AIMS2012 Macrina Slides

Asset Pricing:
Utility Theory and Pricing Kernel Models
Andrea Macrina
Department of Mathematics
Kings College London
African Institute for Mathematical Sciences

16-18 February 2012
Cape Town
Asset Pricing: Utility Theory and Pricing Kernel Models -2- AIMS, Cape Town
Contents
1. Dynamic asset pricing; pricing kernels; stochastic interest rate and market
price of risk; numeraire and change-of-measure martingales; asset pricing
formulae.
2. Utility functions; marginal utility; risk aversion and psycological discount
factor; utility maximization; optimal consumption.
3. Marginal utility of consumption and the pricing kernel; equilibrium asset
pricing; Sidrauski utility, money supply and liquidity benefit; inflation.
4. Pricing kernels driven by timeinhomogeneous Markov processes; explicit
AZ example of asset pricing including stochastic interest rates and market
risk premium, stochastic volatility, and option pricing for various asset classes.
16-18 February 2012 Andrea Macrina, Kings College London

Dynamic asset pricing
The financial market is modelled by a probability space (, F, P), and the

market information is modelled by the filtration {Ft}0t.
We may assume that the market filtration is generated by a multidimensional

Brownian motion {Wt}0t.
Asset prices. Each asset is characterised by (a) its price process and (b) the
dividend or cashflow stream that it generates. In the case of an asset
generating a continuous cash flow, the asset is defined by a pair of
{Ft}adapted processes {St} and {Dt} which we call the price process and the
dividend flowrate process, respectively.
Martingale. A process Mt is a martingale if it is {Ft}adapted and satisfies

E [|Mt|] < and E [Mt | Fs] = Ms for any 0 s t.

Pricing kernel. A process {t} is called a pricing kernel for the asset pair
{St, Dt} if the process {Mt} defined by
Z t
Mt = t St + u Du du (1)
0
is a martingale.
The existence of a (universal) pricing kernel ensures the absence of arbitrage in

the market.
The pricing kernel establishes the intertemporal relationship between asset cash
flows.
Money market account. We assume the existence of a money market

account {Bt}0t that is a nondividendpaying asset with a price process of the
form Z t
Bt = B0 exp rs ds , (2)
0
where {rt}0t is the {Ft}adapted short rate of interest.

Market price of risk. The pricing kernel {t} satisfies the dynamical equation
dt
= rt dt t dWt, (3)
t
where {t}0t is the (vectorvalued) market price of risk process.
Proof: Let {t}0t be defined by t = tBt. Then:
t
dt = d = t d(Bt1) + Bt1dt. (4)
Bt
Also,
d(Bt1) = Bt2dBt = rtBt1dt. (5)
By the martingale representation theorem, we may write
dt = t t dWt, (6)
and therefore

t t
dt = rt dt t dWt,
Bt Bt
= t rt dt t t dWt. (7)

Asset pricing formula. Recall that,

Z t
Mt = t St + uDu du, (8)
0
where {Mt}0t is an ({Ft}, P)martingale.
It follows that: Z T
1 P
St = E t St + uDu du Ft . (9)
t t

Market numeraire. The process {t}0t defined by
t = t1, (10)
is called the market numeraire (or natural numeraire).
By an application of Itos formula, we obtain
2t

dt = rt + t dt + t t dWt. (11)
Asset price dynamics. Let {St} denote the asset price process, and let {Dt}
denote the dividend rate process of the asset. The asset price process {St}
satisfies the following SDE:
dSt
= (rt t + tt) dt + t dWt, (12)
St
where t := Dt/St, and where {t}0t is some volatility process.

Proof: We recall that

Z t
Mt = t St + uDu du. (13)
0
Then, we have that

t
Mt 1
Z
St = sDs ds,
t t 0
t
Ds
Z
= t Mt t ds. (14)
0 s
An application of Itos formula gives
St
dSt = dt + t dMt + dt dMt Dt dt. (15)
t
Since {Mt} is a martingale, we may write
dMt = t dWt, (16)
where {t}0t is a vectorvalued {Ft}adapted volatility process.

