betaFeedback

WorksMathematics

Estimation and Inference for the Joint Autoregressive Quantile-Expected Shortfall Model

Abstract

Expected shortfall is defined as the truncated mean of a random variable that falls below a specified quantile level and is widely recognized as an important risk measure. Motivated by the empirical observation of clustering patterns in financial risks, we consider a joint autoregressive model for both conditional quantile and expected shortfall. Existing estimation methods for such models typically rely on minimizing a nonlinear and nonconvex joint loss function, which is challenging to solve. We introduce a two-step regression framework to estimate the model parameters with theoretical guarantees. We further develop a Portmanteau test for model diagnostics and prove its asymptotic validity for Type I error control. Our empirical results on stock market data indicate that the proposed models effectively capture the clustering patterns and leverage effects on the conditional expected shortfall.

Introduction

Financial risk management has been a prominent research area for decades. In recent years, the expected shortfall has gained considerable attention as a risk measure in finance [32]. Notably, the fundamental review of the trading book, an integral part of Basel III reforms, announced its adoption of the expected shortfall to replace the Value-at-Risk as the standard risk measure for regulatory supervision in the year 2016 [4]. Let YtY_t be the return of a financial asset over some time horizon and let Ht\mathcal{H}_t be an information set available by time tt. For a given quantile level τ∈(0,1)\tau\in(0,1), let Qτ(Yt∣Ht)=inf⁡{r∈R;P(Yt≤r∣Ht)≥τ}Q_\tau(Y_t \mid\mathcal{H}_t) = \inf\{r \in\mathbb{R}; \mathbb{P}(Y_t \le r \mid\mathcal{H}_t) \ge\tau\} be the τ\tau-level conditional quantile of YtY_t given the information set Ht\mathcal{H}_t, which is also referred to as the Value-at-Risk. Then, the τ\tau-level conditional expected shortfall (ES) of YtY_t can be defined as

ES⁡τ(Yt∣Ht)=E{Yt∣Yt≤Qτ(Yt∣Ht),Ht}=1τ∫0τQu(Yt∣Ht) du.(1)\operatorname{ES}_\tau(Y_t \mid\mathcal{H}_t) = \mathbb{E}\{Y_t \mid Y_t \le Q_\tau(Y_t \mid\mathcal{H}_t), \mathcal{H}_t\} = \frac{1}{\tau}\int_0^\tau Q_u(Y_t \mid\mathcal{H}_t)\,\mathrm{d}u. \tag*{(1)}

Equation (1) can be interpreted as the expected return of an asset, given that its return is less than or equal to the τ\tau-level conditional quantile. Unlike the Value-at-Risk that focuses on a specific quantile level, the conditional expected shortfall aggregates the tail information of YtY_t beyond the conditional quantile, providing a more comprehensive representation of the tail risk.

Although the importance of the conditional expected shortfall is well recognized in different fields such as finance [22], operations research [30], and insurance [9], its estimation and statistical inference remain challenging tasks. This is primarily due to the non-elicitability of the expected shortfall, that is, there does not exist a loss function for which the minimizer of the expected loss is the expected shortfall [20]. Although the expected shortfall is not elicitable by itself, it is jointly elicitable with the quantile from a class of strictly joint loss functions [17]. Motivated by the joint loss function in Fissler & Ziegel [17], Dimitriadis & Bayer [13] and Patton et al. [29] proposed joint semi-parametric regression frameworks to estimate the τ\tau-level conditional quantile and expected shortfall simultaneously. On the other hand, Barendse [2] introduced a two-step estimation approach by first estimating the conditional quantile via quantile regression, and then estimating the expected shortfall using a least squares regression with a constructed pseudo-response based on the estimated quantile function obtained from the first step. Motivated by Barendse [2], He et al. (2023) proposed estimation and inference procedures for linear expected shortfall models under the heavy-tailed data and the growing dimension regime for the covariates. Furthermore, Barendse [1] and Zhang et al. [40] proposed an ℓ1\ell_1-penalized expected shortfall regression for the high-dimensional regime in which the dimension of the covariates can be larger than the number of samples. Li et al. (2026) proposed an expected shortfall regression approach based on the optimization perspective considered in Rockafellar et al. [30].

In this manuscript, we aim to address the following issues in the context of modeling conditional expected shortfall for financial time series: (i) develop a model that effectively captures the risk patterns of financial markets; (ii) provide an easy-to-implement estimation procedure with theoretical guarantees; and (iii) construct a statistical test to assess the quality of the resulting estimator. We start with modeling the conditional expected shortfall with autoregressive structures, i.e., the current expected shortfall can be expressed as a parametric function of the past expected shortfalls and a set of covariates. We are motivated by the observed clustering pattern of market risks, where a bad trading day can increase the risk level of future trading days, reflecting an important empirical property of the global financial market [8].

The existing literature on other risk measures such as standard deviation and Value-at-Risk with autoregressive structures is rich. The standard deviation models include the Generalized AutoRegressive Conditional Heteroskedasticity (GARCH) and its variants [7, 26]. The model proposed in Engle & Manganelli [16] assumed certain autoregressive structures on the conditional quantile, which has been extensively applied in financial risk management and prediction [12], and various extensions have been developed [36].

However, limited work has been done on estimating and conducting model diagnostics for conditional autoregressive expected shortfall in the presence of dependent data, despite its importance as a risk measure that aggregates tail risk. Most existing work does not model the autoregressive structures of the conditional expected shortfall directly. For instance, Taylor [33] proposed a joint estimation approach in which they modeled the conditional quantile via Engle & Manganelli [16], and then assumed certain parametric structures on the ratio (multiplicative models) or the difference (additive models) between the conditional quantile and the conditional expected shortfall. A sequence of works, including Gerlach & Wang [19], Wang et al. [34], and Patton et al. [29], extends Taylor [33] from the perspectives of Bayesian approaches and the generalized autoregressive score framework [11]. The aforementioned works have several limitations that hinder their practical applicability to financial institutions. First, the joint loss function applied is highly non-linear and non-convex, making the resulting optimization problem computationally challenging to solve in practice. Second, the estimators obtained from this joint loss function are generally not asymptotically efficient, which limits their statistical performance. Moreover, even when estimation is feasible, a reliable diagnostic framework is important to assess model adequacy to meet the regulatory requirements for financial institutions. However, research on model diagnostics for expected shortfall remains relatively unexplored, with existing approaches largely relying on the joint elicitability of quantile and expected shortfall, as in [28], Du & Escanciano (2017), and [3].

This manuscript contributes to the growing financial econometrics literature on the use of expected shortfall in risk management in three key aspects. First, we propose to model the autoregressive conditional expected shortfall directly. Compared with existing literature [33, 19], our modeling framework offers greater flexibilities in learning the tail risk dynamics. We develop a computationally efficient two-step procedure to estimate the parameters of the expected shortfall models with theoretical guarantees. At the first stage, we solve an autoregressive conditional quantile regression problem as described in [16]; and at the second stage, we solve a least squares autoregressive regression problem with a surrogate response variable that depends on the autoregressive conditional quantile function obtained from the first stage. We establish the asymptotic properties of the proposed estimators and show that, owing to the orthogonality structure of our framework (see § 3–§ 4 for details), the quantile estimation error from the first stage is first-order negligible. Second, we develop a novel Portmanteau test to assess the goodness-of-fit of the expected shortfall model, and we establish a theoretical guarantee for Type I error control. We further demonstrate empirically that the proposed test exhibits higher power than existing approaches in § 5. Lastly, in the data application presented in § 6, we uncover a leverage effect in the expected shortfall of the U.S. stock market, i.e., the expected shortfall is more sensitive to previous losses than previous gains. In addition, motivated by the optimal weighted least squares scheme considered in Barendse [2], we construct a weighted least squares estimator for the proposed autoregressive expected shortfall model and show that this weighted estimator is more efficient than estimators obtained by existing approaches [29, 13], i.e., our proposed estimator has a smaller asymptotic variance.

The manuscript is organized as follows. § 2 introduces the proposed joint autoregressive quantile-expected shortfall framework. § 3 details the estimation and inference methodology. § 4 presents the theoretical results for the expected shortfall estimator. § 5 reports the results of the numerical experiments. § 6 encompasses the empirical data analysis. We end with a discussion in § 7.

A Joint Autoregressive Quantile-ES Model

Consider a time series {Yt,Xt}t≥1\{Y_t,\mathbf{X}_t\}_{t\geq1} with YtY_t as the response variable and Xt\mathbf{X}_t as the covariate vector at time tt, respectively. We allow the covariate vector Xt\mathbf{X}_t to be endogenous or exogenous. In the endogenous case, Xt\mathbf{X}_t consists of lagged values of YtY_t, i.e., {Yt−i}i≥1\{Y_{t-i}\}_{i\geq1}. In the exogenous case, Xt\mathbf{X}_t includes additional covariates in addition to the lagged values of YtY_t. For a given quantile level τ∈(0,1)\tau\in(0,1), let q1q_1 and e1e_1 be the τ\tau-level (conditional) quantile and the expected shortfall of Y1Y_1, respectively. Moreover, let Ht\mathcal{H}_t be the sigma-algebra generated by {q1,e1,Y1,X1,…,Yt−1,Xt−1,Xt}\{q_1,e_1,Y_1,\mathbf{X}_1,\ldots,Y_{t-1},\mathbf{X}_{t-1},\mathbf{X}_t\} for any t>1t>1. Let qtτ(βτ∗)q_{t\tau}(\boldsymbol{\beta}_\tau^*) and etτ(θτ∗)e_{t\tau}(\boldsymbol{\theta}_\tau^*) be two Ht\mathcal{H}_t-measurable functions with true underlying parameters βτ∗\boldsymbol{\beta}_\tau^* and θτ∗\boldsymbol{\theta}_\tau^*. We model the conditional quantile and the ES defined in (1) as

Qτ(Yt∣Ht)=qtτ(βτ∗)andES⁡τ(Yt∣Ht)=etτ(θτ∗).(2)Q_\tau(Y_t \mid\mathcal{H}_t) = q_{t\tau}(\boldsymbol{\beta}_\tau^*) \quad\text{and} \quad\operatorname{ES}_\tau(Y_t \mid\mathcal{H}_t) = e_{t\tau}(\boldsymbol{\theta}_\tau^*). \tag*{(2)}

To simplify the notation, we omit the dependence of qt(β∗)q_t(\boldsymbol{\beta}^*) and et(θ∗)e_t(\boldsymbol{\theta}^*) on τ\tau in the rest of the paper. Our general model specification is motivated by empirical observations that risky days in the stock market tend to cluster together and financial risk changes smoothly over time [16]. Hence, it is natural to use certain autoregressive specifications to model such characteristics. To this end, we assume that both qt(β∗)q_t(\boldsymbol{\beta}^*) and et(θ∗)e_t(\boldsymbol{\theta}^*) have the following autoregressive structures:

qt(β∗)=β0∗+∑i=1l1βi∗qt−i(β∗)+∑i=0r1βl1+1+i∗g(Xt−i);et(θ∗)=θ0∗+∑i=1l2θi∗et−i(θ∗)+∑i=0r2θl2+1+i∗h(Xt−i),(3)\begin{aligned} q_t(\boldsymbol{\beta}^*) &= \beta_0^* + \sum_{i=1}^{l_1}\beta_i^*q_{t-i}(\boldsymbol{\beta}^*) + \sum_{i=0}^{r_1}\beta_{l_1+1+i}^*g(\boldsymbol{X}_{t-i});\\ e_t(\boldsymbol{\theta}^*) &= \theta_0^* + \sum_{i=1}^{l_2}\theta_i^*e_{t-i}(\boldsymbol{\theta}^*) + \sum_{i=0}^{r_2}\theta_{l_2+1+i}^*h(\boldsymbol{X}_{t-i}), \tag*{(3)} \end{aligned}

where {l1,l2,r1,r2}\{l_1,l_2,r_1,r_2\} are user-specified autoregressive orders. For notational simplicity, let dq=l1+r1+2d_q=l_1+r_1+2 and de=l2+r2+2d_e=l_2+r_2+2. We have β∗=(β0∗,β1∗,…,βl1+r1+1∗)T∈Rdq\boldsymbol{\beta}^*=(\beta_0^*,\beta_1^*,\ldots,\beta_{l_1+r_1+1}^*)^{\mathsf{T}}\in\mathbb{R}^{d_q} and θ∗=(θ0∗,θ1∗,…,θl2+r2+1∗)T∈Rde\boldsymbol{\theta}^*=(\theta_0^*,\theta_1^*,\ldots,\theta_{l_2+r_2+1}^*)^{\mathsf{T}}\in\mathbb{R}^{d_e}. The functions g(⋅)g(\cdot) and h(⋅)h(\cdot) are pre-determined and used to model relationships between qt(β∗)q_t(\boldsymbol{\beta}^*) and et(θ∗)e_t(\boldsymbol{\theta}^*) and the covariates Xt\boldsymbol{X}_t.

