All topics
capitalevergreen
Research-ready episode

How Lending Learned to See You

放贷如何学会看穿你

What this research must answer

Lending is a four-thousand-year-old machine for deciding who will pay. Every era builds a smarter version — collateral, reputation ledgers, scores, algorithms, collections, packaged debt — and each machine solves one information problem while creating the next. The episode follows that machine from temple loans and pawnshops, through the 1841 credit ledgers and the 1958 FICO score, into the unregulated frontier where a loan decision takes seconds and the daily interest has no ceiling. It ends at the machine's two great blind spots: total debt that no single lender can see, and bad debt that never dies — it is sold, repackaged, and held by someone who never priced it.

放贷是一台存在了四千年的机器，用来决定谁会还钱。每一代人都造出更聪明的版本——抵押、声誉账本、评分、算法、催收、打包债券——而每台机器解决一个信息问题的同时，又制造出下一个。影片跟着这台机器，从寺庙贷款和当铺出发，穿过 1841 年的手写征信账本和 1958 年的 FICO 分数，进入监管真空的前沿：几秒批贷，日利率却没有上限。结尾落在机器两大盲区上：任何单一放贷人都看不见的总债务，以及永远死不了的坏账——它被卖掉、打包，最后落在某个从没给风险定价的人手里。

Opening tension

In Wonogiri, Central Java, a 38-year-old housewife borrowed from twenty-three phone lending apps. Police believe the harassment from those lenders — calls to her contacts, threats, shaming — drove her to take her own life. Every single app approved a small loan. None of them saw the other twenty-two. That is the story of lending in one screen: a machine built to decide who will pay, and what it cannot see.

Opening variants
  1. ·

    The twenty-three apps

    A housewife in Wonogiri borrowed from 23 phone lending apps; each approved a small loan and none saw the other 22. Harassment from the lenders is believed to have driven her suicide.

    Selected opening: the machine's central blind spot — invisible total debt — is visible in one life before any explanation begins.

  2. ·

    The senator's forty-eight percent

    Roman law capped interest at 12 percent in Egypt. In 44 BCE, Marcus Junius Brutus still lent the city of Salamis money at 48 percent a year — because the law was only as strong as the person willing to enforce it.

    Strongest history hook: the rate-cap war is two thousand years old, and the loophole was there from the start.

  3. ·

    The ledgers read aloud

    In 1841, Lewis Tappan's clerks wrote credit reports by hand, and subscribers gathered to hear them read aloud. Reputation became a record, a record became a number, and the number learned to see you.

    Best mechanism hook: makes the risk-recognition machine physical and shows how old the trade in trust really is.

How the episode moves
  1. 01

    What turns a loan into a trap?

    Wonogiri, Central Java: a 38-year-old housewife borrowed from 23 illegal phone lending apps; police believe harassment from the lenders drove her suicide. Each app saw a small, safe loan; none saw the other twenty-two. Lending looks like a product, but it is a prediction business — and prediction needs information.

    Tempo — OJK comments on illegal online lending-related suicide: https://en.tempo.co/read/1514449/ojk-comments-on-illegal-online-lending-related-suicide

    How did lenders decide who to trust before there was any data?

  2. 02

    How did lenders trust strangers before data?

    For most of history the answer was price, collateral, and law. Roman Egypt capped cash-loan interest at 12 percent a year — cut from 24 percent before the Roman conquest — yet in 44 BCE Brutus lent the city of Salamis money at 48 percent. England's 1571 statute legalized lending at a 10 percent ceiling, lowered to 8 percent in 1624. The church banned interest outright, so money hid inside annuities and triple contracts — Zinskauf, rentes, contractum trinius — and pawnshops, from ancient China to the charitable Monte di pietà, lent against things instead of people.

    History Today — Roman Loans: https://www.historytoday.com/print/pdf/node/55404/debug; Bogazici digital archive — Brutus at 48 percent: https://digitalarchive.library.bogazici.edu.tr/server/api/core/bitstreams/cb278ad4-320e-4355-91ba-b78a318de8fe/content; UToronto — England 1571 usury statute: https://www.economics.utoronto.ca/index.php/index/research/workingPaperDetails/439

    If price and collateral are blunt instruments, what else can a lender look at?

