Markov Functional Model

Peter Caspers

IKB

November 13, 2013

Peter Caspers (IKB) Markov Functional Model November 13, 2013 1 / 72

Disclaimer

The contents of this presentation are the sole and personal opinion of the
author and do not express IKB’s opinion on any subject presented in the
following.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 2 / 72

Table of contents

1 Code Example

2 Theoretical Background

3 Model description

4 Calibration

5 Numerics

6 Secondary instrument set calibration

7 Implementation in QuantLib

8 References

9 Questions

Peter Caspers (IKB) Markov Functional Model November 13, 2013 3 / 72

 Code Example

Term Sheet

Consider the following interest rate swap, with 10y maturity.

We receive yearly coupons of type EUR CMS 10y
We pay Euribor 6m + 26.7294bp
We are short a bermudan yearly call right
What is a suitable way to price this deal ?

Peter Caspers (IKB) Markov Functional Model November 13, 2013 4 / 72

 Code Example

Model requirements

A good start1 would be to use a model that

prices the underlying CMS coupons consistently with given swaption
volatility smiles
can in addition be calibrated to swaptions representing the call right
(say for the moment to atm coterminals)
gives us control over intertemporal correlations
A plain Hull White model fulfills #2 and #3 but clearly fails to meet #1.

1
one important requirement - the decorrelation of Euribor6m and CMS10y
rates - is missing here, and actually not satisfied by the Markov 1F model
Peter Caspers (IKB) Markov Functional Model November 13, 2013 5 / 72
 Code Example

The Markov model approach

This is where the Markov Functional Model jumps in. Before going into
details on how it works, we give an example in terms of QuantLib Code.
Let’s suppose you have already constructed a
Handle<YieldTermStructure> yts
representing a 6m swap curve and a
Handle<SwaptionVolatilityStructure> swaptionVol
representing a swaption volatility cube (suitable for CMS coupons pricing).

Peter Caspers (IKB) Markov Functional Model November 13, 2013 6 / 72

 Code Example

Model construction

We create a markov model instance as follows

boost::shared_ptr<MarkovFunctional> markov(
new MarkovFunctional(yts,reversion,sigmaSteps,sigma,
swaptionVol,cmsFixingDates,
cmsTenors,swapIndexBase));
with a mean reversion
Real reversion
controlling intertemporal correlations (see below), a piecewise volatility
(for the coterminal calibraiton) given by
std::vector<Date> sigmaSteps
std::vector<Real> sigma

Peter Caspers (IKB) Markov Functional Model November 13, 2013 7 / 72

 Code Example

Model construction (ctd)

our structured coupon fixing dates and tenors

std::vector<Date> cmsFixingDates
std::vector<Period> cmsTenors
and a swap index
boost::shared_ptr<SwapIndex> swapIndexBase
codifying the conventions of our cms coupons.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 8 / 72

 Code Example

Model calibration

The calibration to the constant maturity swaption smiles is done

automagically (see below). The calibration to the coterminal swaptions is
done as usual by defining a calibration basket
std::vector<boost::shared_ptr<CalibrationHelper> >
coterminalHelpers
and then calibrating the model with
markov->calibrate(coterminalHelpers,optimizer,endCriteria)
where the sigmaSteps are the coterminals’ expiry dates. Note that n + 1
volatilities are needed for n options with the first volatility kept fixed
during calibration (which should become clearer later on).

Peter Caspers (IKB) Markov Functional Model November 13, 2013 9 / 72

 Code Example

Instrument

Our callable swap is represented as two instruments, the swap without the
call right and the call right itself. The former can be constructed as
boost::shared_ptr<FloatFloatSwap> swap(new FloatFloatSwap( ... ));

and the latter as a swaption

boost::shared_ptr<FloatFloatSwap>
underlying(new FloatFloatSwap( ... ));

boost::shared_ptr<Exercise>
exercise(new BermudanExercise(exerciseDates));

boost::shared_ptr<FloatFloatSwaption>
swaption(new FloatFloatSwaption(underlying, exercise));

Peter Caspers (IKB) Markov Functional Model November 13, 2013 10 / 72

 Code Example

Pricing Engine

To do the actual pricing we need a suitable pricing engine which can be

constructed by
boost::shared_ptr<PricingEngine>
engine(new Gaussian1dFloatFloatSwaptionEngine(markov));
and assigned to our instrument by means of
swaption->setPricingEngine(engine);

which then allows to extract the dirty npv

Real npv = swaption->NPV();

The npv of the swap on the other hand can be retrieved with standard cms
coupon pricers implementing a replication in some swap rate model 2 .

