Ứng dụng trí tuệ nhân tạo (AI) trong quản lý rủi ro tín dụng tại Ngân hàng Nông nghiệp và Phát triển Nông thôn Việt Nam (Agribank)

MINISTRY OF EDUCATION STATE BANK OF VIETNAM

AND TRAINING

------------------

NGUYEN TIEN HUNG

SUMMARY OF DOCTORAL THESIS

APPLICATION OF ARTIFICIAL INTELLIGENCE IN CREDIT RISK

MANAGEMENT AT VIETNAM BANK FOR AGRICULTURE AND RURAL DEVELOPMENT

HANOI – 2022

MINISTRY OF EDUCATION STATE BANK OF VIETNAM

AND TRAINING

------------------

NGUYEN TIEN HUNG

SUMMARY OF DOCTORAL THESIS

APPLICATION OF ARTIFICIAL INTELLIGENCE IN CREDIT RISK

MANAGEMENT AT VIETNAM BANK FOR AGRICULTURE AND RURAL DEVELOPMENT

Scientific supervisor: Supervisor 1: Dr. Bui Tin Nghi Supervisor 2: Assoc. Prof. Dr. Nguyen Duc Trung

Specialization: Finance - Banking Code: 9340201

HANOI – 2022

TABLE OF CONTENTS

GENERAL INTRODUCTION .................................................................. 1

CHAPTER 1: LITERATURE REVIEW OF ARTIFICIAL INTELLIGENCE IN CREDIT RISK MANAGEMENT ............................ 4 1.1. Researches on credit risk management models ......................................... 4 1.2. Researches on credit risk assessment ......................................................... 4

1.2.1. Researches on measurement of probability of default ..................... 5

1.2.2. Researches on loss given default (LGD). ......................................... 5

1.2.3. Researches on default risk ................................................................ 6

1.3. Researches on artificial intelligence model in credit risk management. ... 6 1.4. Research gap .............................................................................................. 8 CHAPTER 2: THEORETICAL BASIS FOR APPLICATION OF ARTIFICIAL INTELLIGENCE IN CREDIT RISK MANAGEMENT .. 9 2.1. Theoretical basis of credit risk management ............................................. 9

2.1.1. Concept of credit risk management .................................................. 9

2.1.2. Contents of credit risk management ................................................. 9

2.2. Theoretical basis for application of artificial intelligence in credit risk management .................................................................................................... 10

2.2.1. Overview of artificial intelligence. ................................................. 10

2.2.2. Artificial intelligence in credit risk management ........................... 10

2.2.3. Measurement framework applied to artificial intelligence model in credit risk management. .......................................................................... 11

2.2.4. Data for artificial intelligence models in credit risk management. 12

2.2.5. Criteria to evaluate the effectiveness of artificial intelligence models in credit risk measurement. ........................................................ 13

2.2.6. Conditions for application of artificial intelligence in credit risk management. ........................................................................................... 14

2.3. International experience in research and application of artificial intelligence in credit risk management. .......................................................... 15

2.3.1. Experience from the UK. ................................................................ 15

2.3.2. Experience from the USA............................................................... 16

2.3.3. Experience from India .................................................................... 16

2.3.4. Experience from the Financial Stability Board (FSB) ................... 17

2.3.5. Experience from the World Bank (WB). ........................................ 17

2.3.6. Lessons for commercial banks in Vietnam .................................... 17

CHAPTER 3: CURRENT SITUATION OF CREDIT RISK MANAGEMENT AT VIETNAM BANK FOR AGRICULTURE AND RURAL DEVELOPMENT .......................................................................... 20 3.1. Overview of Agribank .............................................................................. 20 3.2. Current situation of credit risk management at Agribank. ....................... 20 3.3. Assessing the current situation of credit risk management at Agribank . 20

3.3.1. Achievements ................................................................................. 20

3.3.2. Limitations and causes .................................................................... 21

CHAPTER 4: BUILDING ARTIFICIAL INTELLIGENCE MODELS IN CREDIT RISK MANAGEMENT AT VIETNAM BANK FOR AGRICULTURE AND RURAL DEVELOPMENT. ................................ 22 4.1. Model recommendation ........................................................................... 22 4.2. Building PD model ................................................................................... 22

4.2.1. Description of data collected .......................................................... 22

4.2.2. Results of PD model ....................................................................... 23

4.3. Building LGD model ................................................................................ 23

4.3.1. Data description .............................................................................. 23

4.3.2. LGD model results .......................................................................... 23

4.4. Building EAD model ................................................................................ 24 4.5. Conditions for application of artificial intelligence in credit risk management .................................................................................................... 25 CHAPTER 5: SOLUTIONS FOR APPLICATION OF ARTIFICIAL INTELLIGENCE IN CREDIT RISK MANAGEMENT AT VIETNAM BANK FOR AGRICULTURE AND RURAL DEVELOPMENT ........... 26 5.1. Development orientation for credit risk management activities in the Vietnamese commercial banking system. ....................................................... 26 5.2. Development orientation for credit risk management activities at Agribank .......................................................................................................... 26 5.3. Solutions for application of artificial intelligence in credit risk management at Agribank ................................................................................ 26

5.3.1. About the organizational structure for credit risk management ..... 26

5.3.2. About the process of applying artificial intelligence in credit risk management. ........................................................................................... 27

5.3.3. About the group of necessary support solutions. ........................... 28

5.4. Recommendations to the State Bank of Vietnam .................................... 29 CONCLUSION OF THESIS ........................................................................ 31

III

GENERAL INTRODUCTION

1. RATIONALE OF THE RESEARCH

In the explosion trend of the 4.0 revolution, artificial intelligence is gradually affirming its role as a pioneering technology for the banking industry in general and the credit sector in particular. This technology was developed more than 50 years ago, however, with the advancement of computer science, the abundance of data and the needs of the market, artificial intelligence is being developed strongly and gradually shaping the future game for banks.

