Vanna-Volga Method for Normal Volatilities

Volodymyr Perederiy*
October 2018, rev. November 2020

Abstract
Vanna-Volga is a popular method for the interpolation/extrapolation of volatility smiles. The technique
is widely used in the FX markets context, due to its ability to consistently construct the entire
Lognormal smile using only three Lognormal market quotes.

However, the derivation of the Vanna-Volga method itself is free of distributional assumptions. With
this is mind, it is surprising there have been no attempts to apply the method to Normal volatilities
(the current standard for interest rate markets).

We show how the method can be modified to build Normal volatility smiles. As it turns out, only minor
modifications are required compared to the Lognormal case. Moreover, as the inversion of Normal
volatilities from option prices is easier in the Normal case, the smile construction can occur at a
machine-precision level using analytical formulae, making the approximations via Taylor-series
unnecessary.

Apart from being based on practical and intuitive hedging arguments, the Vanna-Volga has further
important advantages. In comparison to the Normal SABR model, the Vanna-Volga can easily fit both
classical convex and atypical concave smiles (‘frowns’). Concave smile patterns are sometimes
observed around ATM strikes in the interest rate markets, particularly in the situations of anticipated
jumps (with an unclear outcome) in interest rates. Besides, concavity is often observed towards the
lower/left end of the Normal volatility smiles of interest rates. At least in these situations, the Vanna-
Volga can be expected to interpolate/extrapolate better than SABR.

Keywords: Vanna-Volga, Option Pricing, Normal Bachelier model, Option Greeks, Delta, Vega, Vanna,
Volga, Volatility Smiles, Volatility Frowns, SABR model, Interest Rates

* Volodymyr Perederiy, PhD (Europa Universität Viadrina, Frankfurt/Oder, Germany)

Email: v.perederiy@gmail.com

1
Introduction
Vanna-Volga (VV) is a popular and market-oriented technique for the modeling and interpolating/
extrapolating of volatility smiles. This technique has been mostly applied in the foreign exchange (FX)
markets. This can be explained by the fact that the technique allows for a consistent deduction of the
entire volatility smile given just three market quotes, and the FX markets happen to have three most
liquid standard quotes (so-called ATM, risk reversal and butterfly).

Besides, the FX focus of the technique seems to be based on the fact that FX is one of the markets
where Lognormal modeling is still deemed feasible. Although the final interpolation and approximation
formulae in the VV methodology (see Castagna and Mercurio [2007]) were indeed derived for the
Lognormal case, the methodology itself is free of distributional assumptions. It is surprising there has
been little or no theoretical nor empirical work on applying the VV method to the Normal (Bachelier)
volatilities. Recently, such models became particularly important for the modeling of interest rates,
given the arrival of the negative interest rates.

In this paper, we reiterate the derivation of the VV method, and obtain its results for the
Normal/Bachelier case using hedging arguments analogous to Castagna and Mercurio [2007]. In doing
so, we retort to forward (rather than spot) hedging instruments, as these are more natural in the
interest rate context. We investigate the volatility patterns generated by the technique and compare
them to those resulting from the classical Normal SABR method.

Bachelier vs. Black Model

Pricing Formulae
In the Normal/Bachelier option pricing model, the forward price of the underlying is assumed to follow
a driftless Brownian motion (in the T-forward measure):

𝑑𝐹𝑡 = 𝜎𝑑𝑊𝑡 (1)

where 𝜎 is the (constant) volatility. Then, the call option price with a strike 𝐾 and expiry 𝑇 can be
shown to equal:

𝐶 = 𝑃(0, 𝑇) [(𝐹 − 𝐾)Φ(𝑑) + 𝜎√𝑇𝜙(𝑑)]

𝐹−𝐾
𝑑= (2)
𝜎√𝑇

with 𝜙 signifying the density, and 𝛷 the cumulative distribution function of the normal distribution,
and 𝑃(0, 𝑇) being the appropriate discount factor. We use the term ‘moneyness’ for the above 𝑑 term.

This Bachelier formula can be compared to the classical Black 76 formula for the Lognormal case,
where the assumption of the underlying process is:

𝑑𝐹𝑡 = 𝜎𝐹𝑑𝑊𝑡 (3)

Here, the call option price is calculated as:

𝐶 = 𝑃(0, 𝑇) [𝐹Φ(𝑑1 ) − 𝐾Φ(𝑑2 )]

𝐹 𝜎2
ln 𝐾 + 2 𝑇
𝑑+ =
𝜎√𝑇 (4)
𝐹 𝜎2
ln − 𝑇
𝑑− = 𝐾 2
𝜎√𝑇
2
with the two moneyness terms 𝑑+ and 𝑑− .

