m3f33chV

Chapter V: MATHEMATICAL FINANCE IN DISCRETE TIME

We follow [BK], Ch. 4 (or see the other sources cited in Ch0).

§1. The Model.

It suffices, to illustrate the ideas, to work with a finite probability space
(Ω, F, P ), with a finite number |Ω| of points ω, each with positive probabil-
ity: P ({ω}) > 0. We will use a finite time-horizon N , which will correspond
to the expiry date of the options.
As before, we use a filtration F0 ⊂ F1 ⊂ · · · ⊂ FN : we may (and shall)
take F0 = {∅, Ω}, the trivial σ-field, FN = F = P(Ω) (here P(Ω) is the
power-set of Ω, the class of all 2|Ω| subsets of Ω: we need every possible sub-
set, as they all (apart from the empty set) carry positive probability.
The financial market contains d + 1 financial assets: a riskless asset (bank
account) labelled 0, and d risky assets (stocks, say) labelled 1 to d. The prices
of the assets at time n are random variables, Sn0 , Sn1 , · · · , Snd say [note that
we use superscripts here as labels, not powers, and suppress ω for brevity],
non-negative and Fn -measurable [at time n, we know the prices Sni ].
We take S00 = 1 (that is, we reckon in units of our initial bank holding).
We assume for convenience a constant interest rate r > 0 in the bank, so 1
unit in the bank at time 0 grows to (1 + r)n at time n. So 1/(1 + r)n is the
discount factor at time n.

Definition. A trading strategy H is a vector stochastic process H = (Hn )N n=0 =

0 1 d N i
((Hn , Hn , · · · , Hn ))n=0 which is predictable (or previsible): each Hn is Fn−1 -
measurable for n ≥ 1.

Here Hni denotes the number of shares of asset i held in the portfolio at
time n – to be determined on the basis of information available before time
n; the vector Hn = (Hn0 , Hn1 , · · · , Hnd ) is the portfolio at time n. Writing
Sn = (Sn0 , Sn1 , · · · , Snd ) for the vector of prices at time n, the value of the
portfolio at time n is the scalar product

Vn (H) = Hn .Sn := Σdi=0 Hni Sni .

1
The discounted value is

Ṽn (H) = βn (Hn .Sn ) = Hn .S̃n ,

where βn := 1/Sn0 and S̃n = (1, βn Sn1 , · · · , βn Snd ) is the vector of discounted
prices.
Note. The previsibility of H reflects that there is no insider trading.

Definition. The strategy H is self-financing (SF), H ∈ SF , if

Hn .Sn = Hn+1 .Sn (n = 0, 1, · · · , N − 1).

Interpretation. When new prices Sn are quoted at time n, the investor adjusts
his portfolio from Hn to Hn+1 , without bringing in or consuming any wealth.

Vn+1 (H) − Vn (H) = Hn+1 .Sn+1 − Hn .Sn

= Hn+1 .(Sn+1 − Sn ) + (Hn+1 .Sn − Hn .Sn ).

For a SF strategy, the second term on the right is zero. Then the LHS, the
net increase in the value of the portfolio, is shown as due only to the price
changes Sn+1 − Sn . So for H ∈ SF ,

Vn (H) − Vn−1 (H) = Hn (Sn − Sn−1 ),

∆Vn (H) = Hn .∆Sn , Vn (H) = V0 (H) + Σn1 Hj .∆Sj

and Vn (H) is the martingale transform of S by H (IV.6). Similarly with
discounting:

∆Ṽn (H) = Hn .∆S̃n , Ṽn (H) = V0 (H) + Σn1 Hj .∆S̃j

(∆S̃n := S̃n − S̃n−1 = βn Sn − βn−1 Sn−1 ).

As in II, we are allowed to borrow (so Sn0 may be negative) and sell short
(so Sni may be negative for i = 1, · · · , d). So it is hardly surprising that if we
decide what to do about the risky assets, the bank account will take care of
itself, in the following sense.

Proposition. If ((Hn1 , · · · , Hnd ))N

n=0 is predictable and V0 is F0 -measurable,
there is a unique predictable process (Hn0 )N 0 1 d
n=0 such that H = (H , H , · · · , H )
is SF with initial value V0 .

2
Proof. If H is SF, then as above

Ṽn (H) = Hn .S̃n = Hn0 + Hn1 S̃n1 + · · · + Hnd S̃nd ,

while as Ṽn = H.S̃n ,

Ṽn (H) = V0 + Σn1 (Hj1 ∆S̃j1 + · · · + Hjd ∆S̃jd )

(S̃n = (1, βn Sn1 , · · · , βn Snd ), so S̃n0 ≡ 1, ∆S̃n0 = 0). Equate these:

Hn0 = V0 + Σn1 (Hj1 ∆S̃j1 + · · · + Hjd ∆S̃jd ) − (Hn1 S̃n1 + · · · + Hnd S̃nd ),

which defines Hn0 uniquely. The terms in S̃ni are Hni ∆S̃ni − Hni S̃ni = −Hni S̃n−1
i
,
which is Fn−1 -measurable. So

Hn0 = V0 + Σn−1
1 (Hj1 ∆S̃j1 + · · · + Hjd ∆S̃jd ) − (Hn1 Sn−1
1
+ · · · + Hnd S̃n−1
d
),

where as H 1 , · · · , H d are predictable, all terms on the RHS are Fn−1 -measurable,
so H 0 is predictable. //

Numéraire. What units do we reckon value in? All that is really necessary is
that our chosen unit of account should always be positive (as we then reckon
our holdings by dividing by it, and one cannot divide by zero). Common
choices are pounds sterling (UK), dollars (US), euros etc. Gold is also pos-
sible (now priced in sterling etc. – but the pound sterling represented an
amount of gold, till the UK ‘went off the gold standard’). By contrast, risky
stocks can have value 0 (if the company goes bankrupt). We call such an
always-positive asset, used to reckon values in, a numéraire.
Of course, one has to be able to change numéraire – e.g. when going
from UK to the US or eurozone. As one would expect, this changes nothing
important. In particular, we quote (numéraire invariance theorem – see e.g.
[BK] Prop. 4.1.1) that the set SF of self-financing strategies is invariant un-
der change of numéraire.
Note. 1. This alerts us to what is meant by ‘risky’. To the owner of a gold-
mine, sterling is risky. The danger is not that the UK government might go
bankrupt, but that sterling might depreciate against the dollar, or euro, etc.
2. With this understood, we shall feel free to refer to our numéraire as ‘bank
account’. The point is that we don’t trade in it (why would a goldmine owner
trade in gold?); it is the other – ‘risky’ – assets that we trade in.

3
§2. Viability (NA): Existence of Equivalent Martingale Measures.

Although we are allowed to borrow (from the bank), and sell (stocks)
short, we are – naturally – required to stay solvent (recall that trading while
insolvent is an offence under the Companies Act!).

