Two-asset HANK

US HANK

The parameters it solves against the paper, what is imposed, and the limits.

Partial replication, not a forecast-accuracy claim

Solved parameters against the paper’s Table B.III; the headline targets are inputs, not results.

What was checked, against what, and the result
CheckAgainstResultVerdict
Solved steady-state parametersAuclert et al. (2021) Table B.III, gate under 5e−4, largest gap 4.2e−4; vphi after × μw²6 of 7 direct, 1 after conversionpass
Goods-market clearing (untargeted)Walras’ law: should be zero3.3e−08pass
Impulse-response signsTheory, three shock kinds, both directions; spending impact multiplier in (0, 1)all gated by testspass
Impulse-response magnitudesThe one magnitude the paper states: impact consumption after a −5pp monetary shock, ρ 0.6 (Fig. 7a), 1.46%; shock units inferred. Every other response is regression-pinned only1.485%, within about 2%partial
First-order accuracyThe model’s own nonlinear solve; the intercept easing shock within 4%, spending 1% of GDP is not0.175 vs 0.251 / 0.125partial
Spending multiplier (ρ 0.8)No published benchmark; 0.376 on impact in the opt-in 0.20-MPC calibration0.175 impact, −0.015 over 20qweak
Hand-to-mouth shareAbout one third of US households (Kaplan, Violante and Weidner 2014)0.537, of which 0.518 wealthyfail
Quarterly MPC out of a rebate0.15–0.25 in US rebate studies; rebates of $500 / $1,000 / $1,500 per household, refined grid0.104 / 0.091 / 0.083fail

Evidence by dimension

Evidence by dimension: the registry’s categorical judgement, and what would change it
DimensionLevelWhat would change it
Implementation fidelitystrongReplication gates across all shocks and variants
Predictive validationnot applicableNone: not a forecaster
Identificationnot applicableNone: calibrated, not an identified VAR
Policy counterfactualsweakA published HANK multiplier to check against
UncertaintyweakSensitivity to price, wage and MPC calibration
Vintage reproducibilitystrongNone while calibration stays frozen
About this replicationIndependent build, not endorsed

An independent replication of a published academic model, built on the authors' own sequence-jacobian toolkit, not produced or endorsed by them. Results are model counterfactuals, not official estimates or forecasts.

What it solves, and what is imposed

1

The parameters it solves for

The ground truth is the published paper, not data.

US HANK: internally solved parameters vs Auclert et al. (2021) Table B.III
parameterpaperours|diff|
betadiscount factor, solved numerically0.9760.97627392.7e-04
chi1portfolio adjustment cost scale, solved numerically6.4166.416424.2e-04
ZTFP, solved analytically0.4680.46778982.1e-04
alphacapital share, solved analytically0.330.32994925.1e-05
mu_pprice markup, solved analytically1.0151.0152282.3e-04
taulabour tax rate, solved analytically0.3560.35606066.1e-05
vphiwage Phillips curve disutility scale, after × μw²2.0732.0733063.1e-04

Six of seven agree to the last published digit directly; vphi solves to 1.7135 and matches after multiplying by μw² = 1.21, because Table B.III follows the working paper's wage-markup convention and this implementation the published appendix's (us-hank-model #15).

Calibration inputs, held fixed through the solve, so their deviations are identities:
Constants equalling themselves: Output Y 1, Capital K 10, Real rate r 0.0125, Government spending G 0.2, Government debt Bg 2.8.
Solver targets: Total wealth A+B 14 hit to 2.7e-13, Liquid assets B 1.04 hit to 9.5e-14 — tolerances, not accuracies.
Market clearing: goods 3.3e-08 untargeted (Walras’ law), assets 3.1e-13 is the solver’s own target.

Calibration, not replication: only the rows the model solves for are the replication. Goods-market clearing is untargeted, so it is the check that counts.

Source: US HANK replication

Source details

Source: papers/us-hank/figures/replication.json.

