💡 왼쪽 원문을 읽으면서 오른쪽에 따라 써보세요. Tab 키로 힌트를 받을 수 있습니다.

원문 렌더가 준비되기 전까지 텍스트 가이드로 표시합니다.

Prologue — FICO 1989 and the BNPL Era

When Fair Isaac first released the FICO Score to the public in 1989, it became the standard for US consumer credit. The score range `<300/850 FICO` set the coordinate system for mortgages, auto loans, and credit card issuance. In 2006, VantageScore arrived as the joint venture of the three big bureaus — Equifax, Experian, and TransUnion — providing an alternative scale. Almost 40 years later, in 2026, those coordinates are being redrawn under the impact of BNPL and AI underwriting.

In the US, Upstart, Affirm, Klarna, and AfterPay offer instant credit to thin-file and no-file users, and feed the outcomes back into training data. Experian Boost lets users add utility, telecom, and subscription payment history to their score by opt-in, while FICO XD builds a score purely from telecom and utility data. The FICO Resilience Index measures how well a borrower will withstand a downturn, serving as a secondary signal in stress periods.

In Korea, KCB and NICE Credit Bureau, the two officially registered credit bureaus, run a 1-1000 point scale that maps to a 1-10 grade scale. KakaoBank built its own ML score combining KakaoTalk, KakaoPay, and financial activity data, while Toss Bank cracked the thin-file market with paperless lending and its proprietary ML model.

In Japan, CIC (信用情報センター), JICC (日本信用情報機構), and KSC (全国銀行個人信用情報センター) hold credit data by sector, with retention periods of 5-30 years. The rise of BNPL — AfterPay Japan, KakaoPay 후불결제, and others — is reshaping the credit information industry itself.

This article draws that map. What is a credit score, why did it move to AI, which data predicts what, how is it validated, and how is it regulated.

1 · Traditional Scoring vs AI Underwriting

Traditional credit scoring was simple. One credit report, one scoring model. The credit report was data collected by Equifax, Experian, or TransUnion from creditors, and the score translated it into the 300-850 range.

AI underwriting splits this flow in two directions.

| Dimension | Traditional underwriting | AI underwriting |

| --- | --- | --- |

| Data | 5-factor credit report | Credit + bank + telecom + BNPL + digital signals |

| Model | Logistic regression (FICO) | GBM/XGBoost/LightGBM + neural nets |

| Decision time | Mortgage 30-45 days, card 7-14 days | Card instant, unsecured loan 5 minutes |

| Thin-file handling | Decline or require co-signer | Alternative data |

| Explainability | 4 reason codes | SHAP/LIME-driven reason codes |

| Regulation | FCRA, ECOA, Reg B | + EU AI Act, NAIC, FSS ADM |

| Pricing | Tier-averaged | Personalized, dynamic |

| Fraud | Post-hoc | Pre-emptive, real-time |

The core shift is **breadth and freshness of data**. Traditional scoring looks at a static snapshot of the credit report; AI scoring combines near-real-time signals from bank flows, telecom, BNPL, subscriptions, and social activity. As a result, no-file and thin-file users can now be scored instantly — and that is the business model of BNPL and Upstart.

2 · FICO Score 5 Factors — Payment History, Credit Utilization, Length, Mix, New Credit

The FICO Score is the weighted average of five factors. The standard consumer weights are:

| Factor | Weight | Description |

| --- | --- | --- |

| Payment history | 35% | Delinquencies, defaults, bankruptcies |

| Credit utilization | 30% | Balance vs limit (30% or less recommended) |

| Length of credit history | 15% | Average account age |

| Credit mix | 10% | Diversity of cards, installment, mortgage |

| New credit | 10% | Recent hard inquiries |

In the US, payment history is by far the dominant factor. A single 30-day delinquency can drop a score by 60-110 points, and a mortgage delinquency drops it more. Utilization moves with similar weight — running a new card at 100% can knock 30-50 points off.