2
this is slightly different from pricing the full structured swap in the markov
model, since there are theoretical differences between the models (in addition to
differences in their numerical solution)
Peter Caspers (IKB) Markov Functional Model November 13, 2013 11 / 72
 Code Example

A full example: Market Data

#include <ql/quantlib.hpp>
using namespace QuantLib;

int main(int, char * []) {

try {

Date refDate(13, November, 2013);

Date settlDate = TARGET().advance(refDate, 2 * Days);
Settings::instance().evaluationDate() = refDate;

Handle<Quote> rateLevel(new SimpleQuote(0.03));

Handle<YieldTermStructure> yts(
new FlatForward(refDate, rateLevel, Actual365Fixed()));

boost::shared_ptr<IborIndex> iborIndex(new Euribor(6 * Months, yts));

boost::shared_ptr<SwapIndex> swapIndex(
new EuriborSwapIsdaFixA(10 * Years, yts));

iborIndex->addFixing(refDate, 0.0200);
swapIndex->addFixing(refDate, 0.0315);

Handle<Quote> volatilityLevel(new SimpleQuote(0.30));

Handle<SwaptionVolatilityStructure> swaptionVol(
new ConstantSwaptionVolatility(refDate, TARGET(), Following,
volatilityLevel, Actual365Fixed()));

Peter Caspers (IKB) Markov Functional Model November 13, 2013 12 / 72

 Code Example

A full example: Cms Swap and Exercise Schedule

Date termDate = TARGET().advance(settlDate, 10 * Years);

Schedule sched1(settlDate, termDate, 1 * Years, TARGET(),

ModifiedFollowing, ModifiedFollowing,
DateGeneration::Forward, false);
Schedule sched2(settlDate, termDate, 6 * Months, TARGET(),
ModifiedFollowing, ModifiedFollowing,
DateGeneration::Forward, false);

Real nominal = 100000.0;

boost::shared_ptr<FloatFloatSwap> cmsswap(new FloatFloatSwap(
VanillaSwap::Payer, nominal, nominal, sched1, swapIndex,
Thirty360(), sched2, iborIndex, Actual360(),
false,false,1.0,0.0,Null<Real>(),Null<Real>(),1.0,0.00267294));

std::vector<Date> exerciseDates;
std::vector<Date> sigmaSteps;
std::vector<Real> sigma;

sigma.push_back(0.01);
for (Size i = 1; i < sched1.size() - 1; i++) {
exerciseDates.push_back(swapIndex->fixingDate(sched1[i]));
sigmaSteps.push_back(exerciseDates.back());
sigma.push_back(0.01);
}

Peter Caspers (IKB) Markov Functional Model November 13, 2013 13 / 72

 Code Example

A full example: Call Right

boost::shared_ptr<Exercise> exercise(
new BermudanExercise(exerciseDates));
boost::shared_ptr<FloatFloatSwaption> callRight(
new FloatFloatSwaption(cmsswap, exercise));

std::vector<Date> cmsFixingDates(exerciseDates);
std::vector<Period> cmsTenors(exerciseDates.size(), 10 * Years);

Peter Caspers (IKB) Markov Functional Model November 13, 2013 14 / 72

 Code Example

A full example: Models and Engines

Handle<Quote> reversionLevel(new SimpleQuote(0.02));

boost::shared_ptr<NumericHaganPricer> haganPricer(
new NumericHaganPricer(swaptionVol,
GFunctionFactory::NonParallelShifts,
reversionLevel));
setCouponPricer(cmsswap->leg(0), haganPricer);

boost::shared_ptr<MarkovFunctional> mf(new MarkovFunctional(

yts, reversionLevel->value(), sigmaSteps, sigma, swaptionVol,
cmsFixingDates, cmsTenors, swapIndex));

boost::shared_ptr<Gaussian1dFloatFloatSwaptionEngine> floatEngine(
new Gaussian1dFloatFloatSwaptionEngine(mf));

callRight->setPricingEngine(floatEngine);

Peter Caspers (IKB) Markov Functional Model November 13, 2013 15 / 72

 Code Example

A full example: Calibration Basket

boost::shared_ptr<SwapIndex> swapBase(
new EuriborSwapIsdaFixA(30 * Years, yts));

std::vector<boost::shared_ptr<CalibrationHelper> > basket =

callRight->calibrationBasket(swapBase, *swaptionVol,
BasketGeneratingEngine::Naive);

boost::shared_ptr<Gaussian1dSwaptionEngine> stdEngine(
new Gaussian1dSwaptionEngine(mf));

for (Size i = 0; i < basket.size(); i++)

basket[i]->setPricingEngine(stdEngine);

