Realized and implied index skews, jumps, and

the failure of the minimum-variance hedging

Artur Sepp
Global Risk Analytics
Bank of America Merrill Lynch, London
artur.sepp@baml.com

Global Derivatives Trading & Risk Management 2014

Amsterdam
May 13-15, 2014

1
Plan
1) Empirical evidence for the log-normality of implied and realized volatil-
ities of stock indices

2) Apply the beta stochastic volatility (SV) model for quantifying implied
and realized index skews

3) Origin of the premium for risk-neutral skews and its impacts on profit-
and-loss (P&L) of delta-hedging strategies

4) Optimal delta-hedging strategies to improve Sharpe ratios

5) Log-normal beta SV model

2
References
Technical details can be found in references
Beta stochastic volatility model:
Karasinski, P., Sepp, A., (2012), “Beta stochastic volatility model,” Risk,
October, 67-73
http://ssrn.com/abstract=2150614
Sepp, A. (2013), “Consistently Modeling Joint Dynamics of Volatility and
Underlying To Enable Effective Hedging”, Global Derivatives conference
in Amsterdam 2013
http://math.ut.ee/~spartak/papers/PresentationGlobalDerivatives2013.pdf
Implied and realized skews, jumps, delta-hedging P&L:
Sepp, A., (2014), “Empirical Calibration and Minimum-Variance Delta
Under Log-Normal Stochastic Volatility Dynamics”
http://ssrn.com/abstract=2387845
Sepp, A., (2014), “Log-Normal Stochastic Volatility Model: Pricing of
Vanilla Options and Econometric Estimation”
http://ssrn.com/abstract=2522425
Optimal delta-hedging strategies:
Sepp, A., (2013), “When You Hedge Discretely: Optimization of Sharpe
Ratio for Delta-Hedging Strategy under Discrete Hedging and Transaction
Costs,” Journal of Investment Strategies 3(1), 19-59
http://ssrn.com/abstract=1865998 3
How to build a dynamic model for volatility?
Suppose we know nothing about stochastic volatility

We want to learn only by looking at empirical data

How do we start?

4
Empirical frequency of implied vol is log-normal
First, check whether stationary distribution of volatility is:
A) Normal or B) Log-normal

Compute the empirical frequency of one-month implied at-the-money

(ATM) volatility proxied by the VIX index for last 20 years

Daily observations normalized to have zero mean and unit variance

Left figure: empirical frequency of the VIX - it is definitely not normal

Right figure: the frequency of the logarithm of the VIX - it does look
like the normal density (especially for the right tail)!
7% Empirical frequency of 7% Empirical frequency of
6% normalized VIX 6% normalized logarithm of the VIX
5%
Frequency

Empirical 5% Empirical

Frequency
4% Standard Normal 4% Standard Normal
3% 3%
2% 2%
1% 1%
VIX Log-VIX
0% 0%
-4 -3 -2 -1 0 1 2 3 4 -4 -3 -2 -1 0 1 2 3 54
Empirical frequency of realized vol is log-normal
Compute one-month realized volatility of daily returns on the S&P 500
index for each month over non-overlapping periods for last 60 years from
1954

Below is the empirical frequency of normalized historical volatility

Left figure: frequency of realized vol - it is definitely not normal

Right figure: frequency of the logarithm of realized vol - again it does

look like the normal density (especially for the right tail)
10% Frequency of Historic 1m 10% Frequency of Logarithm of
8% Volatility of S&P500 returns 8% Historic 1m Volatility of S&P500

Frequency
Empirical
Frequency

6% Empirical 6%
Standard Normal Standard Normal
4% 4%

2% 2%
Vol Log-Vol
0% 0%
-4 -3 -2 -1 0 1 2 3 4 -4 -3 -2 -1 0 1 2 3 4

6
Dynamic model for volatility evolution should
not be based on price-volatility correlation
Now we look for a dynamic factor model for volatility (next slide)

We cannot apply model based on correlation between S&P500 returns

and changes in volatility because using correlation we can only predict
the direction of change, not the magnitude of change

For risk management of options, we need a factor model for volatility

dynamics

7
Factor model for volatility uses regression model for
changes in vol V (tn) predicted by returns in price S(tn)
" #
S(tn) − S(tn−1)
V (tn) − V (tn−1) = β + V (tn−1)n (1)
S(tn−1)
iid normal residuals n are scaled by vol V (tn−1) due to log-normality
Volatility beta β explains about 70% of variations in volatility!
Left figure: scatter plot of daily changes in the VIX vs returns on S&P
500 for past 14 years and estimated regression model
Right: time series of empirical residuals n of regression model (1)
Residual volatility does not exhibit any systemic patterns
Regression model is stable across different estimation periods
20% 30% Time Series of Residual Volatility
Change in VIX

Change in VIX vs Return on S&P500

15% 20%
10% y = -1.08x 10%
R² = 67%
5% 0%
0% -10%
-10% -5% -5% 0% 5% 10% -20%
-10% -30%
Dec-99
Dec-00
Jan-02
Jan-03
Jan-04
Jan-05
Jan-06
Jan-07
Jan-08
Jan-09
Jan-10
Jan-11
Jan-12
Jan-13
-15%
-20% Return % on S&P 500

Volatility beta β: expected change in ATM vol predicted by price return

For return of −1%: expected change in vol = −1.08 × (−1%) = 1.08% 8
More evidence on log-normal dynamics of vol: indepen-
dence of regression parameters on level of ATM vol
Estimate empirically the elasticity α of volatility by:
1) computing volatility beta and residual vol-of-vol for each month using
daily returns within this month
2) test if the logarithm of these variables depends on the log of the VIX
in that month using regression model
 
α
Left figure: test β̂(V ) = βV by regression model: ln β̂(V ) = α ln V + c
 

Right: test ε̂(V ) = εV 1+α by regression model: ln |ε̂(V )| = (1+α) ln V +c

The estimated value of elasticity α is small and statistically insignificant
Indeed the realized volatility is close to log-normal
1.5 ln(VIX residualvol) vs ln(AverageVIX) 0.5
ln(VIX beta) vs ln(Average VIX)
y = 0.15x + 0.14 1.0 y = 0.14x - 0.45
0.0

ln(|VIX residual vol)

R² = 2% R² = 4%
ln(|VIX beta|)

0.5
0.0 -0.5
-0.5
-1.0
-1.0
ln(Average VIX) -1.5 ln(Average VIX)-1.5
-2.5 -2.0 -1.5 -1.0 -0.5 0.0 -2.5 -2.0 -1.5 -1.0 -0.5 0.09
Empirical estimation of volatility elasticity α:
volatility dynamics is log-normal
(maximum likelihood estimation - see my paper on log-normal volatility)
Figure: 95% confidence bounds for estimated value of elasticity α using
realized (RV) and implied (IV) volatilities for 4 major stock indices
1.0 95% confidence bounds for
0.5 estimated elasticity alpha

Alpha
0.0
-0.5
-1.0

IV, FTSE100

RV, FTSE100
VSTOXX, Reg
VIX, ML
VSTOXX, ML

IV, NIKKEI
IV, STOXX50
IV, S&P500

RV, NIKKEI
RV, STOXX50
VIX, Reg

RV, S&P500
Estimation results confirm evidence for log-normality of volatility:
[i] In majority of cases (7 out of 12), bounds for α̂ contain zero
[ii] One outlier α̂ = −0.4 (realized volatility of Nikkei index)
[iii] Remaining are symmetric: two with α̂ ≈ 0.2 and two with α̂ ≈ −0.2
To conclude - alternative SV models are safely rejected:
1) Heston and Stein-Stein SV models with α = −1
2) 3/2 SV model with α = 1
Also, excellent econometric study by Christoffersen-Jacobs-Mimouni (2010),
Review of Financial Studies: log-normal SV outperforms its alternatives
10
Beta stochastic volatility model (Karasinski-Sepp 2012):
is obtained by summarizing our empirical findings for
dynamics of index price S(t) and volatility V (t):
dS(t) = V (t)S(t)dW (0)(t)
dS(t) (2)
dV (t) = β + εV (t)dW (1)(t) + κ(θ − V (t))dt
S(t)
V (t) is either returns vol or short-term ATM implied vol
W (0)(t) and W (1)(t) are independent Brownian motions
β is volatility beta - sensitivity of volatility to changes in price
ε is residual vol-of-vol - standard deviation of residual changes in vol

