Introduction Data and simula- tion methodology VaR models and estimation results Estimation perfor- mance analysis Conclusions Appendix Doctoral School of Finance and Banking Academy of Economic Studies Bucharest Testing and ComparingValue at Risk Models – an Approach to Measuring Foreign Exchange Exposure -dissertation paper- MSc student: Lapusneanu Corin Supervisor: Professor Moisa Alta Bucharest 2001
76
Embed
Doctoral School of Finance and Banking Academy of Economic Studies Bucharest
Doctoral School of Finance and Banking Academy of Economic Studies Bucharest. Testing and ComparingValue at Risk Models – an Approach to Measuring Foreign Exchange Exposure -dissertation paper-. MSc student: Lapusneanu Corina Supervisor: Professor Moisa Altar. Bucharest 2001. - PowerPoint PPT Presentation
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Doctoral School of Finance and Banking
Academy of Economic Studies Bucharest
Testing and ComparingValue at Risk Models – an Approach to Measuring
Foreign Exchange Exposure-dissertation paper-
MSc student: Lapusneanu CorinaSupervisor: Professor Moisa Altar
Bucharest 2001
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
VaR is a method of assessing risk which measures the worst expected loss over a given time interval under normal market conditions at a given confidence level.
Corina Lăpuşneanu - Introduction
Basic Parameters of a VaR Model
Advantages of VAR
Limitations of VaR
IntroductionIntroduction
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
For internal purposes the appropriate holding period corresponds to the optimal hedging or liquidation period.
These can be determined from traders knowledge or an economic model
The choice of significance level should reflect the manager’s degree of risk aversion.
Corina Lăpuşneanu - Introduction
Basic Parameters of a VaR ModelBasic Parameters of a VaR Model
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Introduction
VaR can be used to compare the market risks of all types of activities in the firm,
it provides a single measure that is easily understood by senior management,
it can be extended to other types of risk, notably credit risk and operational risk,
it takes into account the correlations and cross-hedging between various asset categories or risk factors.
Advantages of VAR
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
it only captures short-term risks in normal market circumstances,
VaR measures may be very imprecise, because they depend on many assumption about model parameters that may be very difficult to support,
it assumes that the portfolio is not managed over the holding period,
the almost all VaR estimates are based on historical data and to the extent that the past may not be a good predictor of the future, VaR measure may underpredict or overpredict risk.
Corina Lăpuşneanu - Introduction
Limitations of VaR:
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Statistical analysis of the financial series of exchange rates against ROL (first differences in logs):
Testing the normality assumption
Homoskedasticity assumption
Stationarity assumption
Serial independence assumption
Corina Lăpuşneanu - Data and simulation methodology
where Z() is the 100th percentile of the standard normal distribution
ZVAR
Equally Weighted Moving Average Approach
Exponentially Weighted Moving Average Approach
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Equally Weighted Moving Average Approach
where represents the estimated standard deviation, represents the estimated covariance, T is the observation period,
rt is the return of an asset on day t,
is the mean return of that asset.
ij
r
; 1
2
T
rrT
tt
T
rrrrT
tjjtiit
ij
1
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Graph 4. VaR estimation using Equally Weighted Moving Average
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Exponentially Weighted Moving Average Approach
T
tt
t rr1
211
T
tjjtiit
tij rrrr
1
11
The parameter is referred as “decay factor”.
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Graph 5. VaR estimation using Exponentially Weighted
Moving Average
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Historical Simulation
Graph 6. VaR estimation using Historical Simulation
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
GARCH modelsGARCH models In the linear ARCH(q) model, the conditional variance is postulated to be a linear function of the past q squared innovations:
21
1
22
t
q
iitit L
GARCH(p,q) model:
21
21
1
2
1
22
tt
p
jjtj
q
iitit
LL
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
GARCH (1,1) has the form:
21
21
2 ttt
where the parameters , , are estimated using quasi maximum- likelihood methods
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
The constant correlation GARCH model estimates each diagonal element of the variance-covariance matrix using a univariate GARCH (1,1)and the risk factor correlation is time invariant:
211,, itiitiiiiiitii
tjjtiiijtij ,,,
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Dependent Variable: LDEMMethod: ML - ARCHDate: 07/01/01 Time: 19:32Sample(adjusted): 1/06/1998 12/29/2000Included observations: 779 after adjusting endpointsConvergence achieved after 8 iterationsBollerslev-Wooldrige robust standard errors & covariance
Coefficient Std. Error z-Statistic Prob.