Definition. A strategy H is admissible if it is self-financing (SF), and

Vn (H) ≥ 0 for each time n = 0, 1, · · · , N .

Recall that arbitrage is riskless profit – making ‘something out of noth-

ing’. Formally:

Definition. An arbitrage strategy is an admissible strategy with zero initial

value and positive probability of a positive final value.
Definition. A market is viable if no arbitrage is possible, i.e. if the market
is arbitrage-free (no-arbitrage, NA).
This leads to the first of two fundamental results:

Theorem (No-Arbitrage Theorem: NA iff EMMs exist). The market

is viable (is arbitrage-free, is NA) iff there exists a probability measure P ∗
equivalent to P (i.e., having the same null sets) under which the discounted
asset prices are P ∗ -martingales – that is, iff there exists an equivalent mar-
tingale measure (EMM).

Proof. ⇐. Assume such a P ∗ exists. For any self-financing strategy H, we

have as before
Ṽn (H) = V0 (H) + Σn1 Hj .∆S̃j .
By the Martingale Transform Lemma, S̃j a (vector) P ∗ -martingale implies
Ṽn (H) is a P ∗ -martingale. So the initial and final P ∗ -expectations are the
same: using E ∗ for P ∗ -expectation,

E ∗ [ṼN (H)] = E ∗ [Ṽ0 (H)].

If the strategy is admissible and its initial value – the RHS above – is zero,
the LHS E ∗ [ṼN (H)] is zero, but ṼN (H) ≥ 0 (by admissibility). Since each
P ({ω}) > 0 (by assumption), each P ∗ ({ω}) > 0 (by equivalence). This and
ṼN (H) ≥ 0 force ṼN (H) = 0 (sum of non-negatives can only be 0 if each

4
term is 0). So no arbitrage is possible. //

The converse is true, but harder, and needs a preparatory result – which
is interesting and important in its own right.
Separating Hyperplane Theorem (SHT).
In a vector space V , a hyperplane is a translate of a (vector) subspace
U of codimension 1 – that is, U and some one-dimensional subspace, say R,
together span V : V is the direct sum V = U ⊕ R (e.g., R3 = R2 ⊕ R). Then
H = [f, α] := {x : f (x) = α}
for some α and linear functional f . In the finite-dimensional case, of dimen-
sion n, say, one can think of f (x) as an inner product,
f (x) = f.x = f1 x1 + . . . + fn xn .
The hyperplane H = [f, α] separates sets A, B ⊂ V if
f (x) ≥ α ∀ x ∈ A, f (x) ≤ α ∀x∈B
(or the same inequalities with A, B, or ≥, ≤, interchanged).
Call a set A in a vector space V convex if
x, y ∈ A, 0≤λ≤1 ⇒ λx + (1 − λ)y ∈ A
– that is, A contains the line-segment joining any pair of its points.
We can now state (without proof) the SHT (see e,g, [BK] App. C).
SHT. Any two non-empty disjoint convex sets in a vector space can be sep-
arated by a hyperplane.
A cone is a subset of a vector space closed under vector addition and
multiplication by positive constants (so: like a vector subspace, but with a
sign-restriction in scalar multiplication).
We turn now to the proof of the converse.

Proof of the converse (not examinable). ⇒: Write Γ for the cone of strictly
positive random variables. Viability (NA) says that for any admissible strat-
egy H,
V0 (H) = 0 ⇒ ṼN (H) ∈
/ Γ. (∗)
To any admissible process (Hn1 , · · · , Hnd ), we associate its discounted cu-
mulative gain process
G̃n (H) := Σn1 (Hj1 ∆S̃j1 + · · · + Hjd ∆S̃jd ).

5
By the Proposition, we can extend (H1 , · · · , Hd ) to a unique predictable pro-
cess (Hn0 ) such that the strategy H = ((Hn0 , Hn1 , · · · , Hnd )) is self-financing
with initial value zero. By NA, G̃N (H) = 0 – that is, G̃N (H) ∈ / Γ.
We now form the set V of random variables G̃N (H), with H = (H 1 , · · · , H d )
a previsible process. This is a vector subspace of the vector space RΩ of ran-
dom variables on Ω, by linearity of the gain process G(H) in H. By (∗), this
subspace V does not meet Γ. So V does not meet the subset

K := {X ∈ Γ : Σω X(ω) = 1}.

Now K is a convex set not meeting the origin. By the Separating Hyper-
plane Theorem, there is a vector λ = (λ(ω) : ω ∈ Ω) such that for all X ∈ K

λ.X := Σω λ(ω)X(ω) > 0, (1)

but for all G̃N (H) in V,

λ.G̃N (H) = Σω λ(ω)G̃N (H)(ω) = 0. (2)

Choosing each ω ∈ Ω successively and taking X to be 1 on this ω and zero

elsewhere, (1) tells us that each λ(ω) > 0. So

P ∗ ({ω}) := λ(ω)/(Σω0 ∈Ω λ(ω 0 ))

defines a probability measure equivalent to P (no non-empty null sets). With

E ∗ as P ∗ -expectation, (2) says that

E ∗ [G̃N (H)] = 0 : E ∗ [ΣN

1 Hj .∆S̃j ] = 0.

In particular, choosing for each i to hold only stock i,

E ∗ [ΣN i i
1 Hj ∆S̃j ] = 0 (i = 1, · · · , d).

By the Martingale Transform Lemma, this says that the discounted price
processes (S̃ni ) are P ∗ -martingales. //

§3. Complete Markets: Uniqueness of EMMs.

A contingent claim (option, etc.) can be defined by its payoff function, h
say, which should be non-negative (options confer rights, not obligations, so
negative values are avoided by not exercising the option), and FN -measurable

6
(so that we know how to evaluate h at the terminal time N ).

Definition. A contingent claim defined by the payoff function h is attain-

able if there is an admissible strategy worth (i.e., replicating) h at time N .
A market is complete if every contingent claim is attainable.

Theorem (Completeness Theorem: complete iff EMM unique). A

viable market is complete iff there exists a unique probability measure P ∗
equivalent to P under which discounted asset prices are martingales – that
is, iff equivalent martingale measures are unique.

Proof. ⇒: Assume viability and completeness. Then for any FN -measurable

random variable h ≥ 0, there exists an admissible (so SF) strategy H repli-
cating h: h = VN (H). As H is SF, by §1
0
h/SN = ṼN (H) = V0 (H) + ΣN
1 Hj .∆S̃j .

We know by the Theorem of §2 that an equivalent martingale measure

∗
P exists; we have to prove uniqueness. So, let P1 , P2 be two such equivalent
martingale measures. For i = 1, 2, (Ṽn (H))N
n=0 is a Pi -martingale. So,

Ei [ṼN (H)] = Ei [V0 (H)] = V0 (H),

since the value at time zero is non-random (F0 = {∅, Ω}). So

0 0
E1 [h/SN ] = E2 [h/SN ].

