Key Takeaways

  • Half of US credit card holders carry a revolving balance, averaging $7,000 in compounding debt.
  • Affirm underwrites each purchase individually with a fixed schedule designed to amortize to zero.
  • Affirm retrains its transformer-based underwriting models every six to seven months, driving continuous loss-rate improvements.
  • By pricing transaction risk instead of open-ended credit lines, Affirm posts lower credit losses than prime lenders like Synchrony across broader demographics.

Revolving Credit Relies on Trapping Users

Half of American credit card holders carry a balance from month to month. The average balance sits at $7,000. That math explains why traditional card issuers make so much money. They hand out revolving credit limits, let interest compound, and profit when borrowers fall behind on payments.

Fraser Christie, founder and CIO of Sandstone Global Investment, points out that Affirm was built on the opposite premise. “The clearest and biggest differentiator is that they're underwriting transaction by transaction,” Christie explains. Rather than granting a revolving five-thousand-dollar line of credit to sit open indefinitely, Affirm evaluates each basket at checkout. “The interest rate is fixed, the repayment schedule is fixed, and the balance is paid to zero according to the rundown of the plan.”

That distinction changes consumer incentives. A borrower knows the exact dollar cost before clicking purchase. When they finish paying, the loan terminates. “Every loan is underwritten on a transaction bytransaction basis and they are designed to be paid down to zero,” Christie notes. Traditional banks want the loan to never disappear. Affirm cannot make money if the loan does not settle.

Rapid Model Iteration Beats Prime Balance Sheets

Traditional consumer credit relies on static FICO scores and periodic credit bureau pulls. If an issuer extends an open credit line, it takes on balance sheet exposure for years without knowing what the user buys next.

Affirm attacks risk differently. Because it scores individual transactions at checkout, it sees item-level metadata, merchant data, and immediate repayment cadence. That data feeds transformer-based machine learning models that the company updates constantly.

“And so they're basically seeing every time they release a new underwriting model, which is on a call it six, seven month cadence, they're getting materially better underwriting outcomes,” Christie notes.

This speed creates a compounding edge over legacy finance. Synchrony, for example, focuses primarily on prime borrowers with pristine paper trails. Yet Affirm achieves lower credit losses than Synchrony while approving customers across a much wider demographic spectrum. Running machine learning models on a six-month release cycle lets Affirm react to consumer stress long before legacy card issuers adjust their line sizes. As Affirm scales past $50 billion in GMV, its predictive advantage widens.

What to Do With This

Inspect your core pricing or risk model this week. If you price customer risk or usage on an annual or account-level basis, calculate the variance in margin across individual transactions or customer actions. Identify the top 20% of customer behaviors that generate 80% of your operational downside, and write a rule to underwrite or meter those specific events individually.