Hyperbolic Discounting

Why discounting shape matters

Every intertemporal choice — save or spend, study or scroll, smoke or quit, take the cheap policy this quarter or pay the political cost of reform — compares a stream of utilities at different points in time. To compare them, an economist needs a rule for translating a future utility into a present equivalent. The classical rule is exponential: a unit of utility t periods from now is worth δᵗ today, where δ is a single per-period discount factor. The rule has one defining property — its ratio is constant. The trade-off between today and tomorrow is the same δ as the trade-off between year 30 and year 31. That property is what makes exponential agents time-consistent: a plan made today is still optimal tomorrow.

That assumption produces models that work beautifully on paper and predict behavior that almost nobody actually exhibits. Real humans treat today as categorically different from the future. Real humans plan to start a diet on Monday and break it by Tuesday lunch. Real humans repeatedly sign up for gyms they never visit. Hyperbolic discounting is the empirical, calibrated correction that lets formal intertemporal models track those behaviors. It is the single most-cited departure from rational-choice economics, and it underwrites most of behavioral macro.

The shape of the curve

Behaviorally elicited discount functions decay faster than any exponential at short horizons and slower than any exponential at long horizons. A clean way to write this is the hyperbolic form due to Mazur (1987):

D(t) = 1 / (1 + k t)

where k > 0 is a steepness parameter. The annualised "implicit interest rate" demanded to wait is therefore not constant: it is huge at short delays (you have to bribe me a lot to wait a week) and shrinks toward zero at long delays (a year tomorrow is barely different from a year-and-a-week tomorrow). That single curve already qualitatively explains preference reversal — see below — but it is mathematically inconvenient for dynamic optimisation. The standard tractable approximation, used in almost all modern applied work, is the quasi-hyperbolic β-δ form of Phelps and Pollak (1968) and Laibson (1997):

U_t = u_t + β · Σ_{s≥1} δ^s · u_{t+s},     β ∈ (0,1), δ ∈ (0,1)

Today's utility u_t enters with weight 1. Every future period gets an extra one-time discount β on top of the usual δ. The function is exponential beyond the present period — making it tractable — but with a discrete kink at t = 0 that captures the over-weighting of "right now". Setting β = 1 recovers the standard exponential model. The case β < 1 is called present bias.

The preference reversal

The cleanest demonstration that real discounting cannot be exponential is the preference reversal experiment, run in dozens of variants since the 1980s. The form is always the same. At time t = 0:

• Ask: "Would you rather have $100 in 30 days, or $110 in 31 days?"
• Most respondents choose $110 in 31 days. The extra day is cheap when both options are weeks away.

Now wait 30 days. Re-ask the same offer, which has now slid forward in time:

• "Would you rather have $100 today, or $110 tomorrow?"
• A substantial fraction flip and choose $100 today.

Under exponential discounting this flip is impossible. The two offers were shifted by exactly the same 30 days, so their ratio is unchanged: if you preferred $110 the first time, you must prefer $110 the second time. The shift cannot reorder preferences. Under β-δ, the second comparison crosses the today-versus-future line: the $100 option is now in the present and escapes the β factor entirely, while the $110 option is still in the future and pays β. The ratio is multiplied by 1/β ≈ 1.4, more than enough to flip the ranking. Same person, same money, same delays — different choice — because one delay sits inside today and one sits outside it.

Time inconsistency and the war between selves

Once preferences can reverse as a function of when you ask, plans made today no longer bind tomorrow. Today's self looking at next Monday sees only a discounted version of next Monday's costs; next Monday's self looking at next Monday's costs sees them at full β-less weight. The two disagree. Strotz (1955) showed this formally and laid out three possible responses:

• Naïve. The agent does not know they will defect, repeatedly plans optimal long-run behavior, and is surprised every period when their plan unravels.
• Sophisticated. The agent correctly anticipates that their future selves will defect, and plays an intrapersonal game against them. Equilibrium strategies look like reduced consumption today plus an active demand for commitment devices.
• Resolute. The agent precommits early and binds future selves by external machinery — locked savings, public stakes, technological friction.

Modern applied work usually models a population mixture: most people are partly sophisticated about themselves, which is why commitment-device markets exist at all.

What present bias predicts in the wild

Domain	Behavior under exponential	What hyperbolic predicts	Observed
Retirement saving	Smooth consumption, save to target	Under-saving; saving rate rises only with default-enrollment	Median U.S. household near-zero retirement wealth at 50
Procrastination	Tasks done as soon as cost is justified	Tasks delayed; sophisticates seek deadlines	Tax-filing bunches at deadline; gym attendance falls after January
Smoking & alcohol	Use only if expected lifetime utility is positive	Use today, regret tomorrow, demand quit aids	70% of U.S. smokers say they want to quit; multi-billion-dollar cessation market
Credit-card debt	Borrow only if return on investment > rate	Borrow at 20%+ APR while holding liquid savings at 1%	~50% of U.S. households revolve credit-card balances
Exercise	Optimal frequency given health value	Gym contracts unused; demand for trainers and friends as monitors	Lab studies (DellaVigna-Malmendier 2006): gym pay-per-use rationalised only if β ≈ 0.5
Public policy	Reform when long-run benefits exceed costs	Reform postponed; tomorrow always cheaper	Fiscal-rule erosion, climate inaction, pension under-funding

Each of these is a place where exponential discounting either gives the wrong sign, the wrong magnitude, or fails to predict the demand for commitment in the first place. The β-δ model produces all six patterns from one extra parameter.