Since h is arbitrary, E1 , E2 have to agree on integrating all non-negative

integrands. Taking negatives and using linearity: they have to agree on non-
positive integrands also. Splitting an arbitrary integrand into its positive and
negative parts: they have to agree on all integrands. Now Ei is expectation
(i.e., integration) with respect to the measure Pi , and measures that agree
on integrating all integrands must coincide. So P1 = P2 . //

Before proving the converse, we prove a lemma. Recall that an admissible

strategy is a SF strategy with all values non-negative. The Lemma shows
that the non-negativity of contingent claims extends to all values of any SF
strategy replicating it – in other words, this gives equivalence of admissible
and SF replicating strategies. [SF: isolated from external wealth; admissible:

7
actually worth something. These sound similar; the Lemma shows they are
the same here. So we only need one term; we use SF as it is shorter, but
w.l.o.g. this means admissible also.]

Lemma. In a viable market, any attainable h (i.e., any h that can be repli-
cated by a SF strategy H) can also be replicated by an admissible strategy H.

Proof. If H is SF and P ∗ is an equivalent martingale measure under which

discounted prices S̃ are P ∗ -martingales (such P ∗ exist by viability and the
Theorem of §2), Ṽn (H) is also a P ∗ -martingale, being the martingale trans-
form of S̃ by H (see §1). So

Ṽn (H) = E ∗ [ṼN (H)|Fn ] (n = 0, 1, · · · , N ).

If H replicates h, VN (H) = h ≥ 0, so discounting, ṼN (H) ≥ 0, so the above

equation gives Ṽn (H) ≥ 0 for each n. Thus all the values at each time n are
non-negative – not just the final value at time N – so H is admissible. //

Proof of the Theorem (continued). ⇐ (not examinable): Assume the market

is viable but incomplete: then there exists a non-attainable h ≥ 0. By the
Proposition of §1, we may confine attention to the risky assets S 1 , · · · , S d ,
as these suffice to tell us how to handle the bank account S 0 .
Call Ṽ the set of random variables of the form

U0 + ΣN
1 Hn .∆S̃n

with U0 F0 -measurable (i.e. deterministic) and ((Hn1 , · · · , Hnd ))N

n=0 predictable;
this is a vector space. (Here (H 1 , . . . , H d ) extends to H := (H 0 , H 1 , . . . , H d ),
by the Proposition of §1, and H can be any strategy here.) Then as h is not
0
attainable, the discounted value h/SN does not belong to Ṽ, so Ṽ is a proper
subspace of the vector space R of all random variables on Ω. Let P ∗ be a
Ω

probability measure equivalent to P under which discounted prices are mar-

tingales (such P ∗ exist by viability, by the Theorem of §2). Define the scalar
product
(X, Y ) → E ∗ [XY ]
on random variables on Ω. Since Ṽ is a proper subspace, by Gram-Schmidt
orthogonalisation there exists a non-zero random variable X orthogonal to
Ṽ. That is,
E ∗ [X] = 0.

8
 Write kXk∞ := max{|X(ω)| : ω ∈ Ω}, and define P ∗∗ by
 X(ω)  ∗
P ∗∗ ({ω}) = 1 + P ({ω}).
2kXk∞
By construction, P ∗∗ is equivalent to P ∗ (same null-sets - actually, as P ∗ ∼ P
and P has no non-empty null-sets, neither do P ∗ , P ∗∗ ). As X is non-zero,
P ∗∗ and P ∗ are different. Now
 
E ∗∗ [ΣN
1 Hn .∆S̃ n ] = Σω P ∗∗
(ω) ΣN
1 Hn .∆S̃n (ω)
 X(ω)  ∗  N 
= Σω 1 + P (ω) Σ1 Hn .∆S̃n (ω).
2kXk∞

The ‘1’ term on the right gives E ∗ [ΣN1 Hn .∆S̃n ], which is zero since this is a
∗
martingale transform of the E -martingale S̃n . The ‘X’ term gives a multiple
of the inner product
(X, ΣN 1 Hn .∆S̃n ),

which is zero as X is orthogonal to Ṽ and ΣN

1 Hn .∆S̃n ∈ Ṽ. By the Martingale
∗∗
Transform Lemma, S̃n is a P -martingale since H (previsible) is arbitrary.
Thus P ∗∗ is a second equivalent martingale measure, different from P ∗ . So
incompleteness implies non-uniqueness of equivalent martingale measures. //

Martingale Representation. To say that every contingent claim can be repli-

cated means that every P ∗ -martingale (where P ∗ is the risk-neutral measure,
which is unique) can be written, or represented, as a martingale transform
(of the discounted prices) by the replicating (perfect-hedge) trading strategy
H. In stochastic-process language, this says that all P ∗ -martingales can be
represented as martingale transforms of discounted prices. Such Martingale
Representation Theorems hold much more generally, and are very important.
For the Brownian case, see VII and [RY], Ch. V.
Note. In the example of Chapter II, we saw that the simple option there could
be replicated. More generally, in our market set-up, all options can be repli-
cated – our market is complete. Similarly for the Black-Scholes theory below.

§4. The Fundamental Theorem of Asset Pricing; Risk-neutral val-

uation.
We summarise what we have learned so far. We call a measure P ∗ under
which discounted prices S̃n are P ∗ -martingales a martingale measure. Such

9
a P ∗ equivalent to the true probability measure P is called an equivalent
martingale measure. Then
1 (No-Arbitrage Theorem: §2). If the market is viable (arbitrage-free),
equivalent martingale measures P ∗ exist.
2 (Completeness Theorem: §3). If the market is complete (all contingent
claims can be replicated), equivalent martingale measures are unique. Com-
bining:

Theorem (Fundamental Theorem of Asset Pricing, FTAP). In a com-

plete viable market, there exists a unique equivalent martingale measure P ∗
(or Q).

Let h (≥ 0, FN -measurable) be any contingent claim, H an admissible

strategy replicating it:
VN (H) = h.
As Ṽn is the martingale transform of the P ∗ -martingale S̃n (by Hn ), Ṽn is a
P ∗ -martingale. So V0 (H)(= Ṽ0 (H)) = E ∗ [ṼN (H)]. Writing this out in full:

V0 (H) = E ∗ [h/SN
0
].

