Values at Risk
Uppsala University Hampus Eliasson Bachelor Thesis in Statistics
Fall 2017
Supervisor: Lars Forsberg
1
Abstract
This paper presents a small study on Value at Risk calculation for four different companies. Value at Risk, VaR, is a risk-metric, which measures the least amount that an asset can be expected to deviate from its value at a chosen probability level. A five and one percent VaR calculation is performed for the four American conglomerates chosen for the study, these are: Apple, Alphabet, Exxon Mobil and General Electric. The VaR is calculated in this paper via three different GARCH estimations for the conditional heteroscedasticity coupled with three different probability distributions for the error terms.
The GARCH models used are the Standard GARCH, the Exponential GARCH and the Threshold GARCH. Every GARCH model is tested with Normal, Students-T and Normal Inverse Gaussian errors. In order to evaluate the models, Kupiec’s test and Christoffersen’s test are conducted. Based on passing both tests for the one and five percent VaR simultaneously, the study yields several suitable model combinations for each company except for Apple. The suitable models seem to cluster around what probability distribution is used rather than what GARCH-model. The study lacks in capacity of comparing the efficiency of the suitable models to one another.
2
Contents
Introduction ... 3
Earlier research ... 3
Data and Descriptive Statistics ... 5
Theoretical Background ... 6
Value at Risk ... 6
Modelling the Conditional Heteroscedasticity ... 6
GARCH ... 7
EGARCH ... 7
TGARCH ... 8
Probability Functions ... 8
Kupiec’s Test ... 9
Christoffersen’s Test ... 9
Methodology ... 9
Comparative analysis ... 9
Results ... 10
Analysis and Conclusion ... 14
References ... 16
Appendix 1 ... 17
3
Introduction
In financial economics there exists a need for measuring and managing different sorts of risk. For instance, asset specific risk can be seen as an assets probability to decrease or increase in value and to what magnitude, clearly a study of interest for financial analysts. These changes in price are called volatility, and by studying an assets historical volatility, projections about future volatility can be made. These projections can be used to calculate the risk-metric Value at Risk or VaR which measures the minimum change in asset value which can be expected with a certain degree of confidence. For example, for a five percent VaR estimate, the Value at Risk estimation presents the lowest amount that the value of an asset will shift in the most extreme five percent of outcomes. This risk-metric gained popularity in the 1990s following changes in the Basel Accords prompting banks to perform their own risk calculations where the VaR metric was one of the options (Chang et al. 2011). This paper will evaluate different GARCH approaches for calculating the Value at Risk of the daily returns of stocks of four large American companies: Apple, Alphabet, General Electric and Exxon Mobil. These companies were chosen since they are large well-known companies with easily accessible financial data. In order to calculate the Value at Risk, the assets historical volatility is analyzed. Since volatility in financial assets has been shown to cluster and be serially dependent, the most recent information may be used in order to predict a portion of future volatility. By using an “Autoregressive Conditional Heteroscedasticity” or ARCH model (Engle, 1982), the volatility conditional on past values can be used to predict changes in future volatility. The ARCH models which will be evaluated in this paper are the Generalized ARCH or GARCH (Bollerslev 1986), the Exponential GARCH or EGARCH (Nelson 1991) and the Threshold GARCH or TGARCH (Zakoïan, 1994). Although the conditional part of the volatility can be predicted with the use of GARCH modeling, an unpredictable stochastic part still exists. This random part of the volatility has to be mapped with a probability distribution function, this paper makes use of the Normal, the Students T, and the Normal Inverse Gaussian distributions. In order to evaluate the models, Kupiecs Exceedance test and Christoffersen’s Exceedance Independence Test are performed on the models.
Earlier research
To provide a framework on how statistical time series behave, the first article presented is a survey of previous research. Rama Cont’s article Empirical Properties of Asset Resurns: Stylized Facts and Statistical Issues (2001) provides a good survey of reports on what has become known as stylized facts in financial analysis. These stylized facts can be seen as empirically proven general behaviors of financial assets. The most relevant of these properties for our VaR analysis are presented in short: The first one, ‘Fat tails’, the propensity for financial returns to have non-normally distributed returns with fatter tail risk than the normal distribution would suggest. This can lead to an underestimation of financial risk if not taken into account. Since Value at Risk deals with these tail end probabilities, it is only reasonable to include different probability distributions with higher tail risk in the study. The second “stylized fact” of importance for this paper is that of ‘asymmetric responses’ to positive or
4
negative returns. Simply put, the market seems react stronger towards negative than positive returns and thus higher volatility ought to be expected days following negative shocks than positive ones.
Hence it is reasonable to include a model which takes these asymmetric responses into account. The third one is ‘absence of serial correlation’, this characteristic tells us high returns today do not in any way indicate high returns tomorrow. Which brings us to the final stylized fact of importance for this paper ‘Volatility Clustering’. Contrary to returns, the volatility of those returns have been seen to serially correlated and prone to cluster. Thus a strong positive shock today does not alter the
probability of a positive shock tomorrow, it can however be shown that the probable level of intensity of tomorrows shock is affected. Robert Engle (1982) used this last characteristic with his ARCH model to forecast future volatility by using current data. For example, a sharp spike in volatility today increases the probability of having high volatility tomorrow. But as the third of these financial characteristics state, the direction of the shocks normally cannot be predicted by statistical methods, hence only a part of the intensity of tomorrows shock can be predicted. However by using the characteristic of asymmetric responses to positive and negative shocks this information may also be harnessed in order to predict the intensity of future volatility. The paper ”The use of GARCH models in VaR estimation” released in (2004) by Timotheos Angelidis, Alexandros Benos and Stavros
Degiannakis calculated and tested the VaR for 5 different portfolios constructed out of major stock indices. Highly similar to this paper they used the SGARCH, EGARCH and TGARCH coupled with the Normal, Students-t and GED distributions. The EGARCH and TGARCH models make use of directional information of previous shocks. However after conducting Kupiec´s and Christofferson´s tests no model stood out as significantly fitting more of the portfolios. However after comparing the significantly fitting models the asymmetrical EGARCH and TGARCH produced more accurate forecasts than the SGARCH, with EGARCH as the best performer. For the error distributions, the thicker tailed Students-T and the Generalized Error Distribution performed better than the Normal distribution. These results are well in line with the expectations garnered from studying the characteristics of financial time-series.
