Market Risk

Measurement
Lecture 3:
VaR Computation

Romain DEGUEST

Market Risk Measurement

MSc in Financial Markets

Romain DEGUEST
(EDHEC-Risk Institute)

[email protected]

Lecture 3:
VaR Computation

1/62
 Market Risk
Outline Measurement
Lecture 3:
VaR Computation

Romain DEGUEST

1 Full Valuation Method

Historical Method
Hybrid Method
Analytic/Parametric Method
Monte Carlo Method

2 Option Quantile

3 Local Valuation Method

The Delta-Normal Method
Failure of the Delta Method
The Delta-Gamma Method

2/62
 Market Risk
Summary of Previous Lectures Measurement
Lecture 3:
VaR Computation

Romain DEGUEST

Remember that we saw different measures for risks: Full Valuation Method

Volatility/Variance: good proxy for risk, useful for investment Option Quantile
decisions. Local Valuation
Value-at-Risk (VaR ): good proxy for extreme risk. Method

Volatility/Variance cannot be used to assess extreme risks.

The regulation requires the use of VaR at a given level :

VaR (FL ) = −q (FL ),

where FL is the distribution of the P&L: L∆

t = Pt+∆ − Pt .

So, the right approach is to estimate the portfolio P&L distribution

at a target horizon ∆ = 10D, 1Y .
This distribution can be derived by modeling the P&L as a function
of underlying risk drivers/factors, e.g.: equities, exchange rates,
interest rates, commodities, and option specific (e.g. implied
volatility).

3/62
 Market Risk
Need for a Unified Framework Measurement
Lecture 3:
VaR Computation

Romain DEGUEST

Full Valuation Method

Option Quantile
What procedure should one follow to measure the risk of a Local Valuation
Method
portfolio?
To answer this question, we could rely on techniques that are
specific to each instrument such as:
Bond Duration: sensitivity of a bond price w.r.t. a small parallel
shift of the yield curve. Duration can be a measure of risk for a
portfolio of bonds.
Option Greeks: sensitivity of an option price w.r.t. a small change
in the price of the underlying, interest rate, volatility.
However, a more general/unified approach is needed for general
portfolios.
Any suggestion?

4/62
 Market Risk
Two General Approaches Measurement
Lecture 3:
VaR Computation

Romain DEGUEST

Once the risk factors are identified, the fundamental question is: Full Valuation Method
how to derive/estimate the P&L distribution? Option Quantile

It is important to either know the P&L distribution in closed form Local Valuation
Method
(e.g. Gaussian) or in a form that we can sample from.

Two approaches exist to estimate the P&L distribution:

Local valuation method: derive an approximation to L∆ t by

calculating the corresponding derivatives (sensitivities) w.r.t to the
risk factors. Then, calculate the distribution of portfolio value from
the joint distribution of the underlying risk factors.
Full valuation method: the same as above, without the
approximation step.

The trade-off between the two methods is speed versus accuracy.

5/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
Introduction VaR Computation

Romain DEGUEST

The full valuation risk measurement procedure can be summarized as: Full Valuation Method
Historical Method
Hybrid Method
Identify Risk Factors → Estimate P&L → Compute the VaR Analytic/Parametric
R = (R1 , .., Rn ) distribution F
bL from R bL ) = −q (F
VaR  (F bL ) Method
Monte Carlo Method

Once the risk drivers are identified, there are four methods to Option Quantile

derive/estimate the P&L distribution: Local Valuation

Method

1 The historical method – uses the available historical market data of

P&L and risk factors.
2 The hybrid method – uses the same data, but applies declining
weights.
3 The analytic/parametric method – relies on a closed-form
expression for the P&L distribution.
4 The Monte Carlo method – relies on a simulated P&L distribution
from a specific dynamic model.

6/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
Identifying Risk Drivers VaR Computation

Romain DEGUEST
What are the risk drivers/factors R1 , R2 , ... of a portfolio P
composed by two stocks? Full Valuation Method
Historical Method
Hybrid Method
Pt = Yt1 St1 + Yt2 St2 Analytic/Parametric
Method
Monte Carlo Method
For equity positions, the risk factors are the stocks themselves Option Quantile
(equity risk). Local Valuation
Method
In what follows, we will always make the assumption:

(H) Portfolio positions remain the same between t and t + ∆.

The P&L can thus be written as a function of two risk drivers:

   
L∆
t = Pt+∆ − Pt = Yt1 St+∆1
− St1 + Yt2 St+∆
2
− St2
 1   2 
St+∆ 2 2 St+∆
= Yt1 St1 − 1 + Yt S t − 1
St1 St2
! !
1 2
   
St+∆ St+∆
ln ln
St1 St2
= Yt1 St1 e − 1 + Yt2 St2 e − 1 .

7/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
Identifying Risk Drivers VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
Hybrid Method
Analytic/Parametric
In the sequel we favor arithmetic returns over log-returns for Method
Monte Carlo Method
simplicity.
Option Quantile
We obtain a simple explicit expression for the P&L as a function of Local Valuation
the risk factors log-returns: Method

L∆
t = Yt1 St1 rt,t+∆
1
+ Yt2 St2 rt,t+∆
2
.

Then, we pick one of the four methods to derive/estimate the P&L

distribution FL .
Finally, we compute the portfolio VaR by taking the quantile at
level :
VaR (FbL ) = −q (F
bL ).

8/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
The Historical Method VaR Computation

Romain DEGUEST

First, we assume that we have portfolio observations (instead of

Full Valuation Method
assuming that we build portfolio observations from risk drivers’ Historical Method
Hybrid Method
observations). Analytic/Parametric
Method
It does not require any assumptions on the P&L distribution since Monte Carlo Method
it is a non-parametric method. Option Quantile

Local Valuation
The P&L distribution is estimated from historical P&L data Method
Lt−n∆ , Lt−(n−1)∆ , . . . , Lt−∆ .
The data are sorted in increasing order L(1) ≤ . . . ≤ L(n) , and then
the VaR at level  is simply given by:
bL ) = −L(bnc+1) .
VaR  (F (1)

For example, if there are n = 1, 000 data and if  = 5%, then:

bnc + 1 = b1, 000 × 5/100c + 1 = b50c + 1 = 50 + 1 = 51.