More generally, the same argument gives Ṽn (H) = Vn (H)/Sn0 = E ∗ [(h/SN
0
)|Fn ]:

h
Vn (H) = Sn0 E ∗ [ 0
|Fn ] (n = 0, 1, · · · , N ).
SN

It is natural to call V0 (H) above the value of the contingent claim h at

time 0, and Vn (H) above the value of h at time n. For, if an investor sells the
claim h at time n for Vn (H), he can follow strategy H to replicate h at time
N and clear the claim. To sell the claim for any other amount would provide
an arbitrage opportunity (as with the argument for put-call parity). So this
value Vn (H) is the arbitrage price (or more exactly, arbitrage-free price or
no-arbitrage price); an investor selling for this value is perfectly hedged.
We note that, to calculate prices as above, we need to know only
(i) Ω, the set of all possible states,
(ii) the σ-field F and the filtration (or information flow) (Fn ),
(iii) the EMM P ∗ (or Q).
We do NOT need to know the underlying probability measure P – only its
null sets, to know what ‘equivalent to P ’ means (actually, in this model, only

10
the empty set is null).
Now option pricing is our central task, and for pricing purposes P ∗ is
vital and P itself irrelevant. We thus may – and shall – focus attention on
P ∗ , which is called the risk-neutral probability measure. Risk-neutrality is
the central concept of the subject. The concept of risk-neutrality is due in
its modern form to Harrison and Pliska [HP] in 1981 – though the idea can
be traced back to actuarial practice much earlier. Harrison and Pliska call
P ∗ the reference measure; other names are risk-adjusted or martingale mea-
sure. The term ‘risk-neutral’ reflects the P ∗ -martingale property of the risky
assets, since martingales model fair games.
To summarise, we have the Risk-Neutral Valuation (or Pricing) Formula:

Theorem (Risk-Neutral Valuation Formula). In a complete viable mar-

ket, arbitrage-free prices of assets are their discounted expected values under
the risk-neutral (equivalent martingale) measure P ∗ (or Q). With payoff h,
Vn (H) = (1 + r)−(N −n) E ∗ [VN (H)|Fn ] = (1 + r)−(N −n) E ∗ [h|Fn ].

§5. European Options. The Discrete Black-Scholes Formula.

We consider the simplest case, the Cox-Ross-Rubinstein binomial model
of 1979; see [CR], [BK]. We take d = 1 for simplicity (one risky asset, one
bank account); the price vector is (Sn0 , Sn1 ), or ((1 + r)n , Sn ), where

Sn (1 + a) with probability p,
Sn+1 =
Sn (1 + b) with probability 1 − p
with −1 < a < b, S0 > 0. So writing N for the expiry time,
Ω = {1 + a, 1 + b}N ,
each ω ∈ Ω representing the successive values of Tn+1 := Sn+1 /Sn . The filtra-
tion is F0 = {∅, Ω} (trivial σ-field), FT = F = 2Ω (power-set of Ω: all sub-
sets of Ω), Fn = σ(S1 , · · · , Sn ) = σ(T1 , · · · , Tn ). For ω = (ω1 , · · · , ωN ) ∈ Ω,
P ({ω1 , · · · , ωN }) = P (T1 = ω1 , · · · , TN = ωN ), so knowing the pr. measure
P (i.e. knowing p) means we know the distribution of (T1 , · · · , TN ).
For p∗ ∈ (0, 1) to be determined, let P ∗ correspond to p∗ as P does to p.
Then the discounted price (S̃n ) is a P ∗ -martingale iff
E ∗ [S̃n+1 |Fn ] = S̃n , E ∗ [(S̃n+1 /S̃n )|Fn ] = 1, E ∗ [Tn+1 |Fn ] = 1 + r,

11
since Sn = S̃n (1 + r)n , Tn+1 = Sn+1 /Sn = (S̃n+1 /S̃n )(1 + r). But

E ∗ [Tn+1 |Fn ] = (1 + a).p∗ + (1 + b).(1 − p∗ )

is a weighted average of 1 + a and 1 + b; this can be 1 + r iff r ∈ [a, b]. As

P ∗ is to be equivalent to P and P has no non-empty null-sets, r = a, b are
excluded. Thus by §2:

Lemma. The market is viable (arbitrage-free) iff r ∈ (a, b).

Next, 1+r = (1+a)p∗ +(1+b)(1−p∗ ), r = ap∗ +b(1−p∗ ): r−b = p∗ (a−b):

Lemma. The equivalent mg measure exists, is unique, and is given by

p∗ = (b − r)/(b − a).

Corollary. The market is complete.

Now SN = Sn ΠN n+1 Ti . By the Fundamental Theorem of Asset Pricing,

the price Cn of a call option with strike-price K at time n is

Cn = (1 + r)−(N −n) E ∗ [(SN − K)+ |Fn ]

= (1 + r)−(N −n) E ∗ [(Sn ΠN
n+1 Ti − K)+ |Fn ].

Now the conditioning on Fn has no effect – on Sn as this is Fn -measurable

(known at time n), and on the Ti as these are independent of Fn . So

Cn = (1 + r)−(N −n) E ∗ [(Sn ΠN Ti − K)+ ]

 n+1 
−n N − n
= (1 + r)−(N −n) ΣN
j=0 p∗ j (1 − p∗ )N −n−j (Sn (1 + a)j (1 + b)N −n−j − K)+ ;
j

here j, N − n − j are the numbers of times Ti takes the two possible values
1 + a, 1 + b. This is the discrete Black-Scholes formula of Cox, Ross &
Rubinstein (1979) for pricing a European call option in the binomial model.
The European put is similar – or use put-call parity (II.3).
To find the (perfect-hedge) strategy for replicating this explicitly: write
 
N −n N − n
c(n, x) := Σj=0 p∗ j (1 − p∗ )N −n−j (x(1 + a)j (1 + b)N −n−j − K)+ .
j

12
Then c(n, x) is the undiscounted P ∗ -expectation of the call at time n given
that Sn = x. This must be the value of the portfolio at time n if the strategy
H = (Hn ) replicates the claim:

Hn0 (1 + r)n + Hn Sn = c(n, Sn )

(here by previsibility Hn0 and Hn are both functions of S0 , · · · , Sn−1 only).

Now Sn = Sn−1 Tn = Sn−1 (1 + a) or Sn−1 (1 + b), so:

Hn0 (1 + r)n + Hn Sn−1 (1 + a) = c(n, Sn−1 (1 + a))

Hn0 (1 + r)n + Hn Sn−1 (1 + b) = c(n, Sn−1 (1 + b)).

Subtract:

Hn Sn−1 (b − a) = c(n, Sn−1 (1 + b)) − c(n, Sn−1 (1 + a)).

So Hn in fact depends only on Sn−1 , Hn = Hn (Sn−1 ) (by previsibility), and

Proposition. The perfect hedging strategy Hn replicating the European

call option above is given by

c(n, Sn−1 (1 + b)) − c(n, Sn−1 (1 + a))

Hn = Hn (Sn−1 ) = .
Sn−1 (b − a)

Notice that the numerator is the difference of two values of c(n, x) with
the larger value of x in the first term (recall b > a). When the payoff function
c(n, x) is an increasing function of x, as for the European call option consid-
ered here, this is non-negative. In this case, the Proposition gives Hn ≥ 0:
the replicating strategy does not involve short-selling. We record this as:

Corollary. When the payoff function is a non-decreasing function of the

final asset price SN , the perfect-hedging strategy replicating the claim does
not involve short-selling of the risky asset.

