15418286, 2014, 73, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/wilm.10366 by Cochrane Netherlands, Wiley Online Library on [23/11/2022].

See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Modeling Volatility and Valuing Derivatives
Under Anchoring
Paul Wilmott
Wilmott Associates, e-mail: [email protected]
Alan L. Lewis
OptionCity.net
Daniel J. Duffy
Dataism Education BV
Abstract not being stock-price dependent. This model has the scaling property that if we
We develop a complete-markets model with volatility smiles, tractability, and intui- multiply S by any constant then dS is increased by the same factor and so the return
tive appeal as an anchoring or habit-formation model. Like traditional stochastic dS/S remains unchanged. Immediately this means that models such as CEV and local
volatility models, it is invariant to a multiplicative scaling of the stock price levels. volatility, where the volatility depends on the level of S, are, according to our one
The anchoring effect is that the volatility depends on the relative value of the current modeling clue, unacceptable.
stock price compared to its past history, with an exponential weighting. Rather than simply dismiss models such as CEV and local volatility, we should
ask why they are popular. The answer is that they can give theoretical option values
Keywords that more closely match those option values seen in the market, in particular in terms
anchoring, behavioral finance, skew, volatility modeling of implied volatility skews and smiles. Whether these skews and smiles ought to be
an output of a theoretical model, via a process such as calibration, is open to question
(Wilmott, 2009). Nevertheless it is common experience, regardless of option values
1 Introduction and purely in terms of the behavior of a stock, that if a stock falls significantly from a
In the subject of quantitative finance there are few clues as to how to model the vari- previous high then its volatility does seem to increase. And this does point at a vola-
ables from first principles; how shares, commodities, interest rates, credit risk, and tility that depends on the stock level S.
other instruments ought to behave. In the language of stochastic calculus this means Is there any way that we can reconcile these two apparently conflicting require-
that the functional forms of the drift and volatility in any stochastic differential equa- ments? Can we have a volatility that depends on S in a model that still has our scaling
tion are almost arbitrary. Take interest rates for example, what we want of our model requirement? The answer is yes.1
is that rates should stay positive (even this is open to question), and neither blow up
nor get absorbed. Apart from that, virtually any interest rate model is acceptable. 1.1 Anchoring
Thus, there is rarely any peg on which to hang our modeling hat, so to speak. According to Wikipedia, anchoring is a cognitive bias that describes the common
There is one exception to this observation. And that exception is that there is a human tendency to rely too heavily on one trait or piece of information when mak-
class of financial instruments that ought to be modeled by a process that has certain ing decisions. In the present context we use the word to refer to the natural tendency
lognormal or geometric characteristics. With shares and exchange rates we are not to compare the current level of a stock to some historical level or average. Of course,
concerned about the price level per se, rather we care about its new level compared we may be misusing the word since there may be no bias at all, such a tendency might
with a previous one. It is the return that matters. Clearly this is different from inter- be perfectly rational. Yet it does fly in the face of almost all quantitative models of
est rates or credit risk, where the level makes a big difference to our perception of the financial instruments.
instrument. If we have $1000 to spend, we are not concerned about the share price, In models of financial instruments it is extremely common, almost universal
we just have to buy more or fewer shares. in fact, that information from the past is given less weight than ‘information’ from
This suggests that a good model for a share price S would be the stochastic differ- the future via option values, and implied volatilities. Note the use of inverted com-
ential equation dS = mSdt + sSdX with s being perhaps time dependent, but certainly mas around the second ‘information.’ Obviously no one has a crystal ball. But there
seems to be a preference for using quantitative numbers that anticipate what might
Copyright © 2014 by Paul Wilmott, Alan L. Lewis, and Daniel J. Duffy happen in the future rather than learning from the past. These models even have a

48 WILMOTT magazine

48-56_Wilm_Duff y_TP_Sept_2014_Final.indd 48 09/02/2014 02:4 PM

 15418286, 2014, 73, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/wilm.10366 by Cochrane Netherlands, Wiley Online Library on [23/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
TECHNICAL PAPER

mathematical name, Markov. Yet share prices are governed by human actions and Figure 1: The time series for detrended log SPX from 1950 until March 2012.
emotions (perhaps not at the millisecond time scale!) and those humans do draw on
the past for clues to the future.

2 The model
We are going to work with a model that has a memory. Let us introduce a new vari-
able, A, representing averaging or, in the language of behavioral finance, anchoring
(to an historical level) (see Tversky and Kahneman, 1974):
t
A=𝜆 e−𝜆(t−𝜏 )S(𝜏 ) d𝜏 . (1)
∫∞
Note that this is just about the simplest memory function that has tractable properties.
We now work with a volatility function s(x), a function to be determined, where
S
𝜉= . (2)
A
Of course, if you multiply S everywhere by a constant then the variable x is
unchanged. Thus the model has our scaling requirement.
Our model is thus
dS = 𝜇S dt + 𝜎(𝜉 )S dX. (3)
Such a volatility function can be made consistent with the previously mentioned There may possibly be a minimum of s(x) if mid-values of x are thought of as
observation that volatility increases when prices are low, but crucially this means being more stable than extreme values. (And this in turn may be seen in a resulting
low relative to some historical average, not simply low in an absolute sense since that theoretical smile.)
would be contrary to our scaling requirement. We shall do most of the analytical development for general s(x) and then use a
Note this is not the only memory function that has the required property; simple sigmoidal function when we discuss explicit examples.
anything of the form
( t ( ) ) 2.4 Qualitative observation on share prices generally
S(t)
f g , t, 𝜏 d𝜏 Figure 1 shows two lines, one being the logarithm of Standard and Poor’s 500 index
∫−∞ S(𝜏 ) (SPX) from 1950 until March 2012, detrended, and the second being the inverse of
would also be acceptable, but a general functional would be just too cumbersome for the index, also detrended and shifted vertically. Essentially the latter is just the mirror
practical use. image of the former. If we were dealing with simple, lognormal random walks then
We now explain the parameters in equations (1), (2), and (3). the choppy (that is Brownian Motion) nature of these two lines would be qualitatively
similar. But there is one aspect of the real financial time series that is different from
2.1 m its mirror image: the artificial data, the mirror image, has peaks and troughs like a sea
In principle, we can similarly model m as a functional. This could give extremely rich wave, the peaks being pointy and the troughs being rounded. With the real data this
dynamics to the S process. However, provided the standard technical requirements picture is upside down. The real data has pointed troughs and rounded crests. This
are met this is irrelevant for the valuation of derivatives, since our model is clearly is suggestive of higher volatility for lower relative share prices; relative since we have
complete and so allows dynamic hedging and thus risk-neutral valuation. When detrended the series in this picture. Again, this is consistent with our model.
we come to determine the parameters in our model from historical time series
(not calibration), we will see whether the simplest m being a constant is sufficient to 3 Option valuation equation
represent the statistics of the underlying. Since
dA = 𝜆(S − A) dt (4)
2.2 l
The parameter l is a measure of the extent of memory, the larger it is the shorter the the equation governing the value of anything but strongly path-dependent options is
span of memory. We would guess that in the financial markets (relevant) human obviously2
1 (S) 2
memory would be measured in months or a small number of years, not in weeks or
decades. And by ‘memory’ here we really mean that part of a time series that is deemed Vt + 𝜎 2 S VSS + rSVS − rV + 𝜆(S − A)VA = 0. (5)
2 A
relevant to the stock’s current performance, that is allows for ‘forgiveness’ as it were.
This will have the usual final condition, at time t = T the expiration, representing the
2.3 s payoff. The direction of convection makes boundary conditions in A unnecessary.
For a simple stock we would expect the sigma function to be larger for small values of
x than for large values of x. (And this may result in a theoretical negative skew.) The 4 Observations on x
function s(x), being a power of x, might be appealing since it mimics the CEV model By examining the stochastic differential equation for x in isolation we can tell a great
(although ours is more sensible, having the desirable scaling property). We will deal about the behavior of this model, and get some clues as to how to approach the
^
discuss this case later, where we will see that it results in undesirable properties for x. determination of a functional form for s and for determining a value for l.