Mean-reversion rate κ and mean θ are added for stationarity of volatility

A closer inspection shows that these dynamics are similar to other log-
normal based SV models widely used in industry:
A) in interest rates - SABR model
B) in equities - a version of log-normal based aka exp-OU SV models

We arrived to beta SV model (2) only by looking at empirical data

for realized&implied vols and using factor model for vol dynamics
11
Implied interpretation of volatility beta and residual vol-
of-vol from Black-Scholes-Merton (BSM) volatilities, σBSM (z) as func-
tions of log-strike z = ln(K/S), inferred form option prices
Compute vol skew SKEW and convexity CONV for small maturities:
SKEW = [σBSM (5%) − σBSM (−5%)] / (2 × 5%)
2


CONV = [σBSM (5%) + σBSM (−5%) − 2σBSM (0)] / 5%

Volatility beta β [I] implied by skew:

β [I] = 2 × SKEW

Residual vol-of-vol ε[I] implied by convexity:

q
[I]
ε = 3 × σBSM (0) × CONV + 2 × (SKEW)2

As model parameters, volatility beta (left figure) and idiosyncratic vol-

of-vol (right figure) have orthogonal impact on BSM implied vols
35% Impact of volatility beta on BSM vol 35% Impact of residual vol-vol on BSM vol
Base vols with beta = -1 Base vols with ResidVol=1.0
Down vols with beta = -0.5 Down vols with ResidVol=0.5
25% 25%
Up vols with beta = -1.5 Up vols with ResidVol=1.5
BSM implied vols

BSM implied vols

15% 15%

Strike Strike
5% 5%
0.7 0.78 0.86 0.94 1.02 1.1 0.70 0.78 0.86 0.94 1.02 1.10 12
Topic II: Implied and realized skew using beta SV model
Use time series from April 2007 to December 2013 for one-month ATM
vols and the S&P500 index with estimation window of one month
Figure 1): Implied and realized one
80% 1m ATM Implied Volatility
month volatilities 70%
1m Realized Volatility
60%
ATM volatility tends to trade at 50%
40%
a small premium to realized 30%
20%
10%
Figure 2): One-month average of 0%

Oct-11 Oct-11
Feb-07
Sep-07
Apr-08
Nov-08 Nov-08
Jun-09 Jun-09
Jan-10 Jan-10
Aug-10
Mar-11 Mar-11

May-12 May-12
Dec-12 Dec-12
Jul-13
implied and realized volatility beta
Implied volatility beta consis-

Aug-10
Sep-07
Apr-08
Feb-07

Jul-13
tently over-estimates realized one
0.0 Implied Volatility Beta
Realized Volatility Beta
Figure 3): Average of implied and -0.5

realized residual vol-of-vol -1.0

Implied residual vol-of-vol signifi- -1.5

cantly over-estimates realized -2.0

1.7 Implied Residual Vol-of-Vol

Absolute (Abs) and relative (Rel) 1.4
Realized Residual Vol-of-Vol

spreads between implieds&realizeds 1.1

0.8
Spreads Vol Beta VolVol 0.5
Abs, Mean 0.51% -0.27 0.78 0.2

Oct-11
Feb-07
Sep-07
Apr-08
Nov-08
Jun-09
Jan-10
Aug-10
Mar-11

May-12
Dec-12
Jul-13
Abs, Stdev 6.2% 0.21 0.16
Rel, Mean 7% 21% 57%
Rel, Stdev 24% 17% 11% 13
Explanation of the skew premium in a quantitative way
In a very interesting study, Bakshi-Kapadia-Madan (2003), Review of
Financial Studies, find relationship between risk-neutral and physical skew
using investor’s risk-aversion
Fat tails (not necessarily skewed) of returns distribution under phys-
ical measure P along with risk-aversion lead to increased negative
skeweness under the risk neutral-measure Q
Quantitatively:
SKEWENESSQ = SKEWENESSP − γ × KURTOSISP × VOLATILITYP

SKEWENESSQ is risk-neutral skeweness of price returns

SKEWENESSP is physical skeweness of price returns
KURTOSISP is kurtosis as measure of fat tails of physical distribution
VOLATILITYP is volatility of returns under physical distribution
γ > 0 is risk-aversion parameter of investors

To conclude: the risk-neutral premium arises because risk-averse

investors assign higher value to insurance puts
Important: Volatility skew is proportional to skeweness of returns
14
 Apply Merton Jump-Diffusion (JD) with normal jumps
Figure 1: Use last 14 years of daily returns on S&P 500 index to estimate
skeweness and kurtosis of returns - see column ”Empirical P”
Table 1: Use γ = 22.0 (estimated from time series of implied vols by
inverting BKM formula) and apply BKM to obtain SKEWENESSQ = −2
Figure&Table 2: Fit Merton JD to first four moments of physical and
risk-neutral distribution (jump frequency is set to one jump per month)
From calibration: JumpMean is 0 under empirical P and -5% under Q
Merton JD params P Q
Empirical P Q
Stdev 21% 21% Jump Mean 0% -5%
Jump Volatility 4% 0%
Skeweness 0 -2
Diffusion vol 17% 13%
Kurtosis 8 8
Jump Frequency 12 12

Figure 3: Value one month options - implied volatility from Merton JD

under Q is skewed, while implied volatility under P is symmetric
10% Frequency of S&P500 daily returns 10% Frequency of S&P500 daily returns Implied volatility skew for one
Empirical Empirical 30% month options on S&P500
8% 8%
Frequency Frequency
Frequency

Physical Merton under P

Frequency

6% Normal Density 6% Physical Merton

25% Risk-Neutral Merton under Q
under P

Implied Vol
4% 4% Risk-Neutral
Merton under Q 20%
2% 2%
Daily return Daily return Strike
0% 0% 15%
-9% -7% -5% -4% -2% 0% 2% 4% 6% 7% -9% -7% -5% -4% -2% 0% 2% 4% 6% 7% 0.75 0.85 0.95 1.05 1.15 1.25
15
To summarize our developments so far:
1) Log-normal beta SV model is consistent with empirical distribu-
tion for realized and implied vols

2) Beta SV model is applied to quantify realized and implied skews

and the spread between them, which turns out to be significant

Any option position is mark-to-market so no point of arguing about

market prices
However, hedging strategy is discretionary and can be the ”edge”

By computing the delta-hedge: should we use implied or realized

skews?

This question is analyzed in the third topic of my talk:

Part I - Quantitative analysis of impact of realized and implied skews on

delta-hedging P&L

Part II - Monte-Carlo simulations for empirical analysis

16
Statistically significant spread between realized and im-
plied skews β [R] − β [I] leads to dependence on realized
price returns and invalidates the minimum-variance hedge
Minimum-variance delta ∆ is applied to hedge against changes in price
and price-induced changes in volatility
Given hedging portfolio Π for option U on S
Π(t, S, V ) = U (t, S, V ) − ∆ × S
∆ is computed by minimizing variance of Π using SV beta dynamics (2)
under risk-neutral measure Q (classic approach) with implied vol betaβ [I]
∆ = US + β [I] × UV /S
where US and UV model delta and vega
To see dependence on return δS due to spread between implied vol beta
β [I] and realized β [R]: given δS apply beta SV for change in vol δV under
physical measure P: [R] [R]
√
δV = β × δS + ε × V δt
By Taylor expansion of realized P&L:
h
[R] [I]
i
[R]
√
δΠ(t, S, V ) = β −β × UV × δS + ε × UV × V δt + O(dt)
ε(R) is random non-hedgable part from residual vol-of-vol
O(dt) part includes quadratic terms (δS)2, (δV )2, (δS)(δV )
17
Volatility skew-beta is important for computing correct
option delta
Figure 1) Apply regression modelIt is nearly maturity-homogeneous
(1) for time series of ATM vols for 0.0 Regression Volatility Beta(T)
maturities T = {1m, 3m, 6m, 12m, 24m} -0.2 y = 0.19*ln(x) - 0.37
(m=month) to estimate regression -0.4 R² = 99%

volatility beta βREGRES (T ) using -0.6

S&P500 returns: -0.8
Regression Vol Beta(T)
Decay of Vol Beta in ln(T)
δσAT M (T ) = βREGRES (T ) × δS -1.0 Maturity T
0.08 0.25 0.50 1.00 2