Recalling the dynamics of the natural numeraire {t}0t, we obtain
dSt = (rtSt Dt + tt) dt + t dWt, (17)

where t := Stt + tt.
It follows that
dSt
= (rt t + tt) dt + t dWt, (18)
St
where t = Dt/St and t = t/St.
Example: GBM. Suppose the asset price process {St} is given by
1 2

St = S0 exp (r + )t 2 t + Wt , (19)
and the pricing kernel {t} is given by
1 2

t = exp rt 2 t Wt . (20)
Asset Pricing: Utility Theory and Pricing Kernel Models - 10 - AIMS, Cape Town
Then the pricing formula

T
1 P
Z

St = E T ST + u du Ft (21)
t t
is satisfied. In other words, the process

Z t
Mt = t St + uSu du (22)
0
is an ({Ft}, P)martingale.
Pricing of a discount bond. Let {PtT }0tT denote the price process of a
T maturity zerocoupon bond (nondividendpaying asset) with payoff function
PT T = 1. Then,
1 P
PtT = E [T | Ft] . (23)
t

The dynamical equation of the discount bond price process is given by
dPtT
= (rt + ttT )dt + tT dWt, (24)
PtT
where {tT }0tT is the bond volatility process.
Risk-Neutral Valuation
The price St at time t 0 of a limitedliability nondividendpaying asset is
1 P
St = E [T ST | Ft] , (25)
t
where ST can be viewed as the random cash flow that will occur at (the future)
fixed time T t.
Recalling that t := t/Bt, we have that


Bt P T
St = E ST Ft . (26)
t BT
Assuming that {t}0tT is a changeofmeasure density martingale from
P Q, we obtain

ST
St = B t E Q Ft . (27)
BT
This is the socalled riskneutral valuation formula, and Q is the riskneutral
measure.
Bond price process under Q. Under the riskneutral measure Q, the bond
price process {PtT } takes the form

1
= Bt EQ

PtT Ft ,
BT
Z T
= EQ exp

ru du Ft . (28)
t
Often this is the starting point in interestrate modelling.
However, what about the model for the market price of risk {t}, which is
necessary for asset pricing under the real probability measure P?

Utility functions
Utility is an abstract concept for the notion of happiness, satisfaction, or simply

for the pleasure derived from consumption of a specified amount of goods and
services.
Some goods and services can provide immediate utility, e.g. an ice cream, or a
shower; other goods, like education, provide continuous utility.
Money provides immediate utility since the notion of the value of money is
universally recognised.
However, money provides also a continuous flow of utility in the form of
status, security, flexibility, etc.

Utility function
A function U : R+ R that satisfies
(i) U (x) C 2
(ii) U (x) > 0
(iii) U (x) < 0
(iv) limx0 U (x) = , limx U (x) = 0
is called a utility function.
The derivative U (x) is called the marginal utility.

Example: Logarithmic utility. U (x) = a ln(x), a > 0.

a
Example: Power utility. U (x) = p xp, a > 0, p (, 1)\{0}.
The parameter p in the power utility function can be interpreted as a measure

for how riskaverse an investor is. The parameter p controls the steepness of the
utility function.
The concavity condition on the utility function, U (x) < 0, implies risk aversion.
Consider a gambling game in which a player receives a random amount X in

exchange for a wager equal to E[X]. Jensens inequality implies that
E [U (X)] U (E[X]) . (29)

This says that a riskaverse player requires an incentive to play a risky game,
and requests to receive a positive risk premium as an additional reward for
accepting the risk involved in the game.

Inverse marginal utility

A function I : R+ R characterised by
I [U (x)] = x,
U [I(y)] + y (30)
is called the inverse function of the marginal utility U (x).
The inverse function I(y) always exists for a standard utility function since
U (x) > 0 and U (x) < 0.
The inverse marginal utility function is monotonically decreasing.
Examples: U (x) = ln(x) I(y) = y1 , and

U (x) = p1 xp I(y) = y q1, where p1 + 1q = 1.

Convex dual
A standard utility function has socalled convex dual given by
U (y) = max {U (z) zy} , (31)

z>0
for y > 0.
The convex dual of a standard utility function is guaranteed to exist, and is

given by
U (y) = U [I(y)] yI(y). (32)
This is shown by observing that for fixed y, the maximum of U (z) zy over z is
obtained when
d
[U (z) zy] = U (z) y = 0. (33)
dz
This is a genuine maximum because U (z) < 0.
Therefore we have y = U (z) and z = I(y) at the maximum.

