Saralees Nadarajah and Stephen Chan
School of Mathematics, University of Manchester, Manchester, M13 9PL, UK
In the last few decades, risk managers have truly experienced a revolution. The rapid increase in the usage of risk management techniques has spread well beyond derivatives and is totally changing the way institutions approach their financial risk. In response to the financial disasters of the early 1990s, a new method called Value at Risk (VaR) was developed as a simple method to quantify market risk (in recent years, VaR has been used in many other areas of risk including credit risk and operational risk). Some of the financial disasters of the early 1990s are the following:
Till Guldimann is widely credited as the creator of value at risk (VaR) in the late 1980s. He was then the head of global research at J.P. Morgan. VaR is a method that uses standard statistical techniques to assess risk. The VaR “measures the worst average loss over a given horizon under normal market conditions at a given confidence level” (Jorion, 2001, p. xxii). The value of VaR can provide users with information in two ways: as a summary measure of market risk or an aggregate view of a portfolio's risk. Overall VaR is a forward-looking risk measure and used by financial institutions, regulators, nonfinancial corporations, and asset management exposed to financial risk. The most important use of VaR has been for capital adequacy regulation under Basel II and later revisions.
Let denote a stationary financial series with marginal cumulative distribution function (cdf) and marginal probability density function (pdf) . The VaR for a given probability is defined mathematically as
That is, VaR is the quantile of exceeded with probability . Figure 12.6 illustrates the definition given by (12.1).
Sometimes, VaR is defined for log returns of the original time series. That is, if , are the log returns for some with marginal cdf and then VaR is defined by (12.1). If and denote the mean and standard deviation of the log returns, then one can write
where denotes the quantile function of the standardized log returns .
Applications of VaR can be classified as:
Applications of VaR have been extensive. Some recent applications and application areas have included estimation of highly parallel architectures (Dixon et al., 2012), estimation for crude oil markets (He et al., 2012a), multiresolution analysis-based methodology in metal markets (He et al., 2012b), estimation of optimal hedging strategy under bivariate regime switching ARCH framework (Chang, 2011b), energy markets (Cheong, 2011), Malaysian sectoral markets (Cheong and Isa, 2011), downside residential market risk (Jin and Ziobrowski, 2011), hazardous material transportation (Kwon, 2011), operational risk in Chinese commercial banks (Lu, 2011), longevity and mortality (Plat, 2011), analysis of credit default swaps (Raunig et al., 2011), exploring oil-exporting country portfolio (Sun et al., 2011), Asia-focused hedge funds (Weng and Trueck, 2011), measure for waiting time in simulations of hospital units (Dehlendorff et al., 2010), financial risk in pension funds (Fedor, 2010), catastrophic event modeling in the Gulf of Mexico (Kaiser et al., 2010), estimating the South African equity market (Milwidsky and Mare, 2010), estimating natural disaster risks (Mondlane, 2010), wholesale price for supply chain coordination (Wang, 2010), US movie box office earnings (Bi and Giles, 2009), stock market index portfolio in South Africa (Bonga-Bonga and Mutema, 2009), multiperiod supply inventory coordination (Cai et al., 2009), Toronto stock exchange (Dionne et al., 2009), modeling volatility clustering in electricity price return series (Karandikar et al., 2009), calculation for heterogeneous loan portfolios (Puzanova et al., 2009), measurement of HIS stock index futures market risk (Yan and Gong, 2009), stock index futures market risk (Gong and Li, 2008), estimation of real estate values (He et al., 2008), foreign exchange rates (Ku and Wang, 2008), artificial neural network (Lin and Chen, 2008), criterion for management of stormwater (Piantadosi et al., 2008), inventory control in supply chains (Yiu et al., 2008), layers of protection analysis (Fang et al., 2007), project finance transactions (Gatti et al., 2007), storms in the Gulf of Mexico (Kaiser et al., 2007), midterm generation operation planning in electricity market environment (Lu et al., 2007), Hong Kong's fiscal policy (Porter, 2007), bakery procurement (Wilson et al., 2007), newsvendor models (Xu and Chen, 2007), optimal allocation of uncertain water supplies (Yamout et al., 2007), futures floor trading (Lee and Locke, 2006), estimating a listed firm in China (Liu et al., 2006), Asian Pacific stock market (Su and Knowles, 2006), Polish power exchange (Trzpiot and Ganczarek, 2006), single loss approximation to VaR (Böcker and Klüppelberg, 2005), real options in complex engineered systems (Hassan et al., 2005), effects of bank technical sophistication and learning over time (Liu et al., 2004), risk analysis of the aerospace sector (Mattedi et al., 2004), Chinese securities market (Li et al., 2002), risk management of investment-linked household property insurance (Zhu and Gao, 2002), project risk measurement (Feng and Chen, 2001), long-term capital management for property/casualty insurers (Panning, 1999), structure-dependent securities and FX derivatives (Singh, 1997), and mortgage-backed securities (Jakobsen, 1996).
The aim of this chapter is to review known methods for estimating VaR given by (12.1). The review of methods is divided as follows: general properties (Section 12.2), parametric methods (Section 12.3), nonparametric methods (Section 12.4), semiparametric methods (Section 12.5), and computer software (Section 12.6). For each estimation method, we give the main formulas for computing VaR. We have avoided giving full details for each estimation method (e.g., interpretation, asymptotic properties, finite sample properties, finite sample bias, sensitivity to outliers, quality of approximations, comparison with competing estimators, advantages, disadvantages, and application areas) because of space concerns. These details can be read from the cited references.