5
Data and Descriptive Statistics
The companies chosen for this analysis are the well-known American industrial giants General Electric and Exxon Mobil as well as the tech giants Apple and Alphabet, the parent company of Google. The time series used in this paper have been downloaded from Google Finance via the
“Quantmod” package of the R Studio statistical software. The data has been gathered from 2007-01-03 up to 2017-12-13. The returns which will be analyzed are the log-returns (not dividend adjusted) constructed by taking the logged difference between the closing prices at time t and t-1. The companies are different from one another in what they produce, however all of them can be seen as relatively large.
Apple Log Returns: 2007-01-03 to 2017-12-13 Alphabet Log Returns: 2007-01-03 to 2017-12-13
Exxon Mobil Log Returns: 2007-01-03 to 2017-12-13 General Electric Log Returns: 2007-01-03 to 2017-12-13 Figure 1: Time series on logged daily returns scaled by 100
6
Theoretical Background
Value at RiskValue at Risk or VaR is a risk-metric, which means that it measures a certain type of risk, which is the minimum amount an asset is expected to change in value given previous market volatility. It is
calculated upon a chosen percentage level which shows the tail end risk. This percentage level is chosen by the investor but is usually 5 and 1 percent. Basically the VaR can be seen as a one sided confidence interval indicating what level the investor can expect the asset not to drop 1 – 𝛼 amount of the time. In other words, given an 𝛼 of 5 percent, the investor can expect the stock to stay above the 5 percent loss quantile 95 percent of the time. A VaR forecast based on GARCH estimates results in a dynamic estimation due to the GARCH models ability to constantly change the volatility forecast (Jondeau, Poon and Rockinger 2007). By inspecting the graphs in Figure 2 one might quickly spot the difference in the red dots indicating VaR forecast exceedances. The plot with an 𝛼 = 1% has
significantly less exceedances than its 𝛼 = 5% counterpart. This is fully reasonable and expected as the
‘safety bar’ has been lowered.
𝛼 = 1% VaR Exceedance plot of Apple Returns 𝛼 = 5% VaR Exceedance plot of Apple Returns Figure 2
Modelling the Conditional Heteroscedasticity
Robert F. Engle introduced the “Autoregressive Conditional Heteroscedasticity”, ARCH model, for time series analysis in (1982) and was awarded the Sveriges Riksbanks Prize in Economic Sciences in Memory of Alfred Nobel for this in 2003 (Nobelprize.org 2003). The ARCH model comes into use when there is a timewise correlation in an assets volatility. Where just like a regular MA(p) time series model the previous shocks are used in order to predict the future value of a series. The difference with the ARCH model being that it attempts to predict part of the volatility of the series instead of the real value. Engle showed that since the volatility 𝜖 is serially correlated it can be written as a function, part conditional on past values, the 𝜎 and a strictly random element z such as 𝜖𝑡= 𝜎𝑡𝑧𝑡.
Equation 1.
𝜎𝑡2= 𝜔 + ∑ 𝛼𝑖
𝑞
𝑖=1
𝜖𝑡−𝑖2
7
Where 𝜎 is the conditional volatility at time t dependent on the previous squared returns 𝜖. The constants, 𝛼 explains how much of the conditional volatility at time t depends on each 𝜖 term.
(Jondeau, Poon and Rockinger 2007) By examining the time series presented in Figure 1 one can clearly see the clustering tendencies of the returns. The most obvious cluster is the one at the financial crisis of 08, this forceful increase in volatility can be seen in each time serie.
This paper uses variations of the GARCH model developed from the ARCH model by Tim Bollerslev (1986), which possesses similar qualities to the ARCH model. Three variation are used, the Standard GARCH, the Exponential GARCH and the Threshold GARCH. The latter variants are included since they allow for asymmetric responses between positive contra negative shocks.
GARCH
Since there is a long persistence in volatility the ARCH model needs an inclusion of many of the previous shocks in order to be a good fit. The “Generalized Autoregressive Conditional
Heteroscedasticity” or GARCH model by Tim Bollerslev (1986) reduces the amount of previous shocks necessary to include in an analysis by introducing the previous day forecast into the equation.
Equation 2
𝜎𝑡2 = 𝜔 + ∑ 𝛼𝑖
𝑞
𝑖=1
𝜖𝑡−𝑖2 + ∑ 𝛽𝑗𝜎𝑡−𝑗2
𝑝
𝑗=1
Hence the forecasted conditional volatility 𝜎 is a function of its previous value and the shock of the previous day. Since the shock, defined as 𝜖𝑡 = 𝜎𝑡𝑧𝑡 is a function of the forecast conditional variance and the random element 𝑧𝑡 it becomes clear that the forecasted conditional variance is a product of its last projection and the deviation from it. The GARCH process is covariance stationary iff ∑𝑝𝑖=1𝛼𝑖+
∑𝑞𝑗=1𝛽𝑗< 1, a sufficient condition for the error term 𝜖 to be strictly stationary. (Jondeau, Poon and Rockinger 2007)
EGARCH
The Exponential GARCH (Nelson, 1991) or EGARCH model is an asymmetric model which takes consideration of both the sign and the size of the shocks and can thus predict different effects of negative returns than positive returns. Again let us define 𝜖𝑡 = 𝜎𝑡𝑧𝑡, that is the shock is equal to the conditional volatility times a random error.
Equation 3
𝑙𝑜𝑔 𝜎𝑡2= 𝜔 + ∑ 𝛼𝑖 𝑞
𝑖=1
𝑔(𝑧𝑡−𝑖) + ∑ 𝛽𝑗log 𝜎𝑡−𝑗2
𝑝
𝑗=1
8 Equation 4
𝑔(𝑧𝑡) = {(𝜃 + 𝛾)𝑧𝑡− 𝛾𝐸(|𝑧𝑡|) 𝑖𝑓 𝑧𝑡 ≥ 0 (𝜃 − 𝛾)𝑧𝑡− 𝛾𝐸(|𝑧|) 𝑖𝑓 𝑧𝑡 < 0
The function 𝑔(𝑧𝑡−𝑖) = 𝜃𝑧𝑡+ 𝛾[|𝑧𝑡| − 𝐸|𝑧𝑡|] is easier understood rewritten in Equation 4 where the asymmetric response to a negative or positive shock becomes clear. 𝜃 and 𝛾 are constants determining the asymmetric responses and linearity. The term E(|𝑧𝑡|) is the expectation of the absolute value of 𝑧𝑡. Contrary to a probability distribution with mean 0 the expectation of an absolute value follows a folded probability distribution from the one which 𝑧 is assumed to follow since every value on the positive side becomes twice as likely when the negatives are transformed. For example the expectation of a folded Normal(0,1) distribution would be √2
𝜋 instead of 0 (Rachev et al. 2007).