Maximised expected utility

Let Ht be a random cash flow at the fixed time T > 0. The expected utility
E [U (HT )] is maximised subject to the budget constraint H0 = E[T HT ] if
HT = I(T ), and is chosen such that the budget constraint is satisfied.
Proof:
We consider an alternative investment target ZT that also satisfies the budget
constraint. We next show that the expected utility derived from ZT cannot
exceed that of HT
E [U (HT )] E [U (ZT )] = E [U (HT ) T HT ] E [U (ZT ) T ZT ]
= E [U (I(T )) T I(T )] E [U (ZT ) T ZT ]

= E U (T ) E [U (ZT ) T ZT ]

= E U (T ) (U (ZT ) T ZT ) (34)
However, by virtue of the definition of U (x), we have that

U (T ) > (U (ZT ) T ZT ) ,
and therefore E [U (HT )] E [U (ZT )] 0.
Example: Optimal investment with logutility. Consider U (x) = a ln(x),

so U (x) = a/x. The first order condition is then is given by
a a
U (HT ) = = T HT = . (35)
HT T
From the budget constraint, we obtain
a
H0 = E [T HT ] = , (36)

and therefore
H0
HT = = H0 T . (37)
T
Optimal consumption
The expected utility of consumption over the period [0, T ] given by
Z T
E U (ct)dt (38)
0
is maximised by the random consumption plan ct = I( t), where is chosen

such that the budget constraint

Z T
E tctdt (39)
0
is satisfied.
Example: logutility of consumption. Let U (ct) = ln(ct). Then the
optimal consumption strategy is given by
H0
ct = t (40)
T
where {t} is the price process of the natural numeraire asset.
Optimal consumption with impatience

Let the utility of consumption be discounted by the exponential impatience
factor exp(t). Then subject to the budget constraint
Z T
H0 = E tctdt , (41)
0

the expected total discounted utility

Z T
E etU (ct)dt (42)
0
is maximised by
etU (ct) = t. (43)
Example: logutility with impatience. We consider U (x) = ln(x). It
follows that U (ct) = 1/ct, and that
e t
ct = . (44)
t
The budget constraint simplifies to
1 T

H0 = 1e . (45)

Therefore, we have:
e t H0 t
ct = = T
e t , (46)
t 1e
or
1 eT
t = ct . (47)
H0 et
Monetary Models for Interest Rates and Inflation
1. Modelling framework for inflation

2. Real and nominal pricing kernels
3. Indexlinked bonds
4. Real interest rates
5. Dynamics of the consumer price index
6. Macroeconomic models for inflation
We now have a look at past inflation levels in the UK, the USA, and Japan.

Interest rate models and asset pricing
To begin, lets say what we mean by an interest rate model.

An interest rate model consists of the following ingredients:
(i) A probability space (, F, P) with a market filtration {Ft}0t< to which
asset price processes are adapted.
(ii) A process {Bt}0t< which we call the nominal money market account,
Z t
Bt = exp rsds , (48)
0
where {rt} is called the short rate.
(iii) A system of nominal discount bond price processes {PtT }0tT <.
(iv) A pricing kernel {t}0t< with the property that {tBt} and {tPtT } are
martingales.

More generally, for any asset with price process {St}0t, and continuous
dividend process {Dt} we require that the process {Mt} defined by
Z t
Mt = t St + sDsds. (49)
0
is a martingale.
We assume that the market filtration is generated by a multidimensional
Brownian motion.
With these ingredients in place, we say that we have a model for the nominal
interest rate system.
The pricing of contingent claim with payoff HT is then given by
Et[T HT ]
Ht = . (50)
t

Models for inflation
The modelling of inflation has a number of subtle features.

We need to model nominal interest rates, real interest rates, and the consumer
price index in a consistent way. The basic setup is as follows.
For the nominal discount bond system we have
Et[T ]
PtT = . (51)
t
The associated dynamics are given by:
dPtT
= (rt + ttT ) dt + tT dWt. (52)
PtT
Here rt (nominal short rate) and t (nominal market price of risk) are defined by
dt
= rtdt tdWt, (53)
t
and tT is the nominal discountbond vector volatility. The dynamics of the
discount bond system gives rise to the interpretation of rt and t.
Next we consider the consumer price index {Ct}.

The rate of inflation over a given period is defined to be the percentage increase
in the price index over the given period, expressed on an annualised basis.
The question is: how do we model {Ct}?
One is tempted to write down the usual arbitrage-free dynamics for the price
process of a risky asset.
But since {Ct} is a price index (including the prices of non-durable items such
as a haircut, a pint of beer, a meal in a restaurant, or a night at the cinema) one
cannot immediately conclude that the usual arbitragefree dynamics apply.
Instead, we consider a related family of prices: the prices of indexlinked bonds.
These are financial assets, and hence are subject to the laws of no arbitrage.