The median US FICO Score in 2026 is roughly 715. Approximately 22% score 800+, 24% score 740-799, 23% score 670-739, 18% score 580-669, and 13% score below 580. 800 and above is "exceptional", 740-799 is "very good", 670-739 is "good", 580-669 is "fair", and below 580 is "poor".

3 · VantageScore — Equifax, Experian, TransUnion Joint Venture

VantageScore is the alternative score created by the three bureaus in 2006. Starting with version 4.0, it adopted trended data and a more ML-friendly architecture, and the VantageScore 5.0 released in 2024 directly integrates BNPL data.

Key differences from FICO:

- Identical 300-850 range — deliberately matched to be comparable.

- Can score someone with only one month of history (FICO requires six).

- Treats hard inquiries within a 14-day window as a single inquiry (shopping protection).

- More aggressive integration of BNPL, rent, and telecom.

When the FHFA approved VantageScore 4.0 for use in mortgages in 2024, the market share between the two scores shifted significantly. As of 2026, VantageScore reaches roughly 40-45% of FICO's usage in cards, auto, and personal lending.

4 · FICO XD and Experian Boost — Thin-File Solutions

The US has roughly 50 million thin-file consumers and 28 million no-file consumers. That's 80 million people who cannot be scored instantly by traditional models. FICO and Experian built two complementary solutions.

- **FICO XD**: A score built purely from telecom, utility, and rent payment data. Sources include LexisNexis Risk Solutions and Equifax NCTUE. The score range is 300-850. About 15 million previously no-file consumers received a score.

- **Experian Boost**: An opt-in product that lets users add telecom, electricity, gas, water, and subscription (Netflix, etc.) payments to their credit file. Average gain is 13 points, and more than 75% of users see an increase.

FICO XD style: alternative-data scoring for thin file (conceptual)

ALT_DATA_FEATURES = [

"telecom_months_active", # months of active telecom service

"telecom_late_pmt_24m", # telecom late payments in 24 months

"utility_late_pmt_24m", # utility late payments in 24 months

"rent_payment_consistency", # rent on-time ratio

"subscription_count_active", # number of active subscriptions

"checking_account_age_months", # months since checking account opened

"checking_overdraft_12m", # overdrafts in 12 months

"address_stability_months", # months at same address

]

params = {

"objective": "binary",

"metric": "auc",

"learning_rate": 0.03,

"num_leaves": 63,

}

booster = lgb.train(params, train_set, num_boost_round=1500)

The CFPB tightened guidelines on "alternative data" in 2024-2025. The key points are explicit opt-in, data accuracy, and a clear dispute process.

5 · FICO Resilience Index — Recession Resilience

The FICO Resilience Index (FRI) was introduced in 2020 after COVID. It ranges from 0-99, with lower being more resilient to downturns. Two people with the same FICO 760 can have more than 2x difference in default rate during a recession depending on FRI.

FRI looks primarily at:

- Stability of utilization patterns relative to limit

- Consistency of payment timing and amount

- Average account tenure

- Pre-downturn changes in usage

FRI is rarely used directly in lending decisions; it is mostly used in portfolio monitoring and capital simulation. Under CECL (Current Expected Credit Losses) accounting, it serves as a secondary signal in loss reserve calculations.

6 · TransUnion ECRA — Extended Credit Reporting Attribute

TransUnion's ECRA (Extended Credit Reporting Attribute) is a data mart that uses card payment data on a 24-month timeline. Instead of "current balance" or "current limit", it surfaces **trended data** — the time evolution.

Key signals:

- 24-month balance trajectory (rising, falling, volatile)

- Payment pattern (minimum vs full payment ratio)

- Limit increase or decrease events

- One-year change in average utilization

VantageScore 4.0/5.0 and FICO 10T draw heavily on this trended data. Two people with the same point-in-time score can have very different future risk profiles, and trended data captures it.

