Quick answer: a credit-card rewards system is a second ledger running beside the money ledger. Transactions earn points, miles or cash-back units according to program rules. Redemptions consume those units at values that can depend on redemption type and partner agreements. For programs where the issuer bears the future redemption cost, the issuer can estimate a rewards liability from the number of points expected to be redeemed and the expected cost per redeemed point. The system therefore has to forecast not only how many points customers will earn, but how many will ultimately be redeemed, at what cost, through which channel, and under which program terms.

A reward point is not money, but once the issuer promises a future benefit for it, the point becomes an economic state the system must count, value and eventually extinguish correctly.

Boundary: this article explains public rewards economics, accounting-style liability estimation and ledger mathematics. Program terms differ substantially by issuer and jurisdiction. It is not advice about which card or rewards program to use.

Why this belongs in mathematics

Rewards systems combine piecewise earn functions, probability, expected value, survival/attrition behaviour, cost forecasting, ledger reconciliation and incentive design. They are also a useful example of an invisible balance sheet: a customer can see 100,000 points as an asset-like benefit while the issuer sees an expected future cost whose amount depends on redemption behaviour.

The Federal Reserve’s research on credit-card profitability shows that rewards expenses can exceed interchange income on the transaction side of card economics. It separates credit-card profitability into credit, transaction and fee functions and finds rewards are a major cost of the transaction business. See Federal Reserve — Credit Card Profitability.

1. Earn rules are piecewise functions

Suppose a card earns:

1 point per S$1 on general purchases;
3 points per S$1 on travel;
5 points per S$1 on selected promotional categories up to a cap.

The earn function is not simply “purchase amount × one rate.” It is:

Points earned = f(amount, merchant category, product, promotion, cap, transaction eligibility, posting state).

A S$1,000 hotel purchase and a S$1,000 tax payment can therefore generate different reward quantities even though the card transaction amount is identical.

2. Pending rewards and posted rewards should be separate states

A purchase can be authorised, cleared, later refunded or disputed. If points are granted before the transaction is sufficiently final, the rewards ledger can create benefits on transactions that later disappear.

A strong state machine can use:

eligible purchase → pending reward → posted reward → redeemed / reversed / expired / transferred.

Refunds and chargebacks should reverse the correct earned reward rather than create an unrelated negative adjustment that nobody can trace.

3. The rewards ledger has its own invariant

A simplified points balance is:

Closing points = opening points + earned + bonuses + inbound transfers − redemptions − reversals − expiries − outbound transfers.

Every change should have a source event. If a customer sees 10,000 points vanish without a matching redemption, expiry, reversal or governed adjustment, the ledger has lost economic explainability.

This connects to double-entry bank-ledger algorithms. Rewards points are not ordinary currency, but they still need auditable state transitions.

4. Redemption value is another piecewise function

One point can have different redemption values:

cash-back statement credit;
travel booking;
gift card;
merchandise;
transfer to airline or hotel partner;
merchant-specific redemption.

Let v_k be cost or value per point in redemption channel k. The expected cost per point is then related to the mix of redemptions:

Expected cost per redeemed point = Σ probability(channel k) × cost per point in channel k.

If customer behaviour shifts toward more expensive redemptions, the issuer’s liability can rise even if the number of outstanding points does not change.

5. Expected redemption turns outstanding points into an estimated liability

Not every point issued is necessarily redeemed. A simple conceptual liability model is:

Rewards liability ≈ outstanding earned points × expected redemption rate × expected cost per redeemed point.

American Express publicly describes a similar structure for its Membership Rewards programme: it estimates the future cost of earned points using an Ultimate Redemption Rate (URR) and a weighted average cost per point. Its 2025 annual report reported a US$16.5 billion Membership Rewards liability at year-end 2025, illustrating that rewards obligations can be financially material. See the American Express 2025 Form 10-K.

6. Breakage is the complement of expected redemption—but only under the model definition

If the model expects 96% of a population’s points ultimately to be redeemed, a simplified expected non-redemption share is 4%.

