Lifetime Portfolio Selection under Uncertainty: The Continuous-Time Case

Author(s): Robert C. Merton

Source: The Review of Economics and Statistics, Vol. 51, No. 3 (Aug., 1969), pp. 247-257
Published by: The MIT Press
Stable URL: http://www.jstor.org/stable/1926560 .
Accessed: 25/06/2014 06:35

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

The MIT Press is collaborating with JSTOR to digitize, preserve and extend access to The Review of
Economics and Statistics.

http://www.jstor.org

This content downloaded from 194.29.185.251 on Wed, 25 Jun 2014 06:35:18 AM

All use subject to JSTOR Terms and Conditions
 LIFETIME PORTFOLIO SELECTION UNDER
UNCERTAINTY: THE CONTINUOUS-TIME CASE
Robert C. Merton *

I Introduction tion must be generalized to become a stochastic

OST models of portfolio selection have differential equation. To see the meaning of
M been one-period models. I examine the such an equation, it is easiest to work out the
combined problem of optimal portfolio selec- discrete-time version and then pass to the limit
tion and consumption rules for an individual in of continuous time.
a continuous-time model whzerehis income is Define
generated by returns on assets and these re- W(t) total wealthat time t
turns or instantaneous "growth rates" are sto- Xi(t) priceof the ith asset at timet, (i 1,
chastic. P. A. Samuelson has developed a sim- . . . ,m)
ilar model in discrete-time for more general C (t) consumptionper unit time at time t
w,(t) proportionof total wealth in the ith
probability distributions in a companion paper asset at time t, (i 1,..., m)
[8].
I derive the optimality equations for a multi- Note
m
asset problem when the rate of returns are ( _ w,(t)e1)
generated by a Wiener Brownian-motionproc- j=j t
ess. A particular case examined in detail is The budget equation can be written as
the two-asset model with constant relative risk-
aversion or iso-elastic marginal utility. An W7(t)= [ 1 w ?(to) ]

explicit solution is also found for the case of

constant absolute risk-aversion. The general [W(to) - C(to)h] (1)
technique employed can be used to examine a where t to + k and the time interval between
wide class of intertemporaleconomic problems periods is h. By subtracting W(to) from both
under uncertainty.
In addition to the Samuelsonpaper [8], there sides and using X w(to) = 1, we can rewrite
i=l
is the multi-period analysis of Tobin [9]. (1) as,
Phelps [6] has a model used to determine the
W (t)-w (to)
optimal consumption rule for a multi-period X i(t )
[() - W(t (to)) (
example where income is partly generated by
an asset with an uncertain return. Mirrless [5]
has developed a continuous-time optimal con- IF V(to) - C(to)h ]- C(to)hk
sumption model of the neoclassical type with m
technical progress a random variable. ._E w(to) (e9t(tO)h_ 1)

II Dynamicsof the Model:The BudgetEquation [w (to) - C(to)]h -C(to)h (2)

In the usual continuous-time model under where
certainty, the budget equation is a differential
gi(to)h log [Xi(t)/Xi(to)],
equation. However, when uncertainty is intro-
duced by a random variable, the budget equa- the rate of return per unit time on the ith asset.
*This work was done during the tenure of a National The gi(t) are assumed to be generated by a
Defense Education Act Fellowship. Aid from the National stochastic process.
Science Foundation is gratefully acknowledged. I am in- In discrete time, I make the further assump-
debted to Paul A. Samuelson for many discussions and
his helpful suggestions. I wish to thank Stanley Fischer, tion that g,(t) is determined as follows,
Massachusetts Institute of Technology, for his comments gi(t)h- (ai - oi2/2)h + AYi (3)
on section 7 and John S. Flemming for his criticism of an
earlier version. where a1, the "expected" rate of return, is con-
[ 247 ]

This content downloaded from 194.29.185.251 on Wed, 25 Jun 2014 06:35:18 AM

All use subject to JSTOR Terms and Conditions
248 THE REVIEW OF ECONOMICS AND STATISTICS
stant; and Y (t) is generated by a Gaussian A more familiar equation would be the aver-
random-walk as expressed by the stochastic aged budget equation derived as follows: From
difference equation, (5), we have
Y (t) - Y (to) _ AY* = Z (t) v\k (4) W(t)-W(to) ] = X
E(t0)[ w*(to)a[ W(to)
where each Z (t) is an independent variate
-C(to)k] - C(to)
with a standard normal distribution for every
0
+ (k). (8)
t, o'i2 is the variance per unit time of the process
Now, take the limit as O-z
h 0, so that (8) be-
Yi, and the mean of the increment A1Y is zero.
Substituting for gi(t) from (3), we can re- comes the following expression for the defined
write (2) as, "mean rate of change of wealth":
m
W(to) limit E (to) [Wt-
W(to) = h (o
W1(t) - > W,1(to) (e)(a,-2/2(k+ A Y)
def. h+O
in
(W(to) - C(to)k)
= X qW(to) afW(to) - C(to). (8')
- C(to)h. (5) 1