Volatility beta for SV dynamics is in- 0.0 Implied Volatility Skew (T)
stantaneous beta for very small T -0.2 y = 0.16*ln(x) - 0.29
R² = 99%
Regression vol beta decays in log-T -0.4

due to mean-reversion: long-dated -0.6 Vol Skew (T)

Decay of Skew in ln(T)
ATM vols are less sensitive in abso- -0.8 Maturity, T
lute values to price-returns 1m 3m 6m 1y 2y

2.0 Volatility Skew-Beta(T)

Figure 2) Implied vol skew for ma-
1.5
turity T has similar decay in log-T y = 0.06*ln(x) + 1.41
1.0 R² = 80%
Figure 3) Volatility skew-beta is Vol Skew-Beta (T)
0.5
regression beta divided by skew Decay of Vol Skew-Beta in ln(T)

Skew-Beta(T ) ∝ βREGRES (T )/SKEW(T ) 0.0 Maturity T

1m 3m 6m 1y 2y
18
Technical supplement to compute model implied skew-
beta (omitted during the talk)
Using backward pricers and PDE:
1) Compute the term structure of ATM volatility σAT M (S0; T ) and skew
SKEW(S0; T ), with strike width α%, implied by model parameters
2) Bump the spot price down by α%, S1 = (1 − α%)S0, and apply corre-
sponding bumping rule for model state variables
For the beta SV: β
V1 → V0 + βα , θ → θ + α (3)
2κ
3) Compute new term structure of ATM vols σAT M (S1; T )
4) Compute model implied skew-beta
σAT M (S1; T ) − σAT M (S0; T )
Skew-Beta(T ) = − (4)
α × SKEW(S0; T )
Using Monte-Carlo pricers:
1) Specify number of paths and simulate set of independent Brownians
2) Compute paths starting from {S0, V0}
2A) Evaluate term structure of ATM volatility σAT M (K = S0; T ) and
skew using σ(K = S1; T ), both using Brownians in 1)
3) Compute paths starting from {S1, V1} with S1 = (1 − α%)S0 and V (1)
bumped as in Eq (3), using Brownians in 1)
4) Evaluate ATM vols σAT M (K = S1; T ) and skew-beta by Eq (4)
19
Volatility and Skew contribution to P&L - important for
volatility positions with daily mark-to-market!
Mark BSM implied vol σBSM (K) in %-strike K relative to price S(0):
σBSM (K; S) = σAT M (S) + SKEW × Z(K; S)
Z(K; S) is log-moneyness relative to current price S:
Z(K; S) = ln (K × S(0)/S )
SKEW < 0 is inferred from spread between call and put implied vols

In practice, this form is augmented with extras for convexity and tails
Any SV model implies quadratic form for implied vols near ATM strikes
(Lewis 2000, Bergomi-Guyon 2012) so my approach for vol P&L is generic

Volatility P&L arises from change in spot price S → S {1 + δS}:

δσBSM (K; S) ≡ σBSM (K; S {1 + δS}) − σBSM (K; S)
= δσAT M (S) + SKEW × δZ(K; S)
First contributor to P&L: change in ATM vol δσAT M (S):
δσAT M (S) = σAT M (S {1 + δS}) − σAT M (S)
Second contributor to P&L: change in log-moneyness relative to skew:
δZ(K; S) = − ln(1 + δS) ≈ −δS
20
Example of volatility and skew P&L with regression beta
(omitted during the talk)
σAT M (S(0)) = 15%, δS = −1.0%, SKEW = −0.5, βREGRESS = −1.0

It is very important how we keep log-moneyness Z(K; S):

1) For strikes re-based to new ATM level (forward-based strikes):
S → S{1 + δS} and log-moneyness does not change δZ(K; S) = 0
P&L arises from change in ATM vol predicted by price return com-
puted using βREGRESS :
δσBSM (K) = βREGRESS × δS = −1.0 × −1% = 1%
2) For strikes fixed at old ATM level (vanilla strikes with fixed S(0))
Thus log-moneyness changes by δZ(K; S) ≈ −δS = 1%
P&L is change in ATM vol adjusted for change in money-ness:
δσBSM (K) = βREGRESS ×δS +SKEW×δS = 1%+(−0.5)×(1%) = 0.50%
Change in vols, strikes fixed to ATM 0 BSM vol 0, strikes fixed to ATM 0
Change in vols, strikes re-based to ATM 1 BSM vol 1, strikes fixed to ATM 0
1.0% 21% BSM vol 1, strikes re-based to ATM 1

0.8%

BSM vol(K)
18%
0.5%
15%
0.3%
0.0% Strike K% 12% Strike K%
90% 95% 100% 105% 90% 95% 100% 105%
21
Changes in skew are not correlated to changes in price
and ATM vols - important for correct predict of vol and skew P&L
Empirical observations yet again confirm log-normality dynamics!
(Using S&P500 data from January 2007 to December 2013)
Figure 1: weekly changes in 100% − 95% skew vs price returns for
maturity of one month (left) and one year (right)
Regression slope = 0.13 (1m) & 0.03 (1y); R2 = 0% (1m) & 1% (1y)
Change in 0.3
1m skew vs Price Return 0.04
Change in 1y skew vs Price Return
0.2 0.02
Change in Skew

Change in Skew
0.1
0
0
-0.02
-0.1 y = 0.03x - 0.00
y = 0.13x - 0.00 -0.04 R² = 1%
-0.2 R² = 0%
Price return -0.3 Price return -0.06
-15% -5% 5% 15% -15% -5% 5% 15%

Figure 2: weekly changes in 100% − 95% skew vs changes in ATM

vols for maturity of one month (left) and one year (right)
Regression slope = −0.15 (1m) & −0.06 (1y); R2 = 0% (1m) & 2% (1y)
Change in0.3
1m skew vs 1m ATM vol Change in0.04
1y skew vs 1y ATM vol
0.2 0.02
Change in Skew

Change in Skew
0.1
0
0
-0.02
-0.1
y = -0.09x - 0.00
-0.2 y = -0.15x - 0.00 -0.04
R² = 2%
Change in ATM vol
-0.3 R² = 0%
Change in ATM-0.06
vol
-15% -5% 5% 15% -15% -5% 5% 22
15%
Volatility skew-beta combines the skew and volatility
P&L together
Given price return δS:
S → S {1 + δS}
Volatility P&L is computed by:
1) For strikes re-based to new ATM level
Log-moneyness does not change, δZ(K; S) = 0
P&L follows change in ATM vol predicted by regression beta and vol
skew-beta:
δσBSM (K) ≡ δσAT M (S) = βREGRESS × δS
= SKEWBETA × SKEW × δS

2) For strikes fixed at old ATM level

Log-moneyness changes by δZ(K; S) ≈ −δS
P&L is change in ATM vol adjusted for skew P&L:
δσBSM (K) ≡ δσAT M (S) − SKEW × δS
= [SKEWBETA − 1] × SKEW × δS
Positive change in ATM vol from negative return is reduced by
skew
23
Volatility skew-beta under minimum-variance approach
is applied to compute min-var delta ∆ for hedging against
changes in price and price-induced changes in implied vol
A) We adjust option delta for change in implied vol at fixed strikes
B) The adjustment is proportional to option vega at this strike:
∆(K, T ) = ∆BSM (K, T ) + [SKEWBETA(T ) − 1] × SKEW(T ) × VBSM (K, T )/
∆BSM (K, T ) is BSM delta for strike K and maturity T
VBSM (K, T ) is BSM vega, both evaluated at volatility skew

