You are on page 1of 36

AMERICAN MONTE CARLO FOR PORTFOLIO

CVA AND PFE


HPCFinance Conference
May, 2013
Alexandre Morali
CVA/PFE: New analytics challenges

The CVA/PFE computation for exotics brings new challenges for financial quantitative analysts:
 Given a counterparty, full product coverage including both Vanillas and Exotics is fundamental as PFE/CVA figures must
consider the total exposure.

 For trade aggregation at Counterparty level, there is need for consistent multi-asset joint calibration & diffusion
models.

 CVA/PFE scenarios are generated through Monte Carlo simulations. Nested Monte Carlo is not an option for
performance reasons, as a result Front-office libraries can not be used in a straight forward manner especially for
exotics.

 Exotics require complex modelling and heavy computations leading to time consuming calculation. Performance is
always key to have consistent figures in a timely manner.

In this context there is a need for a generic, accurate and efficient resolution method.

American Monte Carlo method seems to be a good candidate for PFE/CVA computation.This
presentation highlights some of the challenges related to the method in this specific context.

2 Copyright © 2012 Murex S.A.S. All rights reserved


A brief reminder of PFE/CVA computation principle

Computation of the Potential Future Exposure of a counterpart at t = 𝝉𝒊

How to compute those MtM prices?


Worst-case (unexpected)
Exposure Exposure
MTM1 (𝝉𝒊 )

MtM2 (𝝉𝒊 )
.
MTM
. Expected Exposure
(current exposure)
.
.
.
.
MtMN (𝝉𝒊 )
Liability Liability

Horizon


𝑷𝑭𝑬𝒕 (𝒑) = 𝒊𝒏𝒇 𝑴: 𝑷𝒓 𝑴𝒕𝑴(𝒕) > 𝑴 𝒙(𝒕) = 𝒑% 𝑪𝑽𝑨 = 𝟏 − 𝑹 ∗ [𝑬𝑷𝑬(𝒕) ∗ 𝑷𝑫 𝒕, 𝒕 + 𝒅𝒕 ]𝒅𝒕
𝟎

 𝒑 : Quantile value  𝑹 : Recovery rate

 𝒙 : Risk factor  𝑬𝑷𝑬(𝒕) : Expected Positive Exposure at t

 𝑷𝑫 𝒕, 𝒕 + 𝒅𝒕 : Counterpart Default Probability between t & t+dt

3 Copyright © 2012 Murex S.A.S. All rights reserved


MTMs evaluation
American Monte Carlo: Regression method

Goal :
• Compute at a given date the payoff conditional expectation as a function of observables at that date.
• These observables are called the regression basis.
• The function is called the regression function. It can be parametric (polynomial form) or not.

1 2 3
Forward Phase Backward Phase Forward Phase
(CVA specific)

 Model Diffusion  Payoff raw evaluation  Payoff evaluation using regression


 Storage of deal cash flows  Regression function building functions and path dependency
 Construction of regression basis  N MC paths  P Monte Carlo paths
 N Monte Carlo paths

Important Remark: In « Phase 3 », MC diffusion can be completly different from « Phase 1 »

4 Copyright © 2012 Murex S.A.S. All rights reserved


MTMs evaluation
American Monte Carlo
Let us illustrate this with a Bermuda swaption which can be exercised at 2 dates T1 or T2:

Libor
6M

T1 T2 T

5 Copyright © 2012 Murex S.A.S. All rights reserved


MTMs evaluation
American Monte Carlo
Forward phase: we diffuse risk factors and store in memory contract cash flows
and payoff variables as they will be used during the Backward phase.

Libor
6M

T1 T2 T

6 Copyright © 2012 Murex S.A.S. All rights reserved


MTMs evaluation
American Monte Carlo
Backward phase: we discount payoff contract and add cash flows up to T2
U(T)=0 and O(T)=0. We obtain U(T2+) cloud of points. O(T2+) worth 0.
 U(t)/O(t) represents the value of the « Underlying Swap »/ « Option » contract at time t

U(T2+)

Libor
6M

O(T2+)

T1 T2 T

7 Copyright © 2012 Murex S.A.S. All rights reserved


MTMs evaluation
American Monte Carlo
At T2+ we obtain U(T2+) cloud of points. O(T2+)=0.

Regression is performed to obtain U(T2-) and O(T2-) as a function of regression basis.

