Hyperbolic Discounting

Hyperbolic discounting is a mathematical and behavioral model of intertemporal choice in which individuals discount delayed rewards and costs via a hyperbolic (or quasi-hyperbolic) function of waiting time t rather than a constant-rate geometric decay function, causing the effective per-period discount rate to fall steeply as the time horizon moves further into the future. Originating in the experimental operant-conditioning psychology of Richard Herrnstein (1961) and George Ainslie (1975) and formalized for economics by Robert H. Strotz (1955), Edmund Phelps and Robert Pollak (1968), and David Laibson (1997), hyperbolic discounting is the central framework in business and economics theory for analyzing present bias, dynamic (time) inconsistency, self-control problems, household undersaving, credit-card borrowing, procrastination, and demand for illiquid commitment devices.

In Paul Samuelson's (1937) classical Discounted Utility (DU) model—which underpins standard net present value corporate finance, dynamic programming, and the Ramsey-Cass-Koopmans Solow-Swan model of economic growth—an agent discounts utility u(c_t) received t periods in the future by the exponential decay factor D_DU(t) = δ^t = exp(−ρt), where ρ = −ln δ is a constant instantaneous discount rate independent of t. A defining property of Samuelson's constant-rate discounting is stationarity (time consistency): if an agent at date t = 0 prefers receiving 110 USD in 31 days over receiving 100 USD in 30 days, then when 30 days have elapsed, the exact same agent will still prefer waiting 1 more day for 110 USD over taking 100 USD immediately. Empirical and experimental studies consistently refute stationarity: human and animal subjects exhibit high impatience over immediate trade-offs (today vs. tomorrow) alongside substantial patience over trade-offs separated by the same interval in the distant future (day 30 vs. day 31), producing preference reversals as the payoff date approaches.

Mazur's Hyperbolic Law and Loewenstein-Prelec Generalized Hyperbolas

In quantitative behavioral analysis (Mazur, 1987), the subjective present value V(x, t) of a reward of utility u(x) delayed by t ≥ 0 units of time is governed by the single-parameter hyperbolic discount function:

where k > 0 is an individual impatience parameter measuring the degree of hyperbolic steepness. More generally, George Loewenstein and Drazen Prelec (1992) derived the two-parameter generalized hyperbola:

In the limit as α → 0, D_GH(t) converges to the classical exponential decay function exp(−γt); when α = γ = k, it reduces to Mazur's simple hyperbola 1 / (1 + k·t).

To see why hyperbolic discounting generates dynamic inconsistency, compute the instantaneous (hazard) discount rate ρ(t) = −[ dD(t) / dt ] / D(t) at delay horizon t:

• Under Classical Exponential Decay D(t) = exp(−ρ₀t):

• Under Generalized Hyperbolic Discounting D_GH(t) = (1 + αt)^−γ/α:

At t = 0, the instantaneous discount rate is high (ρ_GH(0) = γ), reflecting intense urgency for immediate gratification; as the horizon t grows large, the instantaneous discount rate decays hyperbolically toward zero (lim_t→∞ ρ_GH(t) = 0).

Consider a classic numerical illustration of a preference reversal under Mazur's hyperbola D_H(t) = 1 / (1 + t) (with k = 1 per day) between a Smaller-Sooner (SS) reward of u_SS = 100 available at t₁ and a Larger-Later (LL) reward of u_LL = 110 available one day later at t₂ = t₁ + 1:

• Viewed from the Distant Present (t₁ = 30 days, t₂ = 31 days):

V(SS) = 100 / (1 + 30) = 3.226, whereas V(LL) = 110 / (1 + 31) = 3.438. Because 3.438 > 3.226, at day 0 the agent strictly prefers the patient Larger-Later reward of 110 on day 31.

• Viewed When Day 30 Arrives (t₁ = 0 days, t₂ = 1 day):

V(SS) = 100 / (1 + 0) = 100.0, whereas V(LL) = 110 / (1 + 1) = 55.0. Because 100.0 > 55.0, the moment day 30 becomes "today," the agent's preference reverses and they impulsively seize the Smaller-Sooner reward of 100.