Index-linked bonds
An index-linked discount bond pays out the amount CT at time T .

I
The value PtT of an index-linked bond at time t is therefore
I Et[T CT ]
PtT = . (54)
t
I
If we divide PtT by Ct we get the value (in real terms) of a bond that pays one
unit of goods and services at time T :
R Et[T CT ]
PtT = . (55)
t Ct
Now suppose we define the real pricing kernel tR by writing
tR = tCt. (56)
Then clearly for the price index we have
tR
Ct = . (57)
t
In terms of tR , the real discount bond prices take the form

Et[TR ] R
= R
. PtT (58)
t
The dynamics of the real discount bond system are given by:
R
dPtT R R R
R
R
= r t + t tT dt + tT dWt . (59)
PtT
Here rtR and R

t are defined by
dtR R R
= r t dt t dWt , (60)
tR
and R
tT is the real discount bond volatility.
The dynamics of the real discount bond show that rtR and R t have the
interpretation of being the real short rate, and the real market price of risk,
respectively.

The CPI process
Now we can deduce the dynamics of the consumer price index. In particular, by
an application of Ito calculus to the relation Ct = tR /t we obtain:
dCt R R
R

= rt rt + t t t dt + t t dWt. (61)
Ct
Thus we can write
dCt R

= rt rt + tt dt + tdWt. (62)
Ct
Here the CPI volatility vector is defined by: t = t Rt .
We see that {Ct} has the arbitrage-free dynamics of a risky asset that pays a
continuous dividend at the rate rtR .
Absence of statistical arbitrage opportunities implies t2 > tt > 0.
With these relations at hand, we are now in a position to build models for
pricing inflation-linked derivatives.
In particular, it is apparent that the two required ingredients are t and tR .
Monetary systems
Inflation is a monetary phenomenon. To model a monetary economy we need

three more ingredients:
1. The rate of real consumption {kt}.

2. The money supply {Mt}.
3. The convenience yield or benefit rate {t} associated with the money supply.
The process {t} represents the intangible rate of benefit delivered, in dollars
per unit of time and per unit of money supply.
If the money supply at time t is Mt it confers a benefit at the rate of tMt
dollars per year.
Thus one thinks of the benefit as a kind of consumable commodity.

What really counts is the real rate of benefit derivable from the presence of
the money supply.
This is given by lt = tMt/Ct.
The consumer price index is thus used as a kind of exchange rate to convert
the nominal benefit into a real benefit.
Optimisation
Our goal now is to determine a set of relations between {kt}, {Mt}, {t}, and
the consumer price index {Ct}.
We assume the existence of a standard bivariate utility function U (x, y) such
that the expected total utility
Z T
E etU (kt, lt)dt (63)
0
is maximised over some time horizon.

This is subject to the budget constraint

Z T Z T
W0 = E tCtktdt + ttMtdt . (64)
0 0
Here W0 denotes the total wealth of the economy (including, e.g., the present
value of any forthcoming income) available for the given period. A standard
argument shows that the maximum is achieved if

Ux(kt, lt) = ettCt, (65)
and
Uy (kt, lt) = ettCt (66)
for some constant which is fixed by the budget constraint.
These relations can then be solved to give t, tR , and Ct as functions of kt, Mt,
and t.

A simple example
A simple choice for U (x, y) is given by

U (x, y) = A ln(x) + B ln(y), (67)
where A, and B are constants.
It follows that
A B
Ux(kt, lt) = , Uy (kt, lt) = . (68)
kt lt
Equating these we get
A A tMt
kt = l t = . (69)
B B Ct
Hence, rearranging, we have
A tMt
Ct = . (70)
B kt
Thus the consumer price index is determined completely by the nominal money
supply benefit and the rate of real consumption.
But we can also determine the pricing kernels. A calculation shows that
Bet Aet
t = , tR = . (71)
tMt kt
Thus both the real and the nominal interest rate systems are determined by the
model.
Now suppose that HT is the payoff of a contingent claim.
Then in our monetary policy model it follows that

HT
H0 = 0 M0 eT E . (72)
T MT
This example shows how expectations concerning monetary policy can have a
major effect on the valuation of long-dated derivatives.