The selection of autoregressive orders and the functional forms of g(⋅)g(\cdot) and h(⋅)h(\cdot) in (3) depends on specific financial questions. In practice, it is crucial to select the appropriate model specifications for g(⋅)g(\cdot) and h(⋅)h(\cdot). In § 3.2, we propose a joint quantile-expected shortfall backtesting procedure that allows us to perform statistical tests to assess model fit and determine the appropriate model specification. In the following, we outline three special cases of (3).

Example 1 (Symmetric Absolute Value Model). *For financial data with short-range dependencies, an autoregressive order of one is typically preferred. We set autoregressive orders to l1=l2=r1=r2=1l_1=l_2=r_1=r_2=1. When data exhibit symmetric sensitivities toward stock price impacts, both extremely high/low past returns have the same influence on the current risk level. To capture this pattern, we define the functions g(x)=h(x)=∣x∣g(x)=h(x)=|x|

and let

qt(β∗)=β0∗+β1∗qt−1(β∗)+β2∗∣Yt−1∣,et(θ∗)=θ0∗+θ1∗et−1(θ∗)+θ2∗∣Yt−1∣.q_t(\beta^*) = \beta_0^* + \beta_1^*q_{t-1}(\beta^*) + \beta_2^*|Y_{t-1}|,\qquad e_t(\theta^*) = \theta_0^* + \theta_1^*e_{t-1}(\theta^*) + \theta_2^*|Y_{t-1}|.

Example 2 (Asymmetric Slope Model). Numerous studies have shown that the current risk level exhibits different sensitivities to past losses compared to past gains. This phenomenon is referred to as the leverage effect [5, 10, 39]. To model such an effect, setting g(y)=h(y)={y+,y−}g(y) = h(y) = \{y^+, y^-\} yields

qt(β∗)=β0∗+β1∗qt−1(β∗)+β2∗Yt−1+−β3∗Yt−1−+β4∗Yt−2+−β5∗Yt−2−,q_t(\beta^*) = \beta_0^* + \beta_1^*q_{t-1}(\beta^*) + \beta_2^*Y_{t-1}^+ - \beta_3^*Y_{t-1}^- + \beta_4^*Y_{t-2}^+ - \beta_5^*Y_{t-2}^-,
et(θ∗)=θ0∗+θ1∗et−1(θ∗)+θ2∗Yt−1+−θ3∗Yt−1−+θ4∗Yt−2+−θ5∗Yt−2−,e_t(\theta^*) = \theta_0^* + \theta_1^*e_{t-1}(\theta^*) + \theta_2^*Y_{t-1}^+ - \theta_3^*Y_{t-1}^- + \theta_4^*Y_{t-2}^+ - \theta_5^*Y_{t-2}^-,

where y+y^+ and y−y^- correspond to y1{y≥0}y\mathbb{1}\{y \ge0\} and y1{y<0}y\mathbb{1}\{y < 0\}, respectively, and 1{⋅}\mathbb{1}\{\cdot\} is the indicator function.

Example 3 (GARCH Model). Consider Yt=σtηtY_t = \sigma_t\eta_t, where {ηt}t≥1\{\eta_t\}_{t\ge1} is a sequence of independent and identically distributed standard normal random variables and

σt2=α0∗+α1∗σt−12+α2∗Yt−12.\sigma_t^2 = \alpha_0^* + \alpha_1^*\sigma_{t-1}^2 + \alpha_2^*Y_{t-1}^2.

This model has been widely employed to analyze the autoregressive structure in the conditional volatilities of YtY_t. For τ<0.5\tau< 0.5, the conditional quantile and expected shortfall functions are

qt(β∗)=−β0∗+β1∗qt−12(β∗)+β2∗Yt−12,et(θ∗)=−θ0∗+θ1∗et−12(θ∗)+θ2∗Yt−12.q_t(\beta^*) = -\sqrt{\beta_0^* + \beta_1^*q_{t-1}^2(\beta^*) + \beta_2^*Y_{t-1}^2},\qquad e_t(\theta^*) = -\sqrt{\theta_0^* + \theta_1^*e_{t-1}^2(\theta^*) + \theta_2^*Y_{t-1}^2}.

Remark 2.1. The expected shortfall, as defined in (1), is a statistic that integrates tail information beyond the given quantile level τ\tau. The two separate model specifications in (3) reflect the fact that the conditional expected shortfall may exhibit different dynamic behavior from the conditional quantile function, due to the additional tail information not captured by the quantile. To illustrate, consider

Yt=qt{β(Ut)}=β0(Ut)+β1∗qt−1{β(Ut)}+β2(Ut)∣Yt−1∣+{β3(Ut)−β3(τ0)}∣Yt−2∣.Y_t = q_t\{\beta(U_t)\} = \beta_0(U_t) + \beta_1^*q_{t-1}\{\beta(U_t)\} + \beta_2(U_t)|Y_{t-1}| + \{\beta_3(U_t)-\beta_3(\tau_0)\}|Y_{t-2}|.

Here, {Ut}t≥1\{U_t\}_{t\geq1} is a sequence of Uniform(0,1)(0,1) random variables independent of Ht\mathcal{H}_t. The vector β(u)={β0(u),β1∗,β2(u),β3(u)−β3(τ0)}T\beta(u) = \{\beta_0(u),\beta_1^*,\beta_2(u),\beta_3(u)-\beta_3(\tau_0)\}^{\mathsf{T}} satisfies that given the history Ht\mathcal{H}_t, qt{β(u)}q_t\{\beta(u)\} is strictly increasing with respect to u∈(0,1)u \in(0,1). For a fixed quantile level τ=τ0\tau=\tau_0, we have

qt(β∗)=β0∗+β1∗qt−1(β∗)+β2∗∣Yt−1∣,et(θ∗)=θ0∗+θ1∗et−1(θ∗)+θ2∗∣Yt−1∣+θ3∗∣Yt−2∣,q_t(\beta^*) = \beta_0^* + \beta_1^*q_{t-1}(\beta^*) + \beta_2^*|Y_{t-1}|,\qquad e_t(\theta^*) = \theta_0^* + \theta_1^*e_{t-1}(\theta^*) + \theta_2^*|Y_{t-1}| + \theta_3^*|Y_{t-2}|,

where θ∗={∫0τ0β0(u) du,τ0β1∗,∫0τ0β2(u) du,∫0τ0β3(u) du−τ0β3(τ0)}T/τ0\theta^* = \{\int_0^{\tau_0}\beta_0(u)\,du,\tau_0\beta_1^*,\int_0^{\tau_0}\beta_2(u)\,du,\int_0^{\tau_0}\beta_3(u)\,du-\tau_0\beta_3(\tau_0)\}^{\mathsf{T}}/\tau_0 and β∗={β0(τ0),β1∗,β2(τ0)}T\beta^*=\{\beta_0(\tau_0),\beta_1^*,\beta_2(\tau_0)\}^{\mathsf{T}}. Here, et(θ∗)e_t(\theta^*) exhibits different dynamic behavior compared with qt(β∗)q_t(\beta^*) due to the extra lag quantity ∣Yt−2∣|Y_{t-2}|.

Methodology

In this section, we propose a framework for estimating and performing statistical inference for the expected shortfall parameters θ∗\theta^* in (3). We propose a two-step regression procedure for parameter estimation in Section 3.1. To perform model diagnostics, we construct a Portmanteau test to test whether the working conditional quantile and expected shortfall functions are correctly specified in Section 3.2.

Weighted Two-Step Autoregressive ES Regression

We propose to estimate the expected shortfall parameter θ∗\theta^{*} using a two-step approach.

Let S0(β,θ;Yt,Xt)=et(θ)−qt(β)−1{Yt≤qt(β)}{Yt−qt(β)}/τS_{0}(\beta,\theta;Y_{t},X_{t})=e_{t}(\theta)-q_{t}(\beta)-\mathbf{1}\{Y_{t}\leq q_{t}(\beta)\}\{Y_{t}-q_{t}(\beta)\}/\tau. This quantity can be interpreted as a score function for β\beta and θ\theta under the joint quantile-expected shortfall framework in (2); see also He et al. (2023) and Barendse [2] for similar score functions for the linear quantile-expected shortfall model. In the following proposition, we establish an orthogonality property for the score function S0(β,θ;Yt,Xt)S_{0}(\beta,\theta;Y_{t},X_{t}) in the context of an autoregressive quantile-expected shortfall model with time series data.

Proposition 3.1. For any time tt, we have

E{S0(β∗,θ∗;Yt,Xt)∣Ht}=0and∂E{S0(β,θ;Yt,Xt)∣Ht}∂β∣β=β∗=0.\mathbb{E}\{S_{0}(\beta^{*},\theta^{*};Y_{t},X_{t})\mid\mathcal{H}_{t}\}=0 \quad\text{and}\quad \left.\frac{\partial\mathbb{E}\{S_{0}(\beta,\theta;Y_{t},X_{t})\mid\mathcal{H}_{t}\}}{\partial\beta}\right|_{\beta=\beta^{*}}=0.

Proposition 3.1 states that, given the history Ht\mathcal{H}_{t}, S0(β,θ;Yt,Xt)S_{0}(\beta,\theta;Y_{t},X_{t}) is an identification function for β\beta and θ\theta and is locally robust to small perturbations of β\beta around the true value β∗\beta^{*}. Motivated by the above observation, the estimator for θ∗\theta^{*} can be obtained by first estimating β∗\beta^{*} through solving a nonlinear quantile regression problem, followed by estimating θ∗\theta^{*} using the score function S0(β,θ;Yt,Xt)S_{0}(\beta,\theta;Y_{t},X_{t}) with a plug-in quantile estimator β^T\widehat{\beta}_{T}.

Assume that the conditional quantile and expected shortfall functions qt(β)q_{t}(\beta) and et(θ)e_{t}(\theta) have autoregressive structures as in (3). In the first step, we propose to solve the following non-linear quantile regression problem:

β^T∈argmin⁡β∈B 1T∑t=1Tρτ{Yt−qt(β)},(4)\widehat{\beta}_{T}\in\underset{\beta\in B}{\operatorname{argmin}}\ \frac{1}{T}\sum_{t=1}^{T}\rho_{\tau}\{Y_{t}-q_{t}(\beta)\}, \tag*{(4)}

where ρτ(u)=u[τ−1{u≤0}]\rho_{\tau}(u)=u[\tau-\mathbf{1}\{u\leq0\}] is the check loss function and BB is the parameter space for β\boldsymbol{\beta}. Given β^T\hat{\boldsymbol{\beta}}_T, we propose to estimate the expected shortfall parameters by minimizing the summation of the squared score function S0(β,θ;Yt,Xt)S_0(\boldsymbol{\beta},\boldsymbol{\theta};Y_t,X_t):

θ^T∈argmin⁡θ∈Θ12T∑t=1TS02(β^T,θ;Yt,Xt),(5)\hat{\boldsymbol{\theta}}_T \in\operatorname*{argmin}_{\boldsymbol{\theta}\in\Theta} \frac{1}{2T}\sum_{t=1}^{T} S_0^2(\hat{\boldsymbol{\beta}}_T,\boldsymbol{\theta};Y_t,X_t), \tag*{(5)}

where Θ\Theta is the parameter space for θ\boldsymbol{\theta}. The problem (5) can be seen as a non-linear least squares problem with constructed pseudo-response variables {qt(β^T)+1{Yt≤qt(β^T)}{Yt−qt(β^T)}/τ}t≥1\{q_t(\hat{\boldsymbol{\beta}}_T)+\mathbb{1}\{Y_t\leq q_t(\hat{\boldsymbol{\beta}}_T)\}\{Y_t-q_t(\hat{\boldsymbol{\beta}}_T)\}/\tau\}_{t\geq1}.