Before passing in the limit to continuous

time, there are two implications of (5) which III The Two-AssetModel
will be useful later in the paper. For simplicity, I first derive the optimal
m
E(to) [1W4(t)- W(to)] = { X (to)aWV(to) equations and properties for the two-asset
model and then, in section 8, display the gen-
-C(to) } k+O(h2) eral equations and results for the m-asset case.
(6) Define
and wI (t) w(t) = proportioninvested in the
risky asset
= X w
E ( to) [(W (t)_W (to) 2w(to)w1(to). w2(t) = 1-w(t) = proportioninvestedin the
E (to) (A Yi A j). sure asset
W2(to) + 0(h2) g1(t) = g(t) = return on the risky asset
(7) (Var g. > 0)
g2(t) = r
= return on the sure asset
where E(to) is the conditional expectation
(Var g2 = 0)
operator (conditional on the knowledge of
Then, for g(t)h = (a - 0o2/2) h + AY, equa-
W(to)), and 0(o) is the usual asymptotic order
tions (5), (6), (7), and (8') can be written as,
symbol meaning "the same order as."
The limit of the process described in (4) as W (t) - W(to)
- [W(to) (e(a-02/2(h+AY)_ 1)
h O-* 0 (continuous time) can be expressed by
the formalism of the stochastic differential + (1 - w(to)) (e - 1)].
equation,' (W(to) - C(to)k - C(to)b. (9)
(4') E(to) [W (t)-W(to) ]
dlh, = a,Z, (t) \/ dt
and Y (t) is said to be generated by a Wiener = { [w(to)(a-r) + r]W(to)
process. -C(to) } k+ 0(2). (10)
By applying the same limit process to the
discrete-time budget equation, we write (5) as E(to) [(W(t) -W(to))2]
= w2(to) W2(to) E(to) [ (Ay)2]
dW = [ wi (t)aiW(t) - C(t) dt + O(h2) = W2(to)W2(to)r2h
in
+ 0(k2). (11)
+ X w,(t)0,Z,(t)W(t) \/dt. (5') diV = [(w(t) (a-r) + r)W(t) -C(t) ] dt
+ w(t)crZ(t)W(t) A/ dt. (12)
The stochastic differential equation (5') is the 0
W(t) = [w(t) (a-r) + r]W(t) -C(t). (13)
generalization of the continuous-time budget
eauation under uncertainty. The problem of choosing optimal portfolio
selection and consumption rules is formulated
'See K. Ito [4], for a rigorousdiscussionof stochastic
differentialequations. as follows,

This content downloaded from 194.29.185.251 on Wed, 25 Jun 2014 06:35:18 AM

All use subject to JSTOR Terms and Conditions
 LIFETIME PORTFOLIO SELECTION 249
Max E { fte-Pt U [C (t) ] dt + B [W(T),T] } In (17), take the E(to) operator onto each
(14) term and, noting that I[W(to),to] = E(to)
subject to: the budget constraint (12), I[W(to),to], subtract I[W(to)to] from both
C(t) _ 0; W(t) > 0; W(0) - Wo > 0 sides. Substitute from equations ( 10) and ( 11)
and where U(C) is assumed to be a strictly for E(to) [W(t) - W1(to)] and E(to) [(W(t)
concave utility function (i.e., U'(C) > 0; - W(to))2], and then divide the equation by
U"(C) < 0); where g(t) is a random variable k. Take the limit of the resultant equation as
generated by the previously described Wiener h -O 0 and ( 17) becomes a continuous-time ver-
process. B[W(T),T] is to be a specified "be- sion of the Bellman-Dreyfus fundamentalequa-
quest valuation function" (also referred to in tion of optimality, (17').
production growth models as the "scrap func-
tion," and usually assumed to be concave in 0 = Max [e-Pt U [C(t)] + at
W(T)). "E" in (14) is short for E(O), the {C(t),w(t)} at
conditional expectation operator, given W(O)
= WO as known.
+ + r)W(t) - C(t)]
aWt [ (W(t) (a-r)
To derive the optimality equations, I restate
(14) in a dynamic programming form so that + 1/2 alt2a2w2
Dw2
(t) W2(t)] (17') ]
the Bellman principle of optimality2 can be
applied. To do this, define, where I, is short for I[W(t),t] and the sub-
script on to has been dropped to reflect that
I[W(t),t] - MaxE(t) rT e-ps U[C(s)]ds (17') holds for any tE [0,T].
{C(s),w(s) } If we define b(w,C;W;t) e-Pt U(C) {
+ B[W(T),T]] (15)
where (15) is subject to the same constraints + at +f- DW [(w(t)(a-r) +r)W(t) -

as (14). Therefore, Z)2

I[W(T),T] = B[W(T),T]. (1 5') C(t)] + 1/2 cr2w2(t)W2(t) 3 then }
In general, from definition (15),
(17') can be written in the more compact form,
I[W(to),to] = MaxE(to) [ft e-ps U[C(s)]ds Max p (w,C;W,t) = 0. (17")
{C (s),w(s)} {C,w}
+ I[W(t),t] ] (16) The first-orderconditions for a regular interior
and, in particular, (14) can be rewritten as maximum to (17") are,
I(WO,O) = Max E[fo e-P U[C(s)]ds oa [w*,C*;W,:t] 0 = e-Pt U'(C) - DIt/DW-

