The Lending Business

A loan is a bet that money handed out today comes back with more attached. Every lender, from a money-center bank's card portfolio to a high-yield online lender, runs the same four-line equation on that bet: what the borrower pays (the yield), what the money costs (the cost of funds), what doesn't come back (credit losses), and what it costs to find, decide, fund and collect (operating expense). The lines move together. A lender that charges more attracts riskier borrowers and loses more; one that funds itself with deposits pays less for money but carries a bank's capital and supervision; one that sells its loans books revenue upfront and keeps none of the yield. The choices about funding and accounting decide when the revenue shows up and who holds the loss, which is why two lenders making the same loan can report very different businesses.

This chapter covers the four lines with figures from public filings and Federal Reserve data, the three ways to earn from a loan (holding it, selling it, servicing it), how prime and high-yield lenders differ, the accounting choice between amortized cost and fair value, and what a lending system records so that each loan's economics can be explained after the fact. The products themselves follow: Credit Cards as Credit, Personal & Installment Loans and Auto Lending, then Mortgages & Home Equity, Student Loans and Buy Now, Pay Later. How lenders raise the money is Funding the Book; how losses are reserved is Credit Losses & CECL.

How the Platforms Do It#

As of October 2026, from the latest annual filings and monthly credit disclosures:

Lender	Model	How it's funded	How it earns	A figure that shows the model
Affirm	Point-of-sale installment lender, fiscal year ending June 30	Partner banks originate (Celtic Bank and Lead Bank are its primary originating banks); Affirm buys, holds, sells or securitizes	Merchant fees, card network revenue, interest, gain on sale, servicing	Interest income was $2.05 billion of $4.26 billion revenue in fiscal 2026, 48% of the total
Capital One	Bank: cards, auto, commercial, deposits	Deposits	Net interest income and fees	Domestic card loans of $262.4 billion at December 2025, with a 5.01% annualized net charge-off rate that month
Enova	High-yield online lender to consumers and small businesses (CashNetUSA, NetCredit, OnDeck, Headway)	Debt facilities and securitizations; five bank partner programs were 31.6% of 2025 originations and purchases; Grasshopper Bank acquisition pending since December 2025	Interest and fees, loans carried at fair value	Average yields of 90% on consumer installment loans and 48% on small business installment loans in 2025
Happen (formerly LendingClub)	Nationally chartered digital bank with a loan marketplace	Deposits of about $9.8 billion at December 31, 2025, plus loan sales and its Structured Program	Net interest income on loans it keeps, plus marketplace fees on loans it sells	$9.6 billion of 2025 originations, $7.1 billion of them sold through the marketplace; net interest margin 6.07%
Upstart	Lending marketplace for more than 100 bank and credit union partners	2025 originations: 64% institutional investors, 26% lending partners, 10% its own balance sheet	Platform and referral fees, servicing	91% of 2025 loans fully automated

Three things stand out. The same loan can be four businesses: Affirm's revenue lines (interest, gain on sale, servicing, merchant fees) are the holding, selling and servicing models in one income statement. A charter changes the cost of money: Happen describes itself as "a leading nationally chartered digital bank" and funds the loans it keeps with deposits, and Enova's stated reasons for buying Grasshopper include lower-cost funding and regulatory simplicity. Fee models still carry credit risk: Upstart's partners and investors hold most of the loans, but the 10% it keeps and the commitments it makes to funding partners put losses back on its balance sheet.

The Four Lines#

Yield

Yield is what the portfolio earns as a rate on what's outstanding: interest, plus fees that are part of the cost of credit (origination fees, which accounting spreads over the loan's life), divided by the average balance. It's set by the product, the borrower's risk and the law. The Federal Reserve's G.19 release gives the market level: in August 2026, commercial banks charged 11.90% on 24-month personal loans, 21.19% on all credit card accounts and 22.36% on card accounts actually assessed interest, and 7.54% on 60-month new car loans. At the other end, Enova reports average annualized yields of 90% on consumer installment loans, 147% on consumer lines of credit, 48% on small business installment loans and 49% on small business lines for 2025.

