Value at Risk

MF 731 Corporate Risk Management

1 / 37
 Outline

VaR for normal distributions.

Estimating VaR from data.

Review from last time.

Computing VaR for portfolio losses.

Time aggregation of VaR.

Backtesting VaR.

2 / 37
 Value at Risk (VaR)

As we saw last week: let L be a r.v. with CDF F .

For a confidence α ∈ (0, 1), we define

VaRα := inf [` ∈ R | P [L > `] ≤ 1 − α]

= inf [` ∈ R | F (`) ≥ α]
= “F −1 (α)00

Also called the quantile of F .

While defined for any α ∈ (0, 1), we typically consider

α ∈ (.95, 1).

3 / 37
 VaRα = F −1(α)

For continuous and discrete distributions

4 / 37
 VaR for Normal Random Variables

Claim: L ∼ N(µ, σ 2 ) =⇒ VaRα (L) = µ + σN −1 (α).

N: standard normal cdf.

Proof (easy!)
Note we can write L = µ + σZ for Z ∼ N(0, 1).

Then, using the scaling property of VaR:

VaRα (L) = µ + σVaRα (Z ) = µ + σN −1 (α)

We will use this result A LOT going forward.

5 / 37
 VaR From Data (again)

VaR for normal F is important to know.

Another important case: empirical distribution Fne .
{`i }i=1,...,n : data points.
Fne : empirical distribution.
1
Mass n at each `i (including repetition).

Claim: VaRα (Fne ) = `(dnαe) .

`(1) ≤ `(2) ≤ ... ≤ `(n) ordered values.
dnαe = min [k ∈ N | k ≥ nα]: “ceiling”.
This works even if some data points are the same.

6 / 37
 Proof of VaRα (Fne ) = `(dnαe)

Assume for simplicity `(1) < `(2) < · · · < `(n) .

I.e.: no repetition.
dnαe
Fne (dnαe) = n1 × # i | `i ≤ `(dnαe) =

n
≥ α.

Now, let ` < `(dnαe) . Since Fne is a step function:

dnαe−1
Fne (`) ≤ Fne (`(dnαe−1) ) = n
< α.

Thus, `(dnαe) is the value at risk.

This is how we defined F −1 for discrete distributions.

7 / 37
 VaR for Portfolio Losses
We can use VaR for normal and empirical dist’s to
obtain portfolio loss VaR in a wide range of models.
Recall: Lt+∆ = −(Vt+∆ − Vt ) is the portfolio loss.
Goal: estimate VaRα (Lt+∆ ) at time t.

We estimate portfolio losses and hence VaR using:

Historical data, Analytic methods, Monte Carlo.

For each method, we may use

Full, linear, or quadratic losses.
quad quad
Lt+∆ = l[t] (Xt+∆ ), Llin lin
t+∆ = l[t] (Xt+∆ ), Lt+∆ = l[t] (Xt+∆ ).

Trade-off: accuracy versus ease of implementation.

8 / 37
 VaR from Portfolio Losses

Historical (H), Monte Carlo (MC) methods produce

loss data.
 n
H: `i = l[t] (xt−(i−1)∆ ) i=1 where {xt−i∆ } are the historical
changes in risk factors.
 i n  i
MC: `i = l[t] (Xt+∆ ) i=1 where Xt+∆ are i.i.d. samples
of the risk factor changes Xt+∆ .
lin quad
l[t] : one of l[t] , l[t] or l[t] .

Once we have the data, getting the VaR is easy!

VaR
\α = `(dnαe) : estimator for VaRα (Lt+∆ ).

9 / 37
 Historical Example: Stock Portfolio

t = 9/1/21, Vt = $1M, 33% AAPL, 67% MSFT.

Historical daily log returns from 9/1/16 − 8/31/21.

10 / 37
 MC Example: Apple Call Option
Sell a call option (K , T ), delta-hedge with stock.
BS model, constant hedging position over [t, t + ∆].



Lt+∆ = − ht (St+∆ − St ) − C BS (t + ∆, St+∆ ) − C BS (t, St ) .

Simulation inputs (physical measure!)

10
T − t = .292, ∆ = 252 , St = 152.51, K = 170.
r = 0.119%, µ = 16.91%, σ = 49.07%.

Loss distribution (100 calls). VaR

\ .95 = 300.6.