TGARCH
The Threshold GARCH or TGARCH (Zakoïan, 1994) is an interesting model which models the volatility rather than the squared volatility. Like the EGARCH it is also allows for asymmetric responses between positive and negative shocks.
Equation 5
𝜎𝑡 = 𝜔 + ∑[𝛼𝑖|𝜖𝑡−𝑖| + 𝛾𝑖
𝑝
𝑖=1
Πt−i− |𝜖𝑡−𝑖|] + ∑ 𝛽𝑗
𝑞
𝑗=𝑖
𝜎𝑡−𝑗
The alpha term indicates how much the past absolute errors contribute to volatility and the gamma term indicates how much difference it makes if the shock is negative. Π𝑡− is an indicator which takes the value 1 if 𝜖𝑡 < 0 but takes the value 0 otherwise.
Probability Functions
The probability functions chosen for this paper are the Normal distribution, the Students T distribution and the Normal Inverse Gaussian distribution. The normal distribution was included to since it is the go-to distribution for new statistics students and the distribution that Robert F.Engle assumed in his groundbreaking work on ARCH models (1982). The Students t-distribution was chosen on the same premise as the normal-distribution, it also being well known and adopted. Moreover in (1987) Tim Bollerslev released a paper where he showed that the t-distribution better handles the leptokurtic qualities of financial returns than the normal distribution due to its fatter tails. The last distribution assumed, the Normal Inverse-Gaussian Distribution or NIG, also allows for fatter tail probabilities and has been shown as very successful at modeling financial data. For example a Paper released (2000) by Erik Bølviken and Fred Espen Benth showed that the NIG-distribution fit the log-return data of several Norwegian stocks ‘nearly perfectly’. Thus including the NIG-distribution will examine how the workhorses of basic econometrics, the Normal and Students t-distributions fare against a more leptokurtic distribution, the NIG-distribution.
9 Kupiec’s Test
One of the earliest tests for Value at Risk modeling performance efficiency was Kupiec’s PF Coverage Test (Kupiec, 1995). This test is a likelihood ratio test which examines whether or not the total amount of exceedances by a back tested VaR model is feasible with regards to the chosen 𝛼. Simply put, it tests how likely it is that the VaR model is correct given the actual result.
Christoffersen’s Test
Christoffersen’s (1998) Exceedance Independence Test examines whether or not the VaR exceedances are independent from one another. Just like Kupiec’s test it is a likelihood ratio test. The test is
relevant since the application of GARCH models to the VaR estimates attempt to account for volatility clustering. Therefor the Exceedance Independence test gives a hint of whether or not the model raises the Value at Risk estimate quickly enough in times of turbulence.
Methodology
The VaR estimation and subsequent Kupiec and Christoffersen tests were all conducted using the Rugarch package in R. The GARCH models were all of the order (1, 1) that is they all used a single lag for the ARCH term and a single lag for the GARCH term, this was chosen for simplicity and in order to reduce the number of models to be estimated. The estimation window was set to 500 observations for every test, by the reasoning that in order to properly forecast the rare 1% events, a few of these events have to be included within the estimation window. The forecast length was set to +1 since any forecasts beyond that have such large variances making them difficult to use
productively. The estimation window was set to refit every 50 days in order to refresh the parameter estimations as market conditions changed. The Value at Risk alphas were set to 1 percent and 5 as is common practice (Jondeau, Poon and Rockinger 2007).
Comparative analysis
In order to assess the efficiency of the various Value at Risk prediction models for the four companies, the Kupiec’s Test and Christoffersen’s Exceedance Independace Test are performed. The first test,
“Kupiec’s test” tests whether or not an estimation model might be correct based upon the number of exceedances of the Value at Risk boundary the model exhibits (Kupiec, 1995). The second test,
“Christoffersen’s Exceedance Independence test”, tests whether or not the exceedances are
independent from one another (Christoffersen 1998), which indicates if the model is suitably flexible to swings in volatility. The probabilities of these test results are presented in the results section and will serve as a basis for approving or disapproving the models.
10
Results
Examining the time series plots in Figure 3 gives us a picture of the behavioral differences of the different stocks. Apple and especially Alphabet seem prone to sharp turns with many spikes in volatility both positive and negative. General Electric looks more stable and Exxon Mobil hardly exhibits any spikes in volatility at all apart from the crash in 2008.
Apple Log Returns: 2007-01-03 to 2017-12-13 Alphabet Log Returns: 2007-01-03 to 2017-12-13
Exxon Mobil Log Returns: 2007-01-03 to 2017-12-13 General Electric Log Returns: 2007-01-03 to 2017-12-13 Figure 3: Time series on logged daily returns scaled by 100
In the following tables the Standard GARCH is abbreviated SG, the Exponential GARCH, EG and the Threshold GARCH, TG. The probability distributions are abbreviated Norm for Normal, Std for Students T and Nig for the Normal Inverse Gaussian Distribution. Combinations of the models and distributions are written for example SG-Norm for the Standard Gaussian model paired with assumed Normally-distributed errors.
11
Significant difference from 𝐻0 at 1, 5 and 10 percent level respectively = *, **, ***.
