Professional Documents
Culture Documents
Novembre 2000
Table of Contents
Table of Contents
PART I 1 Standard (One Period) Portfolio Theory . . . . . . . . . . . . . . . . . . . . . 1
Portfolio Choices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 1.A Framework and notations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 1.A.i No Risk-free Asset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 1.A.ii With Risk-free Asset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.B Ecient portfolio in absence of a risk-free asset . . . . . . . . . . . . . . . . . . . . . . 6 1.B.i Eciency criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 1.B.ii Ecient portfolio and risk averse investors . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 1.B.iii Ecient set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 1.B.iv Two funds separation (Black) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.C Ecient portfolio with a risk-free asset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 1.D HARA preferences and Cass-Stiglitz 2 fund separation . . . . . . . . . . . . . . 14 1.D.i HARA (Hyperbolic Absolute Risk Aversion) . . . . . . . . . . . . . . . . . . . . . . . . 14 1.D.ii Cass and Stiglitz separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 Capital Market Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.A CAPM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.A.i The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.A.ii Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.A.iii CAPM as a Pricing and Equilibrium Model . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.A.iv Testing the CAPM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.B Factor Models and APT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.B.i K-factor models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.B.ii APT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.B.iii Arbitrage and Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 2.B.iv References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
PART II Multiperiod Capital Market Theory : the Probabilistic Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 3 Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 3.A Probability Space and Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 3.B Asset Prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 3.B.i Denitions and Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 3.C Portfolio Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 3.C.i Notation: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 3.C.ii Discrete Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 3.C.iii Continuous Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Table of Contents
AoA, Attainability and Completeness. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 4.A Denitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 4.B Propositions on AoA and Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 4.B.i Correspondance between Q and : Main Results . . . . . . . . . . . . . . . . . . . 35 4.B.ii Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 Alternative Specications of Asset Prices . . . . . . . . . . . . . . . . . . . . . . . . . . 39 5.A Ito Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 5.B Diusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 5.C Diusion state variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 5.D Theory in the Ito-Diusion Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 5.D.i Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 5.D.ii Martingales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.D.iii Redundancy and Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 5.D.iv Criteria for Recognizing a Complete Market . . . . . . . . . . . . . . . . . . . . . . . . 44
PART III State Variables Models: the PDE Approach. . . . . . . . . . . . . . . . 45 6 7 Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 Discounting Under Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 7.A Itos lemma and the Dynkin Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 7.B The Feynman-Kac Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 The PDE Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 8.A Continuous Time APT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 8.A.i Alternative decompositions of a return . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 8.A.ii The APT Model (continuous time version) . . . . . . . . . . . . . . . . . . . . . . . . . . 51 8.B One Factor Interest Rate Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 8.C Discounting Under Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 Links Between Probabilistic and PDE Approaches . . . . . . . . . . . . . . . 55 9.A Probability Changes and the Radon-Nikodym Derivative . . . . . . . . . . . 55 9.B Girsanov Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 9.C Risk Adjusted Drifts: Application of Girsanov Theorem . . . . . . . . . . . . 56
PART IV The Numeraire Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 10 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 11 Numeraire and Probability Changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 11.A Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 11.A.i Assets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 ii
Table of Contents 11.A.ii Numeraires . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.B Correspondence Between Numeraires and Martingale Probabilities . 11.B.i Numeraire Martingale Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.B.ii Probability Numeraire . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.C Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
61 62 62 63 63
12 The Numeraire (Growth Optimal) Portfolio . . . . . . . . . . . . . . . . . . . . . . . 65 12.A Denition and Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 12.A.i Denition of the Numeraire (h, H) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 12.A.ii Characterization and Composition of (h, H) . . . . . . . . . . . . . . . . . . . . . . . . 65 12.A.iii The Numeraire Portfolio and Radon-Nikodym Derivatives . . . . . . . . . . . . 69 12.B First Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 12.B.i CAPM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 12.B.ii Valuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 PART V Continuous Time Portfolio Optimization. . . . . . . . . . . . . . . . . . . . 72 13 Dynamic Consumption and Portfolio Choices (The Merton Model) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 13.A Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 13.A.i The Capital Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 13.A.ii The Investors (Consumers) Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 13.B The Solution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 13.B.i Sketch of the Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 13.B.ii Optimal portfolios and L + 2 funds separation . . . . . . . . . . . . . . . . . . . . . . 77 13.B.iii Intertemporal CAPM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 14 THE EQUIVALENT STATIC PROBLEM (Cox-Huang, Karatzas approach) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 14.A Transforming the dynamic into a static problem . . . . . . . . . . . . . . . . . . . . 80 14.A.i The pure portfolio problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 14.A.ii The consumption-portfolio problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 14.B The solution in the case of complete markets . . . . . . . . . . . . . . . . . . . . . . . . 83 14.B.i Solution of the pure portfolio problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 14.B.ii Examples of specic utility functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 14.B.iii Solution of the consumption-portfolio problem . . . . . . . . . . . . . . . . . . . . . . 86 14.B.iv General method for obtaining the optimal strategy x . . . . . . . . . . . . . . . 87 14.C Equilibrium: the consumption based CAPM . . . . . . . . . . . . . . . . . . . . . . . . 88 PART VI STRATEGIC ASSET ALLOCATION . . . . . . . . . . . . . . . . . . . . . . . 90 15 The problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 16 The optimal terminal wealth in the CRRA, mean-variance
iii
Table of Contents
and HARA cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 16.A Optimal wealth and strong 2 fund separation. . . . . . . . . . . . . . . . . . . . . . . 92 16.B The minimum norm return . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 17 Optimal dynamic strategies for HARA utilities in two cases . . . . 93 17.A The GBM case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 17.B Vasicek stochastic rates with stock trading . . . . . . . . . . . . . . . . . . . . . . . . . 93 18 Assessing the theoretical grounds of the popular advice . . . . . . . . . 94 18.A The bond/stock allocation puzzle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 18.B The conventional wisdom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 REFERENCES 95
iv
Chapter 1
Portfolio Choices
1.A
In all the following we consider a single period or time interval (0 1), hence two instants t = 0 and t = 1 Consider an asset whose price is S(t) (no dividends or dividends reinvested). The return of this asset between two points in time (t = 0, 1) is: R= S (1) S (0) S (0)
We now consider the case of a portfolio. and distinguish the case where a riskless asset does not exist from the case where a risk free asset is traded.
1.A.i
No Risk-free Asset
There are N tradable risky assets noted i = 1, ..., N : The price of asset i is Si (t), t = 0, 1. The return of asset i is Ri = Si (1) Si (0) Si (0) 2
Chapter 1
Portfolio Choices
The number of units of asset i in the portfolio is ni . The portfolio is described by the vector n(t); ni can be >0 (long position) or <0 (short position). Then the value of the portfolio, denoted by X (t), is X (t) = n0 S (t) with n (0) = n (1) = n (no revision between 0 and 1), the prime denotes a transpose. S (t) stands for the column vector (S1 (t), ..., SN (t))0 The return of the portfolio is: X (1) X (0) X (0)
RX =
xi (0) = xi =
(Note that xi (1) 6= xi ). Besides the weights sum up to one: x0 1=1 where x= (x1 , x2 , ..., xN )0 and 1 is the unit vector. The return of the portfolio is the weighted average of the returns of its components:
RX = x0 R
Chapter 1
Portfolio Choices
Proof
X (1) X (0) n0 S (1) X (0)
N X ni Si (1) Si (0) X (0) Si (0) i=1 N X i=1 N X i=1
1 + RX
= = =
xi
Si (1) Si (0)
xi (1 + Ri )
N X i=1
= 1+
xi Ri
Q.E.D.
Dene i = E [Ri ] and = (1 , 2 , ..., N ) , then: X = E (RX ) = x0 Denote the variance-covariance matrix of returns NN = ( ij ), where ij = cov (Ri , Rj ), then: var (RX ) = var (x0 R) = x0 x N N XX = xi xj ij
i=1 j=1
1.A.ii
We now have N +1 assets, with asset 0 being the risk-free asset, and the remaining N assets being the risky assets.
Chapter 1
Portfolio Choices
S0 (1) = S0 (0) (1 + r) with r a deterministic interest rate. Again we can dene the portfolio in units, with n= (n0 , n1 , n2 , ..., nN )0 The portfolio can be similarly dened in weights: xi = ni S (0) X (0)
xi
where x= (x1 , x2 , ..., xN )0 denotes the weights in the N risky assets. The return of the portfolio is: RX = x0 r +
N X i=1
xi Ri = r +
N X i=1
xi (Ri r)
The term (Ri r) is the excess return of asset i over r. Moreover: X = E (RX ) = r + x0 where is the risk premium vector of the E (Ri r) Also denote NN as the variance-covariance matrix of the risky assets, then: var (RX ) = x0 x is always positive semi-denite (meaning that x, x0 x 0). In some cases it is positive denite (x 6= 0, x0 x > 0).
that
Denition 1 Assets i = 1, 2, ..., N are redundant if there exist N scalars 1 , 2 , ..., N such P
N i=1 i Ri
Proposition 1
The N assets i = 1, 2, ..., N are not redundant iff singular or invertible). is positive denite (i.e. non-
Chapter 1
Portfolio Choices
Proof
Assume that the assets are redundant, then there exist N scalars 1 , 2 , ..., N such that PN i=1 i Ri = k. Consider the portfolio dened by the weights . The variance of its return = var (k) = 0 = 0 , i.e. is singular and not positive denite. Conversely if is singular and not positive denite there existP non 0 vector such that 0 = 0; Then the return of a N portfolio has zero variance and i=1 i Ri = k Q.E.D.
Remark 1 In the following sections we will assume that the assets are non-redundant (it is always possible to drop redundant assets if any).
1.B
1.B.i
Eciency criteria
Consider any ecient portfolio (x , X ) and let variance(RX ) = k x solves the optimization program (P ) : max E [RX ]
x
s.t. x0 x = k ; x0 1 = 1
Chapter 1
Portfolio Choices
Remark that these rst order conditions are necessary and also sucient for the solution being a maximum since the second order conditions hold (L(x) is strictly concave - positive denite). Theorem 1
A portfolio (x, X) is ecient iff there exist two scalars and such that for all i = 1, 2, ..., N : i = + cov (Rx , Ri )
Proof
The necessary and sucientP condition for x to be ecient is that it satises the rst order condition: for all i: i = + N x ij . We then have: j=1 j i = +
N X j=1
x cov (Ri , Rj ) j
N X j=1
= + cov Ri,
x Rj j
Remark 2 The second term can be considered as the additional required rate of return (risk premium), proportional to cov (Ri , RX ). Remark 3 If cov (Ri , RX ) = 0, then i = . Remark 4 Also note:
var (RX ) =
N N XX i=1 j=1 N X i=1 N X i=1
xi xj ij
N X j=1
xi cov Ri ,
xj Rj
xi cov (Ri , RX )
The covariance term cov (Ri , RX ) indicates the contribution of asset i to the total risk of the portfolio. Therefore, additional required rate of return should be proportional to this induced risk which is what is stated in the theorem. Moreover cov (Ri , RX ) appears to be the relevant measure of risk for any asset i embedded in the portfolio X.
