BEHAVIORAL ECONOMICS · FINTECH · PRODUCT · DECISIONS · SYSTEMS
Hyperbolic Discounting Is Why Fintech Apps Are Designed Against Their Users
Present bias as the operating model of buy-now-pay-later, credit cards, and trading apps.
Present bias as the operating model of buy-now-pay-later, credit cards, and trading apps.
You are at checkout. The cart total reads $120. Underneath the total, a button offers four interest-free payments of $30, the first due today. You tap it. The order ships.
You do not run a discounted utility calculation. You do not pull out the APR comparison sheet. You do not stop to ask why a third party is willing to front the merchant 75% of the purchase price in exchange for the right to bill you on a schedule that did not exist a moment ago. You compare $30 today against $120 today and choose the smaller number.
What just happened has a name. It is one of the most empirically documented features of human cognition. It has been formalized mathematically since 1955. And it is the operating model of every major consumer fintech product launched in the last decade.
The product is not flawed. It is working exactly as designed.
The Math
Paul Samuelson published the standard model of intertemporal choice in 1937. The model assumes a decision maker evaluates a stream of future consumption by applying a constant discount factor δ to each future period:
The factor δ is between 0 and 1. Consumption at time t+1 is worth δ times consumption at time t. Consumption at t+2 is worth δ². The compounding is geometric. A plan made today is mathematically guaranteed to remain optimal tomorrow.
Robert Strotz proved this in 1955. Under exponential discounting, the optimal plan never changes as time passes. Preferences are time-consistent. The Samuelson formulation produces a clean, internally coherent agent who never regrets a choice.
Humans are not that agent.
George Ainslie published the empirical refutation in 1975. Across hundreds of laboratory experiments, subjects systematically reversed their preferences as time approached. A subject would prefer $100 in 365 days over $50 in 364 days. The same subject, asked a year later when the smaller reward had become immediate, would prefer $50 today over $100 tomorrow. The reversal is robust. It does not go away with education. It does not go away with stakes. It shows up in pigeons, in rats, in children, in adults.
The discount function that fits this behavior is not exponential. It is hyperbolic. Future rewards lose value rapidly in the short term, then more slowly over longer horizons.
David Laibson formalized the cleanest tractable version in 1997. The quasi-hyperbolic model, also called β-δ discounting, applies an extra discount to everything in the future relative to the present:
The parameter β captures present bias. When β = 1, the model collapses back to Samuelson and preferences are time-consistent. When β < 1, the agent attaches a sudden additional penalty to anything that is not now. Everything beyond the current period gets uniformly weighted by β. Long-run impatience is captured separately by δ.
The structure is simple and unforgiving. Once β drops below 1, the plan today is no longer the plan that will be executed tomorrow. Today the agent prefers to save. Tomorrow, when “tomorrow” arrives, the agent prefers to spend. The agent regrets in advance and continues anyway.
Ted O’Donoghue and Matthew Rabin published the operational consequence in 1999. They distinguished naive agents from sophisticated agents. Both have the same true β. The difference is in what they believe about their future selves. Naive agents believe their future β will be 1, and so they make plans they will not keep. Sophisticated agents know their β will stay below 1, and so they self-bind. The naive procrastinate. The sophisticated commit.
Most consumers are partially naive. They know they have a self-control problem. They underestimate its size.
Laibson, Repetto, and Tobacman estimated the parameters structurally in 2007 from household consumption data. The benchmark fit implied a 40 percent annualized short-term discount rate and a 4.3 percent long-term rate. The two-rate structure is the empirical signature of present bias. The short-run number is the cost of being human.
Now apply this to product design.
Buy Now, Pay Later
The dominant buy-now-pay-later product is “pay in 4.” The consumer pays exactly 25 percent of the total at checkout. The remaining 75 percent is split into three equal installments two weeks apart. The loan is interest-free if paid on time.
Map this onto the β-δ utility function and the design becomes transparent.