Apple,
sGARCH(1,1)
Std 𝛼 = 1%
Std 𝛼 = 5%
Norm 𝛼 = 1%
Norm 𝛼 = 5%
Nig 𝛼 = 1%
Nig 𝛼 = 5%
Exceedances 0,88 3,72 1,15 3,15 0,80 3,06
Kupiec P 58, 21 0,37* 47,59 0,00* 31,82 0,00*
Christofferson P
26,84 66,09 30,37 39,01 80,19 67,20
Apple,
eGARCH(1,1)
Std 𝛼 = 1%
Std 𝛼 = 5%
Norm 𝛼 = 1%
Norm 𝛼 = 5%
Nig 𝛼 = 1%
Nig 𝛼 = 5%
Exceedances 0,88 3,23 0,93 2,97 0,75 3,05
Kupiec P 58,21 0,00* 73,99 0,00* 21,94 0,00*
Christofferson P
75,45 12,79 59,25 4,76** 72,23 13,50
Apple, tGARCH(1,1)
Std 𝛼 = 1%
Std 𝛼 = 5%
Norm 𝛼 = 1%
Norm 𝛼 = 5%
Nig 𝛼 = 1%
Nig 𝛼 = 5%
Exceedances 0,79 3,28 0,93 2,88 0,75 2,83
Kupiec P 31,82 0,00* 73,99 5,87*** 21,94 3,21**
Christofferson P
44,20 4,81** 59,25 2,77** 42,08 2,34**
Table 1 Apple test results, showing exceedances in percent of the VaR models and the p values of Kupiec’s and Christoffersen’s tests.
Alphabet, sGARCH(1,1)
Std 𝛼 = 1%
Std 𝛼 = 5%
Norm 𝛼 = 1%
Norm 𝛼 = 5%
Nig 𝛼 = 1%
Nig 𝛼 = 5%
Exceedances 0,57 4,21 1,01 2,83 0,48 3,90
Kupiec P 2,81** 7,82** 92,43 3,21** 0,66* 1,30**
Christofferson P
63,12 90,07 79,20 48,10 22,76 97,75
Alphabet, eGARCH(1,1)
Std 𝛼 = 1%
Std 𝛼 = 5%
Norm 𝛼 = 1%
Norm 𝛼 = 5%
Nig 𝛼 = 1%
Nig 𝛼 = 5%
Exceedances 0,93 4,39 0,97 3,01 0,84 3,76
Kupiec P 73,99 17,52 90,69 0,00* 43,99 0,51*
Christofferson P
96,57 77,24 98,16 98,89 87,15 89,66
Alphabet, tGARCH(1,1)
Std 𝛼 = 1%
Std 𝛼 = 5%
Norm 𝛼 = 1%
Norm 𝛼 = 5%
Nig 𝛼 = 1%
Nig 𝛼 = 5%
Exceedances 0,88 4,21 0,93 3,01 0,75 3,68
Kupiec P 58,21 7,82*** 73,99 0,00* 21,94 0,26*
Christofferson P
94,61 84,75 88,95 94,60 63,87 91,72
Table 2 Alphabet test results, showing exceedances in percent of the VaR models and the p values of Kupiec’s and Christoffersen’s tests.
12 Exxon Mobil,
sGARCH(1,1)
Std 𝛼 = 1%
Std 𝛼 = 5%
Norm 𝛼 = 1%
Norm 𝛼 = 5%
Nig 𝛼 = 1%
Nig 𝛼 = 5%
Exceedances 0,88 4,92 1,11 4,43 0,79 4,12
Kupiec P 58,07 86,16 61,18 20,77 31,72 4,88
Christofferson P
80,35 64,53 42,91 55,94 91,72 49,56
Exxon Mobil, eGARCH(1,1)
Std 𝛼 = 1%
Std 𝛼 = 5%
Norm 𝛼 = 1%
Norm 𝛼 = 5%
Nig 𝛼 = 1%
Nig 𝛼 = 5%
Exceedances 1,01 4,87 1,19 4,43 0,66 4,07
Kupiec P 92,60 78,59 36,21 20,77 8,84*** 3,81**
Christofferson P
29,88 6,20 17,38 15,77 14,51 6,72***
Exxon Mobil, tGARCH(1,1)
Std 𝛼 = 1%
Std 𝛼 = 5%
Norm 𝛼 = 1%
Norm 𝛼 = 5%
Nig 𝛼 = 1%
Nig 𝛼 = 5%
Exceedances 0,79 4,65 1,15 4,25 0,62 3,81
Kupiec P 31,72 44,60 47,72 9,61*** 5,14*** 0,69*
Christofferson P
76,33 8,60*** 34,84 19,62 4,66** 5,99***
Table 3 Exxon test results, showing exceedances in percent of the VaR models and the p values of Kupiec’s and Christoffersen’s tests.
General Electric, sGARCH(1,1)
Std 𝛼 = 1%
Std 𝛼 = 5%
Norm 𝛼 = 1%
Norm 𝛼 = 5%
Nig 𝛼 = 1%
Nig 𝛼 = 5%
Exceedances 0,84 5,36 1,24 4,74 0,75 4,96
Kupiec P 43,87 43,34 26,72 57,21 21,86 93,83
Christofferson P
67,47 69,96 3,03** 66,92 58,10 86,06
General Electric, eGARCH(1,1)
Std 𝛼 = 1%
Std 𝛼 = 5%
Norm 𝛼 = 1%
Norm 𝛼 = 5%
Nig 𝛼 = 1%
Nig 𝛼 = 5%
Exceedances 0,93 4,74 1,41 4,34 0,84 4,74
Kupiec P 73,84 57,21 6,02*** 14,39 43,87 57,21
Christofferson P
4,51** 72,16 8,23*** 90,79 33,66 99,10
General Electric, tGARCH(1,1)
Std 𝛼 = 1%
Std 𝛼 = 5%
Norm 𝛼 = 1%
Norm 𝛼 = 5%
Nig 𝛼 = 1%
Nig 𝛼 = 5%
Exceedances 0,97 5,00 1,37 4,43 0,84 4,60
Kupiec P 90,52 98,45 9,10*** 20,77 43,87 38,92
Christofferson P
2,68** 99,13 4,35** 84,42 39,77 87,21
Table 4 General Electric test results, showing exceedances in percent of the VaR models and the p values of Kupiec’s and Christoffersen’s tests.
13
Company VaR 𝛼 = 1% VaR 𝛼 = 5%
Apple All models pass tests. TG-Norm
Alphabet SG-Norm, EG-Std, EG-Norm, EG- Nig, TG-Std, TG-Norm, TG-Nig
SG-Std, EG-Std, TG-Std
Exxon Mobil All pass SG-Std, SG-Norm, EG-Std, EG-Norm,
TG-Std, TG-Norm
General Electric All pass All pass
Table 5 Models that pass Kupiec’s test at 5% level.