WILMOTT magazine 49

48-56_Wilm_Duff y_TP_Sept_2014_Final.indd 49 09/02/2014 02:4 PM

 15418286, 2014, 73, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/wilm.10366 by Cochrane Netherlands, Wiley Online Library on [23/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
From Equations (1), (2), and (3) we have 6 Determining P¥(x) from the time series
d𝜉 = (𝜇 + 𝜆 − 𝜆𝜉 ) 𝜉 dt + 𝜎(𝜉 )𝜉 dX. By finding the probability density function for the data in Figure 1, we find the func-
tion P∞(x). See Figure 5. However, from equation (6) we have a relationship between
We shall find it very useful to work with the Fokker–Planck Equation to relate the
the s(x) we have just found and the P∞(x) we have also just found. So we could for
density function for x to the volatility function s(x). In the steady state (assuming it
example take the found s(x) and use it to determine P∞(x). By integrating (6) we have
exists) we can relate these via the ordinary differential equation
( 𝜉 )
1 d2 ( ) d ( ) 1 𝜇 + 𝜆 − 𝜆s
𝜎(𝜉 )2𝜉 2P∞ (𝜉 ) = D(𝜉 )𝜉 P∞(𝜉 ) , (6) P∞ (𝜉 ) = 2 2 exp 2 ds + C . (7)
2 d𝜉 2 d𝜉 𝜉 𝜎 ∫ s𝜎(s)2

with P∞(x) being the steady-state distribution and D(x) = m + l – lx. Given any two
out of the three functions s(x), m, and P∞(x), we can use this equation to determine Figure 3: s(x) from time series data.
the third. That is, assuming the model is correct. sigma (from dX)
From historical data for S, and hence for A, assuming that we know l, we can 0.8 sigma
determine both s(x) and P∞(x). (The former by differencing the x time series as 0.7
explained in Chapter 36 of Wilmott, 2006 and the latter by looking at the distribution
0.6
of the time series data, as explained in the same chapter. These ideas were first intro-
duced in Apabhai et al., 1995.) And since we are assuming m to be a constant, again 0.5
easily determined from the time series if it assumed constant, Equation (6) can be 0.4
used to check whether the model is internally consistent and robust. 0.3
0.2
5 Determining s(x) from the time series 0.1
As already mentioned, we take daily prices for the Standard and Poor’s 500 index
0
from 1950 until March 2012. First we estimate m, assuming it to be constant. This is 0 0.5 1 1.5 2
ξ
not done in a particularly sophisticated way, just treating the asset price as the con-
stant volatility model. It is found to be 0.0835. Next we wish to analyze the data for
Figure 4: A sigmoidal function fitted to the data.
x. Choose a l that seems plausible – we shall shortly be performing an optimization
that properly determines its value – and calculate a time series for A, using Equation sigma (from dX)
(4) and a starting guess for A. Thus we have a time series for x. See Figure 2. Sigmoidal function
0.8
Difference this to get a time series dx, which is virtually equivalent to a time sigma
series for s(x)x dX since the time between data points is so small. From this we can 0.7
estimate the function s(x) as explained in Wilmott (2006). The results are shown in 0.6
Figure 3. (Note that this shows the s(x) after the (still to be described) optimization 0.5
has been done, so in effect this is our model.)
0.4
This already shows the realistic, and expected, behavior in which the volatility is
large for smaller x. Because we wish to manipulate this function we need to first fit a 0.3
relatively simple function to it. We chose the sigmoidal function 0.2
b−a 0.1
a+ .
1 + e−c(ln (𝜉)−d) 0
The reason for this choice is the behavior for small and large x, where it asymptotes to 0 0.5 1 1.5 ξ 2
two different volatility values. Figure 4 shows the original data and the fitted curve.
(Again, this fitted curve is the one that is optimal as discussed shortly.)
Figure 5: The steady-state PDF for x from time series data.
Figure 2: The time series for x, SPX from 1950 until March 2012. 6 Probability Density Function, steady state

S/I
1.6 5
1.4
4 From empirical data
1.2
1 3
0.8
0.6 2

0.4
1
0.2
ξ
0 0
4/12/1949 9/15/1965 2/18/1982 7/24/1998 12/27/2014 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

50 WILMOTT magazine

48-56_Wilm_Duff y_TP_Sept_2014_Final.indd 50 09/02/2014 02:4 PM

 15418286, 2014, 73, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/wilm.10366 by Cochrane Netherlands, Wiley Online Library on [23/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
TECHNICAL PAPER