The immediate cost equals one quarter of the actual price. The other three quarters fall in future periods. The β term applies to all of them uniformly. For an agent with β around 0.7, which sits inside the field-estimated range for credit-card users, the perceived total cost of the purchase is not $120. It is $30 + 0.7 × $90 = $93. The product reduces the perceived price by 22 percent without changing the actual price by a cent.
The interface reinforces the math. There is no Annual Percentage Rate on the checkout screen, because for short-term installment loans the APR is technically zero. The underwriting is a soft pull, which generates no perceptible friction. The application form is autofilled. The whole loan originates in roughly four taps.
This is not careless UX. It is precision engineering.
The Consumer Financial Protection Bureau reported in 2022 that the five largest U.S. BNPL lenders originated 16.8 million loans in 2019 and 180 million in 2021. Total loan volume grew from $2 billion to $24.2 billion in two years. During the 2023 holiday season alone, consumers originated $16.6 billion in BNPL loans. The heaviest users skew toward consumers under 35, with annual incomes between $20,000 and $50,000, who also carry traditional credit-card balances and use payday loans and overdrafts.
The revenue breakdown completes the picture. Klarna’s 2024 mix was 75 percent merchant fees, 16 percent consumer service fees, and 8 percent advertising. The consumer-service line grew from 12 percent in 2022 to 16 percent in 2024. Late fees are a growing share of revenue at the largest provider. The CFPB documented that 34 to 41 percent of BNPL users have made at least one late payment. The actual default charge-off rate stays around 2 percent.
The numbers describe a product where many users are paying the late fee and very few are defaulting. That is the signature of a population with β strictly less than 1 attempting to comply with a schedule it would never have committed to if the same dollar amount had appeared on the screen at the moment of checkout.
The Original Present-Bias Product
The credit-card industry has been monetizing present bias since the 1980s. Lawrence Ausubel published the foundational analysis in 1991. He documented an industry with 4,000 competing firms, no entry barriers, and persistently elevated returns. Major issuers earned three to five times the return of standard banking from 1983 to 1988. The market satisfied every condition for textbook competition and refused to converge to a competitive rate.
Ausubel’s diagnosis was behavioral. Consumers select credit cards based on features they expect to use, like the introductory rate or the rewards. They do not select on the long-run interest rate, because they do not expect to revolve. They are naive. They intend to pay off the statement each month and they routinely fail.
The product is built around this gap. Three mechanisms compound it.
The first is minimum payment anchoring. Benjamin Keys and Chen Wang exploited formula changes in 2019 to separate liquidity constraints from psychological anchoring. They found that at least 22 percent of near-minimum payers respond to the printed minimum as an anchor, not as a binding cash constraint. The minimum is a number that appears on the bill in large type. It functions as a recommendation. Consumers pay the recommendation and the balance survives another month.
The second is the teaser rate. Haiyan Shui and Lawrence Ausubel analyzed a 2005 randomized experiment by a major issuer. Consumers strictly preferred low introductory rates with short durations over moderate ongoing rates. After the introductory period expired, 60 percent retained the card and continued to revolve at the much higher post-teaser rate. The average borrower paid $50 more per year by choosing the lower teaser. Naive present bias predicts exactly this pattern. The consumer weights the teaser period heavily and the post-teaser period through the β filter.
The third is the late fee. The CFPB has repeatedly documented the share of credit-card revenue from late fees and over-limit charges. These fall disproportionately on consumers in the lowest quintile of financial sophistication. Public-sector workers in the bottom sophistication quintile, when granted expanded credit access, increased their debt-to-income ratio by 5 percentage points more than sophisticated peers. The fees do not subsidize the cardholder population. They redistribute from naifs to sophisticates.
The revolver economics are sharp.
Most credit-card revenue comes from a minority of accounts paying interest. Most accounts do not revolve. The product is structurally a tax on consumers who underestimate their future self-control, paid to subsidize consumers who do not.
The Gamified Channel
Robinhood added a fourth surface. The platform launched commission-free trading in 2014, executed an aggressive interface gamification strategy, and converted retail equities and options trading into a real-time consumer entertainment product.
