I MPERIAL C OLLEGE L ONDON

D EPARTMENT OF M ATHEMATICS

Utility maximization

Author:
Hafsae Tabti (CID: 01807226)

A thesis submitted for the degree of

MSc in Mathematics and Finance, 2019-2020

 Abstract

The aim of this thesis is to study portfolio optimisation problems combining

a risk free asset with risky assets. A single period model with finite probability
space is first considered. The convex duality approach is used to find the optimal
trading strategy that maximises the final wealth.

A time-continuous model is then assumed, where the stock prices are driven
by stochastic differential equations. The Dynamic Programming Principle and
some of the related results are stated and used to solve the classical Merton’s
problem where we seek to maximize the expected utility of consumption under
an infinite time horizon assuming a deterministic interest rate and deterministic
stock volatilities.

These assumptions are then relaxed. A Hobson and Rogers stochastic volatil-
ity model is considered, where the volatility is expressed as an exponentially-
weighted mean of historic log-prices. The HJB equation is used to derive a non
linear ODE of the value function which can be linearised by a change of variable.
Then assuming a finite time horizon T , Feynman-Kac theorem provides a solu-
tion of the ODE, and by taking the limit T → ∞ and considering a transversality
condition, the solution of the ODE can be written as the expectation of an Ito
process. Finally, a numerical example is provided to illustrate the results, where
the expectation is approximated by Monte Carlo simulations and the optimal
consumption and investment strategies and the effects of varying the parame-
ters on these optimums are analysed.

Finally, a consumption-investment problem is considered under a non deter-

ministic interest rate driven by a Vasicek model. The ODE is derived using the
HJB equation, which is again non linear and can’t be solved analytically. The
subsolution-supersolution method developed by Fleming and Pang is used to
find an upper and lower bound of the optimal solution.

Keywords: Optimal consumption-investment, HJB equation, Dynamic Pro-

gramming Principle, Stochastic volatility, Vasicek Interest Rate Model.

2
 Acknowledgements

” Gratitude is the memory of the heart. ”

Hans Christian Andersen

I would like to thank my supervisor Pietro Siorpaes for guiding me and pro-
viding me with the tools that I needed to successfully complete my thesis.

I’m indebted to my parents and my two sisters Fatine and Oumaima, who
unconditionally loved, trusted, supported and believed in me and in my choices.

I would like to thank my dear friends Imtiaz Rafiq, Altynay Myrzabayeva and
Kamil Kakar for the wonderful times we shared, especially during the difficult
period of lockdown.

My sincere gratitude goes to Mohamed Taik for his help and precious advice
through my whole journey in Imperial College.

3
CONTENTS CONTENTS

Contents
1 Introduction 8
1.1 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 A convex optimisation method 12

2.1 Model specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Optimisation problem . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Risk neutral approach . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.1 Complete models . . . . . . . . . . . . . . . . . . . . . . . . . 14

3 A DPP approach 16
3.1 Model specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Structure of optimal control problems . . . . . . . . . . . . . . . . . 18
3.3 Dynamic Programming Principle (DPP) . . . . . . . . . . . . . . . . . 19
3.4 Hamilton Jacobi Bellman Equation (HJB Equation) . . . . . . . . . . 20
3.5 Verification theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4 Merton’s problem 22
4.1 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2 Time-homogeneous property: . . . . . . . . . . . . . . . . . . . . . . 23
4.3 Scaling property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.4 HJB equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.5 Optimal control parameters . . . . . . . . . . . . . . . . . . . . . . . 25
4.6 Value function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.7 Analysis of optimal values . . . . . . . . . . . . . . . . . . . . . . . . 27

5 Stochastic Volatility 28
5.1 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.2 HJB equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.3 Optimal control parameters . . . . . . . . . . . . . . . . . . . . . . . 32
5.4 Solving the ODE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.5 Numerical solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

6 Stochastic interest rate 41

6.1 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.2 HJB equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.3 Optimal control parameters . . . . . . . . . . . . . . . . . . . . . . . 44
6.4 Solving the ODE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.5 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

7 Conclusion 49

4
LIST OF FIGURES LIST OF FIGURES

List of Figures
1 The value function of Merton’s problem as a function of time t and
wealth ωt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1 + y2
2 Function f : y 7→ σ . . . . . . . . . . . . . . . . . . . . . . . . . 37
2 + y2
3 Ns Path simulations of Ŷt with Y0 = 5 and Ns = 1000 . . . . . . . . . 37
4 Ns simulations of ĥ(5) with Ns = 1000 . . . . . . . . . . . . . . . . . 37
5 The numerical solution of the function y 7→ ĝ(y) . . . . . . . . . . . . 38
6 The numerical derivative of the function y 7→ ĝ 0 (y) . . . . . . . . . . 38
7 The optimal consumption proportion of wealth c∗ as a function of y . 38
8 The optimal investment proportion of wealth π∗ as a function of y . . 38
9 Effect of varying R on the optimal consumption proportion of wealth
c∗ evaluated at y = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
10 Effect of varying R on the optimal investment proportion of wealth π∗
evaluated at y = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
11 Effect of varying σ on the optimal consumption proportion of wealth
c∗ evaluated at y = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
12 Effect of varying σ on the optimal investment proportion of wealth π∗
evaluated at y = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
13 Effect of varying λ on the optimal consumption proportion of wealth
c∗ evaluated at y = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
14 Effect of varying λ on the optimal investment proportion of wealth π∗
evaluated at y = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
15 Supersolution f¯(r) and subsolution f (r) of the function f . . . . . . . 48
16 Supersolution q̄∗ (r) and subsolution q∗ (r) of the consumption propor-
tion q∗ (r) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
17 Effect of varying the risk aversion rate R on the supersolution f¯(r) of
the function f . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
18 Effect of varying the risk aversion rate R on the subsolution q∗ (r) of
the consumption proportion q∗ (r) . . . . . . . . . . . . . . . . . . . . 48

