ADBI Working Paper Series

CREDIT RISK DATABASE FOR SME

FINANCIAL INCLUSION

Lan H. Nguyen and Megumi Sagara

No. 1111
April 2020

Asian Development Bank Institute

 Lan H. Nguyen is a senior analyst at the Credit Risk Database Association. Megumi Sagara
is director of the Overseas Division at Credit Risk Database Association.
The views expressed in this paper are the views of the author and do not necessarily reflect
the views or policies of ADBI, ADB, its Board of Directors, or the governments they
represent. ADBI does not guarantee the accuracy of the data included in this paper and
accepts no responsibility for any consequences of their use. Terminology used may not
necessarily be consistent with ADB official terms.
Working papers are subject to formal revision and correction before they are finalized and
considered published.

The Working Paper series is a continuation of the formerly named Discussion Paper series;
the numbering of the papers continued without interruption or change. ADBI’s working papers
reflect initial ideas on a topic and are posted online for discussion. Some working papers may
develop into other forms of publication.

In this report, “$” refers to United States dollars.

Suggested citation:

Nguyen, L. H. and M. Sagara. 2020. Credit Risk Database for SME Financial Inclusion. ADBI
Working Paper 1111. Tokyo: Asian Development Bank Institute. Available:
https://www.adb.org/publications/credit-risk-database-sme-financial-inclusion

Please contact the authors for information about this paper.

Email: sagara@crd-office.net, nguyen@crd-office.net

Asian Development Bank Institute

Kasumigaseki Building, 8th Floor
3-2-5 Kasumigaseki, Chiyoda-ku
Tokyo 100-6008, Japan

Tel: +81-3-3593-5500
Fax: +81-3-3593-5571
URL: www.adbi.org
E-mail: info@adbi.org

© 2020 Asian Development Bank Institute

ADBI Working Paper 1111 Nguyen and Sagara

Abstract

We introduce the Credit Risk Database (CRD) and its contribution to financial inclusion efforts
in Japan. By collecting financial data about small and medium-sized enterprises (SMEs), the
CRD contributes to the overall understanding of the SME sector, to the adaptation of risk-
based lending and to a fairer loan guarantee system. In addition to financial data, the CRD
also includes alternative data, bank account transaction data, when assessing SME credit. A
machine learning model is adopted to process the extremely large body of transaction data.
The best performing predictors of default include cash balance
and cash outflow related to repayments. The machine learning model outperforms the logistic
model and is highly accurate in predicting the probability of short-term default. The alternative
data and model can serve as both an enabling short-term monitoring instrument and a credit
assessment tool for SMEs without financial statements.

Keywords: SME credit assessment, machine learning, transaction data

JEL Classifications: G21, G28, G32

ADBI Working Paper 1111 Nguyen and Sagara

Contents

1. INTRODUCTION ......................................................................................................... 1

2. JAPAN’S SME FINANCIAL INFRASTRUCTURE: CREDIT RISK DATABASE .......... 1

2.1 Background ..................................................................................................... 1

2.2 Impact of the Credit Risk Database on the Financial Inclusion of SMEs......... 2
2.3 Credit Scoring Model for Credit Risk Management ......................................... 4

3. CREDIT SCORING USING BANK TRANSACTION DATA WITH THE MACHINE

LEARNING APPROACH ............................................................................................. 5

3.1 Dataset ............................................................................................................ 5

3.2 Default Definition ............................................................................................. 6
3.3 Modelling Approach ......................................................................................... 7
3.4 Model Performance ......................................................................................... 9

4. CONCLUSIONS AND POLICY IMPLICATIONS ....................................................... 10

REFERENCES ..................................................................................................................... 12
ADBI Working Paper 1111 Nguyen and Sagara