That non-redemption is often called breakage in loyalty modelling. But breakage is not automatically the same thing as point expiration. Points can go unredeemed because accounts close, customers abandon small balances, transfer rules change, or other program states prevent or discourage redemption.

American Express reported an Ultimate Redemption Rate of about 96% for current Membership Rewards participants as of June 30, 2026, showing how high expected redemption can be in a mature programme. See its June 2026 filing.

7. Small changes in redemption assumptions can move large liabilities

Suppose a programme has 1 trillion outstanding points, expected redemption of 90% and expected cost of 1 cent per redeemed point:

1,000bn × 90% × S$0.01 = S$9bn estimated future cost.

If expected redemption rises from 90% to 92%, the estimated cost rises by S$200m before a single additional point is redeemed.

This is model risk: a liability can move because expected customer behaviour changes, not only because actual points balances change.

8. Redemption models can use survival-style thinking

Points can be grouped by cohort and age: how long since they were earned, customer tenure, product, spend level and prior redemption behaviour.

A survival-style model asks:

P(point remains unredeemed after time t | programme/customer state).

The cumulative redemption curve can then estimate ultimate redemption rather than treating one year’s observed redemption percentage as the final lifetime outcome.

9. Sign-up bonuses create cohort distortion

A customer receiving 100,000 bonus points after meeting an introductory spend target can behave differently from a long-tenured customer earning points gradually.

Bonus cohorts can have:

higher immediate redemption;
higher churn;
different partner-transfer use;
different profitability;
different fraud or abuse risk.

A model trained on ordinary earn behaviour can therefore misestimate the liability created by large acquisition campaigns.

10. Rewards economics interact with interchange

On the transaction side of card economics, purchase volume can generate interchange income for issuers while also generating rewards expense.

A simplified transaction margin is:

Net transaction margin ≈ interchange + annual-fee allocation − rewards cost − fraud cost − transaction funding/operating cost.

Federal Reserve research found that rewards and other transaction expenses can exceed interchange revenues on average, making the transaction function slightly negative in aggregate in the studied portfolios even though rewards can attract valuable customers and spending.

11. A generous earn rate can still be expensive after redemption mix changes

Suppose a card earns 2 points per S$1 and expected cost per redeemed point is 0.7 cents with 90% ultimate redemption. Expected rewards cost per S$1 spend is:

2 × 0.90 × 0.007 = 1.26 cents per dollar of spend, or 1.26% of purchase volume.

If redemption mix changes and expected cost per point rises to 0.9 cents, expected rewards cost becomes 1.62% of spend. The advertised earn rate did not change; programme economics did.

12. Partner transfers create a second ledger boundary

When points transfer to an airline, hotel or merchant partner, the issuer must remove or convert its own point obligation and establish the correct partner settlement obligation.

The transfer needs:

customer identity match;
programme account match;
transfer ratio;
minimum/maximum rules;
transaction idempotency;
partner acknowledgement;
reversal/repair process.

A point deduction without a corresponding partner credit is a ledger-integrity failure. The CFPB has specifically highlighted consumer complaints involving rewards disappearing or becoming unavailable during redemption and partner-transfer processes.

13. Devaluation changes customer value and potentially model cost

A programme can change how many points are required for a reward or change available redemption options. That can alter perceived consumer value and future redemption behaviour.

CFPB Circular 2024-07 states that credit-card reward programme operators can face consumer-protection risk when they devalue earned rewards, revoke rewards through buried conditions or deduct points without delivering the corresponding promised benefit. See CFPB Circular 2024-07.

The mathematical lesson is that changing redemption value is not merely updating a catalogue. It can change liability assumptions, redemption timing, customer attrition and programme trust.

14. Rewards fraud and gaming are separate from ordinary high engagement

Some accounts legitimately optimise promotions and categories. Others may generate abusive patterns through synthetic purchases, manufactured returns, account manipulation or compromised credentials.

A defensive system can monitor for anomalous earn-and-burn patterns without equating “high points earned” with abuse. Useful features include transaction reversals, impossible timing, linked accounts and repeated earn/redeem loops, subject to lawful data use and human review.