§6. Continuous-Time Limit of the Binomial Model.

Suppose now that we wish to price an option in continuous time with
initial stock price S0 , strike price K and expiry T . We can use the work
above to give a discrete-time approximation, where N → ∞. We write

13
(temporarily) ρ ≥ 0 for the instantaneous interest rate in continuous time,
and define (again temporarily) r by
ρT N
r := ρT /N : eρT = limN →∞ (1 + ) = limN →∞ (1 + r)N .
N
Here r, which tends to zero as N → ∞, represents the interest rate in discrete
time for the approximating binomial model.
For σ > 0 fixed (σ 2 plays the role of a variance, corresponding in continuous
time to the volatility of the stock – below), define a, b (→ 0 as N → ∞) by
√ √
log((1 + a)/(1 + r)) = −σ/ N , log((1 + b)/(1 + r)) = σ/ N .

We now have a sequence of binomial models, for each of which we can price
options as in §5. We shall show that the pricing formula converges as N → ∞
to a limit. This is the famous Black-Scholes formula, the central result of
the course. We shall meet it later, and re-derive it, in continuous time, its
natural setting, in Ch. VII; see also e.g. [BK], 4.6.2. Fortunately, the con-
tinuous Black-Scholes formula is much neater than its discrete counterpart,
which involves the unwieldy binomial sum above.

Lemma. Let (XjN )N

j=1 be iid with mean µN satisfying

N µN → µ (N → ∞)

and variance σ 2 (1 + o(1))/N . If YN := ΣN N

1 Xj , then YN converges in distri-
bution to normality:

YN → Y = N (µ, σ) (N → ∞).

Proof. Use characteristic functions (CFs), I.4: since YN has mean and vari-
ance as given, it also has second moment σ 2 (1 + o(1))/N , so has CF

φN (u) := E exp{iuYN } = ΠN N N N
1 E exp{iuXj } = [E exp{iuX1 }]

iuµ 1 σ 2 u2 1 1
= (1 + − + o( ))N → exp{iuµ − σ 2 u2 } (N → ∞),
N 2 N N 2
the CF of the normal law N (µ, σ). Convergence of CFs implies convergence
in distribution by Lévy’s continuity theorem for CFs ([W], §18.1). //

14
 We can apply this to pricing the call option above (the details of the
calculation below, which are messy, are not hard, and not crucial; the point
is that this can be done):

(N ) ρT −N ∗
C0 = (1 + ) E [(S0 ΠN
1 Tn − K)+ ]
N
ρT −N
= E ∗ [(S0 exp{YN } − (1 + ) K)+ ], (1)
N
where XN
YN := log(Tn /(1 + r)).
1

Since Tn = TnN above takes values 1 + b, 1 + a, XnN := log(T N

√ n /(1 + r)) takes
values log((1 + b)/(1 + r)), log((1 + a)/(1 + r)) = ±σ/ N (so has second
moment σ 2 /N ). Its mean is
1 + b
∗
1 + a σ σ √
µN := log (1−p )+log p∗ = √ (1−p∗ )− √ p∗ = (1−2p∗ )σ/ N
1+r 1+r N N
√
(we shall see below that 1 − 2p∗ = O(1/ N ), so the Lemma will apply).
Now (recall r = ρT /N = O(1/N ))
√ √
a = (1 + r)e−σ/ N
− 1, b = (1 + r)eσ/ N
− 1,

so a, b, r → 0 as N → ∞, and
√
(b − r) [(1 + r)eσ/ N − 1 − r]
1 − 2p∗ = 1 − 2 =1−2 √ √
(b − a) [(1 + r)(eσ/ N − e−σ/ N )]
√
[eσ/ N
− 1]
=1−2 √ √ .
[eσ/ N − e−σ/ N ]
Now expand the two [· · · ] terms above by Taylor’s theorem: they give

σ 1 σ 2σ σ2
√ (1 + √ + · · · ), √ (1 + + · · · ).
N 2 N N 6N
√
So, cancelling σ/ N ,
1 √σ
∗
2(1 + 2 N
+ ···) 1 σ
1 − 2p = 1 − σ 2 = − √ + O(1/N ) :
2(1 + 6N + ···) 2 N

15
 σ 1 σ 1
N µN = N. √ .(− √ + O(1/N )) → µ := − σ 2 (N → ∞).
N 2 N 2
We now need to change notation:
(i) We replace the variance σ 2 above by σ 2 T . So σ 2 is the variance per unit
time (which is more suited to the work of Ch. VI, VII in continuous time);
the standard deviation (SD) σ is called the volatility. It measures the vari-
ability of the stock, so its riskiness, or its sensitivity to new information.
(ii) We replace ρ in the above by r. This is the standard notation for the
riskless interest rate in continuous time, to which we are now moving.
As usual, we write the standard normal density function as φ and distri-
bution function as Φ:
1 2 Z x Z x − 1 u2
e− 2 x e 2
φ(x) := √ , Φ(x) := φ(u)du = √ du.
2π −∞ −∞ 2π
Note that as φ is even, the left and right tails of Φ are equal:
Z −x Z ∞
φ(x) = φ(−x), so φ(u)du = φ(u)du : Φ(−x) = 1−Φ(x).
−∞ x

Theorem (Black-Scholes formula (for calls), 1973). The price of the

European call option is

ct = St Φ(d+ ) − Ke−r(T −t) Φ(d− ), (BS)

where St is the stock price at time t ∈ [0, T ], K is the strike price, r is the
riskless interest rate, σ is the volatility and
1 √ √
d± := [log(S/K) + (r ± σ 2 )(T − t)]/σ T − t : d+ = d− + σ T − t.
2
For completeness, we state the corresponding Black-Scholes formula for
puts. The proofs of the two results are closely analogous, and one can derive
either from the other by put-call parity.

Theorem (Black-Scholes formula for puts, 1973). The price of the

corresponding put option is

pt = Ke−r(T −t) Φ(−d− ) − St Φ(−d+ ). (BS − p)

The Black-Scholes formula is not perfect – indeed, Fischer Black himself

famously wrote a paper called The holes in Black-Scholes. But it is very

16
useful, as a benchmark and first approximation.

Proof of the Black-Scholes formula.