7 · KCB (Korea Credit Bureau) — Korea's 1-1000

KCB, founded in 1985, is one of Korea's two official consumer credit bureaus, alongside NICE Credit Bureau. Its score is on a 1-1000 scale, which is mapped to a 1-10 grade.

The 2026 average KCB score in Korea is around 858. The grade distribution:

| Grade | KCB Range | Share (approx.) |

| --- | --- | --- |

| 1 | 942-1000 | ~16% |

| 2 | 891-941 | ~14% |

| 3 | 832-890 | ~18% |

| 4 | 768-831 | ~16% |

| 5 | 698-767 | ~13% |

| 6 | 630-697 | ~9% |

| 7 | 530-629 | ~7% |

| 8 | 454-529 | ~4% |

| 9 | 335-453 | ~2% |

| 10 | 0-334 | ~1% |

The drivers are similar to the US — payment history, utilization, balance, length of history, and inquiries. However, Korea integrates **public data** more directly: tax delinquencies, telecom late payments, and unpaid health insurance premiums.

KCB became a MyData operator in 2021, so it can now build scores combining bank, brokerage, card, and telecom data on the basis of user consent. Internet banks like KakaoBank, Toss Bank, and KBank feed this MyData-based score into their ML models.

8 · NICE Credit Bureau — Korea's Other 1-1000

NICE Credit Bureau is the other major bureau alongside KCB. Same 1-1000 scale, same 1-10 grade mapping. But because the model weights and data sources differ, the same person can have a 30-80 point gap between KCB and NICE.

NICE's features:

- Strong card transaction data (merchant level integration)

- Faster integration of fintech and BNPL data than KCB

- Heavier presence in corporate credit through NICE Investors Service

Banks and card issuers typically **query both** KCB and NICE and apply the more conservative score. For consumers, that means managing both. Apps like Toss, KakaoPay, and NICE Jigimi provide integrated lookups.

9 · KakaoBank's In-House Score — the Cake Model

Since launching in 2017, KakaoBank has operated its own credit-scoring model. The internal name is not public, but the stack is understood.

Key data sources:

- External CB reports from KCB and NICE

- KakaoBank's own transactions (deposits, transfers, overseas card use)

- Kakao ecosystem (KakaoTalk usage signals, KakaoPay payment history)

- MyData (consented external financial data)

- Telecom (KT, SKT, LGU+ — consented)

The model uses external CB scores as a baseline and layers proprietary ML signals on top. The output is a per-user "KakaoBank credit limit" that applies to mini-loans, unsecured loans, and negative-balance accounts.

KakaoBank-style: in-house ML credit scoring (conceptual)

FEATURES = [

External CB

"kcb_score",

"nice_score",

"kcb_grade",

"nice_grade",

Internal transactions

"kakao_bank_account_age_months",

"avg_monthly_inflow_krw",

"avg_monthly_outflow_krw",

"salary_inflow_consistency",

"overdraft_12m_count",

Kakao ecosystem

"kakao_pay_txn_12m",

"kakao_pay_late_pmt_12m",

MyData

"mydata_other_bank_avg_balance",

"mydata_total_loan_balance",

"mydata_card_utilization",

]

model = xgb.XGBClassifier(

n_estimators=800,

max_depth=6,

learning_rate=0.04,

subsample=0.8,

eval_metric="auc",

)

KakaoBank's edge is **the ability to score users that external bureaus cannot**. Roughly 30% of unsecured loans go to applicants who would have been declined by traditional CB-only models.

10 · Toss Bank — Paperless Loans and In-House Scoring

Toss Bank launched as an internet-only bank in 2021. It combines payment, transfer, brokerage, and insurance data from the broader Toss (Viva Republica) ecosystem.

The hallmark is **paperless lending**. Without income documentation, a borrower's limit is computed from data in the Toss app — bank flows, spending categories, brokerage balances, card usage. External CB scores are an input, but the proprietary data is the core.

