m3f33chVII

Ch. VII: MATHEMATICAL FINANCE IN CONTINUOUS TIME

§1. Geometric Brownian Motion (GBM)

As before, we write B for standard Brownian motion. We write Bµ,σ for
Brownian motion with drift µ and diffusion coefficient σ: the path-continuous
Gaussian process with independent increments such that

Bµ,σ (s + t) − Bµ,σ (s) is N (µt, σ 2 t).

This may be realised as

Bµ,σ (t) = µt + σB(t).

Consider the process

1 1
Xt = f (t, Bt ) := x0 . exp{(µ− σ 2 )t+σBt } : log Xt = const+(µ− σ 2 )t+σBt .
2 2
(∗)
Here, since
1
f (t, x) = x0 . exp{(µ − s2 )t + σx},
2
1
f1 = (µ − σ 2 )f, f2 = σf, f22 = σ 2 f.
2
By Itô’s Lemma (Ch. VI: dXt = Ut dt + Vt dBt and f smooth implies df =
(f1 + U f2 + 21 V 2 f22 )dt + V f2 dBt ) we have (taking U = 0, V = 1, X = B),

1 1
dXt = df = [(µ − σ 2 )f + σ 2 f ]dt + σf dBt :
2 2
dXt = µf dt + σf dBt = µXt dt + σXt dBt :
X satisfies the SDE

dXt = Xt (µdt + σdBt ) : dXt /Xt = µdt + σdBt , (GBM )

and is called geometric Brownian motion (GBM). We turn to its economic

meaning, and the role of the two parameters µ and σ, below. It will be used
to model price processes in the Black-Scholes model of VII.2. But note that
in (∗), log-prices log Xt are normally distributed.

1
 Note that for µ = 0, (GBM ) shows that X is a martingale (see VII.3, in
connection with Girsanov’s theorem).
We recall the model of Brownian motion from Ch. VI. It was developed
(by Brown, Einstein, Wiener, ...) in statistical mechanics, to model the ir-
regular, random motion of a particle suspended in fluid under the impact of
collisions with the molecules of the fluid.
The situation in economics and finance is analogous: the price of an as-
set depends on many factors (a share in a manufacturing company depends
on, say, its own labour costs, and raw material prices for the articles it
manufactures. Together, these involve, e.g., foreign exchange rates, labour
costs, transport costs, etc. – all of which respond to the unfolding of events
– economic data/political events/the weather/technological change/labour,
commercial and environmental legislation/ ... in time. There is also the
effect of individual transactions in the buying and selling of a traded asset
on the asset price. The analogy between the buffeting effect of molecules
on a particle in the statistical mechanics context on the one hand, and that
of this continuous flood of new price-sensitive information on the other, is
highly suggestive. The first person to use Brownian motion to model price
movements in economics was Bachelier in his celebrated thesis of 1900.
Bachelier’s seminal work was not definitive (indeed, not correct), either
mathematically (it was pre-Wiener) or economically. In particular, Brownian
motion itself is inadequate for modelling prices, as
(i) it attains negative levels, and
(ii) one should think in terms of return, rather than prices themselves.
However, one can allow for both of these by using geometric, rather than
ordinary, Brownian motion as one’s basic model. This has been advocated
in economics from 1965 on by Samuelson1 – and was Itô’s starting-point for
his development of Itô or stochastic calculus in 1944 – and is now standard.
Returning now to (GBM), the SDE above for geometric Brownian mo-
tion driven by Brownian noise, we can see how to interpret it. We have a
risky asset (stock), whose price at time t is Xt ; dXt = X(t + dt) − X(t) is
the change in Xt over a small time-interval of length dt beginning at time t;
dXt /Xt is the gain per unit of value in the stock, i.e. the return. This is a
sum of two components:
(i) a deterministic component µdt, equivalent to investing the money risk-
lessly in the bank at interest-rate µ (> 0 in applications), called the under-
1
Paul A. Samuelson (1915-2009), American economist; Nobel Prize in Economics, 1970

2
lying return rate for the stock,
(ii) a random, or noise, component σdBt , with volatility parameter σ > 0
and driving Brownian motion B, which models the market uncertainty, i.e.
the effect of noise. Note that dBt is a Brownian increment, so is normally
distributed. So: returns are normally distributed.
Return intervals.
That both log-prices and returns are normally distributed just reflects
log(1 + x) ∼ x (x → 0),
or equivalently (as in II.1),
 x
1+ → ex (n → ∞).
n
We can recognise this as being bound up with the passage from discrete time
(time-interval ∆t, small but finite, as in V) to continuous time (time-interval
dt, infinitesimal, and the SDE for GBM as above). Now in investment, there
are many possible time-scales, corresponding to how often we observe prices;
we single out the main three (cf. [BK, §2.9]).
1. Long (macroscopic).
Here we are investing over a time-scale of months (say – or years), and
observe prices daily (say). As the price-change over the month is the sum
of price-changes over the days, and these are independent (as Brownian in-
crements are), the reason we get normality is the Central Limit Theorem
(CLT): if we sum many independent random variables with finite mean and
variance, we get normality (in the limit) after centring and scaling. This is
the phenomenon of aggregational Gaussianity. Note that Gaussian (normal)
tails are extremely thin (‘minus log-density’ grows quadratically). The ‘rule
of thumb’ is that 16 trading days suffice here.
2. Intermediate (mesoscopic).
If our investment time-frame is, say, a day (there are ‘day traders’ out
there!), aggregational Gaussianity does not set in, and the tails observed are
much fatter – typically, ‘minus log-density’ grows linearly. One model com-
monly used here is that of hyperbolic distributions (see e.g. [BK, §2.12]).
3. Short (microscopic).
With the development of the Internet and the intensive computerisation
of trading, high-frequency data – ‘tick data’ – is available; here the interval
may be of the order of seconds or much smaller. Here, the picture is differ-
ent again: the tails are much fatter still: tails decay like a power, so ‘minus

3
log-density’ grows logarithmically. Distributions used include Student t and
stable (see e.g. [BK, §2.9]).
Note. The world’s most famous investor, Warren Buffett, the Sage of Om-
aha, famously invests right, and over a time-frame of many years.