The estimator θ^T\hat{\boldsymbol{\theta}}_T in (5) is generally not asymptotically efficient, i.e., it does not have the minimal asymptotic variance, due to the presence of conditional heteroskedasticity in the regression problem (5). In the existing literature, a weighted least squares estimator has been shown to be generally more efficient than the unweighted least squares estimator, provided that the weighting function in the regression problem is set to be the inverse of the conditional variance of the regression error terms (White [35]; Wooldridge [37]). Hence, to achieve a more efficient estimator than θ^T\hat{\boldsymbol{\theta}}_T, the key lies in consistently estimating the conditional heteroskedasticity terms in the regression problem (5). It can be shown that

var⁡{S0(β∗,θ∗;Yt,Xt)∣Ht}=1τvar⁡{Yt∣Yt≤qt(β∗),Ht}+1−ττ{qt(β∗)−et(θ∗)}2.(6)\operatorname{var}\{S_0(\boldsymbol{\beta}^{*},\boldsymbol{\theta}^{*};Y_t,X_t)\mid\mathcal{H}_t\}=\frac{1}{\tau}\operatorname{var}\{Y_t\mid Y_t\leq q_t(\boldsymbol{\beta}^{*}),\mathcal{H}_t\}+\frac{1-\tau}{\tau}\{q_t(\boldsymbol{\beta}^{*})-e_t(\boldsymbol{\theta}^{*})\}^{2}. \tag*{(6)}

However, there is no explicit parametric form of the term var⁡{Yt∣Yt≤qt(β∗),Ht}\operatorname{var}\{Y_t\mid Y_t\leq q_t(\boldsymbol{\beta}^{*}),\mathcal{H}_t\} under the general data generating process given in (2). Estimating the conditional variance can be challenging without further modeling assumptions.

To this end, we propose a working model for the weighting function under the location-scale family time series. Specifically, the time series {Yt}t≥1\{Y_t\}_{t\geq1} can be written in the following form:

Yt=qt(β∗)+{qt(β∗)−et(θ∗)}ζt,(7)Y_t = q_t(\beta^*) + \{q_t(\beta^*) - e_t(\theta^*)\}\zeta_t, \tag*{(7)}

where Qτ(ζt∣Ht)Q_\tau(\zeta_t \mid\mathcal{H}_t) and ES⁡τ(ζt∣Ht)\operatorname{ES}_\tau(\zeta_t \mid\mathcal{H}_t) are 0 and −1-1, respectively. As shown in Proposition S5.1 of Section S5.1 in the online supplementary material, (7) is equivalent to a general class of location-scale time series models with independent and identically distributed (i.i.d.i.i.d.) innovations, provided that {ζt}t≥1\{\zeta_t\}_{t\geq1} is an i.i.d.i.i.d. sequence of appropriately scaled random variables and ζt\zeta_t is independent of Ht\mathcal{H}_t for any tt. Under the i.i.d.i.i.d. assumption on the innovation process, the conditional variance of the regression error terms in (5) can be expressed as

var⁡{S0(β∗,θ∗;Yt,Xt)∣Ht}={qt(β∗)−et(θ∗)}2{1τE(ζt2∣ζt≤0)−1}.(8)\operatorname{var}\{S_0(\beta^*,\theta^*;Y_t,X_t)\mid\mathcal{H}_t\}=\{q_t(\beta^*)-e_t(\theta^*)\}^2\left\{\frac{1}{\tau}\mathbb{E}(\zeta_t^2\mid\zeta_t\leq0)-1\right\}. \tag*{(8)}

Hence, var⁡{S0(β∗,θ∗;Yt,Xt)∣Ht}∝{qt(β∗)−et(θ∗)}2\operatorname{var}\{S_0(\beta^*,\theta^*;Y_t,X_t)\mid\mathcal{H}_t\}\propto\{q_t(\beta^*)-e_t(\theta^*)\}^2 for t≥1t\geq1. Motivated by the above derivation, we set the weighting function for the tt-th observation as 1/{qt(β∗)−et(θ∗)}21/\{q_t(\beta^*)-e_t(\theta^*)\}^2 with the corresponding true parameters {(β∗)T,(θ∗)T}T\{(\beta^*)^\mathsf{T},(\theta^*)^\mathsf{T}\}^\mathsf{T}.

Let β^T\hat{\beta}_T and θ^T\hat{\theta}_T be estimators obtained from solving (4) and (5), respectively. We propose a weighted expected shortfall estimator θ^TW\hat{\theta}^{W}_T that takes the following form:

θ^TW∈arg⁡min⁡θ∈Θ 12T∑t=1T1{qt(β^T)−et(θ^T)}2S02(β^T,θ;Yt,Xt).(9)\hat{\theta}^{W}_T\in\underset{\theta\in\Theta}{\arg\min}\ \frac{1}{2T}\sum_{t=1}^{T}\frac{1}{\{q_t(\hat{\beta}_T)-e_t(\hat{\theta}_T)\}^2}S_0^2(\hat{\beta}_T,\theta;Y_t,X_t). \tag*{(9)}

For any j∈{1,2,…,de}j\in\{1,2,\ldots,d_e\} and α∈(0,1)\alpha\in(0,1), the 1−α1-\alpha-level confidence interval for θj∗\theta_j^* can then be constructed as

[θ^T,jW±Φ−1(1−α2)1T{(Σ^TW)−1Ω^TW(Σ^TW)−1}jj],(10)\left[\hat{\theta}^{W}_{T,j}\pm\Phi^{-1}\left(1-\frac{\alpha}{2}\right)\sqrt{\frac{1}{T}\left\{(\hat{\Sigma}^{W}_T)^{-1}\hat{\Omega}^{W}_T(\hat{\Sigma}^{W}_T)^{-1}\right\}_{jj}}\right], \tag*{(10)}

where Ω^TW=∑t=1TS02(β^T,θ^TW;Yt,Xt)∇et(θ^TW)∇et(θ^TW)T/[T{qt(β^T)−et(θ^T)}4]\hat{\Omega}^{W}_T=\sum_{t=1}^{T}S_0^2(\hat{\beta}_T,\hat{\theta}^{W}_T;Y_t,X_t)\nabla e_t(\hat{\theta}^{W}_T)\nabla e_t(\hat{\theta}^{W}_T)^\mathsf{T}/[T\{q_t(\hat{\beta}_T)-e_t(\hat{\theta}_T)\}^4] and

Σ^TW=∑t=1T∇et(θ^TW)∇et(θ^TW)T/[T{qt(β^T)−et(θ^T)}2]\widehat{\Sigma}_T^W = \sum_{t=1}^{T} \nabla e_t(\widehat{\theta}_T^W)\nabla e_t(\widehat{\theta}_T^W)^{\mathsf{T}} \big/ \left[T\{q_t(\widehat{\beta}_T)-e_t(\widehat{\theta}_T)\}^2\right]

and Φ(⋅)\Phi(\cdot) is the cumulative distribution function for a standard normal distribution.

Additionally, we also consider an alternative approach to construct confidence intervals based on random weighting bootstrap in § 5 for numerical stability. We defer the details of this construction to Section S2.2 of the online supplementary material.

Moreover, it should be noted that while our choice of the weighting function 1/{qt(β∗)−et(θ∗)}21/\{q_t(\beta^*)-e_t(\theta^*)\}^2 is motivated by the location-scale time series models with i.i.d. innovations, the validity of the proposed inference procedure does not rely on such specific data-generating assumptions, as illustrated by Theorems 4–5 of § 4. Under such an i.i.d. assumption, our weighted estimator attains asymptotic efficiency within the class of joint MM-estimators considered in Fissler & Ziegel [17], as demonstrated in Corollary 4.1 of § 4.

Finally, we provide a generic algorithm to solve (9). Our optimization steps include Broyden–Fletcher–Goldfarb–Shanno (BFGS) Fletcher & Reeves [18] and Nelder-Mead Nelder & Mead [27] algorithms. Our implementation is based on applying optim function in R. We note that (4) and (5) can be solved by slight modifications of Algorithm 1.

Remark 3.1. The optimal weighting function derived in this section relies on the assumption that the data are generated from a location-scale family time series model with i.i.d. innovations. The weighting function is no longer optimal when this assumption is violated. However, since the term {qt(β∗)−et(θ∗)}2\{q_t(\beta^*)-e_t(\theta^*)\}^2 remains an important component of (6), our weighting function can still be viewed as a working approximation to the optimal weights when (7) no longer holds. We conduct numerical experiments to quantify the efficiency loss of our weighted estimator (9) when the data-generating

process deviates from the location-scale family with i.i.d. innovations. Our results show that, even when the working weighting function is suboptimal in such settings, the efficiency loss is very small. See § 5, Model (5.2), and Section S9 of the online supplementary material for details.

Algorithm 1 A Framework for Solving (9)

Step 1: Randomly sample 10,000 initial parameter vectors from a uniform distribution.

Step 2: Among the 10,000 initial parameters, select 10 vectors {θ^(1)0,…,θ^(10)0}\{\hat{\theta}^{0}_{(1)}, \ldots, \hat{\theta}^{0}_{(10)}\} that return the lowest empirical loss value in (9).

Step 3:

for i=1,…,10i = 1, \ldots, 10 do

Initialize: Set the starting value θ^(i)′←θ^(i)0\hat{\theta}^{\prime}_{(i)} \leftarrow\hat{\theta}^{0}_{(i)}.

Refinement loop: Repeat the following procedure 5 times and output θ^(i)\hat{\theta}_{(i)}:

(i) BFGS update: Obtain θ^(i)\hat{\theta}_{(i)} by optimizing (9) using the BFGS algorithm initialized at θ^(i)′\hat{\theta}^{\prime}_{(i)}.

(ii) Perturbation: Add a small Gaussian perturbation

θ^(i)′←θ^(i)+ι,ι∼N(0,0.1I).\hat{\theta}^{\prime}_{(i)} \leftarrow\hat{\theta}_{(i)} + \iota,\qquad\iota\sim N(0, 0.1I).

(iii) Nelder–Mead update: Obtain θ^(i)\hat{\theta}_{(i)} by optimizing (9) with Nelder–Mead algorithm initialized at θ^(i)′\hat{\theta}^{\prime}_{(i)}.

(iv) Perturbation: Repeat step (ii).

end for

Step 4: Compute the empirical loss in (9) for {θ^(1),…,θ^(10)}\{\hat{\theta}_{(1)}, \ldots, \hat{\theta}_{(10)}\}. Select the final estimator θ^TW\hat{\theta}^{W}_{T} as the vector achieving the smallest empirical loss.

Model Diagnostics via a Portmanteau Test

Under the autoregressive quantile-ES framework in (3), it is important to accurately determine the appropriate autoregressive orders and functional forms of g(⋅)g(\cdot) and h(⋅)h(\cdot). In this section, we assume that the data {Yt}t≥1\{Y_t\}_{t \ge1} are generated by the location-scale family time series with i.i.d. innovations. Under this assumption, we develop a Portmanteau test for model diagnostics [23] (Li & Mak 1994).