Option Greeks
We now derive the relevant Greeks (option sensitivities) for the Bachelier case, which will be required
later1. The Delta and Vega can be shown (see e.g. Frankena [2016]) to equal:

(Forward) Delta:
𝜕𝐶
Deltaf = = 𝑃(0, 𝑇)Φ(𝑑) (5)
𝜕𝐹
Vega:
𝜕𝐶
Υ= = 𝑃(0, 𝑇) √𝑇𝜙(𝑑) (6)
𝜕𝜎
From this, the relevant second-order derivatives can be derived as follows:

Volga:

First, we have:

𝜕 2 𝐶 𝜕Υ 𝜕𝜙(𝑑)
= = 𝑃(0, 𝑇) √𝑇
𝜕𝜎 2 𝜕𝜎 𝜕𝜎
and, making use of the derivative of normal density 𝑑𝜙(𝑥)/𝑑𝑥 = −𝑥𝜙(𝑥) and applying the chain rule,
we arrive at:

𝜕2𝐶 𝐹 − 𝐾 −2 𝑑2 𝑑2
= 𝑃(0, 𝑇) √𝑇 [(−𝑑𝜙(𝑑)) (− 𝜎 )] = 𝑃(0, 𝑇) √𝑇𝜙 (𝑑) = Υ (7)
𝜕𝜎 2 √𝑇 𝜎 𝜎
(Forward) Vanna:

Analogously:

𝜕2𝐶 𝜕Υ 1 −𝑑𝜙(𝑑) 𝑑
= = 𝑃(0, 𝑇) √𝑇 [(−𝑑𝜙(𝑑)) ( )] = 𝑃(0, 𝑇) = −Υ (8)
𝜕𝐹𝑑𝜎 𝜕𝐹 𝜎√𝑇 𝜎 √𝑇𝜎
(Forward) Gamma:

𝜕 2 𝐶 𝜕Deltaf 𝑃(0, 𝑇)𝜙(𝑑) Υ

2
= = = (9)
𝜕𝐹 𝜕𝐹 𝜎√𝑇 𝜎𝑇

We now compare the Normal Greeks derived above to the Lognormal Greeks from the classical Black
76 model2 (second-order derivatives are expressed in terms of Vega Υ):

1
The Greeks derived here are for the forward (and not spot) price F. Corresponding Delta, Vanna and Gamma
Greeks related to the spot (instead of forward) S can be easily obtained via multiplying the forward-related
𝜕F 𝜕(𝑆𝑒 𝑐𝑇 )
Greeks with the partial derivative (with repeated multiplication for Gamma) = = 𝑒 𝑐𝑇 and
𝜕𝑆 𝜕𝑆
substituting 𝐹 = 𝑆𝑒 𝑐𝑇 , where 𝑐 stands for the cost of carry.
2
See Xiong (2016) and Frankena (2016) for the derivation of the Lognormal (Black 76) Greeks.
3
 Greeks Normal (Bachelier) Lognormal (Black 76)
Moneyness 𝐹−𝐾 𝐹 𝜎2
𝑑= ln 𝐾 + 2 𝑇
𝜎√𝑇 𝑑+ =
𝜎√𝑇
𝐹 𝜎2
ln − 𝑇
𝑑− = 𝐾 2
𝜎√𝑇
𝜕𝐶 𝑃(0, 𝑇)Φ(𝑑) 𝑃(0, 𝑇)Φ(𝑑+ )
Delta, 𝜕𝐹
𝜕𝐶
Vega, 𝜕𝜎 Υ = 𝑃(0, 𝑇) √𝑇 𝜙(𝑑) Υ = 𝑃(0, 𝑇)𝐹 √𝑇 𝜙(𝑑+ )
𝜕2 𝐶 Υ Υ
Gamma, 𝜕𝐹2
𝜎𝑇 𝐹 2 𝜎𝑇
𝜕2 𝐶
Volga, 𝜕𝜎2 𝑑2 𝑑+ 𝑑−
Υ Υ
𝜎 𝜎
𝜕2 𝐶 𝑑 𝑑−
Vanna, −Υ −Υ
𝜕𝐹𝑑𝜎
√𝑇𝜎 𝐹√𝑇𝜎

There are obvious similarities between the Normal and Lognormal Greeks. In particular, when
expressed in terms of the corresponding moneyness (as defined above), the expressions are actually
identical apart from the inclusion of the forward price in the Lognormal case. Also, given a fixed Vega,
the Vanna Greek is in both cases linearly proportional to the moneyness, and the Volga – to the
squared moneyness.

Vanna-Volga for Normal/Bachelier Volatilities

Portfolio Construction
As in the case of the Lognormal VV method (see Castagna and Mercurio [2007]), we assume that the
option price can be described with a flat (strike-independent) but stochastic implied volatility.
Analogously, we construct a portfolio Π consisting of:

- one long position in the call option under consideration with strike 𝐾0 and value 𝐶0 (according
to the Bachelier model)
- a short position Δ in the delta-hedging asset 𝐻 and
- three short positions w1 , w2 , w3 in some pivot/benchmark call options with strikes 𝐾1 , 𝐾2 , 𝐾3
and values 𝐶1 , 𝐶2 , 𝐶3 (according to the Bachelier model).

Over a small time interval, by application of Ito’s lemma to the function 𝐶(𝐹, 𝜎, 𝑡), we can derive the
following changes in the options’ values (for i=0,1,2,3), based on a change in the forward price 𝑑𝐹 and
a change in the implied volatility 𝑑𝜎:
𝜕𝐶𝑖 𝜕𝐶𝑖 𝜕𝐶𝑖
𝑑Ci = 𝑑𝐹 + 𝑑𝑡 + 𝑑𝜎
𝜕𝐹 𝜕𝑡 𝜕𝜎
2 2 2
1 𝜕 𝐶𝑖 𝜕 𝐶𝑖 𝜕 𝐶𝑖
+ ( 2 (𝑑𝐹)2 + (𝑑𝜎) 2
+ (𝑑𝑡)2 ) (10)
2 𝜕𝐹 𝜕𝜎 2 𝜕𝑡 2
𝜕 2 𝐶𝑖 𝜕 2 𝐶𝑖 𝜕 2 𝐶𝑖
+( (𝑑𝐹𝑑𝑡) + (𝑑𝜎𝑑𝑡) + (𝑑𝐹𝑑𝜎))
𝜕𝐹𝜕𝑡 𝜕𝜎𝜕𝑡 𝜕𝐹𝜕𝜎
Now, because 𝑑𝐹 = 𝜎𝑑𝑊 in the Bachelier case, it holds:

(𝑑𝐹)2 = 𝜎 2 (𝑑𝑊)2

4
So, making use of the following rules of stochastic calculus:

𝑑𝜎𝑑𝑡 = 0

(𝑑𝑡)2 = 0

(𝑑𝑊)2 = 𝑑𝑡
𝑑𝑊𝑑𝑡 = 0
𝑑𝐹𝑑𝑡 = 𝜎𝑑𝑊𝑑𝑡 = 0
the expression (10) can be simplified to:
𝜕𝐶𝑖 𝜕𝐶𝑖 𝜕𝐶𝑖
𝑑Ci = 𝑑𝐹 + 𝑑𝑡 + 𝑑𝜎
𝜕𝐹 𝜕𝑡 𝜕𝜎
1 𝜕 2 𝐶𝑖 𝜕 2 𝐶𝑖
+ ( 2 𝜎 2 𝑑𝑡 + (𝑑𝜎)2 )
2 𝜕𝐹 𝜕𝜎 2

𝜕 2 𝐶𝑖
+( (𝑑𝐹𝑑𝜎))
𝜕𝐹𝜕𝜎

or:

𝜕𝐶𝑖 𝜕𝐶𝑖 1 𝜕 2 𝐶𝑖 2 𝜕𝐶𝑖 1 𝜕 2 𝐶𝑖 2

𝜕 2 𝐶𝑖
𝑑Ci = 𝑑𝐹 + ( + 𝜎 ) 𝑑𝑡 + 𝑑𝜎 + (𝑑𝜎) + (𝑑𝐹𝑑𝜎) (11)
𝜕𝐹 𝜕𝑡 2 𝜕𝐹 2 𝜕𝜎 2 𝜕𝜎 2 𝜕𝐹𝜕𝜎

As for the delta-hedging asset, analogously by Ito’s lemma, but now assuming independence between
the price of this asset and volatility, we have:

𝜕𝐻 𝜕𝐻 1 𝜕 2 𝐻 2
𝑑𝐻 = 𝑑𝐹 + ( + 𝜎 ) 𝑑𝑡 (12)
𝜕𝐹 𝜕𝑡 2 𝜕𝐹 2

The change in the overall portfolio value over a small time interval is then:
3
𝑑Π = 𝑑𝐶0 − Δ dH − ∑ 𝑤𝑖 𝑑𝐶𝑖 (13)
𝑖=1
where 𝑑𝐶0 , 𝑑𝐶1 , 𝑑𝐶2 , 𝑑𝐶3 are defined according to (11) and dH is defined according to (12).

So far, the derivation has been effectively assumption-free, apart from using the Normal relationship
(𝑑𝐹)2 = 𝜎 2 (𝑑𝑊)2 , which only effects the second drift term in front of 𝑑𝑡 in (11) and (12), resulting
1 𝜕 2 𝐶𝑖 2 1 𝜕2 𝐻 2
here in 2 𝜕𝐹 2
𝜎 and 2 𝜕𝐹 2
𝜎 correspondingly. In the Lognormal case, the relationship is (𝑑𝐹)2 =
1 𝜕 2 𝐶𝑖 2 2
𝜎 2 𝐹 2 (𝑑𝑊) , leading therefore to the same results with the terms changed to
2
𝜎 𝐹 and
2 𝜕𝐹 2
1 𝜕2 𝐻 2 2
𝜎 𝐹 .
2 𝜕𝐹 2

Risk Elimination
From (13), and using the Normal Greeks derived in the previous section, we can calculate the
quantities 𝑤1, 𝑤2 and 𝑤3 and Δ such that this makes 𝑑Π insensitive to changes in the forward price 𝑑𝐹
and to changes in volatility 𝑑𝜎.

5
The Delta risk of the portfolio (stemming from the terms in front of 𝑑𝐹) can be eliminated, for arbitrary
weights 𝑤𝑖 , via setting the amount of the delta-hedging asset Δ such that
𝜕𝐶0 𝜕𝐻 3 𝜕𝐶𝑖
−Δ − ∑ 𝑤𝑖 =0 (14)
𝜕𝐹 𝜕𝐹 𝑖=1 𝜕𝐹
The Vega risk of the portfolio (stemming from terms in front of 𝑑𝜎) can be eliminated via enforcing
the following restriction on the weights 𝑤𝑖 :
𝜕𝐶0 3 𝜕𝐶𝑖 3
− ∑ 𝑤𝑖 = Υ0 − ∑ 𝑤𝑖 Υi = 0
𝜕𝜎 𝑖=1 𝜕𝜎 𝑖=1

or:
3
Υ0 = ∑ 𝑤𝑖 Υi (15)
𝑖=1
where Υi signifies the Vega Greek of the i-th option according to the Normal/Bachelier model.