Realizing the importance and potential of artificial intelligence, on January 26, 2021, the Prime Minister issued the Decision No. 127/QD-TTg on the national strategy on research, development and application of artificial intelligence to 2030, which stipulates specific tasks of the banking industry, including: “Analyzing and predicting loan demand, borrowers, supporting credit granting activities, detect fraudulent behaviors; customizing banking services for customers; providing instant support services to customers through virtual assistants and chatbots”. In the commercial banking system of Vietnam, Vietnam Bank for Agriculture and Rural Development is the largest bank in terms of total assets and number of customers. However, in the previous period, this bank showed many negative credit-related cases, greatly affecting the business operations and reputation of the Bank. The consequences of these cases remained for a long time and caused direct impacts on employees’ lives. Considering the above situation, the author finds that the research on the thesis: “Application of artificial intelligence in credit risk management at Vietnam Bank for Agriculture and Rural Development” has great significance in both theory and practice.

2. OBJECTIVES OF THE RESEARCH

2.1. General objectives

The thesis does an overall research on the theory and practice of artificial intelligence in credit risk management, thereby proposing solutions for application of artificial intelligence in risk management at Vietnam Bank for Agriculture and Rural Development.

2.2. Particular objectives

The general objectives are concretized into the following four

objectives:

First, systematize the theoretical basis of artificial intelligence in credit

risk management;

Second, assess the current situation of credit risk management at Vietnam Bank for Agriculture and Rural Development in order to determine the conditions and solutions for applying artificial intelligence in this activity;

Third, apply artificial intelligence to build a credit risk assessment model according to the advanced approach of Basel II at Vietnam Bank for Agriculture and Rural Development;

Fourth, propose groups of solutions and recommendations to apply artificial intelligence in risk management at Vietnam Bank for Agriculture and Rural Development.

3. SUBJECT AND SCOPE OF THE RESEARCH

3.1. Subject of the Research

Research subject of the thesis is artificial intelligence in credit risk

management.

3.2. Scope of the Research

Spatial scope of the research: Applying artificial intelligence in credit risk management at Vietnam Bank for Agriculture and Rural Development, particularly focusing on artificial intelligence modeling in credit risk measurement.

Time scope of the research: 2009-2021. In particular, the data used to build the artificial intelligence model was collected during the period 2009 - 2014.

4. RESEARCH METHODOLOGY

- Methods of statistics, description, analysis and synthesis

- Survey method

- Quantification method: The thesis uses artificial intelligence models including: Decision Tree Model (DT), Neural network model (NN) to measure credit risk and make comparisons with traditional models such as the logit model.

5. NEW FINDINGS OF THE THESIS

The thesis has made new contributions in both theory and practice as

follows:

First, the thesis has systematized the theoretical basis of artificial intelligence in credit risk management. Artificial intelligence models are analyzed and clarified according to the steps of credit risk management including: identification, measurement, use of management and reporting tools, and supervision. The thesis also provides a theoretical framework for building and applying artificial intelligence in credit risk management at commercial banks.

Second, the thesis has used the survey method for leaders and employees to assess the current situation of credit risk management at Vietnam Bank for Agriculture and Rural Development.

Third, the thesis has built an artificial intelligence model to measure credit risk based on the real data at Vietnam Bank for Agriculture and Rural Development. Credit risk measurement models are designed according to the Advanced Internal Ratings- Based Approach (AIRB) of Basel II.

the Fourth,

thesis has proposed a series of solutions and recommendations to apply the artificial intelligence model in credit risk management.

6. CONCLUSION OF THE THESIS

In addition to the general introduction, conclusion, list of references and

appendices, the thesis is structured with 5 chapters including:

Chapter 1: Literature review of artificial intelligence in credit risk

management

Chapter 2: Theoretical basis of artificial intelligence application in credit

risk management

Chapter 3: Current situation of credit risk management at Vietnam Bank

for Agriculture and Rural Development

Chapter 4: Application of artificial intelligence to build credit risk measurement models at Vietnam Bank for Agriculture and Rural Development

Chapter 5: Solutions for application of artificial intelligence in credit

risk management at Vietnam Bank for Agriculture and Rural Developmen

CHAPTER 1: LITERATURE REVIEW OF ARTIFICIAL INTELLIGENCE IN CREDIT RISK MANAGEMENT

1.1. Researches on credit risk management models

The theoretical research, Bullivant (2010) presented comprehensive aspects of credit risk management. Some guidelines and researches by IIA (2020) and Oliver Wyman (2016) were about the three-layer protection model or the “four-layer protection” model as renamed by Basel (2015) that is considered the standard in the field of risk management in general and credit risk management in particular. Also studying this model, Tammenga (2020) approached from the perspective of assessing the suitability of the model when using modern tools such as artificial intelligence.

Nguyen Van Tien (2015) and Ghosh (2012) researched on the banking governance model in which the basic point of the credit risk management model is determined as the independence between the business division, risk management division and internal processing division while still ensuring the centralized credit management process.

Research by Le Thi Huyen Dieu (2010) presented an overview of appropriate credit risk management models in the context of commercial banks in Vietnam. On the same basis, the research by Nguyen Bich Ngan (2020) simulated the portfolio risk management model according to the Foundation Internal Ratings- Based Approach (FIRB) of Basel.

In addition, there are some studies on credit risk management at a particular bank that can be mentioned such as those by Tran Khanh Duong (2019), Nguyen Quang Hien (2016), Le Thi Hanh (2017) and Nguyen Nhu Duong (2018).