The first test performed, Kupiecs test of whether the model exhibits the correct amount of exceedances or not yielded varying results for the different companies. Interestingly enough all models pass the test for Apple with a 1% 𝛼. However for the 5% 𝛼 only the TG-Norm combination proved successful. For Alphabet most models passed the 1% VaR but only the models with assumed t-distributed errors passed the 5% VaR test. For Exxon Mobil all models passed the tests save for the ones with assumed Normally Inverse Gaussian distributed errors at 5% VaR. For General Electric any model seems to do.
Company VaR 𝛼 = 1% VaR 𝛼 = 5%
Apple All pass SG-Std, SG-Norm, SG- Nig, EG-Std,
EG-Nig
Alphabet All pass All pass
Exxon Mobil All pass save for TG-Nig All pass General Electric SG-Std, SG-Nig, EG-Norm, EG-
Nig, TG-Nig
All pass Table 6 Models that pass Christoffersen’s test at 5% level.
Christoffersen’s Exceedance Independence Test show that the exceedances of Apple are all
independent from one another at the 1% VaR. For the 5% VaR however only the Standard GARCH models and two Exponential GARCH models, the EG-Std and the EG-Nig pass the independence test.
For Alphabet all of the models pass the test. All models pass the test save for the Threshold GARCH with NIG distributed standard errors at the 1% VaR. For the 5% VaR all models pass. The models which fit for General Electric with the 1% VaR are all of the models paired with the Nig errors as well as the Standard GARCH with t-distributed errors and the Exponential GARCH with normal-
distributed errors. For the 5% VaR all of the models fit.
Company VaR 𝛼 = 1% VaR 𝛼 = 5%
Apple All pass None
Alphabet SG-Norm, EG-Std, EG-Norm, EG- Nig, TG-Std, TG-Norm, TG-Nig
SG-Std, EG-Std, TG-Std
Exxon Mobil All pass save for TG-Nig SG-Std, SG-Norm, EG-Std, EG-Norm, TG-Std, TG-Norm
General Electric SG-Std, SG-Nig, EG-Norm, EG- Nig, TG-Nig
All pass Table 7 Models that pass both Kupiec’s and Christoffersen’s tests at the 5% level.
14
The models which passed both the ‘Kupiec’s Test’ of the proportion of exceedances and the
‘Christoffersen’s Exceedance Independence Test’ were: For Apple at the 1% VaR, all of the models passed but none at the 5% VaR level. For Alphabet all but the Standard GARCH paired with the Nig and T-distributions passed. At the 5% level however all of the GARCH models proved succesfull but only if performed with assumed T-distributed errors. Exxon Mobil was properly modeled according to both tests at the 1% VaR level with all models and distributions except for the Threshold GARCH with assumed NIG-errors. For the 5% VaR all of the models passed both tests except for those with assumed NIG-errors. For General Electric on the other hand all of the models with NIG-errors passed the tests for a 1% VaR as well as the SGARCH with T-distributed errors and the EGARCH with Normally-distributed errors. At the 5% VaR all of the models passed both tests.
Company
Apple None
Alphabet EG-Std, TG-Std
Exxon Mobil SG-Std, SG-Norm, EG-Std, EG-Norm, TG-Std, TG-Norm General Electric SG-Std, SG-Nig, EG-Norm, EG-Nig, TG-Nig
Table 8 Models which passed both tests for both 1% and 5% VaR estimations.
No GARCH-Probability function combination properly managed to capture both the 1% and 5% VaR for Apples returns. For Alphabet the Exponential GARCH and the Threshold GARCH paired with student T-distributions successfully modelled the VaR at both 1 and 5% with a suitable number of exceedances and exceedance independence. Exxon Mobil was properly modeled by all of the models except those using NIG-distributed errors. General Electric on the other hand was properly modeled for both 1 and 5% VaR estimates by any model paired with the NIG-distribution as well as the combination SGARCH with T-distributed errors and the EGARCH with Normally-distributed errors.
Analysis and Conclusion
The Value at Risk for the two tech giants proved to be more difficult to model than the petroleum producer Exxon Mobil or the conglomerate General Electric. A quick examination of the time series plots for the companies gives a hint of the tech stocks being more unpredictable and volatile with more sharp spikes in volatility. The tests did not yield a single valid candidate model which could accurately predict Apples Value at Risk for alphas of 1 and 5 percent simultaneously. For Alphabet the Students-t distribution combined with either of the direction sensitive EGARCH or TGARCH models proved to be effective predictive models. The Value at Risk of Exxon Mobil was accurately predicted by any one of the SGARCH, EGARCH or TGARCH models given an assumption of either Students-t or Normally-distributed errors however no model with assumed Normal Inversed Gaussian errors was successful. General Electric on the other hand gave clear results favoring the Normal Inverse Gaussian distribution where all of the models with assumed NIG errors were significant and only one model each for the Students-t and Normal distributions.
15
Out of the probability distributions tested, the Students-T distribution yielded the most successfull models with 6, wheras only 4 combinations with normally distributed and 3 combinations with NIG distributed errors were successful. That a fat-tailed distribution won the bout is as should be expected since the error distributions seen in financial timeseries are usually fat-tailed as per the stylized facts presented by Rama Cont (2001). However as pertains to the GARCH models, the EGARCH model scored 5 successful models wheras the other scored 4 each. These small differences can hardly be taken as evidence of any models superiority over another for a broader population of companies.
Although one interesting finding is a clear clustering of results, which seem to indicate that the probability distribution assumed is more important for successful modeling than what GARCH model is used. The asymmetric EGARCH and TGARCH models did not stick out as being particularly more efficient than the Standard GARCH in this small survey. This result is similar to the one that
Timotheos Angelidis, Alexandros Benos and Stavros Degiannakis (2004) presented when they only used Kupiec´s and Christofferson´s tests for evaluating fit. The very differing results for each company suggests that it is highly unlikely that a single model would fit every stock but rather that similar surveys would have to be performed on a case by case basis in order to properly calculate their VaR estimates.
16
References
Bollerslev, Tim (1987) A Conditionally Heteroscedastic Time Series Model for Speculative Prices and Rates of Return. The Review of Economics and Statistics Vol 69 (No 3) pp 542-547.
Bølviken, Erik and Benth, Fred Espen (2000) Quantification of Risk in Norwegian Stocks via the Normal Inverse Gaussian Distribution. Proceedings AFIR 2000. Tromsø, Norway
Chia‐lin Chang, Juan‐Ángel Jiménez‐Martín, Michael McAleer, Teodosio Pérez‐Amaral, (2011) "Risk management of risk under the Basel Accord: forecasting value‐at‐risk of VIX futures", Managerial Finance, Vol. 37 Issue: 11, pp.1088-110
Christoffersen, Peter (1998). Evaluating interval forecasts. International Economic Review. Vol. 39 (No. 4) pp 841-862.