The posted rate and the realized yield differ. Promotional and 0% balances earn nothing (Affirm's 0% monthly installments were 14% of fiscal 2026 GMV, and Pay-in-X another 16%), card balances paid in full within the grace period earn no interest, loans that charge off stop accruing, and prepayment cuts an installment loan's interest short. A yield figure is only comparable when it says which balances are in the denominator.

Cost of Funds

A bank funds loans mostly with deposits, whose cost is the rate it pays depositors plus the cost of running branches and apps; Bank Economics & Balance Sheet Risk covers deposit betas and net interest margin. A nonbank lender borrows: warehouse lines secured by loans, securitizations, forward flow sales to investors, and corporate debt. Its cost of funds is a spread over a benchmark that widens when credit markets tighten, which is why nonbank lenders' margins move with the capital markets in a way bank margins don't. Enova's 2025 interest expense, net, was $339.3 million against $3.15 billion of revenue; Affirm's funding costs were $454.0 million in fiscal 2026.

Happen shows the bank version in one number: a net interest margin of 6.07% in 2025, up from 5.62%, on a book funded by deposits. A nonbank with the same loans would report the spread as interest income minus interest expense, with no deposits to lean on when its facilities reprice.

Credit Losses

Losses are the line that decides whether the other three matter. The Federal Reserve publishes charge-off rates for all US banks, seasonally adjusted: in the second quarter of 2026, credit cards charged off at an annualized 3.82% (4.19% a year earlier), other consumer loans at 1.21%, and all loans and leases at 0.55%. The delinquency rates behind them were 2.85% for cards and 2.33% for other consumer loans. Capital One's domestic card book charged off at 5.01% in December 2025 and its auto book at 2.07%.

Losses arrive late. A loan cohort's charge-offs build over its first year or two and then fade, so a lender growing fast shows low losses on a book that's mostly young, and a lender whose growth stops sees its loss rate rise with no change in underwriting. That's why loss rates are read by vintage (the month or quarter of origination) and why the accounting below matters: under CECL a lender books the expected lifetime loss on the day it makes the loan, so growth costs earnings upfront.

Operating Expense

Finding a borrower, deciding, funding, servicing and collecting all cost money. Marketing is the largest variable piece for direct-to-consumer lenders: Enova's 2025 marketing expense was $621.1 million, 19.7% of revenue by my arithmetic, and its total operating expenses were $1.09 billion. Banks report the efficiency ratio, noninterest expense over revenue: Happen's was 63.1% in 2025, down from 69.1%. Automation moves this line more than any other: Upstart reports 91% of its 2025 loans fully automated, meaning no human touched the decision or the documents.

Three Ways to Earn From a Loan#

Model	Revenue	Who carries the loss	When income shows up	Who does it
Hold it (balance sheet)	Interest and fees over the loan's life, minus the cost of funds	The lender	Over the life; losses reserved upfront under CECL	Banks, credit unions, Happen's retained loans, Affirm's held loans
Sell it (gain on sale, marketplace)	Origination fees and a gain or premium when the loan is sold	The buyer, unless the seller guarantees or retains a piece	At sale	Happen's marketplace, Affirm's loan sales, Upstart's investor channel, mortgage originators
Service it	A servicing fee, usually a share of the outstanding balance, plus late and other fees where allowed	The owner of the loan	Monthly, while the loan is outstanding	Every seller that keeps servicing; dedicated servicers

Most lenders run more than one. Affirm's fiscal 2026 revenue was $4.26 billion: $2.05 billion of interest income, $1.15 billion of merchant network revenue (merchants pay Affirm to offer the financing), $294.0 million of card network revenue, $596.6 million of gain on sales of loans and $173.1 million of servicing income. Happen earned $625.7 million of net interest income in 2025 on the loans it kept, and its marketplace revenue (origination fees from borrowers and fees on loans sold) rose $113.2 million, 47%, with total net revenue of $998.8 million. Upstart's revenue is mostly platform and referral fees from its partners; its top three lending partners originated 83% of its marketplace loans and produced 61% of its fee revenue in 2025.