t C(s),w(s) } (18)
+ I [W (t),t] I ( 14') and
If t = to + h and the third partial derivatives (kw[w*;C*;W;t] - 0 = (a-r) DW
of I[W(to),to] are bounded, then by Taylor's
theorem and the mean value theorem for in-
+ WWr2. (19)
tegrals, (16) can be rewritten as
I[W(to,to] = Max E(to) {
ePtU[C(t)] A set of sufficient conditions for a regular in-
{C,w} DI[W(t0),t0] terior maximum is
+ I[W(to),to] + at
4ww < 0; oce < 0; det[ ww OweJ > 0.
iOew O ea
+ aI[W(to),to] [W(t) - W(to]
Owe = ?tw = 0, and if I[W(t),t] were strictly
1 D2I[W(to),;t] concave in W, then
2 DW2 = U"(c) < 0, by the strictconcavityof U
[W(t) - W(to)]2 + O(h2) } and
(20)
where t-E [to,t]. (17)
2 The basic derivation of the optimality equations in this 30(w,C:W:t) is short for the rigorouso[w,C; 3It/3t;
section follows that of S. E. Dreyfus [2], Chapter VII. aIt/DW; alItlaW'; It: W;t].-

This content downloaded from 194.29.185.251 on Wed, 25 Jun 2014 06:35:18 AM

All use subject to JSTOR Terms and Conditions
250 THE REVIEW OF ECONOMICS AND STATISTICS
32t where a strategically-simplifying assumption
ou7 = W(t)of2 by the strict con- has been made as to the particular
< 04,
DW2 form of the
cavityof It, (21) bequest valuation function, B[W(T),T].5
and the sufficientconditions would be satisfied. To solve (17") of (*'), take as a trial solu-
Thus a candidate for an optimal solution which tion,
causes I[W(t),t] to be strictly concave will be t[W(t),t] = ()e_pt [W(t)],Y (22)
any solution of the conditions (17')-(21).
The optimality conditions can be re-written By substitution of the trial solution into (17"),
as a set of two algebraic and one partial dif- a necessary condition that It [W(t) ,t] be a
ferential equation to be solved for w* (t), C* (t), solution to (17") is found to be that b(t) must
and I[W(t),t]. satisfy the following ordinary differentialequa-
cf[w*;C*;W;t] = 0 (18") tion,
Fa [[w*C*; W;t] = 0
OW[W*,C*;W;t] = 0
(18)
(19)
b(t) = ph(t)-(1-y) [b(t)]-'Y1/-Y (23)
subject to b(T) = E1l-, and where p,u p - y
(*) subjectto the boundarycondition
[(a - r)2/2cr2 (1 -_y) + r]. The resulting de-
I[W(T),T] = B[W(T),T] and
the solutionbeing a feasiblesolu- cision rules for consumptionand portfolio selec-
tion to (14). tion, C* (t) and w* (t), are from equations
(18) and (19) of (*'), then
IV ConstantRelativeRiskAversion C*(t) = [b(t)]1'Y-' W(t) (24)
and
The system (*) of a nonlinear partial dif-
ferential equation coupled with two algebraic w*(t) = (a-r) (25)
equations is difficult to solve in general. How- 2) (25)
The solution to (23) is
ever, if the utility function is assumed to be of
b (t) = { [ 1 + (vE-1) e(t-T) ] /v}l-y (26)
the form yielding constant relative risk-aver-
sion (i.e., iso-elastic marginal utility), then (*) where v ,/( 1 - y).
can be solved explicitly. Therefore, let U(C) A sufficient condition for I[W(t),t] to be a
= C/y, yy< 1 and y70 or U(C) =log C solution to (*') is that I[W(t),t] satisfy
(the limiting form for y = 0) where- U" (C) A. I [W(t),t] be real (feasibility)
C/U'(C) = 1- 8 is Pratt's [7] measure
of relative risk aversion. Then, system (*) can B. 32t (concavity for a maxi-
a W2 mum)
be written in this particular case as C. C*(t) 0 (feasibility)
The condition that A, B, and C are satisfied in
7 the iso-elastic case is that
DIt DIt > 0, 0 - t - T (27)
+ t + @W rW [1 + (vE-l) ev(t-T)]/V
D3t DaW which is satisfied for all values of v when
(a-r) 2 [3It/fW] 2 (l 7ff) T < oo.
(*') - 2r2 32It/DW2
3 Because (27) holds, the optimal consump-
C*(t) ept (18)
= tion and portfolio selection rules are,6
'The form of the bequest valuation function (the bound-
= (a-r) Dlt/DW (19) ary condition), as is usual for partial differential equations,
w*(t) can cause major changes in the solution to (*). The par-
92W D2It/D3W2 ticular form of the function chosen in (*') is used as a
subject to I[W(T),T] = El-ye-pT proxy for the "no-bequest" condition (e = 0). A slightly
more general form which can be used without altering the
[W(T)]Y/y,forO <e << 1 resulting solution substantively is B[W(T),T] = e-PtG(T)
By the substitution of the results of (18) into (19) at [W(T)] 'Y/y for arbitrary G(T). If B is not of the iso-
(C*,w*), we have the condition w*(t) (a- r) > 0 if and elastic family, systematic effects of age will appear in the
32It optimal decision-making.
32
only if < 0. 'Although not derived explicitly here, the special case
The paper considers only interior optimal solutions. The (y = 0) of Bernoulli logarithmic utility has (29) with y = 0
problem could have been formulated in the more general as a solution, and the limiting form of (28), namely
Kuhn-Tucker form in which case the equalities of (18) and C*(t) = + p ' W(t).
(19) would be replaced with inequalities. I+ (pe-1I) ep(t-T)