Chapter 1
Portfolio Choices
1.B.ii
Dene another optimization program (P 0 ), equivalent to (P ) : 0 0 0 (P ) max x x x s.t. : x0 1 = 1 2 ( (P ) and (P 0 ) yield the same solutions since they have the same Lagrangian) (P 0 ) writes, equivalently: M ax E [RX ] var (RX ) , s.t. : x0 1 = 1 2 In (P 0 ) is interpreted as a given risk-aversion while in (P ) it is an unknown lagrangian multiplier. In (P ) we are given 2 and we solve for and as functions of 2 . In (P 0 ) we X X are given the risk-aversion parameter and solve for 2 as function of . X The rst order conditions of (P) write as for (P): x 1 = 0 (with only one multiplier for (P 0 )) Consider the case of minimum variance portfolio where = , i.e. 1 min x0 x s.t. : x0 1 = 1 2
The Lagrangian is then: 1 L (x,) = x0 x x0 1 2 Call k1 the solution. The rst order condition gives: Together with the constraint k01 1 = 1 gives: = Thus: k1 = 1 1 1 = 0 1 1 1 1 1 1 10 1 1 k1 1 = 0
Chapter 1
Portfolio Choices
1.B.iii
Ecient set
Denition 3 The Ecient Set is the set of all x that obey the rst order condition. Equivalently, it is the set of all x that solve the optimization program (P 0 ) 0.
Recall that the rst order condition for (P 0 ) is: x 1 = 0 Dene risk tolerance b as the inverse of risk aversion, i.e. Then x can be solved as: x = b 1 1 b= 1
Then:
1 = 10 x = 10 b 1 1 b 0 1 10 1 1 = 1 b 1 b b b 0 1 10 1 1 = 1 10 1 = 10 1 1
or:
Then: x = b 1 1 10 1 1 = b 1 10 1 1 10 1 1 1 b 1 = 0 1 + 1 1 1 10 1 1 9
Chapter 1
Portfolio Choices
We recognize in the rst term the minimum variance portfolio (k1 ) and we call k2 the second term: k1 k2 1 1 = 0 1 1 1 10 1 1 0 1 1 = 1 1 x = k1 + b 2 k
Since the expected return x0 is linear in b and the variance is quadratic in b in , 2 the ( , R) space the ecient portfolios are represented by the ecient frontier, which is a parabola. Each point on the ecient frontier corresponds to a given , the slope of the parabola at this point being equal to 2 (the shadow price in (P ) of the constraint on variance). In the (, R) space the ecient frontier is an hyperbola. 1.B.iv Two funds separation (Black)
Note that k01 1 = 1 and x0 1 = 1, therefore k02 1 = 0. Any ecient portfolio is thus the sum of k1 (the minimum variance portfolio) and k2 which is a zero weight (zero investment) portfolio. As it could be expected, an investor with a zero risk tolerance will hold only k1 ; If he has a positive risk tolerance b he will add a risk b in order to increase the expected return. The ecient set can taking the form k2 now be caracterized as: o n b b > 0 ES = x |x = k1 + k2
Theorem 2
Consider any two ecient portfolio x and y:
1. 2. 3.
Any convex combination of x and y is ecient, i.e. u [0, 1] , ux+ (1 u) y ES Any ecient portfolio is a combination of x and y (not necessarily a convex combination) The whole parabola (ecient and inecient frontier) is generated by (all) combinations of x and y 10
Chapter 1
Portfolio Choices
h i z = [uk1 + (1 u) k1 ] + ubX + (1 u) bY k2
Let z ES, then z = k1 + bZ k2 for some bZ > 0. For any x ES and y ES: h i ux + (1 u) y = k1 + ubX + (1 u) bY k2 By equating bZ to ubX + (1 u) bY we get: u = Then the combination u x + (1 u ) y = z
Q.E.D.
bZ bY bX bY
1.C
Consider gure 1 where the upper branch of the hyperbola EFR represents, in the (, E) space, the ecient portfolios in absence of a riskless asset. Assume now that exists a risk free asset 0 yielding the certain return r. M stands for the tangency point of the hyperbola EFR with a straight line drown from r representing asset 0. Point M represents a portfolio composed only of risky assets, called the tangent portfolio. 11
Chapter 1
Portfolio Choices
E EFR M r X
Figure 1.1.
12
Chapter 1
Portfolio Choices
1. 2.
Proposition 2 Asset 0 is ecient Consider any portfolio X. Any combination of 0 and X yielding R = uRX + (1 u) r, lies on the straight line connecting 0 and X in the (, E) space Any feasible portfolio which representative point is not on r M (such as X) is dominated by portfolios in r M. The straight line r M is the ecient frontier and is called the Capital Market Line (Tobins Two-fund Separation) Any ecient portfolio is a combination of any two ecient portfolios, for instance 0 and M Any ecient portfolio writes: x = b 1 r1
3.
4. 5.
6.
m = bM 1 r1 1 bM = 0 1 1 r1 Proof
1, 2, 3, 4 are standard and easy to prove. Let us proove 5 and 6: x ES solves: max 1 r + x0 r1 x0 x 2 r1 = x Then: x 1 1 r1 = b 1 r1 = 1 r1
Q.E.D.
13
Portfolio Choices
b < bM , the portfolio is long in 0 and m b > bM , the portfolio shorts 0 Remark 6 We dene later the market portfolio as a portfolio containing all the risky assets present in the market (and only risky assets). In absence of riskless asset the market portfolio is ecient iif its representative point belongs to the hyperbola EFR. In presence of a risk free asset the necessary and sucient condition for the market portfolio to be ecient is that it coincides with the tangent portfolio m (which is the only ecient portfolio of EFR, in presence of a risk free asset). Would all investors face the same ecient frontier (it would be the case under homogeneous expectations and horizon) and would they all follow the mean-variance criteria, they would all hold combinations of 0 and M and the tangent portfolio M would necessarily coincide with the market portfolio.
1.D
A rational agent (in the sense of Von Neumann-Morgenstern) should maximize the expected utility of wealth E [U (W )].
1.D.i
Some restrictions are imposed on the coecients and b and the domain of denition. The absolute risk tolerance (ART) and absolute risk aversion (ARA) are: ART = 1 ARA U0 = 00 U W = b+ 14
Chapter 1
Portfolio Choices
and the relative risk tolerance (RRT) is: RRT = In particular: 1. b=0 b 1 + W
U (W ) =
We obtain CRRA, i.e. constant relative risk aversion. A limit case of CRRA is obtained for = 1 which can be showed to be equivalent to the Log utility 2. = 1 U (W ) = W i.e. the quadratic utility function. 3.
W 1 1
W2 2b
Using a quadratic utility function implies a mean-variance criteria; Indeed: b min var (RX ) s.t. E [RX ] = E (and x0 1 = 1) 2 b min E [RX ] s.t. E [RX ] = E (and x0 1 = 1) i h b min E [X 2 (1)] s.t. E [X (1)] = X (0) 1 + E min E[X 2 (1)] E [X (1)] 1 max E X (1) X 2 (1)
2
4.
Saturation at W = b (for that wealth U (W ) = W Wb is maximum; U (W ) decreases 2 ARA increasing with wealth (it is commonly admitted that ARA decreases for most
agents).
Indierence to skewness (only the two rst moments of W matter), whereas most
investors actually like skewness.
1.D.ii
Cass and Stiglitz showed that all HARA investors sharing the same exponential 15
Chapter 1
Portfolio Choices
parameter can build their optimal portfolios by mixing the two same funds. When a risk free asset exists it can be chosen as one of the two funds. Since all quadratic (mean-variance) investors exhibit the same (= 1) Tobin and Black 2 fund separation are particular cases of Cass and Stiglitz separation. Cass and Stiglitz conditions on the utility functions for separation to hold for investors sharing the same exponential parameter are summarized in the following table @r r Complete Market (under complete markets r) class wider than HARA Incomplete Market quadratic or CRRA2 HARA
in the particular case of CRRA one fund suces (for a given the portfolio is the same for all W
16
Chapter 2
2.A.i
The Model
Consider again N risky assets (a risk free asset may exist or not). The market value of asset i is Vi , then (by denition of the market portfolio) its weight in the market portfolio is: Vi mi = PN Vi
i=1
RM = m0 R
Remark 7 The market portfolio would be ecient if all investors would hold ecient portfolios (since a combination of ecient portfolios is ecient).
Theorem 3
(General CAPM )
1.
If (H) is true, then there exist and such that, for i = 1, ..., N : i = E [Ri ] = + cov (RM , Ri )
2.
Conversely, if there exist and such that, for i = 1, ..., N : i = + cov (RM , Ri ), then (H) is true. 17
Chapter 2
Proof
The proof comes directly from Theorem 1. Q.E.D.
Remark 8 can be interpreted as the risk aversion of the average (representative) investor. Remark 9 CAPM holds for any portfolio (x, X). Indeed, call RX its return and consider the case where no risk free asset exists (x0 1 = 1) : E [RX ] = = =
N X i=1 N X i=1 N X i=1
xi i xi ( + cov (RM , Ri )) xi +
N X i=1
xi cov (RM , Ri )
N X i=1
= + cov RM ,
xi Ri
= + cov (RM , RX )
with weight x0
Remark 10 The proof follows the same lines when the portfolio contains a risk free asset Remark 11 and are the same for all assets or portfolios Remark 12 For the market portfolio:
M Therefore: = Then: i = + cov (RM , Ri ) = + M2 cov (RM , Ri ) M M 2 M = + cov (RM , RM ) = + 2 M
18
Chapter 2 Dene:
i =
cov (RM , Ri ) 2 M
Then we may write the CAPM equation in the alternative form: E [Ri ] = + i (M )
Consider any portfolio z with Z = 0: Z = 0 cov (RM , RZ ) = 0 cov (m0 R, z0 R) = z0 m = 0 z m z [vect (m)]
vect [v1 , v2 , ..., vN ] is the set of all linear combinations of v1 , v2 , ..., vN , or linear subspace generated by v1 , v2 , ..., vN . The dimension of [vect (m)] is thus N 1 and there are an innity of 0-beta portfolios. Now, from the general CAPM, we would have: = Z ; Thus: Corollary 1 (0beta CAPM) If M is ecient, for any zero beta portfolio or asset Z: E [Ri ] = Z + i (M Z ) Corollary 2 (Standard CAPM) : If there exists a risk-free asset yielding r (which is a particular zero beta asset) E [Ri ] = r + i (M r) Note that Z = r for any zero beta portfolio or asset.