1. INTRODUCTION
The access that small and medium-sized enterprises (SMEs) have to finance has been
central to policy discussion in both advanced and developing countries. Maningo (2016)
and Amornkitvikai et al. (2016) describe common situations in developing Asia where
SMEs struggle with limited funding, restrictive repayment periods, high financing costs,
and a complex lending process. One of the major concerns is collateral, which most
SMEs cannot provide.
Efforts to loosen collateral requirements only began after the Asian financial crisis as a
result of declining financial assets and property. These efforts, however, made only a
small impact on closing the SME finance gap. The World Bank reports an unmet SME
finance demand of about $5.3 trillion every year, the largest share of which is from Asia.
Financial institutions are encouraged to further lessen dependence on collateral and to
fast-track risk-based lending for SMEs, to improve their access to finance.
In this paper, we introduce the Japanese approach to SME risk-based lending. We
present Japan’s nationwide financial database, which plays an important role as SME
infrastructure enabling SME credit risk analysis and modeling. We also introduce the
three most important contributions of the database: SME statistical analysis, credit risk
assessment models and the guarantee fee rating scheme. Second, we explore bank
transaction data as an alternative for incorporation into risk models to tackle challenges
in short-term monitoring and the lack of financial information for micro SMEs and
startups. We briefly describe our model, which was developed together with the Bank of
Japan (BOJ) and a Japanese Megabank, along with the main variables, the machine
learning approach and its overall performance. Finally, policy implications for the
improvement of SME access to finance are drawn from the Japanese experience,
indicating the leadership role that the financial sector authority should take in establishing
SME data infrastructure and promoting statistical analysis using both traditional financial
data and alternative data.

2. JAPAN’S SME FINANCIAL INFRASTRUCTURE:

CREDIT RISK DATABASE
2.1 Background
The Credit Risk Database was established following the collapse of the Japanese bubble
economy in the 1990s. The collapse caused a sharp decline in land price and a
consequent credit crunch affecting many businesses, especially SMEs. Banks at
that time needed to move away from lending basing on land-collateral to lending based
on credit risk. To evaluate a business based on credit risk, it is necessary to use
a statistical model with a relatively large pool of data. In 2001, in order to establish
such a large pool of data, the SME Agency, an authorized body of the Ministry of
Economy, Trade and Industry, and the BOJ together established the Credit Risk
Database (CRD) using government funding. The CRD is managed by the CRD
Association, a private not-for-profit organization with data-providing institutions as
membership-paying members.
Data has been duly collected from member financial institutions since 2001. Members
include 51 credit guarantee corporations, 103 private financial institutions and
five government-affiliated institutions. The BOJ, Financial Service Agency (FSA)
and SME Agency are also members. As of November 2019, the database had collected

1
ADBI Working Paper 1111 Nguyen and Sagara

2.5 million SMEs and 1.2 million sole proprietors. There are 21 million financial
statements (data points) for SMEs and 5.6 million financial statements for sole
proprietors.

