The Hyperbolic Consumption Model: Calibration, Simulation, and Empirical Evaluation
📄 Summarized from the full manuscript · Human-reviewed for faithfulness before publication
In brief
Why do people simultaneously run up credit-card debt and build home equity -- behavior that looks both reckless and prudent at once? This paper simulates life-cycle households whose discount rate is much higher short-run than long-run ("hyperbolic" discounting) and compares them with standard households, calibrated to hold the same total wealth before retirement. The hyperbolic households hold far less liquid wealth, borrow much more on revolving credit, and cut consumption much more sharply at retirement -- all matching survey data better than the standard model, showing hyperbolic discounting explains several puzzling saving patterns at once. Sourced from the authors' working paper, not the final published text.
What this paper finds — and why it matters
A note on provenance. No copy of the Journal of Economic Perspectives article itself could be retrieved (AEA’s own site serves it from behind a Cloudflare bot challenge that blocked repeated scripted and in-browser attempts). This page instead rests on the underlying technical working paper by the same five authors, Angeletos, Laibson, Repetto, Tobacman, and Weinberg, “Hyperbolic Discounting, Wealth Accumulation, and Consumption” (Centro de Economia Aplicada, Universidad de Chile, Serie Economia No. 90, July 2000). The working paper’s own abstract is word-for-word the abstract Crossref and OpenAlex deposit for the published JEP article, and the two share all five authors and the same model, calibration, and headline results, so the working paper is almost certainly the technical basis the JEP “Perspectives” piece condenses – but the JEP text itself may re-sequence, retitle, or add exposition not present here. Treat section/page references below as pointing to the 2000 working paper, not to the 2001 JEP pagination (pp. 47-68).
This 2001 Journal of Economic Perspectives paper by George-Marios Angeletos, David Laibson, Andrea Repetto, Jeremy Tobacman, and Stephen Weinberg – sourced here from the authors’ own underlying 2000 working paper (see provenance note above) – presents lifecycle buffer-stock simulations comparing households with quasi-hyperbolic discounting (a short-run discount factor beta = 0.7 applied on top of a long-run discount factor delta) against otherwise-identical exponential-discounting households, with both types calibrated so that each matches the same median ratio of pre-retirement (ages 50-59) wealth to income observed in the Survey of Consumer Finances. Both household types face the same technology: a liquid asset that can be borrowed against up to a credit-card limit set at 30 percent of average income, and a partially illiquid asset calibrated to real-estate-like transaction costs (a $10,000 fixed cost plus 10 percent of value on sale). Despite being calibrated to hold identical total pre-retirement wealth, the hyperbolic simulations hold markedly less liquid wealth and borrow far more on revolving credit than the exponential simulations (mean simulated credit-card debt of $3,408 for sophisticated hyperbolics and $4,777 for naive hyperbolics, versus $900 for exponentials, against an empirical mean of roughly $4,600), and this low-liquidity, high-debt profile in turn generates both substantially higher consumption-income comovement (an excess-sensitivity coefficient near 0.25 for hyperbolics versus about 0.03 for exponentials, against Panel Study of Income Dynamics estimates of 0.19 to 0.33) and a much larger, empirically matched drop in consumption around retirement (14.5 percent for sophisticated hyperbolics and 15.6 percent for naifs versus 3 percent for exponentials, against an estimated 11.6 percent drop in the PSID) than the exponential model produces. The paper’s claim is that hyperbolic discounting, rather than some alternative calibration of the standard exponential model, lets a single set of preference parameters jointly reconcile three separate empirical regularities – low liquid-asset holdings, heavy revolving-credit use, and excess sensitivity of consumption to predictable income changes – that the exponential benchmark cannot match simultaneously.
Summary of a classic paper, AI-assisted and human-reviewed. See the linked original for the authoritative claims and full conditions.
Questions & answers
Q1. What is the paper’s basic research design?
The authors simulate three types of lifecycle households – standard exponential-discounting households, “sophisticated” hyperbolic households (who correctly anticipate that their future selves will not honor current plans), and “naif” hyperbolic households (who mistakenly believe future selves will carry out today’s plans) – and calibrate each type’s discount factor separately so that all three match the same empirical target: the median ratio of total wealth to income among households aged 50-59 in the Survey of Consumer Finances. Because the three types are calibrated to match identical total pre-retirement wealth by construction, any differences that emerge in how that wealth is held (liquid versus illiquid), in credit-card borrowing, or in the comovement of consumption with income, are differences the discounting structure itself generates, not artifacts of different savings rates.
Q2. What is the quasi-hyperbolic discount function, and why do the authors use it instead of the fully general hyperbolic form?
The authors adopt the discrete-time “quasi-hyperbolic” function with values {1, betadelta, betadelta^2, …}, a tractable approximation to the generalized hyperbolic discount function that was first used by Phelps and Pollak (1968) for a different purpose (imperfect intergenerational altruism) and later adopted by Laibson (1997a) to approximate experimentally estimated time preference. Laboratory and field estimates consistently approximate generalized hyperbolas in which the implied one-year discount rate is far higher than the long-horizon annualized discount rate; the quasi-hyperbolic form reproduces this sharp short-run/long-run split with just two parameters, beta (set to 0.7 in the simulations) and delta (set to 0.957), values the authors describe as chosen so that, “using annual periods, these parameter values roughly match experimentally measured discounting patterns.”