Cont, Rama (2001) Empirical properties of asset returns: stylized facts and statistical issues.
Quantitative Finance Vol. 1 pp 223-236
Engle, Robert F. and Bollerslev, Tim (1986) Modelling the Persistence of Conditional Variances.
Econometric Reviews Vol 5 1986 (No. 1) pp 1-50.
Engle, Robert F. (1982) Autoregresive Conditional Heteroscedasticity with estimates of the Variance of United Kingdom Inflation. Econometrica. Vol. 50 (No. 4)
Joneau, Eric, Poon, Ser-Huang, Rockinger, Michael (2007) Financial Modeling Under Non-Gaussian Distributions. London: Springer
Kupiec, Paul H. (1995). Techniques for Verifying the Accuracy of Risk Measurement Models. The J. of Derivatives. Vol. 3 (No. 2)
Nelson, Daniel B. (1991) Conditional Heteroscedasticity in Asset Returns: A New Approach.
Econometric Society. Vol. 59 (No. 2) pp 347-370.
Nobelprize.org (2003) The Prize in Economic Sciences 2003 - Press Release
http://www.nobelprize.org/nobel_prizes/economic-sciences/laureates/2003/press.html [2018-01- 02]
Timotheos Angelidis, Alexandros Benos, Stavros Degiannakis, (2004) ”The use of GARCH models in VaR estimation”, Statistical Methodology, Vol. 1, pp.105-128.
Rachev, Svetlozar T., Mittnik, Stefan, Fabozzi Frank J., Focardi, Jai, Teo (2007) Financial Econometrics From Basic to advanced modeling techniques, Hoboken: Wiley
Zakoïan, Jean-Michel (1994) Threshold heteroskedastic models Journal of Economic Dynamics and Control. Vol. 18 (No. 5) pp 931-955
17
Appendix 1
Parameter estimations for VaR models. Coded as follows, AP: Apple, Alp: Alphabet, Ge: General Electric, Ex: Exxon Mobil. SG: Standard Garch, EG: Exponential Garch, TG: Threshold Garch. Norm:
Normal distribution, STD: Students-T, NIG: Normal Inverse Gaussian.
AP SG Norm omega alpha1 beta1
1 0,306295 0,117368 0,852631
2 0,289923 0,111899 0,860874
3 0,283519 0,104876 0,864882
4 0,232039 0,101662 0,87201
5 0,144027 0,094668 0,888518
6 0,155467 0,088604 0,891508
7 0,161511 0,089527 0,888584
8 0,18896 0,09512 0,879193
9 0,176565 0,091195 0,882734
10 0,142276 0,084339 0,893445
11 0,103314 0,081919 0,901413
12 0,085867 0,079279 0,906624
13 0,07682 0,079435 0,907511
14 0,070917 0,075933 0,911583
15 0,088311 0,082912 0,902009
16 0,087937 0,081006 0,903246
17 0,094326 0,081068 0,901195
18 0,1091 0,086735 0,893265
19 0,108737 0,084589 0,894632
20 0,098549 0,081638 0,89886
21 0,09449 0,078428 0,903627
22 0,092832 0,076696 0,906623
23 0,096978 0,076547 0,905554
24 0,099123 0,076948 0,904054
25 0,099149 0,074594 0,906015
26 0,093776 0,076166 0,904987
27 0,105785 0,074991 0,903243
28 0,131753 0,080346 0,891953
29 0,12363 0,08368 0,89002
30 0,114952 0,078589 0,896195
31 0,10898 0,077542 0,898034
32 0,111094 0,077534 0,897486
33 0,110002 0,077923 0,896885
34 0,108792 0,080962 0,89411
35 0,112934 0,082448 0,891877
36 0,114504 0,082183 0,891371
37 0,117041 0,082789 0,890319
38 0,119638 0,080305 0,891641
39 0,116017 0,080681 0,891484
40 0,128347 0,082191 0,88658
41 0,124604 0,084127 0,885156
42 0,124536 0,082293 0,886202
43 0,108596 0,082884 0,889524
44 0,106879 0,080629 0,891624
45 0,10799 0,080662 0,890925
46 0,103081 0,079111 0,893261
Ap SG STD omega alpha1 beta1 shape
1 0,280452 0,106569 0,865185 7,552712
2 0,250252 0,099732 0,876185 8,628924
3 0,241469 0,091852 0,881578 8,940503
4 0,191784 0,086847 0,890524 8,612583
5 0,099771 0,078179 0,910045 7,932404
6 0,106238 0,071994 0,914422 7,786365
7 0,109212 0,073199 0,912212 7,671384
8 0,116409 0,084319 0,902032 7,148482
9 0,105678 0,080005 0,905979 7,165117
10 0,081673 0,073778 0,915034 6,948246
11 0,045744 0,073736 0,921408 6,558602
12 0,050955 0,074028 0,919727 6,703125
13 0,047392 0,073932 0,919827 6,785395
14 0,052095 0,071887 0,920115 7,192561
15 0,063707 0,077296 0,9133 7,179022
18
16 0,062672 0,07448 0,915561 7,193273
17 0,058988 0,071878 0,918469 6,825677
18 0,065508 0,072168 0,917326 6,584304
19 0,066429 0,072347 0,91654 6,405191
20 0,064916 0,070642 0,917642 6,626739
21 0,067776 0,070185 0,918271 6,510747
22 0,067828 0,065532 0,922754 6,084107
23 0,070254 0,065041 0,922323 6,173944
24 0,069931 0,064951 0,922061 6,055032
25 0,068577 0,062237 0,924872 5,951644
26 0,065729 0,063349 0,923726 6,009311
27 0,063125 0,057996 0,929333 5,746969
28 0,061039 0,060029 0,928134 5,438589
29 0,057945 0,062331 0,926141 5,466586
30 0,047567 0,057502 0,933312 5,277323
31 0,046654 0,05625 0,934185 5,379759
32 0,051165 0,05753 0,931924 5,448957