Selling changes what the lender optimizes. A balance sheet lender wants loans that pay for years; a seller wants loans investors will buy at a premium, and its volume depends on their appetite. A seller that also services keeps a relationship with the borrower and a fee stream, and an obligation: the owner of the loan expects it serviced to the standard the sale agreement sets. Funding the Book covers forward flow, warehouse lines and securitization; Loan Servicing covers the servicer's ledger.

Prime and High-Yield Economics#

A prime lender and a high-yield lender aren't the same business at different prices. The four lines take different shapes:

Line	Prime (cards and installment at banks, Affirm's interest-bearing loans)	High-yield (Enova's consumer products)
Yield	Low to mid 20s for cards, around 12% for bank personal loans	90% to 147% on consumer products
Losses	Card charge-offs around 4% at all banks	A large share of revenue: Enova's change in fair value, which includes charge-offs and the remeasurement of expected losses, was $1.32 billion, 41.9% of 2025 revenue by my arithmetic
Operating cost	Spread over large balances; card marketing and rewards	High per loan: small balances, short terms, heavy marketing (19.7% of Enova's revenue)
Constraint	Competition and, for cards, the CARD Act's fee and rate rules	Rate caps: the Military Lending Act's 36% for covered borrowers, state caps, and the states where a lender doesn't offer a product at all (Enova didn't offer consumer loans in 13 states at the end of 2025)

The high-yield model works when its yield covers losses that would end a prime lender, which takes small exposures, fast feedback (short loans show their losses quickly) and pricing that adjusts by state and by borrower. It's also why the model depends on law more than any other part of lending: a state rate cap below the lender's loss rate closes the market, and partnering with a bank that can export its home state's rate reopens it until a court or a legislature says otherwise. APR, Finance Charges & Rate Caps covers exportation, true lender and the state opt-outs; Lender Licensing & Bank Partnerships covers the partnership itself.

Affirm sits between the two. Its 0% and Pay-in-X products (30% of fiscal 2026 GMV) earn the merchant's fee instead of interest, which makes them the merchant's marketing cost, and its interest-bearing loans (70%) earn interest like any installment lender's. It has charged $0 in late fees since its founding, so its revenue per loan has to come from the rate and the merchant.

Worked Numbers: One Loan's Economics#

Take the representative example Happen publishes for its personal loan (as of July 2026): $15,262 over 36 months at 13.99% interest, with a 6% origination fee of $916, so the borrower receives $14,346 and pays $522 a month. The APR is 18.40% because the fee is a finance charge the borrower pays from the proceeds; Personal & Installment Loans builds the schedule. Over its life, if every payment arrives on time:

Line	Amount	Basis
Interest	$3,513.61	The amortization schedule at 13.99% on $15,262, payment $521.54
Origination fee	$916.00	Happen's example
Revenue	$4,429.61
Cost of funds	$1,130.19	Assumed 4.5% a year on the outstanding balance, which averages $8,371.80 over the 36 months
Expected credit loss	$763.10	Assumed 5% of principal over the life
Acquisition	$400.00	Assumed marketing and verification cost per funded loan
Servicing	$108.00	Assumed $3 a month
Contribution	$2,028.32	About 8.1% a year on the average balance

The assumptions are mine and illustrative; the structure is the point. Three things the arithmetic shows: the fee is a fifth of the revenue and arrives on day one, which is why fee-heavy lenders look profitable early; the cost of funds is charged on a balance that falls every month, so it's smaller than a rate times the original principal; and a doubling of the loss assumption (to 10%) takes away $763 of the $2,028, which is why underwriting, not pricing, is where lenders win or lose. A seller would replace the interest, cost of funds and loss lines with a sale price; a servicer would see only the $108 line from the other side.