The review of value of risk presented here is not complete, but we believe we have covered most of the developments in recent years. For a fuller account of the theory and applications of value risk, we refer the readers to the following books: Bouchaud and Potters (2000, Chapter 3), Delbaen (2000, Chapter 3), Moix (2001, Chapter 6), Voit (2001, Chapter 7), Dupacova et al. (2002, Part 2), Dash (2004, Part IV), Franke et al. (2004), Tapiero (2004, Chapter 10), Meucci (2005), Pflug and Romisch (2007, Chapter 12), Resnick (2007), Ardia (2008, Chapter 6), Franke et al. (2008), Klugman et al. (2008), Lai and Xing (2008, Chapter 12), Taniguchi et al. (2008), Janssen et al. (2009, Chapter 18), Sriboonchitta et al. (2010, Chapter 4), Tsay (2010),Capinski and Zastawniak (2011), Jorion (2001), and Ruppert (2011, Chapter 19).
This section describes general properties of VaR. The properties discussed are ordering properties (Section 12.2.1), upper comonotonicity (Section 12.2.2), multivariate extension (Section 12.2.3), risk concentration (Section 12.2.4), Hürlimann's inequalities (Section 12.2.5), Ibragimov and Walden's inequalities (Section 12.2.6), Denis et al.'s inequalities (Section 12.2.7), Jaworski's inequalities (Section 12.2.8), Mesfioui and Quessy's inequalities (Section 12.2.9), and Slim et al.'s inequalities (Section 12.2.10).
Pflug (2000) and Jadhav and Ramanathan (2009) establish several ordering properties of . Given random variables , , , and a constant , some of the properties given by Pflug (2000) and Jadhav and Ramanathan (2009) are the following:
Let denote the joint cdf of with marginal cdfs and . Write to mean , where is known as the copula (Nelsen, 1999), a joint cdf of uniform marginals. Let have the joint cdf and have the joint cdf , , and . Then, Tsafack (2009) shows that if is stochastically less than , then for .
If two or more assets are comonotonic, then their values (whether they be small, medium, large, etc.) move in the same direction simultaneously. In the real world, this may be too strong of a relation. A more realistic relation is to say that the assets move in the same direction if their values are extremely large. This weaker relation is known as upper comonotonicity (Cheung, 2009).
Let denote the loss of the th asset. Let with joint cdf . Let . Suppose all random variables are defined on the probability space . Then, a simple formula for the VaR of in terms of values at risk of can be established if is upper comonotonic.
We now define what is meant by upper comonotonicity. A subset is said to be comonotonic if for all and whenever and belong to . The random vector is said to be comonotonic if it has a comonotonic support.
Let denote the collection of all zero probability sets in the probability space. Let . For a given , let denote the upper quadrant of and let denote the lower quadrant of . Let .
Then, the random vector is said to be upper comonotonic if there exist and a zero probability set such that
If these three conditions are satisfied, then the VaR of can be expressed as
for and , a comonotonic threshold as constructed in Lemma 2 of Cheung (2009).
In this chapter, we shall focus mainly on univariate VaR estimation. Multivariate VaR is a much more recent topic.
Let be a random vector in with joint cdf . Prékopa (2012) gives the following definition of multivariate VaR:
Note that MVaR may not be a single vector. It will often take the form of a set of vectors.
Prékopa (2012) gives the following motivation for multivariate VaR: “A finance company generally faces the problem of constructing different portfolios that they can sell to customers. Each portfolio produces a random total return and it is the objective of the company to have them above given levels, simultaneously, with large probability. Equivalently, the losses should be below given levels, with large probability. In order to ensure it we look at the total losses as components of a random vector and find a multivariate -quantile or MVaR to know what are those points in the -dimensional space ( being the number of portfolios), that should surpass the vector of total losses, to guarantee the given reliability.”
Cousin and Bernardinoy (2011) provide another definition of multivariate VaR:
or equivalently
where is the boundary of the set .
Cousin and Bernardinoy (2011) establish various properties of MVaR similar to those in the univariate case. For instance,
for , where denotes the th component of ;
for ;
for all ;
for all .
Bivariate VaR in the context of a bivariate normal distribution has been considered much earlier by Arbia (2002).
A matrix variate extension of VaR and its application for power supply networks are discussed in Chang (2011a).
Let denote future losses, assumed to be nonnegative independent random variables with common cdf and survival function . Degen et al. (2010) define risk concentration as
If is regularly varying with index , (Bingham et al., 1989), meaning that as , then it is shown that
as . Degen et al. (2010) also study the rate of convergence in (12.5).
Suppose , are regularly varying with index , . According to Jang and Jho (2007), for ,
for all for some . This property is referred to as subadditivity. If holds as , then the property is referred to as asymptotic subadditivity. For ,
as . This property is referred to as asymptotic comonotonicity. For ,
for all for some . If holds as then the property is referred to as asymptotic superadditivity.
Let denote a counting process independent of with for . According to Jang and Jho (2007), in the case of subadditivity,
for all for some . In the case of asymptotic comonotonicity,
as . In the case of superadditivity,
for all for some .