I classify volatility regimes using vol skew-beta for delta-adjustments:




 ∆BSM (K, T ) + SKEW(T ) × VBSM (K, T )/S, Sticky local

∆
BSM (K, T ), Sticky strike

∆(K, T ) =


 ∆BSM (K, T ) − SKEW(T ) × VBSM (K, T )/S, Sticky delta
1 SKEW(T ) × V

∆ (K, T ) + BSM (K, T )/S, Empirical S&P50

BSM 2

”Shadow” delta is obtained using ratio O (may be different from 1/2):

∆(K, T ) = ∆BSM (K, T ) + O × SKEW(T ) × VBSM (K, T )/S
which is traders’ ad-hoc adjustment of option delta
24
 Volatility skew-beta and vol regimes (also see Bergomi 2009):

2, Sticky local regime: minimum-variance delta in SV and LV


SkewBeta = 1, Sticky strike regime: BSM delta evaluated at implied skew



0, Sticky delta regime: model delta in space-homogeneous SV
Empirical estimates for skew-beta and its lower and upper bounds are
found by regression model (see my paper)

In beta SV model, with empirical estimate of vol beta and adding jumps/risk-
aversion to match skew premium, we fit empirical vol skew-beta:
1) S&P 500: empirical skew-beta of about 1.5
2) STOXX 50: strong skew-beta close to 2
3) NIKKEI: weak skew-beta is about 0.5

As result: beta SV model with jumps can produce the correct delta!
2.50 2.50 2.50
Vol Skew-Beta for S&P500 Vol Skew-Beta for STOXX 50 Vol Skew-Beta for NIKKEI
2.00 2.00 2.00
SVJ Skew-Beta with empirical beta
1.50 1.50 1.50 Sticky local with Min-var delta
1.00 SVJ Skew-Beta with empirical beta 1.00 Empirical bounds
SVJ Skew-Beta with empirical beta 1.00
0.50 Sticky local with Min-var delta 0.50 Sticky local with Min-var delta 0.50
Empirical bounds Empirical bounds
0.00 T in months 0.00 T in months 0.00 T in months
1m
3m
5m
7m
9m
11m
13m
15m
17m
19m
21m
23m

1m
3m
5m
7m
9m
11m
13m
15m
17m
19m
21m
23m

1m
3m
5m
7m
9m
11m
13m
15m
17m
19m
21m
23m
25
Second part of topic III: Monte Carlo analysis of delta-
hedging P&L

Now let’s have some fun and do some number crunching!

We are going to simulate the market dynamics and compare hedging

performance under different specifications of delta

In next few slides I briefly discuss the methodology

Details are provided for the interested for self-studying

Details are important to understand how to improve the performance of

delta-hedging strategies
Application to actual market data produces equivalent conclusions

In my talk, I will only discuss final results and conclusions

26
Apply beta SV for dynamics under physical measure P:
1) Index price S(t),
2) Volatility of returns Vret(t):
3) Short-term implied volatility Vimp(t):

dS(t) = Vret(t)S(t)dW (0)(t)

 
dVret(t) = κ[P ] θ [P ] − Vret(t) dt + β [P ]Vret(t)dW (0)(t) + ε[P ]Vret(t)dW (1)(t)
 
dVimp(t) = κ [I] θ − Vimp(t) dt + β [I]Vimp(t)dW (0)(t) + ε[I]Vret(t)dW (1)(t)
[I]

4) At-the-money (ATM) implied vol Vatm(t) is obtained by computing

model implied ATM vol for maturity T using model dynamics for Vimp(t)

Important: Model parameters are estimated from time series by

maximum likelihood methods - as a rule, parameters for returns vol
[P ] and for implied vol [I] are different
Here, apply the same parameters for clarity
Physical for Returns dVret (t), [P ] Vol dVimp (t), [I]
V. (0) 16% 16.75%
θ[.] 16% 16.75%
κ[.] 3.0 3.0
ε[.] 0.5 0.5
β [.] -1.0 -1.0
27
Volatility and skew premiums are produced using BSM implied
volatility, σBSM (K), as function of % strike K relative to S(0):
σBSM (K) = Vatm(t) + SKEW × ln (K × S(0)/S(t)) (5)
SKEW = −0.5 is vol implied skew specified exogenously by
strike % BSM vol σBSM (K) σBSM (K) − Vret (0)
99% 17.25% 1.25%
100% 16.75% 0.75%
101% 16.25% 0.25%
Market Skew -0.50
Important - option delta is computed using two models:
1) Beta SV model with market implied beta β [I] = -1.1
2) Beta SV model with empirical beta β [I] = -1.0 and jumps (risk-
aversion) to price-in excessive skew −1.1 − 1.0 = −0.1 (discussed later)
Both SV models fit to market skew exactly!
[i] Premium of implied vol to realized vol is:
16.75% − 16% = 0.75% (in line with empirical spread)
[ii] Premium of implied and empirical beta is:
β [I] − β [R] = -1.1 − ( -1.0 ) = -0.1 (empirical is about −0.2)
As we saw using Madan-Merton fits, physical dynamics don’t need to
have asymmetric jumps to produce skew premium - now, skew premium
arises from excess kurtosis produced by empirical SV model for returns
28
Consistency with market skew does not guar-
antee fit to empirical dynamics
Both hedging models are consistent with market implied skew
However, we observe discrepancy:
SV model with market implied beta,called Minimum variance hedge
Implies vol skew-beta about 2.0 , which is inconsistent with empirical
dynamics
SV model with jumps and empirical beta, called Empirical hedge:
Implies vol skew-beta about 1.6 , which is consistent with empirical
dynamics

Important - no re-calibration along a MC path is applied:

Both hedging models are initially consistent with the market skew - as
price S(t) and vol Vimp(t) change, both models remain very close to
market skew
Log-normality assumption - independence of implied&realized skew
from volatility - comes into play
29
Specification for trading in delta-hedged positions:
1) Straddle - short ATM put and call
Figure 1: P&L profile with Delta= 0 is function of realized return squared

Important: P&L/delta of straddle are not sensitive to realized/implied

skew - Benefits from small realized variance of price returns

2) Risk-reversal - short put with strike 99% and long call with strike
101% of forward
Figure 2: P&L profile with Delta= −0.8 is function of realized return

Important: P&L/delta of risk-reversal are very sensitive to real-

ized/implied skew - Benefits from small realized covariance of changes
in price and ATM vol
PayOff+PV-DeltaHedge with Delta=0 PayOff+PV-DeltaHedge with delta=-0.8
5.0% 10.0%
PayOff 7.5% PayOff
2.5%
5.0%
0.0% 2.5%
-2.5% 0.0%
-5.0% -2.5%
-5.0%
-7.5% -7.5%
-10.0% Straddle P&L vs Price return -10.0% Risk-Reversal P&L vsPrice return
-10%-8%-6%-4%-2% 0% 2% 4% 6% 8%10% -10%-8%-6%-4%-2% 0% 2% 4% 6% 8%10%

30
Specification for notionals of delta-hedged positions

Notionals are normalized by CashGamma=(1/2) × (S 2)×OptionGamma

Notionals for straddle:

0.5
P utN otional(tn) = CallN otional(tn) = −
ATM CashGamma(tn)
Notionals for risk-reversal:
0.5 × (Vatm(tn))2T
P utN otional(tn) = −
2% × {Put Vega(tn)}
0.5 × (Vatm(tn))2T
CallN otional(tn) = +
2% × {Call Vega(tn)}
where 2% comes from strike width 2% = 101% − 99%

Important: for Straddle, cash-gamma is 1.0

For Risk-reversal, the vanna (vega of delta) is 1.0

31
Monte-Carlo analysis: P&L accrual
Daily re-balancing at times tn, n = 1, ..., N
At the end of each day, we roll into new position so straddle is at-the-
money and risk-reversal has the same strike width

Realized P&L is P&L on hedges minus P&L on options position:

N
X  
P&L = {∆(tn−1) S(tn) − S(tn−1)
n=1


− Π (T − dt, S(tn), Vatm(tn)) − Π T, S(tn−1), Vatm(tn−1) }
Π (T, S(tn), Vatm(tn)) is options position computed using BSM formula and
implied volatility skew (5) with Vatm(tn), T = 1/12, dt = 1/252
Transaction costs are 2bp (k = 0.0002) per delta-rebalancing:
N
X
TC = k |∆(t0)| S(t0) + k |∆(tn) − ∆(tn−1)| S(tn)
n=1
where ∆(tn) is combined delta for newly rolled position

Important: P&L across different days and paths is maturity-time

and strike-space homogeneous - robust for statistical inference!
32
Monte-Carlo analysis - final notes
Trade notional is 100,000,000$
Realized P&L and explanatory variables are reported in thousands of $

Option maturity: one month

Daily re-hedging with total for each path: N = 21
P&L is annualized by multiplying by 12

Draw 2,000 paths and compute realized P&L and price return, variance,
volatility beta for changes in price and ATM vol, etc

Price and volatility paths are the same for straddle and risk-reversal
and different hedging strategies

A) Analyze realized delta-hedging P&L (Profit and Loss) by

[i] Realized P&L and its volatility, transaction costs
[ii] Sharpe ratios

B) P&L Explain using regression model with explanatory variables

What factors (realized variance, covariance, etc) contribute to P&L
33
1. Analysis of realized P&L for straddle
Figure left - realized P&L with no accounting for transaction costs
Right - realized P&L with transaction costs

Approximately, straddle P&L is spread between implied&realized vols2:

n o
2
P&L = Γ × (Vatm) − (Vret) 2
n o
2 2
= 100, 000 × (16.75%) − (16.00%) = 246
where Γ is cash-gamma notional in thousands $
Realized P&L little depends on the delta hedging strategy
Important is that asset drift is zero, otherwise P&L-s for different hedging
strategies have directional exposure to realized asset drift
300 Straddle P&L, zero trans costs 300 Straddle P&L after trans costs

200 200

244 243
100 100
161 161

0 0
Minimum var Empirical beta Minimum var Empirical beta
34
2. Analysis of realized P&L for risk-reversal

Figure: left - realized P&L with no accounting for transaction costs

Right - realized P&L with transaction costs

Approximately, risk-reversal P&L is spread between implied and realized

co-variance of price and vol returns:
n h i o
2 2
P&L = V × −SKEW × (Vatm) + (Vret) + β [R] × (Vret) 2
n h i o
2 2 2
= 100, 211 × 0.5 × (16.75%) + (16.00%) − 0.88 × (16.00%) = 431
where V is vanna notional in thousands $

Again, realized P&L little depends on the delta hedging strategy when
asset drift is zero
500 Risk-Reversal P&L, zero trans costs 500 Risk-Reversal P&L after trans costs
400 400
300 300
200 423 423 200
100 100 190 192
0 0
Minimum var Empirical beta Minimum var Empirical beta
35
3. Analysis of transaction costs

Transaction costs are 2bp per traded delta notional or 1$ per 5, 000$

Left figure: realized transaction costs

1) Risk-reversal has higher transaction costs due to larger delta notional
2) Minimum variance hedge and empirical hedge imply about equal trans-
action costs for straddle
3) Minimum variance hedge implies higher transaction costs for
risk-reversal because of over-hedging the put side

Right figure: volatility of transaction costs

Volatility is about uniform and very small compared to mean costs
300 Realized Transaction costs Volatility of Transaction costs
6
200
4
233 231 5 5
100 2
83 82 2 2
0 0
Min var for Empirical Min var for Empirical Min var for Empirical Min var for Empirical
straddle beta for risk-reversal beta for straddle beta for risk-reversal beta for
straddle risk-reversal straddle risk-reversal

36
4. Volatility of Realized P&L

Left figure: P&L volatility without accounting for transaction costs

Empirical hedge implies lower P&L volatility for:
[i] Risk-reversal (about 20%)
[ii] Straddle (about 2 − 3%)

Because Minimum Variance delta over-hedges for put side and make delta
more volatile

Right figure: volatility of realized P&L accounting for costs

1) Transaction costs increase P&L slightly by about 1 − 2%
2) Contrast with reduction of realized P&L by about 50%
400 P&L Volatility, zero transaction costs 400 P&L Volatility, after transaction costs
300 300

200 200
328 320 331 323
100 100
122 102 122 102
0 0
Min var for Empirical Min var for Empirical Min var for Empirical Min var for Empirical
straddle beta for risk-reversal beta for straddle beta for risk-reversal beta for
straddle risk-reversal straddle risk-reversal

37
5. Sharpe ratios of realized P&L-s

Left figure: Sharpe ratios for delta-hedging P&L without account-

ing for transaction costs
Right figure: Sharpe ratios for P&L accounting for costs

1) For straddle, both Minimum Variance and Empirical hedges imply

about the Sharpe ratio

2) For risk-reversal, Minimum Var hedge implies smaller Sharpe

than Empirical hedge (by about 20%) because of higher P&L volatility
and transaction costs
Sharpe ratio, zero tranaction costs Sharpe ratio, after transaction cost
2.00
4.00

3.00 1.50

2.00 4.14 1.00 1.88

3.46 1.56
1.00 0.50
0.74 0.76 0.49 0.50
0.00 0.00
Min var for Empirical Min var for Empirical Min var for Empirical Min var for Empirical
straddle beta for risk-reversal beta for risk- straddle beta for risk-reversal beta for risk-
straddle reversal straddle reversal

38
P&L Attribution to risk factors is applied to understand
what factors contribute to P&L by using regression
P&L = α + s1X1 + s2X2 + s3X3 + s4X4 + s5X5 + s6X6 (6)
α (”Alpha”) is theta related P&L - P&L we would realize if nothing would
move  2
S(tn )
−1
P
X1 (”Var”) is returns variance: X1 = S(t )
n−1

X2 (”VolChange”) is change in ATM vol: X2 =  Vatm(tn) − Vatm(tn−1)
P

S(tn )

− 1 Vatm(tn) − Vatm(tn−1)
P
X3 (”Covar”) is covariance: X3 = S(t )
n−1

2
Vatm(tn) − Vatm(tn−1)
P
X4 (”VarVol”) is variance of vol changes: X4 =
 3
S(tn )
X5 (”Return3”) is cubic return: X5 = −1
P
S(tn−1 )
 
S(tn )
−1
P
X6 (”Return”) is realized return: X6 = S(tn−1 )
P
Summation runs from n = 1 to n = N , N = 21

R2 indicates how well the realized variables explain realized P&L (not
accounting for transaction costs) - we should aim for R2 = 90%
Some explanatory variables are correlated so it is robust to test reduced
regressions
39
P&L explain for straddle by realized variance of returns:
Empirical hedge has stronger explanatory power
Is needed to confirm theoretical P&L explain by MC simulations
For P&L of straddle hedged at implied vol, first-order approximation:
" #2
2 −
X S(tn)
Vatm −1
n S(tn−1)
First term is alpha or ”carry” - approximate alpha is
2 = 100, 000 × 0.16752 = 2806
α = Γ × Vatm
Second term is short risk to realized variance - key variable for P&L
Theoretical slope should be −Γ = −100, 000
Figure: explanatory power using only realized variance is weak because
of impact of other variables and skew (for multiple variables, R2 ≈ 90%)
Straddle P&L by Min-Var Hedge Straddle P&L by Empirical hedge
4,000 P&L = -48,768*Var + 1,559 4,000 P&L = -55,132*Var + 1,730
R² = 30% R² = 40%
0 0
P&L

-4,000 -4,000 P&L

-8,000 Realized Variance -8,000 Realized Variance

0.00 0.05 0.10 0.15 0.00 0.05 0.10 0.15
40
P&L explain for risk-reversal by realized vol beta:
Empirical hedge implies that realized vol beta is clear
driver behind P&L of risk-reversal with R2 = 50%
For P&L of risk-reversal hedged at implied vol skew, approximation:
( 2 )
X  S(tn) X  S(tn) 
2
−SKEW × Vatm + −1 + − 1 (Vatm (tn ) − Vatm (tn−1 ))
n
S(tn−1 ) n
S(tn−1 )

