Pith. sign in

REVIEW 3 major objections 4 minor 22 references

The Endogenous Grid Method for Epstein-Zin Preferences

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A power transformation, W=V^(1-ρ), converts the Epstein-Zin Bellman equation into additive form, permitting a closed-form Euler inversion and a root-finding-free endogenous grid method that is 10-50x faster than value function iteration.

desk verdict A genuine extension of EGM to Epstein-Zin preferences with a clean closed-form Euler inversion and credible benchmark speedups; the main repair needed is a convergence proof for the hybrid iteration and a check of the rho>1 robustness claims. read the letter →

arxiv 2601.04438 v2 pith:FA7CBVHH submitted 2026-01-07 econ.GN q-fin.EC

Alan Lujan This is my paper · ORCID
classification econ.GNq-fin.EC MSC 91B6290C39
keywords endogenousgridmethodEpstein-Zinpreferencesrecursiveutilityconsumption-savingsEulerequationinversionpowertransformationvaluefunctioniterationtime
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the apparent incompatibility between the endogenous grid method and Epstein-Zin recursive utility is resolved by a power transformation of the value function. Raising V to the power 1-ρ turns the CES Bellman recursion into an additive one, and the resulting Euler equation can be inverted analytically for consumption as a function of end-of-period assets. This yields an algorithm with no root-finding that tracks consumption and transformed value simultaneously. Benchmarks in a standard consumption-savings model show speedups of one to two orders of magnitude over value function iteration and time iteration at equal grid size, with Euler errors near 1e-5 and up to three orders of magnitude speedup holding accuracy constant. If correct, this makes recursive-utility models dramatically cheaper to estimate and solve.

What carries the argument

The load-bearing object is the power transformation W=V^(1-ρ) and the additive Bellman equation it induces. The certainty equivalent μ(a,z)=(E[W(m',z')^θ])^(1/θ) with θ=(1-γ)/(1-ρ) and the expectation Ξ(a,z)=E[W^(θ-1) c^(-ρ)] together make the Euler equation c^(-ρ)=βR μ^(1-θ) Ξ invertible in closed form (Proposition 2.1). This inversion is what lets EGM compute consumption directly from the end-of-period asset grid, sidestepping root-finding. The algorithm is a hybrid of EGM and Howard policy iteration, updating c and V/W together until both stabilize; the borrowing constraint is anchored by appending (0,0), and non-monotonic grids are handled with the upper envelope.

What would settle it

Run EZ-EGM on a calibration with very patient agents (βR ≥ 1) or with a binding borrowing constraint and high curvature (e.g., ρ=3, γ=10) and check whether the iteration converges to the same fixed point as VFI/TI; specifically, check whether the endogenous grid remains monotone after the upper-envelope step and whether the mean Euler error stays near 1e-5. A failure to converge, or a grid that the upper-envelope cannot repair, would falsify the claim that EGM sidesteps the speed-accuracy tradeoff for Epstein-Zin preferences.

Watch

Extended reading notes

Core claim

The central claim is that the transformation W(m,z)=V(m,z)^(1-ρ) (with θ=(1-γ)/(1-ρ)) converts the Epstein-Zin Bellman equation into the additive form W(m,z)=max_c[(1-β)c^(1-ρ)+β μ(m-c,z)], where μ is the power-mean certainty equivalent of next-period transformed value. Differentiation then yields an Euler equation c^(-ρ)=βR μ^(1-θ) Ξ that is invertible in closed form: c(a,z)=(βR μ^(1-θ) Ξ)^(-1/ρ). Given an exogenous grid over end-of-period assets, this makes consumption computable directly without numerical root-finding, and the endogenous grid is recovered as m=c+a. The algorithm iterates on both the policy and the value function until convergence, and the paper reports large speed and acc

Load-bearing premise

Algorithm 1's convergence to the infinite-horizon fixed point is assumed to hold under 'standard conditions' without a proof that the EGM-plus-Bellman iteration is a contraction; if the iteration fails to converge or the upper-envelope repair cannot fix non-monotonic endogenous grids in some economically relevant region, the reported speed and accuracy gains would not generalize beyond the calibrated case.