Information-sensitive pricing kernels with

Time-Inhomogeneous Markov Processes
Stochastic discount bond systems can be constructed by modelling the pricing

kernel (stochastic discount factor) that we denote {t}0t.
The price PtT at time t of a discount bond with unit payoff at maturity T is
EP [T | Ft]
PtT = , (73)
t
where P is the real probability measure.
Next we model the pricing kernel {t} and the filtration {Ft} following the
scheme of information-based asset pricing developed in
D.C. Brody, L.P. Hughston, A. Macrina (2008) Information-Based Asset Pricing.
International Journal of Theoretical and Applied Finance Vol. 11, 107-142.
We fix a time U > T and introduce a random variable XU with real probability
density p(x).
The random variable XU may represent a macroeconomic factor (e.g. the GDP
level of a country at time U ) revealed at time U .
Suppose that at time t < U , market investors have access only to incomplete
information about the macroeconomic factor XU .
We model this incomplete information by an information process {LtU } with the

property that LU U = G(XU ), where G(x) is an invertible function.
Next, we assume that the market filtration is given by Ft = ({LsU }0sU ).
Let the bond price process {PtT }0tT <U be adapted to {Ft}.
We consider a pricing kernel {t} that is modelled by a function of the value

LtU at time t, and possibly time t:
t := (t, LtU ). (74)
The function (t, x) shall be chosen such that the pricing kernel is guaranteed
to be a positive supermartingale.

Armed with the models for the pricing kernel and the market filtration, the price
PtT of the discount bond is
EP [(T, LT U ) | LtU ]
PtT = . (75)
(t, LtU )
Here we have recalled that the market filtration {Ft} is generated by {LtU }
which is taken to be a time-inhomogeneous Markov process.
To obtain explicit models for the bond price PtT , we need to explicitly construct
(i) pricing kernel models and (ii) information processes {LtU }.
One method to construct information-based pricing kernels is presented in:

L. P. Hughston & A. Macrina (2012) Pricing Fixed-Income Assets in an
Information-Based Framework. Applied Mathematical Finance.
arXiv.org: No. 0911.1610.
However, we consider another method, today.

Weighted heat kernel approach with time-inhomogeneous

Markov processes
The details of the interest-rate framework, which is presented in this lecture, can
be found in:
J. Akahori & A. Macrina (2012) Heat Kernel Interest Rate Models with
Time-Inhomogeneous Markov Processes. International Journal of Theoretical
and Applied Finance.

We consider a time-inhomogeneous Markov process {LtU }0tU , and introduce

a real-valued measurable function p(u, t, x).
A so-called propagator {p(u, t, LtU )} associated with the process {LtU } has the
property that
E [p (u, t, LtU ) | LsU ] = p (u + t s, s, LsU ) , (76)
for s < t, 0 < u, and 0 < u + t < U .
An example of a propagator is
p(u, t, LtU ) = E [F (u + t, Lu+t,U ) | LtU ] , (77)

where F (t, x) is taken to be a measurable positive function.
Next we introduce a so-called weight function w(t, u) that is positive and

measurable, and has the property that
w(t, u s) w(t s, u), (78)
where t, u [0, U ) and s t u.
A weighted heat kernel is then defined by

Z U t
g(t, LtU ) = p(u, t, LtU ) w(t, u) du, (79)
0
for 0 t < U < . In the case of the propagator (77), we have

Z U t
g(t, LtU ) = E [F (u + t, Lu+t,U ) | LtU ] w(t, u) du. (80)
0
It can be proved that {g(t, LtU )} is a positive supermartingale by showing that

Z U s
E [g(t, LtU ) | LsU ] = p(s, u, LsU ) w(t, u t + s) du (81)
Zts
U s
p(s, u, LsU ) w(s, u) du, (82)
0
= g(s, LsU ), (83)

for s t.

Applying the pricing kernels constructed by weighted heat kernels, the bond
price PtT can be expressed by
R U t
T t p(u, t, LtU ) w(T, u T + t) du
PtT = R U t . (84)
0 p(u, t, LtU ) w(t, u) du
In the case that the propagator is given by (77), we have
R U t
T t E [F (u + t, Lu+t,U ) | LtU ] w(T, u T + t) du
PtT = R U t (85)
0 E [F (u + t, Lu+t,U ) | LtU ] w(t, u) du
Explicit formulae for the bond price are obtained by specifying the functions
F (t, x) and w(t, u), and the information process {LtU }.