2.2 Impact of the Credit Risk Database on the Financial

Inclusion of SMEs
The database contributes to SME financial inclusion in several ways. First, it helps
mitigate the information asymmetry problem by providing information and benchmark
statistics for SME sectors. Information such as the business performance of a typical
SME in a specific industry in a specific region is now available. Policy makers such as
the SME Agency, BOJ, and FSA are frequent users of such overall statistical data and
use it for policy analysis. The recent SME Agency white paper (Japan SME Agency 2019)
included an interesting study of how business performance could be improved through
a shift in management from elderly to younger managers. A recent study of the database
also provides insights into the impact of subsidies policy in an evidence-based policy
making (EBPM) study sponsored by the Japan SME Agency (Deloitte Tohmatsu
Consulting 2018). 1
The database also makes the credit risk evaluation of SMEs possible. Credit scores are
given to an SME based on the credit risk model developed by the CRD Association. All
SMEs who are clients of a member bank are eligible for the scoring service. Where a
member bank already has an internal risk model, the scores given by the CRD
Association can be used to cross-checking or validate the internal model, a practice that
is strongly recommended by the Japanese financial authority. Figure 1 presents the
distribution of credit scores for a sample set of data. The horizontal axis represents the
credit score range, in which 0 indicates the highest risk and 100 indicates the lowest risk.
The company introduced in this example has a score of 38. To determine whether 38 is
good or bad, it is useful to look at its position relative to
the mean score of defaulted companies. The mean score takes a value of 37, as depicted
in green. According to its score, this company’s credit position is relatively close to
default, it ranks 7,200th in its industry and 220,000th against similar sized sales groups.
Benchmarking in the region and further breakdowns by industry are
also feasible.
The third impact of CRD is its contribution to improving the guarantee policy for SMEs.
A SME in Japan can apply to be guaranteed by a credit guarantee corporation if deemed
necessary by the lending institution. The SME pays a guarantee fee in exchange for the
guarantee service. In case of default, the credit guarantee corporation shall bear all
responsibility to reimburse the lending institution. In 2006, the credit guarantee
corporation changed their fee structure from charging a universal guarantee rate to
charging a different rate based on credit risk measured by CRD risk evaluation model.
In this fee scheme, the riskiest SME pays a guarantee fee of 2.2% of the total loan
amount, while the least risky SME pays only 0.5%. This fairer fee system, considering
credit quality, has helped to lower the subrogation rate and mitigate the adverse selection
problem. Figure 3 describes the nine levels corresponding to the nine ranks of credit risk
and their respective fee rates.

1 The study investigates firm performance before and after receiving subsidies from the government. Firms
are categorized by their level of creditworthiness measured by default probability, and by growth potential
measured by growth scores. Both default probability and growth scores are build-in parameters calculated
using CRD models. For further details, refer to the Japan SME Agency sponsored paper (in Japanese):
https://www.meti.go.jp/meti_lib/report/H30FY/000316.pdf.

2
ADBI Working Paper 1111 Nguyen and Sagara

As part of the SME inclusion efforts, the credit grantee system for SMEs is being
discussed and designed in many developing nations, such as the Philippines, Myanmar,
Thailand, and so on. Maningo (2016) pointed out that the lack of SME information is a
challenge for the Philippines’ Credit Surety Fund, in order to correctly assess and
manage default risk, as the past due ratio increased to 1.6% in 2014 compared to the
2013 level of 0.6%. A database such as the CRD could both provide SME information
for guarantee decision-making and create an enabling instrument to strategically decide
guarantee fees.

Figure 1: Credit Score and Its Uses

Source: CRD Association.

Figure 2: Subrogation Rate and Outstanding Guaranteed Liabilities

* Subrogation rate = Subrogation Amount/Outstanding Guarantee Liabilities.

Source: Japan Federation of Credit Guarantee Corporation.

3
ADBI Working Paper 1111 Nguyen and Sagara

Figure 3: Credit Guarantee Fees Scheme before and after 2006

Source: Japan Federation of Credit Guarantee Corporation.

2.3 Credit Scoring Model for Credit Risk Management

The conventional credit scoring model normally applies logistic regression to predict
default, which often takes a binary value: default or not default. A common approach
of the various default definitions is to follow the monetary authority’s borrower
classification definition. Borrowers in Japan are typically categorized into six main
categories: Normal, Needs Attention, Special Attention, In Danger of Bankruptcy,
De facto Bankrupt and Bankrupt. In addition to borrower classification, loan repayment
records can be complimentarily used as a default definition. Such records are
conventionally updated monthly by most lending institutions.
The traditional model mainly examines the correlation between default and financial
ratios derived from a financial statement. At the time of the loan screening, three-
consecutive-year of financial statement or more are often requested for credit checking.
These statements are then verified by lending officers who examine the statement
accountability. By collecting financial statements from the lending institutions, we
are likely to receive information with high accountability. In fact, collecting financial
statements from lending institutions has been the main CRD approach.
Building the model ideally involves a wide range of financial ratios to capture profitability,
efficiency and stability, as well as the growth potential of the business. Given the volume
of the CRD database, we have included financial ratios representing all above categories
without having to resort to using non-financial data to obtain a model with high accuracy. 2
The output of the model is default probability within 1 to 3 years which conveniently
reflects a business’s chance of bankruptcy within the respective time horizon. This time
horizon is a common time frame that lenders can look at: 1 year for a credit line
and 2–3 years for a term loan. The current model could not provide such analysis
if lenders need to look at default probability within a few months, due to the lack of input
variables, because interim financial statements are often unavailable. The bank
transaction model which outputs default probability within 3–6 months could resolve this
problem. Details of the model are discussed in Section 3.
To ensure the model’s adaptation to the latest portfolio, it is validated yearly under the
supervision of an independent committee of scholars, financial sector practitioners and
policy makers. Committee members should understand the technical aspects of the
model, as well as validation procedures. Model accuracy is tested against an out-of-