The mechanics were specific. Confetti animation on trade execution. Digital scratch-off tickets for free shares. Push notifications on price moves. A “Top Movers” list updated continuously throughout the day, showing the 20 stocks with the largest absolute percentage price changes.
Brad Barber, Xing Huang, Terrance Odean, and Christopher Schwarz published the definitive analysis of Robinhood user behavior in the Journal of Finance in 2022. They used a regression discontinuity design around the $300 million market-cap threshold for inclusion on the Top Movers list. Robinhood users disproportionately purchased stocks just above the threshold, because those stocks appeared on the list. The 20-day average abnormal return on the top stocks purchased by Robinhood users was negative 4.7 percent.
Users were not buying losers because they thought losers would rebound. They were buying losers because the interface showed them losers.
The revenue model required this trading velocity. Robinhood’s S-1 disclosed that payment for order flow and transaction rebates accounted for 75 percent of revenue in 2020, rising to 80.5 percent in the first quarter of 2021. Market makers paid Robinhood for the privilege of executing retail orders, because retail orders are less informed than institutional orders and more profitable to internalize. The interface design that maximized trading volume was the same interface design that maximized revenue.
The Massachusetts Securities Division filed an administrative complaint in December 2020 alleging that the gamification tactics manipulated inexperienced customers into continuous interaction. The complaint specifically named the confetti animation. Robinhood settled for $7.5 million and removed the celebratory imagery. FINRA followed in June 2021 with a $70 million order on supervisory failures and options approvals.
The removal of the confetti was the legal acknowledgment of the mechanism.
The Historical Baseline
Payday lending established the empirical baseline against which all present-bias products are measured. The loans are small, short-term, high-interest, and structurally cousin to the BNPL “pay in 4” product if you squint. The fee is $15 to $20 per $100. The maturity is two to four weeks. The borrower writes a post-dated check.
Marianne Bertrand and Adair Morse published the field experiment in the Journal of Finance in 2011. They tested whether disclosure could correct payday borrowing behavior. Standard APR disclosures had zero effect. The Annual Percentage Rate is the textbook fix for asymmetric information and it changed nothing. What worked was reframing the dollar cost. When borrowers saw the cumulative dollar cost of the loan rolled over three times, in the same denomination as the loan itself, take-up dropped by 11 percent over four months. The absolute reduction was 5.4 percentage points.
The intervention worked because it raised the present-tense salience of the future cost. It moved the future dollar penalty back through the β filter and onto the same side of the cognitive ledger as the immediate dollar reward.
Brian Melzer published the welfare analysis in the Quarterly Journal of Economics the same year. He used geographic variation in payday-loan access to test whether credit access reduced economic hardship. The standard prediction is that liquidity smoothing should help households absorb shocks. The actual finding was the opposite. Low-income households with proximate access to payday loans were 5.3 percentage points more likely to experience severe hardship, defined as inability to pay rent, mortgage, or utilities. Access to high-interest credit did not smooth consumption. It tightened the constraint.
Payday lending is the older, less politely designed version of the same product.
The Same Insight Across Surfaces
The four products share one structural property. Each of them has a revenue model that is mathematically more profitable when the user has β < 1.
BNPL collects late fees from users who agreed to a payment schedule their present-biased selves would not have agreed to in the present. Credit cards collect interest from revolvers who intended to be transactors. Trading apps collect order flow from users who trade more than the rational frequency. Payday lenders collect rollover fees from borrowers whose two-week plan extends across three or more cycles.
None of these products fail when the user behaves rationally. They simply earn less. A BNPL business with no late fees and no missed payments earns merchant fees only. A credit-card business with no revolvers earns interchange only. A trading app with patient buy-and-hold users still earns order flow, but at a tiny fraction of the volume.
The user behaviors that produce harm are the same user behaviors that produce profit.
This is the part that makes the framing morally awkward but technically sharp. The companies are not designed to harm users in some general malicious sense. They are designed to maximize revenue given the product they sell. The revenue function is monotonically increasing in β-related deviations from the user’s own stated plan. The math forces the design.
What Actually Works
The implication for product correction is uncomfortable.