2.A.ii missing
Geometry
2.A.iii
e e For a security delivering V (1) at time 1(the pdf of V (1) is given, thus e e E(V (1)) and cov(V (1) , RM ) are known), what is its price V (0) at time 0? 19
Chapter 2
Lets assume that there exists a risk-free asset, then: h i ! e E V (1) e V (1) = E [1 + R] = 1 + r + cov , RM V (0) V (0) with = Then: h i e e E V (1) = (1 + r) V (0) + cov V (1) , RM V (0) = h i e (1) cov V (1) , RM e E V 1+r M r 2 M
and
i.e. V (0) is the present value of its certainty equivalent at time 1 discounted at the risk-free rate. However this asset may be an element of the market portfolio M (unless this claim is in zero net supply ..) and therefore the previous pricing formula is not a closed form general equilibrium relation. In fact CAPM is an equilibrium condition stemming from the demand side; The equilibrium price can only be otained by specifying the supply side (in the previous example the supply was a right on an exogeneous cash ow X). General equilibrium requires a specication of the supply of all securities traded in the market. Consider the N risky assets together and we look for their equilibrium prices. We
e assume rst an inelastic supply. Assume that asset i delivers Vi (1), an exogenous cash ow, at time 1, what is its price at time 0? h i ! e E Vi (1) e M r Vi (1) =1+r+ , RM cov Vi (0) 2 Vi (0) M
For i = 1, 2, ..., N . We have N equations with N unknowns Vi (0) (i = 1, ..., N ). e PN e V (1) (1 + RM = P i=1 Vi (0) allows to compute M , 2 , cov Vi (1) , RM as functions of the N M Vi (0) i=1 i Vi (0))
Consider again the N risky assets and an elastic supply with constant returns to scale,
where the joint pdf of the Ri is given and independent of the scale Vi (0) to be invested in technology i. The CAPM determines the scale Vi (0) of investment in technology i
20
i = E [Ri ] = r + and 1 + RM =
M r cov (RM , Ri ) 2 M
PN
2.A.iv
One remark about this important empirical topic. Testing the CAPM is equivalent to testing (H). However, how should we dene the market portfolio and how to measure the market return? Usually the market portfolio is proxied by stock (plus bond) indices. But results on stock indices do not include all assets in M (non tradable assets, art,..). Hence we test the eciency of the index and not that of M (Rolls Critique).
2.B
2.B.i
K-factor models
ik Fk + i
k=1
0 Fk + k
21
In practice, we should have large N and small K, so that in estimating the variance-covariance matrix, cov (Ri , Rj ) =
K X k=1
ik jk 2 k
we only need to estimate K terms of 2 and run N regressions for estimating k the ik . In CAPM or in the Markowitz model, without the factor decomposition, we need to estimate N (N 1) /2 terms. A Particular case: K = 1 boils down into the market model that writes: Ri = i + i F + i Then: RM =
N X i=1 N X i=1 N X i=1
mi i + F
mi i +
mi i
= M + F
PN
i=1
mi i = 1:
Ri = i + i [RM M ] + i Note that the Ri are linked through [RM M ] (since cov (Ri , Rj ) = i j 2 ). M Also, i [RM M ] is the systematic risk, and i is the unsystematic (diversiable) risk; only systematic risk should be priced (CAPM).
2.B.ii
APT
We assume that the returns are generated by a K factors linear process previously dened that writes: 22
Chapter 2
R = + +F + = +
K X k=1
k Fk +
Recall that k is an N dimensioned column vector with an ith component equal to ik Denition 4 A zero investment portfolio, dened by the amount of wealth, x, invested in each asset, satises:
x0 1 = 0 V (0) = 0 V (1) = x0 R
xi (1 + Ri ) =
N X i=1
xi +
N X i=1
xi Ri = x0 R
Absence of arbitrage (AOA) prevails if no arbitrage portfolio can be constructed i.e: x01 = 0 and x0R 0 a.s. implies x0R = 0 a.s. (or equivalently implies E(x0R) = 0) Theorem 4
(AP T ) In AoA there exist K + 1 scalars such that: = 0 1 + 1 1 +... + K K or i = 0 + 1 i1 + ... + K iK
0 is the required rate of return without systematic risk. k is the market price of risk k. k ik is the risk premium imposed to security i because it has a risk k of intensity ik . 23
Chapter 2
Proof
Consider any well-diversied zero investment portfolio satisfying: x0 1 = 0 x0 k = 0 hence: h i x is any element of vect 1,1 , 2 , ..., K RX = x0 R = x0 + = x
0 K X
or x 1 or xk for k = 1, ..., K
Fk x0 k + x0
k=1
Since x0 is certain, in AoA x0 must be zero (if x0 > 0 then x is an arbitrage portfolio and if x0 < 0 then x is an arbitrage portfolio). Thus: x0 = 0 or x, which means that is orthogonal to any element x of [vect(1, 1 , 2 , ..., K )] , i.e. vect(1, 1 , 2 , ..., K ) implying that exist K + 1 scalars such that : = 0 1 + 1 1 + ... + K K Q.E.D.
In the particular case where there is a risk-free asset, then: 0 = 0 = r and i = r + 1 i1 + ... + K iK
2.B.iii
Equilibrium implies AoA, but the inverse is not true. AoA conditions do not involve utility functions.
24
Chapter 2
2.B.iv
References
25
Chapter 3
Framework
Chapter 3 Framework
3.A
We consider the usual probability triplet (, F, P ), where F is a -algebra on representing the observable events at time T .
Information in the period [0, T ] is represented by a ltration {Ft }t[0,T ] , where Ft is the set of observable events at time t (represented by a algebra), and the sequence {Ft }t[0,T ] satises the usual conditions: F0 = {null events and a.s. event} (s < t) (Fs Ft ) FT = F \ Fs = Ft
t>s
In the discrete time setting, all transactions take place at discrete points, i.e., t = 1, 2, ..., T . In the continuous time setting, transactions take place continuously, i.e., t [0, T ].
27
Chapter 3
Framework
3.B
Asset Prices
3.B.i
There are N + 1 assets traded in the market, one being the locally risk-free asset, denoted by 0, and the remaining N being the risky assets. The prices of those assets are noted Si (t) ( for i = 0, 1, ..., N ); S(t) = (S1 (t), .., SN (t))0 or (S0 (t), S1 (t), .., SN (t))0 (depending on the context) is the N (or N + 1) dimensional column vector of asset prices. Without loss of generality it will generally be assumed that Si (0) = 1 It is assumed for the time being that there is no dividend, or that a dividend is reinvested in the asset that delivers it. 1. In the discrete time case S0 (t) = S0 (t 1) [1 + r (t 1)], with r (t 1) being the locally risk-free rate in [t 1, t] , known at time t 1 but unknown before. Remark that S0 (t + 1) = S0 (t)(1 + r(t)) is random at t 1 since r(t) is unknown. 2. In the continuous time context:
r(t) is stochastic but Ft -adapted. For a risk-free asset: dS0 = S0 rdt or S0 (t) = e with S0 (0) = 1. For a risky asset we will usually assume that prices follow Ito processes: dSi = Si i dt + Si i 0 dw with risk induced by w, the vector of standard Brownian Motions. Technical conditions (e.g., the integrability conditions) apply. If Si follows Ito process, we preclude jumps. If jumps are involved, however, then a rather general assumtion is that Si follows a semi-martingale process. A slightly more specic assumption is that asset prices follow processes that yield a.s. Right Continuous and Left Limited (RCLL) paths. When considering the possibility of 28
Rt
0
r(u)du
Chapter 3
Framework
jumps we will assume RCLL processes for the asset prices to avoid the sostication of semi martingales3 . It is worthwhile to note that Ito processes RCLL Semi martingales. Most of the results of the next chapter (On AOA and completeness) hold in the semi-martingale case.
3.C
Portfolio Strategies
3.C.i
Notation:
n(N+1)1 the vector of the N+1 numbers of assets ; xN1 the vector of N weights on risky assets S(N+1)1 the vector of the N+1 asset prices X (t) = n0 (t)S(t) the value of the portfolio at t (n,X) or (x,X) a strategy 3.C.ii Discrete Time
[t 1, t[ is period t 1;at time t S(t) is set and, just after, n(t) is choosen During period t 1, the value of the portfolio will evolve:
X(t) X(t 1) = n0 (t)S(t)n0 (t 1)S(t 1) = n0 (t 1) [S(t) S(t1)] + S0 (t) [n(t) n(t 1)]
R Consider the integral: (u) dS. In a regular integral of this form dS is innitesmal, while in a jump process it can assume some nite value somewhere.
3
The rst term in the right hand side of the equation, n0 (t 1) [S(t) S(t1)], is the gain during the period [t 1, t[ , and is represented as g(t 1, t).
29
Chapter 3
Framework
The second term can be deemed as the net cash inow added to the portfolio at time t. Indeed it can be decomposed into two terms: S0 (t)n(t 1), the value of assets sold at time t, and S0 (t)n(t), the algebric value of assets purchased (may be <0 if sales> purchases).
The cumulative gain in [0, t], dened for t = 1, ..., T , can be represented as: G(t) =
t X u=1
g(u 1, 1)
Denition 6 (Self-nancing Portfolio) When at each time t the net inow is 0, the strategy
is said to be self-nancing, i.e., if (n,X) is self-nancing, then: X(t) X(t 1) = g(t 1, t) = n0 (t 1) [S(t) S(t1)] and X(t) = X(0) + G(t)
Let Si (t) be the value of asset i at time t, and S0 (t) be the numeraire, then the discounted value of i is: Si (t) Sid = S0 (t) Self-nancing is independent of the numeraire used; In particular (n,X) self-nancing implies: X d (t) X d (t 1) = n0 (t 1) Sd (t) Sd (t 1) 3.C.iii Continuous Time
dG(t) = n0 (t)dS(t)
t
dG(u) =
0
n0 (u)dS(u)
0
30
Chapter 3
Framework
with the rst term in the right hand side of the equation being the period gain dG(t) and the second term the net inow at t + dt. Again, in a self-nancing strategy: dX(t) = dG(t), and X(t) = X(0) + G(t); As in the discrete time case, the self nancing property as well as the expression of the gain do not depend on the choosen numeraire .
31
Chapter 4
Denition 7 strategy (n,X) is admissible if: 1. 2. 3. n(t) is Ft adapted and satises some technical conditions4 . X(t) L1,2 . (This is an additional condition imposed sometimes) X(t) is bounded from below to avoid doubling Strategies5 . Denition 8 A is the set of admissible strategies Denition 9 A0 = { Self-nancing and admissible strategies} We now work with A0 , i.e., (n,X) A0 , dX = n0 dS.