Editorial extensions

If this is right

  • An analyst solving an infinite-horizon consumption-savings model with Epstein-Zin preferences can replace root-finding-based time iteration or VFI with EZ-EGM and obtain mean Euler errors around 1e-5 at roughly 1/10 to 1/50 the runtime at equal grid size.
  • Holding solution accuracy constant, EZ-EGM is 150 to 630 times faster than VFI in the reported benchmarks, turning previously intractable estimation or uncertainty-quantification tasks into routine computations.
  • The closed-form Euler inversion eliminates the speed-accuracy tradeoff that characterizes VFI and time iteration: the endogenous grid places evaluation points exactly where the Euler equation holds, so accurate-mode precision is achieved at fast-mode cost.
  • The method extends to the logarithmic limit ρ=1 (unit EIS) and to the risk-sensitive preferences of Hansen-Sargent, and is verified numerically for ρ>1 (EIS<1), covering the empirically relevant low-EIS range.
  • Non-monotonic endogenous grids, when they occur, can be repaired with the upper-envelope procedure, so the approach does not require a monotonicity assumption for the baseline implementation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same inversion strategy may apply to other recursive utility forms whose Bellman equations become additive under a suitable transform — for instance, models with multiple assets or portfolio choice — since the essential requirement is only that the Euler equation be invertible in closed form after the transformation.
  • Because the endogenous grid evaluates the Euler equation exactly at the grid points, combining EZ-EGM with higher-order or adaptive interpolation could push accuracy beyond the reported 1e-5 at negligible additional cost; the paper does not explore this.
  • The reported speedups are hardware- and implementation-dependent (JAX on CPU), but the structural advantage of eliminating root-finding is implementation-independent; a careful comparison on GPU or with alternative interpolation methods could change the constants but not the qualitative ordering.
  • The convergence of the hybrid EGM-plus-Bellman iteration is asserted rather than proven; a formal contraction argument or a counterexample for a pathological calibration would determine whether the approach applies beyond the parameterizations tested.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes a power transformation W = V^{1-rho} that converts the Epstein-Zin Bellman equation into an additive form, derives a closed-form inverted Euler equation c(a,z) = (beta R mu^{1-theta} Xi)^{-1/rho} (Proposition 2.1), and presents Algorithm 1 (EZ-EGM), a root-finding-free endogenous grid method. The authors benchmark EZ-EGM against value function iteration and time iteration, reporting speedups of 10-50x at equal grid size, one order-of-magnitude accuracy improvements, and 150-630x speedups at equal accuracy. The paper also claims the method extends to rho > 1 (EIS < 1) with numerical verification in Appendix F.

Significance. The central insight is attractive and, if correct, practically important: it would make the endogenous grid method available for the widely used Epstein-Zin preferences, substantially accelerating structural estimation and model evaluation. The derivation of the inverted Euler equation is clean and the baseline calibration is standard. The paper also includes thoughtful benchmarking, an equal-accuracy comparison, and an independent welfare-cost check (below 0.1%). These are genuine strengths. However, the absence of a convergence proof for the hybrid EGM-plus-Bellman iteration and unresolved issues in the rho > 1 robustness analysis prevent the claims from being fully established as a general property of the method.

