Numerical approximation of the implied volatility

under arithmetic Brownian motion

Jaehyuk Choi, Kwangmoon Kim , and Minsuk Kwak
June 1, 2007

Abstract
We provide an accurate approximation method for inverting an option
price to the implied volatility under arithmetic Brownian motion. The
maximum error in the volatility is in the order of 1010 of the given option price and much smaller for the near-the-money options. Thus our
approximation can be used as a near-exact solution without further refinements of iterative methods.
Keyword : implied volatility, arithmetic Brownian motion, ABM, Bachelier, rational approximation, closed form approximation

Introduction

The French mathematician, Bachelier (1900) pioneered the study of financial

Mathematics by investigating option pricing in his doctorate thesis several
decades before Black and Scholes (1973) and Merton (1973) did. Under the
Bachelier model, the forward price of an underlying security Ft follows an arithmetic Brownian motion with a volatility ,
d Ft = d W t .

(1)

whereas, in the Black-Scholes-Merton model, the price follows a geometric Brownian motion. The forward prices of call and put options, maturing in time T
and struck at K with the current forward price F0 , are accordingly given by

(2)
C = E{max(FT K, 0)} = (F0 K) N (d) + T n(d)

P = E{max(K FT , 0)} = (K F0 ) N (d) + T n(d)

(3)
Department of Mathematics, Massachusetts Institute of Technology, 77 Massachusetts
Avenue, Cambridge, MA 02139, USA
Department of Mathematical Sciences, Korea Advanced Institute of Science and Tehonology, 335 Gwahangno, Yuseong-gu, Daejeon, Republic of Korea

Electronic copy available at: http://ssrn.com/abstract=990747

where n( ) and N ( ) are the probability density function and the cumulative
distribution function of the standard normal distribution respectively and d is
the moneyness measured in the units of the standard deviation,

d = (F0 K) / ( T ).
(4)
The implied volatility under arithmetic Brownian motion is the volatility that
produces the given option price from Eqns. (2) or (3). For the rest of the article,
this implied volatility will be referred as the normal implied volatility as opposed
to the lognormal implied volatility from the Black-Scholes-Merton model.
Although the Bachelier model is seemingly obsolete because it allows the underlying price to become negative, its use can not be neglected. For the pricing
of options whose underlying asset value is not necessarily positive, arithmetic
Brownian motion is more appropriate for the stochastic process than geometric
Brownian motion. Spread options actively traded in fixed income and commodity markets are such an example (Poitras, 1998; Carmona and Durrleman, 2003).
In those products, the normal implied volatility is naturally used to quote the
option price.
Even when the underlying price is positive, the normal implied volatility
provides greater insight than the lognormal implied volatility if the price process
is closer to an arithmetic Brownian motion. Interest rate derivatives are an
example of such asset classes. The downward sloping volatility skew observed
in the lognormal volatility space of the swaption market is much reduced in the
the normal volatility space. Thus traders often use the normal volatility for the
purpose of the in-house risk management.
No analytic expression is known for the normal or the lognormal implied
volatility. The calculation always depends on iterative root-finding methods
such as bisection or Newton-Raphson methods. With the computation powers
available these days, running such methods for daily risk management takes only
a fraction of a second. However, it is necessary to carefully fine-tune the solver
to ensure convergence for all possible ranges of parameters. Failure to perform
this task can lead to an unavailable hedge ratio, an important operational risk.
For the lognormal volatility Brenner and Subrahmanyam (1988); Chance
(1996); Chambers and Nawalkha (2001); Li (2007) have made several attempts
to find approximations. However, Brenner and Subrahmanyam (1988) only
dealt with the at-the-money case and Chance (1996); Chambers and Nawalkha
(2001); Li (2007) dealt with limited range of moneyness. Even within the proper
range restricted by Li (2007), the error is still large compared to the machine
precision, and thus additional refinements of iterative methods are still required.
To the best of authors knowledge, no such research has been reported on the
normal implied volatility. As it turns out, the task of inverting option prices is
much simpler in the Bachelier model than in the Black-Scholes-Merton model.
In Sec. 2, we show that the option pricing from the normal volatility is reduced
to a one-dimensional problem in a non-dimensional form. We then transform
variables so that they are better suited for approximation. In Sec. 3, we present
the approximation process in detail and the error of our best approximation. In
Sec. 4, we summarize with the actual coefficients.
2

Electronic copy available at: http://ssrn.com/abstract=990747

1
v(d)
2N(d)1
tanh((/2)1/2d)

1
3

0
d

Figure 1: v(d) (solid) compared with 2N (d)1 (dash) and tanh(

/2 d) (cross).