Figure 6: Best fit integrated steady-state PDF and empirical PDF. 8 Numerical results for option values
6 Probability Density Function, steady state We report results from two numerical solvers for the option valuation PDE. Both use
Mathematica’s NDSolve.4 The state space for (S, A) is the (unbounded) positive first
5 quadrant. For the first solver, we map this to the unit square (X, Y ) using the trans-
From empirical data formations X = S/(S + S0) and Y = A/(A + A0), where (S0, A0) are the initial ‘hot spot’
4
From differential equation values. With that, let V(S,A,t) = Ke–rt h(X,Y,t), where t = T – t and K is the option
3 strike price. Then equation (5) becomes ht = AhXX + BhX + ChY, with new (variable)
PDE coefficients that are straightforward to develop. As noted earlier, boundary con-
2 ditions are unnecessary at Y = 0 and Y = 1. Even if you are not a Mathematica user,
our solver call for a put option should be fairly readable. It simply repeats the PDE,
1
sets the initial/boundary conditions, and then specifies the method:
ξ
0 NDSolve[{D[h[X,Y,t],t] == A[X,Y] D[h[X,Y,t],{X,2}]
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
+ B[X,Y] D[h[X,Y,t],X] + C[X,Y] D[h[X,Y,t],Y],
h[X,Y,0] == Max[1-(S0/K)X/(1-X),0], h[Xmin,Y,t] == 1,
h[Xmax,Y,t] == 0}, h, {X,Xmin,Xmax},{Y,Ymin,Ymax},{t,T,T},
This can be done numerically. The constant of integration C is chosen to ensure that Method->{"MethodOfLines","SpatialDiscretization"->
the integral of the PDF over zero to infinity is one.3 {"TensorProductGrid","Coordinates"->{xgrid,ygrid}}}]
Let us recap. We have picked a plausible value for l, which is by no means
necessarily correct. This gives us a time series for x and hence a (fitted sigmoidal) The (X, Y ) spatial grid here is a uniform discretization in both directions: {i/N,
function for s(x) and a steady-state PDF. The latter two are also related via equa- j/N}, (i,j = 1,2,…, N–1), N an integer. Table 1 shows results on the implied volatility
tion (7). There are five parameters in our model, l and four (a, b, c, d) for the sig- surface for various strikes with S0 = A0 = 100. The associated put option values and
moidal s(x) function. We also have the parameter m estimated directly from the S parameter set are found in Table 2. Note that the parameter set is different from the
time series. earlier example. The results are based on spatial grids of size 250 × 250, except for
T = 100, which uses a mesh of size 200 × 200. Entries for T = 0 come from a separate
theoretical analysis of the model’s asymptotics, as discussed below.
7 The optimization As a second test of the numerical results we discuss how to approximate the
We choose l and a, b, c, d to minimize a combined, weighted error by finding the solution of the PDE (5) using the finite difference method (FDM) and subsequent
best fits for both:

1. the sigmoidal function to the empirical s function and Table 1: Anchoring model: Black–Scholes implied volatility (%)
2. the steady-state PDF obtained from integrating the sigmoidal function to the T K=60 K=80 K=100 K=120 K=140
empirical mean and standard deviation of the steady-state distribution.
0 29.46 21.17 17.13 15.99 15.61
The results of the best-fit sigmoidal function have already been shown in Figure 4. 0.25 29.03 21.15 17.18 16.07 15.76
The best-fit steady-state PDF is shown in Figure 6. 1 27.78 21.07 17.50 16.29 15.78
The best-fit parameters are 3 25.68 20.91 18.15 16.85 16.22
a = 0.60, b = 0.12, c = 10.82, d = −0.27 and 𝜆 = 0.90. 10 24.04 21.62 19.92 18.77 17.97
100 25.60 25.16 24.82 24.54 24.31
The last of these confirms a typical memory of the order of 1 year, as anticipated.

Table 2: European-style option prices for Table 1 (p=put, c=call)

7.1 A note on the goodness of fit
T K=60 K=80 K=100 K=120 K=140
We can fit the s function as accurately as we like, by choosing suitably flexible func-
tions with sufficient degrees of freedom. But from that function follows the steady- 0.25 p 0.0004298 0.04264 2.832 18.593 38.296
state distribution P∞. There is no guarantee that this distribution should be anywhere c 40.731 21.017 4.049 0.0540 0.00003977
near matching the empirically calculated distribution given that we have allowed 1 p 0.1700 0.8486 4.681 16.280 33.551
no freedom whatsoever in the parameter m. Had we given ourselves the freedom c 43.039 24.674 9.463 2.019 0.2460
to choose m to be a function of x or something even more general, then such a good 3 p 0.9670 2.379 6.035 13.695 25.444
fit would probably have been possible. But this was not guaranteed with the simple c 49.169 33.315 19.706 10.099 4.583
choice of m as a constant, and especially not with m being the constant given directly 10 p 1.928 3.534 5.988 9.558 14.397
by the time series data. Considering this, the match of the two functions to the data is c 65.171 54.524 44.725 36.043 28.629
remarkable. 100 p 0.05594 0.08064 0.1073 0.1358 0.1658
As a test of the robustness of the model we input different values for m and per-
c 99.609 99.485 99.363 99.242 99.123
formed the same optimization. The fits were in no case anywhere near as impressive
^
as when we used the historical value for m. Parameter set: a = 0.6, b = 0.15, c = 10, d = –0.3, r = 0.049, l = 0.5.