Peter Caspers (IKB) Markov Functional Model November 13, 2013 16 / 72

 Code Example

A full example: Model Calibration and Pricing

LevenbergMarquardt opt;
EndCriteria ec(2000, 500, 1E-8, 1E-8, 1E-8);
mf->calibrate(basket, opt, ec);

std::cout << "model vol & swaption market & swaption model \\\\" << std::endl;
for (Size i = 0; i < basket.size(); i++) {
std::cout << mf->volatility()[i] << " & "
<< basket[i]->marketValue() << " & "
<< basket[i]->modelValue() << " \\\\" << std::endl;
}
std::cout << mf->volatility().back() << std::endl;

Real analyticSwapNpv = CashFlows::npv(cmsswap->leg(1), **yts, false) -

CashFlows::npv(cmsswap->leg(0), **yts, false);
Real callRightNpv = callRight->NPV();
Real firstCouponNpv = - cmsswap->leg(0)[0]->amount() * yts->discount(cmsswap->leg(0)[0]->date()) +
cmsswap->leg(1)[0]->amount() * yts->discount(cmsswap->leg(1)[0]->date());
Real underlyingNpv = callRight->result<Real>("underlyingValue") + firstCouponNpv;

std::cout << "Swap Npv (Hagan) & " << analyticSwapNpv << "\\\\" << std::endl;
std::cout << "Call Right Npv (MF) & " << callRightNpv << "\\\\" << std::endl;
std::cout << "Underlying Npv (MF) & " << underlyingNpv << "\\\\" << std::endl;
std::cout << "Model trace : " << std::endl << mf->modelOutputs() << std::endl;
}
catch (std::exception &e) {
std::cerr << e.what() << std::endl;
return 1;
}
}

Peter Caspers (IKB) Markov Functional Model November 13, 2013 17 / 72

 Code Example

A full example: Model Calibration (Coterminals)

model vol swaption market swaption model

0.01 0.0273707 0.0273706
0.0107835 0.0337373 0.0337373
0.0106409 0.0354739 0.0354741
0.0109936 0.0344716 0.0344715
0.0109297 0.0315097 0.0315096
0.0111361 0.0270958 0.0270957
0.0111917 0.0215014 0.0215015
0.0112365 0.0150493 0.0150493
0.0113241 0.00783272 0.00783278
0.00998595

Peter Caspers (IKB) Markov Functional Model November 13, 2013 18 / 72

 Code Example

A full example: Pricing

The expectation is to get a similar price of the underlying cms swap both
in the Markov and the replication model, since both are consistent with
the input swaption smile. Note that the underlying of the swaption is
receiving the cms side while the cms swap is paying.

Swap Npv (Hagan) -0.00

Call Right Npv (MF) 604.50
Underlying Npv (MF) 1.00

The match is very close (0.1 bp times the nominal). It should be noted
that from theory we can not even expect a perfect match since the rate
dynamics is not the same in the Hagan model and the Markov model
respectively.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 19 / 72

 Code Example

A full example: Benchmarking against Hull White

We change the code to replace the Markov model with a Hull White
model. This is easily done by
std::vector<Date> sigmaSteps2(sigmaSteps.begin(), sigmaSteps.end() - 1);
std::vector<Real> sigma2(sigma.begin(), sigma.end() - 1);
boost::shared_ptr<Gsr> gsr(new Gsr(yts,sigmaSteps2,sigma2,reversionLevel->value()));

and replacing mf by gsr (we could have use a generic name of course...).
The calibration now reads
gsr->calibrate(basket, opt, ec, Constraint(), std::vector<Real>(), model->FixedReversions());

or also
gsr->calibrateVolatilitiesIterative(basket, opt, ec);

Peter Caspers (IKB) Markov Functional Model November 13, 2013 20 / 72

 Code Example

A full example: Pricing in the Hull White model

The fit to the coterminals is exact, just as above (with different model
volatilities of course). The pricing compares ot the Markov model as
follows:

Swap Npv (Hagan) -0.00 -0.00

Markov Hull White
Call Right Npv 604.50 1012
Underlying Npv 1.00 657.81

The underlying and the option are drastically overpriced (both) by over
60bp in the Hull White model.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 21 / 72

 Code Example

A full example: A more realistic smile surface

Swaption smiles are usually far from being flat. We test the model with a
SABR volatility cube (with constant parameters α = 0.15, β = 0.80,
ν = 0.20, ρ = −0.30) by setting
Handle<SwaptionVolatilityStructure> swaptionVol(
new SingleSabrSwaptionVolatility(refDate, TARGET(), Following, 0.15,
0.80, -0.30, 0.20,
Actual365Fixed(), swapIndex));

Swap Npv (Hagan) 246.00 246.00

Markov Hull White
Call Right Npv 696.96 1004.79
Underlying Npv 259.96 648.26

Again the underlying and the option is overpriced (by 40bp resp. 53bp) in
the Hull White model. In the Markov model the fit is still good (1.4bp
underlying price difference).