In terms of returns vol Vret and implied vol beta β R :

 2 2
+ β [R] × Vret
2

−SKEW × Vatm + Vret
First term is ”carry” or alpha
Second term is risk to realized beta between returns and vol - key variable
In our example: α = 0.5 × V × {(16.75%)2 + (16.00%)2} = 2, 682
Slope= V × (16.00%)2 = 2, 560
Risk-Reversal P&L 3000 Risk-Reversal P&L 3000
by Min-Var Hedge by Empirical Hedge
P&L = 2129*Beta + 2570 1500 P&L = 2050*Beta + 2490 1500
R² = 37% R² = 49%
0 0
P&L

P&L
Realized Volatility Beta -1500 Realized Volatility Beta -1500
-2.0 -1.5 -1.0 -0.5 -2.0 -1.5 -1.0 -0.5 41
Important: vol beta (for skew) is comparable to Black-
Scholes-Merton (BSM) implied volatility (for one strike)
1) Volatility and vol beta are meaningful and intuitive model pa-
rameters which can be inferred from both implied and historical data
Implied vol σ [I] is inferred from option market price
Realized vol σ [R] is volatility of price returns
Implied vol beta β [I] is inferred from market skew (β [I] ≈ 2 × SKEW)
Realized vol beta β [R] is change in implied ATM volatility predicted by
price returns: β [R] = hdS(t)dVatm(t)i /(σ [R])2
2) Both serve as directs input for computation of hedges
3) Both allow for P&L explain of vanilla options in terms of implied
and realized model parameters:
Implied/realized volatility- P&L of delta-hedged straddle:
 2  2
σ [I] − σ [R]

Implied/realized volatility beta- P&L of short delta-hedged risk-reversal

(more noisy because 
of contribution from σ [R] ):
1 2 i2 2
 h 
[I]
−β × σ [I] + σ [R] +β [R] × σ [R]
2
42
Conclusion: existing practical approaches for hedging
improvement are not fully satisfactory - we need proper
model for dynamic delta-hedging!
A) Hedge all vega exposure
B) Recalibration for computing delta-risks (most common):
⊗ Project change in implied volatility using empirical backbone
(For example, by applying empirical volatility skew-beta)
⊗ Re-calibrate valuation model to bumped volatility surface
⊗ Re-valuate and compute delta by finite-differences
However runs into problems:
1) A) - vega-hedging is (very) expensive and unprofitable unless
implied skew and vol-of-vol are sold at large premiums to future realizeds

2) B) - re-calibration works poorly for path-dependent and multi-

asset products and it makes P&L explain very noisy
Recall applying regression for P&L explain of straddle and risk-reversal
3) any mix of A) and B) becomes very tedious for CVA computations

Important: the choice between local vol (LV) or stoch vol (SV) is irrel-
evant when hedging using minimum variance hedge at implied vol skew -
any combination of LV and SV produces almost the same deltas! 43
Beta SV model with jumps is fitted to empirical&implied
dynamics for computing correct delta (Sepp 2014):
dS(t)
= (µ(t) − λ(eη − 1)) dt + V (t)dW (0)(t) + (eη − 1) dN(t)
S(t)
dV (t) = κ(θ − V (t))dt + βV (t)dW (0)(t) + εV (t)dW (1)(t) + βη dN(t)
1) Consistent with empirical dynamics of implied ATM volatility by
specifying empirical volatility beta β
2) Has jumps, as degree of risk-aversion, to make model fit to both
empirical dynamics and risk-neutral skew premium
Only one parameter with simple calibration! - explained in a bit
Jumps/risk-aversion under risk-neutral measure Q produced by:
Poisson process N (t) with intensity λ:
negative&positive jumps in returns&vols with constant size η < 0&βη > 0

3) Easy-to-implement (with no extra parameters) extension to multi-

asset dynamics using common jumps - produces basket correlation skew

4) Beta SVJ model is robust to produce optimal hedges for path-

dependent and multi-asset trades and CVA
44
Third to last topic: closed-form solution for log-normal
Beta SV
Mean-reverting log-normal SV models are not analytically tractable
I derive a very accurate exp-affine approximation for moment generating
function (details in my paper)
Idea comes from information theory: apply Kullback-Leibler relative en-
tropy for unknown PDF p(x) and test PDF q(x) with moment constraints:
R k R k
x p(x)dx = x q(x)dx, k = 1, 2, ...
Now let’s think in terms of moment function:
[i] MGF for Beta SV model with normal driver for SV (as in Stein-Stein
SV model) has exact solution, which has exp-affine form
[ii] Correction for log-normal SV has an exp-affine form
35% Implied vol for 1y S&P500 35% Implied vol for 1y S&P500
30% options, beta SV, NO JUMPS 30% options, beta SV, WITH JUMPS
25% 25%
20% 20%
15% Analytic for Normal SV 15%
Analytic for Normal SV
10% Closed-form for Log-normal SV 10% Closed-form for Log-normal SV
Monte-Carlo for Log-normal SV Strike Monte-Carlo for Log-normal SVStrike
5% 5%
0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3

45
Proof that closed-form MFG for log-normal model pro-
duces theoretically consistent probability density
1) Derive solutions for excepted values, variances, and covariances of the
log-price and quadratic variance (QV) by solving PDE directly

2) Prove that moments derived using approximate MGF equal to theo-

retical moments derived in 1)

Using closed-form MFG for log-normal model, we apply standard valuation

methods for affine SV models based on Lipton-Lewis formula

Implementation of closed-form moment function (MGF), MC, and PDE

pricers produce values of vanilla options on equity and quadratic variance
that are equal within numerical accuracy of these methods
35% Implied vol for 1y S&P500 35% Implied vol for 1y S&P500
30% options, beta SV, NO JUMPS 30% options, beta SV, WITH JUMPS
25% 25%
20% 20%
15% Closed-form MGF 15% Closed-form MGF
10% Monte-Carlo 10% Monte-Carlo
PDE, numerical solver PDE, numerical solver
5% Strike 5% Strike
0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3

46
Second to last topic - optimal hedging under discrete
trading and transaction costs

As we saw in simulation of P&L, we need quantitative framework that

incorporates discrete hedging and optimizes trade-off between:
the reward - higher P&L and lower transaction costs
the risk - higher P&L volatility

47
Illustration of trading in implied&realized vol with strad-
dle: unique optimal hedging frequency can be found!
Figure 1) Forecast expected upside: 3%
Expected/Forcasted P&L(N)
the spread between implied and real-
2%
ized vol for given maturity T
This is independent of valua- 1%
tion&hedging model and hedging
N - hedging frequency
frequency 0%
10 160 310 460 610 760 910

Figure 2) Forecast P&L volatility 3% Expected/Forcasted P&L

Volatility and Costs (N)
and transaction costs P&L Volatility
2%
These depend on valuation&hedging Transaction costs
model and hedging frequency 1%

N - hedging frequency
Part of P&L volatility is not hedge- 0%
10 160 310 460 610 760 910
able due to vol-of-vol and jumps -
1.4 Expected/Forcasted
Not optimal to hedge too fre- 1.2 Sharpe Ratio (N)
quently 1.0
0.8
0.6
Figure 3) Obtain Sharpe ratio as ra- 0.4
tio of forecast P&L after costs and 0.2
N - hedging frequency
P&L volatility 0.0
10 160 310 460 610 760 910
48
Solution for optimal Sharpe ratio with dynamics under
physical measure driven by Diffusion and SV with jumps
Expected P&L − TransactionCosts(N )
Sharpe(N ) =
P&L Volatility(N )
N is hedging frequency - for details see my paper on optimal delta-hedging
Using this solution we can analyze:
Figure 1) What maturity is optimal to trade given the forecast spread
between implieds and realizeds
(longer maturities have higher spreads but their P&L is more volatile
because of higher risk to ATM vol changes)
Figure 2) What is optimal hedging frequency for each maturity
Translate into approximations of optimal bands for price and delta triggers
Naturally, results are sensitive to assumed price dynamics
Under SV with jumps: lower Sharp ratio and less frequent hedging
Optimal Sharpe ratio 12 Optimal hedging frequency in days
1.6 Diffusion Diffusion
10
Stochastic volatility with jumps Stochastic volatility with jumps
1.3 8
6
1.0
4
0.7 2
0.4 Option maturity in months 0 Option maturity in months
1m
4m
7m
10m
13m
16m
19m
22m
25m
28m
31m
34m
37m