If now n = 100, and  = 0.5%, then

bnc + 1 = b100 × 0.5/100c + 1 = b0.5c + 1 = 0 + 1 = 1.

9/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
The Historical Method VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
Hybrid Method
Analytic/Parametric
Method
Monte Carlo Method
Option Quantile

Local Valuation
Method

An illustration for  = 5% (Jorion (2007)).

10/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
The Historical Method in Practice VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
Hybrid Method
Simple case: when portfolio positions remain constant over time Analytic/Parametric
Method
(past and future), then one just has to look at old portfolio values Monte Carlo Method
to get historical P&L data (no need of risk factors). Option Quantile

Local Valuation
If not, risk factors need to be identified to compute old P&L values. Method

We can adopt the following strategy to carry out the calculations:

1 Identify risk drivers and compute their historical returns.

2 Take the returns of the risk drivers on the historical dates and
compute what would have been the P&L at these dates given the
current values of the portfolio positions and risk drivers.
3 Calculate the VaR of the portfolio using the historical method
formula in Equation (1).

11/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
The Historical Method in Detail VaR Computation

Romain DEGUEST
Assume a portfolio of two positions and two risk drivers.
Full Valuation Method
Historical Method
1. n historical dates t1 < t2 < . . . < tn < t are selected, which could Hybrid Method
Analytic/Parametric
be the last one-year of daily data for example: Method
Monte Carlo Method

Dates t1 t2 t3 ... tn Option Quantile

Returns of Risk Driver 1 rt1 ,t +∆ rt1 ,t +∆ rt1 ,t +∆ ... rt1n ,tn +∆ Local Valuation
1 1 2 2 3 3 Method
Returns of Risk Driver 2 rt2 ,t +∆ rt2 ,t +∆ rt2 ,t +∆ ... rt2n ,tn +∆
1 1 2 2 3 3

2. Apply the returns to the current values of the risk drivers St1 and St2
(Top table for log-returns, and bottom table for arithmetic returns):
Scenarios 1 2 3 ... n
r1 r1 r1 r1
Risk Driver 1 St1 e t1 ,t1 +∆ St1 e t2 ,t2 +∆ St1 e t3 ,t3 +∆ ... St1 e tn ,tn +∆
r2 r2 r2 r2
Risk Driver 2 St2 e t1 ,t1 +∆ St2 e t2 ,t2 +∆ St2 e t3 ,t3 +∆ ... St2 e tn ,tn +∆

Scenarios 1 2 3 ... n
Risk Driver 1 St1 (1 + rt1 ,t +∆ ) St1 (1 + rt1 ,t +∆ ) St1 (1 + rt1 ,t +∆ ) ... St1 (1 + rt1 ,t +∆ )
1 1 2 2 3 3 n n
Risk Driver 2 St2 (1 + rt2 ,t +∆ ) St2 (1 + rt2 ,t +∆ ) St2 (1 + rt2 ,t +∆ ) ... St2 (1 + rt2 ,t +∆ )
1 1 2 2 3 3 n n

12/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
The Historical Method in Detail VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
2. (Cont.) Calculate the P&L values for each scenario: Hybrid Method
Analytic/Parametric
Method
Scenarios 1 2 3 ... n Monte Carlo Method
Portfolio L1 L2 L3 ... Ln Option Quantile

where Lk = Yt1 St1 rt1k ,tk +∆

+ Yt2 St2 rt2k ,tk +∆
in the case of simple Local Valuation
Method
linear positions in risk factors 1 and 2, or more generally for any
function f of the risk factors, we can write:
 h i h i
Lk = f St1 1 + rt1k ,tk +∆ , St2 1 + rt2k ,tk +∆ .

3. Finally, sort the P&L data in increasing order L(1) ≤ . . . ≤ L(n) , and
then compute the VaR at level :
bL ) = −L(bnc+1) .
VaR  (F

13/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
Advantages of the Historical Method VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
Hybrid Method
Analytic/Parametric
Method
The historical method is relatively simple to implement if historical Monte Carlo Method
data on risk drivers have been collected in-house. Option Quantile

There is no need to estimate any expected returns, volatility or Local Valuation

Method
more generally covariance matrix; the method is non-parametric. It
simplifies calculations for large portfolios.
The method can account for fat tails to the extent that they are
present in the sample.
There is no need to apply rules for time aggregation because the
historical return frequency can be selected to match the target
horizon ∆.

14/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
Disadvantages of the Historical Method VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
The main assumption of the method is that the past represents the Hybrid Method
immediate future fairly. If the sample omits important events, the Analytic/Parametric
Method
tails will not be represented properly. Monte Carlo Method
Option Quantile
The sampling variation is big. Most institutions use between 250
Local Valuation
and 750 scenarios. This implies between 3 and 8 observations Method

beyond VaR1% which may lead to imprecise quantiles.

The method assumes equal weights of the past returns, where
practitioners might give more weight to more recent data.
The hybrid approach can be used to alleviate this last drawback
(like the EWMA versus MA models for the historical volatility
estimators).
The estimation is pro-cyclical, meaning that good past P&L values
will lead to small VaR values. People might expect the opposite in
practice.