Q3. How is the life-cycle model’s asset structure set up?
Households choose net investment in a liquid asset X, which can be borrowed against up to a credit limit set at 30 percent of average income at that age (calibrated from actual 1995 SCF credit limits), and in a partially illiquid asset Z that is bounded below by zero and pays no return but a consumption flow equal to 5 percent of its value, with transaction costs of a flat $10,000 plus 10 percent of value charged only on sales (not purchases), “chosen to match those that arise in real estate transactions, including brokerage commissions, legal fees, bank fees, moving costs, and search costs.” Interest rates differ by liquid-asset position: credit-card debt carries a real rate 3 percentage points below the Federal Reserve’s mean debt-weighted real interest rate (to implicitly capture bankruptcy risk), while positive liquid balances earn 3.75 percent, reflecting a diversified stock/bond portfolio taxed at an effective 25 percent rate.
Q4. How does the calibration pin down the discount factors for each household type?
Holding the credit limit, transaction costs, demographics, and income process fixed (estimated from the Panel Study of Income Dynamics, matching Hubbard, Skinner, and Zeldes 1994’s parameter values closely), the authors search over the discount factor for each household type until the simulated median wealth-to-income ratio for ages 50-59 matches the SCF target of 3.2. This yields delta = 0.9437 for the exponential model, delta = 0.9571 for the sophisticated hyperbolic model, and delta = 0.9495 for the naif model – the hyperbolic deltas must exceed the exponential delta because hyperbolic consumers have a second source of discounting (beta) pulling against long-run patience, so that – in the authors’ words – the calibration is meant to “equalize” the underlying willingness to save between the exponential and hyperbolic consumers.
Q5. What is the difference between “sophisticated” and “naif” hyperbolic consumers, and how is each modeled as an equilibrium?
Sophisticated hyperbolic consumers are modeled, following Strotz (1956), as a sequence of rational players in an intra-personal game: each self picks the strategy optimal from its own perspective taking later selves’ (correctly anticipated) strategies as given, and the equilibrium is solved by numerically implemented backwards induction. Naif consumers, by contrast, “have optimistic forecasts in the sense that they believe that future selves will carry out the wishes of the current self”: the current self computes and implements the first action of a full plan that maximizes the current self’s preferences, only to have each subsequent self recompute and deviate from that plan in turn. The paper reports that “the results of the sophistication case and the naif case are generally quite similar”, so most of the discussion focuses on the exponential and sophisticated-hyperbolic cases.
Q6. How can hyperbolic households simultaneously look highly impatient (willing to borrow at 15-20 percent on credit cards) and highly patient (willing to accumulate illiquid wealth for decades)?
The model’s Hyperbolic Euler Relation shows that the effective discount factor relevant to a marginal dollar of liquid saving is a weighted average of the short-run factor (beta*delta) and the long-run factor (delta), with weights given by the future marginal propensity to consume out of liquid wealth; because sophisticated hyperbolics deliberately keep themselves liquidity-constrained to prevent future overspending, that future MPC is high, pushing the effective liquid-saving discount rate toward roughly 41 percent a year. Illiquid wealth, by contrast, is spent far more slowly, so the effective discount rate governing illiquid accumulation converges (in the limit of holding wealth for many periods) to -ln(delta), about 5 percent a year – “illiquid wealth accumulation is primarily driven by” the long-run factor “not” the short-run one, which is why the same household can rationally run up revolving-credit balances while also building a large stock of home equity or retirement assets.
Q7. What do the simulations find about liquid-wealth holdings and credit-card borrowing, compared to the data?
At any point in time, only 19 percent of simulated exponential consumers borrow on credit cards, compared to 51 percent of hyperbolics and 60 percent of naifs, while the 1995 SCF shows that 70 percent of households with credit cards did not pay their bill in full the last time they made a payment – so hyperbolics and naifs come far closer to the observed incidence of revolving debt. In dollar terms, simulated exponential households average $900 of credit-card debt (including zero-debt households), versus $3,408 for sophisticated hyperbolics and $4,777 for naifs, against an actual average of roughly $4,600 per household with a credit card; on liquid-asset shares, using the intermediate definition of liquid assets the empirical lifecycle profile of liquid-wealth shares runs 11%, 6%, 5%, 5%, 10%, 12% across age groups 20-29 through 70+, versus 97%, 65%, 35%, 20%, 27%, 57% for simulated exponentials (average gap +42 points) and 86%, 46%, 24%, 13%, 12%, 56% for hyperbolics (average gap +31 points) – hyperbolics outperform exponentials but do not fully close the gap to the data.
Q8. What do the comovement regressions find?