major comments (3)
  1. [§3, Algorithm 1, lines 11-18] Convergence of Algorithm 1 to the infinite-horizon fixed point is asserted, not proved. The iteration is a hybrid: the policy is updated by the analytic Euler inversion (lines 7-11), then the value function is updated by exactly one Bellman evaluation (lines 12-16), and the loop stops when ||c^n - c^{n-1}|| < epsilon. No contraction mapping, monotonicity, or policy-iteration theorem is established for this operator. The stopping criterion does not by itself imply that the limit is a fixed point of the Bellman equation, and non-monotone endogenous grids are only addressed by a heuristic upper-envelope fix. Since all Section 4 speed/accuracy claims rest on this iteration, this is load-bearing. Please either supply a convergence theorem (for example by adapting policy-iteration theory for recursive preferences) or explicitly present the algorithm as heuristic and temper the benchmark claims
  2. [Appendix F vs Footnote 2] The robustness analysis for rho > 1 conflicts with the paper's own existence condition. Footnote 2 states that a sufficient condition for a well-defined infinite-horizon problem is beta R^theta < 1. For rho > 1 and gamma = 10, theta = (1-gamma)/(1-rho) is positive (e.g., theta = 90 for rho = 1.1), and beta R^theta is approximately 5.7, violating the condition. Yet Appendix F reports convergence with Euler errors near -5 for rho in {1.1, 1.5, 2, 3}. This suggests the finite-grid runs may be solving a truncated or ill-posed infinite-horizon problem. The paper must clarify whether a genuine infinite-horizon solution exists for these parameter values, or whether the reported numbers are for a modified (e.g., bounded-state or finite-horizon) problem.
  3. [§2.1, Eq. (3) and Footnote 3] For rho > 1, W = V^{1-rho} is decreasing in V, so the transformed Bellman equation should be a minimization, not a maximization. The paper acknowledges this only in a footnote and continues to write 'max' in Eq. (3). The first-order condition is necessary, but the paper provides no convexity/concavity argument to show it is sufficient for the optimum in the minimization problem. Since Proposition 2.1 is the analytical foundation of the method, the statement of the transformed Bellman equation for rho > 1 should be corrected and the sufficiency conditions stated, or the formal claims should be restricted to rho < 1 with rho > 1 presented only as a numerical extension.
minor comments (4)
  1. [Algorithm 1, line 10] Appending the point (0,0) at the constraints is not sufficient to represent the borrowing-constrained policy. For m below the first endogenous grid point, the constrained policy is c(m,z) = m, not a linear interpolation through (0,0). Please specify that the constrained segment is imposed separately, as stated later in the text.
  2. [Table 1 and Table 2] The header 'Euler Error' in Table 1 is ambiguous; it should be labeled 'Mean Euler Error (log10)' to match Table 2.
  3. [Footnote 2] The statement that V > 0 follows by induction from a terminal condition does not directly apply to the infinite-horizon problem; the limiting argument should be stated explicitly.
  4. [Appendix F] The appendix reports results for rho in {0.5, 0.9, 1.1, 1.5, 2, 3} but does not provide a table of errors or iteration counts. Including the detailed numbers would help readers assess the robustness claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the EZ-EGM inversion is an algebraic consequence of the transformed Bellman equation, with benchmark claims grounded in external calibration and independent welfare checks.

full rationale

The central derivation is self-contained algebra: Definition 2.2 defines W=V^(1-rho); substituting into the Epstein-Zin Bellman equation gives the additive transformed Bellman equation (3); the first-order condition (4), envelope condition (5), and derivative of the certainty equivalent (6) combine into the Euler equation (7)-(8), which Proposition 2.1 inverts in closed form as c(a,z) = (beta R mu^(1-theta) Xi)^(-1/rho). No parameter is fitted to the quantities the paper claims to predict; the calibration (beta=0.96, R=1.02, gamma=10, EIS=1.5) comes from Bansal and Yaron (2004) and related literature. The speed and accuracy results are benchmark measurements, not retrodictions of fitted inputs. The only mildly self-referential element is that EGM satisfies the Euler equation by construction at its own endogenous grid points (Appendix E), but the reported Euler errors are evaluated on a simulated ergodic distribution rather than on those grid points, and the consumption-equivalent welfare cost check (<0.1%) provides an independent accuracy grounding. No load-bearing self-citation appears: all cited methods and theorems (Carroll 2006; Epstein and Zin 1989; Rendahl 2015; Santos 2000) are external prior work. The paper does have correctness risks that are not circularity: Algorithm 1's convergence is asserted rather than proved (Section 3, lines 11 and 18), and Appendix F's rho>1 robustness cases violate the sufficient existence condition beta R^theta < 1 quoted in footnote 2 (e.g., rho=1.1, gamma=10 gives theta=90 and beta R^theta approximately 5.7). These are open mathematical or numerical issues, not input-output circularity.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central derivation relies on standard model assumptions and the benchmark calibration; no physical entities are invented and no parameters are fitted to the target speed/accuracy result. The main unproven elements are convergence of the hybrid iteration and monotonicity of the endogenous grid in the Epstein-Zin case.