U(T2-) U(T2+)

Libor
6M O(T2-) O(T2+)

T1 T2 T

8 Copyright © 2012 Murex S.A.S. All rights reserved


MTMs evaluation
American Monte Carlo
We discount Underlying and Option variables again to get U(T1+) and O(T1+)
adding cash flows from table.

U(T1+) U(T2-) U(T2+)

Libor
6M O(T1+) O(T2-) O(T2+)

T1 T2 T

9 Copyright © 2012 Murex S.A.S. All rights reserved


MTMs evaluation
American Monte Carlo
Regression is performed to obtain U(T2-) and O(T2-) as a function of regression basis

U(T1-) U(T1+) U(T2-) U(T2+)

Libor
6M O(T1+) O(T2-) O(T2+)
O(T1-)

T1 T2 T

10 Copyright © 2012 Murex S.A.S. All rights reserved


MTMs evaluation
American Monte Carlo
Forward phase: We then move forward again and invoke regression functions with the new set of MC paths.

We update the trade state “S” at any “Call date” by comparing regression function values.

Libor
6M
U(T1) U(T2)

O(T1) O(T2)

S(T1) S(T2)

T1 T2 T

11 Copyright © 2012 Murex S.A.S. All rights reserved


MTMs evaluation
American Monte Carlo
We can compute distribution of MTMs, exposure, PFE, etc…

Libor MtM1 (T2)


6M
MtM2 (T2)

.
.
.
.
.
.
MtM8 (T2)

T1 T2 T

12 Copyright © 2012 Murex S.A.S. All rights reserved


American Monte Carlo for Portfolio CVA and PFE

1 Local versus parametric regression

2 Regression basis choice

3 Convergence to digital features

4 Performance
Parametric versus Local regression
Limitation of parametric approach

Let us study and check on a PFE profile the accuracy of the AMC when a closed-form solution exists.

We plot PFE(95) and PFE(99) profiles for the following European swaption.

Swaption characteristics Model framework


1Y european swaption in delivery mode (1M notional) 1 Factor Hull and White
Underlying wap Tenor: 4y, K = 3% (Forward swap rate: 2.44%) Calibration Basket: coterminal swaptions
Frequency Fixed/Floating : Semi Annual Monte Carlo number of paths= 8K MC paths
Parametric Regression: Polynomial Form Degree 2

European Swaption USD: Profiles


80,00%

70,00%

60,00%
USD (%Notional)

50,00%

40,00%

30,00%

20,00%
PFE(5): Parametric PFE(5): Closed Form
10,00%
PFE(1): Parametric PFE(1): Closed Form
0,00%

Horizon
14 Copyright © 2012 Murex S.A.S. All rights reserved
Parametric versus Local regression
Limitation of parametric approach

The following table expresses the maximum difference observed on one of the 15 PFE dates for
PFE(95%) and PFE(99%) between closed-form formula and American Monte Carlo

Test Case Max error (PFE 95%) Max error (PFE 99%)

HW: K = 2% 9.62% 12.24%


HW: K = 3% 6.20% 9.04%
HW: K = 4% 1.48% 5.74%
HW: K = 2%; High vol factors (*5) 38.69% 48.45%
HW: K = 3%; High vol factors (*5) 35.76% 45.82%
HW: K = 4%; High vol factorsl (*5) 32.83% 43.19%
LMM: K = 2% 3.63% 5.60%
LMM: K = 3% 2.67% 4.70%
LMM: K = 4% 1.41% 3.01%

Observations:
 The higher the quantile, the higher the difference

 The more ITM, the higher the difference for Hull and White model

 Higher volatility factors leads to higher differences

15 Copyright © 2012 Murex S.A.S. All rights reserved


Parametric versus Local regression
Limitation of parametric approach

Comparing parametric function shape induced by the American Monte Carlo when regressing on
the Underlying swap rate highlights differences especially on tail events impacting PFE
computations.

Parametric Regression Results:


Option Value at Expiry 1y (Second order Polynomial Form)

250 000
MC Realization
200 000
Parametric Regression

150 000 Closed Form

100 000
Option Value

50 000

-
-1% 0% 1% 2% 3% 4% 5% 6% 7% 8%
(50 000)

(100 000)

(150 000)
Swap Rate

16 Copyright © 2012 Murex S.A.S. All rights reserved


Parametric versus Local regression
Introduction of Local regression approach

Local Regression (LOWESS: LOcally Weighted Scatterplot Smoothing) approach is based on a


definition of subspaces for the cluster of points generated through Monte Carlo method where we
perform locally regression.
Bandwidth(α)

The method depends on:


Number of MC paths: N

Number of subspaces

Size of subcloud: α
Link function between subcoulds

Let us apply this method to American Monte Carlo and our previous test case.