5
LIST OF TABLES LIST OF TABLES

List of Tables
1 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Coefficients of PDE of h . . . . . . . . . . . . . . . . . . . . . . . . . 34
3 Parameter values of the numerical solution of SV model . . . . . . . 36
4 Parameter values of the numerical solution of stochastic interest rate
model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

6
LIST OF TABLES LIST OF TABLES

List of Symbols
The notations below will be used in this thesis:

a.s Almost surely

> Transpose

−1 Inverse

tr Trace

∇ Gradient

D2 Hessian matrix

E Expected values

Ω Probability space

P Probability measure

Q Risk-neutral probability measure

R Real numbers

Rn Real vectors of dimension n

Mn,d (R) Real matrices of dimensions n ∗ d

u.v Scalar product of u and v

Vx First order partial derivative of V

Vxy Second order partial derivative of V

Table 1: Notations

7
 1 INTRODUCTION

1 Introduction

1.1 Literature review

” Don’t gamble; take all your savings and buy some good stock and hold it till it goes
up, then sell it. If it don’t go up, don’t buy it. ” - Will Rogers

The objective of every investor is to gain the highest possible return and take
the lowest risk. He can build a portfolio by considering two types of assets: a risk-
free asset that returns a fixed low rate and consequently doesn’t imply any risk and a
risky asset that gives a preeminent expected return against a higher risk. Depending
on his preferences, the investor will either invest all of his wealth in one of these
assets or allocate it between the two assets and have a trade-off between high return
and riskiness. But how can he assess the performance of a given strategy?

The first approach that has been used is the expected value of the payoff, un-
til Nicholas Bernoulli introduced the ”St. Petersburg Paradox” in 1713. Suppose a
gambler can enter the following game: a fair toss coin is tossed until a head appears
and the gambler gets 2n where n is the number of times the coin was flipped. The
outcome of this game is Y = 2n with probability P (T T T T · · · H) = 2−n .
| {z }
n−1 times
The question now is: how much this game is worth ? According to the expected
value approach, the gambler can pay :
∞
X 1
E(Y ) = ∗ 2n = ∞
2n
n=1

This is absolutely not a reasonable price of the game. In 1738, Nicholas’s cousin
Daniel Bernoulli came up with a solution that revolutionized the world of finance.
He suggested to alter the nominal amount and replace it by the utility of this amount,
and proposed to take a logarithmic utility function U (x) = log(x):
∞
X 1
E(U (Y )) = ∗ U (2n ) < ∞
2n
n=1

Arrow [2] suggested to take a bounded utility function and De Buffon [10]
argued that some sufficiently improbable outcomes are “morally impossible” and
should be ignored.
Bernoulli’s idea gave birth to the Marginalist Revolution in 1817, and a growth
in interest about utility maximization raised since then.

Some researchers were interested in the single period model. In this setting,
Markowitz [28] and Tobin [52] showed the ”Efficient Set Theorem” stating that if the
returns are normally distributed and the utility function is concave, then the problem

8
1 INTRODUCTION 1.1 Literature review

of maximising the expected utility is equivalent to minimizing the standard devia-

tion of the returns for every concave utility function and maximizing the expected
returns for any standard deviation of the returns. If the returns are also indepen-
dent, the theorem holds for the quadratic utility function.

Fama [22] showed an extension of this theorem for a wider class of distribu-
tions: symmetric stable Paretian returns [26]. This class of stable distributions was
first introduced by Lévy [38].

Elton and Guber [20] proved a necessary condition for expected utility maxi-
mization under log-normal distribution: the optimal portfolio should lie on an exte-
rior boundary in mean variance space for all utility function U .
They have also found [19] that maximizing the expected utility is equivalent to max-
imising the geometric mean in the particular case of logarithmic utility.

The statistical properties of the efficient frontier have been investigated by Job-
son [35], Broadie [9], Chopra and Ziemba [15], Best and Ding [5] and MacLean and
Weldon [39]. They studied distribution tests and confidence sets to assess the as-
sumption of independence and normal distribution of log-returns.

The continuous time modeling has seen a grown interest after the introduction
of stochastic calculus. Merton has formulated and solved the optimal consumption-
investment model [42] under log-normally distributed returns with deterministic
volatility and interest rate and a Constant Relative Risk Aversion (CRRA) utility func-
tion using the Dynamic Programming approach. It was developed by Bellman [3] in
1975 and had been used in Optimal Control theory.

However, the assumptions made in Merton’s problem aren’t realistic. For in-
stance, interest rates aren’t deterministic in real life and should be modeled by a
stochastic differential equation. The main interest rate models that have been intro-
duced in the literature are Vasicek [54], CIR [16] and Hull–White [32] models.

Some papers studied utility maximisation under stochastic interest rates. In

the finite time horizon case, Chang and Chang [13] derived a closed-form solution
in a complete market under a Hyperbolic Absolute Risk Aversion (HARA) utility
function using the Legendre transform. Dong [34] derived the optimal solution of
the consumption-investment problem under CRRA utility function for a market that
consists of a risk-free asset where the interest rate is driven by a Vasicek model, a
zero-coupon bond and a risky stock that follows a log-normal model. Korn and Kraft
[13] and Grasselli [14] used the CIR interest rate dynamics to derive the optimal
solution under HARA utility.