1.2. Researches on credit risk assessment

Koulafetis (2017) comprehensively studied credit risk measurement models, clearly specifying portfolio risk measurement models, from the ones recommended by Basel according to the standard approach (SA), the Foundation Internal Ratings- Based Approach (FIRB) and the Advanced Internal Ratings- Based Approach (AIRB) to those developed by the world’s oldest financial institutions such as CreditMetrics by JP Morgan (1997), KMV by Moody's (2002), CreditRisk+ by Credit Suise (1997), CreditPortfolioView developed by Wilson (1997) and used by McKinsey.

Many studies focusing on analyzing credit risk measurement models according to the Basel's approach include those by Acharya et al. (2006), Carey & Gordy (2007), Hibbeln (2010), Engelmann & Rauhmeier (2006), Witzany (2017), and Jacob (2010).

In this thesis, the author inherits the theoretical arguments about credit risk management model based on the Advanced Internal Ratings- Based Approach (AIRB) to build an internal credit rating system for Vietnam Bank for Agriculture and Rural Development.

1.2.1. Researches on measurement of probability of default

Some typical international researches include: Altman (1968); Arminger et al (1997); Vasanthi & Raja (2006); Autio et al (2009); Kocenda & Vojtek (2011); and Nwachukwu (2013).

Besides, there are researches using data from Vietnam such as: Dinh and Kleimeier (2007); Tra Pham and Lensink (2008); Linh et al (2020); and Thu et al (2020).

1.2.2. Researches on loss given default (LGD).

Table 1.1: Summary of researches on LGD

Model

Research Data Coun try

Perio d

Macroeconom ic variable

Influence of macro factors

GDP

Altman et al (2005)

USA 1982- 2001

Multivariate regression

No influence

---

Bastos (2010)

Portu gal

1995- 2000

All macro variables

1999- 2005

(LR),

1000 observ ations 374 observ ations 55,000 observ ations

Bellotti and Crook (2011)

bank UK interest rates, unemployment UK rate, income index

Logit (LR), Decision Tree (DT) Linear regression, logit tobit Decision Tree (DT) Multivariate regression

Caselli et al (2008)

Italy 1990- 2004

GDP, employme nt rate

11,649 observ ations

GDP, employment rate, household consumption, annual investment,

Logit (LR) No

Portu gal

1995- 2000

influence

374 observ ations

of by

GDP, frequency default sector GDP

OLS, Logit GDP

Dermine and Carvalho (2006) Khieu et al (2012)

USA 1987- 2007

GDP

Tobit

GDP

USA 1982- 2009

Rosh and Sheule (2012)

1364 observ ations 1653 observ ations

Source: Synthesized by the author

1.2.3. Researches on default risk

Valvonis's research (2008) showed that the determination of EAD depends largely on the credit product and the content of the commitment between the bank and the customer. EAD can be divided into two groups: fixed group and variable group.

For the group with variable loan balance, EAD needs to be estimated according to the model. The studies by Barakova and Parthasarathy (2013), Leow and Crook (2016), Tong et al (2016), Luo and Murphy (2020) examined a variety of EAD affecting factors and EAD forecasting techniques on diverse data sets from individual customers to business customers.

1.3. Researches on artificial intelligence model in credit risk management.

International researches on artificial intelligence in the credit field were carried out in a variety of ways, from different aspects, including general theoretical studies on artificial intelligence and building empirical models.

Table 1.2: Summary of researches on artificial intellectual models in credit risk measurement

Research Date West (2000)

Australia:

Model MDA, LR, KNN, NN (MLP, MOE, RBF, LVQ, FAR);

- Germany, 1000 observations (30% of bad debts) - 690 observations (56 % of bad debts)

Best model Germany: MOE (correct ratio classification 78.6%) Australia: RBF (correct ratio classification 88.78%), MLP (correct classification ratio 87.68%)

Desai et al - USA: 962 NN (MLP, LR (correct classification

(1996)

MNN), MDA, LR

ratio 81.7%) MLP (correct classification ratio of bad debts 42.08%)

(MLP,

Abdou (2008) NN PNN) (correct ratio

Brown et al (2012) (18.42% observations of bad debts), 918 observations (25.98% of bad debts), 853 observations (21.15% of bad debts) 581 Egypt: observations (25.5% of bad debts) Germany, Australia, 03 other datasets

MDA, LR, DT, KNN, SVM, NN, RF, GB

MLP, MMLP MMLP classification

Tsai and Wu (2008)

Australia:

Japan:

(correct ratio

MDA MLP classification 94.84%) the Best SVM with dataset of 30% of bad debts Best RF with the dataset of 1% of bad debts Germany: (correct ratio 83.38%) Australia: MLP (correct ratio classification 97.32%) Japan: MLP classification 87.94%) 95%

( MDA, LR

(PNN,

Altman (1968) Wiginton (1980) Tang et al (2018) NN MLP)

LR (correct classification ratio 61.84%) Australia: PNN (85.64%) Japan: PNN (85.54%)

Japan:

MLP 87%

Zhao et al (2015)

- Germany, 1000 observations (30% of bad debts) - 690 observations (56 % of bad debts) - 690 observations (56 % of bad debts) 66 observations (48.4% of bankruptcy) 1908 observations 41.8 % of bad debts) 690 Australia: - observations (56 % of bad debts) - 690 observations (56 % of bad debts) - Germany, 1000 observations (30% of bad debts)

Source: Synthesized by the author

1.4. Research gap

After doing the literature review, the thesis identified a number of

research gaps as follows:

First, domestic researches have mentioned some aspects of risk management such as organizational models, principles, management processes, and measurement methods, but only few researches on the use artificial intelligence in credit risk management. So far, there is no research on the theoretical framework for building artificial intelligence models in credit risk management.

Second, almost neither

international nor domestic studies have comprehensively studied the use of artificial intelligence in credit risk measurement according to the Basel's advanced approach.

Third, there are no domestic studies on the application of artificial intelligence in risk management at a particular bank. This is the practical gap that the thesis will focus on.