§2. The Black-Scholes Model

For this section only, it is convenient to be able to use the ‘W for Wiener’
notation for Brownian motion/Wiener process, thus liberating B for the al-
ternative use ‘B for bank [account]’. Thus our driving noise terms will now
involve dWt , our deterministic [bank-account] terms dBt .
We now consider an investor constructing a trading strategy in continu-
ous time, with the choice of two types of investment:
(i) riskless investment in a bank account paying interest at rate r > 0 (the
short rate of interest): Bt = B0 ert (t ≥ 0) [we neglect the complications
involved in possible failure of the bank – though banks do fail – witness Bar-
ings 1995, or AIB 2002!];
(ii) risky investment in stock, one unit of which has price modelled as above
by GM B(µ, σ). Here the volatility σ > 0; the restriction 0 < r < µ on the
short rate r for the bank and underlying rate µ for the stock are economically
natural (but not mathematically necessary); the stock dynamics are thus

dSt = St (µdt + σdWt ).

Notation. Later, we shall need to consider several types of risky stock - d

stocks, say. It is convenient, and customary, to use a superscript i to label
stock type, i = 1, · · · , d; thus S 1 , · · · , S d are the risky stock prices. We can
then use a superscript 0 to label the bank account, S 0 . So with one risky
asset as above, the dynamics are

dSt0 = rSt0 dt,

dSt1 = µSt1 dt + σSt1 dWt .

We shall focus on pricing at time 0 of options with expiry time T ; thus the
index-set for time t throughout may be taken as [0, T ] rather than [0, ∞).
We proceed as in the discrete-time model of V.1. A trading strategy H is
a vector stochastic process

H = (Ht : 0 ≤ t ≤ T ) = ((Ht0 , Ht1 , · · · , Htd )) : 0 ≤ t ≤ T )

4
which is previsible: each Hti is a previsible process (so, in particular, (Ft− )-
adapted) [we may simplify with little loss of generality by replacing previsi-
bility here by left-continuity of Ht in t]. The vector Ht = (Ht0 , Ht1 , · · · , Htd )
is the portfolio at time t. If St = (St0 , St1 , · · · , Std ) is the vector of prices at
time t, the value of the portfolio at t is the scalar product

Vt (H) := Ht .St = Σdi=0 Hti Sti .

The discounted value is

Ṽt (H) = βt (Ht .St ) = Ht .S̃t ,

where βt := 1/St0 = e−rt (fixing the scale by taking the initial bank account
as 1, S00 = 1), so
S̃t = (1, βt St1 , · · · , βt Std )
is the vector of discounted prices.
Recall that
(i) in V.1 H is a self-financing strategy if ∆Vn (H) = Hn .∆Sn , i.e. Vn (H) is
the martingale transform of S by H,
(ii) stochastic integrals are the continuous analogues of mg transforms.
We thus define the strategy H to be self-financing, H ∈ SF , if

dVt = Ht .dSt = Σd0 Hti dSti .

The discounted value process is

Ṽt (H) = e−rt Vt (H)

and the interest rate is r. So

dṼt (H) = −re−rt dt.Vt (H) + e−rt dVt (H)

(since e−rt has finite variation, this follows from integration by parts,
1
d(XY )t = Xt dYt + Yt dXt + dhX, Y it
2
– the quadratic covariation of a finite-variation term with any term is zero)

= −re−rt Ht .St dt + e−rt Ht .dSt = Ht .(−re−rt St dt + e−rt dSt ) = Ht .dS̃t

5
(S̃t = e−rt St , so dS̃t = −re−rt St dt + e−rt dSt as above).
Summarising: for H self-financing,

dVt (H) = Ht .dSt , dṼt (H) = Ht .dS̃t ,

Z t Z t
Vt (H) = V0 (H) + Hs dSs , Ṽt (H) = Ṽ0 (H) + Hs dS̃s .
0 0

Now write Uti := Hti Sti /Vt (H) = Hti Sti /Σj Htj Stj for the proportion of the
value of the portfolio held in asset i = 0, 1, · · · , d. Then ΣUti = 1, and
Ut = (Ut0 , · · · , Utd ) is called the relative portfolio. For H self-financing,

Hti Sti dSti

dVt = Ht .dSt = ΣHti dSti = Vt Σ . : dVt = Vt ΣUti dSti /Sti .
Vt Sti

Dividing through by Vt , this says that the return dVt /Vt is the weighted
average of the returns dSti /Sti on the assets, weighted according to their pro-
portions Uti in the portfolio – as one would expect.
Note. Having set up this notation (that of [HP]) – in order to be able if
we wish to have a basket of assets in our portfolio – we now prefer – for
simplicity – to specialise back to the simplest case, that of one risky asset.
Thus we will now take d = 1 until further notice.