For the infinite horizon time case, Fleming and Pang [24] developed a subsolution-
supersolution method and provided upper and lower bounds for the value function
under a CRRA utility function. The drawback of this method is that some of the

9
1.1 Literature review 1 INTRODUCTION

supersolutions/subsolutions are not explicitly derived but they are rather given in a
parametric form. Trybula [53] used Fleming and Pang’s approach to derive explicit
bounds in the case of a consumption problem without investment. We use the same
approach as Fleming and Pang to derive a numerical approximation of the optimal
solution under a Vasicek interest rate model.

The volatility is a very important parameter in pricing and hedging financial

securities. It has been proved that it is not constant (see [6] and [50]) and models
based on deterministic volatility don’t match the implied market distributions. Sev-
eral models have been suggested to model the randomness that drives the volatility
processes.

Cox and Ross [17], Geske [27]), Rubinstein [49] and Bensoussan et al [4] sug-
gested to model it by a diffusion process that is correlated with stock prices and/or
with a firm’s debt. Johnson and Shanno [36], Scott [50], Hull and White ([33] and
[31]) and Wiggins [55] described the randomness of the volatility by a Brownian
motion independent of the price process.

In 1982, Engle [21] introduced ARCH processes and in 1986, Bollerslev [7]
introduced GARCH processes to model asset volatilities. Theses models capture the
stylized facts that log-returns volatility is clustered and highly persistent. These
proporties are also captured by Hobson and Rogers [29] model. They introduced
in 1998 an endogenous model of volatility that is expressed as an exponentially-
weighted average of the moments of historic log-prices. This model had the advan-
tage of being observable and also consistent with the ‘smiles’ and ‘skews’ implied by
the market.

Kraft [37] addressed the optimization problem under a Heston stochastic volatil-
ity model and the power utility function. Chacko and Viceira [12] investigated
the case of a dynamic hedging portfolio and derived closed-form formulas of op-
timal consumption and investment strategies for an infinite horizon. Fleming and
Hernandez-Hernandez [23] and Fouque, Papanicolaou and Sircar [25] were also in-
terested in the optimization problem with stochastic volatility.

In this paper, we model the volatitlity process by Hobson and Rogers [29]
model, and write the solution of the HJB equation as an expectation using Feynman-
Kac theorem and by imposing a transversality condition ([1] and [18], [42]), then
this expectation is computed using Monte Carlo simulations. This is a powerful
method that provides high level results. This idea has been suggested by Ravi in his
paper [40].

Other researchers combined the stochastic volatility model together with stochas-
tic interest rates. For example, Chang and Rong [14] used a CIR interest rate model
and Heston’s stochastic volatility model to derive a closed-form solution of the ex-
pected utility under a finite time horizon for both power utility and logarithmic

10
1 INTRODUCTION 1.1 Literature review

utility.

Noh and Kim [44] addressed the case of a Vasicek asset price volatility model
and a stochastic interest model such that the mean return µ(rt ) and volatility σ (rt )
satisfy some conditions using HARA utility and logarithmic utility. They character-
ized the optimum by a supersolution and subsolution.

Finally, there is a plenty of variants of Merton’s problem that take into ac-
count transaction costs, taxes, random stopping time, etc. These problems have
been briefly discussed by Rogers [48], and some researchers were interested in these
problems and tackled them in a more detailed manner. We refer to the recent pa-
per of Hobson, Tse and Zhu’s [30] for the multi-asset investment and consumption
problem with transaction costs.

11
 2 A CONVEX OPTIMISATION METHOD

2 A convex optimisation method

2.1 Model specifications
In this first chapter, we consider a single period model with initial time t = 0 and
final time t = 1, and finite probability space Ω = {ω1 , ..., ωK }. P is the historical prob-
ability measure and Q is the risk-neutral measure.

Let B = {B0 , B1 } denotes the bank account process, such that B0 = 1, and
B1 = 1 + r where r is the interest rate.

We consider N securities, and denote the price process of the i-th security by
Si = (Si (0), Si (1)). The time t = 0 price Si (0) is a deterministic positive number and
Si (1) is a non-negative random variable representing the time t = 1 price.

Si (t)
We also consider the discounted price process Si∗ (t) := for t ∈ {0, 1}.
Bt
We denote by the vector H = (H1 , ..., HN ) the trading strategy, where Hi is the
number of units held of the i-th security, and H the set of all trading strategies.

Finally, we denote the P wealth process of the portfolio by V = (V0 , V1 ), its value
at time t ∈ {0, 1} is Vt = Bt + N
n=1 Hn Sn (t).

2.2 Optimisation problem

The problem consists of finding the optimal strategy, hence the need of quantifying
the performance of a given strategy H. One of the performance measures used is the
expected utility.

Let u : R × Ω → R be a function such that w → u(w, Ω) is differentiable, con-

cave, and strictly increasing for each ω ∈ Ω. This function represents the utility of
the wealth ω.

The optimisation problem is formulated as:

maximize Eu (V1 (H1 , · · · , HN ))

H∈H
subject to V0 = v

Why this is a convex optimisation problem?