11 / 37
 Analytical Method
F
Assumes Xt+∆ ∼t F and computes VaR off F .
F
E.g.: equities, log returns Xt+∆ ∼t N (µt+∆ , Σt+∆ ).
 (i)

Pd (i) Xt+∆


Lt+∆ = − i=1 θt e − 1 = −θtT e Xt+∆ − 1 .
Pd (i) (i)
Llin
t+∆ = − i=1 θt Xt+∆ = −θtT Xt+∆
(i) (i) (i)
θt = λt St : dollar position.

VaR for linearized losses:

Ft

Llin T T
t+∆ ∼ N −θt µt+∆ , θt Σt+∆ θt .
p
θtT Σt+∆ θt N −1 (α).


VaRα Llin T
t+∆ = −θt µt+∆ +

12 / 37
 Equity Example Continued
VaR for full losses estimated via simulation:
For m = 1, . . . , M:
m F
Sample Xt+∆ ∼t N(µt+∆ , Σt+∆ ).
m

Set `m = Lm T
t+∆ = −θt e
Xt+∆
−1 .

VaRα (Lt+∆ ) = `(dMαe) .

Ex: AAPL/MSMF portfolio from above.

Estimate µt+∆ , Σt+∆ using data from 9/1/16 − 8/31/21
VaR estimates
Empirical: $24, 138.5.
Full normal simulated: $25, 707.1.
Linearized normal: $26, 090.3.

13 / 37
 VaR and Time Aggregation

The basic problem:

Our methods for estimating VaR are typically applied to

one-day horizons.

The regulatory requirement is to calculate VaR for a ten-day

horizon.

14 / 37
 Time-Aggregation

How do we go from a one day to ten day horizon?

One idea, using historical data:
Use current methodology taking data from non-overlapping
10 day periods.
Problem: not enough data. E.g.: 4 years historical data
yields just 100 such periods.
VaR.99 is second largest loss.

Work-around(??): use overlapping periods. BUT

This introduces complicated dependencies.

What do we do?
15 / 37
 Square Root of Time
Industry convention:
√
J day VaRα = J× one day VaRα .

Where is this coming from?

Equity portfolio, constant holdings over [t, t + J∆],
linearized losses, log-return risk factors.
P 
J
Llin
t+J∆ = −θtT j=1 Xt+j∆ .

F
Next, assume {Xt+j∆ }Jj=1 i.i.d. ∼t N(0, Σ).
PJ F
This implies j=1 Xt+j∆ ∼t N (0, JΣ).

16 / 37
 Square Root of Time

Ft

Thus, Llin
t+J∆ ∼ J 0, Nθt
T
Σθt .

q T
VaRα Llin
t+J∆ = Jθt Σθt N −1 (α);
√ q
= J × θtT Σθt N −1 (α);
√
= J × VaRα Llin


t+∆ .

Square root of time horizon multiplied by 1 period VaR.

Industry convention: assume this scaling in general.

17 / 37
 Square root of Time Rule

Included in Basel-I (1996 ammend.) and Basel-II.

However: only from going from 1 to 10 day VaR.
Not from e.g. 1 to 100 day VaR.

Little theoretical evidence to support the rule in

non-Gaussian, i.i.d. models.
Kaufmann, et. al. (’04,’05,’07): OK for unconditional VaR,
longer time horizons.
Frey, et. al. (’00): not OK for conditional VaR, shorter time
horizons.

18 / 37
 Time Aggregation: MC
We can avoid such approximations using MC.
Idea: if Vt = f (t, Zt ) then
   
Lt+J∆ = − f t + J∆, zt + Jj=1 Xt+k∆ − f (t, zt )
P
P 
J J
= l[t] j=1 Xt+j∆
J
l[t] (x) = − (f (t + J∆, zt + x) − f (t, zt )).

Method:
Postulate a model for the process {Xt+j∆ }Jj=1 .
For m = 1, . . . , M
n oJ
m
Sample Xt+j∆ .
j=1
P 
J J m
Compute `m = l[t] j=1 t+j∆ .
X

Output the estimator VaR

\α = `(dMαe) .
19 / 37
 MC and Time Aggregation

J
How do we obtain the loss operator l[t] (x)?
Issue: for J >> 1 we must allow for position changes.

Ex: equity portfolio.

Set Tj = t + j∆ and Lj as loss over [Tj−1 , Tj ].
Positions over [Tj−1 , Tj ] chosen at Tj−1 .
 