§3. The (continuous) Black-Scholes formula (BS): derivation via

Girsanov’s Theorem

The Sharpe ratio.

There is no point in investing in a risky asset with mean return rate µ,
when cash is a riskless asset with return rate r, unless µ > r. The excess
return µ − r (the investor’s reward for taking a risk) is compared with the
risk, as measured by the volatility σ, via the Sharpe ratio

θ := (µ − r)/σ,

also written λ, and also known as the market price of risk. This is important,
both here (see below), in CAPM (II.3), and in asset allocation decisions.
Consider now the Black-Scholes model, with dynamics

dBt = rBt dt, dSt = µSt dt + σSt dWt .

6
The discounted asset prices S̃t := e−rt St have dynamics given, as before, by
dS̃t = −re−rt St dt + e−rt dSt = −rS̃t dt + µS̃t dt + σ S̃t dWt
= (µ − r)S̃t dt + σ S̃t dWt = σ S̃t (θdt + dWt ).
We summarise the main steps briefly as (a) - (f) below:
(a) Dynamics are given by GBM , dSt = µSdt + σSdWt (VI.1).
(b) Discount: dS̃t = (µ − r)S̃dt + σ S̃dWt = σ S̃(θdt + dWt ) (above).
We work with the discounted stock price S̃t . We would like this to be
a martingale, as in Ch. V, where we passed from P -measure to Q- (or
P ∗ )-measure, so as to make discounted asset prices martingales. Girsanov’s
theorem (below) accomplishes this, in our new continuous-time setting: it
maps P to P ∗ (or Q), and µ to r, so θ to 0. This kills the dt term on the
right in (b). If we then integrate dS̃t = σ S̃dWt , we get an Itô integral, so a
martingale, on the right. Assuming this for now:
(c) Use Girsanov’s Theorem to change µ to r, so θ := (µ − r)/σ to 0: under
P ∗ , dS̃t = σ S̃dWt .
(d) This and dṼt (H) = Ht dS̃t (where V is the value process and H the
trading strategy replicating the payoff h – VII.2) give dṼt (H) = Ht .σ S̃t dWt
(VII.2 above). Integrate: Ṽt is a P ∗ -mg, so has constant E ∗ -expectation.
(e) This gives the Risk-Neutral Valuation Formula (RNVF), as in V.4.
(f) From RNVF, we can obtain BS, by integration, as in V.6.
It remains to state and discuss Girsanov’s theorem. We cannot prove it
in full (only the finite-dimensional approximation below) – this is technical
Measure Theory. But we must expect this in this chapter: in discrete time
(Ch. V) we could prove everything; here in continuous time, we can’t.
Consider first ([KS], §3.5) independent N (0, 1) random variables Z1 , · · · , Zn
on (Ω, F, P ). Given a vector µ = (µ1 , · · · , µn ), consider a new probability
measure P̃ on (Ω, F) defined by
1
P̃ (dω) = exp{Σn1 µi Zi (ω) − Σn1 µ2i }.P (dω).
2
R
This is a positive measure as exp{.} > 0, and integrates to 1 as exp{µi Zi }dP =
E[eµi Zi ] = exp{ 12 µ2i } (normal MGF – Problems 4 (bivariate normal), or Prob-
lems 8 Q1), so is a probability measure. It is also equivalent to P (has the
same null sets), again as the exponential term is positive (the exponential on
the right is the Radon-Nikodym derivative dP̃ /dP ). Also
1
P̃ (Zi ∈ dzi , i = 1, · · · , n) = exp{Σn1 µi zi − Σn1 µ2i }.P (Zi ∈ dzi , i = 1, · · · , n)
2
7
(Zi ∈ dzi means zi ≤ Zi ≤ zi + dzi , so here Zi = zi to first order)
1 1 1 1 1
= (2π)− 2 n exp{Σµi zi − Σµ2i − Σzi2 }Πdzi = (2π)− 2 n exp{− Σ(zi −µi )2 }dz1 · · · dzn .
2 2 2
This says that if the Zi are independent N (0, 1) under P , they are inde-
pendent N (µi , 1) under P̃ . Thus the effect of the change of measure P 7→ P̃ ,
from the original measure P to the equivalent measure P̃ , is to change the
mean, from 0 = (0, · · · , 0) to µ = (µ1 , · · · , µn ).
This result extends to infinitely many dimensions – i.e., stochastic pro-
cesses. We quote (Igor Vladimirovich GIRSANOV (1934-67) in 1960):

Theorem (Girsanov’s Theorem). Let (µt : 0 ≤ t ≤ T ) be an adapted

RT
process with 0 µ2t dt < ∞ a.s. such that the process L with
Z t
1 t 2
Z
Lt := exp{ µs dWs − µ ds} (0 ≤ t ≤ T )
0 2 0 s
is a martingale. Then, under the probability PL with density LT relative to
P , the process W ∗ defined by
Z t
∗
Wt := Wt − µs ds, (0 ≤ t ≤ T )
0
Rt
is a standard Brownian motion (so W is BM + 0
µs ds).