4
(a) G(t) =
(c) (predictability of n (t)) n(t) LCRL so that if there is a jump in S(t), rebalancing must take place in t+ but never in t , the latter being equivalent to insider trading, i.e., a rebalancing, or jump, in n(t) takes the advantage of a jump in S(t) that has just occured. This condition is not necessary when S(t) is continuous.
5
(b) (Integrability) Rt i. 0 kn0 (u)k2 du< a.s. Rt ii. 0 |n0 (u)| du< a.s.
In a Doubling Strategy the gambler bets 2 when losing 1 and bets 4 when losing 2...
32
Chapter 4
It is also possible to dene a strategy by a vector of weights xN1 . The weight of the risk-free asset in the portfolio is then 1 x0 1. Denition 10 (a,A) is an arbitrage if: 1. 2. 3. 4. (a,A) A0 . A(0) = n0 (0)S(0) =0, (i.e., zero initial investment). A(T ) 0 a.s. (i.e., non-negative cash ow at the end). E [A(T )|F0 ] > 0 There is an arbitrage opportunity each time that a strategy (x, X) in A0 dominates another strategy (y, Y ) in A0 (i.e. X(T ) Y (T ) a.s. and E[X(T )] E[Y (T )] for the same initial investment X(0) = Y (0); or X(T ) = Y (T ) a.s.with X(0) < Y (0)). Arbitrage is built by being long in (x, X) and short in (y, Y ). Example 1 X(T ) S0 (T ) = e
RT
0
r(u)du
Example 2 X(T ) = K, a constant, while X(0) < KBT (0) where BT (0) denotes the value
at time 0 of a zero-coupon bond yielding 1 at time T.
1. 2. 3.
e CT is FT measurable.
33
Chapter 4
Example 3 The terminal values of N + 1 primitive assets are contingent claims. Example 4 A FT , the indicator function 1A is a contingent claim.
e e e Denition 13 CT C is attainable if (c,C) A0 with C(T ) = CT a.s.. We say CT is
Denition 14 Ca = {attainable contingent claims} Denition 15 Cn = {non-attainable contingent claims} Denition 16 The market is (dynamically) complete when all contingent claims are attainable, i.e., Ca = C or Cn = .
Remark 13 Market completeness is unrealistic in discrete time, but less unrealistic in continuous time. In continuous time the possibility of rebalancing at each point of time allows a much larger spanning. When completeness is obtained through continuous rebalancing, the market is said dynamically complete.
Denition 17 A pricing formula maps C onto R. To be viable, must satisfy: 1. e e0 is linear, i.e., 1 , 2 , CT C, and CT C: e e0 e e0 1 CT + 2 CT = 1 CT + 2 CT
2.
3.
e CT C
Denition 18 = {| viable}
34
Chapter 4
Denition 19 Two probability measures P and Q are equivalent if they have the same null
sets (the impossible as well as the certain events are the same for P and Q)
Z(t) is a martingale if E[Z(t)/F s ] = Z(s) for any s and t such that 0 s t T Denition 21 Q = Q|Q P and (x,X) A0 , E Q
n h
X(T ) S0 (T ) |F0
o X(0) = S0 (0) . Equivalently, Q is a set of P -equivalent probability measures Q under which the asset 0 discounted asset X(T prices X d (T ) = S0 (T) are martingales. )
h i X(T It should be noted that X(0) 6= E P S0 (T) |F0 because investors are risk-averse ) and expect a return dierent than the risk-free rate r (usually higher since, in general, holding a risky asset increases the risk of their portfolio). However, this does not mean that there is no such a probability measure as Q that yields Qmartingale discounted prices. In the following we will consider the problems: Are Q and empty? What is the relation between Q and ?
4.B
Recall in the following that S0 (0) = 1 4.B.i Correspondance between Q and : Main Results
Theorem 5
Assume Q and are not empty. There exists a one-to-one relation between Q and .
Q Q , dened by:
e e CT C : Q CT = E Q 35
"
e CT |F0 S0 (T )
Chapter 4
1.
2.
e e CT 0 a.s., Q CT 0;Indeed:
e Q is linear (because the expectation operator E is linear), i.e., XT Ca a e and YT C : # " e e Q XT + YT e e Q XT + YT = E S0 (T ) " # " # e e XT Y T Q Q = E +E S0 (T ) S0 (T ) e e = Q XT + Q YT eT > 0 Q CT = E Q e C " # eT C >0 S0 (T ) " # e CT =0 S0 (T )
and
3.
eT = 0 a.s. Q CT = E Q e C
"
e CT S0 (T )
Chapter 4
4.
It has been shown that Q is a viable pricing formula that maps Q into . Moreover, this mapping is injective, i.e., Q0 6= Q and Q0 , Q Q, Q0 6= Q . Indeed: Q0 Q A FT s.t. Q0 (A) 6= Q (A) 0 E Q [1A ] 6= E Q [1A ] 1A S0 (T ) Q (1A S0 (T )) = E S0 (T ) Q = E [1A ]
Q
6=
and Q0 (1A S0 (T )) = E
Q0
0
= E Q [1A ]
1A S0 (T ) S0 (T )
Therefore we obtain dierent prices for this particular contingent claim, hence, Q0 6= Q . 5. The proof ends by checking that when is a viable price system Q (A) = (1A S0 (T )) denes a probability measure which has the same null sets than P .
Q.E.D.
1. 2. 3.
Theorem 6 In AoA, a viable pricing formula on Ca exists and is unique. Market is complete and AOA Q is a singleton is a singleton AoA Q 6= 6= Proof : 37
Chapter 4
e 1. Assume AoA and consider any CT Ca attained by (x,X). Because the eT by CT = e compatibility condition it is only possible to dene the price of C X (0). If another strategy y,Y attains f T , X(0) = Y (0) because AOA, Hence C there is only one viable pricing of an attainable claim under AOA. 2. Under AOA, if the market is complete all claims are attainable, hence there is one and only one viable price for any contingent claim in C. 3. Under AOA and incomplete markets there are an innite number of viable prices for a non-attainable contingent claim ( 6= but is not a singleton). When AOA does not prevail no pricing system meets the compatibility condition, hence is empty. 4.B.ii Extensions
4.B.ii.a
Extension I.
e At time 0, CT = E Q
S0 (0) = E Q
e CT S0 (T ) |F0
= X (0)
4.B.ii.b
Extension II.
The portfolio is not self-nancing: Assume an adapted and integrable dividend payment (t) in [t, t + dt], then:
X (0) =
Q E0
(Z
(Z
38
Chapter 5
There are N + 1 assets in the market: r (t) being the adapted, locally risk-free rate, asset 0 is the corresponding risk-free asset with: dr = r (t) dt + 0r (t) dw
Rt
0
dS0 (t) = S0 (t) r (t) dt S0 (t) = e At t : dr (t) is not known, but dS0 is known. The N risky assets follow the process:
r(u)du
or for the i
th
where wM1 is the vector of standard Brownian Motions and N1 (t) and NM (t) are the two adapted processes 0i is the ith row of . The coecients of all these Ito processes are stochastic processes that satisfy integrability conditions. In terms of returns: dRi = or in vector form: dR = () dt+ () dw 39
dSi = i () dt + 0i () dw Si
Chapter 5
Rt
0
0 (.)dw i
The integrability conditions on the coecients are: Rt Rt (IC) |i (.)| du and 0 k i (.)k2 du dened a.s for 0 i = 1, ..., N They will be refered as the integrability conditions (IC) in the following chapters Ito process yields continuous sample paths, but they are not necessarily Markovian.
5.B
Diusions
S(t) follows a diusion process if: dS = (t, S (t) , r(t)) dt + (t, S (t) , r(t)) dw or dSi =i (t, S (t) , r(t)) dt + 0i (t, Si (t) , r(t)) dw or dR = (t, S (t) , r(t)) dt+ (t, S (t) , r(t)) dw or, equivalently
t 2 1 Si (t) = Si (0) e 0 [i 2 ki k ]du+
Rt
0
0 dw i
The process for the risk-free rate is: dr = r (t, r, S) dt + 0r (t, r, S) dw with the coecients ( () , () , () , ..), being a deterministic function of stochastic variables r, S and the deterministic t. The diusion process is an Ito process, hence it exhibits continuous sample paths. Moreover it is Markovian since the next increment depends on t and S(t) , r(t) only. Technical conditions to be satised bythe coecients of a diusion process are the Lipschitz condition and the linear growth condition. 40
Chapter 5
5.C
The state of the economy is dened by L state variables Y obeying the diusion SDE: dY = Y (t, Y (t)) dt + Y (t, Y (t)) dw (the coecients meet the integrability conditions). The dynamics of all the nancial variables depend on (t, Y (t)) ,i.e: dR = (t, Y (t)) dt+ (t, Y (t)) dw ; dr = r (t, Y) dt + 0r (t, Y) dw Remark that - The processes are Markovian - The simple diusion case is a particular case of the state variable diusion case (where S and r are the state variables); the state variable diusion case is a particular case of the Ito case. 5.D Theory in the Ito-Diusion Case
All the results on AOA, martingale measures, viable prices, completeness,.., presented in the case of RCLL asset prices and LCRL strategies hold of course when they follow Ito or diusion processes (which are continuous). We present in the following some specic results valid in these last cases. 5.D.i Framework
Assume the usual probability triplet [, F , P ] . Let wM1 denote the sources of uncertainties. The observable events at t are the events w(t0 ) a for all t0 t and all the real vectors a: roughly, information at t is represented by the path of w between 0 and t. {Ft , t [0, T ]} F w is then called the ltration generated by w. dR = (t)dt+(t)dw and dSi = Si i dt + Si 0i dw ; dr = r (t) dt + 0r (t) dw ; the coecients ((t), (t), ..) are F w adapted and satisfy the integrability conditions. Let () denote the instantaneous variance-covariance matrix, then: = dRdR0 dt dw dw0 0 = dt 0 = 41
Chapter 5
Let C be the set of contingent claims dened as L2 , as previously. If X, Y C, then E [XY ] is a scalar product and C is an Hilbert space A strategy can be dened by weights on risky-assets xN1 . The strategy will be denoted by (x,X) and the weight of the risk-free asset in the portfolio would be x0 = 1 x0 1 (x,X) is self-nancing if f dX X = x0 rdt + x0 dR = rdt + x0 (dRrdt1) i.e., the increment in value comes only from returns. Equivalently: dX = X dt + x0 dw X with X = rdt + x0 r1 5.D.ii Martingales
Theorem 7
(martingale representation theorem, stated without proof ). Consider any F w -adapted Martingale Z(t) : there exists an integrable process M1 () such that, for t (0T ) : Rt dZ = 0 (t)dw(t) Z(t) = Z0 + 0 0 (.)dw
In particular, for any (x, X) in A0 , under any Q Q: dXd = 0 dw (X d = X/S0 ) X and dX = r(t)dt+ 0 dw X We will see later that the probability change (from P to Q for instance) changes only the drift of the process but not the diusion term (this follows from Girsanov theorem stated further on). Since this diusion part would be x0 dw for a d portfolio (x, X),we can write under Q: dX d = x0 dw and dX = r(t)dt + x0 dw X X 5.D.iii Redundancy and Completeness
0 dimensional vector (t) = (1 , 2 , ..., N ) such that 0 dR= (t) dt a.s., i.e., a linear combination of risky assets gives a locally risk-free result.