Let ζt={Yt−qt(β∗)}/{qt(β∗)−et(θ∗)}\zeta_t = \{Y_t - q_t(\beta^*)\}/\{q_t(\beta^*) - e_t(\theta^*)\}. Under the assumption that {Yt}t≥1\{Y_t\}_{t\ge1} is generated by the location-scale family time series with i.i.d. innovations, and that both quantile and expected shortfall functions are correctly specified with true parameters β∗\beta^* and θ∗\theta^* known, it can be shown that {ζt}t≥1\{\zeta_t\}_{t\ge1} is a sequence of i.i.d. random variables, see details in Proposition S5.1 in the online supplementary material. Therefore, the dependence structure of {ζt}t≥1\{\zeta_t\}_{t\ge1} reflects whether qt(β∗)q_t(\beta^*) and et(θ∗)e_t(\theta^*) correctly specify the conditional quantile and expected shortfall of YtY_t, respectively. Motivated by this observation, we propose the following model diagnostic procedure based on detecting serial dependence in {ζt}t≥1\{\zeta_t\}_{t\ge1}.

For any positive integer mm, let rm=corr⁡(ζt2,ζt−m2)r_m = \operatorname{corr}(\zeta_t^2, \zeta_{t-m}^2). We develop a Portmanteau test to test the following hypothesis problem:

H0:rm=0 ∀ m≥1versusHa:∃ m≥1 such that rm≠0.(11)H_0: r_m = 0\ \forall\ m \ge1 \quad\text{versus} \quad H_a: \exists\ m \ge1\ \text{such that}\ r_m \ne0. \tag*{(11)}

In practice, it is infeasible to examine all lag levels. To this end, we pick a positive integer MM and consider testing the hypothesis problem

H0:rm=0, ∀ m∈[M]versusHa:∃ m∈[M] such that rm≠0,(12)H_0: r_m = 0,\ \forall\ m \in[M] \quad\text{versus} \quad H_a: \exists\ m \in[M]\ \text{such that}\ r_m \ne0, \tag*{(12)}

where [M]={1,…,M}[M] = \{1,\ldots,M\}. The integer MM should be chosen such that YtY_t and Yt+MY_{t+M} are approximately independent. Common choices of MM include {4,6,8,10}\{4,6,8,10\}. Our simulation results in § 5 show that the empirical performance of the Portmanteau test is robust to the choice of MM. In the real data applications in § 6, we set M=10M = 10, which corresponds to the length of two trading weeks. We now outline the procedure for performing hypothesis testing on (12); detailed derivations of the proposed testing procedure are deferred to Section S2.3 of the online supplementary material. The test is based on the estimators β^T\hat{\beta}_T and θ^T\hat{\theta}_T obtained by (4) and (5), respectively. We construct the estimated noise term as ζ^t=Yt−qt(β^T)qt(β^T)−et(θ^T)\hat{\zeta}_t=Y_t-\frac{q_t(\hat{\beta}_T)}{q_t(\hat{\beta}_T)-e_t(\hat{\theta}_T)}.

Let r~=(r~1,…,r~M)T\tilde{r}=(\tilde{r}_1,\ldots,\tilde{r}_M)^{\mathsf{T}} be the vector of the sample autocorrelation coefficients of {ζ^t2}t≥1\{\hat{\zeta}_t^2\}_{t\geq1} up to the lag MM. The derivation steps given in the proof of Theorem 6 show that under H0H_0, Tr~\sqrt{T}\tilde{r} is asymptotically normally distributed with mean 0\mathbf{0} and variance Ωtest\boldsymbol{\Omega}^{\mathrm{test}}. Furthermore, Tr~T(Ωtest)−1r~T\tilde{r}^{\mathsf{T}}(\boldsymbol{\Omega}^{\mathrm{test}})^{-1}\tilde{r} converges in distribution to a χM2\chi^2_M random variable.

The general procedure for testing (12) is summarized in Algorithm 2. The closed form expression of Ωtest\boldsymbol{\Omega}^{\mathrm{test}} involves many intermediate quantities and is omitted from the main manuscript. We refer the reader to Section S2.3 in the online supplementary material for the details on the derivation and estimation of Ωtest\boldsymbol{\Omega}^{\mathrm{test}}.

Algorithm 2 The Portmanteau Test for Testing (12).

  1. Step1: Set a positive integer MM. Typical choices of MM include {4,6,8,10}\{4,6,8,10\}.

  2. Step2: Obtain β^T\hat{\beta}_T and θ^T\hat{\theta}_T by solving (4) and (5) with dataset {Yt,Xt}t≥1\{Y_t,X_t\}_{t\geq1}.

  3. Step3: Compute {ζ^t}t≥1=[Yt−qt(β^T)qt(β^T)−et(θ^T)]t≥1\{\hat{\zeta}_t\}_{t\geq1}=\left[\frac{Y_t-q_t(\hat{\beta}_T)}{q_t(\hat{\beta}_T)-e_t(\hat{\theta}_T)}\right]_{t\geq1} and r~\tilde{r}.

  4. Step4: Estimate Ω^test\hat{\boldsymbol{\Omega}}^{\mathrm{test}} from data.

  5. Step5: Let ψ^=Tr~T(Ω^test)−1r~\hat{\psi}=T\tilde{r}^{\mathsf{T}}(\hat{\boldsymbol{\Omega}}^{\mathrm{test}})^{-1}\tilde{r}. Reject the null hypothesis in (12) if the test statistic ψ^≥χM,1−α2\hat{\psi}\geq\chi^2_{M,1-\alpha} for a given confidence level α∈(0,1)\alpha\in(0,1).

Theoretical Results

We study the theoretical properties of the proposed two-step expected shortfall regression estimators in Section 3. We first show that the unweighted expected shortfall estimator θ^T\hat{\theta}_T is weakly consistent, and provide the convergence rate and asymptotic normality of θ^T\hat{\theta}_T in Section 4.1. Parallel theoretical results for the weighted estimator θ^TW\hat{\theta}_T^W are presented in Section 4.2. The theoretical guarantee of the proposed Portmanteau test is included in Section 4.3. We defer all assumptions required for the theoretical results to Section S3 of the online supplementary material.

We start with defining some notation. For a vector r=(r1,…,rp)T∈Rp\mathbf{r}=(r_1,\ldots,r_p)^{\mathsf{T}}\in\mathbb{R}^p, let

∥r∥q=(∑j=1p∣rj∣q)1/q\|\mathbf{r}\|_q=\left(\sum_{j=1}^{p}|r_j|^q\right)^{1/q} for 1≤q<∞1\le q<\infty. For sequences {an}\{a_n\} and {bn}\{b_n\}, we write an=O(bn)a_n=O(b_n) if there exists a constant CC such that an≤Cbna_n\le Cb_n for all nn and an=Op(bn)a_n=O_p(b_n) if an/bn=Op(1)a_n/b_n=O_p(1). We write an=o(bn)a_n=o(b_n) if lim⁡n→∞an/bn=0\lim_{n\to\infty}a_n/b_n=0 and an=op(bn)a_n=o_p(b_n) if an/bn→p0a_n/b_n\xrightarrow{p}0.

Let Et(⋅)=E(⋅∣Ht)\mathbb{E}_t(\cdot)=\mathbb{E}(\cdot\mid\mathcal{H}_t). For a function f(β):Rdim⁡(β)→Rf(\beta):\mathbb{R}^{\dim(\beta)}\to\mathbb{R}, let ∇f(β)=∂f(β)/∂β\nabla f(\beta)=\partial f(\beta)/\partial\beta be the gradient of f(β)f(\beta) with respect to the parameter β\beta.

Unweighted Expected Shortfall Estimator

We derive the asymptotic properties of θ^T\hat{\theta}_T obtained by solving (5). The first step of our proposed approach involves solving an autoregressive quantile regression problem (4). Let

E1T={β^T∈B(β∗,rT)={β∈B; ∥β−β∗∥2≤rT}; rT=o(T−1/4)}(13)\mathcal{E}_{1T}=\left\{\hat{\beta}_T\in\mathbb{B}(\beta^*,r_T)=\{\beta\in B;\ \|\beta-\beta^*\|_2\le r_T\};\ r_T=o(T^{-1/4})\right\} \tag*{(13)}

be an event characterizing the convergence rate of quantile estimator β^T\hat{\beta}_T. The event E1T\mathcal{E}_{1T} has been shown to hold with high probability tending to 1 as TT increases (Xu & Härdle 2026), which implies that β^T\hat{\beta}_T is weakly consistent with a mild rate. In the following, we will conduct our analysis, conditioned on the high-probability event E1T\mathcal{E}_{1T}.

We start by showing that θ^T\hat{\theta}_T obtained from (5) is weakly consistent. Our theoretical results require a series of regularity assumptions on the smoothness of quantile-expected shortfall functions, the existence of certain envelope functions, the stationarity and ergodicity conditions, and identifiability conditions. Specifically, our theoretical analysis requires the existence of sixth-moment bounds on YtY_t and some corresponding envelope functions. This assumption allows the data to be heavy-tailed to a certain extent, but not without limitation. We refer readers to Assumption S1 of the online supplementary material for details.

Theorem 1. Under Assumption S1 and the event E1T\mathcal{E}_{1T}, the estimator θ^T\hat{\boldsymbol{\theta}}_T is weakly consistent: θ^T→pθ∗\hat{\boldsymbol{\theta}}_T \xrightarrow{p} \boldsymbol{\theta}^{*}.

Next, we characterize an upper bound for the estimation error of θ^T\hat{\boldsymbol{\theta}}_T under the ℓ2\ell_2 norm and establish its asymptotic distribution. For notational simplicity, let St(β,θ)=S0(β,θ;Yt,Xt)S_t(\boldsymbol{\beta},\boldsymbol{\theta})=S_0(\boldsymbol{\beta},\boldsymbol{\theta};Y_t,\boldsymbol{X}_t). We define the event E2T\mathcal{E}_{2T} as:

{sup⁡β∈B(β∗,rT)∥1T∑t=1T(1−Et){St(β,θ∗)∇et(θ∗)−St(β∗,θ∗)∇et(θ∗)}∥2≤O(sTT)}.\left\{\sup_{\boldsymbol{\beta}\in\mathbb{B}(\boldsymbol{\beta}^{*},r_T)}\left\Vert\frac{1}{T}\sum_{t=1}^{T}(1-\mathbb{E}_t)\left\{S_t(\boldsymbol{\beta},\boldsymbol{\theta}^{*})\nabla e_t(\boldsymbol{\theta}^{*})-S_t(\boldsymbol{\beta}^{*},\boldsymbol{\theta}^{*})\nabla e_t(\boldsymbol{\theta}^{*})\right\}\right\Vert_2\le O\left(\frac{s_T}{\sqrt{T}}\right)\right\}.

where lim⁡T→∞sT=0\lim_{T\to\infty}s_T=0. The event E2T\mathcal{E}_{2T} establishes the stochastic equicontinuity condition for the quantity 1/T∑t=1T(1−Et)S0(β,θ∗;Yt,Xt)∇et(θ∗)1/T\sum_{t=1}^{T}(1-\mathbb{E}_t)S_0(\boldsymbol{\beta},\boldsymbol{\theta}^{*};Y_t,\boldsymbol{X}_t)\nabla e_t(\boldsymbol{\theta}^{*}). This property is a crucial condition to obtain the estimation error bound and the asymptotic normality of θ^T\hat{\boldsymbol{\theta}}_T. This condition can be verified to hold under some sufficient conditions. In Section S4 of the online supplementary material, we provide a proof demonstrating that P(E2T)→1\mathbb{P}(\mathcal{E}_{2T})\to1 under certain moment bounds and convergence rate conditions.

The derivation of the asymptotic properties of θ^T\hat{\boldsymbol{\theta}}_T relies on an upper bound for the conditional density of the quantile error ϵt\epsilon_t and the existence of a positive definite covariance matrix. These requirements are summarized in Assumption S2.

Theorem 2. Under Assumptions S1–S2 and the event E1T∩E2T\mathcal{E}_{1T}\cap\mathcal{E}_{2T}, we have (i) ∥θ^T−θ∗∥2=Op{(1+sT)/T+rT2}\left\Vert\hat{\boldsymbol{\theta}}_T-\boldsymbol{\theta}^{*}\right\Vert_2=O_p\left\{(1+s_T)/\sqrt{T}+r_T^2\right\} and (ii) TΩT−1/2ΣT(θ^T−θ∗)→dN(0,Ide)\sqrt{T}\boldsymbol{\Omega}_T^{-1/2}\boldsymbol{\Sigma}_T(\hat{\boldsymbol{\theta}}_T-\boldsymbol{\theta}^{*})\xrightarrow{d}\mathcal{N}(0,\boldsymbol{I}_{d_e}).