The pipeline is roughly:

1. Sign up to Toss, verify identity (via NICE identity verification).

2. Opt in to MyData (bank, card, brokerage, insurance).

3. Combine with Toss-internal activity (transfers, brokerage, insurance auto-debits).

4. Run the ML model to compute limit and rate.

5. Provide an appeal path (per FSS ADM guideline).

Toss Bank gained 6 million users in its first year, and loan balances rose quickly. During 2023-2024 its NPL ratio rose temporarily, exposing the limits of ML underwriting, and the bank later tuned the model and adjusted limit policy.

11 · Japan CIC (信用情報センター) — 30 Years of Card and Installment

CIC (株式会社シー・アイ・シー) is one of Japan's three credit information centers. It holds data on credit cards, installment plans, and BNPL. Members of the Japan Credit Association (JCA) report data to CIC.

CIC data covers:

- Applications, contracts, and payment history for cards, installment plans, BNPL

- Default, bankruptcy, and legal proceedings

- Identity (name, date of birth, address, phone)

Retention windows:

- Applications: 6 months

- Contract and payment history: 5 years after contract end

- Default and legal proceedings: 5-10 years

- Anonymized statistics: up to 30 years

Unlike the US and Korea, Japan does not have a standardized "credit score". Instead, raw data is delivered to lenders and BNPL operators, who compute their own scores. So in Japan, the question is not "what is the standard FICO equivalent?" but "how does each lender turn CIC data into its own internal model?".

12 · JICC (日本信用情報機構) and KSC (全銀協) — Consumer Finance and Banks

JICC (Japan Credit Information Reference Center) members are consumer finance companies, some card issuers, and a few credit firms. While CIC is card- and installment-centric, JICC is consumer-loan-centric.

KSC (全国銀行個人信用情報センター) is the credit information center of the Japanese Bankers Association. Banks are direct members, and it holds mortgage, auto, and unsecured loan data.

Japan's credit market is split across these three agencies, and members cannot freely query other bureaus. The same user's data is therefore fragmented. Some lenders join multiple bureaus, but it costs more.

-- Japan CIC/JICC/KSC integrated lookup — conceptual underwriting mart

SELECT

a.applicant_id,

a.dob,

a.residence_years,

a.employment_years,

-- CIC: cards, installment, BNPL

cic.card_count,

cic.card_late_pmt_24m,

cic.bnpl_active_count,

cic.bnpl_late_pmt_12m,

-- JICC: consumer finance, unsecured loans

jicc.consumer_loan_balance_jpy,

jicc.consumer_loan_late_pmt_24m,

-- KSC: bank loans

ksc.mortgage_balance_jpy,

ksc.auto_loan_balance_jpy,

ksc.bank_late_pmt_24m

FROM applications a

LEFT JOIN cic_data cic ON cic.applicant_id = a.applicant_id

LEFT JOIN jicc_data jicc ON jicc.applicant_id = a.applicant_id

LEFT JOIN ksc_data ksc ON ksc.applicant_id = a.applicant_id

WHERE a.created_at >= '2026-05-01'

AND a.product_code = 'CARD_LOAN_JPY'

The Japanese FSA started a partial cross-bureau sharing pilot in 2022. As of 2026, some information (default, bankruptcy) is shared across the three, but everyday payment data remains separated.

13 · BNPL Part 1 — Affirm, the US Standard

Affirm, founded in 2012 by PayPal co-founder Max Levchin, is one of the standard BNPL operators in the US. It went public on the NYSE in 2021. The core idea is "installments without a credit card", and key partners include Amazon, Walmart, Peloton, and Shopify.

Affirm's distinguishing traits:

- A mix of **0% APR and interest-bearing** options, chosen at checkout.

- Unlike credit cards, the limit is **evaluated transaction by transaction**. Every purchase is underwritten.

- Affirm reports can affect **credit reports**: TransUnion and Experian integrate Affirm data.