2 35–85 financial items in financial statements, approximately 174 financial ratios, are calculated and
become candidates for model variables.

4
ADBI Working Paper 1111 Nguyen and Sagara

sample dataset segmented into industries and regions. If the model doesn’t fit to the out-
of-sample dataset and accuracy falls, adjustments to the model shall be advised by the
committee. The results of validation are available yearly to all member institutions.
Validation framework is described in more details by Kuwahara et al. (2019).

3. CREDIT SCORING USING BANK TRANSACTION

DATA WITH THE MACHINE LEARNING APPROACH
This section summarizes the CRD’s published research, jointly undertaken with the Bank
of Japan and a Japanese megabank to adapt alternative data for credit risk evaluation. 3
Miura et al. (2019) address the need for short-term monitoring by using bank transaction
data to predict short-term default probability. The approach involves applying machine
learning to high-frequency transaction data to study its patterns and correlation with
defaults.
Apart from its monitoring function, the model innovatively reveals new financing
opportunities for micro SMEs and startups who are bank account owners but have
not prepared financial statements. The presence of such micro SMEs is especially
noticeable in developing nations where the number of personal bank accounts is rapidly
increasing but access to finance is still limited.
Bank transaction data is abundant and relatively easy for a deposit-taking bank
to gather, as such data is recorded automatically. Almost 90%–100% of adults in
developed countries have one or more accounts. In developing Asia, the number is
lower, at approximately 78% in Thailand, 36% in Indonesia, 31% of adults in the
Philippines, and 31% in Viet Nam (World Bank 2017). With increasing financial inclusion
efforts in developing nations, the number of bankable individuals is expected to increase
soon. Notable efforts to increase account ownership include full-service transaction
accounts for government cash grants and the promotion of private sector wage payments
by the government of the Philippines (Banko Sentral ng Pilipinas 2019). The abundance
of available data allows machine learning to perform its best, and to provide reliable
predictions without much human intervention.
Another merit of transaction data is its reliability and accountability. Large data is
often associated with errors, but full automation in bank transaction recording, means
that transaction data is typically free from missing values and nearly impossible
to manipulate.

3.1 Dataset
The transaction dataset consists of data ranging from October 2014 to May 2018
(44 months). 4 The original frequency of the transaction data was measured in seconds
but later transformed into monthly frequency for analysis purposes. The two main
reasons for this transformation are that default information is often available in monthly
frequency; and that some major transaction data, such as fixed revenues and fixed costs,
are often monthly payments.
Table 1 describes the three categories of data in our dataset: cash-inflow, cash-outflow
and cash balance. The cash-inflow subcategory mainly consists of cash flow from

3 The original study is published in Bank of Japan Working Paper series (Japanese):
https://www.boj.or.jp/research/wps_rev/wps_2019/wp19j04.htm/.
4 This dataset is prepared for research purpose only.