WILMOTT magazine 51

48-56_Wilm_Duff y_TP_Sept_2014_Final.indd 51 09/02/2014 02:4 PM

 15418286, 2014, 73, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/wilm.10366 by Cochrane Netherlands, Wiley Online Library on [23/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
implementation in C++. The PDE (5) is defined on the positive quarter-plane in Consider general bivariate stochastic volatility models of the form
S and A. In order to reduce it to a problem on a bounded domain, we can choose ̃ S dW (1)
dS = (r − q) S dt + 𝜎(y)
between the popular domain truncation and domain transformation techniques. We
choose the latter approach, as already introduced in this section. We then transform dy = b(y) dt + a(y) dW (2) ,
equation (5) in the variables (S, A) to a convection–diffusion–reaction PDE in the (1) (2)
where dW √ dW = r(y)dt. For example, the Heston model and many others fit, with
variables (X, Y). In other words, we solve a transformed PDE on the unit square. In 𝜎(y)
̃ = y , so that y is the variance rate. Our anchoring model fits with y = x = S/A,
order to specify the full problem we define an initial condition (in this case the pay- s̃(y) = s(x), and r = 1, a degeneracy which is not a problem. Note that s̃(·) is a general
off function) and boundary conditions. For the X (transformed underlying) vari- function, while s(·) is the sigmoidal function. For simplicity, assume y is positive.
able we prescribe well-known boundary conditions for call and put options while In traditional stochastic volatility models, for example (Heston, etc.), St transfor-
no boundary conditions are needed for the transformed anchoring variable Y, as mations do not affect yt, and one has scaling: invariance under St → kSt. Under the
already mentioned. anchoring model, you need to transform the entire historical series: {St} → {kSt}. This
In order to approximate the solution of the transformed PDE we have chosen is translational invariance for {xt} = {log St}. Under both types
[ of scaling one] has this
to use the Alternating Direction Explicit (ADE) method (for the original source see key property: the marginal characteristic function Φ = E eiz(xt −x0 ) |x0 , y0 is condi-
Saul’yev, 1964 and for applications to computational finance see Duffy, 2009). There tionally independent of x0, the conditioning being on y0.5 Thus, Φ = Φ (t, y0, z). This
are two main ADE variants: first, the original Saul’yev scheme that is unconditionally leads to the Fourier pricing formula for call options:
stable and explicit and second, the Barakat–Clark (1966) scheme that is also uncon-
H (T, y0, z) dz
ditionally stable and explicit but more accurate than the Saul’yev scheme. The main C(T, S0 , y0 ; K) = S0 e−qT − Ke−rT e−izm . (8)
∫ℑz∈(0,1) z2 − iz 2𝜋
advantages of the ADE method are: it is fast, easy to program, and can be applied
to a wide range of linear and nonlinear time-dependent PDEs (some examples are Here, m ≡ log{S0e–qt/Ke–rt} is a natural measure of ‘moneyness’ and
discussed in Pealat and Duffy, 2011, and Duffy and Germani, 2013). When applied to H(T, y0, z) ≡ eiz(r–q)T Φ (T, y0, –z). The integral notation means: integrate along a line
the problems in the current work we found excellent agreement between the results parallel to the real z-axis, such that the imaginary part ℑz ∈ (0, 1). We had some
produced by ADE and those produced by Mathematica. We note that ADE uses a discretion about where to integrate. We chose ℑz ∈ (0, 1) because that condition
constant mesh size in time while NDSolve uses adaptive time-stepping. coincides with a strip of analyticity of the integrand whenever the process has both
We discuss the accuracy and performance of the Saul’yev scheme, in particular a bounded norm and a bounded expected stock price.6 It has been shown in Lewis
as applied to put option pricing and results in Table 2. This scheme has a truncation (2000) that H(t, y, z) is a solution to the PDE problem
( )2
error and we note the presence of the inconsistent term Δt 1 2
h H𝜏 = a (y)H yy + [b(y) − iz𝜌(y)𝜎̃ (y)a(y)]H y − c(z)𝜎̃ 2 (y)H , (9)
that affects accuracy, where Δt is the step-size in time and h is the step-size in space. 2
We can thus take a large number of steps in order to reproduce the results in Table 2. where c(z) = (z2 – iz)/2. The solution we want has the initial condition H(0, y; z) =1.
In general, we can get two-digit accuracy behind the decimal point using 200 mesh For our anchoring model (and many others), no boundary conditions are needed at
points in each direction, for example. We can improve the accuracy by using the y = 0 or y = ∞. For anchoring, H = H (t, x; z) and (9) becomes
Barakat–Clark scheme that has a truncation error O(Δt2,h2) (note that there is no
1 2
inconsistent term in this case). This scheme is discussed in Duffy (2009). H𝜏 = 𝜎 (𝜉 ) 𝜉 2 H 𝜉𝜉 + [(𝜆 + r − 𝜆 𝜉 ) − iz𝜎 2 (𝜉 )] 𝜉 H 𝜉 − c(z)𝜎 2 (𝜉 ) H . (10)
2
In order to compute the implied volatility values in Table 1 we use the Black–
Scholes formula for calls and puts using the option values that we found using ADE.
Figure 7: Anchoring model implied volatility vs. strike and maturity. The
There are many methods to solve this nonlinear equation and in this case we used the
strikes range from 60 to 140, and the times range from 0 to 3 years. The
fixed-point method in combination with Aitken’s delta-squared process to accelerate
numerical method uses equation (8) with a numerical PDE solver to obtain
convergence. It is worth noting that the fixed-point method converges for any ini-
the integrand.
tial guess because the corresponding iteration function is a contraction (Haaser and
Sullivan, 1991). This is in contrast to the Newton–Raphson and bisection methods
(for example), where it can sometimes be an issue to find good starting values for
these iterative methods.
Finally, the ADE method is very efficient and does not need matrix solvers. In fact,
most of the computational overhead lies in computing the sigmoidal function that we
introduced in Section 5. This involves calling the exponential and logarithm functions,
which are computationally expensive. We can then consider faster approximations
to these functions or pre-computing lookup tables of the sigmoidal function. For an
introduction to the finite difference method in finance, see Duffy (2006).

9 Dimensional reduction: pricing by Fourier

transform
We introduce an alternative which reduces the PDE problem from two spatial fac-
tors to one factor. Apart from its application to general time numerics, we need this
development for our discussion of model asymptotics.