Suppose is multivariate regularly varying with index according to Definition 2.2 in Embrechts et al. (2009a). If is a measurable function such that
then it is shown that
see Lemma 2.3 in Embrechts et al. (2009b).
Let denote a random variable defined over , with mean and variance . Hürlimann (2002) provides various upper bounds for : for ,
for ,
for ,
The equality in (12.6) holds if and only if .
Now suppose is a random variable defined over , with mean , variance , skewness , and kurtosis . In this case, Hürlimann (2002) provides the following upper bound for :
where is the percentile of the standardized Chebyshev–Markov maximal distribution. The latter is defined as the root of
if and as the root of
if , where
where , , , and .
Let denote a portfolio return made up of asset returns, , and the nonnegative weights . Ibragimov (2009) provides various inequalities for the VaR of . They suppose that are independent and identically distributed and belong to either , the class of distributions that are convolutions of symmetric stable distributions with and , or , convolutions of distributions from the class of symmetric log-concave distributions and the class of distributions that are convolutions of symmetric stable distributions with and .
Here, denotes a stable distribution specified by its characteristic function
where , , , , and . The stable distribution contains as particular cases the Gaussian distribution for , the Cauchy distribution for and , the Lévy distribution for and , the Landau distribution for and , and the dirac delta distribution for and .
Furthermore, let . Write to mean that for and , where and denote the components of and in descending order. Let and .
With these notations, Ibragimov (2009) provides the following inequalities for . Suppose first that and belong to . Then,
Suppose now that and belong to . Then,
Further inequalities for VaR are provided in Ibragimov and Walden (2011) when a portfolio return, say, , is made up of a two-dimensional array of asset returns, say, . That is,
where are referred to as “row effects,” are referred to as “column effects,” and are referred to as “idiosyncratic components.”
Let , , , , , and .
With these notations, Ibragimov and Walden (2011) provide the following inequalities for :
Ibragimov and Walden (2011, Section 12.4) discuss an application of these inequalities to portfolio component VaR analysis.
Let denote prices of financial assets. The process could be modeled by
where is a Brownian motion; is a compound Poisson process independent of ; are jump times for ; is an adapted integrable process; and , are certain random variables.
Denis et al. (2009) derive various bounds for the VaR of the process
The following assumptions are made:
almost everywhere for all . In this case, let
for .
With these assumptions, Denis et al. (2009) show that
For , Denis et al. (2009) show that
If the jumps follow a simple Poisson process, Denis et al. (2009) show that
If the jumps follow an exponential distribution with parameter , Denis et al. (2009) show that
About the issue of continuity/discontinuity of the market with jumps, see Walter (2015).
Jaworski (2007, 2008) considers the following situation: suppose , are the quotients of the currency rates at the end and at the beginning of an investment; suppose that the joint cdf of is , where is a copula (Nelsen, 1999) and is the marginal cdf of ; suppose is the part of the capital invested in the th currency, where are nonnegative and sum to one. Then, the final investment value is
where . Jaworski (2007, 2008) defines the value of risk for a given and a probability as
Jaworski (2007) shows this VaR can be bounded as
for portfolios consisting of only one currency, where and .
Suppose a portfolio is made up of assets and let denote the losses for the assets. Suppose also that the joint cdf of is , where is a copula (Nelsen, 1999) and is the marginal cdf of . Furthermore, define the dual of a given copula (Definition 2.4, Mesfioui and Quessy, 2005) as
With these notations, Mesfioui and Quessy (2005) derive various inequalities for the VaR of . If is such that and for some copulas and , then
where
and
If are identical random variables with common cdf and if is such that is nonincreasing for , then it is shown under certain conditions that
where is the diagonal section of .
Mesfioui and Quessy (2005) also show that if is a random variable with mean and variance , then
where
and
where . If , have means , and variances , , then it is shown that
where and .
Suppose a portfolio is made up of assets. Let denote the losses for the assets. Let and denote the cdf and the pdf of . Let denote the value for which is nonincreasing for all . Given this notation, the total portfolio loss can be expressed as for some nonnegative weights summing to one. Slim et al. (2012) show that the VaR of can be bounded as follows:
where
and
for . The use of the earlier results allows easy computation for explicit VaR bounds for possibly dependent risks.
This section concentrates on estimation of VaR when data comes from a parametric distribution, and we want to make use of the parameters. The parametric methods summarized are based on Gaussian distribution (Section 12.3.1), Student's distribution (Section 12.3.2), Pareto-positive stable distribution (Section 12.3.3), log-folded distribution (Section 12.3.4), variance–covariance method (Section 12.3.5), Gaussian mixture distribution (Section 12.3.6), generalized hyperbolic distribution (Section 12.3.7), Fourier transformation method (Section 12.3.8), principal components method (Section 12.3.9), quadratic forms (Section 12.3.10), elliptical distribution (Section 12.3.11), copula method (Section 12.3.12), Gram–Charlier approximation (Section 12.3.13), delta–gamma approximation (Section 12.3.14), Cornish–Fisher approximation (Section 12.3.15), Johnson family method (Section 12.3.16), Tukey method (Section 12.3.17), asymmetric Laplace distribution (Section 12.3.18), asymmetric power distribution (Section 12.3.19), Weibull distribution (Section 12.3.20), ARCH models (Section 12.3.21), GARCH models (Section 12.3.22), GARCH model with heavy tails (Section 12.3.23), ARMA–GARCH model (Section 12.3.24), Markov switching ARCH model (Section 12.3.25), fractionally integrated GARCH model (Section 12.3.26), RiskMetrics model (Section 12.3.27), capital asset pricing model (Section 12.3.28), Dagum distribution (Section 12.3.29), location-scale distributions (Section 12.3.30), discrete distributions (Section 12.3.31), quantile regression method (Section 12.3.32), Brownian motion method (Section 12.3.33), Bayesian method (Section 12.3.34), and Rachev et al.'s method (Section 12.3.35).