5
ADBI Working Paper 1111 Nguyen and Sagara

various types of revenue, investing activities and financing activities. Cash-outflow

comprises cash-out flow for payments for cost of goods and services, variable cost, fixed
cost, investing activities and financing activities. Finally, cash balance is calculated on
the daily average of a monthly balance instead of the end-of-month balance for a better
reflection of cash movements.
Banks typically assign a code for each type of transaction, to identify its purpose for
internal monitoring. Each transaction is classifiable to the three large categories noted,
and to various subcategories by referring to this code. It should be noted, however, that
such classification is based on the authors’ experience in the banking industry and is not
a universal approach. Furthermore, internal codes and the way a bank monitors
transaction purposes could also be very different from bank to bank. To achieve a more
standardized approach to classifying transactions, it is necessary to discuss them with
other financial institutions and investigate their practices.
Table 2 describes the data transformations of the data given in Table 1. Level data, first
difference data, frequency data and standardized data is constructed. The number of
cash flow data points, including both money amounts and frequency, totals 984. The
standardized cash-flow data using sales, total assets and outstanding loans data as the
basis adds another 2,952 variables. The growth in cash-flow data in terms of year-on-
year and month-on-month data adds another 1,435 variables. Finally, the period total
and volatility of cash in-flows and out-flows data adds another 615 variables. In total
there are approximately 6000 variables.

3.2 Default Definition

Default is defined in this study as borrowers classified as “Needing Special Attention”
including those who have been late with repayments for 3 months or more. Records of
default status are updated on monthly basis, as a common practice in the banking sector.
Our transaction data will be matched with defaults observed within 3 months
to produce a prediction of default probability within 3 months. In practice, we also
work with other time horizons and calculate default probability within 1, 2, 6, 9, and
12 months. The dataset includes the data of 7,000 anonymous borrowers, of whom 1,400
are associated with default.
The matching process for default information and transaction data is straightforward. For
example, the January 2016 transaction data will be matched with default information from
February to April 2016 if our time horizon is 3 months and will be matched with default
data from February to July 2016 if our time horizon is 6 months.

Table 1: Categories of Transaction Data

Cash-inflow Revenue Bank Transfers for Revenue, FX…
Investing Activities Dividend, Proceeds from Trust Fund…
Financing Activities Borrowing (Loans)…
Other Cash-inflow Cash, Credit Card, Interest…
Cash-outflow Cost of Goods and Services Bank Transfers…
Variable Cost Credit Card, Insurance, Tax…
Fixed Cost Utility (Electric, Gas, Water), Cable…
Investment Activities Investment Trust Fund…
Financing Activities Loan Repayment, Interest Payment, Guarantee
Fees…
Other Cash-outflow Penalties…
Cash Balance Cash Balance Average Monthly Cash Balance…

6
ADBI Working Paper 1111 Nguyen and Sagara

Table 2: Summary of Variables

Variables Data Type Data Source
Cash-inflows, Cash-outflows and Cash Balance Level Transaction
Money Amount Frequency Data
Sales Level Financial Data
Total Assets Level Financial Data
Outstanding Loan Amount Level Loan Data
Standardized Transaction Data by Sales, Total Assets and Standardized
Outstanding Loan Amount.
YoY of Cash-inflows First Difference
MoM of Cash-outflows First Difference
Period Total of Cash-inflows, Cash-outflows and Cash Balance Level
Standard Deviation of Cash-inflows, Cash-outflows and Cash
Balance
YoY Period Total of Cash-inflows, Cash-outflows and Cash Balance First Difference

Figure 4: Example of a Decision Tree

Source: Author.

3.3 Modelling Approach

Given the volume and the quality of the data, a machine learning model is a good choice
compared to a traditional logistic model. Machine learning is powerful in dealing with a
large volume and information-rich dataset to find certain patterns, and specifically a
pattern of default in our case. It has the capacity to process many variables and the
flexibility to accommodate numerous data patterns. The model uses Random Forest, a
popular classification model in machine learning, which consists of a large number of
uncorrelated decision trees (models). Each tree in the forest produces a prediction, and
the average probability becomes the model’s final prediction. The prediction errors of
many uncorrelated models cancel out to output the best prediction.