Here, Lt is the Radon-Nikodym derivative of PL w.r.t. P on the σ-algebra

Ft . In particular, for µt ≡ µ, change of measure by introducing the RN
derivative exp{µWt − 21 µ2 } corresponds to a change of drift from 0 to µ.
Exponential martingale.
The martingale condition in Girsanov’s theorem is satisfied in the case
µt ≡ µ is constant. For, write
1
Mt := exp{µWt − µ2 t}.
2
This is a martingale. For, if s < t,
1
E[Mt |Fs ] = E[exp{µ(Ws + (Wt − Ws )) − µ2 (s + (t − s))}|Fs ]
2
1 2 1
= exp{µWs − µ s}.E[exp{µ(Wt − Ws ) − µ2 (t − s)],
2 2
8
as the conditioning has no effect on the second term, by independent incre-
ments of Brownian motion. The first term on the right is Ms . The second
term is 1. For (normal MGF), if Z ∼ N (0, 1),
1
E[exp{µZ}] = exp{ µ2 };
2
√
Wt − Ws = t − sZ, Z ∼ N (0, 1)
(properties of BM). Combining, M is a mg, as required. //
So the case µt constant = µ of Girsanov’s theorem passes between BM
and BM + µt. The argument above uses this with µ − r for µ.
Girsanov’s Theorem (or the Cameron-Martin-Girsanov Theorem: R. H.
Cameron and W. T. Martin, 1944, 1945) is formulated in varying degrees of
generality, and proved, in [KS, §3.5], [RY, VIII].
Stochastic exponential.
The SDE for GBM, dSt /St = µdt + σdWt , with solution St = S0 exp{(µ −
1 2
2
σ )t + σWt } as above, is a special case of the Doléans-Dade exponential
(or stochastic exponential: Cathérine Doléans-Dade (1942-2004)). It extends
from Brownian motion to semi-martingales M , when it is written E(M ).

Theorem (Risk-Neutral Valuation Formula, RNVF). The no-arbitrage

price of the claim h(ST ) is given by

F (t, x) = e−r(T −t) Et,x

∗
[h(ST )|Ft ],

where St = x is the asset price at time t and P ∗ is the measure under which
the asset price dynamics are given by

dSt = rSt dt + σSt dWt .

Proof (Step (e) in the above: (a) – (d) are already done). Change measure
from P , corresponding to GBM (µ, σ), to P ∗ , corresponding to GBM (r, σ),
by Girsanov’s Theorem. Then as above, dS̃t = σ S̃t dWt . So by VII.2, dṼt =
Ht dS̃t = Ht .σ S̃t dWt , where V is the value process following strategy H to
replicate payoff h. Integrating, Vt is a P ∗ -martingale, as it is an Itô integral.
So it has constant expectation. So if St = x is the asset price at time t,
∗ ∗
Et,x [Ṽt (H)|Ft ] = Et,x ṼT (H) = e−rT Et,x
∗
h(ST ) :
∗
F (t, x) = Et,x Vt (H) = e−r(T −t) Et,x
∗
h(ST ). //

9
Theorem ((Continuous) Black-Scholes Formula, BS).
1 √
F (t, S) = SΦ(d+ )−e−r(T −t) KΦ(d− ), d± := [log(S/K)+(r± σ 2 )(T −t)]/σ T − t.
2

Proof (Step (f) in the above). After the change of measure P 7→ P ∗ , µ 7→ r

by Girsanov’s Theorem, St has P ∗ -dynamics as in GBM (r, σ):

dSt = rSt dt + σSt dWt , St = s, (∗)

with W a P ∗ -Brownian motion. So (VII.1) we can solve this explicitly:

1
ST = s exp{(r − σ 2 )(T − t) + σ(WT − Wt )}.
2
√
Now WT − Wt is normal N (0, T − t), so (WT − Wt )/ T − t =: Z ∼ N (0, 1):
1 √
ST = s exp{(r − σ 2 )(T − t) + σZ T − t}, Z ∼ N (0, 1).
2
So by the Risk-Neutral Valuation Formula, the pricing formula is
∞ 1 2
e− 2 x
Z
−r(T −t) 1 1
F (t, x) = e h(s exp{(r − σ 2 )(T − t) + σ(T − t) 2 x}). √ dx.
−∞ 2 2π
For a general payoff function h, there is no explicit formula for the integral,
which has to be evaluated numerically. But we can evaluate the integral for
the basic case of a European call option with strike-price K:

h(s) = (s − K)+ .

Then
∞ 1 2
e− 2 x
Z
−r(T −t) 1 1
F (t, x) = e √ [s exp{(r − σ 2 )(T − t) + σ(T − t) 2 x} − K]+ dx.
−∞ 2π 2

We have already evaluated such integrals in Chapter V, where we obtained

the BS formula from the binomial model by a passage to the limit. Complet-
ing the square in the exponential as before gives the result, as in V.6. //

10
4. Related results.

The Black-Scholes PDE and the Feynman-Kac formula.

The original proof of the Black-Scholes formula was via the Black-Scholes
PDE; this approach is closely linked with the Feynman-Kac formula, which
links PDEs with SDEs. For details, see the website for last year’s course.
Risk-neutral measure.
We call P ∗ the risk-neutral probability measure. It is equivalent to P
(by Girsanov’s Theorem, which gives the Radon-Nikodym derivative show-
ing equivalence), and is a martingale measure (as the discounted asset prices
are P ∗ -martingales, by above), i.e. P ∗ (or Q) is the equivalent martingale
measure (EMM).
Fundamental Theorem of Asset Pricing (FTAP). The above continuous-time
result may be summarised just as the FTAP in discrete time: to get the
no-arbitrage price of a contingent claim, take the discounted expected value
under the equivalent mg (risk-neutral) measure.
Completeness.
In discrete time, we saw that absence of arbitrage corresponded to exis-
tence of risk-neutral measures, completeness to uniqueness. We have obtained
existence and uniqueness here (and so completeness), by appealing to Gir-
sanov’s Theorem, which we have not proved in full. Completeness questions
are linked to the Representation Theorem for Brownian Martingales, below.