C 0.000963 0.000281 3.429349 0.0006DUMMY 0.066992 0.000852 78.62253 0.0000
D199 0.001102 0.000834 1.321575 0.1863
Variance Equation
C 4.73E-06 4.06E-06 1.163966 0.2444ARCH(1) 0.032776 0.023325 1.405153 0.1600
GARCH(1) 0.882684 0.088609 9.961515 0.0000
R-squared 0.095860 Mean dependent var 0.001297Adjusted R-squared 0.090012 S.D. dependent var 0.007949S.E. of regression 0.007582 Akaike info criterion -6.933546Sum squared resid 0.044442 Schwarz criterion -6.897669Log likelihood 2706.616 F-statistic 16.39123Durbin-Watson stat 1.885839 Prob(F-statistic) 0.000000
Table 3.1. Estimation results with GARCH(1,1)
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Dependent Variable: LUSDMethod: ML - ARCHDate: 07/01/01 Time: 19:41Sample(adjusted): 1/27/1998 12/29/2000Included observations: 764 after adjusting endpointsConvergence achieved after 14 iterationsBollerslev-Wooldrige robust standard errors & covariance
Coefficient Std. Error z-Statistic Prob.
C 0.000445 8.22E-05 5.416127 0.0000LUSD(-1) 0.348223 0.038563 9.029914 0.0000
R-squared 0.128217 Mean dependent var 0.001511Adjusted R-squared 0.114287 S.D. dependent var 0.004452S.E. of regression 0.004190 Akaike info criterion -9.741589Sum squared resid 0.013182 Schwarz criterion -9.662661Log likelihood 3734.287 F-statistic 9.204373Durbin-Watson stat 2.142489 Prob(F-statistic) 0.000000
Table 4. Estimation results with GARCHFIT
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Graph 8. . VaR estimation results with GARCHFIT
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Orthogonal GARCH
X = data matrixX’X = correlation matrixW = matrix of eigenvectors of X’X
The mth principal component of the system can be written:
WXWX '
kkmmmm xwxwxwp ........2211 Principal component representation can be write:
mimiii ppy *1
*1 .......
where iijij w *
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
The time-varying covariance matrix (Vt) is approximated by:
'AADV tt where is the matrix of normalised factor weights
is the diagonal matrix of variances of principal components The diagonal matrix Dt of variances of principal components is estimated using a GARCH model.
*ijA
mpVpVdiagD ,......1
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Graph 9. VaR estimation results with Orthogonal GARCH
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Estimating the pdf of portfolio returns
- Gaussian
- Epanechnikov , pentru - Biweight , pentru
where
n
i n
i
nn h
xxK
nhxf
1
1
2exp
2
1 2t
203.015.05 t
2201.025.03 t
52 t
5t
h
xxt i
Kernel Estimation
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Estimating the distribution of percentile or Estimating the distribution of percentile or order statisticorder statistic
jnjj xFxFxf
jnj
nxg
1
!!1
! 1
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Graph 10a. VaR estimation results with Gaussian kernel
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Graph 10b. VaR estimation results with Epanechnikov kernel
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Graph 10c. VaR estimation results with biweight kernel
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Structured Monte Carlo
dzSdtSdS ttttt If the variables are uncorrelated, the randomization can be performed independently for each variable:
ttSS tjjjtjtj ,1,,
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
But, generally, variables are correlated. To account or this correlation, we start with a set of independent variables , which are then transformed into the , using Cholesky decomposition. In a two-variable setting, we construct:
2
2/1212
11
1
where is the correlation coefficient between the variables .
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
AppendixGraph 11. VaR estimation results using Monte Carlo Simulation
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Generalized Pareto Distribution:
0 /exp1
0 /111
,
y
yyG
= “shape parameter” or “tail index” = “scaling parameter”
yGyFu ,
Extreme value method
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Tail estimator:
uux
n
NxF u
for x ,ˆ
ˆ11ˆˆ/1
F(u)q where, 11ˆ
ˆVAR
ˆ^
qN
nu
u
q
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - Value at Risk models and estimation resultsValue at Risk models and estimation results
Graph 12. VaR estimation results using Extreme Value Method
Equally Weighted Moving Average is a conservative risk measure, which produce the second great average estimation of risk, with a medium variability, good accuracy and a medium efficiency. As the window length is increased (Appendix D), the conservatism and variability will increase. Exponentially Weighted Moving Average tends to produce estimates over all model average, a low variability, good accuracy and medium efficiency. This method is more efficient when calibrated on smaller data window lengths.
ConclusionsConclusions
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
GARCH models produce estimates over all model average, a medium variability. good accuracy and efficiency.
Historical Simulation tends to produce estimates below all model average, a low variability, medium accuracy and efficiency. It is more efficient when calibrated on smaller data window.
Corina Lăpuşneanu - ConclusionsConclusions
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Structured Monte Carlo Simulation: presents the highest level of conservatism, high variability, the least accurate estimates, and low efficiency.
Kernel density estimation: produce estimates below all model average, a high variability, a good accuracy and efficiency except the Gaussian kernel that has a low accuracy and efficiency.
Corina Lăpuşneanu - ConclusionsConclusions
Introduction
Data and simula-tion methodology
VaR models andestimation results
Estimation perfor-mance analysis
Conclusions
Appendix
Corina Lăpuşneanu - ConclusionsConclusions
Extreme Value: produce the least conservatives VaR estimates except 5% threshold with produce estimates over all model average, low variability, good accuracy except 15% threshold, the most efficient models after that was scaling with the multiple to obtain coverage. It’s preferred a model with a low threshold (5%).