This content downloaded from 194.29.185.251 on Wed, 25 Jun 2014 06:35:18 AM

All use subject to JSTOR Terms and Conditions
 LIFETIME PORTFOLIO SELECTION 251
C*(t) = [v/(1 + (vE-1) ep(t-T))] W(t), To examine some of the dynamic properties
for v 0 of C*(t), let E= 0, and define V(t) = [C*(t)/
= [1/(T -t + e)] W(t), for v = 0 W(t)], the instantaneous marginal (in this
(28) case, also average) propensity to consume out
and of wealth. Then, from (28),
w*(t) - ((I) = a constant independ-
V(t) = [V(t)]2 ev(t-T) (30)
ent of W or t. (29) and, as observed in figure 1 (for E= 0), V(t)
is an increasing function of time. In a generali-
V DynamicBehaviorand the Bequest zation of the half-life calculation of radioactive
ValuationFunction decay, define X as that tE [0,T] such that V(X-)
The purpose behind the choice of the par- = nV(O) (i.e., X is the length of time required
ticular bequest valuation function in (*') was for V(t) to grow to n times its initial size).
primarily mathematical. The economic motive Then, from (28),
is that the "true" function for no bequests is
Tlog[ eT (1 ) + ] /v; for v 0
B[W(T),T] = 0 (i.e., E= 0). From (28),
C*(t) will have a pole at t = T when E = 0.
So, to examine the dynamic behavior of C*(t) - ( ( )) T , for v =0.
(n)
and to determine whether the pole is a mathe- (31)
matical "error" or an implicit part of the eco-
To examine the dynamic behavior of W(t)
nomic requirementsof the problem, the param-
under the optimal decision rules, it only makes
eter E was introduced.
sense to discuss the expected or "averaged"
From figure 1, (C*/W)t=T -> oo asE--> 0. How-
behavior because W(t) is a function of a ran-
ever, one must not interpret this as an infinite
dom variable. To do this, we consider equation
FIGURE 1.
(13), the averaged budget equation, and evalu-
Ge/w ate it at the optimal (w*,C*) to form
0
M
'V I W( = V- V(t) (13')
w (t)
whereatr).= [ (a-
~2(1 +r], and, in sec-

Vv'E I tion VII, a* will be shown to be the expected

0~~ it
return on the optimal portfolio.
By differentiating (13') and using (30), we
get
0

rate of consumption. Because there is zero dt [ W ] -V(t) < o (32)

utility associated with positive wealth for t >
which implies that for all finite-horizonoptimal
T, the mathematics reflects this by requiring paths, the expected rate of growth of wealth is
the optimal solution to drive W(t) -> 0 as
a diminishing function of time. Therefore, if
t -> T. Because C* is a flow and W(t) is a
a* < V(O), the individual will dis-invest (i.e.,
stock and, from (28), C* is proportional to he will plan to consume more than his expected
W(t), (C*/W) must become larger and larger income, a*W(t)). If a* > V(O), he will plan
as t ->T to make W(T) = 0.7 In fact, if to increase his wealth for 0 < t < 1, and then,
W(T) -) > 0, an "impulse" of consumption dis-invest at an expected rate a* < V(t) for
would be required to make W(T) = 0. Thus, K t < T where t is defined as the solution to
<
eauation (28) is valid for E = 0.
a*
7The problem described is essentially one of exponential 1 = T + 'log v (33)
decay. If W(t) = Woe-(t), f(t) > 0, finite for all t, and v a*
WO> 0, then it will take an infinite length of time for Further, t/3a* > 0 which implies that the
W(t) = 0. However, if f(t) -* oo as t -* T, then W(t) -> 0
as t-* T. length of time for which the individual is a net