Bond pricing with Brownian bridge information
We consider the case where {LtU }0tU is a so-called Brownian bridge

information process:
LtU = XU t + tU , (86)
where XU is taken to be independent of the Brownian bridge process {tU }.
We assume that the market filtration is defined by

Ft = {LsU }0st . (87)
Remarks:
(i) {LtU } is a time-inhomogeneous {Ft}-Markov process.
(ii) XU is FU -measurable.
(iii) The information flow rate in (86) is constant.
(iv) Var[tU ] = t(U t)/U .

We recall that the bond price can be expressed by
E [(T, LT U )| LtU ]
PtT = . (88)
(t, LtU )
In order to work out the conditional expectation, we assume with no loss of
generality, that
(t, LtU ) = Mt g(t, LtU ), (89)
where the (P, {Ft})-martingale {Mt} is defined by
dMt U
= E [XU | LtU ] dWt, (90)
Mt U t
for 0 t < U and where
t t
LsU 1
Z Z
Wt = LtU + ds U E [XU | LsU ] ds. (91)
0 U s 0 U s
The martingale {Mt} induces a change of measure from P to the so-called
bridge measure B under which {LtU } has the distribution of a Brownian bridge.

The bond price can thus be expressed as follows:
EP [ (T, LT U ) | LtU ]
PtT = , (92)
(t, LtU )
EP [MT g(T, LT U )| LtU ]
= , (93)
Mt g(t, LtU )
EB [g(T, LT U )| LtU ]
= . (94)
g(t, LtU )
We emphasize that the pricing kernel {(t, LtU )} is a P-supermartingale if
{g(t, LtU )} is a supermartingale under B (and vice versa).
We now may make use of

Z U t
g(t, LtU ) = EB [F (u + t, Lu+t,U ) | LtU ] w(t, u) du, (95)
0

and hence of
R U t
T t EB [F (Lt+u,U ) | LtU ] w(T, u T + t) du
PtT = R U t . (96)
B
E [F (Lu+t,U ) | LtU ] w(t, u) du
0
The information process {LtU }0t<U has the distribution of a Brownian bridge
under B so that the conditional expectation can be worked out explicitly.
Example: quadratic family

Let F (x) = x2, and w(t, u) = U t u.
A calculation shows that
i u(U t u) 2
h U tu
EB (Lu+t,U )2 LtU = L2tU .

+ (97)
U t U t
With this intermediate result at hand, we can write the weighted heat kernel

process as follows:
Z U t
g(t, LtU ) = EB [F (Lu+t,U ) | LtU ] w(t, u) du, (98)
0
Z U t " 2 #
u(U t u) U tu
= + L2tU (U t u)du.
0 U t U t
(99)
The integral in the expression for the weighted heat kernel can be calculated in
closed form, so that we obtain the supermartingale
1 1
g(t, LtU ) = (U t) + (U t)2 L2tU .
3
(100)
12 4
The bond price PtT at time t, derived in this example, is thus given by
1 3 1 (T t)(U T )3 1 (U T )4
12 (U T ) + 4 (U t) + 4 (U t)2 L2tU
PtT = 1 3 1 2 2
. (101)
12 (U t) + 4 (U t) LtU
The simulation of the bond price is straightforward since the process {LtU },
LtU = XU t + tU , (102)
is Gaussian conditional on the outcome of the underlying economic factor XU .
The short rate process {rt} can be worked out by calculating the instantaneous
forward rate associated with the bond price {PtT }0tT <U .
The result is:

L2tU
r(t, LtU ) = 1 1 2
, (103)
4 (U t) 3 (U t) + LtU
for 0 t < U . We emphasize that this is a positive interest rate model.
The market price of risk {t} associated with the quadratic family is
1 2
U P 2 (U t) LtU
(t, LtU ) = E [XU | LtU ] 1 . (104)
U t 12 (U t)3 + 41 (U t)2 L2tU

Example: exponential quadratic family

1 2

Let F (x) = exp 2 t+u x , where t+u is deterministic.
In this case the propagator takes the form

B 1 2

E exp 2 t+u Lt+u,U | LtU
2

1 t+u at+u
= exp L2tU , (105)
1 u t+u at+u 2 (1 u t+u at+u)
where at+u = (U t u)/(U t).
By setting t+u = (U t u)1, and by choosing the weight function to be
1
w(t, u) = (U t u) 2 ( > 21 ), (106)
we obtain an analytical expression for the supermartingale {g(t, LtU )}:
L2tU

