Proposal for a New Service Using AI:
Accelerating Business with Data Analysis
Table of contents
⮚ Current State of the Consumer Finance Industry
⮚ Prediction Using Machine Learning Models
⮚ Our Business Proposal
Current State of the Consumer Finance Industry
It has become an era where customers demand the 4 S’s.
Speed Simplicity Secret Safety
Review can be Get a loan A trustworthy
Apply in just three
completed within simple steps on without your lending
a few minutes your smartphone. colleagues company
after application finding out
In order to improve profitability, the consumer finance industry
⮚ wants to provide services to more customers while managing risk.
⮚ wants to offer larger loans to customers with high creditworthiness.
Accurate risk management and identification of low-risk customers through customer
data analysis are required
Utilize AI
From large volumes of information such as customer data, it is now possible
to represent complex relationships that humans could not discover.
2
Table of contents
⮚ Current State of the Consumer Finance Industry
⮚ Prediction Using Machine Learning Models
⮚ Our Business Proposal
3
Introduction of Data
Approximately 200 types of customer information have been accumulated as
data.
A portion of the
data you provided
Data Contents
⮚ Key data for each customer (gender, occupation, family information, presence/absence of Proportion of Non-Defaulters
default, etc.)
⮚ Information on customers’ past borrowings at other financial institutions (delinquent
amounts, credit limits, etc.)
⮚ Monthly balance status of customers
⮚ Credit card usage information and withdrawal amounts for the month prior to the application
date
⮚ Key data at the time of past applications (contract type, application amount, purpose of use)
⮚ POS cash balance information
⮚ Installment payment information
n= 307506
4
Data Visualization
We conducted various visualizations of the customer data.
Type of job Non-Defaulters
Defaulters
There are many NULL (unknown) entries; next
most common is “Labor.”
No difference in credit balance is
observed between non-defaulters
and defaulters.
Credit balance
It is difficult to distinguish non-
defaulters from defaulters
using individual features alone.
Advanced analysis using machine
Age learning methods is necessary.
There is a wide range of customers from their 5
20s to their 60s.
Overview of the Machine Learning Model
From customer information such as gender, occupation, family information, and credit card usage
amounts, the proposed model predicts the probability of default.
Machine Learning
Input Output
Methods
Input data into Output Prediction
Gender Probability of Default
machine learning
LightGBM
Occupati methods
on
XGBoost
Family info
… Neural Net
Account info
…
Defined as a credit score that is easy for humans to interpret*
Credit Score
Credit Score =1- Probability of default
*When calculating the credit score considering the distribution of default probability, a square root is taken. 6
Important Features for Prediction
Many of the features important for prediction are those related to values from external sources.
Values from external sources
EXT_SOURCES_○○
Feature
s
Importan
ce
Features derived from processed data and statistical measures are important for AI predictions,
indicating that the AI takes a different approach from human decision-making.
By utilizing both AI and human decision-making, highly accurate lending decisions
become possible.
7
Table of contents
⮚ Current State of the Consumer Finance Industry
⮚ Prediction Using Machine Learning Models
⮚ Our Business Proposal
8
Our Business Proposals
Business proposals utilizing credit scores and long-term support.
Businesses Utilizing Credit Scores Long-Term Support
①Risk Management Time Series Analysis
Enables risk management for customers with a Enables lending decisions aligned with
high likelihood of default economic trends in each country
②Estimation of Low-Risk Customers Model Updates
Enables proactive approaches to “low-risk Through various analyses and training,
customers” with repayment capability further accuracy improvement is possible
Loan Application
Loan
Complete in 1
min
9
①Risk Management Using Credits Scores
By utilizing AI, fast and accurate lending becomes possible.
Overview Loan Process Workflow
Using the credit score, predict customers with a high Enter required information for the
probability of default. application online
Novelty NO
AI Screening
Decisions that were traditionally made based on
human intuition and experience can be accelerated
with higher accuracy. YES Expert
Effects YES Screening
Customer Compa NO
ny
Fast and responsive Careful responses Loan Approved Loan Not
lending. become possible for Approved
loans with a potential
risk of default.
Point
Ideally, we would like to make “loan not approved” decisions based solely on AI screening; however, because there
is a certain number of customers who can actually repay among those judged “NO” by AI, expert screening is
necessary to lend to them.
10
Validity of the Credit Score
Using the credit score, it is possible to distinguish non-defaulters and defaulters.
Difference in credit scores by non-default vs. default.
Non-Defaulters From the graph on the left, the credit score
tends to be:
Non-defaulters: Often 0.71 or higher
Defaulters Defaulters: Often below 0.71
If we define credit score ≥ 0.7→☑︎Non-defaulters
Among defaulters,
< 0.7→×︓
Prediction
Defaulter, thenNumber of Data
Prediction Total
57% can be detected
On the other hand, 44,801 cases (14% of all data) were
mistakenly predicted x;
Results
Total Therefore, classification by experts with specialized
knowledge is necessary.
11
Impact on Revenue
Reducing the customer’s default rate leads to improved profitability.
3%
Source :Home Credit annual report
Assuming: 2017
Median loan amount : 204,515 yen*
Annual number of loans: 41,608,720 people**
Difference in default rate: 3%,
annually, Apx. 255.2 billion yen can be reduced
* As shown in the graph on the left, the distribution of loan amounts is right-
skewed and the mean becomes large, so we estimate using the median
(calculated with 1 tenge = 0.25 yen).
** Annual number of loans in 2017 listed on company’s website was used 12
②Estimation of Low-Risk Customers by AI
It becomes possible to propose loan increases to customers whose probability of default is extremely
low.
Proportion of defaulters
Proportion of non-defaulters Proportion of non-defaulters
←Low-risk customers
Credit
Score
99.4% of customers judged by AI to be low-risk*1 are non-defaulters.
Of the annual number of loans, 41,608,720 people*2, if we estimate low-risk customers at 5.5% of the
total*3, then proactive approaches become possible to
Apx. 2.28million people
*1 Defined as customers with a credit score of 0.9 or higher
*2 Annual number of loans in 2017 listed on company’s *3 Calculated from the data you provided 13
website was used (17,035/307,506 = 0.055)
Long-Term Support
Through long-term consulting, lending that adapts to each country’s financial markets and economic
trends becomes possible.
Source: Partially excerpted from BTMU (China) Weekly
Economic Report 2017
The 0.6% increase in the default rate in 2017 is attributed
to the rapid growth of the consumer finance market due to
yuan depreciation in China, while the credit system for
China was insufficient.
→To account for each country’s economic trends, utilize
foreign exchange data.
Existing Data
＋ Input
Foreign Machine Lending adapted to financial markets and
economic trends becomes possible.
Exchange Data Learning
In addition, model updates can be implemented to improve the accuracy of lending decisions.
Quotation
In addition to the AI credit score, we propose long-term support.
AI Credit Score Long-Term
Support
Content Content
⮚ Risk Management ⮚ Time Series Analysis
Reduce the default rate by 57% and Lending decisions can be made in line with
prevent an estimated 25.52 billion yen economic trends.
in annual losses. ⮚ Model Updates
⮚ Estimation of Low-Risk Customers Introduce economic indicators and text
Enable approaches to 2.28 million reliable mining to improve accuracy.
customers with a non-default rate of 99.4%.
Price Price
We propose a consulting fee of
We propose a fee of
30 2.5 million
million yen
yen per month.
for providing AI.
This price corresponds to about 0.01% of the annual
revenue improvement of 255.2 billion yen from risk
management, plus the revenue from estimating low-risk
To accelerate your business, our staff specializing
customers. in Finance × AI will utilize customer
data to support you. 15