CHAPTER 2: THEORETICAL BASIS FOR APPLICATION OF ARTIFICIAL INTELLIGENCE IN CREDIT RISK MANAGEMENT

2.1. Theoretical basis of credit risk management

2.1.1. Concept of credit risk management

Credit risk management can be expressed in different forms, but the concepts and views are generally about the nature of credit risk: it is the process of formulating and implementing risk management policies and strategies regarding risk identification and measurement, risk prevention, risk handling and risk control measures in order to maximize profits within acceptable risk levels.

2.1.2. Contents of credit risk management

2.1.2.1. Principles of credit risk management

To ensure effective credit risk management, Basel (2000) set out 17 principles of credit risk management, divided into five groups: establishing an appropriate credit risk environment; operating under an effective credit granting process; maintaining appropriate credit management, measurement and monitoring processes; ensuring a safe control system against credit risks; and the role of supervisors.

2.1.2.2. Credit risk management models at commercial banks

* Credit risk management models according to the organizational

structure

Centralized credit risk management model

This model separates credit activities at a branch and a head office. The branch only performs business/ sales/ customer relations functions, while the head office performs credit risk management/ credit appraisal/ analysis and credit approval as well as operational support functions. Distributed credit risk management model The distributed risk management model gives each bank branch a position, which is very independent from the head office as a subsidiary bank in the parent bank. This model does not separate between risk management, business and operations functions. In particular, the bank's credit department fully performs 3 functions and is responsible for all stages of loan preparation.

*Three-layer risk management organization model

The three-layer protection model is built with the aim of distinguishing risk management tasks by different functional departments in the bank. This model provides a simple and effective approach, to enhance communication between risk management and control by clarifying the roles and responsibilities of stakeholders. One advantage of this model is that it is suitable for all credit institutions regardless of the size and complexity of business operations.

2.1.2.3. Credit risk management process

The credit risk management process is built

in four stages: identification, measurement, use of risk management tools and risk reporting (Ghosh, 2012).

2.2. Theoretical basis for application of artificial intelligence in credit risk management

2.2.1. Overview of artificial intelligence.

Artificial intelligence (AI) can be understood as the intelligence demonstrated by machines. In other words, the term AI is used when a machine mimics human “cognitive” functions, such as “learning” and “problem solving”.

The main applications of AI mentioned include: analysis, automated advice, process automation and reporting, in which the applications of risk analysis including credit risk are considered as bringing the greatest benefit to the banks.

2.2.2. Artificial intelligence in credit risk management

According to the guidance of the World Bank (2019), artificial intelligence in credit risk management is divided into 3 categories according to training algorithms including: supervised learning, unsupervised learning and other related techniques.

Table 2.1: Classification of artificial intelligence models

 Decision

tree and

 Automated

Feature

mixed models.

Engineering - AFE

vector

AI for supervised learning AI for other related techniques

 Support machine

 Reinforcement Learning - RL

AI for unsupervised learning  K-means clustering  Hierarchical clustering

 Neural network

 Natural

Language

Processing - NLP

Source: Synthesized by the author based on the WB guidelines (2019)

With the diversity of the above algorithms, artificial intelligence can participate in almost contents of the credit risk management process. The risk management process is normally established in 4 basic steps including: identification, measurement, use of risk management tools, monitoring and reporting. In addition, the use of artificial intelligence also allows banks to control credit risk right at the step of customer search.

Figure 2.1: Artificial intelligence in credit risk management

Source: Synthesized by the author

2.2.3. Measurement framework applied to artificial intelligence model in credit risk management.

including: Standard Approach internalization

Basel II introduced a series of credit risk approaches with increasing complexity and (SA), Foundation Internal Ratings- Based Approach (FIRB), and the Advanced Internal Ratings- Based Approach (AIRB).

In the research scope of the thesis, the author uses the measurement framework according to the advanced approach of Basel II in which the PD, LGD and EAD models are estimated through historical data of the bank.

education,

type, field of business,

Legal informatio n about customers

2.2.4. Data for artificial intelligence models in credit risk management.

Individual: age, gender, marriage, employment, residence, insurance, dependents, membership status in related organizations and groups Enterprise: address, size

Assets, capital,

income, expenditure,

PD

financial performance, manageability,

solvency

Informatio n about financial status

Limit, outstanding balance, interest rate,

term, disbursement method, repayment

Information about the loan

method

LGD

Types of services used at the bank, usage

behavior & habits

Information about transaction history

Type, form of ownership, value, volatility,

commitment related to collaterals

Information about Collaterals

EAD

Bad debt & overdue debt handling policy,

credit control & supervision policy,

Internal information of the Bank

collateral handling policy and customer

support policy GDP,

import-export

balance,

Outside Information

unemployment

rate,

inflation

index,

business industry & market index

Figure 2.2: Data used in PD, LGD and EAD models

Source: Synthesized by the author

2.2.5. Criteria to evaluate the effectiveness of artificial intelligence models in credit risk measurement.

2.2.5.1. For the PD model

Correct classification ratio matrix

This is one of the most commonly used criteria in the researches on

credit rating models (Abdou, 2009; Zheng et al. 2004).

Model Training data CCG % CCB % Test data CCG % CCB % All data CCG % CCB %

Aver age ratio Aver age ratio Aver age ratio A B C

In particular, CCG% is the correct classification ratio of good debts, representing the number of correctly classified good debts out of the total number of actual good debts; CCB% is the correct classification ratio of bad debts, representing the number of correctly classified bad debts out of the total number of actual bad debts; The average correct classification ratio is calculated based on the total number of correctly classified cases out of the total number of observations.

Misclassification cost

The misclassification cost of each model will be determined according

to the West (2000) formula given as follows:

Misclassification cost = 𝐶1 x𝑃1 x 𝜋1 + 𝐶2 x 𝑃2 x 𝜋2

Where:

𝐶1 and 𝐶2 are misclassification costs for type I and II errors, respectively; 𝑃1and 𝑃2are the probabilities of type I and II errors; 𝜋1and 𝜋2are the ratios of bad debts and good debts.