Theorem (Representation Theorem for Brownian Martingales). Let

(Mt : 0 ≤ t ≤ T ) be a square-integrable martingale with respect to the
Brownian filtrationR (Ft ). Then there exists an adapted process H = (Ht :
0 ≤ t ≤ T ) with E Hs2 ds < ∞ such that
Z t
Mt = M0 + Hs dWs , 0 ≤ t ≤ T.
0

That is, all Brownian martingales may be represented as stochastic integrals

with respect to Brownian motion.

We refer to, e.g., [KS], [RY] for proof.

The economic relevance of the Representation Theorem is that it shows
(see e.g. [KS, I.6], and below) that the Black-Scholes model is complete – that

11
is, that EMMs are unique, and so that Black-Scholes prices are unique (we
know this already, from FTAP/RNVF above). Mathematically, the result is
purely a consequence of properties of the Brownian filtration. The desirable
mathematical properties of BM are thus seen to have hidden within them
desirable economic and financial consequences of real practical value.
Hedging.
To find a hedging strategy H = (Ht0 , Ht ) (Ht0 for cash, Ht for stock) that
replicates the value process V = (Vt ), itself given by RNVF (VII.3):

Vt = Ht0 + Ht St = E ∗ [e−r(T −t) h|Ft ].

Now
Mt := E ∗ [e−rT h|Ft ]
is a martingale (indeed, a uniformly integrable mg: IV.4, V.2) under the
filtration Ft , that of the driving BM in (GBM ) (VII.1, VII.2), and the fil-
tration is unchanged by the Girsanov change of measure (we quote this).
So by the Representation Theorem for Brownian Martingales, there is some
adapted process K = (Kt ) with
Z t
Mt = M0 + Ks dWs (t ∈ [0, T ]).
0

Take
Ht := Kt /(σ S̃t ), Ht0 := Mt − Ht S̃t .
Then
Kt
dMt = Kt dWt = .σ S̃t dWt = Ht dS̃t ,
σ S̃t
and the strategy given by K is self-financing, by VII.2. This is of limited
practical value:
(a) the Representation Theorem does not give K = (Kt ) explicitly – it is
merely an existence proof;
(b) we already know that, as Brownian paths have infinite variation, exact
hedging in the Black-Scholes model is too rough to be practically possible.
So to hedge in practice, we need to go back to discrete time, where we
can compute things and where such roughness questions do not arise. But
this is familiar by now (and is why we have Chapters IV, V in discrete time
and Chapters VI, VII in continuous time). We need to go back and forth at
will between continuous time – where we can do calculus, in particular, Itô

12
calculus – and discrete time – where we can calculate, using computers.
Comments on the Black-Scholes formula.
1. The Black-Scholes formula transformed the financial world. Before it (see
Ch. II), the expert view was that asking what an option is worth was (in
effect) a silly question: the answer would necessarily depend on the attitude
to risk of the individual considering buying the option. It turned out that –
at least approximately (i.e., subject to the restrictions to perfect – frictionless
– markets, including No Arbitrage – an over-simplification of reality) there
is an option value. One can see this in one’s head, without doing any math-
ematics, if one knows that the Black-Scholes market is complete (above). So,
every contingent claim (option, etc.) can be replicated, by a suitable com-
bination of cash and stock. Anyone can price this: (i) count the cash, and
count the stock; (ii) look up the current stock price; (iii) do the arithmetic.
2. The programmable pocket calculator was becoming available around this
time. Every trader immediately got one, and programmed it, so that he
could price an option (using the BS model!) in real time, from market data.
3. The missing quantity in the Black-Scholes formula is the volatility, σ. But,
the price is continuous and strictly increasing in σ (options like volatility!).
So there is exactly one value of σ that gives the price at which options are
being currently traded. This – the implied volatility – is the value that the
market currently judges σ to be, and the one that traders use.
4. Because the Black-Scholes model is the benchmark model of mathematical
finance, and gives a value for σ at the push of a button, it is widely used.
5. This is despite the fact that no one actually believes the Black-Scholes
model! It is an over-simplified approximation to reality. Indeed, Fischer
Black himself famously once wrote a paper called The holes in Black-Scholes.
6. This is an interesting example of theory and practice interacting!
7. Black and Scholes has considerable difficulty in getting their paper pub-
lished! It was ahead of its time. When published, and its importance under-
stood, it changed its times.
8. Black-Scholes theory and its developments, plus the internet (a global
network of fibre-optic cables – using photons rather than electrons), were
important contributory factors to globalization. Enormous sums of money
can be transported round the world at the push of a button, and are every
day. This has led to financial contagion – ”one country’s economic problem
becomes the world’s economic problem”. (The Ebola virus comes to mind
here.) The resulting problems of systemic stability are very important, and
still largely unsolved; they dominate the agenda at international meetings.

