What is a robust stochastic volatility model

Artur Sepp* Parviz Rakhmonov„

27th November, 2023

Abstract
We address specification of the functional form for the dynamics of stochastic volatility (SV)
driver including affine, log-normal, and rough specifications. We propose the four principles
which, in our opinion, determine the applicability of an SV model for valuation of derivative
securities for different asset classes including equities, rates, commodities, FX and cryptocur-
rencies. We emphasise that the invariance of an SV under different numeraires is crucial for
the model applications for modeling volatility of different asset classes. We argue that cur-
rently only the two SV dynamics satisfy these universality conditions: affine Heston SV model
and log-normal SV model with quadratic drift. We discuss that both models are analytically
tractable for valuation of vanilla options and model calibration when applying these models
in different asset classes. We also present some empirical evidence for the considered models
and discuss their link with contemporary research topics such as volatility skew-stickiness. We
conclude that log-normal SV model with quadratic drift is robust because it does not require
special conditions (such as Feller condition for Heston model) for numerical implementation of
the model using MC and PDE methods.

1 Introduction
We formulate the following principles for universality and feasibility of a stochastic volatility (SV)
model. Our primary focus is based on specifying the parametric form of the dynamics of the
volatility driver so that we leave aside important but, in our opinion, secondary features of a
universal volatility model including jumps, local volatility, etc.

1. The dynamics of volatility must have the same marginal distribution under statistical measure
P and risk-neutral valuation measure Q. This point ensures that the model can be used under
the both statistical and pricing measures.
More generally, this requirement implies that the model can be used with different numéraires
specific to different asset classes, including equities, rates, commodities, FX and cryptocurren-
cies. For universality of a SV driver, the SV model dynamics must be functionally invariant
under different numéraires. For an example, for interest rates derivatives it is necessary that
the volatility dynamics are invariant under the annuity measure, while for options on FX and
cryptocurrencies the model must be invariant under the price numéraire.

2. The price process augmented with stochastic volatility must remain a strict martingale under
different model specification. This particular point is important for the model application for
assets with positive implied volatility skews and, as a result, with positive return-volatility
correlation1 .
*
Clearstar Labs, artursepp@gmail.com
„
Marex, parviz.msu@gmail.com
1
Lions and Musiela (2007) show that most of one-factor SV models fail to produce strict martingale dynamics
when return-volatility correlation between SV and return drivers is positive. This point can be overlooked in equity
derivatives where return-volatility correlation is strongly negative but it cannot neglected for other asset classes where
return-volatility is positive on many occasions.

Electronic copy available at: https://ssrn.com/abstract=4647027

 3. The dynamics of volatility must be well behaved: the volatility process must be strictly
positive without explosions, the stationary distribution of the volatility must exist. This point
ensures that the model can implemented efficiently with analytical and numerical methods.

4. The model is relatively easy to implement both analytically (for model calibration to market
data) and numerically (through Monte-Carlo and PDEs) for valuation of exotic options and
structured products.

We also mention that another important condition is the model consistency with empirical
features observed for a given market of interest. While this condition is important, it is also a
subjective topic due to multiple possibilities in estimation: using a specific data set, a single asset
class, a specific frequency or sampling period, etc 2 . In Section 3 we provide our empirical analysis
for volatility indices including VIX index (for options on the S&P500 index), MOVE index (for
options on US treasury bonds), and OVX index (for options on US Oil Fund USO).
In Section 2 we review the following most popular SV models and we check the validity of these
conditions for each of these models:

1. Stein-Stein model (2.1),

2. Exp-OU SV model (2.2),

3. Bergomi model (2.3),

4. Heston Model (2.4),

5. Rough Volatility (2.5),

6. Log-normal volatility model (2.6),

7. Log-normal SV model with quadratic drift (2.7).

In Section 4 we conclude with a discussion.

2 Common SV models
As noted in Herdegen and Schweizer (2018), stochastic volatility models are incomplete by con-
struction and completeness is achieved by fixing volatility risk premium to complete the market,
as outlined in Lewis (2000) for the case of Heston model. If the model is not invariant under the
change from statistical measure P to risk-neutral measure Q, the model may be misspecified from
the start.

2.1 Stein-Stein model

Stein-Stein model dynamics under P for the price St and volatility σt are given by:
dSt = St (µt dt + σt dWt ) , S0 = S,
(1)
dσt = κ (θ − σt ) dt + εdZt , σ0 = σ,

where µt is the price drift, κ is the mean-reversion speed and θ is the long-run variance, ε is the
volatility of volatility, Wt and Zt are Brownian motions with return-volatility correlation ρ.
2
We only mention substantive study of Christoffersen, Jacobs and Mimouni (2010) who examine the empirical
performance of Heston, log-normal and 3/2 SV models for options on the S&P 500 index using three sources of
market data: the VIX index, the implied volatility for options of the S&P500 index, and the realized volatility of
returns on the S&P500 index. They found that, for all three sources of data, the log-normal SV model outperforms
its alternatives. Such a study could be generalized to all other asset classes, but we have not yet seen any extensions
in this direction.

Electronic copy available at: https://ssrn.com/abstract=4647027

 As shown in Wong and Heyde (2006), Stein-Stein model is ”immediately arbitrageable”. To
present their result, we rewrite the dynamics (1) in terms of independent Brownian motions as
follows
(0)
dSt = µt St dt + σt St dWt , S0 = S,

(0) (1)
 (2)
dσt = κ (θ − σt ) dt + ε ρdWt + ρ∗ dWt , σ0 = σ,
p
where W (0) and W (1) are independent Brownian motions and ρ∗ = 1 − ρ2 . We assume that P
and Q are related by the Radon-Nikodym derivative
 Z t 
dQ (0) (1)
ζ(t) := =E − λ0 (u) dWu + λ1 (u) dWu (3)
dP Ft 0

where E(·) denotes Itô exponential, and λ0 (t) and λ1 (t)are asset and volatility risk-premia functions.
If process ζ(t) was a true martingale, then the processes
Z t
(0) (0)
Ŵt = Wt + λ0 (u) du,
0
Z t
(1) (1)
Ŵt = Wt + λ1 (u) du
0

would define independent Brownian motions under Q by Girsanov theorem. Then, the dynamics
of price process under risk-neutral measure Q becomes
(0)
dSt = (µt − λ0 (t)σt ) St dt + σt St dŴt

As discounted price process must be martingale under Q, we must necessarily have µt −λ0 (t)σt = rt ,
where rt is a risk-free rate. However, if µt ̸= rt , such λ0 (t) might not even exist as volatility process
σt , which is driven by an Ornstein-Uhlenbeck (OU) process, attains 0 for any t > 0.