The Vanna risk of the portfolio (stemming from the terms in front of 𝑑𝐹𝑑𝜎) can be eliminated via
enforcing the following restriction on the weights 𝑤𝑖 :

𝜕 2 𝐶0 3 𝜕 2 𝐶𝑖 𝑑0 3 𝑑𝑖
− ∑ 𝑤𝑖 = −Υ0 + ∑ 𝑤𝑖 Υi =0
𝜕𝐹𝜕𝜎 𝑖=1 𝜕𝐹𝜕𝜎 √𝑇 𝜎 𝑖=1 √𝑇 𝜎
which is equivalent to
3
Υ0 𝑑0 = ∑ 𝑤𝑖 Υi 𝑑𝑖 (16)
𝑖=1
where 𝑑𝑖 signifies the moneyness of the i-th option according to the Normal/Bachelier model.

The Volga risk of the portfolio (stemming from the terms in front of (𝑑𝜎)2 ) can be eliminated via
enforcing the following restriction on the weights 𝑤𝑖 :

𝜕 2 𝐶0 3 𝜕 2 𝐶𝑖 𝑑02 3 𝑑𝑖2
− ∑ 𝑤𝑖 = Υ0 − ∑ 𝑤𝑖 Υi =0
𝜕𝜎 2 𝑖=1 𝜕𝜎 2 𝜎 𝑖=1 𝜎
which is equivalent to
3
Υ0 𝑑02 = ∑ 𝑤𝑖 Υi 𝑑𝑖2 (17)
𝑖=1

The drift term in front of 𝑑𝑡 for the portfolio is:

𝜕𝐻 1 2 𝜕 2 𝐻 𝜕𝐶0 3 𝜕𝐶𝑖 1 𝜕 2 𝐶0 2 3 1 𝜕 2 𝐶0 2
−Δ ( − 𝜎 ) + − ∑ 𝑤𝑖 + 𝜎 − ∑ 𝑤𝑖 𝜎
𝜕𝑡 2 𝜕𝐹 2 𝜕𝑡 𝑖=1 𝜕𝑡 2 𝜕𝐹 2 𝑖=1 2 𝜕𝐹 2

𝜕𝐻 1 2 𝜕 2 𝐻 𝜕𝐶0 3 𝜕𝐶𝑖 1 Υ0 2 3 Υi 2
= −Δ ( − 𝜎 2 )+ − ∑ 𝑤𝑖 + 𝜎 − ∑ 𝑤𝑖 𝜎
𝜕𝑡 2 𝜕𝐹 𝜕𝑡 𝑖=1 𝜕𝑡 2 𝜎𝑇 𝑖=1 𝜎𝑇

𝜕𝐻 1 2 𝜕 2 𝐻 𝜕𝐶0 3 𝜕𝐶𝑖 1 𝜎 2 3
= −Δ ( − 𝜎 ) + − ∑ 𝑤𝑖 + (Υ − ∑ 𝑤𝑖 Υi )
𝜕𝑡 2 𝜕𝐹 2 𝜕𝑡 𝑖=1 𝜕𝑡 2 𝜎𝑇 0 𝑖=1

1
Because of the already implied restriction Υ0 = ∑3𝑖=1 𝑤𝑖 Υi , the term after the last 2 also vanishes.

6
Thus, provided all constraints are imposed and fulfilled as described above, the change in portfolio
value 𝑑Π becomes (instantaneously) deterministic and free of the risks related to the underlying
(forward) price and its volatility:

𝜕𝐻 1 2 𝜕 2 𝐻 𝜕𝐶0 3 𝜕𝐶𝑖
𝑑Π = (−Δ ( − 𝜎 2 ) + − ∑ 𝑤𝑖 ) 𝑑𝑡 (18)
𝜕𝑡 2 𝜕𝐹 𝜕𝑡 𝑖=1 𝜕𝑡

We summarize the weight constraints as a system of linear (in 𝑤𝑖 ) equations:

Υ1 𝑤1 + Υ2 𝑤2 + Υ3 𝑤3 = Υ0
Υ1 𝑤1 𝑑12 + Υ2 𝑤2 𝑑22 + Υ3 𝑤3 𝑑32 = Υ0 𝑑02
(19)
Υ1 𝑤1 𝑑1 + Υ2 𝑤2 𝑑2 + Υ3 𝑤3 𝑑3 = Υ0 𝑑0

This system is, again, remarkably similar to that in the Lognormal case, with just both Lognormal
moneyness terms 𝑑+ and 𝑑− replaced by the Normal moneyness 𝑑.

The solution is analogously, by Cramer’s rule:

Υ0 Υ2 Υ3
|Υ0 𝑑0 Υ2 𝑑2 Υ3 𝑑 3 |
Υ0 𝑑02 Υ2 𝑑22 Υ3 𝑑32
w1 = Υ1 Υ2 Υ3
|Υ1 𝑑1 Υ2 𝑑2 Υ3 𝑑 3 |
Υ1 𝑑12 Υ2 𝑑22 Υ3 𝑑32

Υ0 Υ2 Υ3 (𝑑2 − 𝑑0 )(𝑑3 − 𝑑0 )(𝑑3 − 𝑑2 )

w1 =
Υ1 Υ2 Υ3 (𝑑2 − 𝑑1 )(𝑑3 − 𝑑1 )(𝑑3 − 𝑑2 )

which results in:

Υ0 (𝑑2 − 𝑑0 )(𝑑3 − 𝑑0 )
w1 = (20)
Υ1 (𝑑2 − 𝑑1 )(𝑑3 − 𝑑1 )

and, using the definition of the Normal moneyness:

Υ0 (𝐾2 − 𝐾0 )(𝐾3 − 𝐾0 )
𝑤1 = (21)
Υ1 (𝐾2 − 𝐾1 )(𝐾3 − 𝐾1 )

Again, this solution is very similar to the solution for the Lognormal case, where just the logarithms of
the strikes appear instead of the strikes in (21) (see Castagna and Mercurio [2007]).