The conclusion of Theorem 2 consists of two parts. The first part characterizes the convergence rate of θ^T\hat{\boldsymbol{\theta}}_T. Our results demonstrate that the estimation error induced by β^T\hat{\boldsymbol{\beta}}_T is first-order negligible and does not affect the asymptotic distribution of θ^T\hat{\boldsymbol{\theta}}_T. This finding aligns with the linear expected shortfall regression results in He et al. (2023).

The second part establishes the asymptotic normality of θ^T\hat{\theta}_T. The variance-covariance matrix ΣT−1ΩTΣT−1\Sigma_T^{-1}\Omega_T\Sigma_T^{-1} exhibits a structure similar to that of the results in Patton et al. [29], although they are not asymptotically equivalent.

Weighted Expected Shortfall Estimator

In this section, we derive the asymptotic properties of the weighted estimator θ^TW\hat{\theta}^{W}_T obtained by solving (9). We start by imposing two high-level events:

E3T={(β^TT,θ^TT)T∈B[{(β∗)T,(θ∗)T}T,dT]; dT=o(1)}(14)\mathcal{E}_{3T}=\left\{\left(\hat{\beta}^{\mathsf{T}}_T,\hat{\theta}^{\mathsf{T}}_T\right)^{\mathsf{T}}\in\mathbb{B}\left[\left\{\left(\beta^{*}\right)^{\mathsf{T}},\left(\theta^{*}\right)^{\mathsf{T}}\right\}^{\mathsf{T}},d_T\right];\ d_T=o(1)\right\} \tag*{(14)}

and the event E4T\mathcal{E}_{4T} is defined as

{sup⁡(βT,θT)T∈B∗∥1T∑t=1T(1−Et){St(β,θ∗)∇et(θ∗){qt(β)−et(θ)}2−St(β∗,θ∗)∇et(θ∗){qt(β∗)−et(θ∗)}2}∥2≤O(sTT)}.\left\{\sup_{(\beta^{\mathsf{T}},\theta^{\mathsf{T}})^{\mathsf{T}}\in\mathbb{B}^{*}}\left\|\frac{1}{T}\sum_{t=1}^{T}(1-\mathbb{E}_t)\left\{\frac{S_t(\beta,\theta^{*})\nabla e_t(\theta^{*})}{\{q_t(\beta)-e_t(\theta)\}^{2}}-\frac{S_t(\beta^{*},\theta^{*})\nabla e_t(\theta^{*})}{\{q_t(\beta^{*})-e_t(\theta^{*})\}^{2}}\right\}\right\|_2\leq O\left(\frac{s_T}{\sqrt{T}}\right)\right\}.

Here, we use B∗\mathbb{B}^{*} to denote B[{(β∗)T,(θ∗)T}T,dT]\mathbb{B}\left[\left\{\left(\beta^{*}\right)^{\mathsf{T}},\left(\theta^{*}\right)^{\mathsf{T}}\right\}^{\mathsf{T}},d_T\right] for notational simplicity. When the parameters for the weighting function can be consistently estimated, the event E3T\mathcal{E}_{3T} holds with high probability converging to one as TT increases, which is a common assumption required by weighted least squares in the existing literature ([35, 31]). For a given model specification (7), E3T\mathcal{E}_{3T} can be satisfied through E1T\mathcal{E}_{1T} and Theorem 1. The event E4T\mathcal{E}_{4T} extends E2T\mathcal{E}_{2T} by incorporating the weighting functions. A detailed discussion of the validity of E4T\mathcal{E}_{4T} is provided in Section S4 of the online supplementary material.

In the following, we derive a weak consistency result for the weighted least squares expected shortfall estimator θ^TW\hat{\theta}^{W}_T. The regularity conditions are in Assumption S3, which generalizes Assumption S1 by incorporating the weighting functions.

Theorem 3. Under Assumptions S1S1–S3S3 and the event E1T∩E3T\mathcal{E}_{1T} \cap\mathcal{E}_{3T}, the estimator θ^TW\widehat{\theta}^{W}_{T} is weakly consistent: θ^TW→pθ∗\widehat{\theta}^{W}_{T} \xrightarrow{p} \theta^{*}.

In § 3.1, we estimate θ\theta using an unweighted least squares regression, where β∗\beta^{*} is replaced by β^T\widehat{\beta}_{T} in the score function S0(β,θ;Yt,Xt)S_{0}(\beta,\theta;Y_{t},X_{t}). The orthogonality property for β\beta is established in Proposition 3.1. We extend this orthogonality condition for the weighted least squares regression problem (9), as formalized in Proposition S5.2 in Section S5.1 of the online supplementary material. Based on Proposition S5.2, we analyze the asymptotic behaviors of θ^TW\widehat{\theta}^{W}_{T}. The derivation follows closely from the proof of Theorem 2.

Theorem 4. Under Assumptions S1S1–S3S3 and the event E1T∩E3T∩E4T\mathcal{E}_{1T} \cap\mathcal{E}_{3T} \cap\mathcal{E}_{4T}, we have (i) ∥θ^TW−θ∗∥2=Op{(1+sT)/T+rT2}\lVert\widehat{\theta}^{W}_{T}-\theta^{*}\rVert_{2}=O_{p}\left\{(1+s_{T})/\sqrt{T}+r_{T}^{2}\right\}; and (ii) T(ΩTW)−1/2ΣTW(θ^TW−θ∗)→dN(0,Ide)\sqrt{T}\left(\Omega^{W}_{T}\right)^{-1/2}\Sigma^{W}_{T}\left(\widehat{\theta}^{W}_{T}-\theta^{*}\right)\xrightarrow{d}\mathcal{N}(0,I_{d_{e}}).

Theorem 4 shows that the estimation errors of the components of the quantile and weighting function do not affect the asymptotic distribution of θ^TW\widehat{\theta}^{W}_{T}. The second part indicates that the variance-covariance matrix of θ^TW\widehat{\theta}^{W}_{T} differs from that of Theorem 2 due to the existence of the weighting functions. A properly constructed weighting function can improve the asymptotic efficiency of θ^TW\widehat{\theta}^{W}_{T}, leading to narrower confidence intervals for statistical inference in θ∗\theta^{*}. This result is summarized in the following corollary.

Corollary 4.1. Suppose that {qt(β∗)−et(θ∗)}2∝var⁡{S0(β∗,θ∗;Yt,Xt)∣Ht}\{q_{t}(\beta^{*})-e_{t}(\theta^{*})\}^{2}\propto\operatorname{var}\{S_{0}(\beta^{*},\theta^{*};Y_{t},X_{t})\mid\mathcal{H}_{t}\} and Theorem 4 hold. Then, the estimator θ^TW\widehat{\theta}^{W}_{T} is asymptotically efficient among the class of joint MM-estimators introduced in Fissler & Ziegel [17].

In the following theorem, we show that our proposed variance-covariance matrix estimator given in § 3.1 is consistent.

Theorem 5. Under Assumptions S1S1–S3S3 and the event E1T∩E2T∩E3T\mathcal{E}_{1T} \cap\mathcal{E}_{2T} \cap\mathcal{E}_{3T}, the estimator of the variance-covariance matrix in Theorem 4 is weakly consistent:

(Σ^TW)−1Ω^TW(Σ^TW)−1⟶p(ΣTW)−1ΩTW(ΣTW)−1.\left(\widehat{\Sigma}^{W}_{T}\right)^{-1}\widehat{\Omega}^{W}_{T}\left(\widehat{\Sigma}^{W}_{T}\right)^{-1}\overset{p}{\longrightarrow}\left(\Sigma^{W}_{T}\right)^{-1}\Omega^{W}_{T}\left(\Sigma^{W}_{T}\right)^{-1}.

The above theorem validates the inference procedure given in (10). Note that Theorem 5 also provides theoretical guarantees for the confidence intervals given in Theorem 2 by setting the weighting function equal to one for every tt, hence the variance-covariance matrix of θ^TW\widehat{\theta}^{W}_{T} will reduce to that of θ^T\widehat{\theta}_{T}.

Validity of Portmanteau Test

In this section, we provide theoretical guarantees for the Portmanteau test proposed in § 3.2 for model diagnostics. Specifically, we characterize the asymptotic behavior of the test statistic ψ^\hat{\psi} introduced in Algorithm 2, under the null hypothesis H0H_{0} in (12). The following two theorems establish the validity of the proposed Portmanteau test.

Theorem 6. Suppose {Yt}t≥1\{Y_{t}\}_{t\geq1} is generated by the location-scale family model (7) and let {ζ^t}t≥1=[{Yt−qt(β^T)}/{qt(β^T)−et(θ^T)}]t≥1\{\hat{\zeta}_{t}\}_{t\geq1} = \left[\left\{Y_{t}-q_{t}(\widehat{\beta}_{T})\right\}/\left\{q_{t}(\widehat{\beta}_{T})-e_{t}(\widehat{\theta}_{T})\right\}\right]_{t\geq1} where β^T\widehat{\beta}_{T} and θ^T\widehat{\theta}_{T} are obtained by solving (4) and (5), respectively. For a fixed constant MM, let r~=(r~1,r~2,…,r~M)T\widetilde{\mathbf{r}}=(\widetilde{r}_{1},\widetilde{r}_{2},\ldots,\widetilde{r}_{M})^{\mathsf{T}} be the vector of sample autocorrelation coefficients of {ζ^t2}t≥1\{\hat{\zeta}_{t}^{2}\}_{t\geq1} up to lag level MM. Then for the matrix Ωtest\Omega^{\mathrm{test}} defined in (S5.13) in the online supplementary material, under Assumptions S1S1–S3S3, and the event E1T∩E2T\mathcal{E}_{1T}\cap\mathcal{E}_{2T}, we have

Tr~T(Ωtest)−1r~→dχM2.T\widetilde{\mathbf{r}}^{\mathsf{T}}\left(\Omega^{\mathrm{test}}\right)^{-1}\widetilde{\mathbf{r}}\xrightarrow{d}\chi^{2}_{M}.

Theorem 7. Let Ω^test\widehat{\Omega}^{\mathrm{test}} be the estimator given in Algorithm S2 in the online supplementary material. Under Assumptions S1S1–S3S3 and the event E1T∩E2T\mathcal{E}_{1T}\cap\mathcal{E}_{2T}, we have

Ω^test→pΩtest.\widehat{\Omega}^{\mathrm{test}}\xrightarrow{p}\Omega^{\mathrm{test}}.

Numerical Studies

We conduct extensive numerical experiments to assess the finite-sample performance of our proposed expected shortfall estimator θ^TW\hat{\theta}^{W}_{T} obtained from (9). We also compare to a class of estimators obtained from minimizing the different joint loss functions given in Patton et al. [29] and Dimitriadis et al. [14]. § 5.1 introduces the data generation processes. § 5.2 presents the results for the different estimators and illustrates the efficiency gain of the proposed weighted expected shortfall estimator relative to existing approaches. We conduct numerical studies to assess the performance of the proposed Portmanteau test for model diagnostics in § 5.3.