2.2 Exp-OU SV model

Exp-OU model dynamics under P for the price St and volatility σt :

dSt = µt St dt + eyt St dWt , S0 = S,

(4)
dyt = κ (θ − yt ) dt + εdZt , y0 = y,

where Wt and Zt are Brownians with return-volatility correlation ρ.

Before we study the properties of the exp-OU model, we rewrite the dynamics in terms of
volatility σt = eyt :
(0)
dSt = µt St dt + σt St dWt , S0 = S,

ϑ2
 
(0) (1)
 (5)
dσt = σt κθ + − κ ln σt dt + εσt ρ dWt + ρ∗ dWt
2
p
where W (0) , W (1) are independent Brownian motions, ρ∗ = 1 − ρ2 . We see that the volatility
has non-linear mean reversion due to the presence of logarithmic term in the drift.
We assume that statistical P and risk-neutral measures Q are related by
 Z t 
dQ (0) (1)
ζ(t) := =E − λ0 (u)dWu + λ1 (u)dWu (6)
dP Ft 0

Electronic copy available at: https://ssrn.com/abstract=4647027

where E(·) denotes Itô exponential, and λ0 (t) and λ1 (t) are asset and volatility risk-premia func-
tions. We assume that the market price of risk is proportional to the volatility3 :

λ0 (t) = λ̄0 σt , λ1 (t) = λ̄1 σt . (7)

Proposition 2.1. We assume that κ ≥ 0, θ > 0. Then measures P and Q are equivalent, iff

ρλ̄0 + ρ∗ λ̄1 ≥ 0 (8)

Proof. The dynamics of σt under Q becomes

ϑ2
     
(0) (1)
dσt = σt κθ + − κ ln σt − σt (ρλ̄0 + ρ λ̄1 ) dt + εσt ρ dŴt + ρ∗ dŴt
2 ∗
(9)
2

where Ŵ (0) , Ŵ (1) are Brownian motions under Q. We see that the process in (9) explodes if the
qudratic ”‘mean-reversion”’ term with ρλ̄0 + ρ∗ λ̄1 is negative.

Proposition 2.2 (Martingale property). We assume that κ1 ≥ 0, θ > 0. Then discounted price
process in Eq (5) is martingale under Q, iff

ρ≤0

Proof. The proof is very similar to Proposition 2.1. We consider measure Q̃ where price process is
chosen as a numéraire. Radon-Nikodym derivative, linking Q and Q̃, equals
Z t 
dQ̃
=E σu dWu(0)
dQ 0
Ft

Using same arguments, we observe that discounted price process is a Q-martingale if and only
if the process

ϑ2
     
(0) (1)
dσt = σt κθ + − κ ln σt + ρεσt dt + εσt ρ dW̃t + ρ∗ dW̃t
2
(10)
2

has unique strong solution under Q̃. We conclude that σt in (10) explodes as long as ρ > 0.

We note that due to the presence of non-zero quadratic term in the drift of the volatility in Eqs
(9) and (10), Exp-OU SV model has different functional form than the original process in Eq (5).
Thus Exp-OU SV model is not functionally invariant under different numéraires. Also, exp-OU
suffers from the ”loss of martingality” when return-volatility correlation is positive, i.e. ρ > 0. We
conclude that Exp-OU model is not robust for applications under different numéraires and when
return-volatility correlation is positive. This conclusion is the same for log-normal SV model with
linear drift which we study in Section 2.6.

2.3 Bergomi model

We consider the one-factor Bergomi model expressed using the forward variance ξt (u)
p
dSt = µt St dt + ξt (t)St dWt , S0 = S,
 
−κ(u−t) 1 2 −2κ(u−t) 2
ξt (u) = ξ0 (u) exp ηe Yt − η e E[Yt ] (11)
2
dYt = −κYt dt + dZt
3
Such linear specification is the simplest functional form that allows to establish simple relationship between gains
on a delta-hedged option portfolio and level of the market volatility. As shown in Bakshi-Kapadia (2003) such form
allows to get precise information about the sign and strength of the risk premium.

Electronic copy available at: https://ssrn.com/abstract=4647027

where Wt and Zt are Brownians with return-volatility correlation ρ.
As shown in Drimus (2012), instantaneous variance ξt (t) in one-factor Bergomi model is log-OU
process so that the logarithm of the volatility follows an OU process. As a result, the model suffers
from the same issues which we concluded for Exp-OU SV model in Section 2.2. In particular,
results of Propositions 2.1 and 2.2 are carried over directly to the one-factor Bergomi model.