7
ADBI Working Paper 1111 Nguyen and Sagara

The study further employs XGBoost, a more advanced version of Random Forest, which
can customize the creation of each decision tree. In XGBoost, new trees are added to
correct or minimize the errors of sequential trees. The best model is created when no
further improvement can be made. XGBoost is easier to overlearn than Random Forest
and so the regularization and cross-validation of the model is necessary to prevent such
overlearning.

Figure 5: XG Boost Model

Source: Authors.

Table 3: Parameter Tuning Result for Random Forest and XGBoost

(Out-of-sample Dataset)
Random Forest XGBoost
Transaction Default Default Hyper Hyper
Data Definition Horizon Parameter AR Parameter AR
Monthly Needs Special 3 months (150, 300)* 0.699 (3, 0.01)** 0.7266
Attention
Monthly Needs Special 3 months (150, 350)* 0.707 (4, 0.01)** 0.7329
Attention
Monthly Needs Special 3 months (150, 400)* 0.702 (5, 0.01)** 0.7181
Attention
Note: * (number of trees, number of information features), ** (tree’s max depth, learning rate).

Table 4: Out-of-sample Testing Results

Dataset Model Default Horizon: 3 Months
Testing Dataset Random Forest 0.7070
(parameter tuning) XGBoost 0.7329
Logistic 0.7113
Back test Random Forest 0.7482
XGBoost 0.7728
Logistic 0.7174

8
ADBI Working Paper 1111 Nguyen and Sagara

One drawback of machine learning in general is overlearning or overfitting to the dataset.

Parameter tuning and out-of-sample cross validation help tackle this problem. For
random forest, parameter tuning is undertaken for a number of trees and a number of
information features (financial variables to be used), and in XGBoost, the hyper
parameters are the number of information features, maximum depth of trees and the
learning rate.
Given the abundance of data, cross-validation in the form of back-testing both the
random forest model and XG Boost model is possible. Back-test datasets involve twelve
out-of-sample datasets, each of which comprises monthly data from June 2017 to May
2018.

3.4 Model Performance

We first examine the prediction power of each variable by looking at the accuracy of a
single-variable model. Accuracy ratio (AR) is a commonly used indicator for a model’s
performance among Japanese banks. It is closely related to, and is calculated from,
the area under the curve of the ROC curve. 5 We first observe that monthly average cash
balance has strong predicting power for defaults. Since the monthly average cash
balance plays such an important role, we further examine whether the transformation of
such data could be as powerful. Level data, period sum data and standardized data all
produce a high accuracy ratio, some higher than 0.5.
In addition to cash balance, cash outflow for loans repayment, cash inflows from
revenue, cash outflows for variable costs, cash outflow for cost of goods and services
and cash inflows from loans, all demonstrate relatively strong correlations with defaulting
by producing high AR.
We observe the same set of best performing variables when applying a different
time horizon for default information, although their performance tends to deteriorate
as the time horizon widens. Specifically, the cash balance AR drops from 0.52 for a one-
month horizon to 0.46 for a twelve-month horizon while loan payment cash flow AR drops
from 30th place to 88th place in the ranking of best performing variables.
Table 3 reports the results of parameter tuning, a grid-search procedure to configure the
optimal combination of the hyper parameters producing highest model accuracy.
Specifically, for random forest, when the default horizon is 3 months, AR peaks at 0.707
with the number of trees equals 350, the maximum number of features (financial
variables) equals 150. Cash balance and cash loan payment have the highest feature
importance calculated in terms of the mean decrease gini, 6 consistent with their high
single-model AR. Of 150 features, the first high-performing 84 features are variables
related to cash balance and cash loan payment. For XGBoost, AR peaks at 0.733 when
the time horizon equals 3 months, the tree’s maximum depth equals four nodes, and the
learning rate, a rate indicating how fast the model moves towards to minimum error,
equals 0.01.