- No hard collection on delinquency in many products. Future approval is denied instead.

Affirm's underwriting is ML-based. It combines applicant data (name, DOB, SSN tail, address) with the credit report to produce a risk grade, then underwrites instantly based on merchant, amount, and term. Affirm funds itself through ABS (asset-backed securities) to maintain a balanced business model.

14 · BNPL Part 2 — Klarna, From Sweden to the World

Klarna (2005, Sweden) is the global #1 BNPL operator. It runs in 45 countries across Europe, the US, and Australia, and completed an IPO in 2024-2025.

Klarna's product menu:

- **Pay in 4**: 4 installments over 6 weeks, interest-free.

- **Pay later 30**: lump-sum payment in 30 days.

- **Financing**: 6-36 month installment plans with interest.

- **Klarna One**: a unified card combining Pay in 4 with regular debit.

Klarna blends in-house ML underwriting with a global data pool. Decisions come back in under 5 seconds at checkout, and decline rates range from 5-15% by category and country.

On the regulatory side, the UK FCA classified BNPL as a regulated credit product in 2024, and the EU Consumer Credit Directive 2.0 covers BNPL starting in 2026. The US CFPB declared BNPL similar to credit cards for consumer protection purposes in 2024.

15 · BNPL Part 3 — AfterPay, Zilch, Klarna One, AfterPay Japan

- **AfterPay** (2014, Australia): Acquired by Square (now Block) in 2021. Pay-in-4 centric. US, Australia, New Zealand, Canada, UK. About 24 million active users cumulatively in 2026. Entered Japan as AfterPay Japan, integrated into unmanned-store and EC checkout.

- **Zilch** (2018, UK): Pay-in-4 plus a rewards card. Europe-focused.

- **Klarna One**: Klarna's unified card launched in 2024. Apple Pay and Google Pay compatible. Pay in 4 or regular debit, selected at checkout.

- **KakaoPay 후불결제** (Korea): Launched in 2022. Subscribers to KakaoPay can pay later without a credit card. The limit is computed from KakaoBank's credit scoring.

- **Toss Card** (Korea): The Toss Bank check card. It is not a credit card, but it is linked to Toss Bank's credit limit, producing a BNPL-like experience.

`$300B BNPL market` is the estimated 2026 global BNPL transaction volume (SimilarWeb, FIS, etc.). Up roughly 6x from `$50B` in 2020. The split is approximately US 30%, Europe 25%, Asia-Pacific 35%, Other 10%.

16 · Upstart — The Symbol of AI Underwriting

Upstart (2012, California) is a fintech built on ML credit underwriting. NASDAQ IPO in 2020. Cumulative originations have exceeded `$30B+`, and Upstart provides the underwriting engine to over 100 partner banks via API.

Upstart's differentiators:

- Over **1,600 variables** in the model (FICO uses about 20).

- Education, major, occupation, employer, and city as alternative data inputs.

- Claims to raise approval rates by 70%+ at the same default level as FICO alone.

- Delivers underwriting as an API to partner banks.

The regulatory issue is the disparate impact of alternative data. The CFPB issued a No-Action Letter to Upstart in 2017 and has continuously monitored fair lending since. Variables like education and major can be proxies for race and income, so Upstart publishes fair-lending audit results on a regular cadence.

17 · The AI Underwriting Algorithm Stack

In 2026, the workhorse of credit modeling remains gradient boosting. Neural nets play a secondary role in processing text and time-series signals.

Selection criteria:

- Tabular data, mid-scale (hundreds of thousands to hundreds of millions of rows): GBM is almost always faster and more accurate than neural nets.

- Text, time series, images: CNN and transformer architectures are strong (transaction-memo analysis, receipt OCR).