The solutions for w2 , w3 are derived analogously, and are in the Bachelier case:

Υ0 (𝐾1 − 𝐾0 )(𝐾3 − 𝐾0 )
𝑤2 =
Υ1 (𝐾1 − 𝐾2 )(𝐾3 − 𝐾2 )
(22)
Υ0 (𝐾1 − 𝐾0 )(𝐾2 − 𝐾0 )
𝑤3 =
Υ1 (𝐾1 − 𝐾3 )(𝐾2 − 𝐾3 )

Choice of Delta-Hedging Instrument

Currently, the Normal/Bachelier model is mostly used for option pricing in the interest rate markets,
with the underlying here being an interest rates index. As two examples: for caplets/floorlets the
7
underlying is an IBOR interest rate, for swaptions - a swap rate. There is no spot market for such
interest rates indices. The rates are not directly tradable assets and thus cannot serve as the delta-
hedging instrument H.
𝜕𝐻
However, there are tradable assets 𝐻 without vega risk (𝜕𝜎 = 0) with their price linked to these rates
and which can thus be used for the delta hedge. These are typically forward-like instruments whose
current fair value is determined as:

𝐻𝑡 = 𝑃(𝑡, 𝑇)(𝐹𝑡,𝑇 − 𝑋)

where 𝐹𝑡,𝑇 is the forward price/rate at time t for a delivery in T and 𝑋 is the strike (of the forward
instrument). An example of such instrument is e.g. forward rate agreement (FRA), as the delta hedge
for a caplet/floorlet option.

Generally, such forward-like instruments can be used as follows. By Ito’s lemma, we have:

𝜕𝐻 𝜕𝐻 1 2 𝜕 2 𝐻
𝑑𝐻 = 𝑑𝐹 + ( + 𝜎 ) 𝑑𝑡
𝜕𝐹 𝜕𝑡 2 𝜕𝐹 2

𝜕𝑃(𝑡, 𝑇)
𝑑𝐻 = 𝑃(𝑡, 𝑇) 𝑑𝐹 + ( (𝐹𝑡,𝑇 − 𝑋) + 0) 𝑑𝑡
𝜕𝑡

Choosing a forward contract with zero fair value (i.e. with 𝑋 = 𝐹𝑡,𝑇 ) as the delta-hedging instrument,
we obtain:

𝑑𝐻 = 𝑃(𝑡, 𝑇) 𝑑𝐹
which can be substituted into (13) to obtain the quantity Δ of the delta-hedging instrument needed
for the risk elimination.

Smile Construction
So far, we have dealt with pricing in a world with a constant Bachelier/Normal volatility. However, the
already mentioned value Π of the resulting weighted portfolio in this world can be shown (see
Shkolnikov [2009]) to be equal to the market value Πmkt of the portfolio in the real world, with a
stochastic (but strike-independent) implied volatility. As the value of the delta-hedging instrument H
does not depend on volatility, we have:
3 3

𝐶0,𝑚𝑘𝑡 − ∑ 𝑤𝑖 𝐶𝑖,𝑚𝑘𝑡 = 𝐶0 − ∑ 𝑤𝑖 𝐶𝑖 (23)

𝑖−1 𝑖−1
The above relationship between the option market prices 𝐶0,𝑚𝑘𝑡 and 𝐶𝑖,𝑚𝑘𝑡 can be transformed into
a relationship between Normal volatilities 𝜎0 , 𝜎𝑖 implied by these prices. After rearranging
3

𝐶0,𝑚𝑘𝑡 − 𝐶0 = ∑ 𝑤𝑖 (𝐶𝑖,𝑚𝑘𝑡 − 𝐶𝑖 ) (24)

𝑖−1
the differences in market vs. theoretical prices (𝐶0,𝑚𝑘𝑡 − 𝐶0 and 𝐶𝑖,𝑚𝑘𝑡 − 𝐶𝑖 ) can be approximated
based on the differences between the implied market volatilities and the reference flat volatility (𝜎0 −
𝜎 and 𝜎𝑖 − 𝜎 ) using Taylor expansions. The first-order approximation (with derivation analogous to
that in Castagna and Mercurio [2007] for the Lognormal case) results in the following expression for
the implied market volatility of the option with a strike 𝐾0 :

𝜎0 (𝐾0 ) ≈ 𝑦𝑖 𝜎1 + 𝑦2 𝜎2 + 𝑦3 𝜎3 (25)
with (for Bachelier model):
8
 (𝐾2 − 𝐾0 )(𝐾3 − 𝐾0 )
𝑦1 =
(𝐾2 − 𝐾1 )(𝐾3 − 𝐾1 )
(𝐾1 − 𝐾0 )(𝐾3 − 𝐾0 )
𝑦2 =
(𝐾1 − 𝐾2 )(𝐾3 − 𝐾2 )
(𝐾1 − 𝐾0 )(𝐾2 − 𝐾0 )
𝑦3 =
(𝐾1 − 𝐾3 )(𝐾2 − 𝐾3 )
and 𝑦1 + 𝑦2 + 𝑦3 = 1

Obviously, 𝜎0 (𝐾0 ) is a quadratic function of 𝐾0 which intersects the points (𝐾1 , 𝜎1 ), (𝐾2 , 𝜎2 ) and
(𝐾3 , 𝜎3 ).