52 WILMOTT magazine

48-56_Wilm_Duff y_TP_Sept_2014_Final.indd 52 09/02/2014 02:4 PM

 15418286, 2014, 73, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/wilm.10366 by Cochrane Netherlands, Wiley Online Library on [23/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
TECHNICAL PAPER

As promised, we have achieved a dimensional reduction – at the expense of requiring Figure 8: Asymptotic T → 0 implied volatility vs. strike. The figure shows a
complex numbers and an integration. The recipe here: write a one-dimensional solv- plot of the formula in equation (13) for the parameter set of Table 2.
er for (10) and feed the solutions to (8), integrating along z = i/2 + u, u Î (0, ∞). This
σimp(K)
Fourier method is used to produce the implied volatility surface shown in Figure 7.
One check: consider the case where the sigmoidal function is degenerate, so
that 𝜎 2(𝜉 ) = 𝜎02 , a constant. Then, the solution to (10), subject to H(0, x; z) = 1, is 0.5
H (𝜏 ,𝜉 ; z) = e−c(z)𝜎0 𝜏 . Note that this has no x-dependence. As shown in Appendix
2

2.1 of Lewis (2000), and as expected, (8) then yields the Black–Scholes formula with 0.4
volatility s0.
Finally note that, unlike ‘traditional’ stochastic volatility models, the interest rate 0.3
r appears in the ‘volatility’ SDE and consequently in the fundamental PDE (10). This
0.2
has the consequence that the BS implied volatility takes on an additional dependence
on r that cannot be absorbed into the moneyness variable m above. This nuance is
0.1
discussed further below.
K
20 40 60 80 100 120 140
10 Model asymptotics
Some natural questions are: how does the model behave for large and small l, recall-
ing that l sets the time scale for which history matters? In addition: how does the then (3) (with m = r) is the Black–Scholes SDE. Reassuringly, in that case, (13) cor-
implied volatility surface behave for large and small T, where T is the time to an rectly regenerates Σ(0) = s0.
option maturity? Asymptotics help clarify the mathematical behavior of the model, Figure 8 shows a plot of the formula in (13) for the parameter set of Table 2. In the
which generally requires a numerical solution. Asymptotics may help create rela- sigmoidal function, a = 0.60, which is the maximum ‘local’ volatility; this is also seen
tively faster calibrations, using them (when appropriate) as a substitute for our PDE in the figure to be the maximum implied volatility, as one would expect. Similarly, b =
solvers. 0.15 in the minimum local and implied volatility.
At-the-money relationship. In practice, as T grows small, liquidity in options
becomes concentrated near-the-money. By taking successive K-derivatives of equa-
10.1 The implied volatility smile as T → 0
tion (13), and then setting K = S0, one can derive the following at-the-money rela-
Option prices equal their payoff function at T = 0. What is interesting is that, for any
tions. To this end, abbreviating Σ(K) ≡ Σ(0)(K, S0, A0) and with x0 = S0/A0, we find
diffusion model (like ours), there is a nontrivial limit as T → 0 for the Black–Scholes
(BS) implied volatility, call it Σ. Recall the Black–Scholes formula for a call option on Σ(S0 ) = 𝜎(𝜉 0),
a non-dividend-paying stock:
S0 Σ′ (S0 ) = 12 𝜉 0 𝜎 ′(𝜉 0 ),
cBS (T, S0 , K, 𝜎) = S0 Φ(d 1 ) − Ke−rT Φ(d 2),
[𝜎 ′(𝜉 0 )]2
with well-known d1, 2. Consider any (time-homogeneous Markov) n-dimensional S20 Σ′′ (S0 ) = 13 𝜉 02 𝜎 ′′ (𝜉 0) − 32 𝜉 0 𝜎 ′ (𝜉 0) − 76 𝜉 02 .
𝜎(𝜉 0 )
diffusion model for a call option price. The model has stochastic factors x = (S, Y)
with x an n -vector and Y an (n – 1)-vector. (In our anchoring model, n = 2 and Y = A Derivation. Briefly, here is how (13) is obtained. We start with call option values
or, Y = x take your pick.) Let C(T, S0, Y0; K) denote that general model option price, in general n-dimensional models:
at time t = 0, for option maturity T and strike K. Then, Σ is defined by equating the
C(T, S0 , Y0 ; K) = e−rT max(ST − K, 0) q(T, S0 , Y0 ; ST ) dST ,
model price to the BS price: ∫
C(T, S0 , Y0 ; K) = cBS (T, S0 , K, Σ). (11) where q is the transition density for a particle starting at x0 = (S0, Y0) to arrive
at ST with any value of YT. Now, the asymptotic smile is obtained by taking two
Of course, for this to work, we must allow Σ = Σ (T, K, S0, Y0). Remarkably, the K-derivatives of both sides of (11) and comparing them as T → 0. First suppose the
complicated nonlinear inversion associated with (11), in general diffusion models, underlying n-factor diffusion has variance–covariance matrix a(x) = {aij(x)}. Denote
admits an (asymptotic) expansion of the form the inverse of this matrix by g(x), so g(x) = a–1(x).
First, two K derivatives of the right-hand-side. Since the BS formula is based on
Σ = Σ(0) + T Σ(1) + T 2 Σ(2) + · · · , as T → 0. (12)
the lognormal density, one has
So, for the limiting expiration smile, we have Σ(0) (K, S0, Y0) in general and Σ(0) (K, { }
𝜕 2cBS (logS0 ∕K)2
S0, A0) for our particular model here. Notice that in Table 1, we show entries for T = 0. (T, S 0 , K, Σ) ≈ exp − .
They were obtained from the following formula: 𝜕K 2 2 (Σ(0) )2 T (14)
log S0 ∕K (Note: all “≈” denote leading behaviors as T → 0, up to ignorable factors.)
Σ(0) (K, S0 , A 0 ) = S ∕A
, (13)
dx
∫K ∕A
0
0
0
x𝜎(x)
Now the left-hand-side. From Breeden–Litzenberger’s (1978) relation and a clas-
sic argument due to Varadhan (1967):
where we recall that s(x) is our sigmoidal function. At K = S0, an application of { 2 }
l’Hospital’s rule is implied, which yields the asymptotic at-the-money implied volatil- 𝜕 2C −rT
−d (S0 , Y0 ; K)
(T, S , Y ; K) = e q(T, S , Y ; K) ≈ exp , (15)
^
0 0 0 0
ity s (S0/A0). [For Tables 1 and 2, S0 = A0 and s(1) ≈ 0.171342.] If s(x) = s0, a constant, 𝜕K 2 2T