17 Copyright © 2012 Murex S.A.S. All rights reserved


Parametric versus Local regression
Results of the local regression approach

Here are the results on the previous cluster of points:

200 000
MC Realization
150 000
Closed Form

100 000 Local Regression


Option Value

50 000

-
-1% 0% 1% 2% 3% 4% 5% 6% 7% 8%

(50 000)

(100 000)

(150 000)
Swap Rate

Conclusion: Local regression method allows us to fit the theoretical value of the payoff profile
through the cloud of points.

18 Copyright © 2012 Murex S.A.S. All rights reserved


Parametric versus Local regression
Results of the local regression approach

New profiles with the local regression method on the same case test case:

European Swaption USD: Profiles


80,00%

70,00%

60,00%
USD (%Nominal)

50,00%

40,00%

30,00%

20,00%
PFE(5): Closed Form PFE(1): Closed Form
10,00%
PFE(1): Local PFE(5): Local

0,00%

Horizon

Conclusion: Visual inspection shows that discrepancies between closed form and American Monte
Carlo have been significantly reduced.

19 Copyright © 2012 Murex S.A.S. All rights reserved


Parametric versus Local regression
Results of the local regression approach

The following table expresses the new results on the European swaption test case

Max error (PFE 95%) Max error (PFE 99%)


Test Case
Parametric vs Local Parametric vs Local
HW: K = 2% 9.62% 0.40% 12.24% 0.28%
HW: K = 3% 6.20% 0.34% 9.04% 0.28%
HW: K = 4% 1.48% 0.18% 5.74% 0.27%
HW: K = 2%; High vol factors (*5) 38.69% 2.43% 48.45% 1.17%
HW: K = 3%; High vol factors (*5) 35.76% 2.48% 45.82% 1.16%
HW: K = 4%; High vol factorsl (*5) 32.83% 2.53% 43.19% 1.33%
LMM: K = 2% 3.63% 3.44% 5.60% 4.35%
LMM: K = 3% 2.67% 2.41% 4.70% 3.82%
LMM: K = 4% 1.41% 1.17% 3.01% 2.55%

Conclusions: In all cases the local regression method allows a significant reduction of the maximum
error between American Monte Carlo and Closed form expression.

20 Copyright © 2012 Murex S.A.S. All rights reserved


Parametric versus Local regression
Results on CVA

The following graph illustrates the difference between CVA results using local regression versus
parametric

5,00 ∞
𝑪𝑽𝑨 = 𝟏 − 𝑹 ∗ 𝟎
[𝑬𝑷𝑬(𝒕) ∗ 𝑷𝑫 𝒕, 𝒕 + 𝒅𝒕 ]𝒅𝒕
4,50
4,00
 𝑹 : Recovery rate
3,50
CVA (bps)

3,00
 𝑬𝑷𝑬(𝒕) : Expected Positive Exposure at t
2,50
2,00
1,50  𝑷𝑫 𝒕, 𝒕 + 𝒅𝒕 : Counterpart Default Probability between t & t+dt
1,00
0,50 250 000
MC Realization
-
200 000 Parametric Regression
Closed Form
150 000
100 000
50 000
-
CVA (bps) -2% 0% 2% 4% 6% 8%
(50 000)
(100 000)
Conclusion: (150 000)

 While impacts on PFE are very significant, impacts on CVA are lower.

 PFE considers extreme market data scenario whereas CVA computing is based on a integral of average exposure and
consequently it is less impacted by the lack of accuracy of regression functions at distribution tails.

21 Copyright © 2012 Murex S.A.S. All rights reserved


American Monte Carlo for Portfolio CVA and PFE

1 Local versus parametric regression

2 Regression basis choice

3 Convergence to digital features

4 Performance
Choice of the regression basis
Limitation of using only market observables
 We consider a 3y FX TARN product paying annual coupons and submitted to early redemption
Fx Tarn characteristics Model framework
1M USD Nominal EUR/USD 3y Fx Tarn 1 Factor Hull and White
Coupon Frequency: Annual; Fx log normal calibrated on ATM Fx Options
Strike=1.3 Target=8% Monte Carlo number of paths = 130K
Coupon Type: Max(Fx-Strike;0) Regression Type: Local

 The following four graphs illustrate the shape of the cloud of points and the local regression
function at dates T=0.5y,T= 1.5y,T=2.5y, T=2.99y as functions of Fx spot.