1m
4m
7m
10m
13m
16m
19m
22m
25m
28m
31m
34m
37m
49
 Last topic: why the beta stochastic vol model with
jumps is better than its alternatives (for stock indices)

The most important feature for dynamic hedging model:

1) Ability to produce different volatility regimes as observed in the
market and to imply empirically consistent delta
Recall definition of volatility skew-beta: change in term structure of ATM
volatility, σAT M (T ), predicted by price return times SKEW(T )
We saw that vol skew-beta is very important to account for correct P&L
arising from change in BSM implied vols
Skew-consistent SV and LV models imply skew-beta of 2
Empirical vol skew-beta: S&P 500 ≈ 1.5; STOXX50 ≈ 1.8; Nikkei≈ 0.5
2.50 2.50 2.50
Vol Skew-Beta for S&P500 Vol Skew-Beta for STOXX 50 Vol Skew-Beta for NIKKEI
2.00 2.00 2.00
SVJ Skew-Beta with empirical beta
1.50 1.50 1.50 Sticky local with Min-var delta
1.00 SVJ Skew-Beta with empirical beta 1.00 1.00 Empirical bounds
SVJ Skew-Beta with empirical beta
0.50 Sticky local with Min-var delta 0.50 Sticky local with Min-var delta 0.50
Empirical bounds Empirical bounds
0.00 T in months 0.00 T in months 0.00 T in months
1m
3m
5m
7m
9m
11m
13m
15m
17m
19m
21m
23m

1m
3m
5m
7m
9m
11m
13m
15m
17m
19m
21m
23m

1m
3m
5m
7m
9m
11m
13m
15m
17m
19m
21m
23m
50
Why the beta SV with jumps is better than its alterna-
tives
Extra arguments to look at apart from implied volatility skew-beta

2) Fit to empirical distribution of implied and realized volatilities

3) Interpretation of model parameters in terms of impact on model implied

BSM vols

4) P&L explain for delta-hedging strategies of vanilla options in terms of

implied and realized model parameters

5) Stability of model parameters

Calibration to vanilla options is not a problem in practical applications -

it is easy to achieve by introducing a (small) local vol part

Calibration problem is solved by Dupire (1994) for diffusions, Andersen-

Andreasen (2000) for jump-diffusions, Lipton (2002) for SV with jumps

51
I. Non-parametric local volatility model - textbook im-
plementation of Dupire local volatility using discrete set
of option prices and interpolation

52
 II. Industry-standard alternative (in equity derivatives)
Implied volatility@strike-into-density@price approach (my terminology)
Conceptually:
σimpl (K; T ) → Pimpl (S(T ) = K) (7)
where→ is Dupire LV formula in terms of implied vols at strike K&mat T
Figure 1A) Given parametric form for implied vols σimpl (K; T )
Figure 1B) Given backbone function fbackbone(δS; K, T ) to map price
changes δS into changes in vols δσimpl (K; T ) according to specified regime

Figure 2) → in Eq(7) serves as interpolator from implied vols in strike

space to implied densities in price space
Figure 3) LV model projects densities to option prices in ”model-independent”
way using MC or PDE methods
50% Implied Volatility (S0=1.00) 2.25% Implied Density from LV mapping (S0=1.00) 0.50 PV Risk-Reversal 95-105% (S0)
Implied Volatility (S1=0.95) Implied Density from LV mapping (S1=0.95) PV Risk-Reversal 95-105% (S1)
40% 0.40 Change in PV Risk-Reversal 95-105%
Change in Density from LV mapping
Change in IV from backbone function 1.50%
Implied Volatility

0.30
30%
Density

PV
0.20
20% 0.75%
0.10
10% 0.00% 0.00
Strike
0% 0.6 0.7 0.8 0.9 1.0 1.1 1.2 -0.10 0.6 0.7 0.8 0.9 1.0 1.1 1.2
Spot Spot
0.6 0.7 0.8 0.9 1.0 1.1 1.2 -0.75% -0.20 53
1) Hedging performance for local vol approach are pri-
mary driven by parametric form for implied vols σimpl (K; T )
and empirical backbone function

2) No consistency with empirical distribution of implied and realized vol

3) & 4) Model interpretation and P&L explain are possible only in terms
of parameters of functional form for implied volatility

Key drawback of implied volatility-into-density approach:

[i] For computation of delta it requires a re-calibration of local vol and

re-valuation for any change in market data

[ii] Lacks vol-of-vol so it is inconsistent for hedging of path-dependent

options sensitive to forward vols and skews

54
Alternatives for local vol or σimpl (K; T ) → Pimpl (S(T ) = K)
approach do not produce improvements
Instead of LV to map implied vol into price density, it is also customary
to use SV or LSV models as interpolators with extra degree of freedom
Hereby hodel choice is typically motivated by availability of a ”closed-
form” solution, not empirical consistency!
Figure: SV and LSV models are not applied for hedging as dynamic
models since their model delta is wrong - with and without minimum
variance hedge - but through re-calibration to empirical backbone
Change in implied vol, S1-S0=-0.05 Delta for 1y call option on 1.00
3.5%
SV model delta S&P500 0.80
2.5% SV with Min Var Hedge
Empirical backbone 0.60
1.5%
SV model 0.40
0.5% SV with Min Var hedge
Strike Sticky-Strike BSM delta 0.20
-0.5% 0.70 0.80 0.90 1.00 1.10 1.20 SV re-calibrated to empirical backbone
0.00
-1.5% 0.70 0.80 0.90 1.00 1.10 1.20 Strike

To conclude I use a quote from Richard P. Feynman:

It doesn’t matter how beautiful your theory is, it doesn’t matter
how smart you are. If it doesn’t agree with experiment, it’s wrong!
55
III. Arguments in favor of Beta SVJ model:
1) the model has ability to fit empirical vol skew-beta
and produce correct option delta without re-calibration
Figure: delta from SVJ model fits empirical backbone
1.00 Delta for 1y call option on
0.80 S&P500

0.60
SV model
0.40
SV with Min Var hedge
0.20 SV re-calibrated to Market backbone

0.00 SVJ (without re-calibration)

0.70 0.80 0.90 1.00 1.10 1.20
Strike

2) Consistent with the empirical distributions of implied and realized

volatilities, which are very close to log-normal
3) It has clear intuition behind the key model parameters:
Volatility beta is sensitivity to changes in short-term ATM vol
Residual vol-of-vol is volatility of idiosyncratic changes in ATM vol
4) P&L explain is possible in terms of implied and realized quantities
of key model parameter - vol beta
5) Stability and calibration - next slide
56
Calibration of beta SV model is based on econometric
and implied approaches without large-scale non-linear
and non-intuitive calibrations
1) Parameters of SV part are estimated from time series
2) Jump/risk-aversion params are fitted to empirical vol skew-beta
Params in 1) & 2) are updated only following changes in volatility regime

3) Small mis-calibrations of the SV part and jumps are corrected

using local vol (LV) part
Contribution to skew from LV part is kept small (no more than 10-15%)
Local vol part is re-calibrated on the fly to reproduce small variations in
some parts of implied vol surface, which are caused by temporary supply-
demand factors specific to that part
It is also robust to compute bucketed vega risk in this way
In practical terms:
1) Local volatility part accounts for the noise from idiosyncratic
changes in implied volatility surface
2) Stochastic volatility and jumps serve as time- and space-homogeneou
factors for the shape of the implied volatility surface
57
More details on calibration of beta SV model (technical
part omitted during the talk)
1) Parameters of SV part are calibrated using maximum likelihood meth-
ods from time series of 1m implied ATM volatility (or the VIX)
[i] mean-reversion κ is estimated over longer-period, at least 5 years, -
better to keep it constant at 3.00
[ii] vol beta β and residual vol-vol ε are estimated over shorter periods,
1y, - typically β ≈ −1.00 and ε ∈ [0.60, 1.00]