2.4 Heston Model

Heston SV model dynamics under statistical measure P are specified for the price St and returns
variance Vt as follows:
p
dSt = µt St dt + Vt St dWt , S0 = S,
p (12)
dVt = κ (θ − Vt ) dt + ε Vt dZt , V0 = V,
where µt is price drift, κ is mean-reversion speed and θ is long-run variance, ε is volatility of
volatility, Wt and Zt are Brownian motions with return-volatility correlation ρ.
We first address the measure changes in Heston model. To have consistent notation, we rewrite
the dynamics (12) in the following form
(0)
p
dSt = µt St dt + Vt St dWt ,
p  (0) (1)
 (13)
dVt = κ (θ − Vt ) dt + ε Vt ρdWt + ρ∗ dWt
p
where W (0) and W (1) are independent Brownian motions and ρ∗ = 1 − ρ2 . We assume that
statistical measure P and risk-neutral measure Q are related by the Radon-Nikodym derivative
 Z t 
dQ (0) (1)
ζ(t) := =E − λ0 (u) dWu + λ1 (u) dWu (14)
dP Ft 0

where E(·) denotes Itô exponential, and λ0 (t) and λ1 (t) are risk-premiums. Assuming that process
(14) is a true martingale, the processes
Z t
(0) (0)
Ŵt = Wt + λ0 (u) du,
0
Z t
(1) (1)
Ŵt = Wt + λ1 (u) du
0

define independent Brownian motions under Q according to Girsanov theorem. Thus, the dynamics
of price process under risk-neutral measure Q become
 p  (0)
p
dSt = µt − λ0 (t) Vt St dt + Vt St dŴt

As discounted price process must be martingale under Q, we need to require that

p µt − rt
µt − λ0 (t) Vt = rt ⇐⇒ λ0 (t) = √ (15)
Vt
Next, the dynamics of the variance process under risk-neutral measure Q become
p  h (0)
i h
(1)
i
dVt = κ (θ − Vt ) dt + ε Vt ρ dŴt − λ0 (t) dt + ρ∗ dŴt − λ1 (t) dt =
h  p  i p  
(0) (1)
= κθ − κVt + ρελ0 (t) Vt + ρ∗ ελ1 (t) Vt dt + ε Vt ρdŴt + ρ∗ dŴt
p

Following Heston (1993), we assume that market price of risk λ0 (·) and market price of volatility
risk λ1 (·) are proportional to the volatility, i.e.
p p
λ0 (t) = λ̄0 Vt , λ1 (t) = λ̄1 Vt (16)

Electronic copy available at: https://ssrn.com/abstract=4647027

Then, using Eqs (15) and (16) the Q-dynamics of price and volatility becomes
(0)
p
dSt = rt St dt + Vt St dŴt ,
p  (0) (1)
 (17)
dVt = κθ − (κ + ρελ̄0 + ρ∗ ελ̄1 )Vt dt + ε Vt ρdŴt + ρ∗ dŴt
 

We can relabel the parameters

κ
κ̂ := κ + ρελ̄0 + ρ∗ ελ̄1 , θ̂ := θ
κ + ρελ̄0 + ρ∗ ελ̄1
and rewrite dynamics in a familiar form
(0)
p
dSt = rt St dt + Vt St dŴt ,
p  (0) (1)
 (18)
dVt = κ̂(θ̂ − Vt ) dt + ε Vt ρdŴt + ρ∗ dŴt

We now specify the domain of model parameters. First, we need to impose Feller condition
2κθ ≥ ε2 to ensure that Vt does not reach 0, which is important for stability of the equity risk-premia
in Eq (15).

Proposition 2.3. Measures P and Q are equivalent, if κ > 0, θ > 0, ε > 0.

Proof. As seen from (17), dynamics of Vt under Q is

dVt = κθ − (κ + ρελ̄0 + ρ∗ ελ̄1 )Vt dt + ε Vt dŴt

  p
(19)

where Ŵ is a Brownian motion under Q. Thus, measures remain equivalent, if process (19) does
not explode. The result follows from Feller boundary classification, see Andersen and Piterbarg
(2007).

Finally, we establish the domain of model parameters to ensure that price process in Eq (18) is
well-behaved.

Assumption 2.1. We assume that model parameters κ̂, θ̂, ε in (18) satisfy

κ̂ > 0, θ̂ > 0, ε>0 (20)

Thus the variance under the Heston model satisfies our third principle about ”well-behaved”
condition only partially because the variance can hit zero when Feller condition is not satisfied.

Proposition 2.4 (Martingale property). Discounted price process is martingale under Q, if con-
ditions (20) are satisfied.

Proof. See Andersen and Piterbarg (2007).

Thus, Heston model preserves the functional form under different numéraires and it is not
sensitive to the sign of return-volatility correlation. Having said that, we note that volatility
process in Heston model might become mean repelling under price numéraire when return-volatility
correlation is positive,

Electronic copy available at: https://ssrn.com/abstract=4647027

2.5 Rough Volatility
Gatheral-Jaisson-Rosenbaum (2015) introduce the so called rough fractional SV model (RFSV).
The volatility dynamics are modeled using the driver
Z t
dXt = ν e−α(t−s) dWtH + m (21)
−∞

where W H is a fractional Brownian motion with Hurst parameter H such that H < 1/2, α > 0
is mean-reversion rate, m is the mean of the volatility, ν is the volatility of volatility. The model
dynamics are then specified by:
dSt = µt St dt + σt St dWt , S0 = S,
(22)
σt = eXt
Further Gatheral-Jaisson-Rosenbaum (2015) consider the limiting case α → 0 arguing that: “con-
sequently, although the RFSV model is technically stationary, its ergodic behaviour is of no interest
for us”.
In our opinion, the discussion about the rough is predominantly focused on the model behaviour
at the very short scales for high frequency data. Cont-Das (2022) conclude that “the origin of the
roughness observed in realized volatility time-series may be due to the estimation error rather than
the volatility process itself”. We take a pragmatic view to check the rough volatility model would
satisfy our conditions 1 to 4.
Since model dynamics (21) and (22) require to full path of fractional Brownian motion WtH ,
Bayer-Friz-Gatheral (2015) consider the following kernel-based specification:

P
√ Z u
dWsP
W̃t (u) = 2H γ
(23)
t (u − s)

where γ = 1/2 − H. So that the variance is defined by

 
vu = E [vu | Ft ] E η W̃tP (u) (24)
√
where η = 2νCH / 2H and E is the Itô exponential.
This specification assumes that the conditional distribution of vu depend on the whole filtration
Ft only through the forecast of the variance E [vu | Ft ], u > t. When pricing options, we can compute
the implied forecast so that the entire history of WsP is complimentary.
As a result, Bayer-Friz-Gatheral (2015) introduce the so-called rough Bergomi or rBergomi
model with the following dynamics under P:
√
dSt = µt St dt + vt St dZtP , S0 = S,
  (25)
vu = E [vu | Ft ] E η W̃tP (u)

where WtP = ρZtP + ρ∗ Z̄tP , ZpP and Z̄ P are independent Brownian motions under P and ρ is return-
volatility correlation, ρ∗ = 1 − ρ2 .
We assume that statistical P and risk-neutral measures Q are related by
 Z t 
dQ P P
ζ(t) := =E − λ0 (u)dZu + λ1 (u)dZ̄u (26)
dP Ft 0

where E(·) denotes Itô exponential. Assuming that process (26) is a true martingale, the processes
Z t
Q P
Zt = Zt + λ0 (u) du,
0
Z t
Q P
Z̄t = Z̄t + λ1 (u) du
0

Electronic copy available at: https://ssrn.com/abstract=4647027

are independent Brownian motions under Q. As a result, Q-dynamics of price process become
(0)
dSt = (µt − λ0 (t)σt ) St dt + σt St dŴt

Since we require the discounted price to be martingale under Q, we impose the following condi-
tion
µt − rt
µt − λ0 (t)σt = rt ⇐⇒ λ0 (t) = (27)
σt
Bayer-Friz-Gatheral (2015) consider deterministic change of measure assuming that functions λ0 (u)
and λ1 (u) are deterministic in Eq (26). They state that this is the simplest change of measure
attainable in this model which allows to retain analytical tractability. Under this assumption, the
variance under Q has a form  
vu = EQ [vu | Ft ] E ηWtQ (u) (28)
where
√ u
 Z 
Q P −γ
E [vu | Ft ] = E [vu | Ft ] × exp η 2H (u − s) λ(s) ds .
t

In general, we cannot hope to derive analytic expression for the forward variance curve EQ [vu | Ft ]
for realistic specifications of risk-premia processes λ0 (u), λ1 (u) and for non-zero return-volatility
correlation ρ ̸= 0. A deterministic change of measure, in itself, is not realistic because in our discus-
sion we consider the traditional way to specify risk-premia proportional to the volatility following,
among others, Heston (1993) and Bakshi-Kapadia (2003). Additionally, in case of price numéraire,
which is essential for valuation of options on cryptocurrencies, λ0 (t) must have the functional form
√
λ0 (t) = vt . Furthermore, as noted by Bayer-Friz-Gatheral (2015), the deterministic change of
measure implies that the log-normal distribution of vu under P is carried over to the log-normal
distribution under Q, which is at odds with positive skews observed in options on the VIX.
Finally, we highlight the ”loss of martingale” property for rough volatility models when return-
volatility correlation is positive, ρ > 0, in Eq (25), as shown in Gassiat (2019).

2.6 Log-normal volatility model

Karasinski and Sepp (2012) introduce following log-normal SV dynamics under statistical measure
P for the price St and volatility σt :
(0)
dSt = µt St dt + σt St dWt , S0 = S,
(0) (1)
(29)
dσt = κ (θ − σt ) dt + β dWt + εσt dWt , σ0 = σ,

where W (0) and W (1) are independent Brownian motions, κ and θ are the mean-reversion and
mean level of the volatility, respectively, and β is volatility beta which measures the change in the
volatility due to change in the price.
In Section 2.7 we study more general case of log-normal SV model with quadratic drift, where
we derive conditions when measures P and Q are equivalent and when the discounted price process
is a true martingale. For brevity, we only state the relevant results and refer to Section 2.7 for
detailed proofs.
We assume that statistical P and risk-neutral measures Q are related by
 Z t 
dQ (0) (1)
ζ(t) := =E − λ0 (u) dWu + λ1 (u) dWu (30)
dP Ft 0

where E(·) denotes Itô exponential, and λ0 (t) and λ1 (t) are equity and volatility risk-premia,
respectively. As for Exp-OU and Heston models in Eqs (7) and (15), we specify the following
functional form for λ0 (t) and λ1 (t)

λ0 (t) = λ̄0 σt , λ1 (t) = λ̄1 σt (31)

Electronic copy available at: https://ssrn.com/abstract=4647027

Proposition 2.5. Assume that κ ≥ 0, θ > 0. Then measures P and Q are equivalent, iff

β λ̄0 + ελ̄1 ≥ 0

Proof. Follows by setting κ2 = 0 in Proposition 2.7.

Proposition 2.6 (Martingale property). Assume that κ ≥ 0, θ > 0. Then discounted price process
is martingale under Q, iff
β≤0
Proof. Follows by setting κ2 = 0 in Proposition 2.8.

2.7 Log-normal SV model with quadratic drift

We extend log-normal SV model in (29) by adding a quadratic mean-reversion term to the drift of
the volatility, which plays a crucial role for the stability of the model. We consider the price and
volatility dynamics under statistical measure P as follows
(0)
dSt = µt St dt + σt St dWt ,
(0) (1)
(32)
dσt = (κ1 + κ2 σt )(θ − σt ) dt + βσt dWt + εσt dWt

where W (0) , W (1) are Brownian motions under P, κ2 is the quadratic mean-reversion rate and other
parameters are same as for (29).
We assume that statistical measure P and risk-neutral measure Q are related by
 Z t 
dQ (0) (1)
ζ(t) := =E − λ0 (u)dWu + λ1 (u)dWu (33)
dP Ft 0

where E(·) denotes Itô exponential, and λ0 (t) and λ1 (t) are equity and volatility risk-premias,
respectively, specified in Eq (31).
Assuming that process (33) is a true martingale, the following processes
Z t
(0) (0)
Ŵt = Wt + λ0 (u) du,
0
Z t
(1) (1)
Ŵt = Wt + λ1 (u) du
0

define independent Brownian motions under Q by Girsanov theorem. Then, the dynamics of price
process under risk-neutral measure Q becomes
(0)
dSt = (µt − λ0 (t)σt ) St dt + σt St dŴt

