cvgtest {quarks}

Unconditional and Conditional Coverage Tests, Independence Test

Description

The conditional (Kupiec, 1995), the unconditional coverage test (Christoffersen, 1998) and the independence test (Christoffersen, 1998) of the Value-at-Risk (VaR) are applied.

Usage

cvgtest(obj = list(loss = NULL, VaR = NULL, p = NULL), conflvl = 0.95)

Arguments

obj

a list that contains the following elements:

loss

a numeric vector that contains the values of a loss series ordered from past to present; is set to NULL by default.

VaR

a numeric vector that contains the estimated values of the VaR for the same time points of the loss series loss; is set to NULL by default.

p

a numeric vector with one element; defines the probability p stated in the null hypotheses of the coverage tests (see the section Details for more information); is set to NULL by default.

conflvl

a numeric vector with one element; the significance level at which the null hypotheses are evaluated; is set to 0.95 by default. Please note that a list returned by the rollcast function can be directly passed to cvgtest.

Details

With this function, the conditional and the unconditional coverage tests introduced by Kupiec (1995) and Christoffersen (1998) can be applied. Given a return series r_t with n observations, divide the series into n-K in-sample and K out-of-sample observations, fit a model to the in-sample data and obtain rolling one-step forecasts of the VaR for the out-of-sample time points.

Define

I_t = 1,

if -r_t > \widehat{VaR}_t (\alpha) or

I_t = 0,

otherwise,

for t = n + 1, n + 2, ..., n + K as the hit sequence, where \alpha is the confidence level for the VaR (often \alpha = 0.95 or \alpha = 0.99). Furthermore, denote p = \alpha and let w be the actual covered proportion of losses in the data.

1. Unconditional coverage test:

H_{0, uc}: p = w

Let K_1 be the number of ones in I_t and analogously K_0 the number of zeros (all conditional on the first observation). Also calculate \hat{w} = K_0 / (K - 1). Obtain

L(I_t, p) = p^{K_0}(1 - p)^{K_1}

and

L(I_t, \hat{w}) = \hat{w}^{K_0}(1 - \hat{w})^{K_1}

and subsequently the test statistic

LR_{uc} = -2 * \ln \{L(I_t, p) / L(I_t, \hat{w})\}.

LR_{uc} now asymptotically follows a chi-square-distribution with one degree of freedom.

2. Conditional coverage test:

The conditional coverage test combines the unconditional coverage test with a test on independence. Denote by w_{ij} the probability of an i on day t-1 being followed by a j on day t, where i and j correspond to the value of I_t on the respective day.

H_{0, cc}: w_{00} = w{10} = p

with i = 0, 1 and j = 0, 1.

Let K_{ij} be the number of observations, where the values on two following days follow the pattern ij. Calculate

L(I_t, \hat{w}_{00}, \hat{w}_{10}) = \hat{w}_{00}^{K_{00}}(1 - \hat{w}_{00})^{K_{01}} * \hat{w}_{10})^{K_{10}}(1 - \hat{w}_{10})^{K_{11}},

where \hat{w}_{00} = K_{00} / K_0 and \hat{w}_{10} = K_{10} / K_1. The test statistic is then given by

LR_{cc} = -2 * \ln \{ L(I_t, p) / L(I_t, \hat{w}_{00}, \hat{w}_{10}) \},

which asymptotically follows a chi-square-distribution with two degrees of freedom.

3. Independence test:

H_{0,ind}: w_{00} = w_{10}

The asymptotically chi-square-distributed test statistic (one degree of freedom) is given by

LR_{ind} = -2 * \ln \{L(I_t, \hat{w}_{00}, \hat{w}_{10}) / L(I_t, \hat{w})\}.

—————————————————————————–

The function needs four inputs: the out-of-sample loss series obj$loss, the corresponding estimated VaR series obj$VaR, the coverage level obj$p, for which the VaR has been calculated and the significance level conflvl, at which the null hypotheses are evaluated. If an object returned by this function is entered into the R console, a detailed overview of the test results is printed.

Value

A list of class quarks with the following four elements:

p

probability p stated in the null hypotheses of the coverage tests

p.uc

the p-value of the unconditional coverage test

p.cc

the p-value of the conditional coverage test

p.ind

the p-value of the independence test

conflvl

the significance level at which the null hypotheses are evaluated

model

selected model for estimation; only available if a list returned by the rollcast function is passed to cvgtest

method

selected method for estimation; only available if a list returned by the rollcast) function is passed to cvgtest

References

Christoffersen, P. F. (1998). Evaluating interval forecasts. International economic review, pp. 841-862.

Kupiec, P. (1995). Techniques for verifying the accuracy of risk measurement models. The J. of Derivatives, 3(2).

Examples

prices <- DAX$price_close
returns <- diff(log(prices))
n <- length(returns)
nout <- 250 # number of obs. for out-of-sample forecasting
nwin <- 500 # window size for rolling forecasts
results <- rollcast(x = returns, p = 0.975, method = 'age', nout = nout,
                     nwin = nwin)
cvgtest(results)