The function u is assumed to be a concave function and the expectation is

linear and increasing, so:

12
2 A CONVEX OPTIMISATION METHOD 2.3 Risk neutral approach

u(λV + (1 − λ)W ) ≥ λu(V ) + (1 − λ)u(W ) ( By concavity of u )

Eu(λV + (1 − λ)W ) ≥ E(λu(V ) + (1 − λ)u(W )) ( Expectation is increasing )
Eu(λV + (1 − λ)W ) ≥ E(λu(V )) + (1 − λ)E(u(W )) ( By linearity of expectation )

Thus the objective function is concave and the set of constraints is clearly con-
vex.

The objective function can be written as:

X
Eu (V1 ) = P(ω)u (V1 (ω), ω)
ω∈Ω
X
P(ω)u (B1 v + H1 ∆S1∗ + · · · + HN ∆SN
∗

= , ω)
ω∈Ω

We compute the gradient as follows:

 
 X 
P(ω)u 0 (B1 (ω) v + H1 ∆S1∗ (ω) + · · · + HN ∆SN∗ (ω) , ω) B1 (ω)∆Sn∗ (ω)

∇(V1 ) = 
ω∈Ω 1≤n≤N
0
= (E [B1 u (V1 ) ∆Sn∗ ])1≤n≤N

So that the first order condition can be written as:

X
P(ω)u 0 (B1 (ω) v + H1 ∆S1∗ (ω) + · · · + HN ∆SN
∗
(ω) , ω) B1 (ω)∆Sn∗ (ω) = 0,

n = 1, . . . , N
ω∈Ω

This system of N equations is not always easy to solve, this is why we will in-
troduce the risk neutral approach.

2.3 Risk neutral approach

First, let’s get some intuition about this approach.

Let’s define Qu (ω) := P(ω)u 0 (V1 (ω), ω) B1 (ω). If Qu is proportional to the risk-
neutral probability measure Q, then the first order condition is equivalent to:

EQ (∆Sn∗ ) = 0, n = 1, . . . , N (1)

By the property of the risk neutral probability measure, Sn∗ is a Q-martingale,

i.e EQ (∆Sn∗ ) = 0, for n = 1, . . . , N , which is the same as the system (1). Hence the link
between the risk neutral measure and our optimal investment problem. We will look
at more details in the following sections.

13
2.3 Risk neutral approach 2 A CONVEX OPTIMISATION METHOD

Recall the objective function of the optimisation problem:

(H1 , · · · , HN ) 7→ Eu (V1 (H1 , · · · , HN )) ,

which is the decomposition of the two functions:

f : Wv → R and V1 : H → W
Pv
V1 7→ Eu(V1 ) (H1 , · · · , HN ) 7→ B1 (v + Hn ∆Sn∗ )

Where Wv is the set of attainable wealths.

The function V1 maps the trading strategies into the real random variables
representing the final wealth, and the function f maps these wealths to the real ex-
pected utility which we aim to maximise.

The approach is the following [46]:

1. Find the optimal attainable wealth V1∗ which maximises the objective function.

2. Replicate the wealth V1∗ and find the corresponding hedging vector H = (H1 , · · · , HN ).

2.3.1 Complete models

Recall that a model is complete if every payoff V1 is replicable, and that there exists
a unique risk-neutral probability measure, under which the time 0 fair value of the
claim V1 is V0 = EQ (V1 ).

Thus, under the assumption of complete market, a wealth V1 is attainable iff

the discounted final wealth verifies v = EQ (V1∗ ).

The set of attainable wealths can be then written as:

( ! )
Q V
Wv = V r.v : E =v
B1
.

Step 1:

According to the above approach, one needs to solve the following optimisa-
tion problem:

maximize Eu (V )
subject to EQ (V /B1 ) = v

To do so, we introduce the Lagrange multiplier:

14
2 A CONVEX OPTIMISATION METHOD 2.3 Risk neutral approach

!
V Q
L(V , λ) = (V ) − λ E
B1
Q
and the density L = , so that for any random variable Y , one has:
P

EQ (Y ) = EP (LY )

and
" #
LV
L(V , λ) = E u(V ) − λ
B1
" #
X L(ω)V (ω)
= P(ω) u(V (ω)) − λ
B1 (ω)
ω∈Ω

So now we are interested in the following problem:

maximize L(V , λ)
subject to EQ (V /B1 ) = v

The necessary conditions can be written as:

L(ω)
u 0 (V ∗ (ω)) = λ , for all ω ∈ Ω
B1 (ω)
i.e !
L(ω)
V ∗ (ω) = u 0−1 λ , for all ω ∈ Ω (2)
B1 (ω)
where u 0−1 denotes the inverse function of u 0 .

Solving the equation 2 would allow to have the expression of the optimal
wealth as a function of λ, then λ is determined such that the constraint is not vi-
olated: " !#
Q 1 0−1 L
E u λ =v
B1 B1

Step 2:

In this step, we find the trading strategy H = (H1 , · · · , HN ) to replicate the

wealth V of equation 2. Thus, we need to solve a system of K equations:
 N

 X 
B1 (ωi ) v + Hn ∆Sn∗ (ωi ) = V (ωi ) i ∈ {1, · · · , K}
n=1

15
 3 A DPP APPROACH

3 A DPP approach
In this chapter, we will consider an other approach to solve the control problems
based on the Dynamic Programming Principle (DPP) that we will introduce later.

3.1 Model specifications

We consider in this chapter a continuous-time setting.

Let (Ω, F , {Ft }t≥0 , P) a filtered probability space and W = (W 1 , · · · , W d ) a d-

dimensional Brownian motion.