As discounted price process must be martingale under Q, we need to require that

µt − rt
µt − λ0 (t)σt = rt ⇐⇒ λ0 (t) = (34)
σt
Next, under martingale measures Q, the dynamics of volatility process will be
(0) (1)
dσt = κ1 θ − (κ1 − κ2 θ)σt − κ2 σt2 − βλ0 (t)σt − ελ1 (t)σt dt + βσt dŴt + εσt dŴt
 
(35)

where Ŵ (0) , Ŵ (1) are Brownian motions under Q.

Similar to Eq (31), we assume that market price of risk λ0 (·) and market price of volatility risk
λ1 (·) are proportional to the volatility:

λ0 (t) = λ̄0 σt , λ1 (t) = λ̄1 σt (36)

Electronic copy available at: https://ssrn.com/abstract=4647027

which allows us to rewrite Eq (35) as
(0) (1)
dσt = κ1 θ − (κ1 − κ2 θ)σt − (κ2 + β λ̄0 + ελ̄1 )σt2 dt + βσt dŴt + εσt dŴt
 
(37)

To retain the functional form of the volatility process, we rescale the parameters as follows

κ1 θ = κ̂1 θ̂,
κ1 − κ2 θ = κ̂1 − κ̂2 θ̂,
κ2 + λ̄0 + λ̄1 = κ̂2

Thus, under martingale measure Q, the dynamics of price and volatility processes will be
(0)
dSt = rt St dt + σt St dŴt ,
(0) (1)
(38)
dσt = (κ̂1 + κ̂2 σt )(θ̂ − σt ) dt + βσt dŴt + εσt dŴt

Proposition 2.7. We assume that κ1 ≥ 0, θ > 0. Then measures P and Q are equivalent, iff

κ2 ≥ max(−β λ̄0 − ελ̄1 , 0) (39)

Proof. We use measure-change argument and argue that measures P and Q are equivalent if the
volatility process under Q does not explode. This idea has been utilized already in Sin (1998),
Lewis (2000), Lucic (2004), Lions and Musiela (2007), among others. As we already derived the
dynamics of the volatility σt under Q in (37), using Sepp and Rakhmonov (2024a), we see that it
does not explode iff κ2 + β λ̄0 + ελ̄1 ≥ 0.

We specify the domain of model parameters to ensure that price process in (38) is well-behaved,
see Sepp and Rakhmonov (2024a) for further details.

Proposition 2.8 (Martingale property). Assume that κ̂1 ≥ 0, θ̂ > 0. Then discounted price
process is true martingale under Q, iff

κ̂2 ≥ max(β, 0) (40)

Proof. See Sepp and Rakhmonov (2024a).

As a result, we obtain that the functional form of log-normal SV model with quadratic drift is
invariant under different numéraires and the model produces price dynamics that are true martin-
gales as long as (40) holds when volatility beta is positive.

3 Empirics
We now discuss a few empirical parts related to using SV models in practice4 . We use the end-of-day
data of following assets.

1. S&P 500 index and its implied volatilities proxied with VIX index;

2. 10y US treasury rate and its implied volatilities proxied with MOVE index;

3. Oil futures (using USO ETF) and its implied volatilities proxied with OVX index5 .
4
Github project https://github.com/ArturSepp/StochVolModels provides Python code for data and the empir-
ical analysis in this section
5
Move and OVX volatility indices are constructed following the same methodology as for the VIX index.

10

Electronic copy available at: https://ssrn.com/abstract=4647027

 4. Bitcoin (denoted as BTC) and its implied at-the-money (ATM) volatility for options with
time-to-maturity of 7 days6 .

There are subtle differences when the volatility indices are constructed from the stochastic
volatility, but we believe the time series of the volatility indices are the best proxies for a model-
free dynamics of stochastic (implied) volatilities. In Figure 1, we show time series and empirical
PDFs of these implied volatilities.

Figure 1: Implied volatility indices: A) Time series from inception (set to 31Dec1999 for the VIX index), B)
Empirical PDFs

3.1 Auto-correlation
Auto-correlation of the volatility measures the “memory” or the longevity of periods with high
volatility. Indeed, rough volatility could allow for flexibility in modeling the empirical autocorrela-
tion.
We define the auto-correlation function (ACF) of the volatility path σt observed at regular
sampling times {tn } by
ACF (h) = Corr(σt+h , σt ), (41)
where h is the lag.
It is well-known that in one-factor Markovian SV models admit the exponential functional form
of the ACF :
ACF (h) = ce−qh , |h| → 0 (42)
where c > 0 and q > 0 are constants.
Bennedsen et al (2022) suggest that rough ACF, arising in rough SV models, follows the fol-
lowing asymptotic:
ACF (h) = 1 − c |h|2α+1 , |h| → 0, (43)
where c > 0 is a constant and α, α ∈ (−1/2, 1/2), is called the roughness index of σt . Negative
values of α imply auto-correlation rougher than that of a Brownian motion. For a given empirical
ACF, we estimate the values of c and α by minimizing squired differences empirical ACF and ACF
given in Eq (43).
6
We use historical options data of Deribit exchange with the data set starting on April 2019.

11

Electronic copy available at: https://ssrn.com/abstract=4647027

 To estimate the auto-correlation implied by log-normal SV model, we need to evaluate the
following equation using (41):

M odel ACF (h) = EP [Corr(σt+h , σt )] (44)

for a vector of lags h = (0, 1, ...).