13
§5. Infinite time-horizon; American puts

We sketch here the theory of the American option (one can exercise at
any time), over an infinite time-horizon. We deal first with a put option (see
VII.6 below under Real options for the corresponding ‘call option’) – giving
the right to sell at the strike price K, at any time τ of our choosing. This
τ has to be a stopping time: we have to take the decision whether or not to
stop at τ based on information already available – no access to the future,
no insider trading. As above, we pass to the risk-neutral measure.
Recall that for American puts over a finite time-horizon, we used the
Snell envelope, which is also the least supermartingale majorant. Now for the
Snell envelope, we need the finite time-horizon T to begin, as we build up the
Snell envelope by working backwards in time from T – dynamic programming.
Although this does not apply when T = ∞, the least supermartingale majo-
rant does. We quote this. We refer for the theory here, and for the pricing
argument below, to the standard work on this subject, Peskir & Shiryaev:
[PS] G. PESKIR and A. N. SHIRYAEV: Optimal stopping and free-boundary
problems. Birkhäuser, 2006.
Under the risk-neutral measure, the SDE for GBM becomes
dXt = rXt dt + σXt dBt . (GBMr )
To evaluate the option, we have to solve the optimal stopping problem
V (x) := sup Ex [e−rτ (K − Xτ )+ ]
τ

where the sup is taken over all stopping times τ and X0 = x under Px .
The process X satisfying (GBMr ) – a diffusion – is specified by a second-
order linear differential operator, called its (infinitesimal) generator,
1
LX := rxD + σ 2 x2 D2 , D := ∂/∂x.
2
Now the closer X gets to 0, the less likely we are to gain by continuing. This
suggests that our best strategy is to stop when X gets too small: to stop at
τ = τb , where
τb := inf{t ≥ 0 : Xt ≤ b},
for some b ∈ (0, K) (the only range in which we would want to exercise an
option to sell at K). This gives the following free boundary problem for the

14
unknown value function V (x) and the unknown point b ∈ (0, K):

LX V = rV for x > b; (i)

V (x) = (K − x)+ = K − x for x = b; (ii)
V 0 (x) = −1 for x = b (smooth fit); (iii)
V (x) > (K − x)+ for x > b; (iv)
V (x) = (K − x)+ for 0 < x < b. (v)
Writing d := σ 2 /2 (‘d for diffusion’), (i) is

dx2 V 00 + rxV 0 − rV = 0. (i∗ )

This ODE is homogeneous. So (Euler’s theorem): use trial solution:

V (x) = xp .

Substituting gives a quadratic for p:

r r
p2 − (1 − )p − = 0.
d d
Trial solution: V (x) = xp . Substituting gives a quadratic for p:
r r
p2 − (1 − )p − = 0 : (p − 1)(p + r/d) = 0.
d d
One root is p = 1; the other is p = −r/d. So the general solution is V (x) =
C1 x + C2 x−r/d . But V (x) ≤ K for all x ≥ 0 (an option giving the right to
sell at price K cannot be worth more than K!): V (x) is bounded. Taking x
large (x < b is covered by (v)), we must have C1 = 0. So (with C := C2 )
r
V (x) = Cx−r/d ; V 0 (x) = − .C −r/d−1 . (∗)
d
From (ii),
Cb−r/d = K − b,
while from (iii),

r 1 bd
− .Cb−r/d . = −1 : Cb−r/d = .
d b r

15
Equating, this gives C and b:
bd
= K − b, K = b(1 + d/r), b = K/(1 + d/r).
r
Then (∗) and (iii) give
d  K 1+r/d
C= .
r 1 + d/r
So
d  K 1+r/d −r/d
V (x) = x if x ∈ [b, ∞)
r 1 + d/r
= K −x if x ∈ (0, b].

This is in fact the full and correct solution to the problem; see [PS], §25.1.
The ‘smooth fit’ in (iii) is characteristic of free boundary problems. For
a heuristic analogy: imagine trying to determine the shape of a rope, tied
to the ground on one side of a convex body, stretched over the body, then
pulled tight and tied to the ground on the other side. We can see on physical
grounds that the rope will be:
straight to the left of the convex body;
continuously in contact with the body for a while, then
straight to the right of the body, and
there should be no kink in the rope at the points where it makes and then
leaves contact with the body. This corresponds to ‘smooth fit’ in (iii).

6. Real options (Investment options).

For background and details, see e.g. Peskir & Shiryaev [PS], and
[DP] Avinash K. DIXIT and Robert S. PINDYCK: Investment under uncer-
tainty. Princeton University press, 1994.
The options considered above concern financial derivatives (so called be-
cause they derive from the underlying fundamentals such as stock). We turn
now to options of another kind, concerned with business decision-making.
Typically, we shall be concerned with the decision of whether or not to make
a particular investment, and if so, when. Because these options concern the
real economy (of manufacturing, etc.) rather than financial markets such as
the stock market, such options are often called real options. But because they
typically concern investment decisions, they are also often called investment

