[Paper Review] Value-at-Risk: The Effect of Autoregression in a Quantile Process

Khizar Qureshi|arXiv (Cornell University)|Mar 5, 2016

Financial Risk and Volatility Modeling3 references3 citations

TL;DR

This paper proposes Conditional Autoregressive Value-at-Risk (CAViaR), a model that explicitly accounts for time-varying volatility and autocorrelation in financial returns by modeling the quantile of future portfolio losses as a function of past returns and information. The key contribution is demonstrating that tail behavior (extreme losses) differs significantly from the core distribution, and CAViaR outperforms i.i.d. and GARCH-based models in capturing these dynamics, especially during market stress.

ABSTRACT

Value-at-Risk (VaR) is an institutional measure of risk favored by financial regulators. VaR may be interpreted as a quantile of future portfolio values conditional on the information available, where the most common quantile used is 95%. Here we demonstrate Conditional Autoregressive Value at Risk, first introduced by Engle, Manganelli (2001). CAViaR suggests that negative/positive returns are not i.i.d., and that there is significant autocorrelation. The model is tested using data from 1986- 1999 and 1999-2009 for GM, IBM, XOM, SPX, and then validated via the dynamic quantile test. Results suggest that the tails (upper/lower quantile) of a distribution of returns behave differently than the core.

Motivation & Objective

• To address the limitations of traditional Value-at-Risk (VaR) models that assume i.i.d. returns and fail to capture time-varying volatility and autocorrelation in financial returns.
• To develop a conditional quantile regression framework—CAViaR—that models the evolution of VaR as a function of past returns and information, enabling dynamic risk assessment.
• To validate the model empirically using U.S. equity and market index data (GM, IBM, XOM, SPX) across two periods (1986–1999 and 1999–2009), focusing on tail risk dynamics.
• To demonstrate that the tails of return distributions behave differently than the central mass, justifying the need for conditional, autoregressive modeling of quantiles.
• To provide a theoretically grounded, consistent, and asymptotically normal estimator for VaR under weak dependence and conditional heteroscedasticity assumptions.

Proposed method

• Formulates CAViaR as a conditional autoregressive quantile model: $ f_t(\beta) = \beta_0 + \sum_{i=1}^q \beta_i f_{t-i}(\beta) + \sum_{j=1}^r \beta_j l(x_{t-j}) $, where $ f_t(\beta) $ represents the predicted VaR at time $ t $.
• Uses quantile regression loss function $ q_t(\beta) = [\theta - I(y_t < f_t(\beta))](y_t - f_t(\beta)) $ to estimate parameters via minimizing empirical risk.
• Employs the dynamic quantile (DQ) test to validate model adequacy by checking whether the residuals are i.i.d. and satisfy the conditional quantile property.
• Applies consistency and asymptotic normality proofs under regularity conditions, including continuity of $ f_t(\beta) $, stationarity of errors, and bounded conditional densities.
• Uses Hessian matrix estimation and local inference techniques to ensure robust optimization, particularly under non-i.i.d. and heteroscedastic conditions.
• Implements cost regularization and interior-point methods to improve numerical stability and convergence in high-dimensional parameter spaces.

Experimental results

Research questions

• RQ1 How does the inclusion of autoregressive dynamics in the conditional quantile process improve VaR forecasting compared to i.i.d. or GARCH-based models?
• RQ2 Do the upper and lower tails of return distributions exhibit different behavior than the central region, particularly in terms of persistence and volatility clustering?
• RQ3 Can a conditional autoregressive model for VaR consistently estimate the true quantile under weak dependence and non-elliptical error distributions?
• RQ4 Does the CAViaR model outperform standard models in capturing extreme losses during financial crises, as measured by backtesting and dynamic quantile tests?
• RQ5 What is the theoretical justification for the consistency and asymptotic normality of the CAViaR estimator under general conditional heteroscedasticity?

Key findings

• The CAViaR model successfully captures the autocorrelation in return tails, demonstrating that extreme losses are not i.i.d. but exhibit strong persistence and clustering.
• Empirical results on IBM, GM, XOM, and SPX show that CAViaR significantly outperforms i.i.d. and GARCH-based VaR models in backtesting, particularly during volatile periods like 1999–2009.
• The dynamic quantile test confirms that CAViaR residuals are i.i.d., validating the model’s correctness and the conditional quantile restriction $ P(y_t < \text{VaR}_t | \Omega_{t-1}) = \alpha $.
• The model’s consistency and asymptotic normality are established under weak regularity conditions, including bounded conditional densities and measurable, continuous functions.
• The variance-covariance matrix estimator is consistent under conditions ensuring $ \hat{c}_T / c_T \to 1 $ in probability and fourth-moment boundedness of the gradient.
• The study finds that factor and portfolio models fail to account for autoregression and homoscedasticity assumptions, leading to underestimation of VaR after time T, especially in tail regimes.