If are observations from a Gaussian distribution with mean and variance , then VaR can be estimated by
where is the sample mean and is the sample variance
The estimator in (12.7) is biased and consistent. If the in (12.8) is replaced by , then (12.7) becomes unbiased and consistent.
If are observations from a Student's distribution with degrees of freedom, then VaR can be estimated by (Arneric et al., 2008)
where is the excess sample kurtosis and is the percentile of a Student's random variable with degrees of freedom.
Sarabia and Prieto (2009) and Guillen et al. (2011) introduce the Pareto-positive stable distribution specified by the cdf
for , , and . Here, and are shape parameters and is a scale parameter. The Pareto distribution is the particular case of (12.9) for .
The Pareto-positive stable distribution has been applied to risk management; see, for example, Guillen et al. (2011). If is a random variable having the cdf (12.9), then it is easy to see that
for . So, if are maximum likelihood estimators of , then
for .
Brazauskas and Kleefeld (2011) introduce the log-folded distribution specified by the quantile function
for , where is a scale parameter, is a shape parameter, and denotes the quantile function of a Student's random variable with degrees of freedom. Brazauskas and Kleefeld (2011) also provide an application of this distribution to risk management.
Suppose is a random sample from the log-folded distribution with order statistics . Brazauskas and Kleefeld (2011) show that the VaR can be estimated by
where
or
where
where and are integers such that and as , where and are trimming proportions with .
Suppose the portfolio return, say, , is made up of asset returns, , , as
where are nonnegative weights summing to one. Suppose also E, Var, and Cov. The variance–covariance method suggests that the VaR of can be approximated by
An estimator can be obtained by replacing the parameters , , and by their maximum likelihood estimators.
Let denote the financial asset prices and let denote the log return corresponding to the original financial series. Zhang and Cheng (2005) consider the model that have a Gaussian mixture distribution specified by the pdf
for , where the mixing coefficients sum to one. Let denote the VaR corresponding to the th component, that is,
Let denote the VaR corresponding to the mixture model, that is,
Then, Theorem 1 in Zhang and Cheng (2005) shows that
always holds.
Furthermore, let denote the significance level of VaR corresponding to the th component, that is,
Let denote the significance level of VaR corresponding to the mixture model, that is,
Then, Theorem 2 in Zhang and Cheng (2005) shows that
always holds.
Suppose the log returns, , follow the model
where is the volatility process and are independent and identical random variables with zero mean and unit variance. Let denote the corresponding VaR. Suppose are independent and identical and have the generalized hyperbolic distribution specified by the pdf
where is a location parameter, is a shape parameter, is an asymmetry parameter, is a scale parameter, , , and denotes the modified Bessel function of order .
Tian and Chan (2010) propose a method based on saddlepoint approximation for computing . It can be described as follows:
for , where are some nonnegative weights summing to one.
where
Siven et al. (2009) suggest a method for computing VaR by approximating the cdf by a Fourier series. The approximation is given by the following result due to Hughett (1998): suppose
Then, for constants , and , the cdf can be approximated as
where , denotes the real part, and
An estimator for is obtained by solving the equation
for .
Brummelhuis et al. (2002) use an approximation based on the principal component method to compute VaR. If is a vector of risk factors over time and if is a random variable, they define VaR to be
This equation is too general to be solved. So, Brummelhuis et al. (2002) consider the quadratic approximation
and assume that is normally distributed with mean and covariance matrix . Under this approximation, we can rewrite (12.10) as
Let denote the Cholesky decomposition and let
Also let denote the principal components decomposition of , , and . With these notations, Brummelhuis et al. (2002) show that VaR can be approximated by
where is the root of
Suppose the financial series are realizations of a quadratic form
where is a standard normal vector, , and . Examples include nonlinear positions like options in finance or the modeling of bond prices in terms of interest rates (duration and convexity). Here, s are the eigenvalues sorted in ascending order. Suppose there are distinct eigenvalues. Let denote the highest index of the th distinct eigenvalue with multiplicity . For , let
Let denote the moment-generating function of evaluated at . With this notation, Jaschke et al. (2004) derive various approximations for VaR. The first of these applicable for is
where denotes the percentile of a noncentral chi-square random variable with degrees of freedom and noncentrality parameter . The second of the approximations applicable for and is
where
The third of the approximations applicable for and is
where
Suppose a portfolio return, say, , is made up of asset returns, say, , , as , where are nonnegative weights summing to one, and . Kamdem (2005) derives various expressions for the VaR of by supposing that has an elliptically symmetric distribution.
If has the joint pdf , where is the mean vector, is the variance–covariance matrix, and is a continuous and integrable function over , then it is shown that
where is the root of
where
If follows a mixture of elliptical pdfs given by
where is the mean vector for the th elliptical pdf, is the variance–covariance matrix for the th elliptical pdf, and are nonnegative weights summing to one, then it is shown that the VaR of is the root of
where is defined as in (12.11).