Denition 22 The N + 1 assets are redundant at time t if there exists a non zero N -
42
Chapter 5
Proposition 4
The assets are not redundant if f Rank () = N, or, equivalently, iff the N rows of are linearly independent, or iff is a positive denite (invertible) matrix for all t a.s..
Proof
Assume that the assets are redundant, i.e., 0 dR = i i = N gives: dRN = dt + Apply the processes followed by Ri : N dt + 0 dw=edt + N For all dw; This implies: 0 = N
N1 X i=1 N1 X i=1 N1 X i=1
PN
i=1 i dRi
i dRi
i 0 dw i
i 0 i
i.e., the N th row of is a linear combination of the other rows. Therefore, Rank () < N . Q.E.D.
Remark 14 A result follows directly: a necessary condition for the assets to be non-redundant
is M N.
Theorem 8
Assume AOA, that M = N , that the coecients are adapted w.r.t. the ltration F w generated by w and that the N + 1 assets are non-redundant (hence Rank () = N t a.s..), then the market is complete w.r.t.. the ltration F w .
Proof 43
Chapter 5
Assume that the number of non redundant assets is equal to the number of brownians: N = M . Consider any contigent claim CT . We must proove that it is attainable by some strategy (n, X) in A0 . We use all along this proof a martingale measure Q Q and asset 0 as numeraire. Therefore we consider the discounted prices and values Sd and X d (which. are Q-martingales) Their dynamics, under Q, write: dSd = (t) dw. T We consider also the Q-martingale C d dened as: C d (t) = E Q [ SC(T ) /Ftw ]; Remark 0 T that C d (T ) = SC(T ) . We know (from martingale representation theorem) that a 0 Rt process Mx1 exists such that: C d (t) = C d (0) + 0 0 (t)dw a.s. Consider a strategy (n0 , n, X d ) (n represents here the N-dimensional vector of numbers of risky assets contained in the portfolio and n0 the number of riskless securities) satisfying at each time t the relations: (S) n0 (t)(t) = 0 (t); X d (0) = C d (0) Such a strategy is possible and is unique; Indeed: (S) is a system of M equations for N unknowns n0 (t) which yields the unique solution : n(t) = [0 (t)]1 (t) since M = N and is a square non singular matrix (the assets are non redundant); X d (0) = C d (0) sets the required initial investment. In a second step n0 (t) can be choosen in order to satisfy the self nancing condition. Then X d (t) is a Q-martingale and its dynamics write (under Q): dX d = (.)dt+ n0 (t)dS =n0 (t)(t)dw = 0 (t)dw = dC d with X d (0) = C d (0) d C d (t) = X d (t) for any t (0T ) Hence (n, X d ) duplicates C d (t) and reaches CT , This implies that (n, X) reaches CT . Any contingent claim is thus attainable (by only one strategy) and the market is (exactly) complete. When M > N it is not possible, in general, to solve the system (S), hence it is not possible to duplicate any CT , and therefore the market is not complete. QED 5.D.iv Criteria for Recognizing a Complete Market
Check the non-redundancy condition, i.e. check Rank () = N. Check if M = N M > N , the market is not complete M = N , the market is exactly complete if f the assets are non redundant M < N , necessarily there are redundant assets
44
Chapter 6
Framework
Chapter 6 Framework
The state of the economy depends on a vector Y of state variables Let wM1 denote the M -Brownian Motions vector and Y L1 denote the L-state variables vector with dY = Y (t, Y(t))dt + LM t,Y(t) dw Y(t) represent the random variable, Yt will denote a particular realization at t We consider N + 1 primitive securities (one risk-free, N risky). The returns of the N risky assets follow the diusion process: dRN1 = (t, Y(t))dt + NM (t, Y(t))dw or, for a single asset: dSi = i (t, Y(t)) dt + 0i (t, Y(t)) dw Si
dRi =
( 0i is the ith row of ) The risk-free rate follows the diusion process: dr (t) = 0 (t, Y(t)) dt + 0r (t, Y(t)) dw The price of the locally riskless asset follows: dS0 = r (t) S0 (t) dt 46
Chapter 6
Framework
The diusion process implies continuous sample paths and Markovian properties; In addition to the N+1 primitive assets we will deal with an indeterminate number of other assets (derivatives, funds, contingent claims). Consider X(t) one of them, a contingent claim attainable through (x, X). Since S0 (t) is a Q-martingale, if the state of the economy at t is Yt (the realization at t of Y(t) is Y t ) : Q X (T ) Q X (T ) E |Ft = E |Y(t) =Y t S0 (T ) S0 (T ) X (t) = f (t, Yt ) = S0 (t) The price at t of the claim writes thus: X (t) = (t, Y t ) The particular case where L = M and is invertible will sometimes be considered .
47
Chapter 7
Recall that we consider the variables Y satisfying; dY = Y (t, Y)dt + LM (t, Y) dw Consider v (t, Y) : [0, T ] RL R, with v C 1 w.r.t. t and v C 1,2 w.r.t. Y. Itos lemma writes in alternative forms: 0 v v 1 2v dv = dt + dY+ dY 0 dY t Y 2 YY 0 This gives: " # M L L L L X v X v X v 1 X X 2v + dv = + Vij dt + ij dwj t Yi Yi 2 i=1 j=1 Yi Yj Yi j=1 i=1 i=1
with Vij being the common term of V , 0 . Dene the Dynkin operator as:
t DY v =
7.B
Consider Y and v (t, Y) as previously dened. For given functions of (t, Y (t)), b (t, Y (t)), and l (Y), we try to nd the solution to the following problem (PDE 48
Chapter 7
Feynman-Kac theorem:
as an expectation: v (t, Y t ) = E
P
(u) e
(x)dx b
du + l (YT ) e
RT
t
(x)dx b
|Y (t) = Yt
The nancial interpretation of this is: v is the price of a nancial asset giving a dividend stream of and a terminal value of l (Y) is the required rate of return b
t DY v+ v
t is the expected instantaneous return with (DY v)dt being the capital gain and (t)dt the dividend during the period [t, t + dt].
The PDE states that the expected return is equal to the required ; Its solution b is the conditional expected value of the discounted stream of dividends + the terminal value, the discount rate being . This is also CIR(1985), lemma III. b Feynman-Kac theorem provides a link between the PDE approach and the martingale approach. However since we do not know the required expected return b the PDE or its solution interpreted as a discounting at rate does not give the b value v. But in the following we are going to provide an APT condition on . b
49
Chapter 8
8.A.i
Consider an asset yielding a dividend stream of and a terminal value of l (Y). We have derived that: 0 v t dv = (DY v)dt + dw Y Divide both sides by v gives: 1 dv 1 t = DY v dt + v v v 0 = v dt + v dw v Y 0 dw
t Here = 1 DY v can be considered as the expected rate of return and 0v = v 1 2v v , v , ..., M the volatility vector or sensitivity w.r.t.. w. v
Also, dene
1 = v v then
v Y
0 dv 1 v = v dt + dw v v Y 0 1 v = v dt + dYY dt v Y 0 1 v = dt + dY v Y 50
Chapter 8
More explicitly we get two alternative decompositions of the return: X dv = v dt + i dwi v v i=1
L M
1 X v = dt + dYi v i=1 Yi
1 v v Yi
the sensitivity
8.A.ii
In the following () denotes (t, Y(t)). The following proposition is the continuous time version of APT and can be justied as the discrete time version. Proposition 5
(APT)
1.
There exist M scalars: 1 () , 2 () , ..., M () such that, for any asset (value v return stream= , required expected instantaneous return = ): X () + () = r () + i () i () v v i=1 i () is the market price of the risk wi and is the same for all assets. The equation above can be deemed as a decomposition of the expected rate of return into the riskless rate and M risk premiums: i () is the market price of risk (MPR) wi ; The MPR vector is the same for all assets.
M
2.
There exist L scalars: 1 () , 2 () , ..., L () s.t., for any asset: X 1 v () + () = r + j () () () v v Yj j=1 This is an alternative decomposition of the expected rate of return with L risk premia (relative to risks Y ). j is the market price of risk (MPR) Yj , and is also the same for all assets. We will drop () in the following for simplicity 51
L
Chapter 8
In the particular case that L = M and then is invertible, can be solved as: = 1 Furthermore, if we apply APT to the ith primitive asset: dRi = i dt + 0i dw i = r + 0i where 0i is the ith row of . In vector form: = r 1+ This equation may be used in two ways: - To obtain the required returns for a given MPR .Then, the returns of the primitive risky assts follow: dR= [r(.)1 + (.)(.)]dt + (.))dw - To obtain (or estimate) the MPR assuming that the risk premiums r1 are known (or estimated). This is only possible when is invertible (M = N and non redundant assets, implying market completeness), in which case: = 1 r1
Under incomplete markets an innite number of MPR vectors are compatible with the risk premia on the primitive securities. It is important to note that:
t DY v + 1 =r+ v v
v Y
This PDE means that the expected rate of return equals the required rate of return. It must be followed by any asset in a world described by Y. The only dierence between assets is the boundary condition v (T, Y) = l (Y) specic to each asset. Example 5 In the Black-Scholes framework Y = S, L = M = N = 1,for a call: l(Y ) =
(Y K)+ , = ( r)/
52
Chapter 8
Example 6 A one factor interest rate model with a stochastic risk-free rate r, which is
also the state variable. Only bonds are considered and one bond is sucient (the others are redundant), therefore, L = M = N = 1. We consider such models in the following section.
8.B
L = M = 1, and now Y1 =r dS0 = S0 r (t) dt Let BT (t, r (t)) be the price of a zero coupon bond at t that delivers 1 at T . The Several BT may be traded (but they are redundant). dr = a r] dt + r (t, r)dw. In Vasicek model r (t, r) = constant, and in CIR model [b
r = r duration of the bond is then T t
This gives: BT BT BT 1 2 BT + 2 + a (b r) = rBT + r t 2 r2 r r with the boundary condition that BT (T ) = 1 T . The PDE can be solved in both Vasicek and CIR settings. 8.C Discounting Under Uncertainty
+ = rv +
v Y
; LC : v(T, Y) == l(Y)
Chapter 8
Note that left hand side of this equation can be interpreted as the Dynkin operator b computed w.r.t. a drift (Y ) dierent from Y . Hence now dene Y (t) = Y (t) i h b and dY (t) = Y dt + dw, we now have an equivalent but simplied writing:
t DY v + = rv b
By Feynman-Kac, the solution is: Z T R R tu rdx tT rdx b P b e du + l(Y(T )) e |Y (t) = Y (t) = Yt v t, Y t = E t h i b and Y (t) = Y dt + dw Note that we now discount with r instead of ! (CIR 1985, lemma IV) We can safely state that the value of any asset is the expected discounted value of future cash ows with r as the discount factor provided that the drift of Y is adjusted by the MPR of the risks Y, which is . Alternatively we could express the valuation formulae in function of the MPR of the risks w (substitute in the formulae for )
54
Chapter 9
1.