15/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
The Hybrid Method VaR Computation

Romain DEGUEST

This method is a modification of the historical method. Full Valuation Method

Historical Method
Hybrid Method
Analytic/Parametric
To each observation of the P&L Lt−k∆ , k = 1, .., n, we assign a Method
weight αk = cλk−1 where 0 < λ < 1 and c = Pn 1λk−1 = 1−λ 1−λ
n.
Monte Carlo Method
k=1 Option Quantile
The weights sum up to 1 and decline as we go further back in time. Local Valuation
Method
As for the historical method, the observations are sorted in
increasing order L(1) ≤ . . . ≤ L(n) .
Then, the VaR equals the first observation in the sorted sample
such that  is smaller than or equal to the sum of the weight of the
current observation and the weights of the smaller ones.
As an example, consider the case where n = 100 and  = 5%. Then
imagine that the losses and their associated weights are as follows:
L(1) = −$3M < L(2) = −$2.7M < L(3) = −$2.4M < ...
(0.02) (0.01) (0.04)
What is the VaR of this portfolio if one uses the hybrid estimation
method?

16/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
The Hybrid Method VaR Computation

Romain DEGUEST

Full Valuation Method

The losses are already estimated and ordered in increasing order, so Historical Method
Hybrid Method
we just need to look at the associated weights: Analytic/Parametric
Method
Monte Carlo Method
0.02 < 5% Option Quantile

0.02 + 0.01 = 0.03 < 5% Local Valuation

Method
0.02 + 0.01 + 0.04 = 0.07 > 5%

Therefore, the Value-at-Risk is the 3rd value of the sorted P&L

values
VaR 5% (FbL ) = −L(3) = $2.4M

The method treats the data closer to the present as more relevant
by assigning a higher weight.
In our example, the 3rd loss has been obtained with more recent
risk factors’ returns than the first two losses because its weights
(= 0.04) is significantly higher than the other ones.

17/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
The Parametric Method VaR Computation

This approach relies on a particular distributional hypothesis, which Romain DEGUEST

means that it is a parametric approach. Full Valuation Method

Having assumed a distribution, then VaR becomes a function of Historical Method
Hybrid Method
the corresponding distribution parameters. Analytic/Parametric
Method
The P&L is expressed as a function of the P&L return viewed from Monte Carlo Method
time t: Option Quantile
 
Pt+∆ Local Valuation
L∆
t = P t+∆ − Pt = P t − 1 = Pt rt,t+∆ Method
Pt
For example, we assume that rt,t+∆ follows a Gaussian distribution
2
N (µt,∆ , σt,∆ ).
Then, we have:
 √ 
bL ) = −Pt (µt,∆ + σt,∆ z ) = −Pt µt ∆ + σt ∆z
VaR  (F
| {z }
2 )
Quantile of N (µt,∆ ,σt,∆

where z is the -quantile of the standard normal distribution.

If  = 5%, then z = −1.645 and
 √ 
bL ) = −Pt µt ∆ − 1.645 × σt ∆
VaR 5% (F

Effectively, in this case VaR  (F

bL ) is a function of σt and µt .
18/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
The Parametric Method: Normal VaR VaR Computation

Romain DEGUEST
This approach is widely used in the industry because it relies on an
Full Valuation Method
easy distribution (Gaussian), which makes the VaR computation
Historical Method
straightforward. Hybrid Method
Analytic/Parametric
Nonetheless, the estimation of the parameters requires to have an Method
Monte Carlo Method
estimation method for the distribution mean µt , and for the Option Quantile
standard deviation σt . Local Valuation
Method
Apart from the Gaussian assumption, the other rule typically used
is the square-root of time scaling.
The relative VaR is then easy to obtain (and useful in practice if ∆
is small):
 √ 
VaR abs
 (F L )
b = −Pt µt ∆ + σt ∆z
√
VaR rel
 (F L )
b = −Pt σt ∆z .

How is this approach different from using volatility as a risk

measure?
Assume, you have to report the relative VaR , how would you
estimate the volatility σ?

19/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
The Parametric Method: Normal VaR VaR Computation

In practice µt is often set to 0 (practitioners communicate relative Romain DEGUEST

VaR ) in very short horizon problems, and σt would be estimated Full Valuation Method
by some appropriate methods and would therefore vary from one Historical Method
Hybrid Method
day to another. Analytic/Parametric
Method
P −P
This illustrates the fact that the P&L return t+∆Pt t is assumed Monte Carlo Method

conditionally Gaussian (conditionally on the current time t). Option Quantile

Local Valuation
The P&L return might not be Gaussian unconditionally. Method

If we use one of the volatility estimation (MA, EWMA, GARCH)

methods from Lecture 2, we obtain:
√
VaR rel
 (FL )
b = −Pt σ b(t, n) ∆z ,

where the estimation volatility σ

b(t, n) would require the same procedure
as for the historical method for the VaR :
1 Identify risk drivers and compute their historical returns.
2 Take the returns of the risk drivers on the historical dates and
compute what would have been the P&L return at these dates
given the current values of the portfolio positions and risk drivers.
3 Estimate the volatility of the portfolio using the MA, EWMA, or
GARCH method.
20/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
The Monte Carlo Method VaR Computation

Romain DEGUEST

This method is also parametric, except that now, you do not need Full Valuation Method
to know the distribution in closed-form to compute the VaR . You Historical Method
Hybrid Method
just need to be able to sample from this distribution. Analytic/Parametric
Method
For example, if we assume a distribution for the risk drivers, the Monte Carlo Method

P&L distribution of the portfolio may not be known in closed form. Option Quantile

Local Valuation
This allows for more modeling flexibility but also implies more Method
computational complexity.
If we know how to sample from the P&L distribution, then we can
generate scenarios and calculate numerically VaR .
The following steps are followed:
1 An assumption is made about the dynamic of the risk drivers.
2 The parameters of these dynamics are fitted to historical data.
3 The risk drivers are simulated to represent future outcomes.
4 The P&L is computed in each simulated scenario (using current
positions).
5 VaR is calculated as a quantile of the simulated distribution F
bL

21/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
The Monte Carlo Variability VaR Computation

Romain DEGUEST
The result strongly depends on the number of simulated scenarios. This
Full Valuation Method
variability can be mitigated by increasing the number of scenarios: Historical Method
Hybrid Method
Number of Scenarios 99% VaR 95% Conf interval Analytic/Parametric
Method
500 2.067 [1.75, 2.38] Monte Carlo Method
1,000 2.406 [2.14, 2.66]
Option Quantile
5,000 2.286 [2.18, 2.38]
10,000 2.297 [2.22, 2.36] Local Valuation
Method
100,000 2.314 [2.29, 2.33]

The table illustrates the VaR at  = 1% of the standard Gaussian

distribution computed through the Monte Carlo method. (This is
only for illustration, because it can be directly computed with the
parametric method)
The true value is VaR 1% (F bL ) = −z = 2.326.
The confidence intervals are calculated from 100 repetitions.
As one can notice, if the number of scenarios increases, the results
gets more accurate. Unfortunately, this comes at an increasing
computational cost.