ROC (Receiver operating characteristic) curve

ROC is a commonly used graph in binary classification models. In the field of credit ratings, this curve is generated by representing the correct

classification ratio of bad debts (sensitivity) based on the misclassification ratio of good debts (– 1-specificity) (Abdou, 2009).

2.2.5.2. For the LGD and EAD models

For the LGD and EAD models, the dependent variable is a continuous variable, so error measures will be used to evaluate the effectiveness of the model (Scandizzo, 2016). One of the most used indicators to evaluate artificial intelligence models is RSE (Relative squared error) using the following formula:

Where: 𝑦𝑖̂ and 𝑦𝑖 are the predicted and observed values; mean (y) is the

mean observed value.

2.2.6. Conditions for application of artificial intelligence in credit risk management.

Many researches have shown significant advantages of artificial intelligence over traditional methods in credit risk management. However, for application of these models in practice, it requires comprehensive attention on the conditions and requirements set forth with these models.

Competition "Race" among credit institutions

Benefits Ability to reduce costs, increase income, increase risk management

Credit related factors Availability of infrastructur e and data for adoption

Technology Computation al Improvemen ts, Algorithms, Costs

Regulation Regulations on safety, transparency and data used.

Figure 2.3: Conditions for deployment of artificial intelligence model

Source: FSB (2017)

2.3. International experience in research and application of artificial intelligence in credit risk management.

2.3.1. Experience from the UK.

About strategy for research and application of artificial intelligence

technology and hardware technology. Artificial

Investment and research in artificial intelligence in the UK boomed in the early years of the 21st century due to the strong development of Machine learning intelligence applications are utilized in most stages from customer contact (front-office) activities to internal (back-office) activities with focus on anti-money laundering and customer related services, credit risk management, valuation, insurance and underwriting.

About selection of artificial intelligence models

Regulatory agencies such as the Bank of England, the Prudential Regulation Authority (PRA) and the Financial Conduct Authority (FCA) are completely neutral on technology application in financial institutions, and no restrictions or bans are imposed on the artificial intelligence models applied. According to statistics, the artificial intelligence model based on Decision Tree (DT) is most commonly used, followed by the natural language processing and neural network approaches.

About data for artificial intelligence models

Financial institutions in the UK use a variety of data types for artificial intelligence models including: structured data, semi-structured data and unstructured data. Structured data refers to data contained in columns and rows whose elements can be linked by predefined fields.

About the process of applying artificial intelligence models into

practice

The process of applying the artificial intelligence model introduced by the Bank of England includes 5 steps as follows: data collection and input; data selection and classification; model running; model appraisal; implementation and safeguards.

About risks and challenges to be solved when applying artificial

intelligence models

First are risks associated with the lack of explanatory power of artificial intelligence models. This raises more and more questions about the appraisal of the model structure and performance. Second are risks related to data quality where biased data may occur. Third are risks associated with negative effects on customers and damages to the bank's reputation.

2.3.2. Experience from the USA

About strategy for research and application of artificial intelligence

Most of the major banks in the US have used artificial intelligence in their key business and risk management activities. US Federal Reserve Governor - Lael Brainard (2021) said that artificial intelligence is being used to analyze traditional data in credit approval and credit risk analysis activities for the purpose of collecting detailed information that cannot be done by traditional credit risk management methods.

About risks and challenges to be solved when applying artificial

intelligence

First, the existence of pre-existing management systems has consumed a large amount of construction cost and its vulnerabilities have slowed down the adoption of artificial intelligence.

Second, the research and application of artificial intelligence requires

the ability to find and recruit talented personnel in this field.

Third, challenges come from the outside, such as customer trust in the security capabilities of artificial intelligence and the explanation of how to use artificial intelligence to managers.

2.3.3. Experience from India

About strategy for research and application of artificial intelligence

Many credit institutions in India have developed and applied AI in risk management and credit approval including the State Bank of India (SBI), HDFC, ICICI and other credit institutions. While banks tend to apply artificial intelligence interacting with customers (chatbots) and performing transactions, the strategy of emerging credit institutions is to focus on the credit market for low-income people and they have made further strides in application of artificial intelligence in risk management and credit approval processes.

About data used for artificial intelligence models

According to Singh & Prasad (2020), credit institutions in India are assessing a variety of data about customers including personal information, messages, online social network activities, consumption and spending habits for the purposes of reviewing borrowers and assessing risks.

2.3.4. Experience from the Financial Stability Board (FSB)

About risks and challenges to be solved when applying artificial

intelligence

According to FSB (2017), besides the benefits and efficiency it brought, the artificial intelligence model itself may also have errors in the construction process. These errors may come from many causes such as inappropriate data, inappropriate algorithms or inappropriate modeling strategy.

2.3.5. Experience from the World Bank (WB).

About strategy for research and application of artificial intelligence

According to the research by Strusani & Houngbonon (2019) from IFC of the World Bank, artificial intelligence is transforming credit operations in emerging markets including the poorest. This is a good lesson for Vietnam, where the development potential of the credit market is very large, while access to credit sources is still limited due to traditional credit risk management methods.

About risks and challenges to be solved when applying artificial

intelligence

First, artificial intelligence models using non-traditional alternative data can capture legally and ethically limited data such as race, gender and region discrimination data. Second, artificial intelligence models often have a complex structure and it is difficult to explain the interactions within the model. Third, artificial intelligence places requirements on credit institutions in collecting and storing large amounts of personal data. This may lead to risks associated with the theft and use of these personal data for other purposes. Fourth, artificial intelligence models may cause bias in decisions.