Laibson's Quasi-Hyperbolic (β-δ) Discrete-Time Model

While continuous hyperbolas capture experimental choice data with high precision, they are analytically cumbersome in multi-period macroeconomic consumption-saving models. To preserve recursive tractability, David Laibson (1997)—adapting the intergenerational altruism function of Phelps and Pollak (1968)—introduced the discrete-time quasi-hyperbolic (or β-δ) discount model. At calendar date t, the agent's intertemporal utility U_t over the consumption stream (c_t, c_t+1, c_t+2, ...) is:

This produces the discrete sequence of discount factors { 1, βδ, βδ², βδ³, ... }, governed by two distinct parameters:

• Long-Run Discount Factor δ ∈ (0, 1): Captures standard time-consistent patience between any two future periods t + τ and t + τ + 1 (where the ratio of discount factors is βδ^τ+1 / (βδ^τ) = δ, typically calibrated around δ ≈ 0.96–0.99 annually).
• Present-Bias Parameter β ∈ (0, 1]: Captures the immediate drop in valuation between the present period t ("now") and the very next period t + 1, where the discount ratio is βδ < δ (empirically estimated across laboratory and field studies at β ∈ [0.60, 0.85]; for β = 1, the model collapses back to Samuelson's time-consistent geometric discounting δ^t = exp(−ρt)).

Strotz-Pollak Equilibrium, Generalized Euler Equation, and O'Donoghue-Rabin Types

Because a hyperbolic discounter's preferences change over time, an individual cannot be modeled as a single unified optimization agent; instead, following Strotz (1955) and Laibson (1997), the individual is modeled via game theory as a sequence of autonomous temporal selves (Self 0, Self 1, ..., Self t, ...), where Self t controls consumption c_t in period t but disagrees with future selves over how much of remaining wealth W_t+1 should be saved at t + 1. In a Markov Perfect Nash equilibrium among temporal selves, consumption follows the Generalized Euler Equation (Harris and Laibson, 2001):

where R = 1 + r is the gross return on savings and c′(W_t+1) ∈ (0, 1) is the next self's marginal propensity to consume out of wealth. Because current Self t recognizes that future Self t + 1 will overconsume at rate c′(W_t+1), the effective discount factor lies between βδ (when c′ = 1) and δ (when c′ = 0).

In their influential 1999 analysis, Ted O'Donoghue and Matthew Rabin demonstrated that behavioral outcomes under quasi-hyperbolic discounting depend critically on the agent's self-awareness (beliefs about future present bias b ∈ [β, 1]):

• Naive Agents (b = 1): Believe that while they are present-biased today (β < 1), starting tomorrow all future selves will behave like time-consistent geometric maximizers (b = 1). Consequently, naifs repeatedly procrastinate unpleasant tasks with immediate effort costs (such as filing taxes, exercising, or canceling high-fee subscriptions), perpetually planning to do the task "tomorrow" while failing to follow through when tomorrow arrives.
• Sophisticated Agents (b = β): Correctly foresee that their future selves will also exhibit present bias β < 1. Sophisticates avoid multi-period procrastination traps through backward induction and exhibit a strictly positive willingness to pay for commitment devices—arrangements that restrict the choice set or impose penalties on future Self t + 1 to prevent impulsive defection (parallel to Ulysses having himself bound to the mast to resist the Sirens).
• Partially Naive Agents (β < b < 1): Recognize that they have self-control problems (b < 1) but underestimate their true severity (b > β), often purchasing weak commitment devices (such as expensive annual gym memberships, studied by DellaVigna and Malmendier, 2006) that they subsequently underutilize.