22/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
The Monte Carlo Method in Practice VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
Hybrid Method
In the following slides, we consider in detail each of these steps: Analytic/Parametric
Method
Monte Carlo Method
Option Quantile
1 Parametric stochastic processes are specified for all risk drivers.
Local Valuation
2 The parameters of these processes are fitted to historical data. Method

3 Random paths are generated for all risk drivers from the current
time t to the target horizon. We generate states of the world in
which the risk drivers take some particular values.
4 The future portfolio values are evaluated in each state of the world
using the current portfolio positions.
5 The VaR is calculated as the corresponding quantile of the
simulated distribution.

23/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
1. Parametric Assumptions VaR Computation

Romain DEGUEST

Full Valuation Method

The first step is perhaps the most important one – we have to Historical Method
Hybrid Method
choose particular stochastic processes for the risk drivers. Analytic/Parametric
Method
Different classes of risk drivers are modeled by different classes of Monte Carlo Method
processes and there is no universal solution: Option Quantile

Stock returns – e.g. the classical Black-Scholes model, or more Local Valuation
Method
sophisticated models (stochastic volatility model, models with
jumps).
Yield curve – e.g. the Vasicek model, the Cox-Ingerson-Ross model.
Volatilities – mean-reversion models.
If needed, also model exchange rates, commodities, or other
alternatives.
By selecting a particular model, our goal is to capture the most
important empirical characteristics (e.g. fat tails).
The typical trade-off is between model sophistication and
estimation complexity. Big institutions with complex portfolios face
significant modeling challenges.

24/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
1. Parametric Assumptions VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
If for example, we go back to the portfolio P composed of two Hybrid Method
Analytic/Parametric
stocks: Method
Monte Carlo Method
Pt = Yt1 St1 + Yt2 St2
Option Quantile

Then both risk factors are the stocks and one can assume that Local Valuation
Method
each of them follows the B&S model:
 1 2 !
1 1 σ
d ln(St ) = µ − dt + σ 1 dBt1
2
 2 2 !
2 2 σ
d ln(St ) = µ − dt + σ 2 dBt2 ,
2

when the two Brownian motions B 1 and B 2 have correlated

increments with correlation ρ.

25/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
2. Parameter Estimation aka Calibration VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
Hybrid Method
Analytic/Parametric
Method
We choose a time window in the (immediate) past and use this Monte Carlo Method
historical sample to estimate the model parameters. Option Quantile

For example, in the B&S assumptions for our 2-stock portfolio, the Local Valuation
Method
parameters to calibrate are: µ1 , µ2 , σ 1 , σ 2 , and ρ.
We have to be careful if the historical sample has any features
violating the model assumptions made in Step 1.
The particular parameter estimation methods depend on the
models selected; the method of maximum likelihood is used often
because of its good large-sample properties.
The trade-off is between computational burden and accuracy.

26/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
3. Path Generation VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
Note that the risk drivers are dependent variables. The model Hybrid Method
Analytic/Parametric
specified in Step 1 should be capable of capturing not only the Method
Monte Carlo Method
stand-alone characteristics of the drivers, but also their dependence
Option Quantile
structure.
Local Valuation
Method
Having estimated the model parameters, we decide how many
paths we need (e.g. N = 10,000). Then, we generate sample paths
for all risk drivers (dependently) and we obtain N states of the
world in which we know the particular realizations of all risk drivers.
For simpler models, we can sample directly from the distribution at
the target horizon t + ∆, i.e paths are not needed.
In fact, paths are needed when
There are path-dependent derivatives in the portfolio.
The distributions of the risk drivers at the target horizon are not
known explicitly.

27/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
3. Path Generation: an Example VaR Computation

Romain DEGUEST

Suppose that the price process of a given stock is assumed to Full Valuation Method
Historical Method
follow the Black Scholes model. Then, Hybrid Method
Analytic/Parametric
σ2 Method
 
Monte Carlo Method
d ln(St ) = µ − dt + σdBt
2 Option Quantile

Local Valuation
We know that the solution of this SDE takes the form Method

σ2 √
  
St+∆ = St exp µ− ∆ + σ ∆Zt
2

where Zt ∈ N (0, 1).

Suppose also that the calibration of the parameters µ and σ has
been done.
If paths are not needed, one way to run Monte Carlo simulations is
to use the closed-form solution above:
1 Draw N scenarios from N (0, 1).
2 For each scenario, calculate St+∆ using the estimated values for µ
and σ and the current stock price St .

28/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
3. Path Generation: an Example VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
If paths are needed, another approach consists in discretizing the SDE Hybrid Method
Analytic/Parametric
and generating sample paths: Method
Monte Carlo Method
Option Quantile
1 Split the time frame [t, t + ∆] into K − 1 smaller time intervals Local Valuation
t = t1 < . . . < tK = t + ∆ (usually days). Method

2 Starting from St , calculate sequentially Stk using the following

discretized SDE:
σ2
 
ln(Stk ) − ln(Stk−1 ) = µ − (tk − tk−1 ) + σ(Btk − Btk−1 ),
2
where the Brownian increments are independent of each other and
follows the Gaussian distribution N (0, tk − tk−1 ) for each tk .
3 Repeat the above N times.