This content downloaded from 194.29.185.251 on Wed, 25 Jun 2014 06:35:18 AM

All use subject to JSTOR Terms and Conditions
252 THE REVIEW OF ECONOMICS AND STATISTICS
saver increases with increasing expected re- o= (1y)
_W__ p1(W)
turns on the portfolio. Thus, in the case a. >
V(O), we find the familiar result of "hump (a-r)2 [J'(W)]2 +rWI(W) (36)
saving." 8 2of2 J"(W)
where the functional equations for C* and w*
VI InfiniteTime Horizon have been substituted in equation (36).
Although the infinite time horizon case (T = The first-order conditions corresponding to
oo) yields essentially the same substantive re- (18) and (19) are
sults as in the finite time horizon case, it is O = U'(C) -I'(W) (37)
worth examining separately because the opti- and
mality equations are easier to solve than for O= (a-r)J'(W) + I"'Wo2 (38)
finite time. Therefore, for solving more com- and assuming that limit B[W(T),T] = 0, the
plicated problems of this type, the infinite time
horizon problem should be examined first. boundary condition becomes the transversality
The equation of optimality is, from section condition,
III, limit E[I[W(t),t]] = 0 (39)

O= Max Le-Pt U(C)+ + t or

{C,w} t limitE[e-PtJ[W(t)]] = 0
t->(*

+ it [ (w(t) (a-r) + r)W(t) - C(t)]which is a condition for convergence of the in-

tegral in (14). A solution to (14) must satisfy
t (39) plus conditions A, B, and C of section IV.
+ 1/2 2w2(t)W2(t) (17')
Conditions A, B, and C will be satisfied in the
However (17') can be greatly simplified by iso-elastic case if
eliminating its explicit time-dependence. De-
fine V* V P (a[ -
r)2 + r ]
1--y 2or2(1_y)2 1-y
J [W (t) t] ePt I [W(t) t] >0 (40)
=Max E(t) fooe-P(sYt) U[C]ds holds where (40) is the limit of condition (27)
{C,w}
in section IV, as T -> oo and V* = C*(t)/W(t)
= Max EfO e-Pv U[C]dv,
when T = oo. Condition (39) will hold if p >
0 0
{C,w}
independent of explicit time. (34) y W/W where, as defined in (13), W(t) is the
Thus, write J[W(t),t] = J[W] to reflect this stochastic time derivative of W(t) and W(t)/
independence. Substituting J[W], dividing by W(t) is the "expected" net growth of wealth
e-Pt, and dropping all t subscripts, we can re- after allowing for consumption. That (39) is
write (17') as, satisfied can be rewritten as a condition on the
O = Max [U(C)-pi + J'(W). subjective rate of time preference,p, as follows:
{C,w} for y < 0 (boundedutility), p> 0
+ r)W-C} = 0 (Bernoulli log case), p> 0
{(w(t)(a-r)
+ 1/2 1"(W)aU2w2W2]. (35) 0 < y < 1 (unboundedutility), p> y
Note: when (35) is evaluated at the optimum
[ (a-r)2(2- y) + r
(C*,w*), it becomes an ordinary differential 2L 2 ( 1) _j.
equation instead of the usual partial differential (41)
equation of (17'). For the iso-elastic case, Condition (41) is a generalization of the usual
(35) can be written as assumption required in deterministic optimal
consumption growth models when the produc-
8"Hump saving" has been widely discussed in the litera- tion function is linear: namely, that p > Max
ture. (See J. De V. Graaff [3] for such a discussion.)
Usually "hump saving" is discussed in the context of work [0, y ,8] where /3 = yield on capital.9 If a "di-
and retirement periods. Clearly, such a phenomenon can
occur without these assumptions as the example in this 'If one takes the limit as o-*2 0 (where a.2 is the
paper shows. variance of the composite portfolio) of condition (41), then

This content downloaded from 194.29.185.251 on Wed, 25 Jun 2014 06:35:18 AM

All use subject to JSTOR Terms and Conditions
 LIFETIME PORTFOLIO SELECTION 253
minishing-returns," strictly-concave "produc- generates the price changes (independent in-
tion" function for wealth were introduced, then crements assumption of the Wiener process).
a positive p would suffice. With these two assumptions, the only feed-
If condition (41) is satisfied, then condition backs of the system, the price change and the
(40) is satisfied. Therefore, if it is assumed resulting level of wealth, have zero relevance
that p satisfies (41), then the rest of the deri- for the portfolio decision and hence, it is con-
vation is the same as for the finite horizon case stant.
and the optimal decision rules are, The optimal proportion in the risky asset,10
w*, can be rewritten in terms of Pratt's relative
C,X*(t) =__
Cx*(t) { p_
-
7Y [ 22(a-r)2
2o2( risk-aversion measure, 8, as

+ ]XWt) (42) w* = (a- r) (29')