Non-dimensionalization and transformation

Without loss of generality we will only invert the price of the straddle, a combination of one call and one put option at the same strike, into the implied
volatility. Straddle is traded as much as call or put options in the market as
a bet on the volatility. If a call or put option is to be inverted, the price of
the straddle with the same strike can be obtained from the put-call parity. The
price of the straddle,

C +P = (F0 K) (2N (d) 1) + 2 T n(d),

(5)
has the attractive property that it is symmetric on the sign of the moneyness
F0 K. Thus, we only consider F0 K (d 0).
As d is non-dimensional, the natural non-dimensional form of Eq. (5) is
v=

F0 K
d
=
.
C +P
d(2N (d) 1) + 2n(d)

(6)

We define the variable v in such a way that it is bounded in the range [1, 1],
since the straddle price is always worth more than the intrinsic value |F0 K|.
Thus, finding for given C +P , F0 , K and T is now equivalent to finding the
inverse of v(d).
The shape of v(d) is similar to that of 2N (d) 1 (see Fig. 1); both functions
asymptotically approach 1 in the same manner as d increases:
log(1 v(d)) log(1 N (d)) d2 /2

as

d .

(7)

We could follow the steps used to obtain the inverse of the normal cumulative
function by Acklam (2004). However, the direct inversion of v(d) causes a slight
3

problem for at-the-money options. When F0 = K, the straddle price is given as

r
2T
C +P =

(8)

and the implied volatility is obvious. To ensure the continuity from the near-thep
money to the at-the-money options, it is required that limd0 d(v)/v = 2/.
As theories on the approximation are mostly on the value of the function, the
condition on the derivative could be difficult to impose.
Our alternative approach is to view the not-at-the-money case as a perturbation from the at-the-money case; we modify Eq. (8) as
r
r
2T
2 v
C +P =
or
h=
,
(9)
h
d
and approximate h as a function of only v. The implied volatility is then calculated as
r

(C +P ) h.
(10)
=
2T
The introduced perturbation h is bounded between [0, 1]. Now the continuity
condition, h = 1 at v = 0 (d = 0), becomes a part of the approximation of the
function value.
To better approximate h, it is necessary to properly transform v. As shown
in Fig. 2, the decay of h near v = 1 is extremely steep, thus making the direct
approximation, h = h(v), quite difficult. Under the following transformation,
=

v
2v
,
=
1
log((1
+
v)/(1
v))
tanh (v)

(11)

we find that the behavior of h as a function of is benign p

(see Fig. 3). The
intuition behind the transformation is that v(d) and tanh( /2 d) are very
similar (see Fig. 1) and the inverse of tanh( ) is analytically known. Therefore,
defined above is almost equal to h itself. The function h() has a diverging
derivative at = 0. However, the singularity
pcan be easily removed since we
know the leading asymptotic term, h() /(2), from Eqs. (7) and (9).

Therefore, we approximate h()/ (see Fig. 3) rather than h().

Numerical approximation and error

To find a good approximation of h()/ , we use a rational Chebyshev approximation, (Press et al., 1992)
m
n
X
X
h()
ak k /
bk k
g() =

k=0

(b0 = 1).

(12)

k=0

It is known that the rational function produces a better approximation than the
polynomial approximation with the same number of coefficients. The coefficients
4

1
h(v)
0.8

h(v)

0.6
0.4
0.2
0
0

0.2

0.4

0.6

0.8

Figure 2: The perturbation h as a function of v.

1
=0.049
0.8

h()

0.6
0.4
0.2

h()
1/2 h()

0
0

0.2

0.4

0.6

0.8

Figure 3: The perturbation h as a function of the transformation .