The more precise second-order approximation results in the following quadratic equation:

𝑑0 2 𝑄
(𝜎0 − 𝜎)2 + (𝜎0 − 𝜎) − (𝑃 + ) ≈ 0
2𝜎 2𝜎
with:
3

𝑄 = ∑ 𝑦𝑖 𝑑𝑖 2 (𝜎𝑖 − 𝜎)2
𝑖−1
3

𝑃 = −𝜎 + ∑ 𝑦𝑖 𝜎𝑖
𝑖−1

This quadratic equation can be easily solved for 𝜎0 − 𝜎, which results in3:

−𝜎 + √𝜎 2 + 𝑑0 2 (2𝜎𝑃 + 𝑄)
𝜎0 (𝐾0 ) ≈ 𝜎 +
𝑑0 2
with:
𝐹 − 𝐾0
𝑑0 =
𝜎 √𝑇 (26)
3

𝑄 = ∑ 𝑦𝑖 𝑑𝑖 2 (𝜎𝑖 − 𝜎)2
𝑖−1
3

𝑃 = −𝜎 + ∑ 𝑦𝑖 𝜎𝑖
𝑖−1

Again, these solutions for Bachelier/Normal case are actually identical to the Lognormal case (as in
Castagna and Mercurio [2007]), with strike prices in the former replacing the log-strikes in the latter,
and moneyness terms redefined.

That said, the above Taylor-series approximations are actually not needed for the Bachelier/Normal
model. Instead, the market price of the option with a strike price 𝐾0 can be calculated directly as:

3
For the particular case 𝐹 = 𝐾0 (or 𝑑0 = 0), it immediately follows from the quadratic equation: 𝜎0 =
∑3 2
𝑖−1 𝑦𝑖 𝑑𝑖 (𝜎𝑖 −𝜎)
2
∑3𝑖−1 𝑦𝑖 𝜎𝑖 +
2𝜎

9
 3

𝐶0,𝑚𝑘𝑡 = 𝐶0 + ∑ 𝑤𝑖 (𝐶𝑖,𝑚𝑘𝑡 − 𝐶𝑖 ) (27)

𝑖−1

and the corresponding Normal volatility 𝜎0 (𝐾0 ) can be simply backed out from this market price. This
inversion is significantly easier for the Normal volatilities, with several analytical methods delivering
machine precision results. E.g. one such method was proposed in Choi et al. [2007] and proceeds for a
call price 𝐶0,𝑚𝑘𝑡 and the corresponding put price4 𝑃0,𝑚𝑘𝑡 (with the same expiry T and strike K) as
follows5:
𝜋 (28)
𝜎0 (𝐾0 ) = √ (𝐶 + 𝑃0,𝑚𝑘𝑡 )ℎ(𝜂)
2𝑇 0,𝑚𝑘𝑡
(𝐹 − 𝐾0 )/(𝐶0,𝑚𝑘𝑡 + 𝑃0,𝑚𝑘𝑡 )
𝜂=
tanh−1 ((𝐹 − 𝐾0 )/(𝐶0,𝑚𝑘𝑡 + 𝑃0,𝑚𝑘𝑡 ))
∑7𝑘=0 𝑎𝑘 𝜂 𝑘
ℎ(𝜂) = √𝑛 9
∑𝑘=0 𝑏𝑘 𝜂 𝑘

a0 =3.994961687345134 e-1 b0 =1.000000000000000 e+0

a1 =2.100960795068497 e+1 b1 =4.990534153589422 e+1
a2 =4.980340217855084 e+1 b2 =3.093573936743112 e+1
a3 =5.988761102690991e+2 b3 =1.495105008310999 e+3
a4 =1.848489695437094 e+3 b4 =1.323614537899738 e+3
a5 =6.106322407867059 e+3 b5 =1.598919697679745 e+4
a6 =2.493415285349361 e+4 b6 =2.392008891720782 e+4
a7 =1.266458051348246 e+4 b7 =3.608817108375034 e+3
b8 =-2.067719486400926 e+2
b9 =1.174240599306013 e+1

Choice of the Reference Volatility

The application of the VV method requires (for the second-order approximation in (26) and the exact
inversion as in (28)) the input of the reference (flat) volatility 𝜎. Normally, either the ATM-implied
volatility or the mid implied pivot volatility 𝜎2 (with 𝜎1 < 𝜎2 < 𝜎3) are used. There are, however, no
solid theoretical underpinnings which would advocate such usage. This can be regarded as the main
drawback of the Vanna-Volga method.

Instead, the reference volatility can be treated as yet another degree of freedom. In particular, for a
reference volatility given, the standard VV methodology fits the smile to three arbitrary market pivotal
quotes. Treating the reference volatility as yet another parameter would normally result in the VV
methodology being able to fit a smile to an additional fourth market quote.