and The qualitative results that 3w*/3a > 0,
w*(t) = (a- r) (43) Dw*/Dr < 0, Dw*/Dor2 < 0, and Dw*/D8 < 0
are intuitively clear and need no discussion.
The ordinary differential equation (35), However, because the optimal portfolio selec-
J" = f (J,J'), has "extraneous" solutions other tion rule is constant, one can define the opti-
than the one that generates (42) and (43). mum composite portfolio and it will have a
However, these solutions are ruled out by the constant mean and variance. Namely,
transversality condition, (39), and conditions a. = E[w*(a+AY) + (1-w*)r] = W*a
A, B, and C of section IV. As was expected,
limit C* (t) = C.* (t) and limit w* (t) = WOO* + (1-w*)r (ar)2 + r (44)
T + oo T-+ o
(t). *2= Var [w*(a+AY) + (1-w*)r]
The main purpose of this section was to
show that the partial differentialequation (17') w*2a2 = (a-r)2 (45)
can be reduced in the case of infinite time
horizon to an ordinary differentialequation. After having determined the optimal w*, one
can now think of the original problem as having
been reduced to a simple Phelps-Ramsey prob-
VII EconomicInterpretation of the Optimal lem, in which we seek an optimal consumption
DecisionRulesfor PortfolioSelection rule given that income is generated by the
and Consumption uncertain yield of an (composite) asset.
An important result is the confirmation of Thus, the problem becomes a continuous-
the theorem proved by Samuelson [8], for the time analog of the one examined by Phelps [6]
discrete-time case, stating that, for iso-elastic in discrete time. Therefore, for consistency,
marginal utility, the portfolio-selection deci- C.*(t) should be expressible in terms of a*,
sion is independent of the consumption deci- 0J*27 8, p, and W(t) only. To show that this is,
sion. Further, for the special case of Bernoulli in fact, the result, (42) can be rewritten as,"
logarithmic utility (y = 0), the separation goes
10Note: no restriction on borrowing or going short was
both ways, i.e., the consumption decision is imposed on the problem, and therefore, w* can be greater
independent of the financial parameters and is than one or less than zero. Thus, if a < r, the risk-averter
only dependent upon the level of wealth. This will short some of the risky asset, and if a > r + a% he
is a result of two assumptions: (1) constant will borrow funds to invest in the risky asset. If one wished
to restrict w*e[O,1], then such a constraint could be intro-
relative risk-aversion (iso-elastic marginal util- duced and handled by the usual Kuhn-Tucker methods
ity) which implies that one's attitude toward with resulting inequalities.
" Because this section is concerned with the qualitative
financial risk is independent of one's wealth changes in the solution with respect to shifts in the param-
level, and (2) the stochastic process which eters, the more-simple form of the infinite-time horizon
case is examined. The essential difference between Co,*(t)
(41) becomes the condition that p > max[O,ya*] where a* and C*(t) is the explicit time dependence of C*(t) which
is the yield on the composite portfolio. Thus, the deter- was discussed in section V. For simplicity, the "oo" on
ministic case is the limiting form of (41). subscript C,,*(t) will be deleted for the rest of this section.

This content downloaded from 194.29.185.251 on Wed, 25 Jun 2014 06:35:18 AM

All use subject to JSTOR Terms and Conditions
254 THE REVIEW OF ECONOMICS AND STATISTICS

t+z* = + -
t vo1) - 2 (t)z
[C* C*2a I (52)
= Wo >
V W(t)
= (46) Therefore, by combining the effects of (51)
where V = the marginal propensity to consume and (52), one can see that individuals with low
out of wealth. relative risk-aversion (O < 8 < 1) will choose
The tools of comparative statics are used to to consume less now and save more to take ad-
examine the effect of shifts in the mean and vantage of the higher yield available (i.e., the
variance on consumptionbehavior in this mod- substitution effect dominatesthe income effect).
el. The comparison is between two economies For high risk-averters (8 > 1), the reverse is
with different investment opportunities, but true and the income effect dominates the sub-
with the individuals in both economies having stitution effect. In the borderline case of Ber-
the same utility function. noulli logarithmic utility (8 = 1), the income
If 0 is a financial parameter, then define and substitution effect just offset one another.'2
[ DC ] the partial derivative of consump-
In a similar fashion, consider the case of
0 = - T*2213 then from, (46) and (48), we de-
tion with respect to 0, IO[WO]being held fixed, rive
as the intertemporal generalization of the
effect,[ a
Hicks-Slutsky"substitution" ]U ( ( -2) ) = W2V (53)
DO U and
for static models. [DC*/DO- (DC*/DO)1O]
will be defined as the intertemporal "income" ( )) = 2 < , the substitution
or "wealth" effect. Then, from equation (22) effect. (54)
with Io held fixed, one derives by total differen- Further, DC*/D(-Go*2) =
(8-1)WO/2, and so
tiation, D DC* C* A
=-^1
0 Db(f Wo + b(O) (DW o J- 3(-(r2 ) 3( r2
* ) 10
-8-1
)