Amortized Cost or Fair Value#

The accounting choice decides when the economics above show up in earnings:

• Amortized cost with a CECL allowance. The loan sits at its principal, adjusted for deferred fees and costs. On the day it's made, the lender books a provision for expected lifetime losses (the current expected credit loss model of ASC 326), then recognizes interest and fees over the life. Affirm reports a provision for credit losses of $796.7 million in fiscal 2026 against $9.47 billion of loans held for investment; Happen's 2025 provision was $191.3 million on $4.27 billion of loans at amortized cost.
• Fair value. The loan is measured each period at what it would sell for, with changes in earnings. Expected losses and prepayments are in the price, so there's no separate allowance and origination fees aren't deferred. Enova carries its loans and finance receivables at fair value; Upstart holds loans at fair value; Happen elected the fair value option for newly originated held-for-investment loans from January 1, 2026, leaving most of the existing book under CECL.

The choice changes the ledger, not the cash. Under CECL, a growing lender's earnings are depressed by day-one provisions; under fair value, they swing with the discount rates and loss assumptions in the mark. Credit Losses & CECL works the two through the same portfolio. For the system, the consequence is that the loan record has to carry its measurement basis, because the same payment posts differently under each.

Banks, Nonbanks and Partners#

Who makes the loan decides which of the four lines it can move:

• A bank or credit union funds with deposits, exports its home state's rate (national banks under section 85 of the National Bank Act, state banks under DIDMCA section 521, which a state may opt out of for loans made in it, as Colorado did in 2023 and Oregon in April 2026), and answers to its prudential regulator for capital and credit risk management. Happen, the former LendingClub, now calls itself a nationally chartered digital bank; Enova's pending purchase of Grasshopper is a move in the same direction.
• A licensed nonbank lender holds state licenses where it lends, prices within each state's caps, and funds itself in the capital markets. Enova lends this way in many states.
• A bank partnership puts a bank's name on the note and a nonbank's technology, marketing and often its funding behind it. Affirm's loans are mostly originated by Celtic Bank and Lead Bank; Enova runs five bank partner programs and says it doesn't guarantee the performance of the bank-originated loans; Upstart's partners set hard credit criteria its models operate inside. The legal question of who the "true lender" is belongs to APR, Finance Charges & Rate Caps, and the operating relationship to Sponsor Bank Oversight.

How to Build It#

These are deltas against The Core Data Model: products (versioned loan products), accounts (one per loan or line), loans (terms extending an account), credit_decisions, interest_accruals, fee_assessments, and journal_entries with journal_lines. This chapter adds the records that let anyone explain a loan's price and its economics after the fact.

CREATE TABLE pricing_grids (                 -- versioned: the price a product offers by risk and term
  id                 UUID PRIMARY KEY DEFAULT uuidv7(),
  product_id         UUID NOT NULL REFERENCES products (id),
  version            INT NOT NULL,
  effective_from     TIMESTAMPTZ NOT NULL,
  effective_to       TIMESTAMPTZ,
  approved_by        TEXT NOT NULL,           -- credit or pricing committee, by reference
  partner_criteria_ref TEXT,                  -- the bank partner's or funder's criteria it was approved against
  UNIQUE (product_id, version)
);

CREATE TABLE pricing_grid_cells (
  grid_id            UUID NOT NULL REFERENCES pricing_grids (id),
  risk_grade         TEXT NOT NULL,           -- the decision engine's grade, not a raw score
  term_months        SMALLINT NOT NULL,
  amount_band_low    BIGINT NOT NULL,         -- minor units
  amount_band_high   BIGINT NOT NULL,
  jurisdiction       TEXT NOT NULL,           -- state or country: caps differ
  interest_rate_bps  INT NOT NULL,
  origination_fee_bps INT NOT NULL DEFAULT 0,
  max_apr_bps        INT NOT NULL,            -- the cap this cell must never exceed after fees
  PRIMARY KEY (grid_id, risk_grade, term_months, amount_band_low, jurisdiction)
);