free parameters (2)
  • Baseline preference/technology calibration (beta, R, gamma, EIS) = beta=0.96, R=1.02, gamma=10, EIS=1.5 (rho=2/3)
    Taken from Bansal-Yaron (2004) and standard calibrations; used in benchmarks but not fitted to the accuracy/speed results.
  • Grid size (asset and cash-on-hand) = 100 points, exponentially spaced, upper bound 20x mean income
    Hand-chosen for the headline benchmark; equal across methods in Table 1-2, so it does not favor EGM. The equal-accuracy comparison deliberately varies grid size.
assumptions (6)
  • domain assumption The Epstein-Zin Bellman recursion has a unique well-defined fixed point under beta R^theta < 1 and V > 0 for all states.
    Invoked in Section 2 / Footnote 2 to ensure the value function exists and the power transformation W=V^(1-rho) is well-defined.
  • domain assumption The value and policy functions are differentiable and the optimization is concave, so the FOC and envelope theorem are sufficient for the optimum.
    Used to derive the Euler equation in Section 2.1; not proven for the transformed problem.
  • domain assumption Tauchen discretization of AR(1) income with 10 states faithfully approximates the continuous process.
    Used in Section 4 benchmarks; standard but affects the reported accuracy and speed.
  • standard math Euler equation residuals bound policy error (Santos 2000).
    Used in Section 4.2 to justify the accuracy metric.
  • domain assumption For rho > 1, the same Euler equation characterizes the optimum of the minimization problem; no modification is needed.
    Stated in Appendix F and Remark 3.1; asserted without a formal proof.
  • domain assumption The endogenous grid is monotone for the Epstein-Zin problem, or the upper-envelope fix restores monotonicity.
    Assumed in Section 3 for 'typical calibrations'; non-monotonicity is handled by an external envelope procedure but no general proof is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Endogenous Grid Method for Epstein-Zin Preferences." pith.science (2026). https://pith.science/paper/FA7CBVHH

@misc{pith2026260104438,
  author       = {Pith},
  title        = {Pith review of: The Endogenous Grid Method for Epstein-Zin Preferences},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FA7CBVHH}},
  note         = {Machine review of arXiv:2601.04438}
}
read the original abstract

The endogenous grid method (EGM) accelerates dynamic programming by inverting the Euler equation, but it appears incompatible with Epstein-Zin preferences where the value function enters the Euler equation. This paper shows that a power transformation resolves the difficulty. The resulting algorithm requires no root-finding, achieves speed gains of one to two orders of magnitude over value function iteration, and improves accuracy by more than one order of magnitude. Holding accuracy constant, the speedup is two to three orders of magnitude. VFI and time iteration face a speed-accuracy tradeoff; EGM sidesteps it entirely.

Figures

Figures reproduced from arXiv: 2601.04438 by the authors.

Figure 1
Figure 1. Consumption policy c(m, z) for different income states. Higher income states (red) allow more consumption at each wealth level. 4.2 Accuracy We assess accuracy using the normalized Euler equation error, a standard metric whose magnitude bounds the policy function error. Santos [2000] establishes the theoretical foundation: ε(m, z) = log10 [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Speed-accuracy Pareto frontier across grid sizes. Marker size reflects grid resolution [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Speed-accuracy tradeoff. VFI and TI face a tradeoff between fast modes (open markers) and [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 17 canonical work pages

  1. [6]

    Philippe Weil

    doi:10.1086/261750. Philippe Weil. The equity premium puzzle and the risk-free rate puzzle.Journal of Monetary Economics, 24 (3):401–421, 11

    work page doi:10.1086/261750

  2. [7]

    David M Kreps and Evan L Porteus

    doi:10.1016/0304-3932(89)90028-7. David M Kreps and Evan L Porteus. Temporal Resolution of Uncertainty and Dynamic Choice Theory. Econometrica, 46(1):185–200, 1

    work page doi:10.1016/0304-3932(89)90028-7

  3. [11]

    Lars Peter Hansen and Thomas J Sargent

    doi:10.1016/j.jmoneco.2019.01.022. Lars Peter Hansen and Thomas J Sargent. Discounted linear exponential quadratic Gaussian control.IEEE Transactions on Automatic Control, 40(5):968–971, 5

    work page doi:10.1016/j.jmoneco.2019.01.022 2019

  4. [17]

    George Tauchen

    doi:10.1086/383105. George Tauchen. Finite state Markov-chain approximations to univariate and vector autoregressions.Eco- nomics Letters, 20(2):177–181,

    work page doi:10.1086/383105

  5. [20]

    Wilbur John Coleman

    doi:10.1111/ecoj.12115. Wilbur John Coleman. Equilibrium in a Production Economy with an Income Tax.Econometrica, 59(4): 1091–1104, 7

    work page doi:10.1111/ecoj.12115

  6. [21]

    Manuel S Santos

    doi:10.2307/2938175. Manuel S Santos. Accuracy of Numerical Solutions Using the Euler Equation Residuals.Econometrica, 68(6): 1377–1402, 11

    work page doi:10.2307/2938175

  7. [22]

    Lilia Maliar and Serguei Maliar.Numerical Methods for Large-Scale Dynamic Economic Models, volume 3, pages 325–477

    doi:10.1111/1468-0262.00165. Lilia Maliar and Serguei Maliar.Numerical Methods for Large-Scale Dynamic Economic Models, volume 3, pages 325–477. Elsevier,

    arXiv

  8. [1978]