We consider a financial market with two types of assets: risk-free asset repre-
senting the bank account, and a vector of risky assets St = (St1 , · · · , Stn ) which are the
solutions of the following SDEs:

dBt = rt Bt dt, dSt = µ(t, St ) dt + σ (t, St ) dWt , (3)

where µ(t, St ) ∈ Rn and σ (t, St ) ∈ Mn∗d (R).

The components of the second equation are:

d
j
X
dSti = µi (t, St ) dt + σij (t, St ) dWt , 1≤i≤n
j=1

In the next sections, we consider the following drifts µi (t, St ) and volatilities
σij (t, St ) functions:
µi (t, St ) = µi Sti , σij (t, St ) = σij Sti ,
where µi and σij are constants.

We denote by Xt and Yt the money invested in the bank account and risky
assets respectively, and ωt the total wealth. We have then:

ωt = Xt + Yt ,

Yt = nt . St ,

where n is a n-dimensional previsible process, its component nit represents the num-
ber of shares of the i-th asset hold in the portfolio at time t and (u.v) denotes the
scalar product of u and v.

Assuming that the stocks S deliver a n-dimensional adapted vector of dividends

δ , and that the agent makes an endowment of a process e and consumes at the rate
process c, we can write the dynamics of X and Y as:

16
3 A DPP APPROACH 3.1 Model specifications

dXt = rt Xt dt + et dt − ct dt,

dYt = nt . ( dSt + δt dt),

so that the final wealth has the dynamics:
dωt = dXt + dYt

= rt Xt dt + et dt − ct dt + nt . ( dSt + δt dt)

= rt (ωt − nt .St ) dt + et dt − ct dt + nt . ( dSt + δt dt)

The processes e, δ, S and r are assumed given. We will take e = 0 and δ = 0, i.e
there is not an endowment of money or dividends:

dωt = rt ωt dt + θt . (σ dWt + (µ − rt ) dt) − ct dt (4)

where θt := nt .St is the vector of each asset worth.

The equation (4) that describes the dynamics of the wealth process can be
written as:
dωt = f (ω(t), ν(t)) (5)
given that: ω(t0 ) = ω0 and such that :
ν(t) = (c(t), θ(t))
Theorem 1 (Carathèodory)
Suppose that:
1. f (·, ·) is continuous,
2. f (·, ν) satisfies the Lipschitz condition, i.e. there exists a constant Lf > 0 such
that:
|f (z, ν) − f (z0 , ν)| ≤ Lf |z − z0 |
for all z, z0 ∈ Rd and ν ∈ A,
3. f (z, ν(t)) is measurable with respect to t.

Then, there is a unique absolutely continuous function y : [t0 , T ] → Rd that satisfies

Zs
y(s) = z0 + f (y(τ), ν(τ))dτ
t0

Proof 1
The proof can be found in [41].

17
3.2 Structure of optimal control problems 3 A DPP APPROACH

The theorem 1 guarantees, under some conditions, the existence and unique-
ness of the solution of equation (5) and that the solution has the same properties as
the function y.

Finally, the agent has to control the processes c and θ in a way such that her
wealth ω is always non-negative. Thus, the couple (θ, c) is constrained to live in the
set of admissible controls A := { (θ, c) | ω ≥ 0 a.s. }.

3.2 Structure of optimal control problems

In an optimal control problem, we aim at finding the control processes (c and n in
section (3.1) for example). To do so, one needs to pay particular attention to the
four following elements [51]:

• State process Z (.): This process describes the state and hypotheses of the
problem. In the setting of the previous section, the wealth of the investor is
the state process and it is characterized by equation (4):

Z (.) := ω(.)

Any other information about the parameters of the problem should be taken
into account in the state process.

• Control process ν(.): This process is the one that the investor can control in
order to optimize his objective. It is represented by the processes (θ, c) in the
previous section:
" #
θ(.)
ν(.) :=
c(.)

• Set of admissible controls A : The control process ν(.) has to respect certain
constraints in order that the strategy can be admissible. We call the set of such
control processes the admissible controls:

A := { (θ, c) | ω≥0 a.s. }

• Objective function: The aim of the agent is to achieve a certain optimal cri-
terion. To assess the performance of a given control process, one needs to set
an optimisation measure. In the majority of optimal control problems, this
function is taken as the expectation of a certain utility function of the control
process and the investor seeks to maximise that quantity:
"Z T #
J(ν(·)) = E u(t, ν(t))dt + G(T , ωT ) | Z0 = z
0

18
3 A DPP APPROACH 3.3 Dynamic Programming Principle (DPP)

where T is the time horizon, it can be finite or infinite. The function u will
mostly be taken as:
u(t, x) = e−ρt U (x)
x1−R
where U is the Constant-Relative-Risk-Aversion (CRRA) function U (x) :=
1−R
s.t R > 0 and R , 1 .

The discount factor e−ρt guarantees the convergence of the integral.

We can notice that the problem suggests a feedback process: the control changes
if there is a change in the state process. Thus, we will use the Dynamic Programming
Principle (DPP) since it leads to solutions in a feedback format [41].

The aim of the agent is to attain the supremum of the above functional. The
value function is then defined as:

V (z) := sup J(ν(·))

ν∈A

Some control problems aim to minimise a cost function:

V (z) := inf J(ν(·))

ν∈A

Then instead of solving the problem only at the initial time t = 0, we will be
interested in finding the optimal strategy at every time t, i.e:
"Z T 1−R
#
−ρs cs
V (t, z) = sup E e ds | Z(t) = z
ν∈Az t 1−R
where Az denotes the set of admissible control processes of a state process Z such
that Zt = z.