- Explainability is paramount: logistic regression or GAM is safe. Pairing SHAP with GBM is the second-best.

| --- | --- | --- | --- |

18 · An XGBoost Credit Model — Conceptual Code

XGBoost remains the standard in 2026. The shape of a typical underwriting model:

XGBoost credit underwriting (conceptual)

from sklearn.model_selection import StratifiedKFold

FEATURES = [

Credit report

"credit_history_length_months",

"open_accounts_count",

"credit_utilization_pct",

"late_pmt_24m",

"hard_inquiries_12m",

"bankruptcy_flag",

Bank (MyData)

"avg_monthly_inflow",

"avg_monthly_outflow",

"overdraft_12m",

"salary_consistency_score",

BNPL

"bnpl_active_count",

"bnpl_late_pmt_12m",

Alternative data

"telecom_late_pmt_24m",

"utility_late_pmt_24m",

"rent_payment_consistency",

]

params = {

"objective": "binary:logistic",

"eval_metric": "auc",

"max_depth": 6,

"learning_rate": 0.04,

"n_estimators": 1200,

"subsample": 0.85,

"colsample_bytree": 0.85,

"scale_pos_weight": 6.0, # default is the minority class

}

dtrain = xgb.DMatrix(X_train, label=y_train)

bst = xgb.train(params, dtrain, num_boost_round=1200)

pd_score = bst.predict(xgb.DMatrix(X_test))

Operationally, the key is **not to use raw probabilities**. The model output (default probability) is mapped to a score, and score bands map to limits and rates. That mapping is the pricing policy, and it changes frequently.

19 · Adverse Action Notices — A Duty on Decline

The US ECOA (Equal Credit Opportunity Act) and FCRA (Fair Credit Reporting Act) require an **adverse action notice** when credit is denied or offered on unfavorable terms. Key items:

- Statement of decline (or unfavorable terms)

- Specific reasons (typically four or more reason codes)

- If a credit report was used: bureau name, address, and toll-free number

- Right to obtain a free credit report

- Creditor contact and dispute process

[Adverse Action Notice — conceptual English template]

Dear [Applicant Name],

We regret to inform you that we are unable to approve your application

dated [YYYY-MM-DD] for [Product Name].

Principal reasons for this decision:

1. Insufficient credit history length

2. High credit utilization on existing accounts

3. Recent delinquencies within the past 24 months

4. Multiple recent credit inquiries

This decision was based in whole or in part on information obtained

from the following consumer reporting agency:

[CRA Name, Address, Toll-Free Number]

The agency did not make the credit decision. You have the right to

obtain a free copy of your consumer report within 60 days, and the

right to dispute inaccurate or incomplete information.

If you have questions, please contact us at [Creditor Contact].

Sincerely,

[Creditor Name]

Korea imposes **similar duties** under the FSS ADM (Automated Decision-Making) guideline. Reason for decline, human-review escalation path, and dispute process are all mandatory. Japan's 2025 financial AI guideline follows the same pattern.

20 · CFPB Reg B and Disparate Impact

CFPB Reg B (the implementation of ECOA) prohibits discrimination in credit. Core concepts:

- **Disparate treatment**: explicit use of protected attributes (race, sex, religion, national origin, marital status, age, receipt of public assistance, etc.). Almost always illegal.

- **Disparate impact**: not explicitly used, but the outcome disadvantages a protected group. Burden to prove business necessity rests with the creditor.

- **Proxy discrimination**: using a variable strongly correlated with a protected attribute (zip code, last name, school) so that discrimination occurs in effect.

ML models raise large disparate-impact concerns. Alternative data like education, employment, and social media can be a strong proxy for race or income. The CFPB clarified reason-code duties for ML-based decisions in its 2023 "Adverse Action Notices for AI-based Decisions" guidance.

Standard validation patterns:

- Demographic parity: gap in approval rate across protected groups

- Equal opportunity: gap in TPR (true positive rate) across protected groups

- Calibration by group: whether default rate at the same score is consistent across groups

21 · The Korean FSS ADM Guideline

The Korean FSS released the "Financial AI Guideline" in 2024, followed by the "ADM (Automated Decision-Making) Guideline" in 2025. It applies to credit, insurance, and automated investment advice.