Household Finance, Golden Eggs, and Save More Tomorrow (SMarT)

Hyperbolic discounting resolves major empirical puzzles in household macroeconomics and contract mechanism design:

1. The Credit Card Debt Puzzle and Illiquid Wealth ("Golden Eggs"): Laibson, Repetto, and Tobacman (2007) explained why typical U.S. households simultaneously hold substantial illiquid wealth (home equity and 401(k) retirement accounts earning ~5–7% real returns) while carrying revolving credit-card debt at 15–20% real interest rates. Under β ≈ 0.70 and δ ≈ 0.96, the long-run self values retirement accumulation and uses illiquid "golden egg" assets (houses and penalty-locked pensions) as a commitment device that the impatient short-run self cannot liquidate, while the short-run self borrows up to its liquid credit-card limit for immediate consumption.
2. Save More Tomorrow (SMarT, Thaler and Benartzi, 2004): Because workers with β < 1 and loss aversion λ > 1 resist immediate cuts to take-home pay, Thaler and Benartzi designed the SMarT pension program, which asks employees today (t = 0) to commit to allocating a portion of their future nominal pay raises (t ≥ 1) into 401(k) savings. Because both t and t + 1 lie in the future when the contract is signed, the decision is governed by the patient long-run factor δ rather than βδ, more than tripling average employee savings rates from 3.5% to 13.6% over four pay raises.

What is the magnitude effect in empirical intertemporal discounting?

Thaler (1981) and Loewenstein and Prelec (1992) documented that implied discount rates are not only declining in time t (hyperbolic discounting), but also declining in the monetary magnitude of the reward: subjects might demand 60 USD in one year to forego 15 USD today (an implied annual discount rate of 300%), but demand 4,000 USD in one year to forego 3,000 USD today (an implied discount rate of 33%). Similarly, the sign effect shows that individuals discount delayed gains at a much higher rate than delayed losses.

How does dual-system neuroeconomics explain the β and δ parameters?

In a landmark fMRI study, Samuel McClure, David Laibson, George Loewenstein, and Jonathan Cohen (2004) scanned human subjects making choices between immediate and delayed monetary rewards. Choices involving an immediate reward (t = 0, activating the β system) selectively engaged paralimbic structures associated with midbrain dopamine reward processing (ventral striatum, medial orbitofrontal cortex, and medial prefrontal cortex), whereas all intertemporal choices regardless of delay (activating the δ system) engaged the lateral prefrontal cortex and posterior parietal cortex associated with deliberative executive planning.

Can uncertainty or mortality hazard produce hyperbolic discounting among rational agents?

Yes. As Sozou (1998) and Dasgupta and Maskin (2005) proved using Bayesian inference, if a delayed reward faces an unknown constant hazard rate λ of cancellation or non-delivery and the agent holds an exponential prior distribution p(λ) = k⁻¹ exp(−λ/k) over the unknown hazard rate, the marginal survival probability of the reward after waiting time t integrates to ∫₀^∞ exp(−λt) k⁻¹ exp(−λ/k) dλ = 1 / (1 + k·t). Thus, Mazur's exact hyperbolic discount function emerges as the Bayesian optimal valuation of a reward under uncertain environmental hazard rates.

1. Strotz, R. H. (1955). "Myopia and inconsistency in dynamic utility maximization." Review of Economic Studies, 23(3), 165–180. https://doi.org/10.2307/2295722
2. Loewenstein, G., & Prelec, D. (1992). "Anomalies in intertemporal choice: Evidence and an interpretation." Quarterly Journal of Economics, 107(2), 573–597. https://doi.org/10.2307/2118482
3. Laibson, D. (1997). "Golden eggs and hyperbolic discounting." Quarterly Journal of Economics, 112(2), 443–478. https://doi.org/10.1162/003355397555253
4. O'Donoghue, T., & Rabin, M. (1999). "Doing it now or later." American Economic Review, 89(1), 103–124. https://doi.org/10.1257/aer.89.1.103
5. McClure, S. M., Laibson, D. I., Loewenstein, G., & Cohen, J. D. (2004). "Separate neural systems value immediate and delayed monetary rewards." Science, 306(5695), 503–507. https://doi.org/10.1126/science.1100907