4
The put price 𝑃0,𝑚𝑘𝑡 can be derived from the call price 𝐶0,𝑚𝑘𝑡 using the put-call parity 𝐶0,𝑚𝑘𝑡 − 𝑃0,𝑚𝑘𝑡 =
𝑃(0, 𝑇)(𝐹 − 𝐾0 ).
5
In the particular case of ATM (for F=K), the implied Normal volatility can be inverted directly from (2) as
2 1
𝜎𝐴𝑇𝑀,𝑚𝑘𝑡 = Φ−1 ( (𝐶𝐴𝑇𝑀 , 𝑚𝑘𝑡 / 𝐹 + 1))
√𝑇 2

10
Comparison to Normal SABR
Normal SABR Model
During the last decade, the SABR model has established itself as the market standard for the
interpolation/extrapolation of interest rate volatility smiles. The technique is based on the forward
rates 𝐹𝑡 modeled with stochastic instantaneous volatility 𝛼𝑡 (see Hagan et al [2002] for details):

𝑑𝐹𝑡 = 𝛼𝑡 𝐹𝑡 𝛽 𝑑𝑊𝑡
𝑑𝛼𝑡 = 𝜈 𝛼𝑡 𝑑𝑍𝑡 (29)
𝑑𝑊𝑡 𝑑𝑍𝑡 = 𝜌𝑑𝑡
with four parameters

0 ≤ 𝛽 ≤ 1, 𝜈 ≥ 0, −1 < 𝜌 < 1 and 𝛼0 > 0.

Setting 𝛽 = 0 results in the so-called Normal SABR model, and allows modeling of negative forwards,
similarly to the Normal/Bachelier framework. However, the SABR volatility dynamics is always
assumed to be Lognormal. This differentiates this method from the VV technique, which does not make
any assumptions as to the dynamics of the stochastic volatility.

The Normal implied volatility smile resulting from the Normal SABR model can be approximated as6:

𝜁 2 − 3𝜌2 2
𝜎0 (𝐾0 ) = 𝛼0 (1 + 𝜈 𝑇)
𝑥(𝜁) 24
𝜈
𝜁 = (𝐹 − 𝐾0 ) (30)
𝛼
√1 − 2𝜌𝜁 + 𝜁 2 − 𝜌 + 𝜁
𝑥(𝜁) = log ( )
1−𝜌

Generally, the 𝛼0 parameter controls the height of the smile, 𝜈 - its deepness, and the 𝜌 parameter -
its skewness/asymmetry. Having three free parameters, the SABR smile in most situations can fit three
market quotes exactly (similarly to the VV technique) and can then be used for inter- and extrapolation.
However, due to the model design (in particular, the Lognormal volatility dynamics), the smiles
resulting from SABR are always convex (or flat in the extreme case of 𝜈 = 0). Thus, the method fails
in the case of concave smiles (so-called ‘frowns’, where the maximum implied volatility is near ATM)
which sometimes occur. As VV is free of assumptions as to the dynamics of the stochastic volatility, it
can also be expected to deal well with such situations.

Practical Examples
The examples below compare the smiles fitted via Normal VV vs. Normal SABR for some constellations
of market data.

The first example refers to a classical situation with three pivots where ITM/OTM volatilities exceed
the ATM volatility, with slight skewness. The VV smile is fitted using the exact inversion method (see
(28)).

6
See Frankena [2016]. The original derivation in Hagan et al [2002] contains a typo in the relevant section
‘A5.Special Cases’ for the particular case of the Normal implied volatility for Normal SABR.
11
Smile construction with Normal VV (with alternative reference volatilities of 40, 45, 50, 55, and 60) vs Normal SABR. For 𝐹 =
0, 𝑇 = 1, 𝑃(0, 𝑇) = 0, 𝑎𝑛𝑑 pivots 𝜎1 (−50) = 51, 𝜎2 (0) = 50, 𝜎3 (50) = 52.

As seen in this example, the patterns in the interpolation range are actually identical for VV vs. SABR,
and also do not depend significantly on the VV reference volatility. The extrapolation patterns are quite
different, however. In particular, the VV smiles show the “wings” on both sides, the smile becoming
concave on the edges. The SABR smile remains convex, becoming almost linear on the edges.
Furthermore, the VV smile extrapolation does depend significantly on the VV reference volatility: the
higher it is, the higher the edges are. The closest match between the VV and SABR smile seems to occur
with a VV reference volatility slightly exceeding the implied ATM volatility.

The strong dependence of the VV smile on the reference volatility advocates using a fourth pivotal
quote, so that an exact fit is achieved for four pivots. In the graphical example above, this fourth quote
is indicated by a circle at the strike 100, resulting in the selection of the optimal reference volatility
equal to 55. Generally, inferring the optimal VV reference volatility via a fit to a fourth market quote
can be achieved via simple numerical methods. Generally, such a four-pivot fit is impossible with SABR.

We now regard a less classical situation of an inverted smile (‘frown’), where the ITM/OTM volatilities
fall below the ATM volatility.

12
Smile construction with Normal VV (with alternative reference volatilities of 40, 50, and 60) vs Normal SABR. For 𝐹 = 0, 𝑇 =
1, 𝑃(0, 𝑇) = 0, 𝑎𝑛𝑑 pivots 𝜎1 (−50) = 48, 𝜎2 (0) = 50, 𝜎3 (50) = 49

The best SABR fit is linear, as SABR is generally unable to fit concave smiles. The VV method
interpolates the concave smile (‘frown’) well. However, the robustness of the VV extrapolation seems
to depend on the reference volatility. For (relatively) high reference volatilities, the method fails on
the edges. Technically, in these cases the calculated VV call prices (see (27)) fall below the intrinsic
values of the options. For the reference volatilities well below the implied ATM volatility, the VV
extrapolation seems to remain robust and consistent on the edges.