  3. 03

    What else can a lender look at?

    The record. In 1841 Lewis Tappan's Mercantile Agency began writing credit reports by hand; subscribers heard them read aloud, and reputation became a file. In 1958 Fair, Isaac sold the first scorecard, turning the file into a number, and in 1970 the Fair Credit Reporting Act gave the number rules. The machine could finally see a stranger — but only if the stranger had a record: 26 million American adults are credit invisible and 19 million more are unscorable because their files are too thin.

    Harvard Baker Library — The Mercantile Agency: https://www.library.hbs.edu/hc/credit/credit3b.html; MarketRealist — Why credit scores were created (FICO 1958): https://marketrealist.com/p/why-were-credit-scores-created/; FTC — 50 years of the FCRA: https://www.ftc.gov/business-guidance/blog/2020/10/50-years-fcra

    What happens when there is no record to score?

  4. 04

    What happens when there is no record to score?

    The lender prices the unknown or refuses. This is the real business of lending: the rate must cover funding cost, operations, expected default, and margin — probability of default, loss given default, exposure at default — so more uncertainty means a higher price or a refusal. Behind the score sits the back office that keeps the machine alive: provisioning for expected losses, collections, write-offs, and capital floors set by Basel. Scoring is the front end; bad-debt management is the engine.

    BIS — Basel framework, credit risk components PD/LGD/EAD: https://www.bis.org/basel_framework/chapter/CRE/32.htm

    What does the machine fail to see?

  5. 05

    What does the machine fail to see?

    The borrower's total debt. From 2014 to 2016, 2.7 million Kenyans were negatively listed by credit reference bureaus — about a tenth of the adult population — including roughly 400,000 who defaulted on digital loans under two dollars. A third of Kenyan digital borrowers were juggling loans from two or more providers at once. Every lender models its own loan; no single model sees the other twenty-two. And yet the same instrument helps: M-Shwari loans measurably improved Kenyan households' ability to absorb shocks like illness or school fees.

    MicroSave — Setting digital credit right (2.7M blacklisted): https://www.microsave.net/2017/08/22/setting-digital-credit-right-is-it-time-for-a-major-re-think/; CGAP/FSD — Digital Credit Revolution: https://www.mfw4a.org/sites/default/files/resources/A_Digital_Credit_Revolution_-_Insights_from_Borrowers_in_Kenya_and_Tanzania.pdf; NBER 25604 — M-Shwari and household resilience: https://www.nber.org/papers/w25604

    What happens where nobody caps the price of that blindness?

  6. 06

    What happens where nobody caps the price?

    The fee becomes the harvest. Indonesia's legal cap for short-term consumer loans is 0.2 percent a day, yet the competition authority is investigating 97 digital lenders for jointly setting rates around 0.8 percent a day; one borrower who took 9.4 million rupiah was told to repay 18 to 19 million. China's 714 high-cannon loans run 7 or 14-day cycles at crushing effective rates until a state broadcaster exposed them. In the United States, payday loans average 391 percent APR, 80 percent are rolled over, and 75 percent of fees come from borrowers who take ten or more loans a year. In the Philippines, apps shamed defaulters to their contact lists until regulators pulled them down.

    Tempo — The online lending trap (0.8 percent a day, KPPU): https://en.tempo.co/read/2027195/the-online-lending-trap; Kompas — Rp9.4 million borrowed, 18-19 million owed: https://www.kompas.id/artikel/en-pinjol-lintah-darat-yang-dilegalkan; Business Insider — Pew, 12 million Americans a year, 75 percent of fees: https://www.businessinsider.com/how-small-dollar-lending-flexible-pay-could-challenge-payday-lending-2021-6

    When the price is capped, does protection or exclusion win?

  7. 07

    When the price is capped, does protection or exclusion win?

    Kenya capped interest rates in September 2016 to protect borrowers; credit to small and medium firms collapsed, with SMEs' share of bank lending falling from about 25 to 15 percent, and the cap was repealed in November 2019. China's 2020 private-lending cap moved from a 24/36 percent two-line rule to four times the one-year LPR — about 15.4 percent at the time — and the same lever that protects the borrower prices the riskiest ones out. Protection and exclusion are the same dial.

