Mercury

An Intelligent Banking Risk-Management System

Based on Fintech

Southwestern University of Finance and Economics

July 23rd. 2021
Bank of SWUFE
The Credit Department …
Mercury
 Current Situation of the Loan Market
- Pain points of traditional banks

• Complex and time-consuming credit investigation

• Adverse selection of loan interest rates and terms

• Hard to detect risk connections based on industry and geography

• Difficult to track company credits after the loan origination

 Product Functions Risk associations identification

Loan interest rates and terms

suggestions based on risks

Comprehensive real-time
loan risk monitoring

Dynamic track of risk changes

How Mercury Works? Data Resources

Web
Web
Crawling
Server

Data
(API) Purchase
Client DB
Middleware Server
Side Server
User
Behaviors
Model Server
··· ···
How Mercury Works? Data Resources

Web
Web
Crawling
Server

Data
(API) Purchase
Client DB
Middleware Server
Side Server
User
Behaviors
Model Server
··· ···
 Algorithm Logic
- Learn from the Past

137 15 230k+ 17%

Attributes Sources Records Labels

 Algorithm Logic
- Feature Engineering

Quick Ratio Asset Rank

Employee Growth
Asset / Liability Government Subsidy
Employee Welfare
Income Growth Government Honors

Asset / Liability Ratio

✘
✔
Income Growth Rate
✘
✔ Risk
Normal
Risk
Government Subsidy
✘
✔
 Algorithm Logic
- Problem 1: Feature Interaction

Income vs. Subsidy Income vs. Tax

✔ ✔

Risk
Normal Risk
Normal
✔ ✔
 Algorithm Logic
- Solution: Tensor Multiplication
𝒲 h𝑖
𝑥𝑎𝑐𝑐

[ 1: 𝑘]
𝒯
+ [ ||
 Algorithm Logic
- Problem 2: Risk Propagation
 Algorithm Logic
- Problem 2: Risk Propagation

- Graph Attention Network

 Algorithm Logic
- Graph Attention Network: Relation Inferring

h𝑖

e 𝑗, 𝑖 [||
hj
 Algorithm Logic
- Graph Attention Network: Information Propagation

h0
e 0,𝑖
h1
e 1,𝑖
s𝑖
¿𝜎 ¿
h2
e 2,𝑖
 Algorithm Logic
- Graph Attention Network
 Algorithm Logic
- Multi-Relation Network

𝐼𝑛𝑑𝑢𝑠𝑡𝑦
𝑠 𝑖

𝑅𝑒𝑔𝑖𝑜𝑛
𝑠 𝑖

𝑇𝑦𝑝𝑒
𝑠 𝑖
 Algorithm Logic
- Final Classifier

MLP Risk
Risk
 Algorithm Logic
- Workflow: Data Preprocess

Data Missing Values

Preprocess
Outliers

Normalization
 Algorithm Logic
- Workflow: Feature Engineering

Feature
Engineering
 Algorithm Logic
- Workflow: Ensamble

Risk
Risk
 Algorithm Logic
- Workflow: Overview

Data Missing Values Feature

Preprocess Engineering

Outliers

Normalization

Risk
Risk
 Algorithm Logic
- Comparison

Feature Engineering Models

0.5 0.53
0.51
0.45 0.49
0.47
0.4 0.45
0.43
0.35 0.41
0.39
0.3
0.37
0.35
0.25 SVM Random Xgboost Proposed Ensambled
Before After Forest Method

avg. F1 score avg. F1 score

How Mercury Works? Data Resources

Web
Web
Crawling
Server

Data
(API)
Client DB Purchase
Middleware Server
Side Server
User
Behaviors
Model Server
··· ···
 Front-end Logic

Financial Institution
Risk Management

Risk State of
Log in Dashboard Search Page
Corporates

Risk Proportion
&
Industry Risk Level
 Risk State of
Corporates
Financial Institution
Risk Management

Corporate Risk
Based on Industry
and Category
 Front-end Logic

Financial Institution
Risk Management

Risk State of
Log in Dashboard Search Page
Corporates

Corporate Risk
Based on Industry
and Category
 Front-end Logic

Corporate List
Corporate
Search Page Information
Page
Search Function
 Reset Go

Search Function

Corporate List
Front-end Logic
Financial Risk
Trends

Corporate Financial Risk Risk

Information Index Measurement
Page

Risk-Spreading
Path
Financial Risk Financial Risk
Trends Index

Risk-Spreading
Path
 Front-end Logic
Loan Interest Rate

Loan term Loan

Risk
Expert Decisions Approval
Measurement
Result
Loan Limit

Loan Discount
 Future Prospects

• A timely warning system based on richer time-series data

• A more scientific and accurate suggestion via

external audit data, etc.

• More abundant sources of data

 Thanks
Southwestern University of Finance and Economics