29/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
3. Path Generation: an Example VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
Hybrid Method
Analytic/Parametric
Method
Monte Carlo Method
Option Quantile

Local Valuation
Method

Path generation in MS Excel.

30/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
3. Path Generation: an Example VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
Hybrid Method
Analytic/Parametric
We can model two risk drivers jointly assuming a two-dimensional Method
Monte Carlo Method
GBM,
Option Quantile

Local Valuation
[σ 1 ]2 Method
  !   
d ln(St1 ) µ1 − d B̄t1
p
1 2 σ1 ρ
= 2 dt+ σ 1 − ρ
d ln(St2 ) [σ 2 ]2 0 σ2 d B̄t2
µ2 − 2

where ρ is the correlation between the log-returns of S 1 and S 2 ,

also d B̄t1 and d B̄t2 are independent.
Paths for S 1 and S 2 can be generated jointly discretizing the
process and following the same algorithm as in the one-dimensional
case.

31/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
3. Path Generation: an Example VaR Computation

Romain DEGUEST

Full Valuation Method

130 130
Historical Method
120 120 Hybrid Method
Risk Driver 1

Analytic/Parametric
110 110 Method
Monte Carlo Method
100 100
Option Quantile
90 90 Local Valuation
80 80 Method
0 10 20 30 0 1000 2000
Steps into the future
100 100
Risk Driver 2

90 90

80 80

70 70

60 60
0 10 20 30 0 1000 2000
Steps into the future

Assume ρ = 0.6, S1t = 80, S2t = 100 and a target horizon ∆ of 25 days. The
colors indicate different states of the world.

32/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
3. Path Generation: an Example VaR Computation

Romain DEGUEST

Full Valuation Method

1500 Historical Method
Hybrid Method
Analytic/Parametric
1000 Method
Monte Carlo Method
Option Quantile
500
Local Valuation
Method
0
60 80 100

130 130
120 120
Risk Driver 1

110 110
100 100
90 90
80 80
60 80 100 0 500 1000 1500
Risk Driver 2

The joint distribution at the target horizon.

33/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
4. Portfolio Valuation VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
Hybrid Method
Calculate the position and the portfolio values for each scenario: Analytic/Parametric
Method
Monte Carlo Method

Scenarios 1 2 3 ... N Option Quantile

Risk Driver 1 at t + ∆ S11 S21 S31 ... SN1 Local Valuation

Method
Risk Driver 2 at t + ∆ S12 S22 S32 ... SN2
..
.
Risk Driver M at t + ∆ S1M S2M S3M ... SNM
⇓
Portfolio L1 L2 L3 ... LN

where Lk = f (Sk1 , Sk2 , . . . , SkM )

−f (St1 , St2 , . . . , StM )
is the value of the
P&L in scenario k expressed as a function of the M risk drivers.

34/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
5. VaR Calculation VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
Hybrid Method
The last row of the table represents a sample drawn from the P&L Analytic/Parametric
Method
distribution at the target horizon t + ∆. Monte Carlo Method
Option Quantile
By construction, the scenarios are equally likely.
Local Valuation
Portfolio VaR is calculated as a sample quantile from the portfolio Method

value scenarios.
The number of scenarios N is a parameter that can be chosen.
The higher N is, the closer the calculated VaR is to the true
VaR implied by the parametric assumptions we have made.
Note that through the Monte Carlo method, we are able to
generate scenarios from a distribution that may be unknown in
closed-form, i.e. it would be impossible to derive the formula from
its density or distribution function.

35/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
Advantages of the Monte Carlo Method VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
Hybrid Method
Analytic/Parametric
Method
Monte Carlo Method
The Monte Carlo method is by far the most powerful method to Option Quantile
compute VaR . Local Valuation
Method
This method allows for the most flexibility: the risk factors can
incorporate time variation volatility or fat tail distributions.
For the instruments in the portfolio, it can account for non-linear
price exposure and complex pricing models.
The accuracy of the estimation can be controlled by increasing the
number of scenarios.

36/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
Disadvantages of the Monte Carlo Method VaR Computation

Romain DEGUEST

Full Valuation Method

A big drawback is the computational burden. Suppose that Historical Method
Hybrid Method
N = 10, 000, then each position needs to be evaluated 10,000 Analytic/Parametric
Method
times. Many derivatives do not allow for closed-form pricing Monte Carlo Method
formulas or easy numerical approximations (binomial models, Option Quantile
solving numerically PDEs) can be computationally demanding. Local Valuation
Method
The method is costly to implement in terms of infrastructure.
Calculations can be accelerated by the technique of parallel
computing.
Weakness of Parametric Approaches in General:
A potential weakness is model risk. It should be verified how
sensitive the results are to changes in the model. Otherwise the
approach turns into a black box which may be dangerous.
Also keep in mind that even if N increases, the VaR estimate only
converges to the VaR implied by your model, and not the true
VaR of the portfolio.

37/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
Summary VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
The full valuation approach does not rely on any approximations of Hybrid Method
Analytic/Parametric
the instruments as functions of the risk drivers. Therefore, it can Method
Monte Carlo Method
be seen as an accurate risk measurement approach.
Option Quantile
The main advantage is that the non-linear exposures of positions Local Valuation
Method
are fully taken into account and underestimation of portfolio risk
due to bad approximation is impossible.
The main disadvantages of full valuation approaches are the lack of
analytic transparency and the increase of computational complexity.
We have considered four different methods of calculation of VaR in
the full valuation approach.
Keep in mind that the use of the same risk measure VaR can lead
to different risk measurement procedures, implying different risk
estimates (depending of the method used).