We estimate the auto-correlation of log-normal SV model by simulation of paths σt using the
following algorithm.
Algorithm 1: Fitting parameters of Log-normal SV Model to empirical ACF and empirical
PDF of the logarithm of the volatility.
Data: Empirical ACF of the volatility, empirical PDF of log volatility
0. Simulate (M, N ) normals, where M is the number of time steps and N is the number of
paths, the simulation horizon is 10y with daily steps and M = 2600; the number of path
is set to 10000.
while the convergence between empirical and model ACFs is not reached do
1. Given optimiser’s suggested values of κ1 and κ2 , estimate θ and ε by minimizing the
difference between the empirical PDF of the logarithm of volatility and the model
produced PDF in Eq (46).
2. Simulate N paths of σt using (κ1 , κ2 , θ, ε) with σ0 = θ using pre-simulated (M, N )
normals.
3. Compute ACF for each path and the average across paths as the estimate for model
ACF, and compute the squared norm between empirical and model ACFs.
end
Output: Fitted model parameters (κ1 , κ2 , θ, ε)
In Figure (2) we show the empirical autocorrelation, the autocorrelation estimated using log-
normal SV model, and rough autocorrerlation in Eq (43) with fitted α and c for each time series
(it is remarkable that estimated α is about −0.22 for the three time series). In Table (1) we show
the fitted parameters of κ1 and κ2 for log-normal SV model.
We see that for the VIX index, the empirical auto-correlation behaves closer to the power low.
The ACF produced by the Log-normal SV model decays too slow for first 60 lags and too strong
for longer lags. This is know feature of one-factor SV models which produce exponentially decaying
ACFs. The fit to the empirical ACF can be improved by using a two factor SV model in which case
the model ACF is a linear combination of two exponential kernels. However the absolute deviation
of the one-factor Log-normal SV model does not exceed 0.1 so that the one-factor model is robust
enough for most practical purposes.
For the MOVE index, the empirical ACF is very strong suggesting the longevity of high volatility
regimes for interest rates. In this case, Log-normal SV and power kernel does not capture the
empirical ACF closely, but the absolute deviation does not exceed 0.1 as well.
For the OVX index, the empirical ACF decays fast up for the first 80 lags. The ACF produced
by the Log-normal SV model fits very well the empirical ACF for the first 80 lags. The rough ACF
is not flexible for OVX underestimating the empirical ACF for small lags and overestimating it for
longer lags.
For ATM implied volatilities of Bitcoin options, both Log-normal SV model and rough AFC
provide a close fit to the empirical ACF.
As a result, we can conclude that one-factor Log-normal SV model is sufficient enough for
reproducing the empirical features of ACF. We observe no material difference between the model
implied ACF and either the empirical or the fitted rough ACFs. Again, our analysis is focused on
daily data with relatively long horizons which is typical for most derivatives desks and prop trading
shops.

12

Electronic copy available at: https://ssrn.com/abstract=4647027

Figure 2: Auto-correlation of implied volatilities as function of lag periods (in days) for the three implies
volatility indices: A) VIX for the S&P 500 index, B) MOVE for the 10y UST rate, C) OVX for oil ETF,
D) Bitcoin (BTC) ATM implied volatilities for options with maturities of 7 day. Empirical is the empirical
estimate, Log-normal SV is the fitted auto-correlation of the lognormal SV model, Rough is the rough
auto-correlation with fitted α.

In Table (1) we show the fitted parameters of the log-normal SV model. The values of mean-
reversion speed κ1 and κ2 affect the decay of the auto-correlation: from low values for MOVE index
to high values for OVX index with slow and fast decay rate of the auto-correlation, respectively.

VIX MOVE OVX BTC

θ 0.20 0.91 0.39 0.71
κ1 1.29 0.10 2.78 2.21
κ2 1.93 0.41 2.24 2.18
ε 0.72 0.36 0.83 0.92

Table 1: Fitted parameters of steady-state distribution of volatility under Lognormal SV model.

Here we assume that β = 0 and ε is the total volatility of the volatility

3.2 Steady-State Distribution

The tails (in particular, the right tail) of the distribution of the volatility access whether the model
can produce extreme values of volatility observed in real data.
The steady state distribution of the volatility is given by the Generalized Inverse Gaussian
distribution (see Sepp and Rakhmonov (2024a)):
∞
n q o
Gσ (σ) = cσ η−1 exp − + bσ , σ > 0,
σ
κ1 θ κ2 κ2 θ − κ1 (b/q)η/2 (45)
q = 2 2 , b = 2 2, η = 2 − 1, c(b) = √
ϑ2


ϑ ϑ 2Kη 2 qb
where b > 0, q ≥ 0, η ∈ R and Kη is a modified Bessel function of the second kind. The distribution
of the logarithm of the volatility y = ln σ is given by:
∞ ∞
Gy (y) ≡ ey Gσ (ey ) = c exp ηy − qe−y + bey


(46)
We can estimate the model parameters using the empirical PDF of the logarithm of the observed
We notice that all thee coefficients q, b, and η are proportional to κ1 /ϑ2 and κ2 /ϑ2 so that mean-
reversion parameter κ1 and κ2 may not be identifiable from the steady-state distribution. Thus