Suppose a portfolio return, say, , is made up of two asset returns, and , as , where is the portfolio weight for asset 1 and is the portfolio weight for asset 2. Huang et al. (2009) consider computation of VaR for this situation by supposing that the joint cdf of is , where is a copula (Nelsen, 1999), is the marginal cdf of , and is the marginal pdf of . Then, the cdf of is
where is the copula pdf. So, can be computed by solving the equation
In general, this equation will have to be solved numerically or by simulation.
Franke et al. (2011) consider the more general case that the portfolio return is made up of asset returns, , ; that is,
for some nonnegative weights summing to one. Suppose as in the preceding text that the joint cdf of is , where is the marginal cdf of and is the marginal pdf of . Then, the cdf of is
where
and
So, can be computed by solving the equation
Again, this equation will have to be computed by numerical integration or simulation.
Simonato (2011) suggests a number of approximations for computing (12.2). The first of these is based on Gram–Charlier expansion.
Let denote the skewness coefficient and the kurtosis coefficient of the standardized log returns. Simonato (2011) suggests the approximation
where is the inverse function of
where denotes the standard normal cdf and denotes the standard normal pdf.
Let denote a vector of returns normally distributed with zero means and covariate matrix . Suppose the return of an associated portfolio takes the general form . It will be difficult to find the value of risk of for general . Some approximations are desirable. The delta–gamma approximation is a commonly used approximation (Feuerverger and Wong, 2000).
Suppose we can approximate for a vector and a matrix. Let denote the Cholesky decomposition. Let and denote the eigenvalues and eigenvectors of . Let denote the entries of , where . Then, the delta–gamma approximation is that
where are independent standard normal random variables. The value of risk can be obtained by inverting the distribution of the right-hand side of (12.12).
Another approximation suggested by Simonato (2011) is based on Cornish–Fisher expansion. With the notation as in Section 12.3.13, the approximation is
where is the inverse function of
where denotes the standard normal quantile function.
A third approximation suggested by Simonato (2011) is based on the Johnson family of distributions due to Johnson (1949).
Let denote a standard normal random variable. A Johnson random variable can be expressed as
where
Here, , , , and are unknown parameters determined, for example, by the method of moments; see Hill et al. (1976).
With the notation as in the preceding text, the approximation is
where
where denotes the standard normal quantile function.
Jiménez and Arunachalam (2011) present a method for approximating VaR based on Tukey's and family of distributions.
Let denote a standard normal random variable. A Tukey and random variable can be expressed as
for and . The family of lognormal distributions is contained as the particular case for . The family of Tukey's distribution is contained as the limiting case for .
With the notation as in Section 12.3.13, the approximation suggested by Jiménez and Arunachalam (2011) is
where and are location and scale parameters. For and , is a normal random variable with mean and standard deviation , so and . For and , is an exponential random variable with parameter , so and . For and , is a Student's random variable with ten degrees of freedom, so and .
Trindade and Zhu (2007) consider the case that the log returns of are a random sample from the asymmetric Laplace distribution given by the pdf
for , , and . The maximum likelihood estimator of is derived as
where are the maximum likelihood estimators of . Trindade and Zhu (2007) show further that
in distribution as , where and .
Komunjer (2007) introduces the asymmetric power distribution as a model for risk management. A random variable, say, , is said to have this distribution if its pdf is
for , where , and . Note that is a shape parameter and is a scale parameter. The cdf corresponding to (12.13) is shown to be (Lemma 1, Komunjer, 2007)
where . Inverting (12.14) as in Lemma 2 of Komunjer (2007), we can express as
where denotes the inverse function of . An estimator of can be obtained by replacing the parameters in (12.15) by their maximum likelihood estimators; see Proposition 2 in Komunjer (2007).
Gebizlioglu et al. (2011) consider estimation of VaR based on the Weibull distribution. Suppose is a random sample from a Weibull distribution with the cdf specified by for , and . Then, the estimator for VaR is
Gebizlioglu et al. (2011) consider various methods for obtaining the estimators and . By the method of maximum likelihood, and are the simultaneous solutions of
and
where is the sample mean and is the sample variance. By Cohen and Whitten (1982)'s modified method of maximum likelihood, and are the simultaneous solutions of
and
where are the order statistics in ascending order. By Tiku (1967 and 1968) and Tiku and Akkaya (2004)'s modified method of maximum likelihood,
where
By the least squares method, and are those minimizing
with respect to and . By the weighted least squares method, and are those minimizing
with respect to and . By the percentile method, and are those minimizing
with respect to and .
ARCH models are popular in finance. Suppose the log returns, say, , of follow the ARCH model specified by
where are independent and identical random variables with zero mean, unit variance, pdf , and cdf , and is an unknown parameter vector satisfying and , . If are the maximum likelihood estimators, then the residuals are
where
Taniai and Taniguchi (2008) show that VaR for this ARCH model can be approximated by
where
where , , , , , , and , .
Suppose the financial returns, say, , satisfy the model
where are independent and identical standard normal random variables, is the return at time , denotes the lag operator satisfying , is the polynomial , is the polynomial , is the conditional variance, and are independent and identical residuals with zero means and unit variances. One popular specification for is
This corresponds to the GARCH model.