Peter Caspers (IKB) Markov Functional Model November 13, 2013 22 / 72

 Code Example

A full example: A more realistic smile surface ctd

Example smile 10y into 10y, α = 0.15, β = 0.80, ν = 0.20, ρ = −0.30
0.8

0.7
implied lognormal black volatility

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.02 0.04 0.06 0.08 0.1
strike
Peter Caspers (IKB) Markov Functional Model November 13, 2013 23 / 72
 Code Example

A full example: Model diagnostics

The markov model can generate information on the calibration process

with
std::cout << "Model trace : " << std::endl << mf->modelOutputs()
<< std::endl;
The output contains information on
numerical model parameters
settings for smile preconditioning
yield term structure calibration results
volatility smile calibration results
and can help identifying calibration problems, resp. confirm a successful
calibration.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 24 / 72

 Code Example

A full example: Model diagnostics (model parameters /

smile settings)

Markov functional model trace output

Model settings
Grid points y : 64
Std devs y : 7
Lower rate bound : 0
Upper rate bound : 2
Gauss Hermite points : 32
Digital gap : 1e-05
Adjustments : Kahale SmileExp
Smile moneyness checkpoints:

Peter Caspers (IKB) Markov Functional Model November 13, 2013 25 / 72

 Code Example

A full example: Model diagnostics (yield term structure fit)

The raw output

Yield termstructure fit:
expiry;tenor;atm;annuity;digitalAdj;ytsAdj;marketzerorate;modelzerorate;diff(bp)
November 13th, 2014;10Y;0.03047961644430512;8.267635590910791;1;1;0.03000000000000008;0.03008016185429107;-0.8
November 12th, 2015;10Y;0.0304803836917478;8.023747530203391;1;1;0.02999999999999999;0.03003959599137232;-0.39
[...]

is meant to be analyzed in another application like office

Peter Caspers (IKB) Markov Functional Model November 13, 2013 26 / 72

 Code Example

A full example: Model diagnostics (volatility smile fit)

Peter Caspers (IKB) Markov Functional Model November 13, 2013 27 / 72

 Theoretical Background

Axiomatic Hull White

Any gaussian one factor HJM model which satisfies separability, i.e.

σf (t, T ) = g(t)h(T ) (1)

for the instantaneous forward rate volatility with deterministic g, h > 0,

necessarily fulfills

dr(t) = (θ(t) − a(t)r(t))dt + σ(t)dW (t) (2)

for the short rate r, which means, it is a Hull White one factor model.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 28 / 72

 Theoretical Background

The T-forward numeraire

Set x(t) := r(t) − f (0, t) and fix a horizon T , then in the T -forward
measure the numeraire can be written

P (0, T ) −x(t)A(t,T )+B(t,T )

N (t) = P (t, T ) = e (3)
P (0, t)
with A, B dependent on the model parameters. The Hull White model is
called an affine model.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 29 / 72

 Theoretical Background

Smile in the Hull White Model

The distribution of N (t) is lognormal. The shape of the distribution

can not be controlled by any of the model parameters.
For fixed t you can calibrate the model to one market quoted interest
rate optoin (typically a caplet or swaption).
You can choose the strike of the option, but the rest of the smile is
implied by the model.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 30 / 72

 Theoretical Background

Callable vanilla swaps

Pricing of callable fix versus Libor swaps may be done in a Hull White
model which is calibrated as follows:
For each call date find a market quoted swaption which is equivalent
to the call right (in some sense, e.g. by matching the npv and its first
and second derivative of the underlying at E(x(t))).
Calibrate the volatility function σ(t) to match the basket of these
swaptions.
Choose the mean reversion of the model to control serial correlations.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 31 / 72

 Theoretical Background

Intertemporal correlations

To understand the role of the reversion parameter assume σ and a

constant for a moment. Then it is easy to see
s s
e2aT2 −1 −a(T2 −T1 ) 1 − e−2aT1
corr(x(T1 ), x(T2 )) = = e (4)
2aT
e 1 −1 1 − e−2aT2
p
which shows that for a = 0 the correlation is T1 /T2 and goes to zero if
a → ∞ and to one if a → −∞.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 32 / 72

 Theoretical Background

Callable cms swaps

The calll rights in a callable cms swap are options on a swap exchanging
cms coupons against fix or Libor rates. Such underlying swaps are
drastically mispriced in the Hull White model in general.
cms coupons are replicated using swaptions covering the whole strike
continuum (0, ∞)
The swaption smile in the Hull White model is generally not
consistent with the market smile and so are the prices of cms coupons
Obviously we need a more flexible model to price such structures

Peter Caspers (IKB) Markov Functional Model November 13, 2013 33 / 72

 Theoretical Background

Model requirements

The wishlist for the model is as follows

We want to be capable of calibrating to a whole smile of (constant
maturity) swaptions, not only to one strike, for all fixing dates of the
cms coupons. This is to match the coupons of the underlying.
In addition we would like to calibrate to (possibly strike / maturity
adjusted) coterminal swaptions to match the options representing the
call rights.
Finally we need some control over intertemporal correlations, i.e.
something operating like the reversion parameter in the Hull White
model
The idea to do so is to relax the functional dependency between the state
variable x and the numeraire N (t, x).