16
options. There is a good introductory treatment in [DP].
The key features are as follows. We are contemplating making some ma-
jor investment – buying or building a factory, drilling an oil well, etc. While
if the decision goes wrong it may be possible to recoup some of the cost,
much or most of it will usually be irrecoverable (a sunk cost – as with an oil
well). So the investment is irreversible – at least in part. Just as stock prices
are uncertain – so we model them as random, using some stochastic process
– here too, the future profitability of the proposed investment is uncertain.
Finally, we do not have to act now, or indeed at all. So we have an open-
ended – or infinite – time-horizon, T = ∞.
We may choose to delay investment,
(a) to gather more information, to help us assess the project, or
(b) to continue to generate interest on the capital we propose to invest.
So we must recognize, and feed into the decision process, the value of waiting
for further information. When we commit ourselves and make the decision
to invest, it is not just the sunk cost that we lose – we lose the valuable
option to wait for new information.
This situation is really that of an American call option with an infinite
time-horizon. With such an American call, we have the right to buy at a
specified price at a time of our choosing (or indeed, not to buy). Following
Dixit & Pindyck [DP, Ch. 5], we formulate an optimal stopping problem, and
solve it as a free boundary problem, using the principle of smooth fit.
We suppose the cost of the investment is I, and that the value of the
project is given by a GBM, X = (Xt ) ∼ GBM (µ, σ) (the value of a project
is uncertain for the same reasons that stock prices are uncertain; we model
them both as stochastic processes; GBM is the default option here, just as
in the BS theory of Ch. IV). If we invest at time τ , we want to maximize

V (X) := max E[(Xτ − I)e−rτ ],

τ

with r the riskless rate (discount rate) as before. Now if µ ≤ 0 the value
of the project will fall, so we should invest immediately if X0 > I and not
invest if not. If µ > r, the growth of X will swamp the investment cost I
and more than offset the discounting, so we should invest and there is no
point in waiting. So we take µ ∈ (0, r]. We invest iff the value x∗ at the time
of investment is large enough; finding x∗ is part of the problem; x∗ is a free
boundary (between the continuation region and the investment region).

17
 We need the following four conditions:
1 2 2 00
σ x V (x) + µxV 0 (x) − rV = 0, (i)
2
V (0) = 0, (ii)
∗ ∗
V (x ) = x − I, (iii)
V 0 (x∗ ) = 1 (smooth pasting). (iv)
For (i): this comes from the generator of the diffusion GBM (r, σ) (cf. the
SDE for GBM (r, σ), and Black-Scholes PDE, VI.2); for details, see [DP
Ch. 5], or Peskir & Shiryaev [PS, Ch. III]. For (ii) (”Nothing will come of
nothing”): the GBM does not hit 0, but if it approaches 0, so will the value
of the project, so (ii) follows from this by continuity). For (iii), this is the
value-matching condition: on investment, the firm receives the net pay-off
x∗ − I. For (iv) (smooth pasting): think of a rope stretched tightly over a
convex surface.
Again, the ODE (i) is homogeneous (cf. Euler’s theorem). So we use
a trial solution V (x) = Cxp . So (i) gives that p satisfies the fundamental
quadratic
1
Q(p) := σ 2 p(p − 1) + µp − r = 0.
2
The product of the roots is negative, and Q(0) = −r < 0, Q(1) = µ − r < 0.
So one root p1 > 1 and the other p2 < 0. The general solution is V (x) =
C1 xp1 + C2 xp2 , but from V (0) = 0, C2 = 0, so V (x) = C1 xp1 , = Cxp1 say.
With x∗ the critical value at which it is optimal to invest, (iii) and (iv) give

V (x∗ ) = x∗ − I, V 0 (x∗ ) = 1.

From these two equations, we can find C and x∗ . The second is

V 0 (x∗ ) = Cp1 (x∗ )p1 −1 = 1, C = (x∗ )1−p1 /p1 .

Then the first gives

p1
C(x∗ )p1 = x∗ − I, x∗ /p1 = x∗ − I, x∗ = I.
(p1 − 1)
The main feature here is the factor

q := p1 /(p1 − 1) > 1

18
by which the value must exceed the investment cost I before investment
should be made (q is used because this is related to ”Tobin’s q” in Economics).
One can check that q increases with σ (the riskier the project, the more
reluctant we are to invest), and also q increases with r (as then investing
our capital risklessly becomes more attractive). Then the critical threshold
above which it is optimal to invest is

x∗ = qI.

Also
C = (qI)1−p1 /p1 , V (x) = (qI)1−p1 xp1 /p1 .
The results above show that the traditional net present value (NPV –
accountancy-based) approach to valuing real options is misleading – see [DP].
This is no surprise: our methods (arbitrage pricing technique, etc.) are su-
perior to NPV!

7. Stochastic volatility (SV).

The Black-Scholes theory above – in discrete or continuous time – has
involved the volatility – the parameter that describes the sensitivity of the
stock price to new information, to the market’s assessment of new infor-
mation. Volatility is so important that it has been subjected to intensive
scrutiny, in the light of much real market data. Alas, such detailed scrutiny
reveals that volatility is not really constant at all – the Black-Scholes theory
over-simplifies reality. (This is hardly surprising: real financial markets are
more complicated than the contents of this course, as they involve investor
psychology, rather than straight mathematics!) One way out is to admit that
volatility is random (stochastic), and then try to model the stochastic pro-
cess generating it. Volatility exhibits clustering, linked to mean reversion, so
Ornstein-Uhlenbeck models are useful here. Such stochastic volatility models
are topical today.
Stylised facts.
There are a number of stylised facts in mathematical finance. E.g.:
(i). Financial data show skewness. This is a result of the asymmetry between
profit and loss (large losses are lethal!; large profits are just nice to have).
(ii). Financial data have much fatter tails than the normal (Gaussian). We
have discussed this in II.5.
(iii) Financial data show volatility clustering. This is a result of the economic
and financial environment, which is extremely complex, and which moves