WILMOTT magazine 53

48-56_Wilm_Duff y_TP_Sept_2014_Final.indd 53 09/02/2014 02:4 PM

 15418286, 2014, 73, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/wilm.10366 by Cochrane Netherlands, Wiley Online Library on [23/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
where 1 Figure 9: ‘Large T asymptotics’ is an eigenvalue problem. As explained in the
d 2 (S0 , Y0; K) ≡ min gij (x(s)) ẋ i (s) ẋ j (s) ds. text, the asymptotic implied volatility as T → ∞ is associated with an eigen-
x(0)=(S0 ,Y0) ∫0
x(1)=(K,free) value problem Lu0(x) = Ù0u0 (x), where Ù0 = Ù0(y *). For the parameter set of
In the integral shown for d2, the ‘min’ is a functional minimization over all parame- Table 2, but with r = 0, the figure shows Ù0(y) and the associated eigenfunc-
terized path functions {x(s): 0 ≤ s ≤ 1} subject to the endpoint conditions shown. The tion at y = y*, the maximum point.
scalers xi are simply the components of x, ẋ i(s) = dxi/ds, and there is an implied sum- Λ(y) u0(ξ )
mation over repeated indices. With that notation, d(S0,Y0; K) has a fascinating inter- 0.0070 1.0
pretation. It is the minimum geodesic distance in the (S, Y ) space from (S0, Y0) to the 0.0065 0.9
(hyperplane) S = K. The geodesics are those found by treating g = a–1 as a metric and 0.0060 0.8
doing deterministic (that is, non-stochastic) path calculations. Comparing the right- 0.0055 0.7
hand sides of (14) and (15) yields a general result for the asymptotic smile: 0.0050 0.6
y ξ
| log S0 ∕K | 0.3 0.4 0.5 0.6 0.7 1 2 3 4
Σ (K, S0 , Y0 ) = ||
(0) |.
| (16)
| d(S0 , Y0 ; K) |
Thus it comes down to calculating geodesic distance functions d(x, A), where x is
Table 3: Black–Scholes implied volatilities (%): large T and T → ∞
again an n -dimensional point and A is any set in the factor variable space that does
not contain x. There are a variety of ways to do this. One of the most straightforward T r=0 r = 0.02 r = 0.04 r = 0.06 r = 0.08 r = 0.10
ways is to solve the Eikonal equation,7 namely 1000 23.58 21.61 19.94 18.56 17.50 16.71
𝜕d 𝜕d ∞ 23.61 21.68 20.00 18.62 17.54 16.74
aij (x) = 1,
𝜕xi 𝜕xj
subject to the boundary condition that d(x, A) vanishes when x touches A. For
Table 4: Option prices corresponding to the T = 1000 entry in Table 3
example, start with n = 1 and local volatility evolutions dSt = 𝜎̃ (St )St dWt . A
drift may be present, but it plays no role in the limiting smile, so we ignore it. r=0 r = 0.02 r = 0.04 r = 0.06 r = 0.08 r = 0.10
Then, a(S) = 𝜎(S)̃ 2 S2 and d = d(S; K ). Thus the Eikonal problem is easy: solve 99.981 99.937 99.838 99.664 99.434 99.176
2 2 2
S 𝜎(S)
̃ (d S ) = 1, subject to d(S = K ) = 0. This has the solution Parameter set: a = 0.6, b = 0.15, c = 10, d = –0.3, l = 0.5. For each r, the strike was
S set to the forward at-the-money value: K = S0erT.
du
d(S; K) = . (17)
∫K u 𝜎̃ (u)
Now our problem is a two-factor problem, dSt = s(St/At)StdWt, but the second factor and that is assumed throughout our discussion.8 With that, one can develop the simple
At has an SDE with a drift but no Brownian term. Given that degeneracy, one would relation for the ultimate implied variance
guess that local volatility formula (17) applies with 𝜎̃ (S) = 𝜎(S∕A 0) . We recall that
V∞ = lim V imp (T, m, 𝜉 0) = 8 Λ0(y∗ ).
imp
s(·) is the sigmoidal function. That is, (18)
T→∞
S0 S ∕A0 √
0
du dx imp imp
d(S0 , A 0 ; K) = = . Then, 𝜎∞ = V∞ . In (18), Λ0(y) is the principal (i.e., smallest) eigenvalue
∫K u 𝜎(u∕A 0) ∫ K ∕A0 x 𝜎(x)
associated with the (negative of the) spatial differential operator of (10), fixing z = iy
Applying (16), this last relation yields our formula (13). Our guess can be justified (0 < y < 1). This leads to the eigenvalue problem  u0 (𝜉 ) = Λ0 u0(𝜉 ) , suppressing the
by setting up the full problem in two-dimensional (S, x) space, which is governed dependence on the fixed y. The function space for u0 is: smooth enough functions
by a bivariate SDE with two (perfectly correlated) Brownian noise terms. Then, one with certain boundary behaviors. Then, y* is a stationary point; i.e., Λ0(y*) is the larg-
shows that the putative solution for the distance function (but written in terms of S est of such principal eigenvalues along the line segment y ∈ (0,1). For the parameter
and x) satisfies the corresponding two-dimensional Eikonal equation. It is tedious, set of Table 2, but with r = 0, Figure 9 shows Λ0(y) and the associated eigenfunction at
but straightforward, and a good exercise for the ambitious reader. y =y*, the maximum point.
Further reading. The geodesic connection is developed in Varadhan (1967). See From (18), and the parameter set of Table 2, except for various interest rates r,
imp
Lewis (2007) for a more elaborate but heuristic derivation of (16) and further referenc- Table 3 shows computed values of 𝜎∞ . To check the results (see Tables 3 and 4), the
es. See also the chapter on ‘Advanced smile asymptotics’ in Lewis (2014, forthcoming). full time-dependent problem has also been solved with T = 1000. This was done by
using the Fourier transform method, solving (10) for H(T, x; z) in Mathematica with
10.2 The implied volatility smile as T → ¥ NDSolve. For the finite-T solver call, no boundary conditions were taken at x = 0 and
The method is to develop the large-T behavior of the Fourier solution representation a zero x-derivative was taken at a cutoff xmax= 2. One can see that the finite maturity
(8) and especially the behavior of the ‘fundamental transform’ H(T, x0;z) in this limit. results are consistent with asymptotic values.
This method was developed in detail in Lewis (2000), so we will be brief. If one carried out this same exercise in, say, the Heston (1993) stochastic
imp
In a nutshell, in general models, the implied volatility surface flattens at large matu- volatility model, all the values in Table 3 for 𝜎∞ would be the same. This is because
rity T. The flattening is to a common value independent of the moneyness and initial the ultimate implied volatility flattens to a value that depends on the parameters
factors. In our case, the initial factors are (S0, x0). One has to be careful about what is in the volatility process, which does not contain r. Here, as noted earlier, such
held fixed as T → ∞. The natural thing is to fix the moneyness m = log {S0e–qT/Ke–rT }, r-dependence is present.