Peter Caspers (IKB) Markov Functional Model November 13, 2013 34 / 72

 Model description

The driving process

We start with a markov process driving the dynamics of the model as

follows:

dx = σ(t)eat dW (t) (5)

and x(0) = 0. The intertemporal correlation of the state variable x is the
same as for the Hull White model, see (4), i.e. the parameter a can be
used to control the correlation just as the reversion parameter in the Hull
White model.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 35 / 72

 Model description

The numeraire surface

The model is operated in the T -forward measure, T chosen big enough to

cover all cashflows relevant for the actual pricing under consideration. The
link between the state x(t) and the numeraire P (t, T ) is given by

P (t, T, x) = N (t, x) (6)

which we allow to be a non parametric surface to have maximum flexibility
in calibration.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 36 / 72

 Calibration

Calibrating the numeriare surface to market smiles

The price of a digital swaption paying out an annuity A(t) on expiry t if

the swap rate S(t) ≥ K in our model is
Z ∞
A(t, y)
digmodel = P (0, T ) φ(y)dy (7)
y ∗ P (t, T )
where y ∗ is the strike in the normalized state variable space (the
correspondence between y and S(t) is constructed to be monotonic).

Peter Caspers (IKB) Markov Functional Model November 13, 2013 37 / 72

 Calibration

Implying the swap rate

Given the market smile of S(t) we can compute the market price
digmkt (K) of digitals for strikes K. For given y ∗ we can solve the equation
Z ∞
A(t, y)
digmkt (K) = P (0, T ) φ(y)dy (8)
y ∗ P (t, T )

for K to find the swap rate corresponding to the state variable value y ∗ .
For this digmkt (·) should be a monotonic function whose image is equal to
the possible digital prices (0, A(0)]. We will revisit this later.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 38 / 72

 Calibration

Computing the deflated annuity

To compute the deflated annuity

n
A(t) X P (t, tk )
= τk (9)
P (t, T ) P (t, T )
k=1

we observe that
 
P (t, u) 1
=E y(t) (10)
P (t, T ) y(t) P (u, T )
i.e. we have to integrate the reciprocal of the numeraire at future times.
Working backward in time we can assume that we know the numeraire at
these times (starting with N (T ) ≡ 1).

Peter Caspers (IKB) Markov Functional Model November 13, 2013 39 / 72

 Calibration

Converting swap rate to numeraire

Having computed the swap rate S(t) we have to convert this value to a
numeraire value N (t). Since

S(t)A(t) + P (t, t∗ ) = 1 (11)

we get (by division by N (t))
1
N (t) = A(t) P (t,t∗ )
(12)
S(t) N (t) + N (t)

all terms on the right hand side computable via deflated zerobonds as
shown above. Note that we use a slightly modified swap rate here, namely
one without start delay.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 40 / 72

 Calibration

Calibration to a second instrument set

Up to now we have not made use of the volatility σ(t) in the driving
markov process of the model. This parameter can be used to calibrate the
model to a second instrument set, however only a single strike can be
matched obviously for each expiry. A typical set up would be
calibrate the numeraire to an underlying rate smile, e.g. constant
maturity swaptions for cms coupon pricing
calibrate σ(t) to (standard atm or possibly adjusted) coterminal
swaptions for call right calibration
Note that after changing σ(t) the numeraire surface needs to be updated,
too.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 41 / 72

 Calibration

Input smile preconditioning

To ensure a bijective mapping

digmkt : (0, ∞) → (0, A(0)) (13)

it is sufficient to have an arbitrage free input smile with a C 1 call price
function. It is possible to allow for negative rates and generalize the
interval (0, ∞) to (−κ, ∞) with some suitable κ > 0, e.g. κ = 1%. In
general input smiles are not arbitrage free, so some preconditioning is
advisable, since arbitrageable smiles will break the numeraire calibration.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 42 / 72

 Calibration

Kahale extrapolation
SABR 14y/1y implied black lognormal volatilities as of 14-11-2012, input (solid)
and Kahale (dashed)
0.9

0.8

0.7
black lognormal volatility

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.05 0.1 0.15 0.2
Peter Caspers (IKB) Markov strike
Functional Model November 13, 2013 43 / 72
 Calibration

Kahale extrapolation
SABR 14y/1y call prices as of 14-11-2012, input (solid) and Kahale (dashed)
0.035

0.03

0.025

0.02
call

0.015

0.01

0.005

0
0 0.05 0.1 0.15 0.2
strike
Peter Caspers (IKB) Markov Functional Model November 13, 2013 44 / 72
 Calibration

Kahale extrapolation
SABR 14y/1y digital prices as of 14-11-2012, input (solid) and Kahale (dashed)
1

0.8

0.6
digital call

0.4

0.2

0
0 0.05 0.1 0.15 0.2
strike
Peter Caspers (IKB) Markov Functional Model November 13, 2013 45 / 72
 Calibration

Kahale extrapolation
SABR 14y/1y density as of 14-11-2012, input (solid) and Kahale (dashed)

40

20
density

-20

-40

0 0.05 0.1 0.15 0.2

strike
Peter Caspers (IKB) Markov Functional Model November 13, 2013 46 / 72
 Numerics

Interpolation of the numeraire

In a numerical implementation of the model we will need to discretize the

numeraire surface on a grid (ti , yj ).
in y-direction we interpolate N (t, y) (normalized by todays market
forward numeraire) with monotonic cubic splines with Lagrange end
condition. Outside a range of specified standard deviations (defaulted
to 7), we extrapolate flat.
in t-direction we interpolate the reciprocal of the normalized
numeraire linearly. This ensures a perfect match of todays input yield
curve even for interpolated times as can easily be seen from (10).
After the horizon T we extrapolate flat (N ≡ 1 there anyway), only
for technical reasons.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 47 / 72

 Numerics

Sample numeraire surface

Numeraire surface for market data as of 14-11-2012

1 1.1
0.8 1
0.9
N(t,y) 0.6 0.8
0.4 0.7
0.2 0.6
0.5
0 0.4
0.3

-3
-2
-1
0 0
y 1 4 2
2 8 6
12 10 t
3 14

Peter Caspers (IKB) Markov Functional Model November 13, 2013 48 / 72

 Numerics

Interpolation of payoffs

Payoffs occuring in the numeraire bootstrap (digitals) or later in pricing

exotics are also interpolated with Lagrange splines. We leave it as an
option to
restrict to integration of the payoff to a specified number of standard
deviations,
extrapolate the payoff flat
extrapolate the payoff according to the Lagrange end condition
The results should not depend significantly on this choice, otherwise the
numerical parameters should be increased.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 49 / 72

 Numerics

Numerical Integration for deflated zero bonds

To compute deflated zerobond prices according to (10) it turns out that it

is fast and accurate to
use the celebrated Gauss Hermite Integration scheme where
32 points are more than enough usually to ensure a good accuracy.
This is because the integrand is globally well approximated by polynomials.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 50 / 72

 Numerics

Numerical Integration for payoffs

For the numerical integration of

digitals during numeraire bootstrapping or
exotic pricing
Gauss Hermite is possible but not leading to satisfactory accuracy. This is
due to the non global nature of the integrand in this case. Here we rely on
exact integration of the piecewise 3rd order polynomials against the
gaussian density, which is possible in closed form only involving the error
function erf

Peter Caspers (IKB) Markov Functional Model November 13, 2013 51 / 72

 Numerics

How is the yield curve matched after all ?

It is interesting to note that the initial yield curve is matched by

calibrating the numeraire to market input smiles. The yield curve is never
a direct input though, as it is e.g. for the Hull White model. It is
reconstructed by the model via the numeraire density coded in and read
off the market smile during numeraire bootstrapping.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 52 / 72

 Numerics

Long term constant maturity calibration

The hardest case of calibration is a long term calibration to constant

maturity swaptions (or caplets).
We start at maturity T and bootstrap the numeraire at some tk < T ,
introducing some numerical noise in N (tk ).
We then boostrap N (tk−1 ) for tk−1 < tk , relying on the already
bootstrapped future numeraire values.
In case of a coterminal calibration we largely still use N (T ) and
N (tk ) only to a smaller degree.
In case of cm calibration however we have to rely on N (tk ) with tk
nearer to our calibration time t.
Therefore in long term cm calibration numerical noise may pile up giving
large errors for the shorter term numeraire surface.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 53 / 72

 Numerics

Yield curve match with standard numerical parameters

Fit to Yts flat @ 3%, 2y cm swaptions @ 20%, 7 standard deviations, 200 % upper rate bound

0.031

0.0308

0.0306

0.0304

0.0302
zero rate

0.03

0.0298

0.0296

0.0294

0.0292

0.029
0 10 20 30 40 50
maturity
Peter Caspers (IKB) Markov Functional Model November 13, 2013 54 / 72
 Numerics

Increasing numerical accuracy for long term cm baskets

The first idea in this case is to increase the numerical accuracy by

increasing
the number of covered standard deviations (e.g. from 7 to 12)
the upper cut off point for rates (e.g. from 200% to 400%)
(maybe the number of discretization points, though less critical)
However this breaks the calibration totally, because the standard
(”double”) 53 bit mantissa numerical precision is not sufficient to do the
computations numerically stable any more.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 55 / 72

 Numerics

NTL high precision computing

The NTL and boost libraries provide support for arbitrary floating point
precision. We incorporated NTL support as an option replacing the
standard double precision by an arbitrary mantissa length precision in the
critical sections of the computation (which turned out to be the integration
of payoffs against the gaussian density, where large integrand values are
multiplied by small density values). NTL is activated by commenting out
// uncomment to enable NTL support
#define GAUSS1D_ENABLE_NTL

in gaussian1dmodel.hpp. The mantissa length is then set (e.g. to

113bit) with
boost::math::ntl::RR:SetPrecision(113)

Peter Caspers (IKB) Markov Functional Model November 13, 2013 56 / 72

 Numerics

Yield curve match using high precision computing

Fit to Yts flat @ 3%, 2y cm swaptions @ 20%, 150bit mantissa, 12 standard deviations, 400 % upper rate bound

0.03001

0.030008

0.030006

0.030004

0.030002
zero rate

0.03

0.029998

0.029996

0.029994

0.029992

0.02999
0 10 20 30 40 50
maturity
Peter Caspers (IKB) Markov Functional Model November 13, 2013 57 / 72
 Numerics

A more pragmatic approach: Adjustment factors

Computations with NTL are slow. A more practical way to stabilize the
calibration in difficult circumstances are adjustment factors forcing the
numeraire to match the market input yield curve. The adjustment factor is
introduced by replacing

P model (0, ti )
N (ti , yj ) → N (ti , yj ) (14)
P market (0, ti )
This option should be used with some care because it may lower the
accuracy of the volatility smile match. In most situations the adjustment
factors are moderate though, in the example we had before:

Peter Caspers (IKB) Markov Functional Model November 13, 2013 58 / 72

 Numerics

Adjustment factors in the example above

Date Adjustment Factor
November 14th, 2013 1.00000029079227
November 14th, 2014 0.999999566861981
November 14th, 2015 0.999999720414697
November 14th, 2016 0.999999838009949
November 14th, 2017 0.999999380770489
... ...
November 14th, 2056 0.999804362556495
November 14th, 2057 0.999904959217797
November 14th, 2058 0.99988000473961
November 14th, 2059 0.999816723715493
November 14th, 2060 0.999830021528368
November 14th, 2061 0.999737006370401
November 14th, 2062 0.999761860227793
November 14th, 2063 0.999887598113548

Peter Caspers (IKB) Markov Functional Model November 13, 2013 59 / 72

 Numerics

MarkovFunctional Options - Numerical Parameters

yGridPoint Number of discretization points for normalized state

variable (default 64)
yStdDevs Number of standard deviations of normalized state variable
covered (default 7)
gaussHermitePoints Number of points used in Gauss hermite
integration (default 32)
digitalGap Numerical stepsize to compute market digitals from call
spreads (default 10−5 )
marketRateAccuracy Accuracy (in model swap rate) when matching
market digitals (default 10−7 )
lowerRateBound Lower bound for model’s (underlying) rates (default
0%)
upperRateBound Upper bound for model’s (underlying) rates
(default 200%)
Peter Caspers (IKB) Markov Functional Model November 13, 2013 60 / 72
 Numerics

MarkovFunctional Options - Adjustments Overview

AdjustDigitals – force the model’s digital at the lower rate bound

to be 1. Limited impact, not recommended, does not seem to be
useful in practice.
AdjustYts – force the model’s yts to match the input yts by
applying a constant adjustment factor per expiry. Recommended only
for very long term calibrations, smile fit may suffer
ExtrapolatePayoffFlat – Instead of extrapolating using the spline,
extrapolate the digital’s payoff flat outside the given standard
deviations of the state variable. Recommended only for sanity checks,
should have no impact, otherwise numerical parameters have to be
adjusted
NoPayoffExtrapolation – Do not extrapolate the payoff at all.
Same as for ExtrapolatePayoffFlat holds

Peter Caspers (IKB) Markov Functional Model November 13, 2013 61 / 72

 Numerics

MarkovFunctional Options - Adjustments Overview ctd.

KahaleSmile – Use Kahale smile preconditioning on arbitrageable

wings, recommended (or SabrSmile) unless you are sure to provide
an arbitrage free smile input
KahaleInterpolation – Use Kahale smile interpolation (implies
KahaleSmile), for naively interpolated smiles (linear, ...), however
generates “Bat Man shaped” Densities, not recommended unless
SabrSmile fails to work or is too inaccurate w.r.t. fit
SmileExponentialExtrapolation – Use Exponential smile
extrapolation on right wing (implies KahaleSmile), for slowly
decreasing call price functions, recommended

Peter Caspers (IKB) Markov Functional Model November 13, 2013 62 / 72

 Numerics

MarkovFunctional Options - Adjustments Overview ctd.

SmileDeleteArbitragePoints – Delete intermediate points from

smile that cause arbitrage (implies KahaleInterpolation),
recommended if you use KahaleInterpolation
SabrSmile – Use a Sabr fitted smile (with Kahale wing sanitization),
recommended for naively interpolated smiles
SmileMoneynessCheckPoints – Custom strike grid for smile
arbitrage check and Kahale interpolation in K/F space (optional),
default is working in most cases

Peter Caspers (IKB) Markov Functional Model November 13, 2013 63 / 72

 Secondary instrument set calibration

Representative Basket Approach

We are coming back to the question to which secondary instrument basket

we should calibrate to represent cms / float (or other exotic) swaptions.
Problem: Find standard swaptions that represent an exotic call right
appropriately in your Markov Functional (or any one factor) model.
Strategy: Match NPV, Delta3 , Gamma around some “central” value
of the model’s state variable. This works well for amortizing, accreting
and step up/down swaptions for example. It is worth a try if we can
use the method also in our initial example with the CMS swaption.

3
which means the first derivative w.r.t. the state variable of the model here
Peter Caspers (IKB) Markov Functional Model November 13, 2013 64 / 72
 Secondary instrument set calibration

Global amortizing vanilla swap fit

(solid = exotic npv, dotted = standard underlying npv)
10

0
npv

-5

-10

-15

-20
-4 -2 0 2 4
y
Peter Caspers (IKB) Markov Functional Model November 13, 2013 65 / 72
 Secondary instrument set calibration

Global CMS swap fit (Flat smile)

(solid = exotic npv, dotted = standard underlying npv)
5000

-5000
npv

-10000

-15000

-20000

-25000
-4 -2 0 2 4
y
Peter Caspers (IKB) Markov Functional Model November 13, 2013 66 / 72
 Secondary instrument set calibration

Global CMS swap fit (SABR smile)

(solid = exotic npv, dotted = standard underlying npv)
5000

-5000
npv

-10000

-15000

-20000

-25000
-4 -2 0 2 4
y
Peter Caspers (IKB) Markov Functional Model November 13, 2013 67 / 72
 Secondary instrument set calibration

Representative Basket Approach - Conclusion

It does not seem to be clear to what secondary instrument set you should
calibrate your model in order to represent calls on a cms (or more generally
float float) swaps, although in practice atm coterminals seem to be a
common choice.

Peter Caspers (IKB) Markov Functional Model November 13, 2013 68 / 72

 Implementation in QuantLib

QuantLib Implementation

A base version of the model is available since release 1.3. Recent

developemts (≥ 1.4 ?) include
support for a (zero volatility) spread between discounting and
forwarding curves
a float float and non standard swaption pricing engine
representative calibration basket generation
support for pricing under credit risk (credit linked swaptions, exotic
bonds)
SABR-Kahale smile preconditioning

Peter Caspers (IKB) Markov Functional Model November 13, 2013 69 / 72

 Implementation in QuantLib

Outlook / Future Tasks

Finite Difference pricing engines

Additional smile preconditioning algorithms

Peter Caspers (IKB) Markov Functional Model November 13, 2013 70 / 72

 References

References

Johnson, Simon: Numerical methods for the markov functional model,

Wilmott magazine, http://www.wilmott.com/pdfs/110802 johnson.pdf
Kahale, Nabil: An arbitrage free interpolation of volatilities, Risk May
2004, http://nkahale.free.fr/papers/Interpolation.pdf
Shoup, Victor: NTL A Library for doing Number theory,
http://www.shoup.net/ntl/
QuantLib A free/open-source library for quantitative finance,
http://www.quantlib.org
Caspers, Peter: Markov Functional One Factor Interest Rate Model
Implementation in QuantLib,
http://papers.ssrn.com/sol3/papers.cfm?abstract id=2183721
Caspers, Peter: Representative Basket Method Applied,
http://papers.ssrn.com/sol3/papers.cfm?abstract id=2320759

Peter Caspers (IKB) Markov Functional Model November 13, 2013 71 / 72

 Questions

Thank you, Questions ?

Peter Caspers (IKB) Markov Functional Model November 13, 2013 72 / 72