1
g(t, LtU ) = (U t) exp . (107)
12 2(U t)

The supermartingale (107) leads to a deterministic bond price, even though the
related pricing kernel is stochastic.
However we can modify slightly the supermartingale {g(t, LtU )}:
Let f0(t) and f1(t) be positive, decreasing and differentiable functions.
Consider the supermartingale

L2tU

g(t, LtU ) = f0(t) + f1(t)(U t) exp . (108)
2(U t)
Then the associated bond price system has stochastic dynamics:
2
L
f0(T ) + f1(T )(U T )1/2(U t)1/2 exp 2(UtUt)
PtT = 2 , (109)
L
f0(t) + f1(t)(U t) exp 2(UtUt)
for t [0, U ) and u [0, U t].
We note that further examples can be constructed: a semi-analytic formula is

obtained for the exponential linear family defined by F (x) = exp( x).
Credit-risky discount bonds with stochastic discounting
In the last part of this talk, we give an idea how the information-based asset
pricing framework can be extended so as to incorporate stochastic discounting.
In particular we aim at generalizing the information-based credit-risk models

presented in
D. C. Brody, L. P. Hughston & A. Macrina (2007) Beyond hazard rates: a new
framework for credit-risk modelling. In Advances in Mathematical Finance,
Festschrift Volume in Honour of Dilip Madan, edited by R. Elliott, M. Fu,
R. Jarrow & J.-Y. Yen. Birkhauser, Basel and Springer, Berlin.
We proceed as follows:
We fix two dates T and U , where T < U , and attach two independent factors
XT and XU to these dates.
The payoff of the credit-risky bond is modelled by making use of the random
variable XT .
We assume that XT is a discrete random variable that takes values in

{x0, x1, . . . , xn} with a priori probabilities {p0, p1, . . . , pn}, where
0 x0 < x1 < . . . < xn1 < xn 1. (110)
We assume that the economic factor XU is a continuous random variable.
With XT and XU we associate the independent information processes

{LtT }0tT and {LtU }0tU defined by
LtT = 1 t XT + tT , LtU = 2 t XU + tU . (111)
The market filtration {Ft} is generated by {LtT } and {LtU }:
Ft = ({LsT }0st, {LsU }0st) (112)

The price BtT at t T of a defaultable discount bond with payoff HT at
T < U is given by
EP[T HT | Ft]
BtT = . (113)
t

Let the pricing kernel {t} be defined by

t = Mt g(t, LtU ) (114)
and let the payoff of the credit-risky bond be given by
HT = H (XT , LT U ) . (115)
The formula for the price BtT of the credit-risky bond is then worked out by
applying the weighted heat kernel approach for the pricing kernel, and eventually
by specifying the payoff function H(XT , LT U ).
However, we leave this task for another time...
Meantime these results can be found in:
A. Macrina & P. A. Parbhoo (2009) Security Pricing with Information-Sensitive

Discounting. In: Recent Advances in Financial Engineering 2009. Proceedings of
the KIER-TMU International Workshop on Financial Engineering 2009 (World
Scientific).
Levy Random Bridges
We fix a probability space (, F, Q).
We say that {LtT }0tT has the law LRB C ([0, T ], {ft}, ) if the following hold:
LT T has marginal law .

There exists a Levy process {Lt} C[0, T ] such that Lt has density ft(x) for all
t (0, T ].
concentrates mass where fT (z) is positive and finite, i.e. 0 < fT (z) < for
-a.e. z.
For every n N+, every 0 < t1 < < tn < T , every (x1, . . . , xn) Rn, and
-a.e. z, we have
Q [Lt1,T x1, . . . , Ltn,T xn | LT T = z ]
= Q [Lt1 x1, . . . , Ltn xn | LT = z ] .

LRBs have the following properties:

We can interpret an LRB to be a Levy process conditioned to have a specified
marginal law at the terminal time T .
LRBs are time-inhomogeneous Markov processes.
LRBs have stationary, dependent increments.
Levy processes and Levy bridges are special cases of LRBs.
Credit risk simulation.