CREATE TABLE loan_pricing (                  -- one row per offer accepted: the price and why
  account_id          UUID PRIMARY KEY REFERENCES accounts (id),
  credit_decision_id  UUID NOT NULL REFERENCES credit_decisions (id),
  grid_id             UUID NOT NULL REFERENCES pricing_grids (id),
  risk_grade          TEXT NOT NULL,
  interest_rate_bps   INT NOT NULL,
  origination_fee     BIGINT NOT NULL,
  apr_bps             INT NOT NULL,           -- as disclosed, from the same calculation as the disclosure
  override_reason     TEXT,                   -- any departure from the grid, with its approver
  override_by         TEXT,
  -- the economics assumed at origination, so vintages can be compared with what was expected
  assumed_cost_of_funds_bps INT NOT NULL,
  assumed_lifetime_loss_bps INT NOT NULL,
  assumed_prepayment_cpr_bps INT,
  assumed_acquisition_cost BIGINT,
  priced_at           TIMESTAMPTZ NOT NULL DEFAULT NOW()
);

ALTER TABLE loans
  ADD COLUMN measurement_basis TEXT NOT NULL DEFAULT 'amortized_cost'
    CHECK (measurement_basis IN ('amortized_cost', 'fair_value', 'held_for_sale')),
  ADD COLUMN basis_elected_at  TIMESTAMPTZ NOT NULL DEFAULT NOW(),
  ADD COLUMN vintage           DATE NOT NULL;  -- first of the origination month

CREATE TABLE loan_economics_monthly (        -- derived, rebuilt from the ledger; never edited
  account_id         UUID NOT NULL REFERENCES accounts (id),
  month              DATE NOT NULL,
  avg_balance        BIGINT NOT NULL,
  interest_income    BIGINT NOT NULL,
  fee_income         BIGINT NOT NULL,         -- including amortized origination fees under amortized cost
  funding_cost       BIGINT NOT NULL,         -- allocated from the funding source's actual cost
  charge_off         BIGINT NOT NULL DEFAULT 0,
  recovery           BIGINT NOT NULL DEFAULT 0,
  servicing_cost     BIGINT NOT NULL DEFAULT 0,
  fair_value_change  BIGINT,                  -- fair value loans only
  PRIMARY KEY (account_id, month)
);

Prices come from data, not code: a grid is a versioned rule set with effective dates (Effective Dates and Versions), and loan_pricing records which grid cell priced each loan and any override (Record Which Rule Decided). The max_apr_bps on each cell is checked against the computed APR before an offer is shown, so a fee change can't push a loan over a state cap or the Military Lending Act's limit unnoticed. The assumptions at origination are stored beside the price because the question every credit committee asks a year later is "is this vintage performing as priced?", and that needs the expected numbers as they were, not as the model says today. loan_economics_monthly is a projection of the ledger, rebuilt by a job and never posted to; the ledger stays the source of truth (Integer Minor Units). measurement_basis lives on the loan, not the product, because an election like Happen's applies by origination date, and a loan moved to held for sale changes basis mid-life.

Auditing#

Examiners, auditors, bank partners and loan buyers each ask for a loan's economics in a different shape, and the general design is in Audit Trails. For the lending business, make sure the record shows:

• The price as offered and why: grid version, cell, grade, overrides with their approver, and the APR computed by the same code as the disclosure.
• The assumptions as they were at origination, unchanged afterward, so performance can be compared with what was priced.
• The measurement basis of every loan, when it was elected or changed, and under which policy.
• Who funded the loan and when ownership moved: originated by whom, sold to whom, at what price, with servicing retained or released.
• Vintage performance: delinquency, charge-offs and recoveries by origination month, rebuilt from the ledger, reconciling to the general ledger each month.
• Partner criteria compliance: for loans made by a bank partner, evidence that each loan met the bank's criteria in force on its decision date.