(i) (i) (i)
Lj = − di=1 λTj−1 STj − STj−1 and Lt+J∆ = Jj=1 Lj .
P P

How do we account for changing positions?

What rule can we apply at t to know position at Tj−1 ?

20 / 37
 MC and Time Aggregation

For constant positions λTj−1 ≡ λ

 
Pd (i) (i) (i)
Lj = − i=1 λ STj − S Tj−1 .

Telescoping sum.
   PJ (i) 
(i) (i) (i) (i) j=1 XTj
Pd (i)
Pd
Lt+J∆ = − i=1 λ STJ − St = − i=1 λ St e − 1 .

For constant weights wTj−1 ≡ w :

 (i) 
Pd (i) XT
Lj = −VTj−1 i=1 w e j −1 .
PJ
Lt+J∆ = j=1 LTj .

21 / 37
 MC Example: Hedged Call Option

Self-financing strategy over [t, t + J∆].

Over [T0 , T1 ] = [t, t + ∆] (first period)
We are short a call option with strike κ, maturity T .
We are long hT0 = ∂x C BS (T0 , ST0 ) shares of the stock.
We hold YT0 = $0 (no position) in the bank account with rate r .

VT0 = hT0 ST0 + YT0 − C BS (T0 , ST0 ).

| {z } |{z} | {z }
stock bank option

VT1 = hT0 ST1 + YT1 = YT0 e r ∆ − C BS (T1 , ST1 ).

| {z } | {z } | {z }
stock bank option

L1 = −(VT1 − VT0 ).

22 / 37
 Hedged Call Option
For j = 2, ..., J consider [Tj−1 , Tj ].
“Yesterday’s closing” dollar positions (end of [Tj−2 , Tj−1 ])
(stock): hTj−2 STj−1 .
(bank): YTj−1 .
“Today’s opening” dollar positions (beginning of [Tj−1 , Tj ])
(stock): hTj−1 STj−1 where hTj−1 = ∂S C BS (Tj−1 , STj−1 ).


(bank) : YTnew
j−1
= YTj−1 + hTj−2 − hTj−1 STj−1 .

At today’s close (end of [Tj−1 , Tj ])

VTj = hTj−1 STj + YTj = YTnew
j−1
e r ∆ − C BS (Tj , STj ).
| {z } | {z } | {z }
stock bank option


Lk = − VTj − VTj−1 .
STj
 
Simulate by sampling XTj = ln STj−1 .
23 / 37
 Hedged Call Option

Simulation inputs.
1
T − t = .292, ∆ = 252
, J = 10, κ = 170.

St = 152.51, r = 0.119%, µ = 16.905%, σ = 49.07%.

VaR
\ .95 ’s (100 calls)

Full 10-day simulation: 86.76.

√
10× one day VaR: 92.97 (slight over-estimation).

Compare with a VaR of 300.6 without rebalancing.

24 / 37
 Backtesting VaR

We have numerous methods for estimating at t the

VaR of our losses over [t, t + ∆].

Whatever method we use, over time we produce a

series of estimates
n to
VaRα
\ .
t=0,∆,2∆,...

We need a way to test our estimates to see if our

methodology/model is sound.

This is where backtesting comes into play.

25 / 37
 Backtesting VaR: Theory

Suppose we have a model for {Lt+∆ }t .

t ∈ {0, ∆, 2∆, . . . }.
Ft : information at t.
Lt+∆ is known using Ft+∆ but not using Ft .

We can then identify

VaRtα = inf τ ∈ R | P Lt+∆ ≤ τ Ft ≥ α .
  

VaRtα is a RANDOM VARIABLE, known using Ft .

We say the event ω ∈ Ω | Lt+∆ (ω) > VaRtα (ω) is


a VaR violation or exception or exceedance.
26 / 37
 Backtesting VaR: Theory
Write It+∆ := 1Lt+∆ >VaRtα .
Assuming a continuous conditional dist:
E It+∆ Ft = P Lt+∆ > VaRtα Ft = 1 − α.
   

F
=⇒ It+∆ ∼t B(1 − α).

Furthermore: {It+∆ }t are i.i.d. (t < s)

Write Gα as the cdf of a B(1 − α) r.v..

  
P [It+∆ ≤ τ1 , Is+∆ ≤ τ2 ] = P It+∆ ≤ τ1 P Is+∆ ≤ τ2 Ft
= Gα (τ2 )P [It+∆ ≤ τ1 ]
= Gα (τ1 )Gα (τ2 ).