Commitment devices: voluntarily binding yourself

The cleanest empirical evidence that hyperbolic discounting is more than a curiosity is the size of the market for commitment devices — voluntary arrangements people pay for in order to constrain their own later behavior. A rational exponential agent has no use for these, since their later self would never defect. Examples:

• Christmas Club accounts. Laibson's signature illustration. Banks offer deposit-only accounts with weekly contributions, no withdrawals until December, and below-market interest. Millions of Americans voluntarily used them through most of the 20th century. The product dominates a higher-rate savings account only if you expect your future self to raid the latter — i.e. only under present bias.
• Stickk.com. Users pledge a goal (exercise three times a week, write a chapter a month, quit smoking), post a financial stake, and name a referee. If they miss the goal, the stake is forfeited — often to an "anti-charity" the user dislikes (e.g. the campaign of a rival political party). Founded 2008 by Yale economists Ian Ayres and Dean Karlan. Empirical success rates are roughly 80% when forfeits are present and a referee is named.
• Retirement defaults. Madrian and Shea (2001) showed that switching a U.S. company's 401(k) from opt-in to opt-out enrollment raised participation from 49% to 86%. The default acts as a commitment device for a future self who will not bother to opt out.
• Pre-paid gyms, prepaid language apps, books bought ahead of vacation. Each pays in money or time today against a behavior the user wants the future self to undertake.
• Antabuse (disulfiram). Pharmacological commitment: ingest the daily pill and any later drink produces violent nausea. The pill is the commitment.
• Ulysses contracts. The literary archetype — tying yourself to the mast so the future self cannot follow the sirens.

The pattern is universal: people facing temptations they recognise will be costly will pay non-trivially to make the tempting option unavailable. That demand is direct revealed-preference evidence for sophisticated, present-biased discounting.

Worked policy example: Save More Tomorrow

The cleanest applied success of hyperbolic discounting in policy is Thaler and Benartzi's Save More Tomorrow (SMarT) plan, deployed first at a mid-sized U.S. manufacturing firm in 1998. The mechanism is elegant:

1. Today, workers commit to increasing their 401(k) contribution by 2-3 percentage points at the time of each future pay raise.
2. Because the increase coincides with the raise, take-home pay never falls — loss aversion is sidestepped.
3. The cost of saving more sits in the future when the commitment is made, where β shrinks it; the cost arrives at the present only after the contribution rate has already shifted.
4. Workers can opt out at any time, preserving freedom of choice (the "libertarian paternalism" core of nudge theory).

Results at the original firm: average savings rates rose from 3.5% to 11.6% after 28 months, and to 13.6% after 40 months. Eighty percent of workers who enrolled in SMarT stayed in through four raises. Variants of SMarT are now embedded in U.S. corporate retirement plan design and the Pension Protection Act of 2006. The mechanism does not fight present bias; it harnesses it, by putting the painful step where β is doing the most discounting.

Estimating β: lab, structural, and field

How confident are we in β ≈ 0.6 to 0.8? Estimates come from three classes of work:

• Lab elicitation. Subjects make choices between dated monetary rewards. Frederick, Loewenstein and O'Donoghue's 2002 meta-review of 42 studies finds individual discount rates ranging from negative to over 100% per year, with strong evidence of hyperbolic decline. Lab estimates of β cluster around 0.7.
• Structural estimation from observed wealth. Laibson, Repetto and Tobacman (2007) estimate β and δ jointly from U.S. wealth, credit-card debt, and consumption data using a sophisticated buffer-stock model. Point estimate: β ≈ 0.703, δ ≈ 0.958 per year. Crucially, the data require a kinked discount function — a single exponential fits poorly.
• Field experiments. DellaVigna and Malmendier (2006) study gym contract choice. Empirical attendance rates imply that consumers who choose monthly contracts behave as if β ≈ 0.5, well below 1. Augenblick, Niederle and Sprenger (2015) elicit time preferences over real-effort tasks (not money, which has a fungibility problem) and again recover β substantially below 1.

The pattern is robust across methods and populations: people in rich economies behave as if they apply roughly 1.3 to 1.7 times extra weight to "today" relative to any nearby tomorrow. The exact number varies — financial literacy, sleep, cognitive load, stress, and substance use all move β — but the qualitative finding does not.