EUR/USD EUR/USD EUR/USD EUR/USD

Conclusions:
• While getting closer to maturity, probability to reach the Target gets higher leading an increasing proportion of
« Dead » Scenario  No Future Cash Flow Expected on those paths (i.e. No exposure).
• Hence, the regression function does not correctly explain the shape of the cluster of points. This is quite obvious when
very close to maturity.
• Some extreme exposures are not taken into account in regression results leading to PFE underestimation.

23 Copyright © 2012 Murex S.A.S. All rights reserved


Choice of the regression basis
Using payoff observable variable

Adding one additional element in the regression basis such as the cumulated coupon leads to the
following 2 dimension profiles at T=1.5y and T=2.99y (cluster of points is in blue and regression
function in red)

T=1.5y T=2.99y

24 Copyright © 2012 Murex S.A.S. All rights reserved


Choice of the regression basis
Using payoff observable variable

Adding one additional element in the regression basis such as the cumulated coupon leads to the
following 2 dimension profiles at T=1.5y and T=2.99y (cluster of points is in blue and regression
function in red).

Fx Spot
Fx Spot

Accumulated
coupon

T=1.5y T=2.99y Accumulated


coupon

Adding this dimension allows us to capture simultaneously:


“Dead” Scenario: points close to zero.
“Alive” Scenario : Non zero exposure contributing to PFE quantiles estimation.

25 Copyright © 2012 Murex S.A.S. All rights reserved


Choice of the regression basis
Numerical results

Looking at the PFE 95% and PFE 99% at each of the four dates considered:

PFE 95% PFE 99% Rsquare


PFE Date Tarn probability
FX spot FX + Accum FX spot FX + Accum FX spot FX + Accum

0.50 0% 19% 19% 25% 25% 29% 29%


1.50 37% 5% 12% 5% 17% 9% 46%
2.50 53% 2% 5% 3% 10% 8% 51%
2.99 53% 3% 6% 3% 13% 11% 81%

If we look the Rsquare, Regression quality has increased significantly with additional element in the
regression basis.

As expected the exposure is increased when adding the second regression basis so this method
allows a better accuracy of the PFE results.

CVA moves from 4.78 bps to 6.04 bps so is impacted by a bit more than 1 bp.

26 Copyright © 2012 Murex S.A.S. All rights reserved


American Monte Carlo for Portfolio CVA and PFE

1 Local versus parametric regression

2 Regression basis choice

3 Convergence to digital features

4 Performance
Convergence to digital features
Limitation of a generic approach
We consider a 1y Worst-of digital on two equity assets

Worst of characteristics Model framework


1M USD Nominal 1y Digital Equity Worst-Of S1 & S2 log normal model: Calibration on ATM vol
Spot 1 = Spot 2 = 10,000. Zero correlation. Monte Carlo number of paths = 16K
Strike = 10,000 Regression Type: Local

We plot PFE 95% and PFE 99% profiles:

Digital On WorstOf PFE profile


~=105%
120%

100%
The method highlights obvious
flaws
USD (%Nominal)

80% ~=87%

60%
 At maturity PFE 99% is greater
(105%) than the engaged nominal
40% which is not possible.
20%
 Knowing that 11% of the paths are in
0% the money at maturity PFE (95%)
0 0,2 0,4 0,6 0,8 1 should also be equal to nominal.
Horizon

PFE(99%) Alpha=30% PFE(95%) Alpha=30%


Final 99% Final 95%

28 Copyright © 2012 Murex S.A.S. All rights reserved


Convergence to digital features
Limitation of a generic approach

Plotting close to Option Expiry, the regression result as a function of [Spot 1,Spot 2] indicates
that the angle is not captured accurately.

Moreover for high spot values the regression function overshoots the digital “expected” profile
explaining why PFE(99%) is above the maximum possible exposition of 100%.