13

Electronic copy available at: https://ssrn.com/abstract=4647027

estimate parameters of the steady-state distribution along with the mean-reversion parameters
using algorithm 1.
For Heston model, the stationary distribution of the variance is given by the gamma distribution
(see, for example, Lipton (2001))
V′
V∞ ′ (V ′ )β−1 e− α
G (V ) = , (47)
αβ Γ(β)

with scale parameter α = ε2 /(2κ), and shape parameter β = 2κθ/ε2 . Applying Eq (47) for V = e2y
we obtain that the distribution of the logarithm of the volatility is given by:
 
y∞ 2y V ∞ 2y 2 1 2y
G (y) ≡ 2e G (e ) = β exp 2βy − e . (48)
α Γ(β) α

We also add the normal distribution of logarithm of the volatility with the sample mean and
standard deviation. The normality of the steady-state distribution arises in exp-OU SV models.
Forde et al (2022) state that, under one-factor exp-OU or Bergomi-type model with rough Gamma
kernel (k(t) = tH−1/2 e−λt , where H is Hurst parameter and λ is relaxation parameter), the steady-
state distribution is log-normal too.
In Figure (3) we show the empirical PDF, the steady state PDF implied by fitted Log-normal
SV model using Eq (46), the PDF produced by fitted Heston model using Eq (48), and the normal
distribution with sampled mean and standard deviation. We observe that both Log-normal SV
and normal distribution fit the empirical PDF quite well apart very heavy tails observed for OVX
volatilities. In opposite, Heston model implies too light right tail because, in comparison to Log-
normal SV PDF in Eq (46) with the right tail decaying in ey , Heston PDF in Eq (48) implies a
much lighter left tail decaying in e2y .

14

Electronic copy available at: https://ssrn.com/abstract=4647027

 Figure 3: Steady-state PDF of the logarithm of the volatility (y-axis is shown in log-scale).

3.3 Volatility Beta

We address the estimation of the volatility beta which measures the linear dependence between
changes in the volatility and price returns. We argue that the estimator of the volatility beta is
not a constant but a time-inhomogeneous variable dynamically inferred from most recent data.
Given discrete time series, we use the following regression model

σt − σt−1 = βt rt + ϵt (49)

where rt is the return over period (t − 1, t], ϵt is noise. We estimate time series of βt using EWMA
regression model with span of 65 days (3 months).
In Figure (4) we show the time series of the volatility beta and the empirical PDF of time series
estimates. We see that, for the VIX index, the volatility beta is consistently negative with the
average of 1.19 and ranging from −2.0 to −0.5. For the rates volatility index MOVE, the volatility
beta is positive on average ranging from positive to negative values. For the oil volatility OVX
index, the volatility beta is negative on average but it takes positive values for short periods of
times (mostly when oil prices rise due to fears of supply disruption). For Bitcoin, the range of
volatility beta is symmetric with both large positive and negative realizations.

15

Electronic copy available at: https://ssrn.com/abstract=4647027

Figure 4: Volatility beta estimated using EWMA regression model in Eq (49) with span of 65 days: (A)
time series from inceptions, (B) Empirical PDF

When volatility beta and volatility-return correlation is positive, it may invalidate certain
volatility models because price dynamics become local martingales and may explode in finite time.
Sepp and Rakhmonov (2024a) show that the exp-OU model results in arbitrages when the volatility-
return correlation is positive.

3.4 Additional Considerations

3.4.1 Volatility Skew-stickiness Ratio
Recently, the so-called volatility skew-stickiness ratio have been utilized when analysis SV models,
in particular, rough SV models. In short, the volatility skew-stickiness measures the change in the
model implied volatility predicted by model implied volatility skew and price return. For example,
it is observed that the volatility skew-stickiness for the S&P500 index is 1.3 averaged over a long-
time. There has been an attempt to align the empirical behavior of the skew-stickiness (notably
rough SV model implies the skew-stickiness of 1.5 + H).
We treat with caution such attempts trying to link the model behavior under the both P and
Q measures. We argue that Q contain the risk-premia observed in market prices of derivative
securities due to risk-preferences and compensation for bearing hard-to-hedge risks for replication
of derivative securities. Sepp (2014) shows that empirically any skew-stickiness ratio (between 0
and 2) can be obtained by mixing stochastic volatility (with estimated volatility beta under P) and
a jump-feature interpreted as risk-premia to match implied market skews observed under Q.
As a result, in our opinion, there is at least one method to reproduce the empirical feature of the
volatility skew-stickiness using any type of one-factor SV model, so that the consistency with the
skew-stickiness is not necessary. The skew-stickiness can be modeled by incorporating additional
factors to SV model dynamics.

3.4.2 Volatility forecast

The forecasting of the volatility for horizon h includes the following problem:

f (σt , h) = EP [σt+h | Ft ] (50)

where Ft contains the information at time t.

16

Electronic copy available at: https://ssrn.com/abstract=4647027

 Also customary for pricing and hedging application is the forecast of the realized volatility:
 Z t+h 
P 1 2 ′
g(σt , h) = E σt′ dt | Ft (51)
h t

where Ft contains the information at time t.

We do not attempt to compare the accuracy of the volatility forecast using considered models,
because this is a broad topic with conclusions heavily dependent on the asset class, forecast period,
forecast penalty function etc. The considered volatility dynamics are typically used for valuation
applications with other features, such as our conditions, playing more important role.