5 The closer that AR is to one, the better the model. A random model’s AR equals 0 and a perfect model’s
AR equals 1. See Annex 3 of “The Study for the Introduction of Credit Risk Database (CRD) in the
Philippines: Data Quality Examination Report” for technical details.
6 Feature importance helps indicate a variable’s predictive power and which variables are the top features
contributing to the prediction of the model. Feature importance is calculated based on the training model,
not testing models, most commonly by the Gini coefficient. Refer to Hastie et al. (2014) for technical
details.

9
ADBI Working Paper 1111 Nguyen and Sagara

Table 4 reports the model performance of the traditional logistic model and machine
learning model. Logistic regression that employs some of the variables with the highest
feature importance performs slightly inferior to machine learning models on out-of-
sample data. The logistic model provides an AR of approximately 0.71 for both testing
data and back test data, lower than XGBoost in both tests and lower than random forest
in back testing.

4. CONCLUSIONS AND POLICY IMPLICATIONS

We introduced the Japan Credit Risk Database (CRD) as part of the Japanese
government’s SME financial inclusion efforts. To date, few databases exclusively for
SMEs have achieved a nationwide coverage and the volume required for robust credit
risk modeling such as the Japan CRD. The impact of CRD is threefold. First, CRD
contributes to the overall understanding of the SME sector by providing general statistics
for benchmarking, policy analysis, and evidence-based policy making (EBPM)
evaluation. Second, to encourage the adaptation of risk-based lending, CRD statistical
scoring models make SME credit risk assessment achievable and easily accessible.
Third, SME credit guarantee becomes less biased by applying nine risk-based fees
schemes, allowing SMEs to pay a guarantee fee according to their risk level.
The CRD framework is being introduced to other Asian countries with the support
of Japan International Cooperation Agency (JICA). The JICA-sponsored feasibility study
consulted by CRD Association confirms that interested parties in the Philippines have
the capacity required to establish the Philippines version of CRD. The Technical
Cooperation project begins between the two governments in the first half of 2020.
Traditionally, a CRD scoring model incorporates financial ratios derived from financial
statements into a logistic model. The output of the model is a prediction of defaults within
one to three years. While the traditional model functions well for decision making
purposes, it lacks the necessary input for analysis in cases when SMEs do not have
financial statements. To tackle this challenge, alternative data, specifically bank
transaction data, is adopted to create a new model. Generally, bank transaction data is
abundant in quantity and credible in quality. It is also relatively easy to retrieve, as the
recording of transaction data is fully automated.
The transaction model introduced in the paper was jointly developed by CRD
Association, BOJ and a Japanese megabank. The machine learning model is adopted
to process the extremely large volume of data. The model outperforms logistic models
and is highly accurate in predicting short-term default probability. It serves as an enabling
short-term monitoring instrument and a credit assessment tool for SMEs without financial
statements. Resona Bank, a Japanese megabank, has recently introduced a credit line
that only requires SMEs to have bank accounts at the bank for a certain period for loan
screening. The credit line does not require collateral, guarantee or the submission of
financial statements. 7
Approximately 6,000 variables are employed in the transaction model, applying two
popular classification modelling approaches in machine learning: random forest and
XGBoost. We find that that monthly average cash balance best explains default, followed

7 The screening for the new credit line will be done by a machine learning model, analyzing the cash
movements of the applicant’s account. Financial cost ranges from 3% to 9%. The lending amount is
capped at 10 million yen, approximately $100,000 (1USD=100Yen). News release of Resona Bank
(in Japanese): https://www.resona-gr.co.jp/holdings/news/hd_c/download_c/files/20200110_1a.pdf?_ga
=2.47077284.2056143362.1579136287-646537328.1579049610.