38/62
 Market Risk
Full Valuation Method Measurement
Lecture 3:
Summary VaR Computation

Romain DEGUEST

Full Valuation Method

Historical Method
The first two methods (historical and hybrid) considered were Hybrid Method
non-parametric and therefore only relied on the past values of the Analytic/Parametric
Method
portfolio and risk factors. Monte Carlo Method
Option Quantile
The last two methods (analytic and Monte Carlo) were parametric,
Local Valuation
and therefore relied on parametric assumptions. Nonetheless, each Method

parameter has to be estimated/calibrated with a procedure that

uses historical data as well.
The parametric methods are better suited for smaller . The main
drawback is model risk.
The most common approach to estimate the risk is to assume a
Gaussian distribution and use the square-root of time rule.
The non-parametric methods are sometimes used as a reference
point. The historical method is known to be more conservative
than the normal VaR .

39/62
 Market Risk
Option Quantile Measurement
Lecture 3:
An Analytical Solution VaR Computation

Consider a position in a European call option and assume that the Romain DEGUEST

only risk driver is the price of the underlying. Full Valuation Method

The call option price C = C (S) is an increasing function of S Option Quantile

because the first derivative (the Delta of the option) is positive. Local Valuation
Method
Below, we show that the quantile of C can be expressed directly
through the quantile of S.
Here, we work with quantiles and not VaR because the stock and
the call option prices are always positive, so it would imply a
negative VaR , which does not make a lot of sense in risk
estimation.
By definition of the quantile q (FS ), we have:

P(S ≤ q (FS )) = 

Using the fact that C is increasing, it implies that:

P(S ≤ q (FS )) = P(C (S) ≤ C (q (FS ))) = 

which is the definition of q (FC ), hence it follows that

q (FC ) = C (q (FS )).

40/62
 Market Risk
Option Quantile Measurement
Lecture 3:
Call Option Price Distribution VaR Computation

Romain DEGUEST

Full Valuation Method

15 15
Option Quantile
Option price

Option price
Local Valuation
10 10 Method

5 5

0 0
80 100 120 0 500 1000 1500
Price of underlying Frequency
1500
Frequency

1000

500

0
80 100 120
Price of underlying

The dashed line corresponds to the 1%-quantile.

41/62
 Market Risk
Option Quantile Measurement
Lecture 3:
Limit of the Analytical Solution VaR Computation

Romain DEGUEST

Full Valuation Method

Option Quantile

Local Valuation
Expressing the option quantile as a function of the underlying Method

quantile is practical when C depends on a single risk driver.

In case of more complex dependence, it is difficult to back out the
values of the risk drivers for which C attains a given quantile.
Finally, it is important to keep in mind that the expression derived
previously:
q (FC ) = C (q (FS ))
does not work for any option C . Indeed it relies on two
assumptions:
1 function C has to be defined at q (FS ),
2 function C has to be monotonic in S.

42/62
 Market Risk
Option Quantile Measurement
Lecture 3:
Short Straddle Price Distribution VaR Computation

Romain DEGUEST

Full Valuation Method

0 0
Straddle price

Straddle price
Option Quantile
−2 −2
Local Valuation
Method
−4 −4
−6 −6
−8 −8
80 100 120 0 2000 4000 6000
Price of underlying Frequency
1500
Frequency

1000

500

0
80 100 120
Price of underlying

The dashed line corresponds to the 1%-quantile.

We see in this case that both sides of the underlying distribution leads to losses:
you cannot directly apply the Option Quantile approach when C is not monotone.

43/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
Definition VaR Computation

Romain DEGUEST

Full Valuation Method

In contrast to full valuation methods, the local valuation approach Option Quantile
relies on approximations of the instruments as functions of the risk Local Valuation
Method
drivers in order to estimate the joint behavior of risk drivers. The Delta-Normal
Method
The main advantage of such a method consists of analytical Failure of the Delta
Method
transparency leading to a significant decrease of computational The Delta-Gamma
Method
complexity.
The main disadvantage is that for non-linear exposures of
positions, approximations need to be done, which may cause
underestimation of portfolio risks.
Keep in mind that low-quality risk estimates can also come from:
bad modeling/estimation of the joint behavior of risk drivers
bad properties of the risk measure
So far we have only seen one risk measure: VaR , but many
different modeling/estimation approaches for the joint behavior of
risk drivers.

44/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
The Delta Method VaR Computation

Romain DEGUEST
The motivation behind using the Delta method is to force the
Full Valuation Method
portfolio P to be a linear function of the risk drivers, by considering
Option Quantile
first order Taylor expansion of P as a function of the risk drivers.
Local Valuation
Denote by P(St ) the value of a given portfolio as a function of one Method
The Delta-Normal
risk driver St . Method
Failure of the Delta
The change of the position value during a small amount of time ∆ Method
The Delta-Gamma
can be approximated by a first order Taylor expansion: Method

∂P
L∆
t := P(St+∆ ) − P(St ) ≈ (St+∆ − St ).
∂S
∂P
The derivative ∂S
(St ) is taken at St and generally denoted by δt .
This method is applied to option positions that are non-linear
functions of an underlying variable, and to fixed-income
instruments that are non-linear functions of a bond yield.
Notice that the VaR of portfolio P is simply the product of δt
times the VaR of the risk driver S:
bP(S )−P(S ) ) ≈ VaR  (F
VaR  (F bδ (S −S ) ) = δt VaR  (F
bS −St )
t+∆ t t t+∆ t t+∆

45/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
The Delta Method VaR Computation

If P represents a position in a bond, then δt = −Dt × P(St ), where Romain DEGUEST

Dt denotes the duration of the bond, and (St+∆ − St ) denotes the Full Valuation Method
change in the interest rates. Option Quantile

If St is a stock price following the Black-Scholes model, and P is Local Valuation

Method
the price of a call option on S with strike K and maturity T , then The Delta-Normal
Method
δt is simply the Black-Scholes delta of the option: Failure of the Delta
Method
The Delta-Gamma
δt = N(d1 ), Method

where N(x) = P(X ≤ x) is the cumulative distribution function of

the standard Gaussian law, and d1 is defined as:
2
ln SKt + r (T − t) + σ2 (T − t)


d1 = p .
σ (T − t)
If we rewrite the portfolio increment as
 
St+∆ − St
L∆
t := P(St+∆ ) − P(St ) ≈ δt St
St
St+∆ −St
then δt St can be interpreted as the dollar exposure and St
denotes the percentage change in the risk driver S.
46/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
The Delta-Normal Method VaR Computation