Data Generating Process and Methodology Comparison

We start with two data-generating processes for numerical studies: (i) an asymmetric slope model; and (ii) a non-location-scale family symmetric absolute value model. The above models correspond to scenarios in which the weighting function in (9) is optimal and suboptimal, respectively. The specifications of the two models are:

  1. The first data generating process combines the quasi-score driven model in Blasques et al. [6] and the asymmetric slope model given in Engle & Manganelli [16]. Recall from § 2 that Yt+=Yt1{Yt≥0}Y^{+}_{t}=Y_{t}\mathbb{1}\{Y_{t}\ge0\} and Yt−=Yt1{Yt<0}Y^{-}_{t}=Y_{t}\mathbb{1}\{Y_{t}<0\}. Let ft(α∗)=α0∗+α1∗ft−1(α∗)+α2∗Yt−1+−α3∗Yt−1−f_{t}(\alpha^{*})=\alpha^{*}_{0}+\alpha^{*}_{1}f_{t-1}(\alpha^{*})+\alpha^{*}_{2}Y^{+}_{t-1}-\alpha^{*}_{3}Y^{-}_{t-1} and

Yt=ft(α∗)+(α4∗+α5∗Yt−1+−α6∗Yt−1−)ηt,(15)Y_{t}=f_{t}(\alpha^{*})+(\alpha^{*}_{4}+\alpha^{*}_{5}Y^{+}_{t-1}-\alpha^{*}_{6}Y^{-}_{t-1})\eta_{t}, \tag*{(15)}

where α∗=(α0∗,α1∗,…,α6∗)T\alpha^{*}=(\alpha^{*}_{0},\alpha^{*}_{1},\ldots,\alpha^{*}_{6})^{\mathsf{T}} and {ηt}t≥1\{\eta_{t}\}_{t\ge1} are i.i.d.i.i.d. standard normal. It can be shown that the conditional quantile and the expected shortfall functions of YtY_{t} follow the asymmetric slope model given in Example 2. Especially, for the expected shortfall component, we have et(θ∗)=θ0∗+θ1∗et−1(θ∗)+θ2∗Yt−1+−θ3∗Yt−1−+θ4∗Yt−2+−θ5∗Yt−2−e_t(\boldsymbol{\theta}^{*}) = \theta_0^{*} + \theta_1^{*}e_{t-1}(\boldsymbol{\theta}^{*}) + \theta_2^{*}Y_{t-1}^{+} - \theta_3^{*}Y_{t-1}^{-} + \theta_4^{*}Y_{t-2}^{+} - \theta_5^{*}Y_{t-2}^{-}, where θ∗={α0∗+(1−α1∗)α4∗ES⁡τ(ηt),α1∗,α2∗+α5∗ES⁡τ(ηt),α3∗+α6∗ES⁡τ(ηt),−α1∗α5∗ES⁡τ(ηt),−α1∗α6∗ES⁡τ(ηt)}T\boldsymbol{\theta}^{*} = \{\alpha_0^{*} + (1-\alpha_1^{*})\alpha_4^{*}\operatorname{ES}_{\tau}(\eta_t), \alpha_1^{*}, \alpha_2^{*} + \alpha_5^{*}\operatorname{ES}_{\tau}(\eta_t), \alpha_3^{*} + \alpha_6^{*}\operatorname{ES}_{\tau}(\eta_t), -\alpha_1^{*}\alpha_5^{*}\operatorname{ES}_{\tau}(\eta_t), -\alpha_1^{*}\alpha_6^{*}\operatorname{ES}_{\tau}(\eta_t)\}^{\mathsf{T}}.

(ii) The second data generating process is based on the symmetric absolute value model in Example 1. However, {Yt}t≥1\{Y_t\}_{t\geq1} generated by the following process does not belong to the location-scale family model introduced in (7). Let

Yt=qt{β(Ut)}=β0(Ut)+β1∗qt−1{β(Ut)}+β2(Ut)∣Yt−1∣,(16)Y_t = q_t\{\boldsymbol{\beta}(U_t)\} = \beta_0(U_t) + \beta_1^{*}q_{t-1}\{\boldsymbol{\beta}(U_t)\} + \beta_2(U_t)|Y_{t-1}|, \tag*{(16)}

where {Ut}t≥1\{U_t\}_{t\geq1} is a sequence of i.i.d. Uniform(0,1)(0,1) random variables. The vector β(u)={β0(u),β1∗,β2(u)}T\boldsymbol{\beta}(u) = \{\beta_0(u), \beta_1^{*}, \beta_2(u)\}^{\mathsf{T}} satisfies that given the history Ht\mathcal{H}_t, qt{β(u)}q_t\{\boldsymbol{\beta}(u)\} is strictly increasing with respect to u∈(0,1)u \in(0,1). For a fixed quantile level τ\tau, we have et(θ∗)=θ0∗+θ1∗et−1(θ∗)+θ2∗∣Yt−1∣e_t(\boldsymbol{\theta}^{*}) = \theta_0^{*} + \theta_1^{*}e_{t-1}(\boldsymbol{\theta}^{*}) + \theta_2^{*}|Y_{t-1}| with θ∗={∫0τβ0(u)du/τ,β1∗,∫0τβ2(u)du/τ}T\boldsymbol{\theta}^{*} = \{\int_0^\tau\beta_0(u)du/\tau, \beta_1^{*}, \int_0^\tau\beta_2(u)du/\tau\}^{\mathsf{T}}.

In this section, we set the coefficient functions to β0(u)=3FBeta−1(u)−1.5\beta_0(u) = 3F_{\mathrm{Beta}}^{-1}(u) - 1.5, β1∗=0.3\beta_1^{*} = 0.3, and β2(u)=FUnif−1(u)\beta_2(u) = F_{\mathrm{Unif}}^{-1}(u), where FBeta−1(⋅)F_{\mathrm{Beta}}^{-1}(\cdot) and FUnif−1(⋅)F_{\mathrm{Unif}}^{-1}(\cdot) denote the quantile functions of a Beta(2,2)(2, 2) distribution and a Uniform(−0.8,0.8)(-0.8, 0.8) distribution, respectively.

In § 5.2, we compare the finite-sample performance of our proposed two-step expected shortfall estimator (9) with two competing joint estimators introduced in Patton et al. [29] and Dimitriadis et al. [14], referred to as the “FZ0” and “Emp” estimators, respectively. Detailed descriptions and implementation procedures for these competing methods are provided in Section S2 of the online supplementary material.

Parameter Estimation and Inference

In this section, we evaluate the finite-sample performance of the estimators described in § 5.1. Specifically, for each estimator, we compute the element-wise bias and the length of its corresponding 95% confidence interval, based on Theorem 4 and the corresponding theoretical results in Patton et al. [29] and Dimitriadis et al. [14].

To improve the finite sample performance of our proposed two-step expected shortfall estimators, we consider using random weighting bootstrap to construct confidence intervals Zhu et al. [41]. We defer the implementation details of this bootstrap procedure to Section S2.2 of the online supplementary material. The average bias and confidence interval width, together with the corresponding standard error of the confidence interval width, summarized over 1000 replications, are reported in Table 1–Table 2 at the quantile level τ=0.1\tau= 0.1. Moreover, to mitigate the influence of outlier confidence interval lengths arising from numerical instabilities in the optimization routine, we compute the average confidence interval length after applying 95% winsorization.

T=5000T = 5000T=7500T = 7500
θ1\theta_1θ2\theta_2θ3\theta_3θ4\theta_4θ5\theta_5θ1\theta_1θ2\theta_2θ3\theta_3θ4\theta_4θ5\theta_5
τ=0.1\tau= 0.1
Avg Bias (WLS)-0.010.010.01-0.02-0.02-0.010.010.01-0.01-0.01
Avg Bias (FZ0)-0.11-0.01-0.01-0.10-0.10-0.07-0.01-0.00-0.06-0.06
Avg Bias (Emp)-0.950.120.04-0.56-0.55-0.690.060.01-0.47-0.45
CI Length (WLS)0.380.190.250.410.500.290.150.210.320.40
CI Length (FZ0)0.390.220.310.440.550.310.180.250.350.44
CI Length (Emp)0.550.250.370.540.630.430.170.250.410.48
(100×)(100\times) Std Err (WLS)0.550.080.120.460.520.370.030.060.320.36
(100×)(100\times) Std Err (FZ0)0.400.050.090.310.330.230.030.050.180.19
(100×)(100\times) Std Err (Emp)1.370.250.490.951.051.010.060.170.720.74
Coverage (WLS)0.930.930.930.940.930.930.940.930.940.95
Coverage (FZ0)0.840.920.920.850.860.870.930.930.880.89
Coverage (Emp)0.260.610.670.400.420.450.730.730.500.50

Table 1. Simulation results for the asymmetric slope model (15) with α∗=(0.45,0.65,−0.25,−0.35,0.8,0.35,0.25)T\boldsymbol{\alpha}^{*} = (0.45, 0.65, -0.25, -0.35, 0.8, 0.35, 0.25)^{\mathsf{T}} are reported above. These include the average element-wise bias, the mean and empirical standard errors of the confidence interval lengths, and the coverage probabilities for expected shortfall regression estimators from three different methods: θ^W\hat{\boldsymbol{\theta}}_{\mathit{W}} from (9) (WLS), the FZ0 minimization approach proposed in Patton et al. [29] (FZ0), the pseudo-efficient joint loss approach proposed in Dimitriadis et al. [14] with plug-in estimated nuisance components (Emp). The mean and standard errors of confidence interval lengths are computed after 95% winsorization.

T=2500T = 2500T=5000T = 5000T=7500T = 7500
θ1\theta_1θ2\theta_2θ1\theta_1θ2\theta_2θ1\theta_1θ2\theta_2
τ=0.1\tau= 0.1
Avg Bias (WLS)-0.000.00-0.000.00-0.000.00
Avg Bias (FZ0)0.00-0.00-0.00-0.000.00-0.00
Avg Bias (Emp)0.00-0.00-0.00-0.000.00-0.00
Avg Bias (Optimal)0.00-0.00-0.00-0.000.00-0.00
CI Length (WLS)0.120.120.090.090.070.07
CI Length (FZ0)0.130.130.090.090.070.07
CI Length (Emp)0.120.120.090.090.070.07
CI Length (Optimal)0.130.130.090.090.070.07
(100×) Std Err (WLS)0.050.030.020.010.020.01
(100×) Std Err (FZ0)0.040.030.020.010.010.01
(100×) Std Err (Emp)0.050.030.020.010.010.01
(100×) Std Err (Optimal)0.050.030.020.010.010.01
Coverage (WLS)0.930.930.940.950.950.95
Coverage (FZ0)0.930.940.950.950.950.95
Coverage (Emp)0.930.940.940.950.950.95
Coverage (Optimal)0.930.940.960.970.940.95

Table 2. Simulation results for model (16). Other details are as in Table 1.

From Table 1, we find that under the data-generating process in (15), the average bias for our proposed estimators θ^TW\hat{\theta}^{W}_{T} approaches zero as we increase the sample size. The empirical coverage probability for our proposed WLS estimator is close to the nominal 95% level as the sample size increases, supporting the validity of the asymptotic properties derived in Theorem 4. For the two joint estimators FZ0 and Emp proposed in Patton et al. [29] and Dimitriadis et al. [14], respectively, while they are asymptotically normal, in the complex model setting (15), the joint estimation approach requires solving a more challenging nonconvex optimization problem and estimating more nuisance quantities (see details in Section S2 of the online supplementary material), which introduce additional numerical instabilities. As a result, their finite-sample performance is worse than that of the two-step estimation approach. Finally, we observe from Table 1 that the proposed weighted two-step estimator has the highest efficiency among all estimators, as it has the smallest 95% confidence interval length. This is due to the use of the optimal weighting function in Corollary 4.1.

Under model (16), the weighting function 1/{qt(β∗)−et(θ∗)}21/\{q_t(\boldsymbol{\beta}^{*})-e_t(\boldsymbol{\theta}^{*})\}^{2} is suboptimal due to the fact that model (16) does not belong to the location-scale family. However, motivated by the construction of the optimal weighting function proposed in Barendse [2], the conditional variance function of the regression problem (5) can be explicitly computed from the data generating process (16), such that

var⁡{S0(Yt,Xt;β∗,θ∗)∣Ht}=∫0τ[qt{β(u)}−qt(β∗)]2 du/τ−{qt(β∗)−et(θ∗)}2.\operatorname{var}\{S_0(Y_t,\boldsymbol{X}_t;\boldsymbol{\beta}^{*},\boldsymbol{\theta}^{*}) \mid\mathcal{H}_t\} = \int_0^\tau[q_t\{\boldsymbol{\beta}(u)\}-q_t(\boldsymbol{\beta}^{*})]^2 \,du/\tau- \{q_t(\boldsymbol{\beta}^{*})-e_t(\boldsymbol{\theta}^{*})\}^2.