Proposition 6 If Q P , then there exists a random variable which is F measurable with E P [] = 1 and > 0 a.s. such that A F: Z Q (A) = ($) dP ($)
A
Then = 2.
dQ dP
= E P [1A ]
Any FT measurable, with E P [] = 1 and > 0 a.s. is a valid Radon Nikodym derivative, meaning that a new probability measure Q can be dened by dQ = or Q (A) = E P [1A ] A F (and Q P ). dP Proof
We proove only part 2. We check rst that Q is a probability measure; Indeed: Q () = E P [1 ] = E P [] = 1 Moreover, for A B = , Q (A B) = = = = E P [1AB ] E P [(1A + 1B ) ] E P [1A ] + E P [1B ] Q (A) + Q (B) 55
Chapter 9
We check that Q P ; Indeed, since > 0 a.s.., Q (A) = E P [1A ] = 0 P (A) = E P [1A ] = 0
Q.E.D.
The intuition behind the changing of probability is that, by considering the probability of an event as a mass, the probability, or the mass, is changed by multiplying a positive ($) (dQ ($) = ($) dP ($)). Also, X with E P [X] < , E Q [X] = E P [X ].
9.B
Girsanov Theorem
Consider the m-dimensional Brownian Motion w under the probability measure P , the ltration Ft (t (0, T )) generated by w, and the m-dimensional adapted Rt process () that satises integrability conditions ( 0 k (s)k2 ds dened,..). Theorem 9
(Girsanov Theorem) a) Dene (t) = e 2 0 k(s)k ds 0 (s)dw , then (T ) is a valid Radon-Nikodym derivative: (meaning that: (T ) is FT measurable; E P [] = 1; 0 a.s.); Moreover (t) is a P-martingale. Rt e b) Dene Q P by dQ = (T ), then w (t) , w (t) + 0 (s) ds is a standard QdP Brownian Motion.
1
Rt
Rt
9.C
Consider: dR = dt + NM dw We had: = r1 + 56
Chapter 9
Therefore: dR = [r1 + ] dt + dw We now look for a probability measure Q under which the dynamics of R have a Si drift of r1 (then the asset 0 denominated values S0(t) would be Q-martingales). (t) We know that Q exists and is unique when the market is complete. Proposition 7
Consider the probability Q , equivalent to P , dened by the Radon-Nikodym derivative:
RT Rt 0 2 1 dQ = e 2 0 k(s)k ds 0 (s)dw dP
where () is the market price of risk. Then the instantaneous Qexpected return of Si any self-nancing asset is r. Moreover, asset 0 denominated values S0(t) (as well as (t) the asset 0 denominated values of self nancing portfolios Q is thus a risk neutral probability.
X(t) S0 (t)
) are Q-martingales.
Proof
e By Girsanov Theorem, w (t) , w (t) + e dw (t) Rt
0
We now dene:
, dw (t) + (t) dt e dR = [r1 + ] dt + [dw (t) (t) dt] e dR = r1dt + dw (t) S (t) b S (t) , S0 (t) = d log = = S 1 + 2 dt S0 2 S 1 dS dS0 1 2 S dt + 2 dt S S0 2 2 S dS rdt S
b dS(t) b S(t)
If the drift of
dS S
are Q-martingales.
57
Chapter 9
Remark that when the set of primitive securities is complete (N = M and is invertible) only one (= 1 r1 ) is compatible with the primitive returns. Under incomplete markets an innite number of s are compatible with the assumed return dynamics and an innite number of risk free probabilities can be constructed.
58
Chapter 10
Introduction
Chapter 10 Introduction
Choose asset 0 as numeraire; then: Si (t) b Si (t) , S0 (t)
In AoA, there exists a set of probabilities Q under which the 0-denominated prices are martingales. Q is a singleton if f market is complete. In this case, write Q as P0 ; P stands as usual for the true probability. 0 is not a compelling choice for numeraire. Alternatives include BT (t). Then, the Si forward adjusted probability yields martingale prices Si (t)= BT(t) (t) The intuition behind the change of numeraire is expressing the value of an asset in units of another asset.
60
Chapter 11
11.A.i
Assets
Asset returns follow are Ito process following the SDE: dR = dt + dw dS0 = rS0 dt dr = r dt + r dw The coecients (, , r , r ) are stochastic but satisfy Ito technical conditions Assume N M, where N is the number of non-redundant risky assets, and M is the number of Brownian Motions.
11.A.ii
Numeraires
Denition 23 A viable numeraire is an admissible self-nancing portfolio with positive values a.s., i.e., (n, N) is a viable numeraire if: (n, N ) A0 N (t) > 0 a.s.
Rt
0
r(u)du
61
Chapter 11
11.B.i
We have studied 0 P0 , now consider any viable numeraire n and the correspondence n Pn . Denition 24 Pn stands for the set of probabilities equivalent to P yielding n-denominated o n
martingale prices: Pn = Pn P | (x, X) A0 , X(t) is a Pn martingale N(t)
Proposition 8
For any admissible numeraire (n, N ) and Pn :
1.
There is a one-to-one correspondence between P0 (the set of risk-neutral probabilities) and Pn . Moreover, P0 P0 , P0 Pn is given by: N(T ) S0 (0) dPn = dP0 S0 (T ) N(0) and Pn Pn , Pn P0 is given by: dP0 S0 (T ) N(0) = dPn N(T ) S0 (0)
2.
Pn 6= AoA ; Pn is a singleton market is complete Proof 1. Consider a viable numeraire (n, N ). Without loss of generality assume N(0) = S(0) = 1. N(T Let = dPn = S0 (T) S0 (0) . is a viable numeraire; indeed it satises the three dP0 ) N(0) requirements: (i) is FT -measurable (since N (T ) and S0 (T ) are FT -measurable); (ii) > 0 a.s. (since N (T ) and S0 (T ) are > 0 a.s.); N(t) N N(0) (iii) EP0 () = 1; Indeed, since S0 (t) is a P0 martingale: EP0 ( S0(T ) ) = S0 (0) . There(T ) N(T fore EP0 () = EP0 S0 (T) S0 (0) = 1. ) N(0) Then denes uniquely a probability Pn P0 . Moeover such Pn belongs to Pn ; 0 indeed, for any (x, X) in A : 62
Chapter 11
N(T X(T EPn ( X(T ) ) = EP0 ( X(T ) ) = EP0 ( X(T ) S0 (T) S0 (0) ) = EP0 ( S0 (T) ) S0 (0) = N(T ) N (T ) N(T ) ) N(0) ) N(0) X(0) , N(0)
hence X is a PN -martingale. An analogous argument shows N denes the correspondance between Pn and P0 and the two sets are thus in a one to one relation through this procedure. 2. Follows from 1. and from the fact that 2. is true for P0 Q.E.D. More generally, given any (n, N ) and (n0 , N 0 ) N , there is a one-to-one correspondence between Pn and Pn0 dened by: dPn0 N 0 (T ) N(0) = dPn N(T ) N 0 (0) with the possible assumption that N (0) = N 0 (0) = 1 11.B.ii Probability Numeraire
11.C
Summary
In AoA: There exists a set P0 of probabilities equivalent to P (the true or historical probability)
such that P0 P0 and (x,X) A0 (the set of admissible and self-nancing strategies), X(t) the S0 (t) denominated price S0 (t) is a P0 -martingale.
P0 is a singleton when the market is complete. Dene N as the set of all admissible numeraires. Then N N iff (n,N) A0 and
N (t) > 0 a.s..
is a Pn -martingale.
63
Chapter 11
In the general case, given any (n, N) and (n0 , N 0 ) N , there is a one-to-one
correspondence between Pn and Pn0 dened by: N 0 (T ) N(0) dPn0 = dPn N (T ) N 0 (0) with the possible assumption that N (0) = N 0 (0) = 1
Then why not consider a portfolio (h,H) that yields martingale prices under the true probability P ? This question is addressed in the following section.
64
Chapter 12
12.A.i
Denition 25 The numeraire portfolio is the unique portfolio (h, H) N such that (x, X)
A0 ,
X(t) H(t)
is a P -martingale, i.e.:
Et
where Et [] = E [|Ft ] is the conditional expectation computed with the true probability.
The price of the asset is then given as: X(t) = H(t) Et with
H(t) H(T )
X(T ) H(T )
(h, H) is also called the growth optimal portfolio (Merton), the log-optimal portfolio, and the numeraire portfolio (Long). 12.A.ii Characterization and Composition of (h, H)
1.
Theorem 10 (h, H) is the strategy that maximizes the expected log of the terminal wealth W (T ), i.e. it solves the program:
(x,X)A0
max E0 [log X (T )]
65
Chapter 12
2. 3. 4.
(h, H) maximizes the expected growth rate in [0, T ] Consider 0 < T < T 0 , the strategy (h, H) that maximizes E [log W (T 0 )] also maximizes E [log W (T )] (the so-called myopic property of log) (h, H) dominates (in probability) asymptotically any other strategy in A0 , i.e. (x, X) A0 : lim Pr oba [H (T ) X (T ) > 0] = 1
T
Remark 15 However, this probabilistic domination does not mean that limT E0 [U (H (T ))] > E0 [U (X (T ))] for all well-behaved utility function U and all (x, X) A0 : the statement that all investors should choose (h, H) in the long run is not valid.
Proof
1.
Starting with the same initial investment H(0) = X(0), H(T ) is preferred by the log investor to X(T ); indeed, by Jensens Inequality for concave functions and the fact that (h, H) is a P numeraire, (x, X) A0 X (T ) X (T ) log E E log H (T ) H (T ) X (0) = log H (0) = 0 (since X(0) = H(0) = 1) Therefore: E log H (T ) E log X (T ) i.e. (h, H) maximizes the expected log of terminal wealth.
2.
Dene the (annualized) growth rate in [0, T ] as r0,T such that X (T ) = er0,T T = eR0,T X (0) or : r0,T =
log[X(T )]log[X(0)] ; T
then 2 is obvious.
3.
Chapter 12
Maximizing E [R0,T 0 ] then implies a two-step maximization procedure: max E [R0,T ] and max E [RT,T 0 ].