How each parameter is solvedThe vphi convention, and when it was solved
  • β 0.9762739 against 0.976, and χ1 6.4164196 against 6.416. χ1 is solved numerically from a 6.5 starting guess, the strongest single check.
  • Z, α, μp and τ are solved analytically.
  • The seventh, vphi, solves to 1.7135 against a published 2.073: a convention, not an error.
  • This implementation follows the published Econometrica appendix, which divides the after-tax wage by μw. The NBER working paper multiplies by μw instead, giving exactly μw² = 1.21 times ours, 2.0733.
  • Both versions print 2.073, so Table B.III follows the working paper. The two values describe the same steady state.
  • The model repo now asserts the conversion (us-hank-model #15).
  • Solved 2026-07-29 at production grids (nB=50, nA=70, nK=50) and re-solved independently 2026-08-13. The seven parameters and the goods-market residual are the same in the model’s docs/VALIDATION.md at the pinned commit 58353e5f.
2

What is gated, and what is not

Imposed inputs carry no information; signs, timing and one published magnitude are gated, other magnitudes are only regression-pinned.

The one published magnitudeFig. 7a: units, source and the nonlinear check
  • One published magnitude is reproduced within about 2%: impact consumption after a −5pp Taylor-rule shock with quarterly persistence 0.6 (Fig. 7a of Auclert et al. 2021).
  • Linear solution: 1.485% against the paper’s 1.46% (+1.7%). Nonlinear: 1.388% against 1.36% (+2.1%) (tests/test_published_magnitude.py).
  • Units are inferred: the paper does not state them. Annualised percentage points (−1.25pp per quarter) fits; the quarterly reading gives 5.94%.
  • The paper’s numbers are read from the working-paper text (NBER w26123, Fig. 8a there); the Econometrica text was not checked for them.
  • Other magnitudes are not gated: output, investment, inflation and the spending and productivity responses are sign-checked and regression-pinned only.

Where this sits among other open HANK implementations: the landscape report.

More detailnotes

Status of each result

Imposed
Y, K, r, G, Bg — identities
Solver-reached
total wealth, liquid assets — tolerances
Asset clearing
about 3e−13, the solver’s own target
IRF signs, timing
gated
Impact consumption, Fig. 7a shock
gated within 3% of the paper
Other IRF magnitudes
regression-pinned on this implementation; no published response to match

Known limitations

Not a forecaster

No data, no estimation, no forecast origin: every result is an impulse response around the paper's calibrated steady state. Nothing in this suite forecasts the US.

Stylised instruments only

Monetary, fiscal-spending and productivity shocks; the labour tax is endogenous. No transfer or tax-rate instrument, so score_reform refuses model="hank".

First-order responses
  • Hosted responses are linear. Hosted IRFs scale with shock size: no state dependence, no zero lower bound, and a rise mirrors a cut.
  • The model is not. On the nonlinear transition a spending rise of 1% of GDP (ρ 0.8) gives an impact multiplier of 0.251, and a cut of 1% of GDP lowers output by 0.125% (a multiplier of 0.125), against the first-order 0.175.
  • Monetary is close to linear. The intercept easing shock (−0.0025, ρ 0.61) gives 0.511 against 0.493.
  • Larger shocks. At +2% of GDP the nonlinear solve does not converge.
  • Uncertainty. Parameter uncertainty is not quantified: the calibration is a single point.
MPCs below the empirical range
  • Default windfall. The aggregate quarterly MPC is 0.058 out of 25% of mean quarterly post-tax labour income (about $2,400 per head or $6,200 per household).
  • Rebate-sized. On a refined liquid grid, rebates of $500, $1,000 and $1,500 per household give 0.104, 0.091 and 0.083, far below the 0.15–0.25 that US rebate studies find. The grid explains little of the gap.
  • Cause. The illiquid account’s adjustment cost is smooth and convex, with zero marginal cost at zero adjustment and no fixed or linear component, so a constrained household that holds illiquid wealth puts most of a rebate into that account.
  • Status. Auclert et al. (2021) report no MPC, so this is a property of the calibration, not a replication failure (us-hank-model #6, PR #7).
Hand-to-mouth share too high
  • Level. About 54% of households (0.537) hold under the hand-to-mouth cutoff of liquid wealth, against about one third in Kaplan, Violante and Weidner (2014).
  • Composition. Almost all of it (0.518) is wealthy hand-to-mouth, against about two thirds in the data.
  • What it counts. Liquid wealth, not spending: most of these households hold illiquid wealth and spend little of a rebate. Out of a $1,000 rebate, 3.1% of households have an MPC above 0.25 (refined grid).
  • Status. The paper reports no hand-to-mouth share, so this is this implementation’s own statistic.
No hump-shaped responses

