Quantile Intertemporal Consumption Model

• Quantile intertemporal consumption model is a framework that replaces expectation-based utility with quantile-based preferences to decouple risk aversion from elasticity of intertemporal substitution.
• It employs quantile Euler equations and smoothed GMM estimation to robustly estimate key parameters such as the discount factor, EIS, and risk attitude even with heavy-tailed outcomes.
• Empirical applications on micro and macro data demonstrate its effectiveness in handling endogeneity and asymmetry, offering improved reliability over traditional estimation methods.

The quantile intertemporal consumption model generalizes the standard dynamic consumption model by replacing expectation-based preferences with quantile-based preferences over future marginal utility, thereby decoupling risk attitudes from elasticity of intertemporal substitution (EIS) and enabling robust estimation even under heavy-tailed outcomes. Using quantile utilities, the approach yields quantile Euler equations, which—when coupled with smoothed generalized method of moments (GMM) estimation—facilitate simultaneous inference of key preference parameters, including the quantile index (capturing downside risk attitude), the discount factor, and EIS, under minimal assumptions on functional form or distributional regularity (Liu et al., 28 Jan 2026, Castro et al., 2017).

1. Quantile Utility and the Quantile Euler Equation

Let $C_t$ denote consumption at time $t$, and $R_{j,t+1}$ the gross return on asset $j$ held between $t$ and $t+1$. At each $t$, an agent allocates consumption and asset portfolios to solve the intertemporal problem under isoelastic utility

$$U(C) = \frac{C^{1-\gamma}}{1-\gamma}, \quad \gamma>0,\, \gamma\neq1.$$

In contrast to models based on expected utility, the quantile model’s objective function is a conditional quantile operator at a prescribed level $\tau\in(0,1)$. The resulting optimality condition for any asset $j$ with positive portfolio share is the quantile Euler equation:

$$Q_\tau\!\left[\, \delta\, U'(C_{t+1})\, R_{j,t+1}\mid \mathcal I_t \right] = 0,$$

where $U'(C_{t+1}) = C_{t+1}^{-\gamma}$ and $\mathcal I_t$ is the agent’s information set at $t$.

Owing to quantile equivariance under monotone transformations, the equation admits a log-linear form. Denoting $\psi = 1/\gamma$ (EIS), one obtains:

$$Q_\tau\!\left[ \ln \delta - \gamma \ln \frac{C_{t+1}}{C_t} + \ln R_{j,t+1} \mid \mathcal I_t \right] = 0.$$

By recentering, this is equivalently rewritten with a zero conditional quantile:

$$Q_{1-\tau}\left[ \ln \frac{C_{t+1}}{C_t} - \frac{1}{\psi}\ln \delta - \frac{1}{\psi}\ln R_{j,t+1} \mid \mathcal I_t \right] = 0.$$

The parameter $\tau$ directly indexes downside risk attitude; $\psi$ determines willingness to substitute intertemporally (Liu et al., 28 Jan 2026, Castro et al., 2017).

2. Unconditional Quantile Restrictions and Structural Moments

The conditional quantile Euler restrictions can be re-expressed as a system of unconditional moment conditions. For each asset $j$, define the structural function:

$$\Lambda_{ji}(\beta) = \ln \frac{C_{t+1,i}}{C_{t,i}} - \frac{1}{\psi}\ln\delta - \frac{1}{\psi}\ln R_{j,t+1,i},$$

with $\beta=(\delta, \psi)$. For suitable instruments $Z_{ji}$,

$Q_{1-\tau}[\Lambda_{ji}(\beta_0) \mid Z_{ji}] = 0.$

This is equivalent to the unconditional moment condition:

$$E\left[Z_{ji}\left\{1\{\Lambda_{ji}(\beta_0)\le0\} - (1-\tau_0)\right\}\right] = 0,\quad j=1,2.$$

The resulting vector-valued moments form the basis for IV-quantile GMM estimation (Liu et al., 28 Jan 2026, Castro et al., 2017).

3. Smoothed Generalized Method of Moments Estimation

Direct GMM estimation is complicated by the discontinuity of the indicator function $1\{u\le0\}$. To ensure numerical tractability and differentiability, a smooth approximation $\tilde{I}(u)$ is used, often specified as a kernel-smoothed indicator of order $r\ge2$ (as in Horowitz, 1998):

$$\tilde{I}(u) = \begin{cases} 0, & u\le-1, \ \int_{-1}^u K(v)dv, & -1 < u < 1, \ 1, & u\ge1, \end{cases}$$

with symmetric kernel $K$. The smoothed empirical moment is:

$$g_{ni}(\beta, \tau) = \begin{pmatrix} Z_{1i}\bigl[\tilde I(-\Lambda_{1i}(\beta)/h_n)-(1-\tau)\bigr] \ Z_{2i}\bigl[\tilde I(-\Lambda_{2i}(\beta)/h_n)-(1-\tau)\bigr] \end{pmatrix},$$

where $h_n\to 0$ is the bandwidth. Averaging over $n$ observations yields the sample moment vector $$\hat M_n(\beta,\tau) = n^{-1} \sum_i g_{ni}(\beta, \tau)$$.