Romain DEGUEST

Full Valuation Method

In order to compute the VaR of the portfolio, we need to estimate
the distribution of the P&L denoted with L∆t
Option Quantile

Local Valuation
The most popular ways to carry on the computation in an easy and Method
The Delta-Normal
transparent way is to assume that the increments of the risk driver Method
S −S 2
are Gaussian: t+∆St t ∼ N (0, σt,∆ ). Failure of the Delta
Method
The Delta-Gamma
We assume a zero-mean µt,∆ = 0 for the risk driver increments, so Method

that all the VaR we compute have to be seen as relative VaR .

Then one can directly use the VaR computation derived for the
Gaussian parametric case and obtained:

VaR  (F
bL ) = δt St VaR  (F
b St+∆ −St )
St

= −δt St (µt,∆ + σt,∆ z )

From the formula above, it is easy to see that this approach is

equivalent to the Gaussian parametric VaR applied to the
linearization of the true portfolio instead of the true portfolio.

47/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
The Delta-Normal Method VaR Computation

Romain DEGUEST

Full Valuation Method

Option Quantile

Local Valuation
Pros: Method
The Delta-Normal
Method
Computationally it is very fast and can be applied in real time. Failure of the Delta
Method
Transparent and and easy to communicate. The Delta-Gamma
Method

Cons:
The assumption of normality is not realistic as a model for the risk
drivers.
The method could be very imprecise for non-linear instruments as
illustrated by the straddle example.

48/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
Failure of the Delta Method VaR Computation

Romain DEGUEST

Full Valuation Method

Option Quantile

Local Valuation
The delta approach can fail miserably if positions are non-linear Method
The Delta-Normal
functions of risk drivers. Method
Failure of the Delta
Method
While the failure can be illustrated with almost any derivative, The Delta-Gamma
straddles are particularly appropriate. Method

A short position in a straddle consists of a short position in a

European put and a short position in a European call option with
the same strike and maturity.
Banks that sell straddle are short straddle and therefore become
exposed to market risks if the price of the underlying deviate from
the strike price.

49/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
Shorting Straddles VaR Computation

Romain DEGUEST
Short straddle
10 Full Valuation Method

Pay−off at maturity Option Quantile

Local Valuation
Method
5 The Delta-Normal
Method
Failure of the Delta
Method
The Delta-Gamma
0 Method
Pay−off

−5

−10

−15
80 85 90 95 100 105 110 115 120
Price of underlying

Payoff of a short position in a straddle with strike $100.

50/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
Shorting Straddles VaR Computation

Romain DEGUEST

Full Valuation Method

Option Quantile

Local Valuation
Method
The Delta-Normal
Method
Assume for simplicity that the only source of risk is the underlying Failure of the Delta
equity. Method
The Delta-Gamma
Method
We consider the following setting where the time-to-maturity T − t
is 3 months (0.25 years), the risk-free rate is 2%, implied volatility
is 18%, strike price is $ 100, and no dividend is paid.
Assume also the current price of the underlying is St = $99.11.
What is the risk of this strategy by the Delta-Normal method?

51/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
Shorting Straddles VaR Computation

Romain DEGUEST

Full Valuation Method

We compute the delta of the straddle which can be seen as the delta of Option Quantile
the two call and put options: Local Valuation
Method
The delta of the call-option position The Delta-Normal
Method
Failure of the Delta
∂Call Method
= N (d1 ) = 0.5 The Delta-Gamma
∂S Method

The delta of the put-option position

∂Put
= N (d1 ) − 1 = −0.5
∂S
Therefore, in that particular setting, the delta of a short position in the
straddle is equal to 0:
∂P ∂Call ∂Put
δt = =− − = 0.
∂S ∂S ∂S

52/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
Shorting Straddles VaR Computation

Romain DEGUEST

Full Valuation Method

Option Quantile

Local Valuation
Method
The Delta-Normal
Method
This result implies that the portfolio VaR is zero for any target Failure of the Delta
Method
horizon: The Delta-Gamma
Method

VaR  (F
bP(S )−P(S ) )
t+∆ t = δt VaR  (F
bS −St ) = 0.
t+∆

But the price of these options are $3.37 for the call and $3.76 for
the put.
Is this a free lunch?

53/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
Shorting Straddles VaR Computation

Romain DEGUEST

2 Full Valuation Method

Straddle value Option Quantile
0
Local Valuation
Method
−2 The Delta-Normal
Method
Failure of the Delta
−4 Method
The Delta-Gamma
Pay−off

Method
−6

−8

−10

−12

−14
80 85 90 95 100 105 110 115 120
Price of underlying

Even if delta-neutral, the straddle position for a target horizon of 1 month shows a
significant potential for losses if the price of the underlying is not between $95 and
$105.

54/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
The Delta-Gamma Method VaR Computation

Romain DEGUEST

Full Valuation Method

An improvement of the Delta method can be constructed by taking Option Quantile
into account the second derivative in the Taylor series expansion of Local Valuation
Method
P(St ). The Delta-Normal
Method
This approach is called the Delta-Gamma method, Failure of the Delta
Method
The Delta-Gamma
∂P 1 ∂2P Method
L∆
t := P(St+∆ ) − P(St ) ≈ (St+∆ − St ) + (St+∆ − St )2 .
∂S 2 ∂S 2
If one denotes the second derivative with Γt , then Taylor expansion
can be rewritten as:
1
L∆
t := P(St+∆ ) − P(St ) ≈ δt (St+∆ − St ) + Γt (St+∆ − St )2 .
2
Bond position: δt = −Dt × P(St ) and Γt = Ct × P(St ) where Dt is
the duration and Ct stands for the convexity of the bond.
Option position: δt is the option delta and Γt is the option gamma.