This quantity can be numerically approximated by solving the quantile regression problem (4) at multiple quantile levels. Consequently, we consider an alternative estimator for this model for comparison, obtained by solving the weighted least squares regression problem in (9) with the estimated optimal weighting function. We refer to this estimator as the “Optimal” estimator and defer the implementation steps to estimate the optimal weighting function to Section S8 of the online supplementary material. Our numerical results in Table 2 suggest that our proposed weighting function, although suboptimal in the general setting, can have competitive empirical performance relative to commonly used joint estimation approaches and the estimator obtained by optimal weights.

Model Misspecification Test

We now assess the empirical performance of the Portmanteau test proposed in § 3.2. To this end, we consider a generalized autoregressive conditional heteroskedasticity model motivated by Example 3 in § 2, which is also considered in Zhu et al. [41]. We start by introducing the true data generating process under the null hypothesis in (12). Recall from Example 3 that

Yt=σt(α∗)ηt,σt2(α∗)=α0∗+α1∗σt−12(α∗)+α2∗Yt−12,(17)Y_t = \sigma_t(\alpha^*)\eta_t,\quad\sigma_t^2(\alpha^*) = \alpha_0^* + \alpha_1^*\sigma_{t-1}^2(\alpha^*) + \alpha_2^*Y_{t-1}^2, \tag*{(17)}

where {ηt}t≥1\{\eta_t\}_{t\ge1} are i.i.d. standard normal and α∗=(0.1,0.1,0.1)T\alpha^* = (0.1, 0.1, 0.1)^{\mathsf{T}}. Under (17), the true conditional quantile and expected shortfall functions are

qt(β∗)=−β0∗+β1∗qt−12(β∗)+β2∗Yt−12,et(θ∗)=−θ0∗+θ1∗et−12(θ∗)+θ2∗Yt−12,(18)q_t(\beta^*) = -\sqrt{\beta_0^* + \beta_1^*q_{t-1}^2(\beta^*) + \beta_2^*Y_{t-1}^2},\quad e_t(\theta^*) = -\sqrt{\theta_0^* + \theta_1^*e_{t-1}^2(\theta^*) + \theta_2^*Y_{t-1}^2}, \tag*{(18)}

where β∗={α0∗Qτ2(η),α1∗,α2∗Qτ2(η)}T\beta^* = \{\alpha_0^*Q_{\tau}^2(\eta),\alpha_1^*,\alpha_2^*Q_{\tau}^2(\eta)\}^{\mathsf{T}} and θ∗={α0∗ES⁡τ2(η),α1∗,α2∗ES⁡τ2(η)}T\theta^* = \{\alpha_0^*\operatorname{ES}_{\tau}^2(\eta),\alpha_1^*,\alpha_2^*\operatorname{ES}_{\tau}^2(\eta)\}^{\mathsf{T}}. Here, we use (18) as our model under the null hypothesis in (12), which corresponds to the scenario in which the conditional quantile and expected shortfall functions are correctly specified. On the other hand, under the alternative hypothesis, we consider the true data generating process to be

Yt=σt(α∗)ηt,σt2(α∗)=α0∗+α1∗σt−12(α∗)+α2∗Yt−12+cYt−22.(19)Y_t = \sigma_t(\alpha^*)\eta_t,\quad\sigma_t^2(\alpha^*) = \alpha_0^* + \alpha_1^*\sigma_{t-1}^2(\alpha^*) + \alpha_2^*Y_{t-1}^2 + cY_{t-2}^2. \tag*{(19)}

We vary the scalar c∈[0,0.6]c \in[0, 0.6]. When c=0c = 0, model (19) reduces to (17), and the null model is correctly specified. For c>0c > 0, however, (19) deviates from the null data generating process (17), leading to a misspecification of the null model (18).

We implement the Portmanteau test described in § 3.2 to evaluate whether the null model given in (17) is correctly specified, with M∈{4,6,8}M \in\{4, 6, 8\} in (12). We start by assuming that the null hypothesis is true and generate T={2500,5000,7500}T = \{2500, 5000, 7500\} observations using (17). The empirical type I error rate is calculated as the rejection frequency across 1000 replications of the numerical studies. The results are summarized in Table 3. We observe that under the null hypothesis, the Portmanteau test controls the Type I error rate around the nominal level of 5% across different choices of MM.

T=2500T = 2500T=5000T = 5000T=7500T = 7500
M=4M = 40.0560.0410.043
M=6M = 60.0540.0430.040
M=8M = 80.0510.0480.051

Table 3. Empirical Type I error rates for different choices of MM and TT under (17).

Next, we evaluate the statistical power of the Portmanteau test. For comparison, we consider another model diagnostics approach proposed by Barendse et al. [3]. Here, we apply the matrix T(4)T^{(4)} given in Barendse et al. [3] as the test matrix and follow their proposed approach to model testing. Under the alternative hypothesis, we set T=5000T = 5000 for our proposed Portmanteau test. For the method given in Barendse et al. [3], T=5000T = 5000 samples are used for in-sample estimation and 50T50\sqrt{T} samples are used for out-of-sample backtesting. The samples are generated using the alternative model (19), by varying the scalar c∈{0,0.1,0.2,0.3,0.4,0.5,0.6}c \in\{0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6\}. The empirical powers of the Portmanteau test and the test proposed in Barendse et al. [3] are calculated as the respective null hypothesis rejection frequencies in (12) and (4) in Barendse et al. [3], respectively. Our empirical power curves are shown in Figure 1. As cc increases, the data-generating process (19) increasingly diverges from the null model (17), causing the true Qτ(Yt∣Ht)Q_{\tau}(Y_t \mid\mathcal{H}_t) and ES⁡τ(Yt∣Ht)\operatorname{ES}_{\tau}(Y_t \mid\mathcal{H}_t) to deviate significantly from the null model (18). This growing distortion leads to higher rejection frequencies. Furthermore, we observe that the power curve for the Portmanteau test consistently dominates the approach in Barendse et al. [3], indicating that our proposed test is empirically more powerful. The overall patterns of the empirical power curves of the Portmanteau test are highly consistent with different choices of MM. Moreover, we include additional numerical experiments on a broader range of model-misspecification settings in Section S9 of the online supplementary material.

Figure 1. The empirical power curve for model misspecification tests with T=5000T = 5000, τ=0.1\tau= 0.1 and M∈{4,6,8}M \in\{4, 6, 8\}.

Financial Risk Inference and Prediction

We implement the proposed method to investigate the dynamic behavior of financial risk during periods of global financial crisis. To this end, we analyze stock index data, specifically the Hang Seng Index and the Dow Jones Industrial Average, from January 2, 2000, to December 31, 2020. Here, our response variable YtY_t is the daily returns of stock indices, which is calculated as 100 times the log differences of daily closing prices in US dollars. The covariates Xt\mathbf{X}_t are set to the lagged values of YtY_t.

We consider two different scenarios: the global financial crisis triggered by the coronavirus pandemic in early 2020, and the real estate-driven financial crisis in 2008. For the first scenario, we use the data from January 2, 2000 to December 31, 2019 as the training set for the Hang Seng Index and the Dow Jones Industrial Average, respectively, to estimate the model parameters. We then treat the remaining observations from January 2, 2020, to June 30, 2020 as the test set for performing one-day-ahead predictions of the expected shortfall. For the second scenario, the model is trained using data from January 3, 2000, to May 30, 2008, and tested on data from June 2, 2008, to December 31, 2008. § 6.1 reports the results for in-sample parameter estimation based on the full in-sample dataset from January 2000 to December 2019. § 6.2 gives the out-of-sample forecasting results.

Parameter Estimation and Model Diagnostics

We begin by determining suitable functional forms for qt(β)q_t(\beta) and et(θ)e_t(\theta) in the model (7). We assume that qt(β)q_t(\beta) and et(θ)e_t(\theta) share the same functional form, and consider five candidate models to describe the joint autoregressive dynamics patterns of quantile and expected shortfall. The models are listed as follows:

  • (i) Lag One Symmetric Absolute Value (SAV1): et(θ)=θ0+θ1et−1(θ)+θ2∣Yt−1∣e_t(\theta) = \theta_0 + \theta_1 e_{t-1}(\theta) + \theta_2 \lvert Y_{t-1} \rvert.

  • (ii) Lag One Asymmetric Slope (AS1): et(θ)=θ0+θ1et−1(θ)+θ2Yt−1+−θ3Yt−1−e_t(\theta) = \theta_0 + \theta_1 e_{t-1}(\theta) + \theta_2 Y_{t-1}^{+} - \theta_3 Y_{t-1}^{-}.

  • (iii) GARCH: et(θ)=−θ0+θ1et−12(θ)+θ2Yt−12e_t(\theta) = -\sqrt{\theta_0 + \theta_1 e_{t-1}^{2}(\theta) + \theta_2 Y_{t-1}^{2}}.

  • (iv) SAV2: et(θ)=θ0+θ1et−1(θ)+θ2∣Yt−1∣+θ3∣Yt−2∣e_t(\theta) = \theta_0 + \theta_1 e_{t-1}(\theta) + \theta_2 \lvert Y_{t-1} \rvert+ \theta_3 \lvert Y_{t-2} \rvert.

  • (v) AS2: et(θ)=θ0+θ1et−1(θ)+θ2Yt−1+−θ3Yt−1−+θ4Yt−2+−θ5Yt−2−e_t(\theta) = \theta_0 + \theta_1 e_{t-1}(\theta) + \theta_2 Y_{t-1}^{+} - \theta_3 Y_{t-1}^{-} + \theta_4 Y_{t-2}^{+} - \theta_5 Y_{t-2}^{-}.

We perform the Portmanteau test proposed in § 3.2 for each of the five candidate models for model diagnostics. We set the lag level M=10M = 10 for the Portmanteau test, which corresponds to the time duration of two trading weeks. The test results indicate that the lag one and two symmetric absolute value models are suitable choices for the Hang Seng Index data, with pp-values 0.28 and 0.61, respectively. On the other hand, the lag two symmetric and asymmetric slope models serve as more appropriate models to fit Dow Jones Industrial Average data, with pp-values of 0.06 for both models. Based on the above results, we apply the weighted expected shortfall estimator proposed in (9) to estimate the parameter θ\theta for the Hang Seng Index and Dow Jones Industrial Average data, respectively. The results for τ=0.05\tau= 0.05 are summarized in Table 4.

θj\theta_jHang Seng IndexDow Jones Industrial Average
SAV1SAV2SAV2AS2
Est ValueStd ErrEst ValueStd ErrEst ValueStd ErrEst ValueStd Err
θ0\theta_0-0.090.02-0.090.03-0.120.04-0.100.03
θ1\theta_10.900.020.900.020.830.040.870.04
θ2\theta_2-0.200.05-0.160.09-0.250.13-0.060.17
θ3\theta_3——-0.040.09-0.100.14-0.440.19
θ4\theta_4——————-0.010.12
θ5\theta_5——————0.010.18

Table 4. Estimated values and standard errors for θj\theta_j in et(θ)e_t(\boldsymbol{\theta}) across different models and indices at τ=0.05\tau= 0.05. Bold entries indicate significance at the 5% confidence level.

From Table 4, we see that the estimated values for θ1\theta_1, the coefficient of the autoregressive term, are always significant across different models and stock indices. Our results align with similar conclusions in the context of autoregressive conditional quantile regression Engle & Manganelli [16], thus supporting the phenomenon of clustering patterns in conditional expected shortfall. Next, we find that the expected shortfall dynamic patterns are different between the Hang Seng Index and the Dow Jones Industrial Average. The symmetric absolute value models fit the Hang Seng Index data well, indicating that the previous positive return and negative return have similar influences on future risk. On the other hand, we find that the asymmetric slope model quantifies the differential impact of previous positive and negative returns for the Dow Jones Industrial Average, as θ3\theta_3, the coefficient for −Yt−1−-Y_{t-1}^{-}, is significantly different from θ2\theta_2, the coefficient of Yt−1+Y_{t-1}^{+}. This indicates a high sensitivity of the financial market to the negative return of the previous day. This asymmetric influence between previous positive and negative returns suggests strong evidence of the leverage effect, which stems from the empirical fact that losses have a greater influence on future risks than gains [10, 39].