54 WILMOTT magazine

48-56_Wilm_Duff y_TP_Sept_2014_Final.indd 54 09/02/2014 02:4 PM

 15418286, 2014, 73, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/wilm.10366 by Cochrane Netherlands, Wiley Online Library on [23/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
TECHNICAL PAPER

Further details of the method, as applied to this particular problem, are rather Large l. Finally, to complete the asymptotics, we note that as l → ∞, this drives
involved and may be reported elsewhere. At → St. So s(St/At) → s(1) and the model prices should converge to the Black–
Scholes’ model prices with constant s = s(1).
10.3 Model prices as l → 0
Using subscripts in S and t for derivatives, and t = T –t, write equation (5) as
 V = –l(S–A)VA. Here, for any smooth function f (t, S, A), we define
11 Calibration challenges and conclusions
In principle, given a parameterized version of s(x), one can estimate the parameters
f (𝜏 , S, A) ≡ −f𝜏 + 12 𝜎 2 ( AS )S2 fSS + rSfS − rf by fits to SPX option chains. Neither of the two Mathematica implementations here
.
are efficient enough for this. However, we can refer the reader to Foschi and Pascucci
By substitution, it is clear that there is a formal power series solution in l; that is, (2009), who have managed to achieve such a calibration in the closely related class of
V = V (0) + 𝜆 V (1) + 𝜆2 V (2) + · · · , models due to Hobson and Rogers (1998).
In general, diffusion models have a very hard time fitting short-dated SPX
where the V(n) are recursively determined by  V (n) = −(S − A)VA(n−1) . Now V(0) is options. That is because the implied volatility smiles for broad-based indexes such as
the solution to SPX tend to become increasingly “V-shaped” as maturity approaches. That behavior is
(0)
V𝜏(0) = 12 𝜎 2 ( AS )S2 VSS + rSVS(0) − rV (0) , (19) strongly suggestive of price jump processes. Since all ‘complete models with stochastic
volatility’ are diffusions, calibrations will likely suggest that these classes of models are
subject to appropriate initial and boundary conditions for the particular payoff best embedded as components in (larger) jump diffusions. This will, of course, break
being considered. Suppose that we found V(0). Since it will satisfy the right initial and the complete-markets property, which is generally contra-indicated anyway.
boundary conditions for the full problem (that is, the problem for V ), that means all
the subsequent V(n)(n ≥ 1) should be solved for with zero initial and boundary condi-
Acknowledgments
tions. This is well-posed because each such subsequent problem has a driving term.
We would like to thank the following members of wilmott.com for their sugges-
Now let us restrict ourselves to the problems for vanilla puts and calls with strike
tions in the thread ‘CEV and SABR for beta outside of [0,1],’ which can be found at
K, writing V = V(t, S, A; K). Take the put; we can write V(0)(0, S, A; K) = (K – S)+ =
www.wilmott.com/messageview.cfm?catid=4&threadid=89643, mtsm, spv205, list,
A(k – x)+, using x = S/A, k = K/A. This suggests we should look for a solution to (19)
SierpinskyJanitor, TinMan, daveangel, japanstar, oislah, Herd, frenchX, croot (some
in the form V(0)(t, S, A; K) = Au(0) (t, x; k). Indeed, u(0) solves
were more helpful than others).
{ }
2 (0) (0)
u(0)
𝜏 −
1 2
2
𝜎 (𝜉 )𝜉 u 𝜉𝜉
+ r 𝜉 u 𝜉
− r u (0)
=0 (20)
Paul Wilmott was born in the quaint little fishing village of Birkenhead and studied
subject to u(0)(0, x; k) = (k – x)+, and similarly for the call. This is readily solved mathematics at St Catherine’s College, Oxford, where he also received his DPhil. He
numerically – certainly much easier than the previous PDEs with two spatial factors. founded the Diploma in Mathematical Finance at Oxford University and the journal
Applied Mathematical Finance. He is the author of Paul Wilmott on Quantitative Finance
Then, at the next order, V(1)(t, S, A; K) = Au(1) (t, x; k). Now u(1) solves
(Wiley, 2006), Paul Wilmott Introduces Quantitative Finance (Wiley, 2007), Frequently Asked
{ }
u(1)
𝜏
− 1 𝜎 2 (𝜉 )𝜉 2u(1)
𝜉𝜉
+ r 𝜉 u(1)
𝜉
− r u(1) = f (𝜏 , 𝜉 ;k), (21) Questions in Quantitative Finance (Wiley, 2009), and other financial textbooks. He has writ-
2 ten over 100 research articles on finance and mathematics. He was a founding partner
where of the volatility arbitrage hedge fund Caissa Capital, which managed US$170 million. His
( )
responsibilities included forecasting, derivatives pricing, and risk management. Dr. Wilmott
f (𝜏 , 𝜉 ; k) = (1 − 𝜉 ) 𝜉 u(0)
𝜉
+ k u (0)
k
− u (0)
,
is the proprietor of www.wilmott.com, the popular quantitative finance community web-
site, and the quant magazine Wilmott, and is the creator of the Certificate in Quantitative
and with u(1)(0, x; k). The same equation applies to all orders, except that the left- Finance. He was a professional juggler with the Dab Hands troupe. He also has three half
hand side uses u(n) and the right-hand side uses u(n–1). We have carried this through blues from Oxford University for Ballroom Dancing. He was the first man in the UK to get an
for the first two orders in Mathematica, and show the following numerical results. online divorce. In his spare time he makes jam and cheese, and plays the ukulele.
Consider the Table 2 entries at T = 1 and T = 3. We show the results for V(0), lV(1) and
their sum in Table 5. Comparing with Table 2, the expansion results are very good at Alan L. Lewis has been active in option valuation and related financial research for
T = 1. But they suggest that at least the O(l2) correction is necessary for a good result over 30 years. He is the author of the book Option Valuation under Stochastic Volatility,
at T = 3. In general, the controlling parameter here is the dimensionless T. a forthcoming second volume with the same title, and articles in many of the leading
financial journals. Previously, he served as Director of Research, Chief Investment Officer,
Table 5: European-style put option prices: l expansion and President of the mutual fund family at Analytic Investment Management, a money
management firm specializing in derivative securities. He was a co-founder and former
T=1 K = 60 K = 80 K = 100 K = 120 K = 140
Board Chairman for Envision Financial Systems, Inc. He received a PhD in physics from
V(0) 0.230 0.986 4.717 16.255 33.545 the University of California at Berkeley, a BS from Caltech, and currently lives in Newport
lV (1) –0.066 –0.155 –0.043 0.047 0.002 Beach, CA.
Sum 0.164 0.831 4.674 16.302 33.547
T =3 K = 60 K = 80 K = 100 K = 120 K = 140 Daniel J. Duffy has been working with numerical methods in finance, industry, and
(0) engineering since 1979. He has written four books on financial models, numerical
V 1.458 2.975 6.249 13.615 25.334
methods, and C++ for computational finance and has also developed a number of new
lV (1) –0.514 –0.766 –0.371 0.054 0.030 schemes in this field. He is the founder of Datasim Education and has a PhD in numerical
^
Sum 0.945 2.209 5.878 13.669 25.365 analysis from Trinity College, Dublin.