    Alper & Clements — Do interest rate controls work? Evidence from Kenya: https://www.semanticscholar.org/paper/Do-interest-rate-controls-work-Evidence-from-Kenya-Alper-Clements/adea279b3cefdf1d6a2e0657348aeabcf7ef6424; Sina Finance — China private-lending cap 4x LPR: https://finance.sina.com.cn/money/bank/bank_hydt/2020-09-09/doc-iivhvpwy5765632.shtml

    What happens to the bad debt the machine finally sees?

  8. 08

    What happens to bad debt the machine finally sees?

    It is sold, and then packaged. Debt buyers purchase charged-off paper at five to nine cents on the dollar and recover twelve to eighteen cents over two to three years; two of the largest buyers, Encore Capital and Portfolio Recovery, have bought the rights to collect more than 200 billion dollars in defaulted consumer debt. The CFPB logged about 207,800 debt-collection complaints in 2024 alone. Above the collectors sits packaging: Ant Group's Huabei and Jiebei microloans dominated consumer-loan asset-backed securities — roughly 85 percent of that market in 2017, 269.2 billion yuan issued cumulatively — until regulators stopped their ABS and then the company's 37-billion-dollar IPO. In 2008, AAA-rated tranches were found to contain mortgages that had never been priced. Bad debt does not disappear; it moves to whoever did not price it.

    ABI — zombie debt buyers Encore and Portfolio Recovery: https://www.abi.org/feed-item/have-you-been-harassed-by-debt-collectors-like-midland-funding-or-portfolio-recovery; Orrick — CFPB FDCPA annual report 2024: https://infobytes.orrick.com/2025-12-04/cfpb-publishes-annual-report-on-fdcpa-findings-from-2024/; Caixin — Ant consumer-loan ABS approved: https://www.caixinglobal.com/2018-01-24/ant-financial-gains-approval-to-issue-securities-backed-by-consumer-loans-101201842.html; Bank Underground — as safe as houses, 2008: https://bankunderground.co.uk/2018/09/17/as-safe-as-houses-how-a-small-corner-of-the-us-mortgage-market-nearly-brought-down-the-global-financial-system/

    And the newest machines — algorithms and AI — who do they see, and who owns the record?

  9. 09

    Who sees you now — and who owns the record?

    Algorithms learned to see people who never had a record: phone-usage patterns predict repayment, M-Shwari lends in seconds, and M-KOPA replaces collateral with a kill switch that locks a financed device when a payment is missed. But the new machine brings the old questions back. An Apple Card algorithm gave a man twenty times his wife's limit despite her higher score, and New York regulators opened a probe. China stopped Ant's IPO two days before listing and rewrote microloan rules; India built the opposite — the Account Aggregator, a consent-based rail through which a borrower authorizes a lender to read verified data, powering 1.67 trillion rupees in loan disbursals in 2025. The ending returns to Wonogiri: the machine is not neutral and not stupid. It is designed by someone, and whoever designs the experiment decides what it is allowed to measure — and who is measured out of the market entirely.

    Björkegren & Grissen — Behavior revealed in mobile phone usage predicts credit repayment: https://documents1.worldbank.org/curated/en/811881575657172759/pdf/Behavior-Revealed-in-Mobile-Phone-Usage-Predicts-Credit-Repayment.pdf; NYT — Apple Card investigation: https://www.nytimes.com/2019/11/10/business/Apple-credit-card-investigation.html; Straits Times — Ant IPO suspended: https://www.straitstimes.com/business/companies-markets/china-suspends-ant-groups-mega-ipo-just-2-days-before-listing; BusinessWorld — India Account Aggregator, Rs 1.67 trillion: https://www.businessworld.in/article/consent-based-finance-pushes-india-toward-rs-1-67-tn-loans-574795

    That is the ending: the machine's blind spots are design choices, not bugs.

Story bank
  1. Opening pressure

    The twenty-three apps

    Wonogiri, Central Java: a 38-year-old housewife borrowed from 23 illegal phone lending apps; police believe lender harassment drove her suicide. Each app approved a small loan; none saw the other 22.