It suffices to take t = 0 – so T is the remaining time to expiry.
We use the Lemma, with µ = − 21 σ 2 T (in our new notation). In (1), we
have YN → Y in distribution and (replacing R in the Lemma by r, as above)
(1 + rT
N
)−N → e−rT as N → ∞. This suggests
(N )
C0 → C0 := EY [(S0 eY − e−rT K)+ ] = e−rT EY [(S0 erT +Y − K)+ ],

where EY is the expectation for the distribution of Y , which is N (− 21 σ 2 T, σ 2 T )

(in our current notation). This can be justified, by standard properties of√con-
vergence in distribution (see e.g. [W], Ch. 18). So if Z := (Y + 21 σ 2 T )/(σ T ),
√
Z ∼ N (0, 1), Y = − 21 σ 2 T + σ T Z, and
1 2 )T +σ
√
C0 = e−rT EZ [(S0 e(r− 2 σ TZ
− K)+ ]
1 2
−rT
Z ∞
1 √ e− 2 x
=e [S0 exp{(r − σ 2 )T + σ T x} − K]+ √ dx.
−∞ 2 2π
Similarly, with payoff h, the time-0 price of the claim, or option is
1 2
−rT
Z ∞
1 2 √ e− 2 x
e h(S0 exp{(r − σ )T + σ T x}) √ dx. (∗)
−∞ 2 2π

To evaluate the integral, note first that [...] > 0 where

1 √ 1 √
S0 exp{(r− σ 2 )T +σ T x} > K : x > [log(K/S0 )−(r− σ 2 )T ]/σ T = c, say. So
2 2
Z ∞
1 2 1 √ √
C0 = S0 e− 2 σ T . exp{− x2 + σ T x}dx/ 2π − Ke−rT [1 − Φ(c)],
c 2
and the last
√ term is Ke−rT Φ(−c) = Ke−rT Φ(d− ) (−c = [log(S/K) + (r −
1 2
2
σ )T ]/σ T = d− , when t = 0). The remaining integral is
Z ∞
1 √ 2 √ Z ∞
1 √
exp{− (x − σ T ) }dx/ 2π = √
exp{− u2 }du/ 2π
c 2 c−σ T 2
√ √
= 1 − Φ(c − σ T ) = Φ(−c + σ T ) = Φ(d+ ),

17
 √ √
as −c + σ T = d− + σ T = d+ when t = 0. So the option price is given
in terms of the initial price S0 , strike price K, expiry T , interest rate r and
volatility σ by
1 √
C0 = S0 Φ(d+ ) − Ke−rT Φ(d− ), d± := [log(S/K) + (r ± σ 2 )T ]/σ T . //
2
Note. 1. Normal approximation to binomial. The proof above starts from a
binomial distribution and ends with a normal distribution. The binomial dis-
tribution is that of a sum of independent Bernoulli random variables. That
sums (or averages) of independent random variables with finite means and
variances gives a normal limit is the content of the Central Limit Theorem
or CLT (the Law of Errors, as physicists would say). This form of the CLT
is the de Moivre-Laplace limit theorem.
The picture for this is familiar. The Binomial distribution B(n, p) has a
histogram with n + 1 bars, whose heights peak at the mode and decrease to
either side. For large n, one can draw a smooth curve through the histogram.
This curve is the relevant approximating normal density.
2. The Cox-Ross-Rubinstein binomial model above goes over in the passage
to the limit to the geometric Brownian motion model of VII.1. We will later
re-derive the continuous Black-Scholes formula in Ch. VII, using continuous-
time methods (Itô calculus), rather than, as above, deriving the discrete BS
formula and going to the limit on the formula, rather than the model.
3. For similar derivations of the discrete Black-Scholes formula and the pas-
sage to the limit to the continuous Black-Scholes formula, see e.g. [CR], §5.6.
4. One of the most striking features of the Black-Scholes formula is that it
does not involve the mean rate of return µ of the stock – only the riskless
interest-rate r and the volatility of the stock σ. Mathematically, this reflects
the fact that the change of measure involved in the passage to the risk-neutral
measure involves a change of drift. This eliminates the µ term; see VII.
5. Volatility. The volatility σ can be estimated in two ways:
a. Historic volatility. Directly from the movement of a stock price in time,
using Time Series methods in discrete time [see Ch. VI for continuous time].
b. Implied volatility. From the observed market prices of options: if we know
everything in the Black-Scholes formula (including the price at which the
option is traded) except the volatility σ, we can solve for σ. Since σ appears
inside the argument of the normal distribution function Φ as well as outside
it, this is a transcendental equation for σ and has to be solved numerically by
iteration (Newton-Raphson method). We quote (see ‘The Greeks’ below, and

18
Problems 7) that the Black-Scholes price is a monotone (increasing) function
of the volatility (more volatility doesn’t make us ‘more likely to win’, but
when we do win, we ‘win bigger’), so there is a unique root of the equation.
In practice, one sees discrepancies between historic and implied volatility,
which show limitations to the accuracy of the Black-Scholes model. But it is
the standard ‘benchmark model’, and useful as a first approximation.
The classical view of volatility is that it is caused by future uncertainty,
and shows the market’s reaction to the stream of new information. How-
ever, studies taking into account periods when the markets are open and
closed [there are only about 250 trading days in the year] have shown that
the volatility is less when markets are closed than when they are open. This
suggests that trading itself is one of the main causes of volatility.
Note. This observation has deep implications for the macro-prudential and
regulatory issues discussed in Ch. II. The real economy cannot afford too
much volatility. Volatility is (at least partly) caused by trading. Conclusion:
there is too much trading. Policy question: how can we reduce the volume
of trading (much of it speculative, designed to enrich traders, and not serv-
ing a more widely useful economic purpose)? One answer is the so-called
Tobin tax (also known as the ”Robin Hood tax”) (James Tobin (1918-2002),
American economist; Nobel Prize for Economics, 1981). This would levy
a small charge (e.g. 0.01%) on all financial transactions. This would both
provide a major and useful source of tax revenue, and – more importantly –
would discourage a lot of speculative trading, thereby (shrinking the size of
the financial services industry, but) diminishing volatility, to the benefit of
the general economy (Problems 7 again).
If the Black-Scholes model were perfect, historic and implied volatility
estimates would agree (to within sampling error). But discrepancies can be
observed, which shows the imperfections of our model.
Volatility estimation is a major topic, both theoretically and in practice.
We return to this in V.7.3-4 below and VII.7.5-8. But we note here:
(i) trading is itself one of the major causes of volatility, as above;
(ii) options like volatility [i.e., option prices go up with volatility].
Recalling Ch. II, this shows that volatility is a ‘bad thing’ from the point
of view of the real economy (uncertainty about, e.g., future material costs
is nothing but a nuisance to manufacturers), but a ‘good thing’ for financial
markets (trading increases volatility, which increases option prices, which
generates more trade ...) – at the cost of increased instability.