4 Discussion
We make the following conclusions about our four principals applied to considered SV models.
Stein-Stein SV model (2.1) does not admit a valid change of measure. While it is still possible
to use this model by directly specifying it under either Q or P measures, the scope of the model is
limited. For example price numéraire (for FX and cryptocurrency derivatives) or annuity numéraire
(for interest rate derivatives) cannot be applied for this model.
Exp-OU SV model (2.2), Bergomi one-factor model (2.3), and log-normal volatility model with
linear drift (2.6) allow for change of measures, but the functional form of model dynamics changes
because of an additional term which arises in the drift of the volatility due to measure change. These
models do not admit strong martingale dynamics when return-volatility correlation is positive. In
our opinion, these models are originally designed for applications in equity derivatives and their
application to other asset classes is rather limited.
Rough SV model (2.5) is an extension of Exp-OU using the power kernel for Brownian driver in
the volatility dynamics. While rough volatility may provide good fit to empirical auto-correlation
function (ACF) as we show in Figure (2), the marginal improvement over a one-factor SV model is
rather low when using ACF fit metric (the absolute difference is less than 0.1). Rough OU-based SV
models inherit drawbacks of Exp-OU models: first, the difficulty in changing measures consistently
and, second, the lack of martingale property when return-volatility correlation is positive. In our
opinion, rough SV models are designed exclusively for equity markets and it may not be feasible to
apply them for other asset classes. On the implementation side, rough Exp-OU models can only
be implemented with MC methods.
Heston SV model (2.4) allows for consistent measure changes under different numéraires. The
model also produces true martingale dynamics when return-volatility correlation is positive and the
variance cannot hit zero as long as the Feller condition is satisfied. On the implementation side,
Heston model admits a closed-form solution for valuation of vanilla options (see Lipton (2001),
Lewis (2000)), which makes it easy for model calibration. These facts undoubtedly have made
Heston model applicable to multiple asset classes. However, numerical implementation of Heston
model using MC or numerical PDE methods is rather complicated, especially when Feller condition
is not satisfied. Furthermore, care must be taken to avoid making volatility process in Heston model
mean-repelling when return-volatility correlation is positive. There is a great deal of literature on
how to make Heston model work in practice.
Log-normal SV model with quadratic drift (2.7) allows for consistent measure changes using
different numéraires. For positive return-volatility correlation, the model produces true martingale
dynamics as long as the quadratic mean-reversion coefficient κ2 exceeds volatility beta. For model
calibration, Sepp and Rakhmonov (2024a) develop a closed-form and accurate solution for valuation
of vanilla options under this model. For numerical implementation using MC methods, Sepp and
Rakhmonov (2024a) develop a first-order MC scheme using the log-transform of the volatility to
unbounded domain. Since in log-coordinates the valuation problem in log-volatility is defined on
unrestricted domain, the problem can be solved efficiently using PDE methods for such domain

17

Electronic copy available at: https://ssrn.com/abstract=4647027

(see Sepp (2011)). As a result, log-normal SV model with quadratic drift can be considered as a
robust choice for modeling price dynamics for different asset classes.
For example of modeling volatilities of interest rates, see Sepp and Rakhmonov (2023) and Sepp
and Rakhmonov (2024b) for application of Heston/CIR SV and log-normal SV dynamics within
one-factor Cheyette model and in Multi-factor HJM framework, respectively.

References
Andersen, L. B. and Piterbarg, V. V. (2007). Moment explosions in stochastic volatility models.
Finance and Stochastics, 11(1), 29-50.

Bayer C., Friz P., and Gatheral J. (2015) Pricing under rough volatility. Quantitative Finance, Vol.
16, No. 6, 887–904

Bakshi, G. and Kapadia, N. (2003). Delta-hedged gains and the negative market volatility risk
premium. The Review of Financial Studies, 16(2), 527–566.

Bennedsen M., Lunde A., and Pakkanen M. (2022) Decoupling the Short- and Long-Term Behavior
of Stochastic Volatility. Journal of Financial Econometrics, Vol. 20, No. 5, 961–1006

Cont R. and Das P. (2022) Rough volatility: fact or artefact? Working paper, arxiv 2203.13820

Christoffersen, P., Jacobs, K. and Mimouni, K. (2010), ‘Models for s&p 500 dynamics: Evi-
dence from realized volatility, daily returns and options prices’, The Review of Financial Studies
23(9), 3141–3189.

Drimus, G. G. (2012). Options on realized variance in Log-OU models. Applied Mathematical

Finance, 19(5), 477-494.

Forde M., Gerhold S., and Smith B. (2022) Small-time VIX smile and the stationary distribution
for the Rough Heston model. Working paper

Gatheral J., Jaisson T., and Rosenbaum M. (2018) Volatility is rough. Quantitative Finance, Vol.
18, No. 6, 933–949

Gassiat, P. (2019). On the martingale property in the rough Bergomi model. Electron. Commun.
Probab. 24: 1-9.

Herdegen, M., & Schweizer, M. (2018). Semi-efficient valuations and put-call parity. Mathematical
Finance, 28(4), 1061-1106.

Heston, S. L. (1993). A closed-form solution for options with stochastic volatility with applications
to bond and currency options. The review of financial studies, 6(2), 327-343.

Karasinski, P., and Sepp, A. (2012). Beta stochastic volatility model, Risk Magazine, 66–71.

Lewis, A. L. (2000), Option Valuation under Stochastic Volatility, Finance Press.

Lions, P.-L. and Musiela, M. (2007), Correlations and bounds for stochastic volatility models,
24(1), 1–16.

Lucic, V. (2004). Forward-start options in stochastic volatility models. The best of Wilmott, 1,
413–420.

Lipton, A. (2001). Mathematical methods for foreign exchange: A financial engineer’s approach.
World Scientific.

18

Electronic copy available at: https://ssrn.com/abstract=4647027

Sepp, A. (2011) Efficient Numerical PDE Methods to Solve Calibration and Pricing Problems in
Local Stochastic Volatility Models, https://www.researchgate.net/publication/265379120

Sepp, A. (2014) Empirical Calibration and Minimum-Variance Delta Under Log-Normal Stochastic
Volatility Dynamics, https://ssrn.com/abstract=2387845

Sepp, A. and Rakhmonov, P. (2024) Log-normal Stochastic Volatility Model with

Quadratic Drift, forthcoming in International Journal of Theoretical and Applied Finance,
https://ssrn.com/abstract=2522425

Sepp, A. and Rakhmonov, P. (2023) ’A Robust Stochastic Volatility Model for Interest Rates’, Risk
Magazine, September 2023, 1–6. Available at https://ssrn.com/abstract=4315906.

Sepp, A. and Rakhmonov, P. (2024) Stochastic Volatility for Factor Heath-Jarrow-Morton Frame-
work. Available at https://ssrn.com/abstract=4646925

Sin, C. A. (1998), ‘Complications with stochastic volatility models’, Advances in Applied Probability
30(1), 256–268.

Wong, B., and Heyde, C. C. (2006). On changes of measure in stochastic volatility models. Inter-
national Journal of Stochastic Analysis, 2006.

19

Electronic copy available at: https://ssrn.com/abstract=4647027