  2. Historical turn

    The senator's forty-eight percent

    Roman law capped interest at 12 percent in Egypt, yet in 44 BCE Brutus lent Salamis money at 48 percent. The rate cap is two thousand years old, and so is the loophole.

  3. Historical turn

    The church's hidden interest

    Medieval bans on usury pushed money into annuities and triple contracts — Zinskauf, rentes, contractum trinius — while pawnshops, from ancient China to the Monte di pietà, lent against things rather than people.

  4. Historical turn

    The ledgers read aloud

    In 1841 Lewis Tappan's Mercantile Agency wrote credit reports by hand; subscribers heard them read aloud. Reputation became a record, and in 1958 FICO turned the record into a number.

  5. Mechanism

    Scoring is the front end

    The real business of lending is bad-debt management: PD/LGD/EAD, provisioning, collections, write-offs, and Basel capital floors. The rate covers expected loss, so uncertainty becomes price — or refusal.

  6. Conceptual reversal

    The two-dollar default

    2.7 million Kenyans were blacklisted between 2014 and 2016 — including 400,000 who defaulted on digital loans under two dollars. A third were juggling multiple loans. No lender saw the total.

  7. Application decision

    The cap that cut credit

    Kenya capped rates in 2016; SME credit share fell from about 25 to 15 percent, and the cap was repealed in 2019. China moved to a 4x LPR private-lending cap in 2020. Protection and exclusion are the same dial.

  8. Application decision

    The harvest without a ceiling

    Indonesia's legal cap is 0.2 percent a day; the competition authority is probing 97 lenders around 0.8 percent. Payday loans average 391 percent APR in the US, and 75 percent of fees come from borrowers taking ten or more loans a year.

  9. Countercase

    The shock that was smoothed

    M-Shwari loans measurably helped Kenyan households absorb shocks; Grameen built social collateral; phone data genuinely predicts repayment. The experiment works — its blind spots are the problem.

  10. Frontier and return

    Who owns the record

    Ant's microloans dominated consumer-loan ABS — 85 percent of the market, 269.2 billion yuan — until regulators stopped the ABS and the 37-billion-dollar IPO. India built the opposite: a consent-based Account Aggregator rail. The ending reinterprets Wonogiri: the machine's blind spots are design choices.

How the system works
Input
A borrower with a thin or missing record, alternative signals (phone usage, transactions), the borrower's other debts that no single lender sees, and the lender's own cost of funds.
Transformation
A risk model converts signals into a predicted default probability and loss; underwriting maps that to a limit, price, and term; repayment updates the record and the next decision; overdue loans move through provisioning, collections, write-off, and sale or securitization.
Output
An approval or refusal, a priced loan, a manufactured credit history, and — for the industry — a portfolio whose bad debt can be sold and repackaged.
Limit
Every machine sees only its own loan, never the borrower's total debt; when the price has no ceiling, the interest itself becomes the harm; and packaged bad debt eventually sits with someone who never priced it.
Mechanism cards

  1. Lending as risk pricing

    Input
    A loan request with unknown repayment probability and a lender's cost of funds.
    Transformation
    The rate must cover funding cost, operations, expected default, and margin; more information lowers the uncertainty that must be priced in.
    Output
    A price and a limit for the loan.
    Limit
    When information is absent, the price rises or the loan is refused; high prices select only the most desperate borrowers.
    Evidence
    Harvard Law Review 2018 — payday, vehicle title, and high-cost installment loans: https://harvardlawreview.org/print/vol-131/payday-vehicle-title-and-certain-high-cost-installment-loans/

  2. Collateral substitutes

    Input
    A borrower with no property to pledge; a financed asset that can be controlled.
    Transformation
    The pawnshop prices the thing, not the person; M-KOPA locks a financed device remotely when a payment is missed.
    Output
    Credit backed by control of an object instead of trust in a person.
    Limit
    Collateral only works for people who own something, and lockout collects the payment by taking the good away.
    Evidence
    Britannica — pawnbroking: https://www.britannica.com/money/pawnbroking; Harvard Baker — Monte di pietà: https://www.library.hbs.edu/hc/credit/credit2c.html