Next we show some simulations of bond price processes driven by various
different LRBs
The associated information processes (LRBs) are plotted in redthe bond price
processes in blue.
In the following simulations, we assume that the bond ends at maturity in a
state of default.

h i
1 x2
Brownian information: ft(x) = 2t
exp 12 t .
XtT
tT 1.0
0.4 0.8
0.2
0.6
Time
1 2 3 4 5 0.4
-0.2
0.2
-0.4
Time
-0.6 0 1 2 3 4 5

mt 1 2 mt2 14
m 2 +4

x
VG information: ft(x) =
(mt) 4 Kmt 1 mx2 , m > 0.
2
XtT
tT 1.0
0.6
0.8
0.5
0.4 0.6
0.3
0.4
0.2
0.1 0.2
Time
0.2 0.4 0.6 0.8 1.0 Time
0.0 0.2 0.4 0.6 0.8 1.0

1 ct
Cauchy information: ft(x) = x +c2 t2 ,
2 c > 0.
XtT
tT 1.0
0.7
0.8
0.6
0.5
0.6
0.4
0.3 0.4
0.2
0.2
0.1
Time Time
0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Further details in:

E. Hoyle, L.P. Hughston, and A. Macrina (2011) Levy Random Bridges and the
Modelling of Financial Information. Stochastic Processes and Their Applications
121, 856-884.

AIMS2012 Macrina Slides

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

AIMS2012 Macrina Slides

Uploaded by

Copyright:

Available Formats

Asset Pricing:

Utility Theory and Pricing Kernel Models

African Institute for Mathematical Sciences

16-18 February 2012 Andrea Macrina, Kings College London

Dynamic asset pricing

The financial market is modelled by a probability space (, F, P), and the

We may assume that the market filtration is generated by a multidimensional

Martingale. A process Mt is a martingale if it is {Ft}adapted and satisfies

16-18 February 2012 Andrea Macrina, Kings College London

The existence of a (universal) pricing kernel ensures the absence of arbitrage in

Money market account. We assume the existence of a money market

where {rt}0t is the {Ft}adapted short rate of interest.

Proof: Let {t}0t be defined by t = tBt. Then:

16-18 February 2012 Andrea Macrina, Kings College London

Asset pricing formula. Recall that,

where {Mt}0t is an ({Ft}, P)martingale.

16-18 February 2012 Andrea Macrina, Kings College London

Market numeraire. The process {t}0t defined by

By an application of Itos formula, we obtain

16-18 February 2012 Andrea Macrina, Kings College London

Proof: We recall that

Then, we have that

where {t}0t is a vectorvalued {Ft}adapted volatility process.

Recalling the dynamics of the natural numeraire {t}0t, we obtain

dSt = (rtSt Dt + tt) dt + t dWt, (17)

Example: GBM. Suppose the asset price process {St} is given by

Then the pricing formula

is satisfied. In other words, the process

16-18 February 2012 Andrea Macrina, Kings College London

The dynamical equation of the discount bond price process is given by

The price St at time t 0 of a limitedliability nondividendpaying asset is

Recalling that t := t/Bt, we have that

16-18 February 2012 Andrea Macrina, Kings College London

Often this is the starting point in interestrate modelling.

16-18 February 2012 Andrea Macrina, Kings College London

Utility is an abstract concept for the notion of happiness, satisfaction, or simply

16-18 February 2012 Andrea Macrina, Kings College London

(ii) U (x) > 0

(iii) U (x) < 0

(iv) limx0 U (x) = , limx U (x) = 0

is called a utility function.

The derivative U (x) is called the marginal utility.

16-18 February 2012 Andrea Macrina, Kings College London

Example: Logarithmic utility. U (x) = a ln(x), a > 0.

The parameter p in the power utility function can be interpreted as a measure

Consider a gambling game in which a player receives a random amount X in

E [U (X)] U (E[X]) . (29)

16-18 February 2012 Andrea Macrina, Kings College London

Inverse marginal utility

is called the inverse function of the marginal utility U (x).

The inverse marginal utility function is monotonically decreasing.

Examples: U (x) = ln(x) I(y) = y1 , and

16-18 February 2012 Andrea Macrina, Kings College London

U (y) = max {U (z) zy} , (31)

The convex dual of a standard utility function is guaranteed to exist, and is

Therefore we have y = U (z) and z = I(y) at the maximum.

16-18 February 2012 Andrea Macrina, Kings College London

Maximised expected utility

However, by virtue of the definition of U (x), we have that

Example: Optimal investment with logutility. Consider U (x) = a ln(x),

is maximised by the random consumption plan ct = I( t), where is chosen

such that the budget constraint

Optimal consumption with impatience