WILMOTT magazine 55

48-56_Wilm_Duff y_TP_Sept_2014_Final.indd 55 09/02/2014 02:4 PM

 15418286, 2014, 73, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/wilm.10366 by Cochrane Netherlands, Wiley Online Library on [23/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ENDNOTES Barakat, H.Z. and Clark, J.A. (1966). On the solution of diffusion equations by numerical
1. Our model is ‘complete with stochastic volatility,‘ in the sense of Hobson and Rogers methods. Journal of Heat Transfer 88, 421–427.
(1998), and has similar features. Breeden, D. and Litzenberger, R. (1978). Prices of state-contingent claims implicit in
2. We generally denote partial derivatives with a subscript, except where clarity option prices. The Journal of Business 51(4), 621–651.
suggests a more explicit notation. Duffy, D.J. (2006). Finite Difference Methods in Financial Engineering. Wiley: Chichester, UK.
3. Not every choice for s(x) admits a stationary density, and a stationary density is not Duffy, D.J. (2009). Unconditionally stable and second-order accurate explicit finite dif-
always needed. When s(0) < ∞, as is true for our sigmoidal parameterization, it is easy ference schemes using domain transformations, Part I. One-factor equity problems.
to see that DOI: 10.2139/ssrn.1552926.
Duffy, D.J. and Germani, A. (2013). C# for Financial Markets. Wiley: Chichester, UK.
Foschi, P. and Pascucci, A. (2009). Calibration of a path-dependent volatility model:
Empirical tests. Computational Statistics and Data Analysis 53(6), 2219–2235.
This leads to the x -stationarity condition: m + l > s(0)2/2. When s(0) < ∞, but the sta- Haaser, N.B. and Sullivan, J.A. (1991). Real Analysis. Dover Publications: New York.
tionary condition is violated, x -mass simply accumulates closer and closer to the natu- Heston, S. (1993). A closed-form solution for options with stochastic volatility with
ral boundary x = 0 (similar to GBM). For some applications, say stocks with bankruptcy applications to bond and currency options. Review of Financial Studies 6(2), 327–343.
risk, unbounded s(·) may be desirable. Then, P∞(x) develops a Dirac mass at x = 0 and Hobson, D.G. and Rogers, L.C.G. (1998). Complete models with stochastic volatility.
the existence of a stationary density in x becomes irrelevant. Finally, nothing much Mathematical Finance 8(1), 27–48.
interesting happens for large x: for integrability there, it is sufficient that l > 0 and s(∞) Lewis, A. (2000). Option Valuation under Stochastic Volatility: With Mathematica Code.
< ∞. The latter condition is uncontroversial for most equities, as this is the regime of Finance Press: Newport Beach, CA.
large S relative to history. Lewis, A. (2007). Geometries and smile asymptotics for a class of stochastic volatility
4. The standard NDSolve algorithm is a ‘method of lines.’ Basically: apply a spatial PDE models. www.optioncity.net/publications.htm.
discretization, which converts the PDE into an ODE system. The latter is then solved by Lewis, A. (2014). Option Valuation under Stochastic Volatility II: With Mathematica Code.
an adaptive time-stepping routine. There are many details. Finance Press: Newport Beach, CA (forthcoming).
5. Such processes are sometimes termed Markov Additive Processes, or MAPs, in the Pealat, G. and Duffy, D.J. (2011). The alternating direction explicit (ADE) method for
applied probability literature. one factor problems. Wilmott magazine, July, 54–60.
6. Those are very weak conditions. Our anchoring model, with a sigmoidal function Saul’yev, V.K. (1964). Integration of Equations of Parabolic Type by the Method of Nets.
bounded away from 0 and infinity, should be both norm-preserving and martingale- Pergamon Press: Oxford, UK.
preserving – although we do not prove it. Tversky, A. and Kahneman, D. (1974). Judgment under uncertainty: Heuristics and
7. Sometimes called the Hamilton–Jacobi equation. biases. Science 185, 1124–1130.
8. Suppose S0 is fixed and we increase T. Note that fixing the moneyness is not the Varadhan, S.R.S. (1967). On the behavior of the fundamental solution of the heat
same as fixing the strike K, except when q = r. equation with variable coefficients; Diffusion processes in a small time interval.
Communications on Pure and Applied Mathematics 20, 431–455; 659–685.
REFERENCES Wilmott, P. (2006). Paul Wilmott on Quantitative Finance, 2nd edn. Wiley: Chichester, UK.
Apabhai, M.Z., Choe, K., Khennach, F., and Wilmott, P. (1995). Spot-on modelling. Risk Wilmott, P. (2009). Where quants go wrong: A dozen basic lessons in commonsense for
magazine, December. quants and risk managers and the traders who rely on them. Wilmott Journal 1(1), 1–22.

56 WILMOTT magazine

48-56_Wilm_Duff y_TP_Sept_2014_Final.indd 56 09/02/2014 02:4 PM

15418286, 2014, 73, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/wilm.10366 by Cochrane Netherlands, Wiley Online Library on [23/11/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License