  3. Reputation records

    Input
    Merchants' scattered impressions of distant trading partners.
    Transformation
    Mercantile Agency clerks turn reports into hand-written ledgers read aloud to subscribers; reputation becomes a file.
    Output
    A record that travels further than the person.
    Limit
    The record is only as good as the reporter, and it leaves out everyone never written down.
    Evidence
    Harvard Baker Library — The Mercantile Agency: https://www.library.hbs.edu/hc/credit/credit3b.html

  4. Credit scoring

    Input
    Payment history, credit use, account ages, inquiries — a record.
    Transformation
    A statistical model converts the record into a standardized score; lenders map the score to decisions.
    Output
    A number the whole industry reads.
    Limit
    The score needs a history: 26 million Americans are credit invisible and 19 million are unscorable because their files are too thin.
    Evidence
    MarketRealist — FICO 1958: https://marketrealist.com/p/why-were-credit-scores-created/; VantageScore — credit invisible: https://vantagescore.com/consumers/blog/financial-product-could-help-consumers-build-credit

  5. Bad-debt management

    Input
    A loan that missed a payment, and the lender's capital.
    Transformation
    Expected loss is priced and provisioned; overdue loans move through collections, write-off, and eventual sale — with Basel capital floors holding the bank's skin in the game.
    Output
    A portfolio whose losses are planned, not discovered.
    Limit
    The plan only sees its own portfolio; systemic, correlated, or invisible debt still arrives as a surprise.
    Evidence
    BIS — Basel framework, PD/LGD/EAD: https://www.bis.org/basel_framework/chapter/CRE/32.htm

  6. Collections economy

    Input
    Charged-off debt that the original lender no longer wants to chase.
    Transformation
    Debt buyers purchase portfolios at five to nine cents on the dollar and recover twelve to eighteen cents over two to three years; zombie debt keeps collecting decades later.
    Output
    A market where other people's bad luck is an asset class.
    Limit
    The recovery machine leans on legal process and pressure — CFPB logged about 207,800 debt-collection complaints in 2024.
    Evidence
    ABI — Encore and Portfolio Recovery, $200 billion: https://www.abi.org/feed-item/have-you-been-harassed-by-debt-collectors-like-midland-funding-or-portfolio-recovery; Orrick — CFPB FDCPA report: https://infobytes.orrick.com/2025-12-04/cfpb-publishes-annual-report-on-fdcpa-findings-from-2024/

  7. Securitization and packaging

    Input
    A pool of loans — microloans, mortgages, consumer credit — with different default risks.
    Transformation
    The pool is tranched and sold as bonds; the highest-rated tranche is paid first and priced as safe.
    Output
    Liquid bonds backed by thousands of small promises.
    Limit
    Packaging moves risk off the originator's books; in 2008, AAA tranches were found to contain mortgages that were never priced — and Ant's Huabei and Jiebei ABS once made up about 85 percent of China's consumer-loan ABS market.
    Evidence
    Caixin — Ant consumer-loan ABS approved: https://www.caixinglobal.com/2018-01-24/ant-financial-gains-approval-to-issue-securities-backed-by-consumer-loans-101201842.html; Bank Underground — 2008: https://bankunderground.co.uk/2018/09/17/as-safe-as-houses-how-a-small-corner-of-the-us-mortgage-market-nearly-brought-down-the-global-financial-system/

  8. Alternative-data scoring

    Input
    Phone usage, transaction behavior, airtime and bill payments — data generated without any credit product.
    Transformation
    A machine-learning model finds predictive patterns in the behavior and maps them to default probability.
    Output
    A credit decision for someone with no file.
    Limit
    Correlation is not capacity: behavior predicts repayment on average, but shocks and total debt stay invisible — and proxies can encode discrimination.
    Evidence
    Björkegren & Grissen — phone usage predicts repayment: https://documents1.worldbank.org/curated/en/811881575657172759/pdf/Behavior-Revealed-in-Mobile-Phone-Usage-Predicts-Credit-Repayment.pdf; NYT — Apple Card: https://www.nytimes.com/2019/11/10/business/Apple-credit-card-investigation.html

