kabir Hassan, Mohamed Elhoseny