The public article deliberately does not publish issuer-specific fraud thresholds or evasion details.

15. Creative-work lens: Up in the Air and the psychology of stored status

Up in the Air is not a banking source, but frequent-flyer status and miles are central to its character world. The story makes one programme-design insight visible: loyalty units can carry emotional value beyond their direct cash equivalent. Customers may hoard points, chase status or change spending behaviour because of a programme’s symbolic structure.

The creative work helps us remember that redemption behaviour is behavioural, not mechanical. Liability estimation still requires real programme data.

16. The rewards algorithmic pipeline

Capture eligible posted transactions.
Apply product, category, promotion and cap rules.
Create pending/posted rewards entries.
Reverse rewards for refunds and invalidated transactions.
Maintain customer points balance by programme/version.
Estimate ultimate redemption by cohort and behaviour.
Estimate weighted average redemption cost per point.
Calculate rewards liability and expense.
Process redemptions and partner transfers atomically.
Relieve liability when rewards are redeemed.
Monitor programme devaluation and partner changes.
Backtest redemption and cost assumptions.
Reconcile points ledger, partner settlements and financial liability.

17. Failure modes

Points=cash assumption. One fixed cash value is assigned despite multiple redemption channels.
Earn-before-finality. Refunded transactions leave permanent rewards behind.
Static redemption rate. Programme changes do not update expected redemption.
Static cost per point. Partner prices and redemption mix change while liability stays fixed.
Bonus-cohort contamination. Acquisition bonuses are modelled like ordinary earn behaviour.
Partner-transfer gap. Customer points are debited without successful partner credit.
Devaluation-only profit view. Lower redemption cost is modelled without customer response or legal risk.
Fraud overreach. High engagement is treated as abuse without sufficient evidence.

18. Diagnostics and falsifiers

Do total points reconcile from opening balance through earn, redeem and reversal events?
What is estimated ultimate redemption by product and tenure?
How sensitive is liability to a 1-percentage-point change in redemption?
Which redemption channel has the highest cost per point?
Do sign-up bonus cohorts redeem faster or churn more?
How much transaction margin remains after rewards and fraud cost?
Which partner-transfer failures create unreconciled points?
Can a liability movement be explained by points volume, redemption rate or cost-per-point changes?

Suppose someone claims, “Outstanding points doubled, so the rewards liability must double.” A falsifier is a material change in ultimate redemption or expected cost per point. Liability depends on expected future redemption cost, not only raw point count.

19. Verification and update triggers

reconcile points balances to transaction histories;
backtest redemption curves by cohort;
re-estimate cost per point after partner changes;
stress large promotional bonus campaigns;
test idempotency and reversals in partner transfers;
review consumer complaints for hidden system failures;
retain programme-version history after earn/redemption changes;
independently reproduce liability sensitivity to redemption assumptions.

Connections across the finance-and-banking algorithms lane

Merchant-acquiring pricing — explains the merchant side of interchange economics.
Card authorisation — transaction state before rewards are finally earned.
Payment idempotency — crucial when rewards and partner transfers retry.
Model validation — expected redemption is a modelled future behaviour.

Research anchors

The deeper lesson

Rewards mathematics is the mathematics of a promise that has not yet been redeemed. The earn engine creates units. The ledger preserves ownership and history. Redemption behaviour determines how many units eventually become cost. Redemption mix determines how expensive that cost is. Interchange and fees fund part of the programme economics. A strong rewards system therefore does not ask only, “How many points are outstanding?” It asks, “Which of those points will become real future obligations, at what cost, and can every point still be traced from the transaction that created it to the redemption that extinguished it?”

Educational note: This article explains credit-card rewards mathematics and public disclosures. It is not card-selection advice, tax advice or a valuation of any individual’s points.

Small Group Tutorials

How Credit-Card Rewards Algorithms Track Points and Liability: Earn Rates, Redemption Value, Breakage, Interchange Economics, Devaluation and Ledger Controls