27 / 37
 Backtesting VaR: Theory

Consequence of i.i.d. {It+∆ }t with It+∆ ∼ B(1 − α).

Fix t0 , set Tj = t0 + j∆ and define

PJ
NtJ0 := j=1 It0 +j∆ .

# of exceedances over the J periods [T0 , T1 ], ..., [TJ−1 , TJ ].

Ft
By construction, NtJ0 ∼0 B(J, 1 − α).
Binomial r.v. with J trials (periods), 1 − α probability of
success (exceedance).

28 / 37
 Backtesting VaR: Practice

We use the binomial property to test our model.

Basic idea:

We should observe bigger losses than our forecasted VaR

with probability 1 − α.

E.g.: for ∆ = 1d and α = .95, in one year we should see

approximately 252 × .05 = 12.6 exceedances.

If we see much more or much less, our model needs

adjusting.

29 / 37
 Backtesting VaR: Practice

The test

VaRtα was computed using our model.

In reality, we use our model to estimate the “true” VaR:

n to
I.e. we produce estimates VaR
\ α .
t

We then use the realized losses L̂t+∆ to compute the

realized violations Iˆt+∆ = 1L̂ >VaR
\ .
t
t+∆ α

30 / 37
 Backtesting VaR: Practice

Null hypothesis: our model is accurate.

n oJ
Iˆt0 +j∆ are samples of i.i.d. B(1 − α) r.v.
j=1

PJ
As such, for large J (recall NtJ0 = j=1 It0 +j∆ )
J
N̂ −J(1−α)
If Zb := √t0 then Zb ≈ N(0, 1) in dist..
Jα(1−α)

1 − β confidence interval (CI)

h i
P τ−β ≤ N̂tJ0 ≤ τ+β ≈ 1 − β.
β
p
τ± := J(1 − α) ± z1− β Jα(1 − α) where zγ = N −1 (γ).
2

31 / 37
 Backtesting VaR: Practice
To test the null hypothesis:
We take in the VaR confidence α, # of sample days J, CI
confidence 1 − β.
We then
Count the number of exceedances N̂tJ0 over [T0 , TJ ].
If this number lies outside the 1 − β CI we reject the null
hypothesis at the 100(1 − β)% confidence.

E.g.: α = .95, J = 250, β = .05.

We want N̂tJ0 ∈ (5.75, 19.25)
If N̂tJ0 ≥ 20: model too optimistic (VaR not high enough).
If N̂tJ0 ≤ 5: model too pessimistic (VaR too high).

32 / 37
 Backtesting Example

S&P 500 daily price data from 8/31/2007 to 8/31/2017.

N = 2518 data points.

Hypothetical portfolio:

100% S&P stock, portfolio size a constant $1.

l[s] (Xs+∆ ) = l(Xs+∆ ) = −(e Xs+∆ − 1).

t
Produce the estimates VaR
\α two ways.

33 / 37
 Backtesting Example

Method one: rolling empirical distribution

For each t, use past M days return data to compute

empirical loss distribution.

 M
Use past returns to obtain `tm = l[t] (xt−(m−1)∆ ) m=1
.

t
t
Estimate VaR
\ α = `(dMαe) .

Compare VaR with realized loss l[t] (xt+∆ ).

34 / 37
 Backtesting Example

Method two: normal returns with rolling EWMA.

For each t update mean/variance estimates via EWMA
µ̂t+∆ = λµ̂t + (1 − λ)xt .
2
σ̂t+∆ = θσ̂t2 + (1 − θ)(xt − µ̂t )2 .
t  
\α = − e µ̂t+∆ +σ̂t+∆ N −1 (1−α) − 1 .
Estimate VaR

F 2
Explicit full loss VaR formula when Xt+∆ ∼t N(µ̂t+∆ , σ̂t+∆ ) is
one-dimensional.
Make sure you can derive this!

Compare VaR with realized loss l[t] (xt+∆ ).

35 / 37
 Backtesting Example

Results for α = .95, β = .05, J = N − M = 1508.

Average: 75.4. Low CI: 58.81. High CI: 91.99.

Exceedances

Empirical: 44.

EWMA: 79.

What do we think is happening?

EWMA ok but Empirical too pessimistic.

36 / 37
 Backtesting Example

Exceedances over time

Independence assumption appears to be violated.

37 / 37