19
between good times/booms/upswings and bad times/slumps/downswings.
Typically, the market ‘gets stuck’, staying in its current state for longer than
is objectively justified, and then over-correcting. As investors are highly
sensitive to losses (see (i) above), downturns cause widespread nervousness,
which is reflected in higher volatility. The upshot is that good times are as-
sociated with periods of growth but low volatility; downturns spark extended
periods of high volatility (and economic stagnation, or shrinkage).
ARCH and GARCH.
We turn to models that can incorporate such features.
The model equations are (with Zt ind. N (0, 1))
p
X
Xt = σt Zt , σt2 = α0 + 2
αi Xi−1 , (ARCH(p))
1

while in GARCH(p, q) the σt2 term becomes

p q
X X
σt2 = α0 + 2
αi Xi−1 + 2
βj σt−j . (GARCH(p, q))
1 1

The names stand for (generalised) autoregressive conditionally heteroscedas-

tic (= variable variance). These are widely used in Econometrics, to model
volatility clustering – the common tendency for periods of high volatility, or
variability, to cluster together in time. They were introduced in 1987 by
Robert Engle (1942) and C. W. J. (Sir Clive) Granger (1934-2009), who re-
ceived the Nobel Prize for this in 2003. From Granger’s obituary (The Times,
1.6.2009): ”Following Granger’s arrival at UCSD in La Jolla, he began the
work with his colleague Robert F. Engle for which he is most famous, and for
which they received the Bank of Sweden Nobel Memorial Prize in Economic
Sciences in 2003. They developed in 1987 the concept of cointegration. Coin-
tegrated series are series that tend to move together, and commonly occur in
economics. Engle and Granger gave the example of the price of tomatoes in
N. and S. Carolina .... Cointegration may be used to reduce non-stationary
situations to stationary ones, which are much easier to handle statistically
and so to make predictions for. This is a matter of great economic impor-
tance, as most macroeconomic time series are non-stationary, so temporary
disturbances in, say, GDP may have a long-lasting effect, and so a permanent
economic cost. The Engle-Granger approach helps to separate out short-term
effects, which are random and unpredictable, from long-term effects, which

20
reflect the underlying economics. This is invaluable for macroeconomic pol-
icy formulation, on matters such as interest rates, exchange rates, and the
relationship between incomes and consumption.”

Volatility Modelling
In the standard Black-Scholes theory we have developed, volatility σ is
constant. Thus a graph of volatility against strike K (or stock price S) should
be flat. But typically it isn’t, and displays curvature. Such volatility curves
often turn upwards at both ends (‘volatility smile’); there may well be asym-
metry (‘volatility smirk’).
As above, it may be useful to model volatility stochastically, and use an
SV model. However, the driving noise in this model will have a volatility of
its own (‘vol of vol’), etc. Practitioners often use computer graphics to repre-
sent volatility surfaces – the three-dimensional equivalents of graphs, where
e.g. σ is graphed against K and S. The subject is too big to pursue further
here; there is a good account (mixing theory with practice) in
J. GATHERAL: The volatility surface: A practitioner’s guide. Wiley 2006.
Volatility is rough.
This is the title of an influential paper by Gatheral, Jaisson and Rosen-
baum in 2014. The message there is that (log-)volatility is not only rough, it
is rougher than Brownian motion. The reasons are the obvious ones: high-
frequency trading, and order splitting. There is a family, fractional Brownian
motion, with a parameter controlling the roughness (called the Hurst index):
BM is ‘in the middle’. This is highly topical today: there is a lot going on
in this area.

21
Postscript.
1. One recent book on Financial Mathematics describes the subject as being
composed of three strands:
arbitrage – the core economic concept, which we have used throughout;
martingales – the key probabilistic concept (Ch. III on);
numerics. Finance houses in the City use models, which they need to cali-
brate to data – a task involving both statistical and numerical skills, and in
particular an ability to programme.
2. You will probably already have experience with at least one general/mathematical
programming language (e.g., Matlab, Python) (if not: get it, a.s.a.p.!), and
for Statistics, R. You may also know some Numerical Analysis, the theory
behind computation. You may have encountered simulation, also known as
Monte Carlo, and/or a branch of Probability and Statistics called Markov
Chain Monte Carlo (MCMC) – computer-intensive methods for numerical
solutions to problems too complicated to solve analytically. The leaders of R
& D teams in the City need to be expert at both stochastic modelling (e.g.,
to propose new products), and simulation (to evaluate how these perform).
Most of the ones I know use Matlab for this. At a lower level, quantitative
analysts (quants) working under them need expertise in a computer language;
C++ is the industry standard. If you are thinking of a career in Mathemat-
ical Finance, learn C++, as soon as possible, and for academic credit.
3. This course deals with equity markets – with stocks, and financial deriva-
tives of them – options on stocks, etc. The relevant mathematics is finite-
dimensional. Lurking in the background are bond markets (‘money markets’:
bonds, gilts etc., where interest rates dominate), and the relevant options –
interest-rate derivatives, and foreign exchange between different currencies
(‘forex’). The resulting mathematics (which is highly topical, and so in great
demand in the City!) is infinite-dimensional, and so much harder than the
equity-market theory we have done. However, the underlying principles are
basically the same. One has to learn to walk before one learns to run, and
equity markets serve as a preparation for money markets.
4. The aim of this lecture course is simple. It is to familiarize the student
with the basics of Black-Scholes theory, as the core of modern finance, and
with the mathematics necessary to understand this. The motivation driving
the ever-increasing study of this material is the financial services industry
and the City. I hope that any of you who seek City careers will find this
introduction to the subject useful in later life. NHB, 2017

22