For the model given by (12.16) and (12.17), Chan (2009b) proposes the following algorithm for computing VaR:
The parameters are determined from the sample moments of Step 3 in a way explained in Chan (2009a) and Rockinger and Jondeau (2002).
Chan et al. (2007) consider the case that financial returns, say, , come from a GARCH specified by
where is strictly stationary with , and are zero mean, unit variance, independent, and identical random variables independent of . Further, Chan et al. (2007) assume that have heavy tails, that is, their cdf, say, , satisfies
for all , where and . Chan et al. (2007) show that the VaR for this model given by
can be estimated by
where
where and as , , are the order statistics of , and and as . Chan et al. (2007) also establish asymptotic normality of .
Suppose the financial returns, say, , , satisfy the ARMA –GARCH model specified by
where are independent standard normal random variables. For this model, Hartz et al. (2006) show that the -step ahead forecast of VaR can be estimated by
where
The parameter estimators required can be obtained, for example, by the method of maximum likelihood.
Suppose the financial returns, say, , , satisfy the Markov switching ARCH model specified by
where are standard normal random variables, is an unobservable random variable assumed to follow a first-order Markov process, and is a typical ARCH process. This model is due to Bollerslev (1986). An estimator of the VaR at time can be obtained by inverting the cdf of with its parameters replaced by their maximum likelihood estimators.
Suppose the financial returns, say, , , satisfy the fractionally integrated GARCH model specified by
where are random variables with zero means and unit variances. This model is due to Baillie et al. (1996). An estimator of the VaR at time can be obtained by inverting the cdf of with its parameters replaced by their maximum likelihood estimators. This of course depends on the distribution of . If, for example, are normally distributed, then , where may be the maximum likelihood estimator of .
Suppose are the log returns of and let denote the information up to time . The RiskMetrics model (RiskMetrics Group, 1996) is specified by
The VaR for this model can be computed by inverting
with the parameters, and , replaced by their maximum likelihood estimators.
Let denote the return on asset , let denote the “risk-free rate,” and let denote the “return on the market portfolio.” With this notation, Fernandez (2006) considers the capital asset pricing model given by
for , where are independent random variables with Var and Var. It is easy to see that
Fernandez (2006) shows that the VaR of the portfolio of assets can be expressed as
where is a vector of portfolio weights, is the initial value of the portfolio, , and diag . An estimator of (12.18) can be obtained by replacing the parameters by their maximum likelihood estimators.
The Dagum distribution is due to Dagum (1977 and 1980). It has the pdf and cdf specified by
and
respectively, for , , , and . Domma and Perri (2009) discuss an application of this distribution for VaR estimation. They show that
where are maximum likelihood estimators of based on being a random sample coming from the Dagum distribution. Domma and Perri (2009) show further that
in distribution as , where and
Here, is the expected information matrix of . An explicit expression for the matrix is given in the appendix of Domma and Perri (2009).
Suppose is a random sample from a location-scale family with cdf and pdf . Then,
where . The point estimator for VaR is
where
and
Bae and Iscoe (2012) propose various confidence intervals for VaR. Based on and asymptotic normality, Bae and Iscoe (2012) propose the interval
where is the confidence level, is the kurtosis of , and is the skewness of . Based on Bahadur (1966)'s almost sure representation of the sample quantile of a sequence of independent random variables, Bae and Iscoe (2012) propose the interval
where is the th quantile and is its sample counterpart.
Sometimes the financial series of interest is strictly positive. In this case, if is a random sample from a log location-scale family with cdf , then (12.19) and (12.20) generalize to
and
respectively, as noted by Bae and Iscoe (2012).
Göb (2011) considers VaR estimation for the three most common discrete distributions: Poisson, binomial, and negative binomial. Let
Then, the VaR for the Poisson distribution is
Letting
the VaR for the binomial distribution is
Letting
the VaR for the negative binomial distribution is
Göb (2011) derives various properties of these VaR measures in terms of their parameters. For the Poisson distribution, the following properties were derived:
For the binomial distribution, the following properties were derived:
For the negative binomial distribution, the following properties were derived:
Empirical estimation of the three VaR measures can be based on asymptotic normality.
Quantile regressions have been used to estimate VaR; see Koenker and Bassett (1978), Koenker and Portnoy (1997), Chernozhukov and Umantsev (2001), and Engle and Manganelli (2004). The idea is to regress the VaR on some known covariates. Let at time denote the financial variable, let denote a vector of covariates at time , let denote a vector of regression coefficients, and let denote the corresponding VaR. Then, the quantile regression model can be rewritten as
In the linear case, (12.21) could take the form
The parameters in (12.21) can be estimated by least squares as in standard regression.
Cakir and Raei (2007) describe simulation schemes for computing VaR for single-asset and multiple-asset portfolios. Let denote the price at time , let denote a holding period divided into small intervals of equal length , let denote the change in over , let denote a standard normal shock, let denote the mean of returns over the holding period , and let denote the standard deviation of returns over the holding period . With these notations, Cakir and Raei (2007) suggest the model
Under this model, the VaR for single-asset portfolios can be computed as follows:
The VaR for multiple-asset portfolios can be computed as follows:
for , where is the number of assets and the notation is the same as that for single-asset portfolios. The standard normal shocks, , need not be correlated.
Pollard (2007) defines a Bayesian VaR. Let denote the financial variable of interest at time . Let denote the posterior pdf of given some parameters and “state” variables . Pollard (2007) defines the Bayesian VaR at time as
The “state” variables are assumed to follow a transition pdf .