In practice, when prior knowledge of the model structure is not available, we recommend that users apply the workflow described in this section for analyzing financial time series data. Specifically, we first implement the Portmanteau test on the time series data over a finite set of candidate joint quantile-expected shortfall models. We then select the models that pass the Portmanteau test as plausible working models and use them to fit the data.

Out-of-Sample Expected Shortfall Prediction

In this section, we evaluate the predictive performance of our proposed dynamic autoregressive expected shortfall model during periods of significant financial turbulence.

We focus on two distinct testing periods, as detailed in the beginning of § 6.

Following the training-testing split described in § 6 and the results given in § 6.1, we select the lag two symmetric absolute value model and lag two asymmetric slope model to make expected shortfall predictions for the Hang Seng index and Dow Jones Industrial Average data, respectively. In addition to these two models, we also consider a rolling window approach for predicting expected shortfall.

Specifically, for a given time point tt and a predetermined window size hh, the conditional τ\tau-level expected shortfall of YtY_t is estimated as the empirical marginal τ\tau-level expected shortfall of the series {Yt−h,Yt−h+1,…,Yt−1}\{Y_{t-h}, Y_{t-h+1}, \ldots, Y_{t-1}\}. In this analysis, we set the window size hh to be 125 for simplicity.

The forecast results are presented in Figure 2–Figure 3. Our initial observation is that the rolling window-based approach provides only a rough prediction and cannot reveal the complex dynamic pattern of financial risk. Conversely, the predictions made by the lag two asymmetric slope and symmetric absolute value models can all reveal more complex dynamic patterns of the expected shortfall throughout the year than the rolling window approach. Both models exhibit very similar trends in predicting one-day-ahead conditional expected shortfall. The prediction results for 2008, as shown in Figure 2, indicate a significant decline in the expected shortfall estimates for both the Hang Seng Index and the Dow Jones Industrial Average following September 29, 2008. This date is particularly important for the Dow Jones Industrial Average, as it recorded the largest single-day point drop in its history. In 2020, the global stock market experienced three major events due to the pandemic: Black Monday I on March 9, Black Thursday on March 12, and Black Monday II on March 16. As illustrated in Figure 3, the expected shortfall predictions for both models and both stock indices show a similar downward trend during this period.

Figure 2. The 5% level expected shortfall forecasting results for different approaches: rolling window (orange curve), lag two symmetric absolute value and asymmetric slope models (blue curves). The upper panel includes the plots for the observed Hang Seng Index and Dow Jones Industrial Average return series in the year 2008. The lower panel plots the one-step expected shortfall predicted values for the aforementioned three models. The dashed vertical black line denotes the date of the largest single-day point drop in the history of the Dow Jones Industrial Average, which occurred on September 29, 2008.

Figure 3. The 5% level expected shortfall forecasting results for different approaches: rolling window (orange curve), lag two symmetric absolute value and asymmetric slope models (blue curves). The upper panel contains plots for the observed Hang Seng Index and Dow Jones Industrial Average return series in the year 2020. The lower panel plots the one-step expected shortfall predicted values for the aforementioned three models. The dashed vertical black lines denote the “Black Monday I”, “Black Thursday” and “Black Monday II” stock market crashes on March 9, 2020, March 12, 2020 and March 16, 2020.

In summary, both the lag two symmetric absolute value and the asymmetric slope models effectively predict the risk levels of various stock indices. Moreover, as discussed in § 6.1, the lag-two asymmetric slope model for Dow Jones data offers the advantage of distinguishing between the differential impacts of past gains and losses on stock returns. Due to its greater flexibility, we recommend adopting the lag-two asymmetric slope model as the preferred working model in this example.

Discussion

We consider a joint model for conditional quantiles and expected shortfall with autoregressive structures. Existing estimation approaches for such models typically involve minimizing a class of joint loss functions [17, 29]. Motivated by Barendse [2], we propose a weighted two-step estimation scheme to estimate the expected shortfall parameters. We start by using an unweighted least squares approach to estimate the expected shortfall parameters under the model (2). We apply weighted least squares regression to improve the asymptotic efficiency of our estimators. The choice of our weighting function is motivated by the location-scale family time series with i.i.d.i.i.d. innovations. Under such a model assumption, our proposed weighting functions are optimal and can achieve asymptotic efficiency among the class of joint estimators proposed in Dimitriadis et al. [14]. Moreover, we develop a Portmanteau test based on (7) with i.i.d. noise assumption, which exhibits great empirical efficiency. Both theoretical analysis and numerical simulations demonstrate that our proposed estimator is more efficient than those of Patton et al. [29].

Although the location-scale family time series models with i.i.d. innovations encompass a wide range of commonly used time series models, they remain a specific subclass of the more general model (2). In some scenarios, our proposed weighted estimator is suboptimal. A potential direction for future research is the development of nonparametric approaches to derive optimal weights without relying on the assumption in (7). Moreover, it is well known that the least squares step involved in the two-step approach is highly sensitive to heavy-tailed and highly skewed data, which are common in financial asset returns. Incorporating a robust loss function into the two-step approach may address this issue and yield a robust estimator (He et al. (2023)). We will leave the aforementioned discussion for future research.

References

  1. [1]Barendse, S. (2024), ‘Expected shortfall LASSO’, arXiv preprint arXiv:2307.01033.arxiv.org/abs/2307.01033
  2. [2]Barendse, S. (2026), ‘Efficiently weighted estimation of tail and interquantile expectations’, Journal of Financial Econometrics 24(2), nbag003.DOI
  3. [3]Barendse, S., Kole, E. & van Dijk, D. (2023), ‘Backtesting Value-at-Risk and expected shortfall in the presence of estimation error’, Journal of Financial Econometrics 21(2), 528–568.
  4. [4]Basel Committee (2016), ‘Minimum capital requirements for market risk’, Consultative Document.
  5. [5]Black, F. (1976), Studies of stock price volatility changes, in ‘Proceedings of the 1976 Meetings of the American Statistical Association’, pp. 171–181.
  6. [6]Blasques, F., Francq, C. & Laurent, S. (2023), ‘Quasi score-driven models’, Journal of Econometrics 234(1), 251–275.DOI
  7. [7]Bollerslev, T. (1986), ‘Generalized autoregressive conditional heteroskedasticity’, Journal of Econometrics 31(3), 307–327.
  8. [8]Bollerslev, T., Chou, R. Y. & Kroner, K. F. (1992), ‘ARCH modeling in finance: A review of the theory and empirical evidence’, Journal of Econometrics 52(1-2), 5–59.DOI
  9. [9]Boonen, T. J. (2017), ‘Solvency II solvency capital requirement for life insurance companies based on expected shortfall’, European Actuarial Journal 7(2), 405–434.
  10. [10]Christie, A. A. (1982), ‘The stochastic behavior of common stock variances: Value, leverage and interest rate effects’, Journal of Financial Economics 10(4), 407–432.
  11. [11]Creal, D., Koopman, S. J. & Lucas, A. (2013), ‘Generalized autoregressive score models with applications’, Journal of Applied Econometrics 28(5), 777–795.
  12. [12]De Gooijer, J. G. & Hyndman, R. J. (2006), ‘25 years of time series forecasting’, International Journal of Forecasting 22(3), 443–473.
  13. [13]Dimitriadis, T. & Bayer, S. (2019), ‘A joint quantile and expected shortfall regression framework’, Electronic Journal of Statistics 13(1), 1823–1871.
  14. [14]Dimitriadis, T., Fissler, T. & Ziegel, J. (2020), ‘The efficiency gap’, arXiv preprint arXiv:2010.144146.
  15. [16]Engle, R. F. & Manganelli, S. (2004), ‘CAViaR: Conditional autoregressive value at risk by regression quantiles’, Journal of Business & Economic Statistics 22(4), 367–381.
  16. [17]Fissler, T. & Ziegel, J. F. (2016), ‘Higher order elicitability and Osband’s principle’, The Annals of Statistics 44(4), 1680 – 1707.
  17. [18]Fletcher, R. & Reeves, C. M. (1964), ‘Function minimization by conjugate gradients’, The Computer Journal 7(2), 149–154.DOI
  18. [19]Gerlach, R. & Wang, C. (2020), ‘Semi-parametric dynamic asymmetric Laplace models for tail risk forecasting, incorporating realized measures’, International Journal of Forecasting 36(2), 489–506.arxiv.org/abs/1805.08653
  19. [20]Gneiting, T. (2011), ‘Making and evaluating point forecasts’, Journal of the American Statistical Association 106(494), 746–762.arxiv.org/abs/0912.0902
  20. [22]Lazar, E. & Zhang, N. (2019), ‘Model risk of expected shortfall’, Journal of Banking & Finance 105, 74–93.
  21. [23]Li, G. & Li, W. K. (2005), ‘Diagnostic checking for time series models with conditional heteroscedasticity estimated by the least absolute deviation approach’, Biometrika 92(3), 691–701.DOI
  22. [26]Lumsdaine, R. L. (1996), ‘Consistency and asymptotic normality of the quasi-maximum likelihood estimator in IGARCH(1,1) and covariance stationary GARCH(1,1) models’, Econometrica 64(3), 575–596.DOI
  23. [27]Nelder, J. A. & Mead, R. (1965), ‘A simplex method for function minimization’, The Computer Journal 7(4), 308–313.
  24. [28]Nolde, N. & Ziegel, J. F. (2017), ‘Elicitability and backtesting: Perspectives for banking regulation’, The Annals of Applied Statistics 11(4), 1833 – 1874.arxiv.org/abs/1608.05498
  25. [29]Patton, A. J., Ziegel, J. F. & Chen, R. (2019), ‘Dynamic semiparametric models for expected shortfall (and Value-at-Risk)’, Journal of Econometrics 211(2), 388–413.
  26. [30]Rockafellar, R. T., Royset, J. O. & Miranda, S. I. (2014), ‘Superquantile regression with applications to buffered reliability, uncertainty quantification, and conditional Value-at-Risk’, European Journal of Operational Research 234(1), 140–154.DOI
  27. [31]Romano, J. P. & Wolf, M. (2017), ‘Resurrecting weighted least squares’, Journal of Econometrics 197(1), 1–19.
  28. [32]Tasche, D. (2002), ‘Expected shortfall and beyond’, Journal of Banking & Finance 26(7), 1519–1533.
  29. [33]Taylor, J. W. (2019), ‘Forecasting Value at Risk and expected shortfall using a semiparametric approach based on the asymmetric Laplace distribution’, Journal of Business & Economic Statistics 37(1), 121–133.
  30. [34]Wang, C., Gerlach, R. & Chen, Q. (2023), ‘A semi-parametric conditional autoregressive joint Value-at-Risk and expected shortfall modeling framework incorporating realized measures’, Quantitative Finance 23(2), 309–334.DOI
  31. [35]White, H. (1980), ‘Nonlinear regression on cross-section data’, Econometrica 48(3), 721–746.
  32. [36]White, H., Kim, T.-H. & Manganelli, S. (2015), ‘VAR for VaR: Measuring tail dependence using multivariate regression quantiles’, Journal of Econometrics 187(1), 169–188.
  33. [37]Wooldridge, J. M. (2010), Econometric analysis of cross section and panel data, MIT Press.
  34. [39]Yu, J. (2005), ‘On leverage in a stochastic volatility model’, Journal of Econometrics 127(2), 165–178.
  35. [40]Zhang, S., He, X., Tan, K. M. & Zhou, W.-X. (2025), ‘High-dimensional expected shortfall regression’, Journal of the American Statistical Association 120(551), 1799–1810.
  36. [41]Zhu, Q., Zeng, R. & Li, G. (2020), ‘Bootstrap inference for GARCH models by the least absolute deviation estimation’, Journal of Time Series Analysis 41(1), 21–40.DOI

Paper details

vlnc:W_9g23sz9f
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents