The SABR model

Daniel Brødsgaard

University of Copenhagen - Department of Economics

14 November 2022
Overview of today

▶ General model considerations

▶ From constant vol to local vol
▶ From local vol to stochastic vol
▶ SABR model
▶ Backbone and skew
▶ Effect on risk measures
▶ SABR parameters and risk
▶ Calibration and concluding remarks
General model considerations
But first... A quiz!

▶ In a casino you realise that the roulette is biased, so that red comes
up 70% of the time.
▶ There is no green, ie. the casino earns nothing in expectation.
▶ The odds are 50/50.
▶ This means that if you bet 50 on black or red and you colour comes
up, you gain 100.
▶ You decide to start a derivatives market where you, the market
maker, quote prices on tickets that pay 100 if red comes up, zero if
not.
▶ Q: Quote a two-way price for the tickets.
▶ Task illustrates again the difference between the P and Q measures:
speculators vs hedgers.
Motivation

▶ The BS assumption is to assume that implied vol is a constant

across strikes and expiries. dS(t) = σS(t)dW (t)
▶ This is, however, contradicted by markets. In the real world markets
imply a skew/smile in the implied vol versus strike plot, and there is
a term-structure in vol-markets as well further contradicting this
simple assumption.
▶ Also, when pricing exotics, which depends on skews, vol-of-vol or
correlation between the underlying and volatility, the BS model are
not able to take such factors into account leading to mis-pricing.
▶ Specifically, in such cases the BS model will imply a price, which
can be arbitraged.
▶ Today we are gonna expand slightly on our knowledge of models,
and how greeks changes when employing different assumptions for
the dynamics of the underlying.
Why do we need models?
▶ We need models for at least three reasons:
1. To calculate risk.
2. For inter- and extra-polation purposes.
3. To price non-vanilla stuff.
▶ Example of 1: We need a model which matches market-implied
moves as good as possible in order to generate good hedge
properties. If my model says I need to hedge a risk of higher implied
vol when the underlying declines this needs to be reflected in how
the market actually moves. Otherwise my hedges will be wrong.
▶ Example of 2: Options are typically only traded liquidly close to
ATM at specific expiries. What if I need to price a 100bp OTM
option? How can I interpolate as good as possible and make sure I
don’t get arbitraged?
▶ Example of 3: Say we want to price a forward-starting option or a
Bermudan swaption. In these cases no closed form solutions exist,
and further they are not liquidly traded. In these cases we need
models calibrated to vanilla markets (to make sure we don’t get
arbitraged) in order to do pricing and hedging, most often by the
use of numerical (simulation) methods.
Term structure chart

Figure 1: Term structure of vol. 10y tail on varying expiries

Vol skew/smile chart

Figure 2: Smile and skew. Varying expiries on 10y tails.

Where does it come from? Note flattening in expiry (jumps).

What happens in badly specified models?
Remember from the last lecture that, when delta hedging a long position
in a European option, we have

"Z #
Z T T
1
Γ(u)S 2 (u) σr2 (u) − σBS
2


Et [ PnL(u)du] = Et du
t 2 t

▶ Under the BS model the σBS 2

is our assumed model implied
volatility, ie. the size of the move implied by the model, whereas σr2
is the (unknown) realized one.
▶ If the model is mis-specified books will bleed or make money slowly
if:
1. where the dealer is long gamma and the model implied moves
are larger than the ones actually realized he will bleed (else
make money).
2. where the dealer is short gamma and the model implied moves
are lower than the ones actually realized he will bleed (else
make money).
Badly specified models continued
▶ Interestingly, the dealer often ends out long gamma if he has a
model implying relatively large moves, since this means ”expensive
options”. So sophisticated clients and others will sell options to him.
▶ Also, the dealer end out short gamma if he has a model implying
relatively small moves, since this means ”cheap options”.
▶ Lesson: Badly specified models hurt the dealer almost always.
▶ In some cases it might be that one dealer has a model superior to
the others (we see such an example later today), which correctly
takes into account relevant risk factors which other dealers’ models
don’t.
▶ Consider for an example a dealer in possession of a model correctly
giving rise to relatively expensive (high implied vol) OTM
swaptions, whereas other dealers’ models don’t. In this case no
clients buy from him, but instead from the other dealers which bleed
PnL every day. Clients and other dealers will sell to him.
▶ He will be able to make a positive PnL (since his model is correct)
but it can take a looooong time. Maybe just publish the better
model?
On Black-Sholes
▶ The BS model is clearly mis-specified in that volatility can obviously
not assumed to be constant. However, the model remains valid in
the sense we are able to explain our PnL stemming from realized
versus implied vol.
▶ Traders spend a lot of time explaining their PnL.
▶ Thus, by using the BS model we take on (known) volatility risks.
The issue here is: is that acceptable to us to take on that risk?
▶ If not, we need to hedge volatility risks.
▶ In order to hedge volatility risks we need a model which can
generate relevant risk figures wrt. volatility. This is not possible in
BS since vol is assumed constant.
▶ Note that the BS model remains highly popular still even in the face
of known drawbacks.
▶ The critique is not entirely fair to the BS model. Their greatest
advance was not to propose a vol specification for the underlying,
but instead to come up with the delta hedging argument seen in the
last lecture, which leads to a PDE which can be solved.
Smile problems
▶ Even if the BS model may indeed remain valid there is a range of
issues presenting itself once handling portfolios or more complex
(exotic) products. It can even lead to arbitrages.
▶ For an example consider having sold a swaption at strike x and
bought another at strike y. Can I then be assured that the greeks
produced at these two strike levels offset each other? Are they even
additive?
▶ Note that in the BS framework when having to make use of two
different implied vols for two different strike levels essentially
amount to employing two different models in order to price two
different swaptions.
▶ We are gonna solve that problem today.
▶ Another issue presents itself in slightly more exotic options which
are skew/smile dependent. Such as digital options. Here the
assumption of constant vol means pricing can be off by a lot for
even a very simple product.
From constant vol to local vol
Implied distributions from option prices
The market disagreeing with BS is the same as the market implying a
different underlying distribution for the terminal value of the underlying
(the value at expiry). To see this let ϕT (x) be the density of a random
variable x with expiry T . Then

Z + inf
C (K , T ) = (x − K )+ ϕT (x)dx
− inf

This can be differentiated twice wrt. to K to yield

Z + inf 2
∂ 2 C (K , T ) ∂ (x − K )+
2
= ϕT (x)dx
∂K − inf ∂K 2
Z + inf
= δ(K )ϕT (x)dx = ϕT (K )
− inf

where δ is Dirac’s delta function. If BS holds then ϕT (K ) = ϕT ie.

independent of strike.
K in C (K , T ) is a strike. K in ϕT (K ) is the terminal stock price.
Implied distributions continued
▶ Full knowledge of call and put prices can be used to infer the
(risk-neutral) terminal distribution of the underlying. Note: Equal
to the price of an (infinitely) tight butterfly which we know (from
the appendix of last lecture) is a strictly positive function.
▶ So if the market prices a higher implied vol (higher price) at some
strike levels relative to others, this is equivalent to saying that
markets assign a larger (risk-neutral) probability mass to these strike
levels relative to what is implied from the log-normal assumption in
BS.
▶ Note that the gamma figure from yesterday sort of looks like a
density and the delta figure a distribution function. Surprised?

Figure 3: Left: Delta of payer. Right: Gamma of payer.

Local vol with spot dependence
Local volatility models expand the BS framework by assuming a diffusion
of the form

dS(t) = σ(S, t)S(t)dW (t)

From the last lecture we know how we can use local vols to generate
implied BS volatilities:

hR i
T
Et t
Γ(u)S 2 (u)σ 2 (S, u)du
2
σBS = hR
T
i
Et t Γ(u)S 2 (u)du
RT R
t
ϕ(S, u)Γ(u)S 2 (u)σ 2 (S, u)dSdu
= RT R
t
ϕ(S, u)Γ(u)S 2 (u)dSdu

where ϕ is the density of S at time u seen from time t.

Ie. BS implied vol is a weighted average of local vols between the spot
today (large density) and strike at maturity (large gamma).
Charts showing where weights are large

Figure 4: Left: Weight large due to density. Right: Weight large due to
gamma. Source: Antoine Savine.
Some useful remarks on local vol
Bruno Dupire (1992) showed that if the local volatility function is

2 2∂
C (K ,T )
∂K 2
σ (K , T ) = ∂C (K ,T )
K 2
∂T

then the local volatility model will be able to fit all European option
prices of all strike as well as maturities. Furthermore, two useful
approximations follow in the case where local vol is either independent of
time or describing short maturities:

σBS (K ) ≈ σ(S(t))
K =S(t)
∂σBS (K ) 1 ∂σ(S(t))
≈
∂K K =S(t) 2 ∂S
Ie. implied ATM vol is equal to local vol, and the implied ATM vol skew
(in the strike dimension) is roughly half of the local vol skew (in the spot
dimension).
Some drawbacks of local vol
▶ Given the formulas above we can infer some dynamics. Assume for
an example that 3m expiry options are currently pricing a smile as
shown below (blue line). Note that the local vol function must be a
more convex function in this case, since implied vol is approximately
a weighted average of local vols between spot and strike.
▶ In this case a local vol model predicts that implied ATM vol is
assumed to increase no matter what direction the underlying moves!
▶ Furthermore, if the underling declines then the ATM skew increases
and vice versa. This is not in line with how the market trades.

Figure 5: LV exacerbates moves in IV.

From local to stochastic vol
Stylized facts on smile/skew

Figure 6: Left: Backbone. Right: Structures across time.

▶ ATM vol is typically higher with lower swap rates. Note, however,
that normal volatilities are closer to constant since σnormal ≈ σBS S
▶ The market inhibits a skew/smile at all times, although it can be
steeper/flatter.
Stylized facts on smile/skew: Backbone and skew effect
Figure 7: Implied vol for given strike.

▶ The backbone effect can be seen as a parallel shift of the implied

vol curve - ATM vol shifts to a higher level. Thus, the 3% strike
swaption is for a given moneyness at a higher implied vol.
▶ However, the moneyness is not constant (closer to ATM). Thus, the
(negative) skew effect implies an lower implied vol. The skew effect
in this case dominates the backbone-effect.
Motivating stochastic vol
▶ We know that local vol models can make a single consistent model
able to fit all prices and maturities.
▶ However, the local vol model implies dynamics different from the
ones in the market. In the market we have a negatively sloped
backbone and ATM skew is not necessarily steepening (more
negative nor more positive) when the underlying changes.
▶ Another class of vol models is the stochastic volatility models. In
these models we are able to, as we will see, to include more greeks
than we have in the BS model. These greeks are vital in making
sure that we do not get arbitraged and that we are able to calibrate
to the smile. As we shall see, stochastic vol models are able to
generate smiles and skews in line with the market.
▶ The idea of stochastic vol models is to introduce an extra state
variable, namely volatility itself in addition to interest rates.
▶ In order to complete the models we need to assume that volatility
can be hedged. How can we hedge vol? By trading ATM straddles
(ie. by treading other options).
How can stochastic vol be able to explain smile and skew?
We start building intuition from using Îto once again on a call option,
which is now both a function of the underlying and volatility (we assume
r=0).

dC dC 1 d2C 2 dC 1 d2C 2 d2C

dC (S, σ, t) = dt + dS + (dS) + dσ + (dσ) + (dSdσ)
dt dS 2 dS 2 dσ 2 dσ 2 dσdS

After delta and now also a vega hedge we get

dC 1 d 2C 2 1 d 2C 2 d 2C
dt = − (dS) − (dσ) − (dSdσ)
dt 2 dS 2 2 |dσ 2
{z } |dσdS
{z }
volga vanna

where we have assumed that the quadratic variation terms: (dS)2 and
(dσ)2 , are functions of dt only.
▶ Thus, we have two new greeks, vanna and volga, to take into
account, which the BS model was not able to.
▶ Signs depends on the vol-of-vol (dσ)2 and on correlation between
spot and vol (dSdσ) besides the original gamma which depends on
the vol of spot (dS)2 .
What does volga and vanna look like?
Figure 8: Top-L: Volga. Top-R: Vanna. Low-L: volga
vega . Low-R:
vanna
vega .
Functions of strike (S=100). See appendix for intuition.
More on vanna and volga
We will see now why we above represented volga and vanna in ratios
relative to vega as well.
▶ We have seen that with stochastic vol there are two more greeks to
take into account.
▶ Thus, in a stochastic volatility model we can, relative to the
constant vol in Black-Sholes, write the excess value (E ) for a strike
K as (where SV (K ) the stochastic vol value of the option)

E = SV (K ) − BS(K , σBS (ATM)) = BS(K , σBS (K )) − BS(K , σBS (ATM))

dBS(K , σBS (ATM))
≈ · (σBS (K ) − σBS (ATM))
dσBS
E
=⇒ σBS (K ) ≈ σBS (ATM) +
VegaBS (K , σBS (ATM))

where E = f (vanna, volga). Thus, implied BS vol at a strike K ̸= ATM

is exactly functions of vanna and volga relative to vega.
How can stoch vol explain skew/smile?
We know that

dC 1 d 2C 1 d 2C d 2C
dt = − 2
(dS)2 − 2
(dσ)2 − (dSdσ)
dt 2 dS |2 dσ
{z } |dσdS
{z }
volga vanna

▶ We know that OTM options have positive and increasing volga

vega , and
since (dσ)2 > 0 wings should always be more expensive. Smile.
▶ We know that high strike options (OTM calls and ITM puts) have
positive vanna
vega . Thus, if dSdσ > 0 (ie. positive correlation), then as
the spot increases, implied vol increases, and these options will have
high delta (close to zero for puts, close to 1 for calls), exactly when
the spot is increasing: positive! Positive skew.
▶ We know that low strike options (ITM calls and OTM puts) have
negative vanna
vega . Thus, if dSdσ < 0 (ie. negative correlation), then as
the spot decreases implied vol increases, and these options will have
lower delta (close to -1, for puts, and 0 for calls) exactly when the
spot is decreasing: positive! Negative skew.
SABR model
The start of the SABR model
▶ It can be shown that trading in three different options in a single
expiry is enough to create a position which is only long volga.
▶ Intuitively, if we buy two OTM options (an OTM receiver and an
OTM payer), we are able to hedge vega in these options with one
ATM option. As vanna is roughly of the same size but with
opposite signs on each side of ATM, we have hedged vanna as well.
▶ Also, since theta and gamma is proportional to vega in BS we have
hedged these risks as well.
▶ Thus we have created a position only long volga - a position which
makes money every time implied vol moves and the larger the move
the better.
▶ This was the birth of the SABR model. In the late 90s the swaption
market traded without any smile despite this being the case in other
markets. Traders at some french banks found that they were able to
buy wings and hedge in ATM options and almost always make a
profit.
▶ They asked their quants to build a model which were able to explain
this phenomenon. When published wings were marked higher...
A stochastic vol specification
In the following I make use of the notation from the lecture note.

dft = σt ftα dWt1 , α ∈ (0, 1) If 0 then normal model.

dσt = ϵσt dWt2 , ϵ>0
dWt1 dWt2 = ρdt, ρ ∈ (−1, 1)
σ0 > 0
▶ α can control the skew (see CEV in appendix). However, in practice
this parameter is freezed to some value as we can get the same
effect from the ρ parameter. Instead α controls the backbone.
Thus, important for sensitivity calculations.
▶ ϵ should intuitively govern the size of the vol-of-vol. As we have
seen this should control the smile - ie. how expensive are the wings?
▶ ρ should intuitively govern the implied correlation between forward
rates and vol. However, works in combination with α, so in cases of
low α may need to be positive despite negative skew.
▶ σ0 is the initial vol. Controls ATM implied vol.
Can only calibrate to one expiry at a time (at least three options).
On the popularity of SABR
▶ The SABR model is able to generate correct skew dynamics, namely
a downward sloping backbone. This is in contrast to local vol
models. In addition parameter values are relatively stable.
▶ Furthermore, the model has good interpolation properties - ie. able
to value options far away from ATM since taking vanna and volga
risks into account.
▶ However, probably the main cause of the SABR popularity is that a
formula exists, which gives implied black vol directly as a ”simple”
function of parameters, ATM forward rate, expiry and strike.
fidSabr does this.

Figure 9: The SABR-to-BS formula.

What do we do about σ0 ?

For ATM options the SABR-to-BS formula collapses to

Figure 10: The SABR-to-BS formula.

▶ Thus, there is a one-to-one mapping between σ0 and σBS (ATM).

▶ This can be done via the fidBlackToSabr, which inverts the above
formula numerically.
SABR parameters and risk
SABR parameters: σ0
Figure 11: SABR-parameter: σ0

▶ In all example below: 3y5y swaptions, ATM rate is 3% and blue

lines are initial smiles calibrated using
(α0 , σ0 , ϵ0 , ρ0 ) = (10%, 0.61%, 18%, 49%).
▶ The σ0 parameter governs the overall level of implied vol given the
other parameters.
▶ In practical applications σ0 is updated intraday in order to hit ATM
implied vol.
▶ In the chart above σBS (ATM) = 14%/23% whereas
σ0 = 0.61%/1%
SABR parameters: α

Figure 12: SABR-parameter: α

▶ α is able to control both the overall level of implied level (lower

value results in higher ATM vol), ATM skew and backbone (the first
two properties follows from the local vol specification - see below).
▶ However, the ATM vol and skew is controlled by σ0 and ρ. Thus,
we leave α to control the backbone.
▶ In practical applications α is set to a fixed level, which generates as
good an empirical hedge as possible. α is not entering the
calibration.
SABR parameters: ϵ

Figure 13: SABR-parameter: ϵ

▶ Epsilon controls the wings. Higher value of ϵ leads to a more

pronounced smile since higher vol-of-vol.
▶ In practical applications ϵ is updated once per month.
▶ In terms of distributions ϵ generates fatter tails (excess kurtosis).
Due to the existance of jumps ϵ tends to decrease with expiry
(reason # 1).
SABR parameters: ρ
Figure 14: SABR-parameter: ρ

▶ As expected, the correlation parameter controls the skew. A more

negative ρ will generate a more negatively skewed smile.
▶ Note here that given the low value of α, ρ is actually positive
despite a negative skew in the initial calibration (+49%).
▶ Thus, ρ has the same effect on the smile as does α. However,
different combinations of (α, ρ) generates different hedge ratios.
▶ In practical applications ρ is updated once per month.
▶ As with ϵ, ρ also tends to zero with expiry.
Bumping ϵ and ρ all else equal
▶ If we bump ϵ upwards all else equal, this increases ATM implied vol,
and thus the entire smile shifts upwards. Why?
▶ Even though volga is zero ATM we cannot guarantee that an option
which is ATM today will remain ATM also tomorrow. Consider
buying an expiry T option today, then that option will most likely
move either ITM or OTM at some point before expiry thus gaining
volga, and thereby benefiting from higher vol-of-vol - ie. higher ϵ.
▶ Also note that for a given ϵ, increasing the time to expiry increases
the ATM implied vol. This is so since the probability of an ATM
option observing high volga at some point in time is proportional to
time to expiry. Thus, in order to calibrate to the market ϵ will
typically have to decrease with expiry (reason # 2).
▶ Bumping ρ higher or lower does not affect ATM implied vol. The
reason is that there is a roughly 50:50 probability that immediately
after trading an ATM option the option is either ITM or OTM and
only one of those scenarios is beneficial for the buyer of the option
(which one depends on the sign of the assumed correlation). Thus,
on average ATM options are independent of rho.
Backbone and skew
α controls the backbone
▶ Since α can be set freely it is in practice used to control the
backbone effect, and thereby risk measures (for an example the
delta).
▶ Remember that the backbone-effect is the predicted change in
at-the-money implied vol...

σBS (f , f ) − σBS (f0 , f0 ), f0 is initial forward rate

and σBS = f (x1 , x2 ), x1 = strike and x2 = ATM forward rate

Figure 15: Left: Backbone α = 0. Right: Backbone α = 1.

How α controls backbone
Leaving out the stochastic vol specification our diffusion for the forward is

dft = ftα σt dWt

=⇒ dft = ft σt ftα−1 dWt

Thus, we have a local vol function of the form σloc (f , t) = σt ftα−1 From
our local vol approximations we know that

∂σBS
K =f (0)
σBS ≈ σ0 f0α−1 and ≈ (α − 1)σ0 f0α−2
K =f (0) ∂f0
∂σBS
K =f (0) σBS (ATM)
=⇒ = (α − 1) <0 if α < 1
∂f0 f0
Thus, if ATM rates move higher the backbone effect implies a lower
implied vol for the new ATM vol when α < 1.
How α controls backbone continued
The effect on risk from a choice of α can be seen by noting that for an
option with strike K , changing the ATM forward rate

∂σBS (K , f0 )
σBS (K , f0 ) ≈ σBS (f0 , f0 ) + (K − f0 )
| {z } ∂K
ATM imp. vol
| {z }
Skew given ATM vol
∂σBS (K , f0 ) ∂σBS (f0 , f0 ) ∂σBS (K , f0 )
=⇒ ≈ −
∂f0 ∂f0 ∂K
| {z } | {z }
backbone skew at strike K

last line follows from differentiating with respect to the ATM rate f0 .
▶ So the change in implied vol for a given strike can be decomposed
into a backbone effect and a skew effect.
▶ If for an example α is close to zero and skew at strike K is close to
flat, then we should expect implied vol to drop at strike K, when
ATMF increases.
▶ If on the other hand the skew at strike K is is highly negative, then
we cannot a priori say if implied vol decreases or increases.
α effect on change in implied vol

▶ In case α < 1 both skew and backbone effect is in play. In case of a

positive bump to forward rates α < 1 implies a lower implied vol, for
a given skew.
▶ In the chart on the next page the blue line is the initial smile, while
the grey and red lines represent the smile following a +100bp bump
to forward rates.
▶ When α = 100% (grey line) the negative skew effect dominates the
backbone effect (there is none). Ie. options that become more away
from ATM (low strike options) see higher implied vol. Also note
that ATM implied vol (the ”bottom of the graph”) is unchanged
following the bump.
▶ When α = 10% (red line) the backbone effect dominates the skew
effect. Ie. options that become more away from ATM (low strike
options) see lower implied vol. Also note that ATM implied vol (the
”bottom of the graph”) falls following the bump.
α effect on change in implied vol: chart

Figure 16: Change in smile for different α. Blue line is initial smile. Red
and grey lines after +100bp bump.
SABR effect on delta
▶ When talking greeks in the SABR model one can either differentiate
directly wrt. to the SABR parameters or alternatively adjust the
greeks in the BS model. We will do the latter as is market practice.
▶ Implied vol in the BS model is now a function of the spot, meaning
that for delta we get

PVSABR = PVBS (f , K , T − t, σBS (K , f ))

∂PVSABR ∂PVBS ∂PVBS ∂σBS
=⇒ = + ·
∂f | ∂f {z } | ∂σ ∂f
{zBS } | {z }
BS delta BS vega Imp. vol
change

▶ Vega is everywhere positive, so the correction to BS delta depends

on the sign of ∂σ
∂f predicted by the SABR model.
BS

▶ And thereby on whether the skew dominates the backbone effect or

vice versa.
SABR effect on delta continued

Figure 17: Left: SABR vs BS delta. Right: Smile after underlying being
bumped more than 100bp.

DV01 for a 10y10y payer on 100m notional on flat curve of 3% with

SABR parameters: (σBS (ATM), α, ϵ, ρ) = (14%, 54%, 70%, 0%).
Note that the negative skew dominates backbone implying a higher
implied vol for all low strike options (< 3%) following positive bump to
forward rates.
Calibration and concluding remarks
How to calibrate the model?

▶ The model is over-specified since both α and ρ affects the smile in

similar ways. Specifically, they both imply a steepening or flattening
of the skew.
▶ A typical choice is to let α be fixed at some value and calibrate the
remaining parameters.
▶ At what value should α be fixed at?
▶ Since, as shown above, α has large implications for risk figures, it
can be fixed at a value generating the delta wanted, and then
calibrate the model.
▶ But what delta do you want? Difficult to answer, but probably the
one closest to the empirical one.
▶ Not the α which suits the dealer the best!
Calibration example in excel

Figure 18: SABR calibration example with different α

▶ Red line: (α, σ0 , ϵ, ρ) = (10%, 0.61%, 18%, 49%)