Pollard (2007) also proposes several methods for estimating (12.24). One of them is the following:
Let denote a portfolio return made up of asset returns, , and the nonnegative weights summing to one. Suppose are independent random variables. Then, it can be shown that (Rachev et al., 2003) , where
and
Hence, the value of risk of can be estimated by the following algorithm due to Rachev et al. (2003):
and
respectively.
This section concentrates on estimation methods for VaR when the data are assumed to come from no particular distribution. The nonparametric methods summarized are based on historical method (Section 12.4.1), filtered historical method (Section 12.4.2), importance sampling method (Section 12.4.3), bootstrap method (Section 12.4.4), kernel method (Section 12.4.5), Chang et al.'s estimators (Section 12.4.6), Jadhav and Ramanathan's method (Section 12.4.7), and Jeong and Kang's method (Section 12.4.8).
Let denote the order statistics in ascending order corresponding to the original financial series . The historical method suggests to estimate VaR by
for .
Suppose the log returns, , follow the model, , discussed before, where is the volatility process and are independent and identical random variables with zero means. Let denote the order statistics of . The filtered historical method suggests to estimate VaR by
for , where denotes an estimator of at time . This method is due to Hull and White (1998) and Barone-Adesi et al. (1999).
Suppose is the empirical cdf of . As seen in Section 12.4.1, an estimator for VaR is . This estimator is asymptotically normal with variance equal to
This can be large if is closer to zero or one. There are several methods for variance reduction. One popular method is importance sampling. Suppose is another cdf and let and
Hong (2011) shows that under certain conditions can provide estimators for VaR with smaller variance.
Suppose is the empirical cdf of . The bootstrap method can be described as follows:
One can also construct confidence intervals for VaR based on the bootstrapped estimates , .
Kernels are commonly used to estimate pdfs. Let denote a symmetric kernel, that is, a symmetric pdf. The kernel estimator of can be given by
where is a smoothing bandwidth and
A variable width version of (12.26) is
where is the distance of from its th nearest neighbor among the remaining data points and . The kernel estimator of VaR, say, , is then the root of the equation
for , where is given by (12.26) or (12.27). According to Sheather and Marron (1990), could also be estimated by
where is given by (12.26) or (12.27) and are the ascending order statistics of .
The estimator in (12.28) is due to Gourieroux et al. (2000). Its properties have been studied by many authors. For instance, Chen and Tang (2005) show under certain regularity conditions that
in distribution as , where
Here, denotes the indicator function.
Chang et al. (2003) propose several nonparametric estimators for the VaR of log returns, say, with pdf . The first of these is , where and , where denotes the greatest integer less than or equal to . This estimator is shown to have the asymptotic distribution
in distribution as . It is sometimes referred to as the historical simulation estimator. The second of the proposed estimators is
This estimator is shown to have the asymptotic distribution
in distribution as . The third of the proposed estimators is
where
This estimator is shown to have the asymptotic distribution
in distribution as .
Jadhav and Ramanathan (2009) provide a collection of nonparametric estimators for . Let denote the order statistics in ascending order corresponding to . For given , define , , , , , and . The collection provided is
where
where denote the incomplete beta function ratio defined by
The last of the estimators in the collection is due to Kaigh and Lachenbruch (1982). The second last is due to Harrell and Davis (1982).
Suppose the log returns, , follow the model, , discussed before. Let denote the corresponding VaR. Jeong and Kang (2009) propose a fully nonparametric estimator for the defined by
where is the -field generated by . Let
and
for some kernel function with bandwidth . With this notation, Jeong and Kang (2009) propose the estimator
where
and
Here, can be determined using a recursive algorithm presented in Section 12.2.1 of Jeong and Kang (2009).
This section concentrates on estimation methods for VaR that have both parametric and nonparametric elements. The semiparametric methods summarized are based on extreme value theory method (Section 12.5.1), generalized Pareto distribution (Section 12.5.2), Matthys et al.'s method (Section 12.5.3), Araújo Santos et al.'s method (Section 12.5.4), Gomes and Pestana's method (Section 12.5.5), Beirlant et al.'s method (Section 12.5.6), Caeiro and Gomes' method (Section 12.5.7), Figueiredo et al.'s method (Section 12.5.8), Li et al.'s method (Section 12.5.9), Gomes et al.'s method (Section 12.5.10), Wang's method (Section 12.5.11), -estimation method (Section 12.5.12), and the generalized Champernowne distribution (Section 12.5.13).
Let denote the maximum of financial returns. Extreme value theory says that under suitable conditions there exist norming constants and such that
in distribution as . The parameter is known as the extreme value index. It controls the tail behavior of the extremes.
There are several estimators proposed for . One of the earliest estimators due to Hill (1975) is
where are the order statistics in descending order. Another earliest estimator due to Pickands (1975) is
The tails of for most situations in finance take the Pareto form, that is,
for some constant . Embrechts et al. (1997, p. 334) propose estimating by .
Combining (12.29) and (12.31), Odening and Hinrichs (2003) propose estimating VaR by
This estimator is actually due to Weissman (1978).
An alternative approach is to suppose that the maximum of financial returns follows the generalized extreme value cdf (Fisher and Tippett, 1928) given by
for , , , and . In this case, the VaR can be estimated by
where are the maximum likelihood estimators of . Prescott and Walden (1990) provide details of maximum likelihood estimation for the generalized extreme value distribution.