Variants and extensions

• True hyperbolic (Mazur 1987): D(t) = 1/(1 + k t). The cleanest fit to lab data, but intractable in dynamic models because every period has its own marginal trade-off.
• Quasi-hyperbolic / β-δ (Phelps-Pollak 1968, Laibson 1997): The workhorse. One kink, otherwise exponential — tractable enough for general-equilibrium analysis.
• Generalised β-δ (O'Donoghue and Rabin 1999): Adds explicit naïve / sophisticated / partial-naïve distinction; characterises procrastination equilibria.
• Dual-self models (Fudenberg-Levine 2006): Reformulate present bias as a long-run "planner" plus a short-run "doer" with bounded self-control. Equivalent at the choice level to many β-δ predictions but richer at the neural level.
• Hyperbolic discounting in production economies: Krusell-Smith (2003), Harris-Laibson (2001) embed quasi-hyperbolic agents in general equilibrium; the steady-state capital stock falls relative to the exponential benchmark, qualitatively matching observed under-saving.
• Behavioral life-cycle (Shefrin-Thaler 1988): Pre-Laibson cousin; separates wealth into "mental accounts" with different propensities to spend, capturing the same intuition without an explicit β.

Why a single δ can't fit the data

A surprisingly common defence of exponential discounting is "just calibrate δ correctly". The problem is that no single δ can match observed short-run impatience and long-run patience simultaneously. To rationalise refusing $100 today for $110 tomorrow you need a daily discount factor below 0.91 — an annual rate of nearly 100,000%. The same agent who would not wait one day at that rate would value $1 of consumption in retirement at essentially zero, predicting zero saving. But the same population does save for retirement at positive rates; structural estimates of δ from long-horizon wealth choices are routinely above 0.95 per year. The two horizons require incompatible exponentials. β-δ resolves the tension with a single extra parameter that lives only at the now/future boundary.

Where the theory is doing work today

• Behavioral macroeconomics. Quasi-hyperbolic households are now standard in heterogeneous-agent DSGE models of consumption, wealth inequality, and monetary policy transmission.
• Public health. Smoking cessation programmes, dietary commitments (StickK, DietBet), and prescription compliance reminders are built around present-bias correction.
• Retirement and savings policy. U.S. auto-enrollment, U.K. NEST, and most European pension defaults are explicit applications of hyperbolic-discounting design principles. The Pension Protection Act of 2006 codified them in U.S. law.
• Climate economics. Climate cost-benefit analysis (Nordhaus, Stern) is sensitive to discount-function choice; recent work uses hyperbolic or declining-rate functions to better match policymaker behavior.
• Consumer credit. Behavioral CFPB rule-making on payday lending, credit-card disclosure, and overdraft defaults explicitly references present bias as the harm being corrected.
• Education. School-savings-match programmes (e.g. SEED OK in Oklahoma) and graduation-incentive nudges target the same intertemporal asymmetry.

Common pitfalls

• Confusing hyperbolic discounting with high impatience. A patient agent with low δ but β = 1 is still time-consistent. The defining property of hyperbolic discounting is the kink at t = 0, not the steepness overall.
• Reading β as a permanent personality trait. Empirically β shifts with sleep, stress, cognitive load, intoxication, and adolescent vs adult brain state. It is a state variable, not a fixed parameter.
• Treating commitment-device demand as irrationality. A sophisticated β-δ agent's demand for commitment is fully rational given the underlying preferences — it is the optimal response of a planner who knows the next-period self has different preferences. Welfare analysis hinges on which self's preferences count.
• Assuming nudges always work. Defaults and SMarT-style programmes succeed because they exploit present bias. They can also entrench bad outcomes (default sub-optimal funds, opt-out subscriptions). The mechanism is neutral; the welfare sign depends on alignment with the long-run preferences.
• Mixing monetary and real-effort discounting. Money is fungible, so monetary discount-rate estimates are confounded by liquidity and arbitrage opportunities. Real-effort and consumption-good discount studies (e.g. Augenblick-Niederle-Sprenger) usually find larger β-deviations than monetary studies — present bias over money is the lower bound.

A brief intellectual history

The mathematical seeds were planted by R. H. Strotz in 1955, who proved that any non-exponential discounting produces dynamic inconsistency, and by Phelps and Pollak in 1968, who introduced the β-δ form in the context of intergenerational savings. Psychological evidence of hyperbolic discounting accumulated through animal behavior studies in the 1970s (Ainslie 1975, Mazur 1987) before crossing into human studies (Thaler 1981). David Laibson's 1997 paper, "Golden Eggs and Hyperbolic Discounting", is the watershed: it embedded β-δ in a workhorse intertemporal consumption-savings model, derived testable comparative statics, and gave the framework its name in economics. The 2017 Nobel Prize to Richard Thaler recognised behavioral economics broadly; hyperbolic discounting is one of the two or three results most often cited in the prize summary.