2) Negative jump in return η is fitted using Merton jump model to put

options with maturity of 6 months and [80% − 100%] OTM strikes -
typically η = −30%

3) Given 1) and 2): 3A) jump intensity λ is calibrated to fit the empirical
sensitivity of implied volatility changes to price changes, aka volatility
skew-beta, - typically λ ∈ [0.03, 0.2]
Given all above: 3B) initial vol V (0) and mean vol θ calibrated to fit the
current term structure of ATM vols

Parameters in 1), 2) 3A) (relatively uniform for major stock indices) are
updated infrequently

4) Local vol part is added to fit daily variations in implied vol surface
58
For risk-neutral pricing, distribution of jumps does not
matter - jumps are only needed to fit skew premium
Recall illustration of emergence of Q- 10% Frequency of S&P500 daily returns
skew using Bakshi-Kapadia-Madan 8%
Empirical
Frequency
formula and Merton jump model:

Frequency
6% Physical Merton
[i] Under P, jumps are symmetric under P
4% Risk-Neutral
with mean of 0% and volatility of 4% Merton under Q
[ii] The risk-neutral mean jump is 2%
Daily return
−5% with zero volatility 0%
-9% -7% -5% -4% -2% 0% 2% 4% 6% 7%

Yet, jumps are needed to fit market prices & compute correct deltas

Also jumps are important to fit market prices of options on realized and
implied volatilities (VIX) - see my presentations at GD in 2011 & 2012
Practical explanation for excessive risk-neutral skew premium:
1) Risk-averse investors always ready to over-pay for insurance ir-
respectively of price changes
2) As part of index correlation skew premium, when holders of
stock portfolios buy index puts for (macro) protection
To make things robust, I assume constant jumps with simple calibration
59
How to explain the difference between implied
and realized dynamics using preference theory
For retail option buyer - option value is derived from his preference/utility
for specific payoffs in certain market scenarios

For institutional option seller - option value is derived from:

[i] Expected hedging costs
[ii] Smooth stream of fees and P&L
[iii] Premium for suffering losses in bad market conditions

As a result:

1) Option prices in the market are set by demand-supply equilibrium

between sellers and buyers

2) Risk-aversion parameters is a degree of demand-supply imbal-

ance

3) Implied and realized vols and, in particular, skews are different

60
To conclude: we can think of jumps as a mea-
sure of risk-aversion for pricing kernel!
Recently, interesting research is made and also presented at Global Deriva-
tives on how to imply the ”expected-implied” physical distribution from
options market prices and specified risk-aversion

Stephen Ross:The Recovery Theorem, GD2012, Journal of Finance 2014

Peter Carr: Can we recover?, Global Derivatives 2013

Computation of ”empirical” delta and calibration of excessive skew are

related concepts:
1) Compute option delta under ”expected-implied” physical distri-
bution using empirical vol beta
2) Fit level of risk-aversion to excessive skew premium observed in
market prices of index options

These concepts and volatility skew-beta are related to the interplay be-
tween the implied and realized risk premiums:
[i] high implied / positive realized risk premiums - sticky strike vol regime
[ii] low implied / negative realized risk premiums - sticky local vol regime

61
Summary 7%
6%
Empirical frequency of
normalized logarithm of the VIX
1) Dynamics of implied and real- 5% Empirical

Frequency
4% Standard Normal
ized vols are log-normal 3%
2%
2) Implied vol beta significantly 1%
Log-VIX
overestimates realized beta 0%
-4 -3 -2 -1 0 1 2 3 4

3) Vol skew-beta is important for

May-12
Mar-11
Nov-08

Aug-10

Dec-12
Sep-07
Apr-08
Feb-07

Oct-11
Jun-09
Jan-10

Jul-13
correct P&L - any dynamic hedging 0.0 Implied Volatility Beta
model should fit empirical skew-beta -0.5
Realized Volatility Beta

Risk-Aversion/Jumps parameter -1.0

is added to fit empirical skew-beta -1.5

SVJ fits empirical skew-beta≈ 1.5, -2.0

unlike Minimum Var delta≈ 2.0 2.50

Vol Skew-Beta for S&P500
2.00
4) Beta SVJ model applied for 1.50
delta-hedging risk-reversal is tool to 1.00 SVJ Skew-Beta with empirical beta
produce P&L from spread be- 0.50 Sticky local with Min-var delta
Empirical bounds
tween implied and realized skews 0.00 T in months

1m
3m
5m
7m
9m
11m
13m
15m
17m
19m
21m
23m
Log-normal beta SVJ model: Risk-Reversal P&L vs 3000

⊗Is consistent with the empirical Realized Vol Beta

P&L = 2050*Beta + 2490 1500
dynamics of ATM volatility R² = 49%

P&L
⊗Produces correct option deltas 0

⊗Can significantly improve Shar- Realized Volatility Beta -1500

pe ratios for delta-hedging P&Ls -2.0 -1.5 -1.0 -0.5

62
Disclaimer

The views represented herein are the author own views and do not neces-
sarily represent the views of Bank of America Merrill Lynch or its affiliates

63
References
Andersen, L., Andreasen, J., (2000), “Jump-Diffusion Processes - Volatility Smile Fitting
and Numerical Methods for Option Pricing ,” Review of Derivatives Research 4, 231-262
Bakshi, G., Kapadia, N., Madan, D., (2003), “Stock return characteristics, skew laws,
and the differential pricing of individual equity options,” Review of Financial Studies 16
(1), 101-143
Bergomi, L., (2009), “Smile dynamics 4,” Risk, December, 94-100
Bergomi,L.,Guyon,J.,(2012),“Stochastic volatility’s orderly smiles,”Risk,May,117-123
Christoffersen, P., Jacobs, K., Mimouni, K., (2010), “Models for S&P 500 Dynam-
ics: Evidence from Realized Volatility, Daily Returns and Options Prices,” Review of
Financial Studies 23(8), 3141-3189
Derman, E., (1999), “Volatility Regimes,” Risk, April, 55-59
Dupire, B. (1994), “Pricing with a smile”, Risk, July, 18-20
Karasinski, P., Sepp, A., (2012), “Beta stochastic volatility model,” Risk, October,
67-73
Lewis, A., (2000), “Option valuation under stochastic volatility,” Finance Press, New-
port Beach, California
Lipton, A., (2002). “The vol smile problem”, Risk, February, 81-85
Ross, S., (2014). “The Recovery Theorem”, Journal of Finance, forthcoming
Sepp, A., (2011), “Efficient Numerical PDE Methods to Solve Calibration and Pricing
Problems in Local Stochastic Volatility Models”, Global Derivatives conference in Paris
Sepp, A., (2012), “Achieving Consistent Modeling Of VIX and Equities Derivatives”,
Global Derivatives conference in Barcelona
Sepp, A., (2013), “Consistently Modeling Joint Dynamics of Volatility and Underlying
To Enable Effective Hedging”, Global Derivatives conference in Amsterdam
Sepp, A., (2013), “When You Hedge Discretely: Optimization of Sharpe Ratio for
Delta-Hedging Strategy under Discrete Hedging and Transaction Costs,” The Journal
of Investment Strategies 3(1), 19-59
Sepp, A., (2014), “Empirical Calibration and Minimum-Variance Delta Under Log-
Normal Stochastic Volatility Dynamics,” Working paper, http://ssrn.com/abstract=2387845
Sepp, A., (2014), “Log-Normal Stochastic Volatility Model: Pricing of Vanilla Options
and Econometric Estimation,” Working paper, http://ssrn.com/abstract=2522425
64