Spot 2

Spot 1

T=1y

The sub cloud of points (alpha = 30% of the total cloud) for each sub space is too big to capture
this shape

29 Copyright © 2012 Murex S.A.S. All rights reserved


Convergence to digital features
Reduction of the size of the sub clouds of points

Reducing the size of each sub-cloud of points allows us to better capture the shape of the digital
worst-of payoff

Spot 2

Spot 1

T=1y

30 Copyright © 2012 Murex S.A.S. All rights reserved


Convergence to digital features
Reduction of the size of the sub clouds of points

Obvious flaws found in the first place are solved by reducing the size of the cloud from 30% to 3%

Digital On Worst-Of PFE profile


120%

100%
USD (%Nominal)

80%

60%

40%

20%

0%
0 0,2 0,4 0,6 0,8 1
PFE(99%) Alpha=30% PFE(95%) Alpha=30%
PFE(99): Alpha = 3% PFE(95): Alpha = 3%

The smaller the size of the sub clouds of points, the better the convergence.

31 Copyright © 2012 Murex S.A.S. All rights reserved


American Monte Carlo for Portfolio CVA and PFE

1 Local versus parametric regression

2 Regression basis choice

3 Convergence to digital features

4 Performance
Performance
CPU vs GPU computation time
Swaption characteristics Model framework
Bermuda Swaption: 10y; First call date: 1y (36 call dates total) 1 Factor LMM or HW model
Underlying Frequency Fixed/Floating : Quaterly Calibration Basket: Underlying Swaptions
Call frequency: Quaterly Regression on 65,000 paths

“Limits pre-deal check” context

 Check no PFE limit break

 Compute CVA fee and reflect it on price

 Arbitrate between different counterparts to optimize CVA consumption

Model CPU (sec) GPU (sec) Speed up


HW 1.81 0.5 4x
LMM 65.51 2.7 25x

33 Copyright © 2012 Murex S.A.S. All rights reserved


Regression Basis Dimension is a headache
Single regression timings as a funtion of MC paths & Regression basis dimension

Local Regression
0,4 0,359 2,5 14
3D 12,203
1D 2D 2,062
12
2
0,3 10
1,5 8
0,2
0,125 1 6
0,609 3,109
0,1 0,078 4
0,5
0,032 0,172 2 0,985
0,063 0,235
0 0 0
100 000 250 000 500 000 1 000 000 100 000 250 000 500 000 1 000 000 100 000 250 000 500 000 1 000 000

Parametric Regression
0,3 1,2 3,5 3,1124
0,253
0,25 0,953 3
1
0,2 2,5
0,8
2
0,15 0,125 0,6 1,5782
0,472
1,5
0,1 0,0624 0,4
0,2312 1 0,75
0,05 0,0218 0,2 0,0624 0,5 0,2468
0
0 0
100 000 250 000 500 000 1 000 000
100 000 250 000 500 000 1 000 000 100 000 250 000 500 000 1 000 000

1 CPU CORE
34 Copyright © 2012 Murex S.A.S. All rights reserved
Regression Basis Dimension is a headache

CPU: 1 core
0,4 0,359 2,5 14
3D 12,203
1D 2D 2,062
12
2
0,3 10
1,5 8
0,2
0,125 1 6
0,609 3,109
0,1 0,078 4
0,5
0,032 0,172 2 0,985
0,063 0,235
0 0 0
100 000 250 000 500 000 1 000 000 100 000 250 000 500 000 1 000 000 100 000 250 000 500 000 1 000 000

GPU
x45 x130 x180
0,009 0,018 0,08
0,0078 0,0157 0,0687
0,008 0,016 0,07
x65
0,007 x25 0,014 0,06 x25
0,006 0,012 x80 x10 0,0484
0,0047 0,05
0,005 x25 0,0406
0,01 0,0344
x40 0,0078 0,04
0,004 x20 0,0032 0,008
0,003
x20 0,03
0,006 0,0046
0,002 0,0015 0,0031 0,02
0,004
0,001 0,002 0,01
0 0 0
100 000 250 000 500 000 1 000 000 100 000 250 000 500 000 1 000 000 100 000 250 000 500 000 1 000 000

 … and for a 10 year Bermuda swaptions on a quaterly swap and 160 PFE dates, (160+36)*2 regressions are
needed.
 Most of Exotics will require high dimension regression basis.

35 Copyright © 2012 Murex S.A.S. All rights reserved


Conclusion
American Monte Carlo is an interesting technique allowing us to model an extensive set of payoffs
in a CVA/PFE context which is absolutely crucial.

But it has many critical parameters when it comes to compute extreme quantiles exposure.
 Choosing the right regression function to match extreme quantiles exposure

 Choosing the right regression basis to explain properly the payoff

 Finding the right way of matching Payoff profile

Performance is a real challenge and switching to GPU allows close to « Real time » computing

36 Copyright © 2012 Murex S.A.S. All rights reserved

You might also like