2.3.6. Lessons for commercial banks in Vietnam

About strategy for research and application of artificial intelligence

in credit risk management

The strategy for application of artificial intelligence models in credit risk management needs to have a roadmap and must be combined with other strategies such as: human resource strategy, technology infrastructure development strategy and business strategy to create synchronization and maximize the effectiveness of the artificial intelligence models.

About data for building artificial intelligence models

Banks need to identify all possible sources of customer-related data that can be collected to build an artificial intelligence model to increase the accuracy of the model in predicting customer credit risk.

About selection of an artificial intelligence model in credit risk

management

Reality shows that no artificial intelligence model is optimal in all circumstances. Each artificial intelligence model will promote its advantages in certain situations, so it is necessary to test many models to choose the most suitable model for the existing data set.

About risks and challenges to be solved when applying artificial

intelligence in credit risk management

The first obstacle is that the old systems have been deeply rooted in the operational process and have also proven to be effective for the credit risk management of the banks. Next, the complexity of the artificial intelligence model makes it difficult to understand and explain the relationship between the inputs and the credit risk prediction results. Besides the above obstacles, the artificial intelligence model itself may also pose risks related to the accuracy and suitability of the model whenever there is a change in the context.

About policy recommendations for the use of artificial intelligence

in credit risk management

First, the Regulatory Authority should issue a legal framework that specifies supervision and accountability in the use of credit risk assessment models.

Second, the Data Collection Process should be made public to

customers.

Third, credit institutions should have a clear understanding of the models and be able to explain to customers about the credit risk assessment results.

Fourth, credit institutions have to explain to regulatory agencies about the process and appropriateness of using artificial intelligence models, and that the applied algorithms and output data are within the estimate.

Fifth, credit institutions should establish a customer data control system to prevent unauthorized access or extraction of data from cyber attacks outside or within the bank.

Sixth, credit institutions should have regulations on model risk

management when applying artificial intelligence in credit risk management.

Seventh, regulatory agencies should establish a mechanism for

cooperation and data sharing among related organizations in the credit field.

CHAPTER 3: CURRENT SITUATION OF CREDIT RISK MANAGEMENT AT VIETNAM BANK FOR AGRICULTURE AND RURAL DEVELOPMENT

3.1. Overview of Agribank

Agribank's business performance is assessed through basic indicators, including: equity, charter capital, total assets, outstanding debt, mobilized capital, bad debt ratio, ROA, ROE. It is compared in some criteria with the five largest banks in the market.

3.2. Current situation of credit risk management at Agribank.

Currently, Agribank still applies a distributed credit management model with regulations on the powers of loan approval and loan management for all levels including General Director, Directors of branches of grades I, II and implementing risk transaction offices. In recent years, Agribank management under the independent three-layer protection model. However, the separation at branches is still limited when there is usually overlap in the performance of these functions by credit officers.

3.3. Assessing the current situation of credit risk management at Agribank

3.3.1. Achievements

3.3.1.1. About the organizational structure and credit risk management policy

Agribank has basically built a system of frameworks and policies for credit risk management and maintained a policy for preventing and limiting credit risks. Accordingly, credit policies and regulations cover quite comprehensively the necessary contents to be implemented by banks in the credit approval process.

3.3.1.2. About credit risk identification and measurement

Agribank has paid much concern on credit risk identification with all stages involving the bank contact with customers including: before, during and after lending.

Agribank's internal credit rating system has met the minimum conditions

for building the SBV's internal credit rating system in the previous period.

3.3.1.3. About use of credit risk management tools.

Agribank uses a variety of risk management tools in all stages: before, during and after credit granting. In fact, the bank's policies and activities for handling and recovering bad debts have been effective in recent years.

3.3.2. Limitations and causes

3.3.2.1. About the organizational structure and credit risk management policy.

Agribank's credit management organization model is a distributed model, which is not suitable for the risk nature and development trends in the banking industry. The reason for this situation is the difficulty in building a centralized credit risk management process due to complicated locations and insufficient staff in many branches.

3.3.2.2. About credit risk identification, assessment and measurement.

The level of customer risk is not determined directly on the customer's probability of default (PD), which causes difficulties for Agribank in meeting the standards of credit rating system proposed by Basel.

There are four main reasons for this situation:

First, the State Bank has not yet issued detailed instructions on building an internal credit rating system and measuring credit risk according to Basel II standards; Second, staff resources for doing credit and credit risk management are limited in both quantity and quality; Third, the credit risk assessment system according to Basel II standards is a comprehensive and highly complex system that requires the bank to have a thorough preparation in terms of both strategy and resources for implementation.

3.3.2.3. About use of risk management tools.

Agribank only uses basic and traditional risk management tools. Agribank lacks an effective credit risk measurement tool, and the financial market has not yet been developed to the extent that modern hedging tools can be used.

3.3.2.4. About credit risk monitoring and reporting

Credit risk supervision at Agribank is inefficient due to no functional separation between the transaction department, and credit appraisal and re- evaluation department while the supervision effectiveness of the internal control department is still limited.

CHAPTER 4: BUILDING ARTIFICIAL INTELLIGENCE MODELS IN CREDIT RISK MANAGEMENT AT VIETNAM BANK FOR AGRICULTURE AND RURAL DEVELOPMENT.

4.1. Model recommendation

AIRB - Advanced internal rating based

LGD

EAD

Logit (LR)

Decision Tree (DT)

Neural Network (MLP)

In order to perform a comprehensive credit risk assessment, in this thesis the author proposes to build a model according to the advanced internal rating – based approach (AIRB) of Basel.

Figure 4.1: Proposed research model

Source: Proposed by the author

4.2. Building PD model

4.2.1. Description of data collected

To build a PD model, the author uses a database that includes information on consumer and business loans at Agribank in the period from 2009 to 2014. The final dataset is used to build the PD model including information about 15,470 customers with 19 related characteristics, including 12118 good debts accounting for 78.3% and 3352 bad debts accounting for 21.7%.