    Darrell Duffie and Larry G Epstein

    doi:10.2307/1913656. Darrell Duffie and Larry G Epstein. Stochastic Differential Utility.Econometrica, 60(2):353–394, 3

    work page doi:10.2307/1913656

Show all 22 references
  1. [1986]

    Tomáš Havránek

    doi:10.1016/0165-1765(86)90168-0. Tomáš Havránek. Measuring Intertemporal Substitution: The Importance of Method Choices and Selective Re- porting.Journal of the European Economic Association, 13(6):1180–1204, 12

    doi

  2. [1989]

    14 Computational Economics Larry G Epstein and Stanley E Zin

    doi:10.2307/1913778. 14 Computational Economics Larry G Epstein and Stanley E Zin. Substitution, risk aversion, and the temporal behavior of consump- tion and asset returns: An empirical analysis.Journal of Political Economy, 99(2):263–286, 4

    doi

  3. [1991]

    Francisco Barillas and Jesús Fernández-Villaverde

    doi:10.2307/2938366. Francisco Barillas and Jesús Fernández-Villaverde. A Generalization of the Endogenous Grid Method.Journal of Economic Dynamics and Control, 31(8):2698–2712, 8

    doi

  4. [1995]

    Ravi Bansal and Amir Yaron

    doi:10.1109/9.384242. Ravi Bansal and Amir Yaron. Risks for the Long Run: A Potential Resolution of Asset Pricing Puzzles.The Journal of Finance, 59(4):1481–1509, 8

    doi

  5. [2000]

    Alberto Giovannini and Philippe Weil

    doi:10.1016/S0304-3932(00)00012-X. Alberto Giovannini and Philippe Weil. Risk Aversion and Intertemporal Substitution in the Capital Asset Pricing Model. Working Paper 2824, National Bureau of Economic Research, 1

    doi

  6. [2004]

    Alexander Meyer-Gohde

    doi:10.1111/j.1540-6261.2004.00670.x. Alexander Meyer-Gohde. Generalized entropy and model uncertainty.Journal of Economic Theory, 183: 312–343, 9

    2004

  7. [2006]

    Angus Deaton

    doi:10.1016/j.econlet.2005.09.013. Angus Deaton. Saving and liquidity constraints.Econometrica, 59(5):1221–1248, 9

    2005 doi

  8. [2007]

    Fedor Iskhakov, Thomas H Jørgensen, John Rust, and Bertel Schjerning

    doi:10.1016/j.jedc.2006.08.005. Fedor Iskhakov, Thomas H Jørgensen, John Rust, and Bertel Schjerning. The Endogenous Grid Method for Discrete-Continuous Dynamic Choice Models with (or without) Taste Shocks.Quantitative Economics, 8 (2):317–365, 7

    2006 doi

  9. [2008]

    Thomas D Tallarini Jr

    doi:10.1057/978-1-349-95121-5_1963-1. Thomas D Tallarini Jr. Risk-sensitive real business cycles.Journal of Monetary Economics, 45(3):507–532, 6

    doi

  10. [2014]

    doi:10.1016/B978-0-444-52980-0.00007-4. 15

    doi

  11. [2015]

    Pontus Rendahl

    doi:10.1111/jeea.12133. Pontus Rendahl. Inequality Constraints and Euler Equation-based Solution Methods.The Economic Journal, 125(585):1110–1135, 6

    doi

  12. [2017]

    Larry G Epstein and Stanley E Zin

    doi:10.3982/QE643. Larry G Epstein and Stanley E Zin. Substitution, risk aversion, and the temporal behavior of consumption and asset returns: A theoretical framework.Econometrica, 57(4):937–969, 7

    doi

  13. [2019]

    David K Backus, Bryan R Routledge, and Stanley E Zin.Recursive preferences

    doi:10.1016/j.jet.2019.06.004. David K Backus, Bryan R Routledge, and Stanley E Zin.Recursive preferences. Palgrave Macmillan, London, 2 edition,

    2019 doi

  14. [2020]

    Nicolas Coeurdacier, Hélène Rey, and Pablo Winant

    URLhttps://arxiv.org/abs/1812.05748. Nicolas Coeurdacier, Hélène Rey, and Pablo Winant. Financial Integration and Growth in a Risky World. Journal of Monetary Economics, 112:1–21,

    arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.