▶ Grey line: (α, σ0 , ϵ, ρ) = (100%, 14.05%, 23%, −24%)
▶ Note how σ0 and ρ changes a lot. Q: Surprised of σ0 -value in ”grey
line”?
SABR can’t answer all our prayers
▶ Not all models are able to calibrate to all smiles/skews.
▶ Opposite to local vol models, stochastic vol models cannot be
guaranteed to hit all implied volatilities for every expiry. This is
especially the case for short expiries, where the market can trade
with a highly pronounced smile to which the parameters of
stochastic volatility models is unable to calibrate. One reason for a
highly pronounced skew/smile for short expiry options is the
existence of jumps.
▶ Jumps are per definition a dis-continuity and renders everything
we’ve done so far wrong. In the cases of jumps we won’t be able to
re-adjust our delta hedges during the jump. In other words: We will
have unhedged risks on our books as the underlying moves
(unexplained PnL).
▶ Overall the existence of jumps should shift upwards all implied vols
at each expiry since jumps will increase average realized vol.
▶ However, it should lift wings more than ATM since jumps towards
the strike level of an option is valuable since the average gamma
experienced during the jump is higher than predicted by our
continuous time analysis.
SABR can’t answer all our prayers continued
▶ Jumps are the primary (only) explanation for the flattening of the
smile over expiries, since it can be shown that both volga vanna
vega and vega
are inversely proportional to T − t whereas var (σ) and cov (S, σ)
are proportional to T − t.
▶ Jumps mean that unrealistic parameter values need to be assumed
in order to be able to calibrate.
▶ In order to make sure that models are able to calibrate to the full
smile one can include local volatility into stochastic vol models.
▶ Furthermore, the SABR model can not be employed in order to
price more exotic types of options since the model can only calibrate
to one expiry at a time and is non-stationary (1y smile 1y forward
is flatter than 1y smile today).
▶ Stationarity can be important in the case of forward starting options
such as cliquets options (also known as ratchets).
▶ In order to price such options correctly one needs a stationary model
such as Heston, which predicts a forward smile in line with spot
smile observed today (forward starting 1y smile equal to 1y smile).
Where to now?

▶ We will not make it to digital options in the lectures. However,

remains in the syllabus. Read section 5.8 and check the lecture 9
slides.
▶ Do assignment 9, now that you have seen the calibration example
▶ This was my last lecture. I hope you have enjoyed it.
▶ Frederik will do the last two lectures on credit and exposure. A new
topic in this course, but in my view a need-to-know.
▶ Note that Frederik will move the ”exam lecture” on 12 December to
some time in the week leading up to your exam in January. I will
offer online sessions for those with questions, which cannot be
answered in writing, in the same week. Check Absalon.
Appendix
Appendix: Local vol w. time dependance

The simplest extension of BS.

▶ As we have seen above there is a term structure in volatility.
▶ Thus, a very simple and BS compliant way to enhance the
assumption of constant vol is to make vol dependent on time.

dS(t) = σ(t)S(t)dW (t)

A closed form solution still exists for european options in the BS model in
this framework since implied vol can be found to equal

s
Z T
1
σBS = σ 2 (u)du
T t
Appendix: Local vol w. time dependance continued

▶ A particular simple, though useful case, is where volatility is a

constant between future points in time, Ti = T0 , T1 , ..., Tn , with
T0 = t.
▶ In this case

Pj
2 i=1 σi2 · (Ti − Ti−1 )
σBS (Tj ) =
Tj

and forward vols from implieds

2 2
σBS (Tj ) · Tj − σBS (Tj−1 ) · Tj−1 = σj · (Tj − Tj−1 )
2 2
σBS (Tj ) · Tj − σBS (Tj−1 ) · Tj−1
=⇒ σj2 =
(Tj − Tj−1 )
Appendix: Other drawbacks of local vol

▶ Unless very simply specified (as the CEV case) the local vol models
a la Dupire requires an entire continuous surface of options prices in
both strike and expiry dimension.
▶ This makes the model somewhat difficult to calibrate, since first we
need to interpolate in both strike and expiry dimension (not as
simple as you might think - we need to keep the model arbitrage
free).
▶ Somewhat of a black box since non-parametric (Dupire calibration).
Lacking intuition as to what factors are relevant for implied vol.
▶ We need frequent re-calibration since, as we have seen, the local
volatility model implies implied vol moves which are different,
sometimes even opposite, of empirics. Thus, the local volatility
model quickly ”moves away” from markets.
▶ Instability of the model is bad for traders and hedge ratios as they
will first of all be wrong, and secondly change often.
Appendix: A simple local vol specification: CEV
A simple example of a local vol model is the CEV (contant elasticity of
volatility) model, which is specified as

dS(t) = λS(t)α dW (t)

dS(t) = S(t)λS(t)α−1 dW (t)