19
§7. More on European Options
1. Bounds. We use the notation above. We also write c, p for the values of
European calls and puts, C, P for the values of the American counterparts.
Obvious upper bounds are c ≤ S, C ≤ S, where S is the stock price (we
can buy for S on the market without worrying about options, so would not
pay more than this for the right to buy). For puts, one has correspondingly
the obvious upper bounds p ≤ K, P ≤ K, where K is the strike price: one
would not pay more than K for the right to sell at price K, as one would not
spend more than one’s maximum return. For lower bounds:
c0 ≥ max(S0 − Ke−rT , 0).
Proof. Consider the following two portfolios:
I: one European call plus Ke−rT in cash; II: one share. Show ”I ≥ II”.
p0 ≥ max(Ke−rT − S0 , 0) (proof: by above and put-call parity).
2. Dependence of the Black-Scholes price on the parameters.
Recall the Black-Scholes formulae for the values ct , pt for the European
call and put: with
1 p
d± := [log(St /K) + (r ± σ 2 )(T − t)]/σ (T − t),
2
ct = St Φ(d+ ) − Ke−r(T −t) Φ(d− ), pt = Ke−r(T −t) Φ(−d− ) − St Φ(−d+ ),
(a). S. As the stock price S increases, the call option becomes more likely to
be exercised. As S → ∞, d± → ∞, Φ(d± ) → 1, so ct ∼ St − Ke−r(T −t) . This
has a natural economic interpretation: as the value of a forward contract
with delivery price K (Hull [H1] Ch. 3, [H2] Ch. 3).
(b). σ. When the volatility σ → 0, the stock becomes riskless, and behaves
like money in the bank. Again, d± → ∞, as above, with the same economic
interpretation.
3. The Greeks.
These are the partial derivatives of the option price with respect to the
input parameters. They have the interpretation of sensitivities.
(i) For a call, say, ∂c/∂S is called the delta, ∆. Adjusting our holdings of
stock to eliminate our portfolio’s dependence on S is called delta-hedging.
(ii) Second-order effects involve gamma := ∂(∆)/∂S.
(iii) Time-dependence is given by Theta is ∂c/∂t.
(iv) Volatility dependence is given by vega := ∂c/∂σ.1
1
Of course, vega is not a letter of the Greek alphabet! (it is the Spanish word for
‘meadow’, as in Las Vegas) – presumably so named for ”v for volatility, variance and
vega”, and because vega sounds quite like beta, etc.

20
Vega.
From the Black-Scholes formula (which gives the price explicitly as a
function of σ), one can check by calculus (Problems 7) that

∂c/∂σ > 0,

and similarly for puts (or, use the result for calls and put-call parity). In sum:
options like volatility. This fits our intuition. The more uncertain things are
(the higher the volatility), the more valuable protection against adversity –
or insurance against it – becomes (the higher the option price).
(v) rho is ∂c/∂r, the sensitivity to interest rates.

§8. American Options.

We now consider an American call option (value C), in the simplest case
of a stock paying no dividends. The following result goes back (at least) to
R. C. MERTON in 1973.

Theorem (Merton’s theorem). It is never optimal to exercise an Ameri-

can call option early. That is, the American call option is equivalent to the
European call, so has the same value:

C = c.

First Proof. Consider the following two portfolios:

I: one American call option plus cash Ke−rT ; II: one share.
The value of the cash in I is K at time T , Ke−r(T −t) at time t. If the call
option is exercised early at t < T , the value of Portfolio I is then St − K from
the call, Ke−r(T −t) from the cash, total

St − K + Ke−r(T −t) .

Since r > 0 and t < T , this is < St , the value of Portfolio II at t. So Portfolio
I is always worth less than Portfolio II if exercised early.
If however the option is exercised instead at expiry, T , the American call
option is then the same as a European call option. Then at time T , Portfolio
I is worth max(ST , K) and Portfolio II is worth ST . So:

before T, I < II,

at T, I ≥ II always, and > sometimes.

21
This direct comparison with the underlying [the share in Portfolio II] shows
that early exercise is never optimal. Since an American option at expiry is
the same as a European one, this completes the proof. //
Second Proof. One can instead use the bounds of §7.1. For details, see e.g.
[BK, Th. 4.7.1].

Financial Interpretation.
There are two reasons why an American call should not be exercised early:
1. Insurance. Consider an investor choosing to hold a call option instead of
the underlying stock. He does not care if the share price falls below the strike
price (as he can then just discard his option) – but if he held the stock, he
would. Thus the option insures the investor against such a fall in stock price,
and if he exercises early, he loses this insurance.
2. Interest on the strike price. When the holder exercises the option, he
buys the stock and pays the strike price, K. Early exercise at t < T loses the
interest on K between times t and T : the later he pays out K, the better.
Economic Note. Despite Merton’s theorem, and the interpretation above,
there are plenty of real-life situations where early exercise of an American
call might be sensible, and indeed it is done routinely. Consider, for exam-
ple, a manufacturer of electrical goods, in bulk. He needs a regular supply
of large amounts of copper. The danger is future price increases; the obvious
precaution is to hedge against this by buying call options. If the expiry is a
year but copper stocks are running low after six months, he would exercise
his American call early, to keep an adequate inventory of copper, his crucial
raw material. This ensures that his main business activity – manufacturing
– can continue unobstructed. Neither of the reasons above applies here:
Insurance. He doesn’t care if the price of copper falls: he isn’t going to sell
his copper stocks, but use them.
Interest. He doesn’t care about losing interest on cash over the remaining six
months. He is in manufacturing to use his money to make things, and then
sell them, rather than put it in the bank.
This neatly illustrates the contrast between finance (money, options etc.)
and economics (the real economy – goods and services).

Put-Call Symmetry.
The BS formulae for puts and calls resemble each other, with stock price
S and discounted strike K interchanged. Results of this type are called put-
call symmetry.

22
American Puts.
Recall the put-call parity of Ch. II (valid only for European options):
c − p = S − Ke−rT . A partial analogue for American options is given by the
inequalities below:

S − K < C − P < S − Ke−rT .

For proof (as above) and background, see e.g. Ch. 8 (p. 216) of [H1].
We now consider how to evaluate an American put option, European
and American call options having been treated already. First, we will need
to work in discrete time. We do this by dividing the time-interval [0, T ]
into N equal subintervals of length ∆t say. Next, we take the values of the
underlying stock to be discrete: we use the binomial model of §5, with a
slight change of notation: we write u, d (‘up’, ‘down’) for (1 + b), (1 + a):
thus stock with initial value S is worth Sui dj after i steps up and j steps
down. Consequently, after N steps, there are N + 1 possible prices, Sui dN −i
(i = 0, · · · , N ). It is convenient to display the possible paths followed by the
stock price as a binomial tree, with time going left to right and two paths, up
and down, leaving each node in the tree, until we reach the N + 1 terminal
nodes at expiry. There are 2N possible paths through the tree. It is common
to take N of the order of 30, for two reasons:
(i) typical lengths of time to expiry are measured in months (9 months, say);
this gives a time-step around the corresponding number of days,
(ii) 230 paths is about the order of magnitude that can be easily handled by
computers (recall that 210 = 1, 024, so 230 is somewhat over a billion).
We now return to the binomial model in §§5,6, with a slight change of
notation. Recall that in §5 (discrete time) we used 1 + r for the discount
factor. Now call this 1 + ρ instead, freeing r for its usual use as the short
rate of interest in continuous time. Thus 1 + ρ = er∆t , and the risk-neutrality
condition p∗ = (b − r)/(b − a) of §5 becomes

p∗ = (u − er∆t )/(u − d).