4.2.2. Results of PD model

Table 4.1: Classification results of representative models

Model

Test data

Training model

General model

CCB %

CCG %

Overal l %

CCB %

CCG %

Overal l %

CCB %

CCG %

Overal l %

32.28

95.43

81.93

31.02

96.50

82.55

31.97

95.75

82.17

DT4

48.50

93.27

83.59

47.93

92.38

82.71

48.33

93.00

83.32

DT16

47.08

94.09

83.85

46.66

94.10

83.95

46.95

94.09

83.88

MLP7*

52.95

95.37

86.21

52.66

94.65

85.48

52.86

95.16

85.99

MMLP

54.32

93.35

85.18

54.07

92.93

84.65

54.25

93.35

85.20

Source: Results calculated by the author

The model with the most accurate classification is MLP7 which is also the most cost-effective model based on the error classification cost ratio of 3:1. However, at the 10:1 error classification cost ratio, the MMLP model shows better loss reduction than the MLP7 model.

The ROC curve analysis results show that the models meet the requirements well with the curve above the random line and the AUC index (area under the ROC curve) both above 0.8. Furthermore, this index also shows that the MLP7 model has the best classification ability when changing the probability threshold for determining bad debts and good debts. The fit of the MLP7 model is excellent for this dataset.

4.3. Building LGD model

4.3.1. Data description

The 2nd dataset was initially collected consisting of 1124 observations. However, after removing some cases where the customers are exceptions with a very long debt grace period, the data set used for building the LGD model includes 1045 loan records of customers transferring bad debts from 1 to 3 years.

4.3.2. LGD model results

Statistical results of relative errors of 20 LGD models show values ranging from 0.13 to 0.246. With such results, the LGD models have good predictive accuracy and the best is the LGD6 model.

Table 4.2: Relative error of LGD models

Model

Sum of Squares Error

Relative Error

Sum of Squares Error

Relative Error

31.63

0.23

LGD11

31.001

0.227

LGD1

38.999

0.242

LGD12

27.127

0.161

LGD2

27.56

0.168

LGD13

28.061

0.2

LGD3

30.921

0.223

LGD14

27.658

0.185

LGD4

32.417

0.242

LGD15

28.657

0.203

LGD5

20.689

0.13

LGD16

33.869

0.215

LGD6*

33.58

0.21

LGD17

23.869

0.144

LGD7

37.13

0.246

LGD18

30.374

0.201

LGD8

27.796

0.189

LGD19

27.765

0.161

LGD9

30.448

0.219

LGD20

29.73

0.166

LGD10

Source: Results calculated by the author

4.4. Building EAD model

The EAD model is built on the MLP neural network with the dependent

variable being the limit at the time of customer default (bad debt transfer) and

the independent variables being all characteristics of the customer and the

loan. The EAD3 model is the proposed model with the lowest relative error.

Table 4.3: Relative error of EAD models

Model

Sum of Squares Error

Relative Error

Sum of Squares Error

Relative Error

33.660

33.177

0.224

0.231

41.018

34.694

0.271

0.179

19.188

15.169

0.116

0.150

33.394

20.072

0.300

0.185

28.723

37.044

0.263

0.292

55.286

36.089

0.314

0.223

74.666

25.101

0.295

0.216

23.881

29.914

0.165

0.151

48.716

26.834

0.380

0.161

27.168

29.656

0.236

0.199

EAD11 EAD12 EAD13 EAD14 EAD15 EAD16 EAD17 EAD18 EAD19 EAD20

EAD1 EAD2 EAD3* EAD4 EAD5 EAD6 EAD7 EAD8 EAD9 EAD10

Source: Results calculated by the author

The training results also show that the limit is a major determinant factor to the predictive ability of the EAD model. The importance of the limit variable does not change in 20 times of estimation of the EAD model and is nearly 4 times higher than the second influential factor, which is the loan term. The next group that has a significant impact on the model is income, interest rate, borrowing purpose and payment method.

4.5. Conditions for application of artificial intelligence in credit risk management

First, in terms of technology, Agribank has not yet implemented any

activities using artificial intelligence.

Second, much of the information contained in a customer's credit profile is stored in the form of hard documents and under distributed management across branches and transaction offices. Such information can only be exploited and used when it is digitized and uploaded to the bank's information system.

Third, economic benefits are an important motivation for Agribank to deploy AI applications in credit risk management to replace the existing credit management system with high costs and great risks.

Fourth, the Vietnamese banking market will witness a drastic change in the strategy of applying artificial intelligence in the credit sector, so in order to maintain the market share and continue to improve the effectiveness of credit risk management, it is essential to consider research and application of artificial intelligence.

Fifth, in terms of legal conditions, neither the State Bank nor Agribank have any documents regulating issues related to the application of artificial intelligence in credit risk management. This is a common difficulty for the entire banking market as it is a relatively new field in Vietnam

CHAPTER 5: SOLUTIONS FOR APPLICATION OF ARTIFICIAL INTELLIGENCE IN CREDIT RISK MANAGEMENT AT VIETNAM BANK FOR AGRICULTURE AND RURAL DEVELOPMENT

5.1. Development orientation for credit risk management activities in the Vietnamese commercial banking system.

On August 8, 2018, the Prime Minister issued the Decision No. 986/QD-TTg on “Development strategy for the Vietnamese banking industry to 2025, with orientation to 2030”. On that basis, the SBV also sets requirements for commercial banks to perfect and apply the risk management system in accordance with the principles and standards of the Basel Committee and the roadmap for applying Basel II in Vietnam.

On January 26, 2021, the Prime Minister issued the Decision No. 127/QD-TTg on the national strategy on research, development and application of artificial intelligence until 2030. In particular, specific tasks of the State Bank are clearly specified.