Regressing log consumption growth on expected log income growth (controlling for a full set of age dummies) using simulated data, the exponential model generates an excess-sensitivity coefficient around 0.03 (0.058 if consumption is restricted to non-durables), while the hyperbolic and naif simulations consistently generate coefficients above 0.16. Replicating the regression on actual PSID data from 1978-1992, the authors obtain statistically significant estimates “between .19 and .33” – “far above the value implied by the exponential simulation (0.03 to 0.06)” and in line with a wider empirical literature the paper cites running from Hall and Mishkin’s (1982) 0.200 to Souleles’s (1999) 0.344.
Q9. What do the simulations find about the drop in consumption around retirement?
Estimating a regression of consumption growth on dummy variables for the years surrounding a fixed, fully anticipated retirement age (63 in the simulations), the exponential model predicts a total consumption drop around retirement of 3 percent, versus 14.5 percent for sophisticated hyperbolics and 15.6 percent for naifs. Running the analogous regression on PSID data (following the retirement-puzzle literature of Banks, Blundell, and Tanner 1998 and Bernheim, Skinner, and Weinberg 1997), the authors estimate an actual total consumption drop around retirement of 11.6 percent – a magnitude “little changed if we exclude households who retired at ages 55 or younger” – so that, as with liquid-asset holdings and comovement, the hyperbolic and naif simulations come much closer to the observed data than the exponential simulations, even though retirement in the model is a completely predictable, non-income-shock event.
Q10. What is the paper’s overall conclusion, and what does it NOT claim?
The stated conclusion is that, calibrated to match identical pre-retirement wealth accumulation, hyperbolic (and naif) households simultaneously reproduce three separate empirical patterns the exponential model misses: low liquid-wealth holdings relative to total wealth, a high incidence and level of revolving credit-card debt, and excess sensitivity of consumption to predictable income changes, including the drop in consumption around retirement – so that “the hyperbolic buffer stock model matches observed consumption and wealth data better than the exponential buffer stock model.” The paper does not claim the fit is perfect: on several measures (e.g., liquid-wealth shares in Table 2) the hyperbolic simulations improve on, but do not close, the gap to the data, and for ages 60-69 the authors note that the exponential and hyperbolic simulations “both do poorly, since the data lie roughly between the predictions of the two models” – only after age 70 does the exponential model outperform the hyperbolic one on the liquid-asset-holding measure.
Key terms in this paper
Definitions below follow the paper's own usage.
- Quasi-hyperbolic discounting
- the discrete-time discount function with values {1, beta*delta, beta*delta^2, beta*delta^3, ...} (Sec. 2, following Phelps and Pollak 1968 and Laibson 1997a), an analytically tractable two-parameter approximation to the generalized hyperbolic discount function that reproduces the sharp wedge laboratory and field studies find between short-run discount rates (governed jointly by beta and delta) and long-run discount rates (governed by delta alone); the paper sets beta = 0.7 and delta calibrated separately for each household type.
- Sophisticated versus naive hyperbolic consumers
- the two ways the paper resolves the dynamic inconsistency hyperbolic discounting creates among a household's temporally distinct 'selves' (Sec. 2-3.6): a sophisticated consumer correctly anticipates that later selves will not honor the current self's plans and solves an intra-personal game by backward induction (following Strotz 1956), while a naive consumer mistakenly believes future selves will carry out the current self's intended plan, computes that plan, and implements only its first action before the next self recomputes and deviates in turn.
- Hyperbolic Euler Relation
- the equilibrium condition (Sec. 4, generalizing the standard Exponential Euler Relation, following Harris and Laibson 2000) under which the effective discount factor governing a sophisticated hyperbolic consumer's marginal liquid saving is a weighted average of the short-run factor beta*delta and the long-run factor delta, with the weight on the short-run factor equal to the expected future marginal propensity to consume out of liquid wealth -- the mechanism by which the model generates simultaneously high effective discount rates on liquid saving (hence heavy credit-card borrowing) and low effective discount rates on illiquid accumulation.
- Liquidity-constraint self-commitment
- the paper's account (Sec. 4) of why sophisticated hyperbolic households deliberately hold a disproportionately low share of wealth in liquid form: by keeping future selves liquidity-constrained, an early self raises the future marginal propensity to consume out of liquid wealth, which -- via the Hyperbolic Euler Relation -- raises the effective discount rate applied to liquid saving and thereby restrains future overspending, making illiquid assets (e.g., housing, retirement accounts) a commitment device against the household's own future impatience.
- Excess sensitivity of consumption to income
- the regression coefficient (Sec. 5.2, eq. 6) on expected income growth in a regression of consumption growth on that expectation plus controls, which should be zero under the frictionless permanent-income hypothesis; the paper's empirical PSID estimates cluster between 0.19 and 0.33, which the calibrated hyperbolic and naif simulations approximate (coefficients above 0.16) but the exponential simulations do not (coefficients of 0.03 to 0.06), because hyperbolic households' low liquid-asset buffers make them more likely to hit binding liquidity constraints when income changes.