Core duties:

1. Impact assessment: consumer impact assessment before deploying an AI system.

2. Explainability: at least five reason codes for decline or high-risk decisions.

3. Human-review path: a mandatory channel to request human review of an automated denial.

4. Dispute resolution: response within 30 days of a complaint.

5. Recordkeeping: input, output, and model version retained for 5 years.

6. Periodic review: model performance and bias reviewed and reported quarterly.

7. Incident reporting: model errors must be reported to the FSS within 5 business days.

Toss, KakaoBank, KBank, Shinhan Card, and KB Kookmin Card all operate ML underwriting under this guideline. Violations attract corrective orders and fines, and some items overlap with the Personal Information Protection Act and the Credit Information Use and Protection Act.

22 · Japan — FSA AI Guideline and BNPL

The Japanese FSA released "金融分野におけるAIガバナンスガイドライン" in 2025, modeled on EU AI Act and NAIC.

Key requirements:

- Risk classification of AI systems (high, medium, low risk)

- Governance committee for high-risk systems (credit scoring, insurance underwriting)

- Transparency (透明性) duty

- Human-review path

- Periodic audits

BNPL is regulated under the revised Installment Sales Act (割賦販売法) from 2022. BNPL above certain thresholds is classified as installment sale, triggering mandatory CIC registration and credit assessment duties. AfterPay Japan, Mercari postpay (メルカリ後払い), and Amazon postpay (後払い) operate under this regime.

23 · BNPL Market Comparison — Global vs Korea vs Japan

The `$300B BNPL market` figure is global, but regional structures vary widely.

| --- | --- | --- | --- | --- |

| Avg ticket size | `$150-200` | EUR 120-180 | KRW 50,000-80,000 | JPY 8,000-12,000 |

Korean BNPL is anchored by KakaoPay postpay, Naver Pay postpay, and Toss postpay; volume is smaller than US/EU, but growth is the fastest. Japan's BNPL is spreading via unmanned-store payments and EC checkout.

24 · Comprehensive Comparison — US/EU/KR/JP Regulatory Matrix

| Item | US | EU | KR | JP |

| --- | --- | --- | --- | --- |

25 · Operational Checklist — Before Launching an AI Credit Model

A practical checklist.

1. **Document data lineage**: source, license, consent record for training and validation data.

2. **Bias review**: demographic parity, equal opportunity, calibration by group.

3. **Proxy scan**: SHAP impact of strong proxies (zip code, education, social).

4. **Explainability**: standardized mapping of 4-5 reason codes.

5. **Human-review path**: escalation workflow for declines and high-risk cases.

6. **Champion/challenger**: launch new models on 5-10% traffic.

7. **PSI/AUC monitoring**: data and performance drift alerts.

8. **Retraining cadence**: quarterly or semi-annual.

9. **Vendor governance**: SLA and test results for third-party scores and models.

10. **Incident response**: SOP for refunds and reprocessing when the model errs.

11. **Logging**: retain inputs, outputs, and model versions per decision.

12. **Consumer disclosure**: disclose AI use and provide the appeal path.

13. **Regulatory reporting**: schedules and formats for CFPB, EU, FSS, FSA.

14. **Capital impact**: simulate Basel III, CECL, K-ICS, and Japan ICS impacts.

These 14 are the 2026 baseline. Above all, **efficiency and accountability must weigh the same**. Upstart's 70% approval-rate lift and the instant decisions of BNPL are marketing copy; behind them are reason codes, fair-lending audits, and human-review paths.

26 · Conclusion — The Next Ten Years of the Score

The credit scoring of 2026 has moved beyond the era where one score decided everything. We are in the era of **decision systems built from multiple scores, models, and data**. FICO and VantageScore are still the standard, but on top of them sit the ML underwriting of Upstart, Affirm, and Klarna. Korean KCB and NICE are evolving with MyData and fintech data, and Japan's CIC, JICC, and KSC are expanding through BNPL and digital payments.