5.2. Development orientation for credit risk management activities at Agribank

the organizational model and personnel according to

In 2020, Agribank finished restructuring phase 2 in association with effective implementation of the business strategy and handling of bad debts for the 2016-2020 period, with a vision to 2030. To accomplish the above mission, Agribank has come up with solutions for further reorganizing the network, the perfecting restructuring plan, in line with the capacity of governance, risk and safety management, and improving the operational efficiency.

5.3. Solutions for application of artificial intelligence in credit risk management at Agribank

5.3.1. About the organizational structure for credit risk management

The bank must establish a risk management structure that ensures the

following principles:

Principle of centralization: Centralized credit approval, credit risks must

be centrally managed at the head office.

Principle of independence and objectivity: The model of preventing and limiting credit risks must be independent in a clear separation between 3 departments: business; risk management and operations.

5.3.2. About the process of applying artificial intelligence in credit risk management.

Based on the results obtained in Chapter 4 on the model for assessing the probability of default (PD) and the model for predicting the loss given default (LGD), Agribank needs to apply the PD and LGD (and EAD) models according to a specific roadmap divided into two phases of model building and model deployment.

5.3.2.1. Proposing the process of building PD, LGD and EAD models

Based on the researches on building PD models in Chapter 2 and the author's empirical research, the author proposes an 8-step process to establish a credit rating model including: determining the scope and objectives; preparing a list of criteria and extracting data; converting and standardizing input criteria; model building and selection; model testing; model calibration.

The proposed internal credit rating system structure will combine qualitative factors, quantitative factors and an appropriate adjustment framework to most accurately assess the credit risk of customers. The structure of the built PD model includes the following main components:

Figure 5.1: Structure of the proposed credit rating model

Source: Proposed by the author

5.3.2.2. Proposing the process for practical application of models.

Figure 5.2: The process of applying PD, LGD, and EAD models in practice

Source: Proposed by the author

5.3.3. About the group of necessary support solutions.

5.3.3.1. About the risk early warning and identification system.

Agribank needs to develop a roadmap to build an early warning system (EWS). In order to maximize the effectiveness of the EWS system, Agribank needs to issue a policy framework on early warning, which clearly defines the functions and duties of each department, implementation process and operating mechanism, to ensure regularly exchanging and updating between operational departments on risk signs and assessment methods must be adaptable with continuous and complex changes of the reality.

5.3.3.2. About human resources

HR solutions are divided into stages of preparation, research,

development and application of artificial intelligence:

Preparation stage: Agribank needs to establish a team of experts from the bank to prepare a plan for the research and application of artificial intelligence.

Research and development stage: this activity is not only conducted in the early stages of artificial intelligence development, but it needs to be continued even after applying the models into reality. Agribank should have research programs on artificial intelligence, connect with the network of scientists in the field through projects and seminars.

Application stage: Agribank needs to have a strategy for training and developing human resources, perfecting and improving knowledge, skills and qualities for bank staff to ensure compatibility with the credit risk management process using artificial intelligence.

5.3.3.3. About information technology

- Improving the internal communication, statistics and reporting systems to build a centralized and unified management information system, ensuring the timely and effective development and implementation of credit risk management policies.

- Building a large and comprehensive database including core information, trending and future information for research and development of artificial intelligence models.

- Increasing the application of automation processes with the core factor

being the artificial intelligence model.

5.4. Recommendations to the State Bank of Vietnam

First, the State Bank should issue strategies for applying artificial intelligence in the banking industry, including specific strategies for the credit and credit risk management sectors.

Second, the State Bank needs to develop a legal framework regulating the research, development and application of artificial intelligence for commercial banks.

Third, building a unified and interconnected database between banks and credit institutions and furthermore, a digital ecosystem in service of banking activities.

Fourth, strengthening research and development of artificial intelligence in credit risk management through projects, plans, seminars and international cooperation programs.

intelligence solutions to provide artificial

Fifth, the State Bank should have policies to support and encourage Fintech companies in risk management in general and credit risk management in particular for commercial banks and financial institutions.

CONCLUSION OF THESIS

Within the scope of choosing a particular commercial bank, namely Agribank which has a large scale and great influence in the market, the thesis has solved the basic objectives when researching the application of artificial intelligence in credit risk management as follows:

First, the thesis has generalized the theoretical issues of credit risk

management and artificial intelligence in credit risk management.

Second, the thesis has synthesized experiences in research and application of artificial intelligence in three countries: UK, US, India and from 2 international organizations, namely the Financial Stability Board (FSB) and the World Bank (WB).

Third, the thesis has generalized the current situation of credit risk and credit risk management at Agribank by using secondary data and results of survey on relevant leaders and employees.

the Fourth, thesis has used

the actual data at Agribank for experimentally building a credit risk measurement model according to the advanced approach of Basel II.

Fifth, the thesis has proposed systematic solutions for research and application of artificial intelligence in credit risk management at Agribank on the basis of the existing conditions of the bank itself and external macro factors.

Besides the goals achieved, the thesis still has some limitations due to the difficulties in data collection and this may be a suggestion for further researches.

LIST OF RESEARCH WORKS OF THE AUTHOR

RELATED TO THE THESIS

1. Nguyen Tien Hung (2015). An internal credit rating system: comparing the artificial neural network model and the logit model. Banking Review, No. 11, pp. 24-30.

2. Nguyen Tien Hung and Le Thi Huyen Trang (2018). A credit scoring model based on combination of Decision Tree, Logit, K nearest neighbor and Artificial Neural Network. Banking Science & Training Review, No. 193, pp. 43-54.

3. Nguyen Tien Hung (2021). Experience of using artificial intelligence in the credit sector in some countries and lessons for Vietnamese commercial banks, Asia-Pacific Economic Review, No. 587, pp. 88-90.