Thus, the CEV specification has a local vol function of the form
σ(S; λ, α) = λS(t)α−1 . Following our approximations above

σBS (K ) ≈ σ(S(t)) = λS(t)α−1

K =S(t)

=⇒ λ = σBS (ATM)S(t)1−α

In order to hit ATM vol. Further we have that this model will only be
able to generate a downward sloping skew. Flat if α = 1 (BS) and steep
if α = 0 (Bachelier/Normal). Distribution is chi-suared for α ∈ (0, 1). No
close-form solution exist (numerical implementation needed but can be
approximated).
Appendix: The CEV model in charts

Figure 19: Left: Model vs. market. Right: Implied skew for different α.
Appendix: A note on the difficulty of calibration and jumps
▶ Not all models are able to calibrate to all smiles/skews. Opposite to
local volatility models stochastic vol models cannot be guaranteed
to hit all implied volatilities for every expiry. This is especially the
case for short expiries, where the market can trade with a highly
pronounced smile to which the parameters of volatility models is
unable to calibrate.
▶ One reason for a highly pronounced skew/smile for short expiry
options is the existence of jumps.
▶ Jumps are per definition a dis-continuity and renders everything
we’ve done so far wrong. In the case of jumps we won’t be able to
re-adjust our delta hedges during the jump. In other words: We will
have unhedged risks on our books as the underlying moves
(unexplained PnL).
▶ Overall the existence of jumps should shift upwards all implied vols
at each expiry since jumps will increase average realized vol.
▶ However, it should lift wings more than ATM since jumps towards
the strike level of an option is valuable due to gamma being largest
close to the strike. As OTM or ITM options are more likely to
observe jumps towards them this can explain highly pronounced
Appendix: Vanna and volga intuition
Volga
▶ Volga is intuitively zero ATM. Why? Implied vol increases when the
market expects a higher average volatility to prevail over some time
period. That is a positive for all options. However, it also changes
the market implied probability distribution at the same time.
▶ When implied vol increases, for at the money options the probability
that the spot is gonna be fairly close to the current spot (ATM) for
at least some time is close to unchanged.
▶ However, for OTM/ITM options the market implied probability that
the spot will at some future point in time be close to the strike level
of these options increases. Thereby it increases the buyer of the
option’s ability to monetize moves in the underlying at the same
time as the market expects a higher average vol. Thus, owners of
OTM/ITM options are exactly exposed towards the high implied vol
when it increases and vice versa!
▶ For VERY OTM and ITM the market implied probability changes by
a very small amount.
Appendix: Vanna and volga intuition continued

Vanna
▶ Vanna is the change in delta given a change in implied vol. Vanna is
positive for high strike options (OTM payers and ITM receivers).
Why? When implied vol increases it increases the probability of
exercise, and thereby results in a more positive delta, for OTM
payers. However, it lowers the probability of exercise for ITM
receivers, and thereby results in a less negative delta.
▶ ATM vanna is close to zero since the probability of exercise doesn’t
change for ATM options, given the higher implied vol gives rise to
symmetric changes in the market implied density.
▶ Vanna is negative for low strike options (ITM payers and OTM
receivers). Why? When implied vol increases the probability of
exercise increases for OTM receivers (more negative delta), while the
probability of exercise decreases for ITM payers (less positive delta).
Skew and backbone decomposed in a drawing

In this example we start out with

an ATM level ATM0 and move the
ATMF rate lower so that the new
ATMF rate is ATM1 . If we for a start
ignore any backbone (assume α = 1)
then we get the first picture. Note
that implied vol is the same for the
two different ATM levels. The new
implied vol for the old ATMF level
ATM0 is now lower at point B (neg-
ative skew).
In the second part we include a back-
bone effect such that the entire smile
moves higher, taking the old ATM
implied vol to the (final) point C.