DO Do'1
(47)
= 2 WO > O, the incomeeffect.
From equations (24) and (46), b(O) = V-1,
in (47), we can
and so solving for (DWO/DO)1O (55)
write it as To compare the relative effect on consump-
tion behavior of an upward shift in the mean
awO -swo av
(48)
versus a downward shift in variance, we ex-
Do Io (8-1)V 30
amine the elasticities. Define the elasticity of
Consider the case where O= a*, then from consumption with respect to the mean as
(46), DC*
aV (8-1) Elct - /C* a*( -1)/8V (56)
_
(9) Da*
Da* 8
and similarly, the elasticity of consumption
and from (48), with respect to the variance as,
(DC* wo (50)
E2 D2aC* / C*= _2(8_ 1)/2V (57)
Da Io V
Thus, we can derive the substitution effect of For graphical simplicity, we plot el
an increase in the mean of the composite port- [VE,/a*] and e2 - -[VE2/a*] and define k
folio as follows,
( I D
[DV W ]Io
'2Many writers have independently discovered that Ber-
noulli utility is a borderline case in various comparative-
Da* Ii0 LDac Dat* i0 static situations. See, for example, Phelps [6] and Arrow
w0 [1].
[1 Because increased variance for a fixed mean usually
=- < 0. (51)
-
(always for normal variates) decreases the desirability of
investment for the risk-averter, it provides a more sym-
Because DC*/Da* = (V/Da*))Wo= [(8-1)/8] metric discussion to consider the effect of a decrease in
Wo, then the income or wealth effect is variance.

This content downloaded from 194.29.185.251 on Wed, 25 Jun 2014 06:35:18 AM

All use subject to JSTOR Terms and Conditions
 LIFETIME PORTFOLIO SELECTION 255
FIGuRE 2. higher future consumption while the high risk
ei
I - -- --- - -- ------ -
averter will always choose to increase the
amount of present consumption.

VIII Extensionto ManyAssets

The model presented in section IV, can be
-k __ __ _ _ ____
l
extended to the m-asset case with little diffi-
culty. For simplicity, the solution is derived in
the infinite time horizon case, but the result is
-k ~~ ~ ~ ~ 1Y similar for finite time. Assume the Mth asset to
be the only certain asset with an instantaneous
rate of return am = r.'4 Using the general
'*2/2a*. e, and e2 are equal at 8 1, l/k. The equations derived in section II, and substituting
n
particular case drawn is for k < 1. for wm(t) =1- : wi (t) where n-m - 1,
For relatively high variance (k > 1), the j=j

high risk averter (8 > 1) will always increase equations (6) and (7) can be written as,
present consumption more with a decrease in E(to) [W(t) -W(to)]
variance than for the same percentage increase = [w' (to) (a-r) + r] W(to) k
in mean. Because a high risk-averter prefers -C(to)k + 0(h2) (6)
a steadier flow of consumption at a lower level and
than a more erratic flow at a higher level, it E(to) [(W(t) -W(to) )2
makes sense that a decrease in variance would = w'(to) a w(to) W2(to)h
have a greater effect than an increase in mean. + 0(h2) (7)
On the other hand, for relatively low variance where
(k < 1), a low risk averter (0 < 8 < 1) will w'(to) [wi(to), .... ,wn(to)], a n-vector
always decrease his present consumption more at - a,, ... 1 a<]

with an increase in the mean than for the same = [r,...,r] a n-vector
percentage decrease in variance because such Q [fj[], the n X n variance-covariance
an individual (although a risk-averter) will matrixof the risky assets
prefer to accept a more erratic flow of con- a is symmetricand positivedefinite.
sumption in return for a higher level of con- Then, the general form of (35) for m-assets is,
sumption. Of course, these qualitative results in matrix notation,
will vary depending upon the size of k. If the O=Max [U(C)-p J(W)
riskiness of the returns is very small (i.e., {C,w}
k < < 1), then the high risk-averter will in- + J'(W) { [w'(a-r) + r] W - C}
crease his present consumption more with an
upward shift in mean. Similarly, if the risk- + 2 J"(W) weQ wW2] (58)
level is very high (i.e., k > > 1) the low risk and instead of two, there will be m first-order
averter will change his consumptionmore with conditions correspondingto a maximization of
decreases in variance. (35) with respect to w, .. . , wn, and C. The
The results of this analysis can be summed optimal decision rules corresponding to (42)
up as follows: Because all individuals in this and (43) in the two-asset case, are
model are risk-averters, when risk is a domi- A A

nant factor (k > > 1), a decrease in risk will

= ty) (a-r)'&Y(a-r)
=
C.O*(t)+~*(t {
have the larger effect on their consumption
{
P__ ] }
-
2( 2

decisions. When risk is unimportant (i.e.,

k < < 1), they all react stronger to an increase + ()(59)
in the mean yield. For all degrees of relative 14 Clearly, if there were more than one certain asset, the
riskiness, the low risk-averterwill give up some one with the highest rate of return would dominate the
present consumption to attain an expected others.