The Gumbel distribution is the particular case of (12.33) for . It has the cdf specified by
for and . If the maximum of financial returns follows this cdf, then the VaR can be estimated by
where are the maximum likelihood estimators of .
For more on extreme value theory, estimation of the tail index, and applications, we refer the readers to Longin (1996, 2000), Beirlant et al. (2017), Fraga Alves and Neves (2017), and Gomes et al. (2015).
The Pareto distribution is a popular model in finance. Suppose the log return, say, , of comes from the generalized Pareto distribution with cdf specified by
for , , and , where is some threshold and is the number of observed exceedances above .
For this model, several estimators are available for the VaR. Let denote the order statistics in ascending order. The first estimator due to Pickands (1975) is
where
for . The second estimator due to Dekkers et al. (1989) is
where
Suppose now that the returns are from the alternative generalized Pareto distribution with cdf specified by
for . Then, the VaR is
If and are the maximum likelihood estimators of and , respectively, then the maximum likelihood estimator of VaR is
There are several methods for constructing confidence intervals for (12.34). One popular method is the bias-corrected method due to Efron and Tibshirani (1993). This method based on bootstrapping can be described as follows:
and
where
and
where is the mean of .
where is the percentile of .
Note that and are the bootstrap replicates of and VaR, respectively.
Several improvements have been proposed on (12.32). The one due to Matthys et al. (2004) takes account of censoring. Suppose only of the are actually observed; the remaining are considered to be censored or missing. In this case, Matthys et al. (2004) show that VaR can be estimated by
where
Here, is a tuning parameter and takes values in the unit interval. Among other properties, Matthys et al. (2004) establish asymptotic normality of .
The improvement of (12.32) due to Araújo Santos et al. (2006) takes the expression
where and
The improvement of (12.32) due to Gomes and Pestana (2007) takes the expression
where
Here, is a tuning parameter. Under suitable conditions, Gomes and Pestana (2007) show further that
in distribution as .
The improvement of (12.32) due to Beirlant et al. (2008) takes the expression
where is as given by Section 12.5.5, and
This estimator is shown to be consistent.
Caeiro and Gomes (2008 and 2009) propose several improvements on (12.32). The first of these takes the expression
where and are as given by Section 12.5.5, and is as given by Section 12.5.6. The second of these takes the expression
where and are as given by Section 12.5.5, and is as given by Section 12.5.6. The third of these takes the expression
where and are as given by Section 12.5.5, is as given by Section 12.5.6, and denotes the incomplete beta function defined by
The fourth of these takes the expression
where , , and are as given in Section 12.5.5. All of these estimators are shown to be consistent and asymptotically normal.
The latest improvement of (12.32) is due to Figueiredo et al. (2012). It takes the expression
where and
with as defined in Section 12.5.5 and as defined in Section 12.5.4.
Let be such that and as . Li et al. (2010) derive estimators for for large . They give the estimator
where
and
where and are the simultaneous solutions of the equations
and
where
and
Li et al. (2010) show under suitable conditions that
in distribution as .
Gomes et al. (2011) propose a bootstrap-based method for computing VaR. The method can be described as follows:
for . Choose the tuning parameter, , as zero if and as one otherwise.
and
for and .
for .
Wang (2010) combined the historical method in Section 12.4.1 with the generalized Pareto model in Section 12.5.2 to suggest the following estimator for VaR:
where and are the maximum likelihood estimators of and , respectively, and is an appropriately chosen threshold.
Iqbal and Mukherjee (2012) provide an -estimator for VaR. They consider a GARCH (1, 1) model for returns specified by
where
and are independent and identical random variables symmetric about zero. The unknown parameters are , and they belong to the parameter space, the set of all with , and . The -estimator, say, , is obtained by solving the equation
where
and
where for some skew-symmetric function and denotes the derivative of . Iqbal and Mukherjee (2012) propose that can be estimated by multiplied by the th order statistic of .
Generalized Champernowne distribution was introduced by Buch-Larsen et al. (2005) as a model for insurance claims. A random variable, say, , is said to have this distribution if its cdf is
for , where , , and is the median. Charpentier and Oulidi (2010) provide estimators of based on beta kernel quantile estimators. They suggest the following algorithm for estimating :
where is given by either
or
where .
Software for computing VaR and related quantities are widely available. Some software available from the R package (R Development Core Team, 2015) are the following:
Some other software available for computing VaR and related quantities are the following:
We have reviewed the current state of the most popular risk measure, VaR, with emphasis on recent developments. We have reviewed 10 of its general properties, including upper comonotonicity and multivariate extensions; 35 of its parametric estimation methods, including time series, quantile regression, and Bayesian methods; 8 of its nonparametric estimation methods, including historical methods and bootstrapping; 13 of its semiparametric estimation methods, including extreme value theory and -estimation methods; and 20 known computer software, including those based on the R platform.
This review could encourage further research with respect to measures of financial risk. Some open problems to address are further multivariate extensions of risk measures and corresponding estimation methods; development of a comprehensive R package implementing a wide range of parametric, nonparametric, and semiparametric estimation methods (no such packages are available to date); estimation based on nonparametric Bayesian methods; estimation methods suitable for big data; and so on.
The authors would like to thank Professor Longin for careful reading and comments that greatly improved the chapter.