  9. Interest-rate regulation

    Input
    High rates on small loans and a regulator's cost-of-credit target.
    Transformation
    A rate cap lowers the price ceiling; lenders respond by rationing credit to higher-risk segments.
    Output
    Cheaper loans for those still served — and fewer loans for those priced out.
    Limit
    Kenya's 2016 cap shrank SME credit from about 25 to 15 percent of lending before repeal in 2019; the protection lever is also the exclusion lever.
    Evidence
    Alper & Clements — Kenya rate controls: https://www.semanticscholar.org/paper/Do-interest-rate-controls-work-Evidence-from-Kenya-Alper-Clements/adea279b3cefdf1d6a2e0657348aeabcf7ef6424

  10. Consent-based data sharing

    Input
    A borrower's own financial data held across banks and platforms.
    Transformation
    An Account Aggregator flow lets the borrower authorize a lender to read specific records, on consent, for a bounded purpose.
    Output
    A portable, user-owned data stream for underwriting.
    Limit
    Adoption and power still concentrate where the data lives; consent design decides whether this is liberation or a new permission fee.
    Evidence
    BusinessWorld — India Account Aggregator, Rs 1.67 trillion in 2025: https://www.businessworld.in/article/consent-based-finance-pushes-india-toward-rs-1-67-tn-loans-574795
Where this changes a real decision

SignalA thin-file applicant's behavior and first repayment; over-indebtedness and collection-abuse indicators; a borrower's consent to share verified data.

Decision ownerDigital lenders and their collections desks, regulators and competition authorities, and the borrower who controls the data.

ThresholdPredicted default above or below the lender's tolerance; harm indicators (rate spikes, contact-list shaming, multiple simultaneous loans) above a regulator's line; consent and purpose limits for data flows.

ActionApprove, decline, reprice, or change the next limit; provision, collect, write off, then sell or securitize; cap rates, take down apps, or stop an IPO; route underwriting through a consent-based data rail.

ConsequenceAccess expands for the unrecorded and shocks get smoothed, while invisible total debt, uncapped prices, and repackaged losses decide who benefits — the machine's blind spots are where the value and the harm both live.

Application chains

  1. Lender underwriting to write-off

    SignalA thin-file applicant's behavior data and first repayment; then a missed payment.

    Decision ownerDigital lender's risk team, collections desk, and portfolio manager.

    ActionApprove a small first loan, set limit and price from predicted loss, update from observed repayment, then provision, collect, write off, and sell or securitize the defaulted paper.

    ConsequenceAccess expands for the unrecorded while invisible cross-provider debt stays hidden — 2.7 million Kenyans were blacklisted between 2014 and 2016, and debt buyers now trade that paper at five to nine cents on the dollar.

  2. Regulator intervention

    SignalRate spikes, collection abuse, contact-list shaming, or over-indebtedness clusters in digital credit.

    Decision ownerCentral banks, competition authorities, and securities regulators (OJK, SEC, PBoC, CFPB).

    ActionCap rates (Kenya 2016), take down abusive apps (Philippines SEC), stop an ABS or an IPO (China 2020), or rewrite microloan rules.

    ConsequenceProtection can ration access — Kenya's cap shrank SME credit from about 25 to 15 percent before its repeal — and stopping Ant's $37 billion IPO reset who may hold the record.

  3. Consent-based credit infrastructure

    SignalA borrower wants a lender to see verified data without handing over the whole record.

    Decision ownerThe borrower, under a regulator-designed data-sharing framework.

    ActionAuthorize a bounded, purpose-limited data flow through an Account Aggregator instead of a bureau or platform sale.

    ConsequencePortable records enable underwriting at scale — India's Account Aggregator processed 1.67 trillion rupees in loan disbursals in 2025 — but the design of consent decides whether power stays with the borrower.

The limit this episode must keep

Digital credit is not uniformly a trap. M-Shwari loans measurably helped Kenyan households smooth shocks like illness and school fees; phone-usage data genuinely predicts repayment; Grameen Bank showed that social collateral can bring credit to people with nothing; and M-KOPA's pay-as-you-go model finances solar kits and smartphones for millions with default near ten percent. The failure is not prediction itself — it is the measurement boundary (total debt invisible), the missing price ceiling, and the ownership of the record. Regulation-void harvesting and over-regulation exclusion are two different failure modes; conflating them is how the debate never moves.

Key sources