After a monetary shock, output, consumption, investment and inflation all peak on impact; empirical monetary responses are usually delayed and hump-shaped. The model has no habits or information frictions to delay them.

Policy rate and real rate
  • The policy rate is an outcome. A monetary shock moves the Taylor rule’s intercept; the rule then sets the nominal rate as intercept + 1.5 × inflation.
  • Its response depends on persistence. For the intercept easing shock of −0.0025 at ρ 0.61, the quarterly policy rate falls about 13bp on impact, less than the 25bp shift. At ρ 0.9 inflation moves by more than the intercept, and the nominal rate moves the opposite way to the intercept (measured in the comparison).
  • The reported r is ex-post. It is the real return on bonds held into the quarter: on impact it moves only with the inflation surprise, and the policy move shows from quarter 1. In the ρ 0.61 example r falls 0.08pp on impact and 0.18pp in quarter 1.
Spending multiplier turns negative
  • Path. A spending rise of 1% of GDP (ρ 0.8) raises output by 0.175% on impact, but output is below baseline from quarter 3 (impact is quarter 0) while spending is still raised; the 20-quarter cumulative multiplier is −0.015.
  • Mechanism. The labour tax that finances the spending enters the union wage Phillips curve: a higher tax raises the wage unions set, which raises inflation, and the Taylor rule answers with a higher real rate that crowds out investment and consumption.
  • A property of the model, not a bug. The household and government budget constraints hold along the path (tests/test_spending_mechanism.py; us-hank-model #5). No published impulse response gates these numbers.
  • Comparability. It is a tax-financed multiplier, not comparable with deficit-financed ones.
  • Sensitivity. The sign depends on the MPC: in the opt-in 0.20-MPC calibration the multiplier is 0.376 on impact and 0.309 over 20 quarters (below).
Persistence of a spending shock
  • No default. The hosted hank_shock has no default ρ for a spending shock: persistence must be given (monetary and productivity shocks default to 0.9). The model’s README quotes ρ 0.8.
  • Why. For the same spending shock the impact multiplier is 0.075 at ρ 0.9 and 0.175 at ρ 0.8, and the 20-quarter cumulative multiplier is −0.032 against −0.015. Quote ρ with any multiplier.
Opt-in calibrations
  • Status. Two opt-in calibrations ship in the model package (us_hank.beta_het) but are not the default and are not hosted. No result on this site comes from them.
  • Discount-factor heterogeneity. Calibrated to a 0.20 MPC out of $1,000: spending impact multiplier 0.376, 20-quarter cumulative 0.309 (ρ 0.8).
  • Median-wealth sensitivity. A $500 MPC of 0.126 and multipliers of 0.274 / 0.213 (PR #9, PR #10, #8).
Approximate distributional output

By-quartile consumption splits the aggregate response across total-wealth quartiles in proportion to steady-state MPCs — the MPC-heterogeneity channel, not household-level dynamics along the transition. The split is nearly flat: for the intercept easing shock, 0.36% in the lowest quartile and 0.28% in the highest.

Calibration, not vintage

The parameterisation is the paper's; there is no data vintage to refresh. Changing it would be a model change.