The two-step smoothed-GMM estimator solves

$$(\hat\beta, \hat\tau) = \arg\min_{(\beta, \tau)\in\mathcal B\times(0,1)} \hat M_n(\beta, \tau)' \hat W\, \hat M_n(\beta, \tau),$$

where $\hat W$ is first set to identity (or a long-run variance estimator), then updated at the preliminary solution for efficient weighting. Global optimization (e.g., via simulated annealing) is recommended due to nonconvexity (Liu et al., 28 Jan 2026, Castro et al., 2017).

4. Identification, Regularity, and Asymptotic Properties

Identification is secured through a full-rank Jacobian condition in the moment derivatives:

$$G = \partial_{(\beta,\tau)} E[g_{i}(\beta, \tau)]\Big|_{(\beta_0,\tau_0)},$$

which, in the two-asset case, has the structure

$$G = \begin{pmatrix} -E[Z_{1} D_{1} f_{\Lambda_1|Z_1,D_1}(0\mid Z_{1},D_{1})] & -E[Z_{1}]\ -E[Z_{2} D_{2} f_{\Lambda_2|Z_2,D_2}(0\mid Z_{2},D_{2})] & -E[Z_{2}] \end{pmatrix}.$$

Full rank requires at least two distinct assets.

Under standard regularity—compact parameter set, smooth structural function, stationary weakly dependent data, appropriate smooth kernel, and bandwidth $h_n=o(n^{-1/(2r)})$—the estimator is consistent and asymptotically normal:

$\sqrt n(\hat\theta - \theta_0) \;\dto\; N(0, (G W G')^{-1} G W \Sigma W G' (G W G')^{-1}),$

with $\Sigma$ the asymptotic variance of the sample moments and efficient $W = \Sigma^{-1}$ (Liu et al., 28 Jan 2026, Castro et al., 2017).

5. Simulation Evidence

Monte Carlo experiments using stylized two-asset designs with random coefficients confirm the finite-sample performance of the estimator. Specifically, when the conditional quantile functions of both assets align at $\tau_0 = 0.70$, and sample size $n$ increases from 1,500 to 5,000:

• The bias in $\hat\tau$ remains below $0.02$ at $n=1,500$ and declines with $n$.
• The root-MSE of $\hat\tau$ decays at approximately $1/\sqrt n$.
• Estimates of $\beta$ and $\gamma$ (with EIS $= 1/\gamma$) have small finite-sample bias and standard errors scaling with $\sqrt n$.
• Results are robust to bandwidth choices (Liu et al., 28 Jan 2026).

6. Empirical Applications and Comparative Outcomes

The quantile intertemporal consumption model has been applied to both micro (household) and macro (country-level aggregate) data.

• In a two-asset application to PSID household data (N=525, 2005–2009), the two-step smoothed-GMM estimates are:
  • Quantile (risk-attitude) $\hat\tau=0.402$ (SE $0.031$; 95% CI $[0.342, 0.463]$)—implying modest downside risk aversion ($\tau < 0.5$).
  • Discount factor $\hat{\delta}=1.197$ (SE $0.075$).
  • EIS $\hat{\psi}=0.829$ (SE $0.248$; 95% CI $[0.342,1.316]$), consistent with established ranges (Liu et al., 28 Jan 2026).
• On country-level macro data (e.g., USA, UK, Australia, Sweden), quantile-GMM estimates at $\tau=0.5, 0.75$ for the discount factor $\beta$ and EIS $\psi$ are economically plausible across countries and exceed the reliability of 2SLS estimates, which can yield implausible results, especially in the presence of heavy tails or weak instruments (Castro et al., 2017).

Country	$\hat{\beta}$ ($\tau=0.5$)	$\hat{\psi}$ ($\tau=0.5$)
USA	1.02	4.1
UK	1.00	3.8
Australia	1.09	4.5
Sweden	0.97	4.1

Low-quantile ($\tau<0.4$) estimates may be nonsensical for $\beta$ or $\psi$ in some contexts, reflecting poor fit to bad consumption shocks. For central and upper quantiles, parameters fall within established macroeconomic ranges (Castro et al., 2017).

7. Salient Features and Practical Implications

The quantile Euler approach offers:

• Robustness to Fat Tails: Quantile-based conditions remain well-behaved for heavy-tailed consumption or return distributions, protecting against the breakdown of moment-based (expected utility) estimators under Cauchy-like shocks.
• Decoupling Risk and Substitution: Unlike expected utility models where risk aversion and EIS are entangled via $\gamma$, quantile preferences partially separate these aspects—$\tau$ governs risk attitude, while $\psi$ reflects substitution, allowing joint estimation and interpretation.
• Exact Log-linearization: Log-linearizing the quantile Euler equation introduces no approximation error due to quantile equivariance, in contrast to the Jensen’s-inequality bias of expectation-based log-linearizations.
• Estimation Under Weak Assumptions: The smoothed GMM framework delivers consistency and root-$n$ asymptotic normality with weak dependence, endogeneity, and minimal distributional conditions, supporting its application in diverse empirical settings (Liu et al., 28 Jan 2026, Castro et al., 2017).

These properties render the quantile intertemporal consumption model a robust alternative to classical expected-utility specifications and particularly suitable for settings with endogeneity or persistent, asymmetric shocks.