√ √
Now recall (§7) (1+a)/(1+r) = exp(−σ/ N ), (1+b)/(1+r) = exp(σ/ N ).
We replaced√σ 2 by σ 2 T (to√make
√ σ the √ volatility per unit time), and T =
N.∆t, so σ/ N becomes σ T / N = σ ∆t. So now
√ √
u/er∆t = eσ/ ∆t
, d/er∆t = e−σ ∆t
: ud = e2r∆t .

23
 √
Since ∆t is small, its square ∆t is a second-order term. So to first order,
ud = 1, which simplifies filling in the terminal values in the binary tree.
We begin again: define our up and down factors u, d so that

ud = 1; (∗)

define the risk-neutral probability p∗ so as to have

p∗ = (u − er∆t )/(u − d)

(to get the mean return from the risky stock the same as that from the
riskless bank account), and the volatility σ to get the variance of the stock
price S 0 after one time-step when it is worth S initially as S 2 σ 2 ∆t:

S 2 σ 2 ∆t = p∗ S 2 u2 + (1 − p∗ )S 2 d2 − S 2 [p∗ u + (1 − p∗ )d]2
√
(using varS 0 = E(S 0 2 ) − [ES 0 ]2 ). Then to first order in ∆t (which is all the
accuracy we shall need), one can check that we have as before
√ √
u = exp(σ ∆t), d = exp(−σ ∆t). (∗∗)

We can now calculate both the value of an American put option and
the optimal exercise strategy by working backwards through the tree (this
method of backward recursion in time is a form of the Dynamic Programming
[DP] technique (Richard Bellman (1920-84) in 1953, book, 1957), which is
important in many areas of optimization and Operational Research (OR)).
1. Draw a binary tree showing the initial stock value and having the right
number, N , of time-intervals.
2. Fill in the stock prices: after one time interval, these are Su (upper) and
Sd (lower); after two time-intervals, Su2 , S and Sd2 = S/u2 ; after i time-
intervals, these are Suj di−j = Su2j−i at the node with j ‘up’ steps and i − j
‘down’ steps (the ‘(i, j)’ node).
3. Using the strike price K and the prices at the terminal nodes, fill in the
payoffs (fN,j = max[K − Suj dN −j , 0]) from the option at the terminal nodes
(where, at expiry, the values of the European and American options coincide)
underneath the terminal prices.
4. Work back down the tree one time-step. Fill in the ‘European’ value at
the penultimate nodes as the discounted values of the upper and lower right
(terminal node) values, under the risk-neutral measure – ‘p∗ times lower right

24
plus 1−p∗ times upper right’ [notation of V.6]. Fill in the ‘intrinsic’ (or early-
exercise) value - the value if the option is exercised. Fill in the American put
value as the higher of these.
5. Treat these values as ‘terminal node values’, and fill in the values one
time-step earlier by repeating Step 4 for this ‘reduced tree’.
6. Iterate. The value of the American put at time 0 is the value at the root -
the last node to be filled in. The ‘early-exercise region’ is the node set where
the early-exercise value is the higher; the rest is the ‘continuation region’.
Note. The above procedure is simple to describe and understand, and simple
to programme. It is laborious to implement numerically by hand, on exam-
ples big enough to be non-trivial. Numerical examples are worked through
in detail in [H1], 359-360 and [CR], 241-242.
Mathematically, the task remains of describing the continuation region -
the part of the tree where early exercise is not optimal. This is a classical
optimal stopping problem. No explicit solution is known (and presumably
there isn’t one). We will, however, connect the work above with that of IV.7
on the Snell envelope. Consider the pricing of an American put, strike price
K, expiry N , in discrete time, with discount factor 1 + r per unit time as
earlier. Let Z = (Zn )N n=0 be the payoff on exercising at time n. We want
to price Zn , by Un say (to conform to our earlier notation), so as to avoid
arbitrage; again, we work backwards in time. The recursive step is
1
Un−1 = max(Zn−1 , E ∗ [Un |Fn−1 ]),
1+r
the first alternative on the right corresponding to early exercise, the second
to the discounted expectation under P ∗ , as usual. Let Ũn = Un /(1 + r)n be
the discounted price of the American option. Then

Ũn−1 = max(Z̃n−1 , E ∗ [Ũn |Fn−1 ]) :

(Ũn ) is the Snell envelope (III.7) of the discounted payoff process (Z̃n ). So:
(i) a P ∗ -supermartingale,
(ii) the smallest supermartingale dominating (Z̃n ),
(iii) the solution of the optimal stopping problem for Z̃.

Note. One can use the Snell envelope to prove Merton’s theorem (equiva-
lence of American and European calls) without using arbitrage arguments.
For details see e.g. [BK, Th. 4.7.1 and Cor. 4.7.1].

25
P -measure and P ∗ − (or Q−) measure.
We use P and P ∗ in the above, as E and E ∗ are convenient, but P and
Q when the emphasis is on Q, for brevity.
The measure P , the real (or real-world) probability measure, models the
uncertainty driving prices, which are indeed uncertain, thus allowing us to
bring mathematics to bear on financial problems. But P is difficult to get
at directly. By contrast, Q is more accessible: the market tells us about Q,
or more specifically, trading does. In addition, trading also tells us about
the volatility σ, via implied volatility, which we can infer from observing the
prices at which options are traded. So Q is certainly more accessible than P .
There is thus a sense in which it is Q, rather than P , which is the more real.
It is as well to bear all this in mind when looking at specific problems, par-
ticularly numerical ones. Now that we know the CRR binomial-tree model,
which gives us the Black-Scholes formula in discrete time (and hence also, by
the limiting argument above, the Black- Scholes formula in continuous time,
the main result of the course), we can recognise the ‘one-period, up or down’
model ($/SFr in II.8, price of gold in Problems 5), though clearly artificial
and stylised, as a workable ‘building block’ of the whole theory. Because P
itself does not occur in the Black-Scholes formula(e), from a purely financial
point of view there is little need to try to construct more realistic, and so
more complicated, models of P . Instead, one can exploit what one can infer
about Q, which does occur in Black-Scholes, from seeing the prices at which
options trade.
From the economic point of view, it is the real world, the real economy,
and so the real probability measure P , that matters. The ‘Q-measure-eye
view of the world’ has a degree of artificiality, in so far as options do. One
can eat food, and needs to. One can’t eat options.
A fuller discussion of Q-measure involves Arrow-Debreu prices, equilibria
etc., but we omit this for lack of time, and because it would take us too far
into Economics.

26