Keywords for the next decade:

1. **Standardization of alternative data**: telecom, utilities, rent, and BNPL may become routine entries in credit reports.

2. **Cross-border score compatibility**: as global BNPL grows, attempts at cross-border score integration will increase.

3. **Convergence of AI governance**: CFPB, EU AI Act, FSS, and FSA core duties are converging.

4. **Personal data rights**: MyData, GDPR, and CCPA are strengthening the right to control one's score.

5. **Standardization of explainability**: tools like SHAP and LIME become the standard for reason-code generation.

Credit scoring is not merely a measure of "can you borrow money". It determines which resources individuals can reach in society. The fairness and explainability of the score are therefore not just regulatory duties — they are basic infrastructure of digital society.

References

- Fair Isaac (FICO) — [https://www.fico.com/](https://www.fico.com/)

- VantageScore — [https://vantagescore.com/](https://vantagescore.com/)

- Experian Boost — [https://www.experian.com/consumer-products/score-boost.html](https://www.experian.com/consumer-products/score-boost.html)

- CFPB — Consumer Financial Protection Bureau — [https://www.consumerfinance.gov/](https://www.consumerfinance.gov/)

- CFPB Reg B (ECOA) — [https://www.consumerfinance.gov/rules-policy/regulations/1002/](https://www.consumerfinance.gov/rules-policy/regulations/1002/)

- FTC FCRA — [https://www.ftc.gov/legal-library/browse/statutes/fair-credit-reporting-act](https://www.ftc.gov/legal-library/browse/statutes/fair-credit-reporting-act)

- Upstart — [https://www.upstart.com/](https://www.upstart.com/)

- Affirm — [https://www.affirm.com/](https://www.affirm.com/)

- Klarna — [https://www.klarna.com/](https://www.klarna.com/)

- AfterPay (Block) — [https://www.afterpay.com/](https://www.afterpay.com/)

- Zilch — [https://www.zilch.com/](https://www.zilch.com/)

- KCB (Korea Credit Bureau) — [https://www.koreacb.com/](https://www.koreacb.com/)

- NICE Credit Bureau — [https://www.niceinfo.co.kr/](https://www.niceinfo.co.kr/)

- KakaoBank — [https://www.kakaobank.com/](https://www.kakaobank.com/)

- Toss Bank — [https://www.tossbank.com/](https://www.tossbank.com/)

- KakaoPay — [https://www.kakaopay.com/](https://www.kakaopay.com/)

- Korea Financial Supervisory Service (FSS) — [https://www.fss.or.kr/](https://www.fss.or.kr/)

- Korea Credit Information Services — [https://www.kcredit.or.kr/](https://www.kcredit.or.kr/)

- CIC 信用情報センター — [https://www.cic.co.jp/](https://www.cic.co.jp/)

- JICC 日本信用情報機構 — [https://www.jicc.co.jp/](https://www.jicc.co.jp/)

- KSC 全国銀行個人信用情報センター — [https://www.zenginkyo.or.jp/pcic/](https://www.zenginkyo.or.jp/pcic/)

- 日本クレジット協会 (Japan Credit Association) — [https://www.j-credit.or.jp/](https://www.j-credit.or.jp/)

- Japan FSA (金融庁) — [https://www.fsa.go.jp/](https://www.fsa.go.jp/)

- EU AI Act — [https://artificialintelligenceact.eu/](https://artificialintelligenceact.eu/)

- SHAP — [https://shap.readthedocs.io/](https://shap.readthedocs.io/)

- LIME — [https://github.com/marcotcr/lime](https://github.com/marcotcr/lime)

- XGBoost — [https://xgboost.readthedocs.io/](https://xgboost.readthedocs.io/)

- LightGBM — [https://lightgbm.readthedocs.io/](https://lightgbm.readthedocs.io/)