33 0,05259 0,057546 0,931035 5,56857
34 0,05503 0,061107 0,927011 5,565364
35 0,058077 0,062753 0,925067 5,493149
36 0,058236 0,061514 0,925926 5,496248
37 0,060546 0,061978 0,925014 5,532824
38 0,062639 0,062314 0,924129 5,413505
39 0,059899 0,062465 0,924232 5,38935
40 0,061753 0,063307 0,923223 5,226812
41 0,059137 0,064321 0,922661 5,206073
42 0,037897 0,058977 0,934009 4,912027
43 0,022132 0,055736 0,941592 4,790828
44 0,022877 0,054237 0,942826 4,723069
45 0,022476 0,05286 0,944099 4,713116
46 0,02096 0,052304 0,944997 4,689189
Ap SG Nig omega alpha1 beta1 skew shape
1 0,271749 0,1024 0,869902 -0,23344 2,801577
2 0,235761 0,096505 0,881012 -0,21243 3,405778
3 0,223342 0,089295 0,886888 -0,21718 3,546147
4 0,185507 0,084791 0,89439 -0,20256 3,182571
5 0,102629 0,077378 0,911476 -0,20249 2,782664
6 0,112232 0,071524 0,914814 -0,16633 2,605645
7 0,110577 0,070257 0,91569 -0,17222 2,478097
8 0,126232 0,081177 0,904094 -0,1436 2,170659
9 0,116294 0,077628 0,907082 -0,13868 2,210168
10 0,09151 0,07222 0,915448 -0,13589 2,088476
11 0,055146 0,07199 0,921703 -0,12335 1,905525
12 0,054335 0,071394 0,922412 -0,14206 1,904541
13 0,04913 0,071241 0,922729 -0,13959 1,974012
14 0,054265 0,069744 0,922501 -0,13367 2,147448
15 0,065635 0,075072 0,915727 -0,1279 2,141287
16 0,064608 0,072618 0,917535 -0,12147 2,160172
17 0,062339 0,070186 0,92031 -0,13222 1,946823
18 0,069364 0,071642 0,917645 -0,10605 1,863893
19 0,070262 0,072006 0,916588 -0,10537 1,800968
20 0,068414 0,0705 0,917776 -0,11157 1,884465
21 0,069994 0,069543 0,918866 -0,10519 1,849406
22 0,069401 0,065051 0,923139 -0,10602 1,709338
23 0,072499 0,064955 0,922259 -0,11422 1,753776
24 0,072475 0,065185 0,921454 -0,10054 1,703756
25 0,071313 0,062684 0,924016 -0,10433 1,661657
26 0,06905 0,063878 0,922719 -0,10546 1,68626
27 0,067426 0,058609 0,927841 -0,10939 1,598871
28 0,069566 0,060893 0,92514 -0,09432 1,457398
29 0,066554 0,063375 0,922992 -0,09506 1,465808
30 0,054728 0,058158 0,930759 -0,10329 1,390271
31 0,053228 0,056827 0,931931 -0,10586 1,436587
32 0,057786 0,058166 0,929681 -0,10482 1,454724
33 0,059223 0,058415 0,92856 -0,10094 1,509052
34 0,061296 0,062117 0,924464 -0,10178 1,509615
35 0,064189 0,0635 0,922654 -0,10246 1,482242
36 0,06415 0,062269 0,923476 -0,09649 1,481956
37 0,066292 0,062586 0,922697 -0,09748 1,50153
38 0,068239 0,06255 0,922054 -0,09889 1,458007
19
39 0,065608 0,063071 0,921839 -0,10382 1,440473
40 0,069222 0,064289 0,919798 -0,09428 1,367699
41 0,066179 0,06561 0,919051 -0,09633 1,353779
42 0,047274 0,059721 0,929979 -0,09459 1,226565
43 0,030333 0,05598 0,938126 -0,08611 1,18111
44 0,02961 0,053706 0,940426 -0,09068 1,154434
45 0,029047 0,052317 0,94173 -0,0895 1,152599
46 0,02724 0,05168 0,942733 -0,08699 1,143515
Ap TG Norm omega alpha1 beta1 eta11
1 0,180715 0,086261 0,871457 0,847349
2 0,175156 0,082026 0,876823 0,870442
3 0,166155 0,072457 0,886957 0,963364
4 0,156576 0,075406 0,887102 0,968461
5 0,129607 0,079976 0,892184 0,872457
6 0,1263 0,079393 0,893149 0,854652
7 0,130884 0,084667 0,88641 0,817514
8 0,151275 0,094154 0,871789 0,828805
9 0,15111 0,096318 0,868536 0,792561
10 0,140502 0,097954 0,870477 0,760016
11 0,124498 0,102256 0,872415 0,669938
12 0,112735 0,104168 0,875098 0,636062
13 0,105775 0,109848 0,872433 0,571643
14 0,097373 0,106896 0,877497 0,543323
15 0,107838 0,111671 0,869386 0,561745
16 0,105144 0,109804 0,871254 0,537431
17 0,114269 0,115595 0,863032 0,552825
18 0,124568 0,120121 0,855347 0,569479
19 0,124656 0,118909 0,855817 0,585547
20 0,12004 0,117755 0,858137 0,58004
21 0,116004 0,111265 0,864783 0,598125
22 0,107732 0,105947 0,873035 0,592577
23 0,108923 0,10561 0,872349 0,592899
24 0,108632 0,108212 0,869796 0,576489
25 0,113794 0,1091 0,866756 0,592427
26 0,109215 0,110366 0,867288 0,575574
27 0,108714 0,10946 0,868032 0,568855
28 0,111893 0,113144 0,863924 0,573524
29 0,102916 0,113487 0,867504 0,559375
30 0,102875 0,110158 0,869581 0,558516
31 0,100948 0,110156 0,870123 0,554997
32 0,101637 0,108883 0,87069 0,567321
33 0,099355 0,108707 0,871327 0,556001
34 0,095779 0,11126 0,870743 0,52616
35 0,096493 0,109845 0,871372 0,538218
36 0,097098 0,109465 0,871075 0,535557
37 0,094443 0,106728 0,87412 0,529881
38 0,095277 0,1049 0,875042 0,550762
39 0,094336 0,107036 0,873379 0,536101
40 0,100658 0,110601 0,867779 0,55965
41 0,097276 0,112749 0,867411 0,540246
42 0,105273 0,110515 0,864965 0,585699