3.3 Dynamic Programming Principle (DPP)

The Dynamic Programming Principle (DPP) is a sine qua non theorem in tackling
optimal investment problems.

Theorem 2 (Dynamic Programming Principle)

For any stopping time τ ≥ t
"Z τ  
#
ν
V (t, z) = sup E u(s, ν(s))ds + V τ, Zt,z (τ) | Ft
ν∈At,z t

ν
where Zt,z (τ) is the value of the state process starting at time t with value z and
controlled by the process ν evaluated at time τ.

19
3.4 Hamilton Jacobi Bellman Equation (HJB Equation) 3 A DPP APPROACH

3.4 Hamilton Jacobi Bellman Equation (HJB Equation)

Theorem 3 (The Davis-Varaiya Martingale Principle of Optimal Control (MPOC) [48])
Assume that:

1. There exists a function V : [0, T ] × R+ → R which is C 1,2 , such that:

V (T , ·) = u(T , ·),

2. For any (n, c) ∈ A (w0 ):

Z t
Yt ≡ V (t, wt ) + u (s, cs ) ds is a supermartingale, (6)
0

3. For some (n∗ , c∗ ) ∈ A (w0 ) the process Y is a martingale.

Then (n∗ , c∗ ) is optimal, and the value of the problem starting from initial wealth w0 is
"Z T #
V (0, w0 ) = sup E u (t, ct ) dt + u (T , wT )
(n,c)∈A (w0 ) 0

By equation (6), and by applying Ito’s formula, we obtain:

1
dYt = Vt dt + Vw dw + Vww (dw)2 + u(t, c)dt
2
1 T 2

= Vw θ · σ dW + u(t, c) + Vt + Vw (rt w + θ · (µ − rt ) − c) + σ θ Vww dt
2
by assuming the dynamics of ω in equation (3).

Then, by integrating from t to t + h:

Z t+h Z t+h 
1 T 2

Yt+h −Yt = Vw θ · σ dW + u(s, c) + Vs + Vw (rs w + θ · (µ − rs ) − c) + σ θ Vww ds
t t 2
| {z }
Local martingale

So,

" Z t+h  #
1 T 2

Yt = E Yt+h − u(s, c) + Vs + Vw (rs w + θ · (µ − rs ) − c) + σ θ Vww ds | Ft
t 2
By setting
1
Lν v := µ(t, x, ν) · ∇v + tr a(t, x, ν)D 2 v
2
where,
d
X
a(t, x, ν) := σ (t, x, ν)σ (t, x, ν)t and tr a := aii
i=1

20
3 A DPP APPROACH 3.5 Verification theorem

and in view of the DPP,

" Z t+h ! #
∂ ν(s)
sup E − v + L v + L ds = 0
ν∈At,x t ∂t

Assuming that µ, a, L are continuous, and dividing the above equation by h and let-
ting h go to zero, we obtain
∂  
− v(t, x) + H x, t, ∇v(t, x), D 2 v(t, x) = 0
∂t
where
1
 
2 T 2
H(ω, ∇V (ω), D V (ω)) := sup µ(ω, θ, c) ∇V (ω) + tr(σ σ (ω, θ, c)D V (ω)) + U (c)
(θ,c)∈A 2
(7)

3.5 Verification theorem

Let β > 0 and f : Rn × A → R a measurable function. For x ∈ Rn , we denote by A(x)
the subset of controls α in A0 such that:
"Z ∞ #
−βs x
E e |f (Xs , αs )| ds < ∞
0

The gain function is defined as:

"Z ∞ #
J(x, α) = E e −βs
f (Xsx , αs ) ds
0

for all x ∈ Rn and α ∈ A(x).

The associated value function is:

v(x) = sup J(x, α)

α∈A(x)

Theorem 4 (Infinite horizon [11])

Let w ∈ C 2 (Rn ) , and satisfies a quadratic growth condition, i.e. there exists a constant
C such that  
|w(x)| ≤ C 1 + |x|2 , ∀x ∈ Rn

i- Suppose that
βw(x) − sup [La w(x) + f (x, a)] ≥ 0, x ∈ Rn
a∈A
lim sup e−βT E w XTx ≥ 0, ∀x ∈ Rn , ∀α ∈ A(x)


T →∞

Then w ≥ v on Rn .

21
 4 MERTON’S PROBLEM

ii- Suppose further that:

- For all x ∈ Rn , there exists a measurable function α̂(x), x ∈ Rn , valued in A

such that

βw(x) − sup [La w(x) + f (x, a)] = βw(x) − Lα̂(x) w(x) − f (x, α̂(x))
a∈A
=0

- The SDE
dXs = b (Xs , α̂ (Xs )) ds + σ (Xs , α̂ (Xs )) dWs
admits a unique solution, denoted by X̂sx , given an initial condition X0 = x,
satisfying h  i
lim inf e−βT E w X̂Tx ≤ 0
T →∞
n   o
- The process α̂ X̂sx , s ≥ 0 , lies in A(x).

Then:
w(x) = v(x), ∀x ∈ Rn
and α̂ is an optimal Markovian control

4 Merton’s problem
4.1 Problem formulation
As mentioned in section 3.2, we need to specify the following elements for the opti-
mal control problem:

• State process:

The model is driven by the differential equations of the wealth:

dωt = rωt dt + θt . (σ dWt + (µ − r) dt) − ct dt, (8)

where r, µ, σ , β, r̄, σr are constants and the correlation parameter η lies in [−1, 1].
The state process is in this case Zt := ωt . It is the unique solution of the follow-
ing equation:

dZt = rZt + θt . (µ − r) − ct dt + θt .σ dWt (9)

where Wt is a standard Brownian motion.