Q.E.D.
The previous results are also valid for semi-martingale prices, but in the following, it is assumed that prices and rates of return follow Ito process, i.e. dR = dt + dw dr = r dt + r dw with R an N 1 vector, an N M matrix with rank N (there is no redundant asset, so N M ), and the coecients obey the usual technical conditions so that there exists a solution for the SDE. Theorem 11
The composition of h is: h = 1 r1
Proof
Because the Log utility is myopic, optimizing over (t,T ) implies optimizing over [t, t + dt] ;In [t, t + dt] max Et [log X (t + dt)]
x x x
max Et [log X (t + dt) log X (t)] max Et [d log X] By Itos lemma, d log X =
dX X
1 2
dX X
= (1 x0 1) rdt + x0 dw = r + x0 r1 dt + x0 dw
dX 2
X
With dwdw0 = I dt
67
Chapter 12
Leave out dt and apply the rst order condition gives: r1 h = 0 h = 1 r1 Q.E.D.
Corollary 5 (h,H) is instantaneously mean-variance ecient, hence homothetical (proportional) to the tangent portfolio (m,M), i.e. h = 1 m kt = 10 1 r1 m 1 10 1 r1
with
kt = and
(the weights in m sum up to one while in h they dont; h is a combination of m and asset 0). We have not excluded the possibility that N < M . But now assume that N = M ; then:
m = kt 1 r1
Corollary 6 When the market is complete and NN is invertible, the market price of risk can be derived as a function of the risk premia6 :
= 1 r1
The composition of the numeraire portfolio can now be expressed in terms of as: h = 1 r1 = 01 1 r1 = 01
6
When the market is not complete, cannot be explicitly specied. For the ith asset, for instance,
M X
i r =
ij j
j=1
With N < M , the system of N equations does not yield an unique solution for 1 ,...,M
68
r + h0 r1 dt + h0 dw h 0 i 0 r + 01 r1 dt + 01 dw r + 0 1 r1 dt + 0 1 dw r + 0 dt + 0 dw r + kk2 dt + 0 dw
t 2 1 H(t) = e 0 {r+ 2 kk }du+
Rt
0
0 dw(u)
2 The risk premium, as well as the instantaneous return variance of (h, H), are kk .
12.A.iii
The Radon-Nikodym derivative has been shown to be the ratio of two numeraires, i.e. N 0 (T ) N(0) dPn0 = dPn N(T ) N 0 (0) and usually with the assumption that N(0) = N 0 (0) = 1. Consider the risk neutral probability P0 and the corresponding numeraire S0 : dP0 S0 (T ) = dP H(T ) =
RT RT
0
r(u)du R
This result was obtained previously through a dierent approach (see Girsanov theorem) 12.B First Applications
69
Chapter 12
12.B.i
CAPM
Theorem 12
Consider S A0 with: dS S dH H (CAPM) In AoA: = S dt + S dwS o n = r + kk2 dt + H dwH S r 1 kt HS = HS = = 1 MS kt
M r 2 M = S H dwS dwH
Proof
S(t) H(t)
is a P -martingale, therefore it should have zero drift. By Itos lemma: dS dH dS(t) = + dH(t) S H dH H 2 dS dH S H
The drift term is: o n S r + kk2 + kk2 HS = 0 S r = HS (The rest remains to be proved). Q.E.D.
Therefore:
12.B.ii
Valuation
H(t) XT H(T )
70
Chapter 12
is called as the pricing kernel or the state price deator. Its product with H(t) P ($), i.e. H(T ) P ($), is the Arrow-Debreu price. For a security with terminal cash ow XT and a dividend stream (t)dt, the pricing formula is: Z T 1 (t) X(0) = Eo XT + Eo dt H(T ) 0 H(t) with the discount factor Rt Rt 0 2 1 1 = e 0 {r(u)+ 2 kk }du 0 dw(u) H(t) In the certainty (riskless) case, 1 = ert H(t)
71
Chapter 13
13.A
Framework
13.A.i
Let L be the number of state variables and Y be the vector describing the states of the economy, with dYL1 = Y (t, Y(t)) dt + LM (t, Y(t)) dwM1
There are N + 1 assets in the economy. The returns of the N risky assets follow the diusion process: dRN1 = (t, Y(t)) dt + NM (t, Y(t)) dwM1 73
Chapter 13
S0 (t) = e
r(u)du
The usual technical (integrability) conditions and a frictionless and continuously open market are assumed.
13.A.ii
where the rst term in the right hand side of the equation is the gain of the portfolio, the second term Consumption, and the third term Income.
For an investor with wealth X(t) at time t, the Budget Constraint in [t, t + dt] is: dX = X(t) r(t) + x0 (t) (t)r(t) 1 dt + x0 (t)dw c(t)dt + y(t)dt
where U () and B () are both utility functions (B () may be interpreted as the utility of a bequest). They are state independent but they do depend on t. Additional restrictions imposed on U () and B () are: U () and B () are strictly concave. U 0 (0) and B 0 (0) = ; U (0) and B (0) = so as to preclude negative consumption
or bequest.
13.B
The Solution
13.B.i
74
Chapter 13
1.
The indirect utility function7 J(t, Xt , Y t ) is the solution (value function) to the objective maximization program M with Xt and Yt being the realization of X(t) and Y(t) at time t and the additional constraint that returns follow the previously specied process. Consider the period (t + dt, T ). Whatever has been done in [t, t + dt], which yields X (t + dt), we optimize between [t + dt, T ], hence deriving a utility J(t + dt, X (t + dt) , Y (t + dt)). Let Et be the conditional operator E(/Xt ,Y t ).Then at time t, we maximize Et [U (c (t) , t) dt + J(t + dt, X (t + dt) , Y (t + dt))]. To express the fact that J(t + dt, X (t + dt) , Y (t + dt)) depends, through X(t + dt),on the choice (c(t), x(t)) made at t for (t, t + dt), we write it also: J c,x (t + dt, X + dX, Y + dY) Now write the maximization program M as: max Et [U (c (t) , t) dt + J c,x (t + dt, X + dX, Y + dY)]
c,x
2.
3.
max Et [U (c (t) , t) dt + J c,x (t + dt, X + dX, Y + dY) J (t, X (t) , Y (t))] c,x h i c,x max Et U (c (t) , t) dt + DY J dt c,x h i c,x max Et U (c (t) , t) + DY J
c,x c,x
The dierence J c,x (t + dt, X + dX, Y + dY) J (t, X (t) , Y (t)) is nothing else but the The maximum of U () dt + J (t + dt) is J (t), therefore, the maximum obtainable value
of () is zero.
c,x Dynkin operator DY J applied to J for given values c and x of the control variables at t. c,x (c (t) , x(t), t, X (t) , Y (t)) stands for Et [U (c (t) , t) + DY J]
X 1 1 XX (J) = Jt + JX X + JXX 2 + JYi JYj dYi dYj + JX JYj dXdYj X 2 2 i=1 j=1 j=1
In the simple static framework the indirect utility is dened as follows: Given unit price pi and wealth y, the optimal allocation (x , x , ..., x ) on N goods is derived from the direct utility 1 2 P N function V (x1 , x2 , ..., xN ) by maxx V s.t. y = pi xi ; The solution, (x , x , ..., x ), is function of y 1 2 N and P The indirect utility function J (y, P), is then dened as V (x , x , ..., x ) i.e: the maximum 1 2 N utility that can be derived with a wealth y given the prices P
75
Chapter 13
where: X = X r () + x0 () r () 1 c (t) 2 = X 2 x0 x X = 0 dYdY0 = 0 dt, with 0 = V or: dYi dYj = Vij dt dXdYi = X (x0 dw) 0i dw = Xx0 i dt where 0i is the ith row of . The problem is now:
c,x max{U (c, t) + DX,Y (J)} c,x
Recall that the maximum of () is zero; We obtain thus two equations (Bellman conditions): max (t, c, x; X (t) , Y (t)) = (t, c (t) , x (t) ; X (t) , Y (t)) = 0
c,x (1) (2)
(a) The rst equation implies the rst order conditions: (t, c (t) , x (t) ; X (t) , Y (t)) = 0 c and (t, c (t) , x (t) ; X (t) , Y (t)) = 0 x (b) The second equation implies that the PDE governing J is: c ,x U (c , t) + DX,Y (J) = 0 76
Chapter 13
The PDE governing J (theoretically) determines J, which in turn allows the comc,x putation of (via DY (J) computed previously). Then the rst order conditions will determine the optimal (c (t) , x (t)). This methodology yields only in some few cases a closed form solution for the optimal portfolio x , but it allows to prove that it has the general form presented in the following paragraphs.
13.B.ii
Proposition 10
The optimal portfolio as a solution of program M writes: x () = A () h + where
L X k=1
Ak () hk ()
Ak depends on the indirect utility function and is its sensitivity w.r.t. dYk h is the (instantaneously mean-variance ecient, ...) log optimal portfolio studied in a previous chapter: h = 1 r1
hk (k = 1, 2, ..., L) are hedge portfolios with a return perfectly correlated with dYk ; they are aimed to hedge against unfavorable shifts of the investment opportunity set induced by movements of Yk . Proof
Follows from the Bellman conditions (1) and (2) (see Merton) Q.E.D.
77
Chapter 13
Corollary 7 When L stands for the number of state variables on which the investment
opportunity set depends, any optimal portfolio is a combination of the following L + 2 funds: {0, h, h1 , h2 , ..., hL }. Moreover, the funds are the same for all the investors with a well behaved utility function: This result is an L + 2 funds separation.
Two particular cases: (a) r is constant and Si are geometric Brownian Motions. Then L = 0 and we have 2-fund separation. (b) r is stochastic and is the only state variable, so L = 1 and we obtain 3-fund separation.
13.B.iii
Intertemporal CAPM
Consider rst the particular case where r is constant and Si are geometric Brownian Motions. Therefore L = 0 and we have 2-fund separation: x () = A () h All investors hold a mix of the same funds 0 and h, though this mix depends on their individual risk tolerance A ().
Recall that the log-optimal portfolio h is homothetical to the instantaneously tangent portfolio m: h=m where m is the tangent portfolio containing only risky assets. Since all investors hold the same risky portfolio m,this portfolio is the market portfolio: all investors hold instantaneously mean-variance ecient portfolios and m is instantaneously mean-variance ecient. Hence an instantaneous version of CAPM holds: S :
S r = S (m r) where S = cov (dRS , dRm ) var (dRm )
78
Chapter 13
Consider now the general case with L state variables (and thus L + 2 fund separation). The generalized CAPM writes: S r = m (m S r) +
L X k=1
k (bk r) S
where k is the expected rate of return of the hedge portfolio hk , and b k = S cov (dRS , dRk ) var (dRk )
79
Chapter 14
We will assume in the following that: The market is AoA S follows a multivariate Ito process: dR=()dt + ()dw 14.A.i The pure portfolio problem
In this section the market may be incomplete Let us specialize the Merton problem into a pure portfolio decision: max E [U (XT )] x dX = r () + x0 r1 dt + x0 dw X s.t. X (0) = X0 where U (XT ) = B (XT ) is the concave bequest function, and the rst constraint is the self-nancing condition. The solution of M is (x , X ), with X (T ) XT . Also consider program P involving two optimization steps (P1 ) and (P2 ): (P1 ) :
XT L2
and
80
Chapter 14
The solution to (P1 ) is the random contingent claim XT . But since XT C a (i.e. the set of attainable contingent claims), it is attainable through at least one strategy (x , X ) A0 with X (0) = X0 .