Financial literacy interventions do not move the β parameter. The literature is consistent on this. Generic APR disclosures do not move it either, as Bertrand and Morse demonstrated. Generic warnings about overspending do nothing. The behavioral economics literature has spent four decades testing interventions that respect consumer autonomy and the result is a graveyard of null findings.
The interventions that work all do the same thing. They move the future cost into the present.
The 2009 CARD Act mandated that credit-card statements display the exact interest savings from paying the balance in 36 months versus making the minimum payment. Sumit Agarwal, Souphala Chomsisengphet, Neale Mahoney, and Johannes Stroebel evaluated the legislation in 2015 using 160 million accounts. The targeted disclosure increased the share of accounts paying the 36-month value by 0.5 percentage points on a 5.7 percent base. The Act cumulatively reduced borrowing costs by 1.7 percent of average daily balances. For consumers below FICO 660, borrowing costs fell by 5.5 percent. Aggregate savings reached $12.6 billion per year with no offsetting reduction in credit volume.
The Bertrand-Morse disclosure worked the same way. So did the Robinhood settlement, which forced the removal of the visual reward mechanism. The 2024 CFPB interpretive rule reclassified BNPL providers as card issuers under Regulation Z, mandating billing dispute rights, periodic statements, and refund protections. The EU Consumer Credit Directive II, applicable from November 2026, requires creditworthiness assessments on all loans including those under €200, eliminating the carve-out that made BNPL frictionless.
Each of these moves the future cost forward in time. Each of them works against the β filter rather than respecting it.
The libertarian-paternalist argument is that disclosure should be enough, that adult consumers can make their own choices given full information. The empirical record is that disclosure does not work unless it shifts the temporal frame. Once you accept that point, you are no longer in the libertarian-paternalist position. You are in the position of designing the choice architecture against the user’s revealed preference and in favor of their stated preference. You are taxing β by force.
There is a steelman to be honest about. Fintech does extend credit to consumers who would otherwise be rationed out of the traditional banking system. Marco Di Maggio and coauthors documented in 2022 that alternative-data lending identifies high-repayment users invisible to traditional bureaus. Some sophisticated users leverage commitment features to self-bind. The claim that BNPL strictly worsens welfare for all consumers is too strong.
The average user is partially naive. The median product is designed against partially naive users. The revenue model is structurally dependent on the gap between stated plans and executed behavior. The marginal welfare gain to invisible primes does not absolve the average welfare loss to everyone else.
The product is working. That is the problem.
The One-Line Version
The reason your fintech app feels like it knows you better than you know yourself is that it has a closed-form mathematical model of your self-control, and its revenue function maximizes when your self-control fails.
Sources
This piece draws on Samuelson (1937), Strotz (1955), Ainslie (1975), Laibson (1997), O’Donoghue and Rabin (1999), Frederick, Loewenstein, and O’Donoghue (2002), Laibson, Repetto, and Tobacman (2007), and Meier and Sprenger (2010) for the formal and empirical foundations of present-bias theory. The credit-card analysis relies on Ausubel (1991), Shui and Ausubel (2005), Gross and Souleles (2002), Keys and Wang (2019), Stango and Zinman (2009, 2011), and Agarwal, Chomsisengphet, Mahoney, and Stroebel (2015) on the CARD Act. The BNPL section uses CFPB reports from 2022 and 2024, Affirm 10-Q disclosures, and Klarna 2024 revenue reporting. The trading-app section draws on Barber, Huang, Odean, and Schwarz (2022), the Massachusetts Securities Division complaint (2020), the FINRA settlement (June 2021), and Robinhood’s S-1. The payday-lending baseline uses Bertrand and Morse (2011), Melzer (2011), and Di Maggio, Ratnadiwakara, and Carmichael (2022) on alternative-data lending. The regulatory section references the CFPB May 2024 interpretive rule, the FCA Deferred Payment Credit framework (effective July 2026), ASIC’s Low Cost Credit Contract regime (effective June 2025), and EU Consumer Credit Directive II (applicable November 2026).