43 0,0942 0,10977 0,870895 0,570139
44 0,095308 0,108662 0,870786 0,563666
45 0,095672 0,109144 0,870042 0,559682
46 0,094378 0,108614 0,87077 0,555273
AP TG STD omega alpha1 beta1 eta11 shape
1 0,148243 0,08717 0,882366 0,714029 8,458949
2 0,14439 0,081568 0,888026 0,765527 10,03576
3 0,137768 0,071204 0,897879 0,858634 10,58898
4 0,128879 0,072599 0,899265 0,897218 10,33871
5 0,100944 0,07575 0,906358 0,824481 9,243207
6 0,098152 0,07359 0,908611 0,824092 9,030634
7 0,103731 0,080584 0,900459 0,780139 8,881434
8 0,124752 0,098349 0,879775 0,747539 8,594958
9 0,122683 0,099481 0,87818 0,717381 8,484088
10 0,111662 0,099428 0,881853 0,693798 8,197386
11 0,092312 0,105491 0,884502 0,597164 7,485707
12 0,087579 0,107667 0,884364 0,584208 7,510735
20
13 0,082012 0,113848 0,881036 0,531041 7,382313
14 0,07552 0,108117 0,887192 0,507344 7,670606
15 0,086211 0,113782 0,87837 0,527845 7,913729
16 0,084735 0,111536 0,880105 0,516158 7,766613
17 0,090149 0,115581 0,874903 0,522796 7,578769
18 0,102137 0,121021 0,865822 0,538623 7,563055
19 0,100363 0,118847 0,867933 0,550177 7,398828
20 0,099236 0,11726 0,869008 0,553417 7,589723
21 0,096544 0,111616 0,874355 0,565739 7,436208
22 0,095779 0,108706 0,876862 0,557199 6,896772
23 0,09558 0,106815 0,877893 0,557456 7,011511
24 0,094357 0,109074 0,87617 0,543267 6,842223
25 0,098462 0,109139 0,874298 0,560737 6,765411
26 0,094535 0,109402 0,87525 0,546053 6,805992
27 0,089332 0,106217 0,87984 0,528333 6,502667
28 0,082804 0,109131 0,880978 0,506973 6,118552
29 0,076581 0,107606 0,884646 0,500364 6,163383
30 0,070388 0,103634 0,890619 0,491342 5,891979
31 0,070819 0,10325 0,890318 0,495259 5,992366
32 0,072351 0,102288 0,890278 0,512287 6,081848
33 0,071918 0,102046 0,890119 0,50855 6,186217
34 0,06951 0,104549 0,889157 0,484965 6,157731
35 0,0699 0,102872 0,890205 0,497387 6,104658
36 0,070403 0,10177 0,890578 0,499372 6,102183
37 0,068548 0,098905 0,893333 0,496304 6,117314
38 0,068789 0,097868 0,893966 0,500009 6,000402
39 0,067423 0,099312 0,893143 0,486616 5,965581
40 0,071574 0,104483 0,887626 0,501944 5,831041
41 0,06899 0,106479 0,88716 0,48925 5,794049
42 0,060414 0,107455 0,891563 0,493113 5,457021
43 0,048605 0,103719 0,900483 0,480456 5,328033
44 0,048549 0,102217 0,901627 0,475196 5,208435
45 0,048789 0,102238 0,901461 0,476339 5,179784
46 0,047073 0,100994 0,903102 0,466645 5,137368
Ap TG Nig omega alpha1 beta1 eta11 skew shape
1 0,143467 0,081675 0,888092 0,719207 -0,19879 3,125936
2 0,139836 0,076184 0,893688 0,793629 -0,19461 4,105617
3 0,130344 0,066501 0,904244 0,879368 -0,19414 4,279005
4 0,123883 0,067534 0,905275 0,927169 -0,18568 4,002017
5 0,096566 0,070127 0,912628 0,848887 -0,18457 3,363061
6 0,094591 0,068001 0,914498 0,857418 -0,15213 3,177279
7 0,097141 0,072728 0,909274 0,812376 -0,15203 3,005501
8 0,123612 0,091655 0,885551 0,775263 -0,1214 2,823638
9 0,122439 0,093412 0,883159 0,741458 -0,11932 2,798109
10 0,111743 0,094028 0,886201 0,711562 -0,11523 2,630058
11 0,092814 0,099951 0,88869 0,61164 -0,10497 2,288208
12 0,086098 0,101276 0,890205 0,5951 -0,12078 2,240762
13 0,080399 0,107459 0,886779 0,535103 -0,11191 2,218307
14 0,074292 0,102152 0,892552 0,514258 -0,11194 2,355579
15 0,085105 0,108089 0,883476 0,53614 -0,10865 2,458972
16 0,084103 0,106421 0,884546 0,525217 -0,10736 2,411954
17 0,091101 0,111173 0,878424 0,534385 -0,11844 2,263739
18 0,103503 0,117633 0,86818 0,547347 -0,0916 2,275186
19 0,1017 0,115865 0,869962 0,557618 -0,09235 2,209395
20 0,100488 0,114418 0,871071 0,560516 -0,098 2,284479
21 0,097308 0,108787 0,876415 0,571934 -0,09054 2,247435
22 0,095909 0,105534 0,879382 0,564362 -0,09427 2,060143
23 0,095448 0,103642 0,880557 0,56252 -0,10089 2,113002
24 0,094469 0,106291 0,87833 0,54811 -0,08771 2,036835
25 0,098414 0,106387 0,876535 0,564793 -0,09206 2,001678
26 0,094862 0,106812 0,877256 0,550203 -0,09394 2,018506
27 0,090375 0,103783 0,881297 0,534296 -0,09694 1,911987
28 0,086162 0,107236 0,880801 0,519569 -0,08034 1,742486
29 0,079891 0,105927 0,884365 0,512819 -0,08004 1,759739
30 0,074048 0,102025 0,890064 0,504208 -0,09154 1,646606
31 0,074404 0,101612 0,889863 0,508092 -0,0945 1,698156
32 0,076125 0,100922 0,889608 0,525892 -0,096 1,72795
33 0,075486 0,100714 0,889498 0,521512 -0,09272 1,778862
34 0,072905 0,103426 0,888433 0,495479 -0,09426 1,767482
35 0,073225 0,101757 0,889421 0,506341 -0,09479 1,745234