As a final check, we verify if the concave smiles fitted by VV produce positive risk-neutral probabilities.
To this end, the density of the risk-neutral distribution of the price x of the underlying at expiry can be
calculated from option prices using the second derivative of the option price with respect to the strike7
(see Breeden and Litzenberger [1978]):

𝜕2
𝑓(𝑥) = 𝑃(0, 𝑇) 𝐶𝑎𝑙𝑙(𝐾) (31)
𝜕𝐾 2

We estimate this density via a discrete approximation using triplets of calls with strikes 𝑥 − 𝛿, 𝑥, 𝑥 +
𝛿 (for a small 𝛿) as:
𝐶𝑚𝑘𝑡 (𝑥 + 𝛿) + 𝐶𝑚𝑘𝑡 (𝑥 − 𝛿) − 2𝐶𝑚𝑘𝑡 (𝑥)
𝑓(𝑥) = 𝑃(0, 𝑇)
𝛿2 (32)

where 𝐶𝑚𝑘𝑡 refers to the VV result (27).

The figures below show the densities for some selected concave volatility smiles (frowns).

7 𝜕2
Alternatively, for a cumulative distribution function 𝐹(𝑥) and put prices: 𝐹(𝑥) = 𝑃(0, 𝑇) 𝑃𝑢𝑡(𝐾)
𝜕𝐾 2
13
Risk-neutral densities with Normal VV (for reference volatility=40) vs with Normal SABR. For 𝐹 = 0, 𝑇 = 1, 𝑃(0, 𝑇) = 0, 𝑎𝑛𝑑
pivots 𝜎1 (−50) = 48, 𝜎2 (0) = 50, 𝜎3 (50) = 49

Risk-neutral densities with Normal VV (for reference volatility=30) vs with Normal SABR. For 𝐹 = 0, 𝑇 = 1, 𝑃(0, 𝑇) = 0, 𝑎𝑛𝑑
pivots 𝜎1 (−50) = 45, 𝜎2 (0) = 50, 𝜎3 (50) = 45

It becomes clear that even in the case of concave smiles (“frowns”) the VV technique can produce
consistent non-negative risk-neutral densities. Moreover, with a frown deep enough, the density
becomes bimodal, in line with expectations for situations with uncertain but significant jumps in the
underlying price happening in the near future.

14
Summary
We have shown that the Vanna-Volga technique is applicable for the Normal/Bachelier volatilities, with
results quite similar to the classical Lognormal case, apart from different expressions for the option
moneyness and for the Vega-related Greeks. The first-order Vanna-Volga approximation of the implied
Normal volatility smile turns out to be just a quadratic function of the strike price. The more precise
second-order approximation can also be easily derived for the Normal case and is similar to Lognormal.
However, quasi-exact analytical solutions for the volatility smile can be easily derived in the
Normal/Bachelier case due to the ease of the quasi-exact (machine-precision) analytical inference of
implied volatilities from option prices, thus making the Taylor approximations unnecessary.

One of the drawbacks of the Vanna-Volga technique is its dependence on the reference volatility as
one of the inputs. This volatility is generally unknown, but, as we show, can significantly influence the
shape of the extrapolated smile. In practice, the ATM-implied volatility or a middle volatility is often
used for this input, as a matter of intuition rather than logical reasoning. Instead, we propose treating
the reference volatility as yet another degree of freedom. With the reference volatility becoming a
model parameter, the Vanna-Volga methodology is generally able to fit a smile to four arbitrary market
quotes.

The Vanna-Volga technique has some other important advantages. We compare the method to the
Normal SABR model, which is the current interpolation/extrapolation standard in the interest rate
context. The Vanna-Volga method seems to perform well in fitting the smile patterns typically
observed in the interest rate markets. In contrast to SABR, the technique is able to fit the inverted
(concave) smile patterns (or corresponding bimodal implied density patterns). These patterns are
sometimes observed before important jump events (such as revisions of interest rates, with option
maturities shortly thereafter). Besides, concavity is often observed towards the lower/left end of the
Normal volatility smiles of interest rates, probably reflecting the strongly diminishing probability of
interest rates becoming too negative. For these situations, the Normal Vanna-Volga technique can be
expected to perform better than SABR.

References
Breeden, D. T. and Litzenberger, R. H. (1978). Prices of State-Contingent Claims Implicit in Option
Prices, Journal of Business 51(4): pp. 621-651

Castagna, A. and Mercurio, F. (2007). The Vanna-Volga Method for Implied Volatilities. Risk, Jan 2007,
pp. 106-111

Choi J., Kim K., and Kwak M. (2007). Numerical Approximation of the Implied Volatility under
Arithmetic Brownian motion, June 1, 2007

Frankena, L.H. (2016). Pricing and Hedging Options in a Negative Interest Rate Environment. Thesis,
Delft University of Technology, Amsterdam, February 29, 2016

Hagan P., Kumar D., Lesniewski A., and Woodward D. (2002). Managing Smile Risk. Wilmott, 1(8), pp.
84-108, 2002

Shkolnikov, Y. (2009). Generalized Vanna-Volga Method and Its Applications, NumeriX Quantitative
Research, June 25, 2009

Xiong C. (2011). Vanna-Volga Method for Foreign Exchange Implied Volatility Smile, June 27, 2016

15