55/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
The Delta-Gamma-Normal Method VaR Computation

Romain DEGUEST

In order to compute the VaR of the portfolio, we need to estimate Full Valuation Method

the distribution of the P&L denoted with L∆t Option Quantile

Local Valuation
Again, we assume that the increments of the risk driver are Method
S −S 2
Gaussian: t+∆St t ∼ N (0, σt,∆ ). The Delta-Normal
Method
Failure of the Delta
But now, it is no longer easy to derive the distribution of the P&L Method
L∆
t , unless we rearrange the terms into:
The Delta-Gamma
Method

1
L∆
t ≈ δt (St+∆ − St ) + Γt (St+∆ − St )2
2
2
1 δt2

1 2 2 St+∆ − St δt
= Γt σt,∆ St + − .
2 σt,∆ St Γt σt,∆ St 2 Γt

Since St+∆ − St ∼ N (0, St2 σt,∆

2
), then it is well known that the
 2
St+∆ −St
random variable σt,∆ St + Γt σδt,∆t
St
follows a non-central
chi-squared distribution χ2λ (1) with 1 degree of freedom, and the
δt2
non-centrality parameter λ is equal to 2 S2 .
Γ2t σt,∆ t

56/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
The Delta-Gamma-Normal Method VaR Computation

Then one can directly compute the VaR of the P&L as: Romain DEGUEST

 − 1 Γt σ 2 S 2 z1− − 1 δt2
 
2 t,∆ t 2 Γt
if Γt < 0 Full Valuation Method
VaR  (FL ) =
b 2 Option Quantile
1 2 2 1 δt
 Γ σ S z − 2 Γt
2 t t,∆ t 
if Γt > 0 Local Valuation
Method
where z and z1− are the -quantile and (1 − )-quantile of the The Delta-Normal
Method
non-central chi-squared distribution χ2λ (1). Failure of the Delta
Method
The Delta-Gamma
Method

57/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
Shorting Straddles VaR Computation

Romain DEGUEST

Full Valuation Method

The Delta method failed to detect risks because of the non-linear Option Quantile

dependence of options on the price of the underlying. Local Valuation

Method

Because the gamma of the European put and call are the same, The Delta-Normal
Method
the gamma of the straddle is non-zero, and can be expressed as Failure of the Delta
Method
The Delta-Gamma
∂2P ∂ 2 Call ∂ 2 Put Method
Γt := = − − = −2ΓCall
t < 0.
∂S 2 ∂S 2 ∂S 2
The Delta-Gamma approximation applied to the straddle leads to
1
P(St+∆ ) − P(St ) ≈ δt (St+∆ − St ) + Γt (St+∆ − St )2
2
= δt (St+∆ − St ) − ΓCall
t (St+∆ − St ) .
2

It is clear here that the risk of shorting the straddle will no longer
be zero.

58/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
Shorting Straddles VaR Computation

Romain DEGUEST

2 Full Valuation Method

Option Quantile

0 Local Valuation
Method
The Delta-Normal
−2 Method
Failure of the Delta
Method
−4 The Delta-Gamma
Method
Pay−off

−6

−8

Straddle value
−10
Delta−gamma approximation

−12

−14
80 85 90 95 100 105 110 115 120
Price of underlying

The straddle value and its Delta-Gamma approximation.

59/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
Shorting Straddles VaR Computation

Romain DEGUEST

Full Valuation Method

Option Quantile
For our specific choice of parameters and variables, we found that:
Local Valuation
δt = 0, leading to: Method
The Delta-Normal
−ΓCall 2 Method
P(St+∆ ) − P(St ) ≈ t (St+∆ − St ) . Failure of the Delta
Method
The Delta-Gamma
Moreover, we know here that Γt = −2ΓCall
t < 0. Method

Therefore, in this case, the VaR calculation using the

Delta-Gamma-Normal method simplifies to
1 2
VaR  (F
bL ) ≈ − Γt σt,∆ St2 z1−
2
= ΓCall 2 2
t σt,∆ St z1− ,

where z1− is the (1 − )-quantile of the χ2 (1) distribution (λ = 0

here).

60/62
 Market Risk
Local Valuation Method Measurement
Lecture 3:
Shorting Straddles VaR Computation

Romain DEGUEST

Full Valuation Method

Option Quantile
Using the numerical values in the example, we calculate the
Local Valuation
Black-Scholes gamma: Method
The Delta-Normal
N 0 (d1 ) Method
ΓCall
t = √ = 0.0446 Failure of the Delta
Method
St σt,∆ T − t The Delta-Gamma
Method

If we set  = 1% (corresponding to 99% VaR ), z99% = 6.635.

√
If the monthly volatility is 5.2% = 18%/ 12, then the 99%
monthly VaR is given by:
bL ) = 0.0446 × (0.052)2 × (99.11)2 × 6.635 = 7.86
VaR 1% (F

At the monthly target horizon, the strategy loses more than $7.86
with a probability less than 1%.
Note that we have ignored any time decay effects.

61/62
 Market Risk
Summary Measurement
Lecture 3:
VaR Computation

Romain DEGUEST

Full Valuation Method

Local valuation methods are attractive because they allow analytic Option Quantile

expressions for portfolio VaR . Local Valuation

Method
We considered the Delta-Normal method which assumes the The Delta-Normal
Method
percentage changes of risk drivers are Gaussian and is based on a Failure of the Delta
Method
linear (first-order) approximation of position values as functions of The Delta-Gamma
Method
risk drivers.
The Delta-Normal approach can be improved by the
Delta-Gamma-Normal method based on a second-order
approximation.
Note that for the Delta-Gamma-Normal method, the distribution
of the P&L is more complex than for the Delta-Normal method,
and analytic expressions for portfolio VaR will not be obtained for
more than one risk driver.

62/62