What This Means for the System You Build#

• Model the four lines per loan, not just per portfolio: yield, cost of funds, losses and cost to serve, from the ledger.
• Keep pricing as versioned data with effective dates, approvals and a computed APR checked against every applicable cap before the offer.
• Store the assumptions at origination beside the price, immutable, so vintages can be judged against what was expected.
• Carry the measurement basis on the loan, because amortized cost and fair value post the same payment differently.
• Read losses by vintage, never as one blended rate, because growth hides losses and slowdown reveals them.
• Expect more than one business model in one ledger: held, sold and serviced loans side by side, with ownership and servicing as attributes that change.
• Know whose charter the loan rests on, because it sets the rate the loan may carry and the regulator who examines it.

References#

Filings

• Affirm: Annual report on Form 10-K for fiscal 2026, $50.2 billion of GMV, product mix (Pay-in-X 16%, 0% APR 14%, interest-bearing 70%), $0 in late fees, Celtic Bank and Lead Bank as primary originating banks, revenue lines (interest income $2,047.5 million, merchant network $1,149.9 million, card network $294.0 million, gain on sales $596.6 million, servicing $173.1 million, total $4,261.1 million), provision for credit losses $796.7 million, funding costs $454.0 million and loans held for investment of $9.47 billion
• Enova International: Annual report on Form 10-K for 2025, $3,151.7 million of revenue, about $7.8 billion of credit extended, average yields by product, five bank partner programs at 31.6% of originations, no guarantee of bank-originated loans, the pending Grasshopper acquisition and its stated reasons, loans at fair value, change in fair value of $1,321.4 million, marketing expense of $621.1 million, operating expenses of $1,090.8 million, net interest expense of $339.3 million and no consumer loans in 13 states
• Happen (formerly LendingClub): Annual report on Form 10-K for 2025, the self-description as a nationally chartered digital bank, deposits of about $9.8 billion, the Structured Program, marketplace and net interest income, $998.8 million of total net revenue, $625.7 million of net interest income, a 6.07% net interest margin, $9,589 million of originations ($7,134 million marketplace), loans at amortized cost of $4.27 billion, a $191.3 million provision, a 63.1% efficiency ratio and the fair value option for new HFI loans from January 1, 2026
• Upstart: Annual report on Form 10-K for 2025, more than 100 lending partners, the 64/26/10 funding split, 91% of loans fully automated, the top three partners' 83% of marketplace loans and 61% of fee revenue, and personal loan terms
• Capital One: Monthly credit metrics for December 2025 (Form 8-K exhibit), domestic card loans of $262.4 billion with a 5.01% net charge-off rate, and auto loans of $83.6 billion with a 2.07% rate

Law and regulation

• Cornell LII: 32 CFR 232.4, Military Lending Act terms of credit, the 36% MAPR limit and the charges it includes
• CRS Legal Sidebar LSB11433: state opt-outs from DIDMCA section 521 (May 22, 2026), national bank exportation under section 85, section 521 parity for state banks, the section 525 opt-out, Colorado's 2023 and Oregon's April 2026 opt-outs, and NAIB v. Weiser's en banc rehearing

Regulators and data

• Federal Reserve: G.19 Consumer Credit (October 7, 2026), August 2026 rates on personal loans, credit card plans and new car loans
• Federal Reserve: Charge-off rates on loans and leases at commercial banks, seasonally adjusted, second quarter 2026 charge-off rates for credit cards, other consumer loans and all loans
• Federal Reserve: Delinquency rates on loans and leases at commercial banks, seasonally adjusted, second quarter 2026 delinquency rates

In practice

• Happen: Personal loans, the representative example ($15,262 over 36 months at 13.99% with a 6% fee, $14,346 received, $522 a month, 18.40% APR), rates and fees as of July 6, 2026