1. 2. 3.
Proposition 11 The solution X (T ) obtained through (M) solves (P1 ) The solution (x , X ) obtained through (P) solves (M)
Under mild technical conditions (e.g., U () strictly concave), XT = XT a.s..
Proof
1.
X (t) H(t)
is a P -martingale,
Hence X (T ) meets the constraints of (P1 ). but, since X (T ) is the solution of (P1 ): E [U (X (T ))] E [U (X (T ))]
2.
On the other hand, (x , X ) is self-nancing and feasible (thereby satisfying the constraints of M) and (x , X ) solves M, so: E [U (X (T ))] E [U (X (T ))] We can conclude from the two last inequalities that: E [U (X (T ))] = E [U (X (T ))] Thus X (T ) solves (P1 ) and (x , X ) solves (M)
3.
(proof by contradiction) Suppose XT and XT are not equal a.s.. Then there exists a subset D of with strictly positive measure where $ D XT () 6= XT () X + XT b XT T 2
Consider
81
Chapter 14
b XT is attainable by a buy-and-hold strategy with initial weight 1 in (x , X ) 2 and 1 in (x , X ). Because of the strict concavity of U (), $ D: 2 XT + XT bT U X U 2 U (XT ) + U (XT ) > 2
Since $ D, XT () = XT (): h i E [U (X )] + E [U (X )] T T b E U XT > = E [U (XT )] = E [U (XT )] 2 That is to say that there exists a feasible and attainable strategy which yields a higher expected utility than either XT or XT . This is in contradiction with our maximization programs M and (P1 ).
Q.E.D.
The static program (P1 ) can be interpreted as a static choice of a contingent claim, or a self-nancing strategy, out of an innite number of them with a static h i budget constraint E X (T)) | F0 = X0 H(T 14.A.ii The consumption-portfolio problem
The dynamic program M writes: Z T M : max E U (c (t) , t) dt + U (XT ) x, c 0 dX = X r + x0 r1 dt + x0 dw c (t) dt s.t. X (0) = X0 The static program P writes:
(P1 ) : max E
XT ,c
"Z
U (c (t) , t) dt + U (XT )
T
s.t. : E0
"Z
# XT c (t) dt + = X0 H (t) H (T )
82
Chapter 14
Call XT the solution of (P1 ) (P2) : nd (c (t), x (t)) yielding a nal wealth = XT a.s.
This problem is conceptually similar to the pure portfolio problem, although more general. As in the pure portfolio problem the solutions of M and P coincide.
14.B
In this section the market is assumed dynamically complete 14.B.i Solution of the pure portfolio problem
Here we assume that U () is strictly concave and dierentiable; Therefore U 0 is decreasing and U 01 exists (we write V U 01 ) Since the market is complete all contingent claims are attainable (C a =L2 ). Thus, (P1 ) becomes: (P1 ) : max E [U (XT )] XT s.t. : E0 = X0 HT
XT L2
Theorem 13
The solution XT of (P1 ) writes: XT
=V
HT
Proof
83
Chapter 14
We prove the theorem is a discrete (countable) state space . Our optimization program is now: X (P01 ) : max p ($) U (XT ($))
XT ($) $
p ($)
XT ($) = X0 HT ($)
p ($)
XT ($) HT ($)
p ($) =0 HT ($)
($) = U
01
HT ($)
where is such that the budget constraint is satised: 1 01 E0 U = X0 HT HT The more general proof for continuous measurable state space involves calculus of variations (Euler Theorem).
Q.E.D.
Remark 16 The optimal terminal wealth XT is the solution to (P1 ). The optimal portfolio
strategy x as solution to (P2 ) remains to be found. This is, however, a dicult task.
XT > 0 a.s.. In general, to make the problem more realistic, an additional constraint such as XT > K a.s. may be involved. It can be shown that in this case the solution is XT , attainable with an initial wealth =X0 P without constraint, plus a put on XT with strike K (P is the price of this put).
Remark 17 XT may be negative, although for some utility functions, such as CRRA,
84
Chapter 14
14.B.ii
= Thus:
1 X0
XT = X0 HT
1 X
HT
1 1
or:
XT
HT
Note that for = 0, the CRRA utility becomes the log utility.
85
Chapter 14
1 2 X 2q
HT
So the optimal portfolio would consist of a long position in q units of zero coupon bond and a 1 short position in H(t) .
14.B.iii
Consider a discrete (countable) state space . The Lagrangian writes: Z T X L (XT , c (t) , ) = p ($) U (c (t, $)) dt + U (XT ($))
$ 0
p ($)
L 0 = p ($) U (c (t)) =0 c ($, t) H (t) and L 0 = p ($) U (XT ($)) =0 XT ($) HT ($) U 0 (c (t)) = H (t)
Thus:
86
Chapter 14
and
U 0 (XT )
HT
14.B.iv
Consider the case of state variable model where the state is described by Y (t): dY (t) = Y(t) () dt + dw
The method can be sketched as follows: (a) By martingale property: X (t) = H (t) E
XT | Yt H (T )
(t, Y t )
The diculty comes from the computation of (t, Yt ). Apply Itos lemma: dX d = X 1 = [] dt + 0
dw
(b) Note also that (x , X ) is a portfolio strategy, so: dX = [] dt + x0 dw X (c) Therefore, by identication: 1 x =
0
Chapter 14
which determines x0 provided that is known. More powerful tools (e.g., Maliavin calculus) are useful for solving this problem.
14.C
Recall that: U 0 (c (t)) = This implies (by dierentiation): [] dt + U 00 (c (t)) dc = dH [H (t)]2 H (t)
Dividing through by U 0 (c (t)) and use the rst order condition again gives: [] dt + or: c From CAPM we have: S S r = HS 1 dS dH = cov , dt S H Thus: 1 S r = cov dt or: dS U 00 (c ) dc , c 0 S U (c ) c dH U 00 (c ) dc + [] dt = 0 (c ) c U H dH U 00 (c (t)) dc = 0 (c (t)) U H
00 U (c ) S r = c c S U 0 (c )
This is the Consumption based CAPM (CCAP M ). Remark 18 The [] dt term does not enter the covariance for it is deterministic. 88
Chapter 14
THE EQUIVALENT STATIC PROBLEM (Cox-Huang, Karatzas approach) h i 00 (c ) c U 0 (c ) is the relative risk aversion of the representative agent. U
Remark 19
89
Chapter 15
The problems
91
Chapter 16
The optimal terminal wealth in the CRRA, mean-variance and HARA cases
Chapter 16 The optimal terminal wealth in the CRRA, mean-variance and HARA cases
16.A Optimal wealth and strong 2 fund separation
16.B
92
Chapter 17
17.B
93
Chapter 18
18.B
94
Chapter 18
References
ARTICLES
BAJEUX I. , R. PORTAIT, 1992, Mthodes probabilistes dvaluation et modles variables dtat: une synthse, Finance BAJEUX I, R. PORTAIT, 1997, The numeraire portfolio : A new perspective on nancial theory, European Journal of Finance December BAJEUX I., R. PORTAIT, 1998, Dynamic asset allocation in a mean-variance framework, Management Science November BAJEUX I. JORDAN J. and R. PORTAIT, 1999, Dynamic asset allocation for stocks, bonds and cash BAJEUX I. JORDAN J. and R. PORTAIT, 2000, An asset allocation puzzle : comment. Forthcoming, American Economic Review CANNER N., N.G. MANKIW AND D.N. WEIL, 1997, An asset allocation puzzle, American Economic Review. COX J. and C.F. HUANG, 1989, Optimal consumption and Portfolio policies when asset prices follow a diusion process, Journal of economic Theory , 49. COX J., J, INGERSOL and S.ROSS., 1985, An intertemporal general equilibrium model of asset prices, Econometrica ,53 DYBVIG P. and C.F. HUANG, 1989, Nonnegative wealth, AOA and feasible consumption plans, Review of Financial studies vol 1, number 4 GEMAN H., EL KAROUI N. and ROCHET J.C, 1995, Changes of numeraire arbitrage and option prices, Journal of Applied Probabilities, 2 HANSEN L.P. and RICHARD S., 1987, the role of conditioning information in deducing testable restrictions implied by dynamic asset pricing models, Econometrica, vol 55, N 3, 587-613. HARRISON J.M. and K. KREPS, 1979, Martingales and arbitrage in multiperiod security markets, Journal of Economic Theory, 20. HARRISON J.M. and S. PLISKA, 1981, Martingales and the Stochastic integrals in the theory of continuous trading, Stochastic processes and their applications ,11 95
Chapter 18
HE H. and PEARSON N.D., 1991, consumption and portfolio policies with incomplete markets and short sale constraints, Journal of Economic Theory, 54, 259-305. KARATZAS, I., J. LEHOCZKY and S. SCHREVE, 1987, Optimal Portfolio and Consumption Decisions for a Small Investor on a Finite Horizon,. SIAM Journal of Control and Optimization , 25 , 1157-86. LONG, J.B., 1990, The Numeraire Portfolio, Journal of Financial Economics, 26, 29-69 NGUYEN P. and R.PORTAIT, 1999, Dynamic Mean Variance Eciency and Asset Allocation with a Solvency Constraint; Forthcoming, Journal of Economics Dynamics and Control RICHARDSON, H., 1989, A Minimum Variance Result in Continuous Trading Portfolio Optimization, Management Science, Vol 35, N 9
BOOKS
DEMANGE G, ROCHET JC, 1992, Mthodes mathmatiques de la nance, Economica DUFFIE D, 1996, Dynamic asset pricing theory Princeton University press. DUMAS B, ALLAZ B, 1995, Les titres nanciers, PUF (English version available) KARATZAS I, SHREVE S, 1999, Methods in Mathematical Finance, New York: Springer Verlag. KARATZAS I, SHREVE S, 1991, Brownian motion and stochastic calculus, New York: Springer Verlag. MERTON R., 1992, Continuous time Finance, Oxford : Basil Blackwell. MUSIELA M, RUTKOWSKI M, Martingale methods in nancial modeling, Applications of Mathematics, Springer, 19
96