10
ADBI Working Paper 1111 Nguyen and Sagara

by cash inflows from revenue, cash outflows for variable costs, cash outflow for cost of
goods and services, and cash inflows from loans.
Model accuracy measured by accuracy ratio peaks at 0.707 for a random forest with 350
decision trees and 150 features, and at 0.733 for XGBoost where the tree’s maximum
depth equals four nodes and learning rate equals 0.01. Both models reach accuracy
levels general approved by the Japanese banking sector.
This paper indicates that SME financial inclusion efforts can incorporate both traditional
and alternative data. Traditional data mainly assists lending decisions, while alternative
data supports monitoring as well as credit assessment for SMEs without financial
statements. Since the establishment of a nationwide database involves many interested
parties and complex procedures, the leadership of the financial sector authority is an
important factor that helps synchronize the efforts of all parties, and directs them towards
a common goal.

11
ADBI Working Paper 1111 Nguyen and Sagara

REFERENCES
Amornkitvikai, Y., and C. Harvie. 2016. The Impact of Finance on the Performance of
Thai Manufacturing Small and Medium-Sized Enterprises. ADBI Working Paper
576. Tokyo: Asian Development Bank Institute. Date of access: 2020/04/09.
Available: http://www.adb.org/publications/impactfinance-performance-thai-
manufacturing-small-and-medium-sized-enterprises/.
Banko Sentral ng Pilipinas. 2019. Financial Inclusion Initiatives. Manila. Date of access:
2020/04/09. Available: http://www.bsp.gov.ph/downloads/Publications/
2019/microfinance_2019.pdf.
Deloitte Tohmatsu Consulting. 2018. Data utilization of SMEs and small businesses
and the optimal information website. Japan SME Agency Report. Date of
access: 2020/04/09. Available: https://www.meti.go.jp/meti_lib/report/
H30FY/000316.pdf.
Hastie T., Tibshirani R. and Friedman J. 2014. The Elements of Statistical Learning.
Springer. California.
Japan International Cooperation Agency (JICA). 2019. The Study for the Introduction of
Credit Risk Database (CRD) in the Philippines: Final Report. Date of access:
2020/04/09. Available: http://open_jicareport.jica.go.jp/pdf/12344552.pdf.
Japan SME Agency, Ministry of Economy, Trade and Industry. 2019. White Paper on
Small and Medium Enterprises in Japan. Date of access: 2020/04/09. Available:
https://www.meti.go.jp/english/press/2019/pdf/0426_010a.pdf.
Japan Federation of Credit Guarantee Corporation, Annual Report 2019. Date of
access: 2020/04/09. Available: https://www.zenshinhoren.or.jp/english/
anual2019.pdf.
Kuwahara, S., N. Yoshino, M. Sagara, and F. Taghizadeh-Hesary. 2019. Establishment
of the Credit Risk Database: Concrete Use to Evaluate the Creditworthiness of
SMEs. ADBI Working Paper 924. Tokyo: Asian Development Bank Institute.
Date of access: 2020/04/09. Available: https://www.adb.org/publications/
establishment-credit-risk-database-evaluatecreditworthiness-smes.
Maningo, G. V. 2016. Credit Surety Fund: A Credit Innovation for Micro, Small, and
Medium-Sized Enterprises in the Philippines. ADBI Working Paper 589. Tokyo:
Asian Development Bank Institute. Date of access: 2020/04/09. Available:
http://www.adb.org/publications/credit-surety-fundcredit-innovation-micro-small-
and-medium-sized-enterprises-philippines/.
Miura S., Ijitsu Y., Takekawa M. 2019. Credit Risk Assessment using Transaction
Information: An Empirical Study with AI Approach. Tokyo: Bank of Japan
Working Paper Series. Date of access: 2020/04/09. Available:
https://www.boj.or.jp/research/wps_rev/wps_2019/wp19j04.htm/.
World Bank. 2017. Global Findex Database. Date of access: 2020/04/09. Available:
https://globalfindex.worldbank.org/basic-page-overview.

12