10

x 10

13

error

error

x 10

8
0.5

3
0

0.95 1

Figure 4: The error in h() from the rational approximation.

are searched with an iterative algorithm so as to minimize the maximum error,

max | h()/ g() | in the present case, in the original implementation of

Press et al. (1992). However, we modify the code so that the error in h(),

max | h() g() |, is minimized instead. We calculate the exact value of h

for a given up to the machine precision using bisection method.
It should be noted that it is sufficient to restrict the approximation domain
of to [0.049, 1]. When d is large, options are far-out-of-the-money. The
straddle price becomes very close to the intrinsic value |F0 K|, and thus v
and approach 1 and 0, respectively. The smallest value of occurs when
the straddle price is barely distinguishable from the intrinsic value, that is, at
v = 1 , where  is the machine epsilon. In double precision, it corresponds
to 2/ log(2/) 0.054 (d 7.7). Our lower bound, 0.049, is the limit
in approaching zero without compromising the accuracy of the approximation.
Since the singularity of h at = 0 is excluded from the domain, we could target

h() for approximation. However, we still use h()/ as it produces a better

result. It appears that the concave shape of h()/ is better suited for a
rational approximation with m > n. With this choice, we can also ensure that

h() 0 as 0 despite the error in h()/ for < 0.049.

The set of coefficients, ak and bk , yielding the best accuracy is reported in
the summary section below. We find the degrees, n = 7 and m = 9, provide
the best fit; extra degrees did not reduce the error significantly. Fig. 4 shows
the error of h. The maximum error is 3.4 1010 . If one reprices a portfolio
of straddles with the implied volatility from the approximation, the error in the
price,
h
(C +P ) =
(C +P ) 3.4 109 (C +P ).
(13)
h
6

is about $3 for a $1 billion portfolio. For > 0.5 (d < 1.46), where most practical
implied volatility calculations fall, the error is much smaller as 8.2 1013 (see
the inset of Fig. 4). For near-the-money options, > 0.95 (d < 0.32), the error
is 8.81015 . Thus, Eq. (8) is practically reduced for the at-the-money options.

summary

The implied volatility under arithmetic Brownian motion can be accurately

obtained from the following expression:
r

=
(C +P ) h(),
(14)
2T
where
=

(F0 K)/(C +P )
,
tanh1 ((F0 K)/(C +P ))

a0 = 3.99496 168734 5134 e-1

a1 = 2.10096 079506 8497 e+1
a2 = 4.98034 021785 5084 e+1
a3 = 5.98876 110269 0991 e+2
a4 = 1.84848 969543 7094 e+3
a5 = 6.10632 240786 7059 e+3
a6 = 2.49341 528534 9361 e+4
a7 = 1.26645 805134 8246 e+4

h() =

P7
ak k
Pk=0
,
9
k
k=0 bk

b0 = 1.00000 000000 0000 e+0

b1 = 4.99053 415358 9422 e+1

(15)
(16)

b2 = 3.09357 393674 3112 e+1

b3 = 1.49510 500831 0999 e+3
b4 = 1.32361 453789 9738 e+3
b5 = 1.59891 969767 9745 e+4
b6 = 2.39200 889172 0782 e+4
b7 = 3.60881 710837 5034 e+3
b8 = -2.06771 948640 0926 e+2
b9 = 1.17424 059930 6013 e+1

References
L. Bachelier. Theorie de la speculation. Annales Scientifiques de LEcole Normale Superieure, 17:2188, 1900.
F. Black and M. Scholes. The pricing of options and corporate liabilities. The
Journal of Political Economy, 81(3):637654, 1973.
R. C. Merton. The theory of rational option pricing. Bell Journal of Economics
and Management Science, 4:141183, 1973.
G. Poitras. Spread options, exchange options and arithmetic brownian motion.
Journal of Futures Markets, 18(5):487517, 1998.
R. Carmona and V. Durrleman. Pricing and hedging spread options. SIAM
Review, 45(4):627685, 2003.

M. Brenner and M. G. Subrahmanyam. A simple formula to compute the implied

standard deviation. Financial Analyst Jounal, 5:8083, 1988.
D. M. Chance. A generalized simple formula to compute the implied volatility.
The Financial Review, 31:859867, 1996.
D. R. Chambers and S. K. Nawalkha. An improved approach to computing
implied volatility. The Financial Reveiw, 38:89100, 2001.
M. Li. Approximate inversion of the black-scholes formula. European Journal
of Operational Research, forthcoming, 2007.
P. J. Acklam. An algorithm for computing the inverse normal cumulative distribution function. http://home.online.no/pjacklam/notes/invnorm/,
2004.
W. H. Press, S. A. Teukolsky, W.T. Vetterling, and Flannery B. P. Numerical
Recipes in C: The Art of Scientific Computings, chapter 5.13, pages 204208.
Cambridge University Press, 2nd ed. edition, 1992.