This content downloaded from 194.29.185.251 on Wed, 25 Jun 2014 06:35:18 AM

All use subject to JSTOR Terms and Conditions
256 THE REVIEW OF E-CONOMICSAND STATISTICS
and their counterpartsfor the constant relative risk-
wX,*(t) =
1
Q- (a-r)
A
(60) aversion case, (42) and (43), one finds that
(l1-y) consumptionis no longer a constant proportion
where w*'(t) - [wl* (t), . . ., w,* (t) of wealth (i.e., marginal propensity to consume
does not equal the average propensity) al-
IX Constant Absolute Risk Aversion though it is still linear in wealth. Instead of the
proportionof wealth invested in the risky asset
System (*) of section III, can be solved ex- being constant (i.e., w* (t) a constant), the
plicitly for a second special class of utility
total dollar value of wealth invested in the
functions of the form yielding constant absolute risky asset is kept constant (i.e., w*(t)W(t)
risk-aversion. Let U(C) = -e-Ocfq, qj> 0, a constant). As one becomes wealthier, the
where -U"(C)/U'(C) = -q is Pratt's [17]
proportion of his wealth invested in the risky
measure of absolute risk-aversion. For con- asset
falls, and asymptotically, as W -> oo, one
venience, I return to the two-asset case and in- invests all his wealth in the certain asset and
finite-time horizon form of system (*) which
consumes all his (certain) income. Although
can be written in this case as, one can do the same type of comparativestatics
0=-J'(W) for this utility function as was done in section
O- ( ) -pJ (W) + J'(W)rW
VII for the case of constant relative risk-aver-
+ I (W) log [J'(W)] sion, it will not be done in this paper for the
sake of brevity and because I find this special
(a-r) 2 [J'(W)]2 (17") form of the utility function behaviorially less
(*"f) 22o2 J"(W)
plausible than constant relative risk aversion.
C*(t)= --log [J'(W)] (18) It is interesting to note that the substitution
w*(t) = -J'(W) (a-r)/r,2 W J"(W) effect in this case, [ a I ; is zero except
subject to limit E [ e-PtJ (W (t) 0
t-oo when r = 0.
(19)
where J(W) - ePt [W(t) ,t] as defined in sec- X OtherExtensionsof the Model
tion VI. The requirements for the general class of
To solve (17") of (*"), take as a trial solu- probability distributionswhich could be accept-
tion, able in this model are,
-p (1) the stochastic process must be Markov-
J(W)= e-qW. (61)
q ian.
By substitution of the trial solution into (17"), (2) the first two moments of the distribution
a necessary condition that J(W) be a solution must be O(At) and the higher-order
to (17") is found to be that p and q must satisfy moments o (At) where o (.) is the order
the following two algebraic equations: symbol meaning "smaller order than."
q = (62) So, for example, the simple Wiener process pos-
and tulated in this model could be generalized to
(r - p - (a-r)2/202 include a, = a, (X1, . .. , Xm,W,t) and =
\~ ~ r / (X1, ... , Xm,W,t), where Xi is the price of the
p=re (63) ith asset. In this case, there will be (m + 1)
The resulting optimal decision rules for port- state variables and (17') will be generated
folio selection and consumption are, from the general Taylor series expansion of
Xm,W,t] for many variables. A
C*(t) = rW(t) + [p- r + (a-r)2/2u2 ] I[X212 ... ,
particular example would be if the ith asset is a
(64) bond which fluctuates in price for t < ti, but
and will be called at a fixed price at time t = ti.
w*() =
(a-r')
(65)
=
Then ata,(Xi,t) andc- = ai(Xi,t) > 0 when
rU2W(t) t < ti and crq = 0 for t > ti.
Comparing equations (64) and (65) with A more general production function of a neo-

This content downloaded from 194.29.185.251 on Wed, 25 Jun 2014 06:35:18 AM

All use subject to JSTOR Terms and Conditions
 LIFETIME PORTFOLIO SELECTION 2 57
classical type could be introduced to replace [2] Dreyfus, S. E., Dynamic Programming and the
the simple linear one of this model. Mirrlees Calculus of Variations (New York: Academic
[5] has examined this case in the context of a Press, 1965).
[3] Graaff, J. De V., "Mr. Harrod on Hump Saving,"
growth model with Harrod-neutral technical Economica (Feb. 1950), 81-90.
progress a random variable. His equations [4] Ito, K., "On Stochastic Differential Equations,"
(19) and (20) correspond to my equations Memoirs, American Mathematical Society, No. 4
(35) and (37) with the obvious proper substi- (1965), 1-51.
tutions for variables. [5] Mirrlees, J. A., "Optimum Accumulation Under
Thus, the technique employed for this model Uncertainty," Dec. 1965, unpublished
can be extended to a wide class of economic [6] Phelps, E. S., "The Accumulation of Risky Capi-
models. However, because the optimality equa- tal: A Sequential Utility Analysis," Econometrica,
30 (1962), 729-743.
tions involve a partial differential equation,
[7] Pratt, J., "Risk Aversion in the Small and in the
computational solution of even a slightly gen- Large," Econometrica, 32 (Jan. 1964), 122-136.
eralized model may be quite difficult. [8] Samuelson, P. A., "Lifetime Portfolio Selection by
Dynamic Stochastic Programming," this REVIEW, L
(Aug. 1969).
REFERENCES
[9] Tobin, J., "The Theory of Portfolio Selection,"
[1] Arrow, K. J., "Aspects of the Theory of Risk- The Theory of Interest Rates, F. H. Hahn and
Bearing," Helsinki, Finland, Yrjo Jahnssonin Saa- F. P. R. Brechling, (ed.) (London: MacMillan
tio, 1965. Co., 1965).

This content downloaded from 194.29.185.251 on Wed, 25 Jun 2014 06:35:18 AM

All use subject to JSTOR Terms and Conditions