Thus, (Zt ) is a diffusion process with drift and volatility:

µ(ω, r, θ, c) := rωt + θt . (µ − r) − ct and σ (ω, r, θ, c) := θt .σ respectively.

22
4 MERTON’S PROBLEM 4.2 Time-homogeneous property:

• Control process:

The state process is controlled by the consumption

" # process ct and the money
c
invested in the risky assets θt , ν is then νt := t .
θt

• Set of admissible controls Aω0 :

The admissible control process for a state process ω which starts at ω0 at time
0 is:
θ
 
Aω0 := { (θ, c) | π := , c bounded, adapted ; ω ≥ 0 a.s. }
ω

The reason behind taking the constraint of bounded processes is the following:
θ
If π := takes a negative value that is too large in absolute value, then the
ω
agent needs to borrow a very big amount. This situation is not realistic: the
agent can’t borrow as much as she wants.

• Objective functional:

The time horizon is infinite in this problem. The objective functional is the
expected discounted utility derived from consumption:
"Z ∞ #
−ρt
J(Z(·), ν(·)) = E e U (c(t))dt | w0 = w0
0

ω1−R
where the utility function is the CRRA function U (ω) := .
1−R
The value function is the supremum of the objective functional:
"Z ∞ 1−R
#
−ρs cs
V (t, ω) = sup E e ds | w(t) = w
(θ,c)∈Aω t 1−R

4.2 Time-homogeneous property:

This problem is time-homogeneous: By changing the variable inside the integral
s 7→ s − t, we can deduce that:

V (t, ω) = e−ρt V (0, ω)

This is why we will focus on V (0, ω) and we will write W (ω) to denote V (0, ω):

V (t, ω) = e−ρt V (ω)

23
4.3 Scaling property 4 MERTON’S PROBLEM

4.3 Scaling property

" #
ĉ
Let’s consider an admissible control process ν̂t := t ∈ Aω̂0 . That means that the
θ̂t
wealth process ω̂t controlled by these processes and driven by the dynamics:

dω̂t = r ω̂t + θ̂t . (µ − r) − ĉt dt + θ̂t .σ dWt

is a.s positive, and starts at ω0 .

" 0# " #
ct ĉ
Now let’s consider λ > 0, the control process νt0 := 0 := λ t and the wealth
θt θ̂t
process ωt0 = λω̂t . It verifies the SDE:

dωt0 = rλω̂t + λθ̂t . (µ − r) − λĉt dt + λθ̂t .σ dWt

= λ dω̂t

Thus the wealth process ω0 is also a.s positive.

θ0
Furthermore, the processes π0 := 0 and c0 are bounded and adapted, so ν 0 is an
ω
admissible control process of a wealth that starts from λω̂0 .

So we can conclude that λν ∈ Aλω0 is equivalent to ν ∈ Aω0 .

We can also notice that, for all ω > 0:

ct1−R
" #
R∞
V (λω) = sup(θ,c)∈Aλω E 0 e −ρt dt | ω(0) = λω
1−R

1−R
" #
R∞ (λct )
= sup(λθ,λc)∈Aλω E 0
e−ρt dt | ω(0) = λω
1−R

(λct )1−R
" #
R∞
= sup(θ,c)∈Aω E e −ρt dt | ω(0) = ω
0 1−R

= λ1−R V (ω)

By taking ω = 1, we obtain V (λ) = λ1−R V (1), which means that V is propor-

tional to the utility function U .

Now, we need to write down the HJB equation and try to solve it. If we find a
solution, then most likely it will be the solution to our optimisation problem.

4.4 HJB equation

The associated HJB equation is given by:

24
4 MERTON’S PROBLEM 4.5 Optimal control parameters

ρV (ω) − H(ω, ∇V (ω), D 2 V (ω)) = 0, ω∈R (10)

where:
1
 
2 T 2
H(ω, ∇V (ω), D V (ω)) := sup µ(ω, θ, c) ∇V (ω) + tr(σ σ (ω, θ, c)D V (ω)) + U (c)
(θ,c)∈A 2
(11)
First, we compute the quantities:

µ(ω, k, c)T ∇V (ω) = (ωr + θ. (µ − r) − c)Vω

and
 
tr σ σ T (ω, θ, c)D 2 V (ω) = tr (θ.σ (θ.σ )> Vωω )

= θ.σ (θ.σ )> Vωω

By substituting these expressions in the equation (11), we obtain:

H(ω, ∇V (ω), D 2 V (ω)) = sup(θ,c)∈A [(ωr + θ. (µ − r) − c) Vω + 21 θ.σ (θ.σ )> Vωω + U (c)]
h i
= ωrVω + supc [−cVω + U (c)] + supθ θ. (µ − r) Vω + 21 θ.σ (θ.σ )> Vωω

4.5 Optimal control parameters

This step consists of solving the optimisation problems:

1
 
sup [−cVω + U (c)] and sup θ. (µ − r) Vω + θ.σ (θ.σ )> Vωω
c≥0 θ≥0 2
By the time-homogeneous and scaling properties (see sections 4.2 and 4.3),
the value function can be written as:
ω1−R
V (t, ω) = e−ρt V (ω) = e−ρt γM
−R
1−R
- Optimal c∗ :

c1−R
Since R > 0, the function c 7→ −cVω + is concave and thus it admits a
1−R
supremum.
The first order condition can be written as:

c1−R
∂c (−cVω + ) = −ω−R γM
−R
+ c−R = 0
1−R
i.e. the optimal proportion of wealth consumed is:
c∗
q∗ = = γM
ω

25
4.6 Value function 4 MERTON’S PROBLEM

- Optimal θ ∗ :

The function ω 7→ θ. (µ − r) Vω + 21 θ.σ (θ.σ )> Vωω is quadratic. Since Vωω < 0
(the value function should be concave), we can deduce that it has a maximum.
The gradient of this function is:

1
∇θ (θ. (µ − r) Vω + θ.σ (θ.σ )> Vωω ) = (µ − r)> Vω + θ > σ σ > Vωω
2

So that the first order condition can be written as:

θ ∗> σ σ > Vωω = − (µ − r)> Vω

i.e
Vω
θ∗ = − (σ σ > )−1 (µ − r)
Vωω
i.e
θ ∗ = ωR−1 (σ σ > )−1 (µ − r)

And the vector of optimal proportions of wealth invested in risky assets is:

πM = R−1 (σ σ > )−1 (µ − r)

4.6 Value function

Let’s plug the optimal values in the equation (10):

R −R −R 1 −R 1−R 2
 
e−ρt (γM w)1−R − ργM−R
u(w) + rwγM w + γM w |κ| /R = 0
1−R 2
−ρt 1−R −R 
e w γM 
1 2

RγM − ρ − (R − 1) r + |κ| /R = 0
1−R 2

where
κ ≡ σ −1 (µ − r)

So:  
ρ + (R − 1) r + 21 |κ|2 /R
γM =
R
Thus the value function is:

ω1−R
V (t, w) = e−ρt γM
−R
1−R

26
4 MERTON’S PROBLEM 4.7 Analysis of optimal values

Figure 1: The value function of Merton’s problem as a function of time t and wealth ωt

4.7 Analysis of optimal values

Both control processes c∗ and θ ∗ are proportional to the wealth.

The optimal controls suggest that the agent should allocate a constant pro-
portion πi∗ of wealth in each asset i. This constant is inversely proportional to the
volatility of the asset: the greater is the volatility, the higher is the risk and the lower
is the optimal proportion of the asset.

We can also notice that it’s proportional to the excess return of stocks over
bonds adjusted by the risk: this is called the sharpe ratio; it measures the profits
associated with risk-taking investment in a given stock.

If for an asset i, µi = r, which means that the expected return is equal to the
risk-free rate, the allocation to that risky asset i is null. This can be expected since
investing that money in the risk-free asset yields to the same return without taking
any risk.

27
 5 STOCHASTIC VOLATILITY

5 Stochastic Volatility
In Merton’s problem, a Black and Scholes model of stocks dynamics is assumed.
Although this model has proven a high tractability, the stocks log-returns prices
are supposed normally distributed with a constant variance and independent in-
crements. This is not consistent with the market implied distributions, and the de-
terministic volatility needs to be altered by a Stochastic Volatility (SV) model.

Different approaches have been suggested for the dynamics of volatility pro-
cess:

• Level dependent volatility models:

dSt = rSt dt + σ (St ) dWt

Cox and Ross [17], Geske [27], Rubinstein [49] and Bensoussan et al [4]
suggested and derived prices formulas for the Constant Elasticity of Variance
−(1−α)
(CEV) model where σ (St ) = σ St such that 0 < α < 1. This model, has the
same drawback as Black and Scholes: the volatility is deterministic.

• Stochastic volatility models:

dSt = µSt dt + σt St dWt

– Stein-Stein:
dσt = κ (θ − σt ) dt + σ̃ dBt
this is the same as Vasicek model for interest rates (see section 6 for pa-
rameters interpretation).

– Hull-White:
√
σt = vt ; dvt = µvt dt + σ̃ vt dBt
where the parameters µ and σ̃ are constants.

– Heston:
√ √
σt = vt ; dvt = κ (θ − vt ) dt + σ̃ vt dBt
where κ, θ, σ̃ are positive constants.

√
– Scott: σt = vt and variance process vt satisfies

vt = exp (yt ) ; dyt = k (ln(θ) − yt ) dt + dBt

where B and W are correlated Brownian motions with correlation coefficient

ρ constant.

28
5 STOCHASTIC VOLATILITY

• Exponentially weighted average of the historic log-price:

dSt = µSt dt + σt St dWt

Hobson and Rogers [29] suggested to model the randomness in the stochas-
tic volatility by the previous log-prices, so that no new Brownian motion is
introduced.
Rt
σt = f (Yt ) ; Yt = −∞ λeλ(s−t) (Xs − Xt ) ds ; Xt = ln(St ) (12)

In this variant of Merton’s problem, we model the volatility by the Exponen-

tially weighted average of the historic log-price model.

Yt can be expressed as:

Rt Rt
Yt = −∞
λeλ(s−t) Xs ds − Xt −∞
λeλ(s−t) ds
(13)
Rt
= e−λt −∞
λeλs Xs ds − Xt

Rt
By setting At and Ht as At := −∞
λeλs Xs ds and Ht := e−λt At , we have then:

dAt = λeλt Xt and dHt = −λe−λt At dt + e−λt dAt (By Ito’s formula)

= −λe−λt At dt + λXt dt
 Rt 
= −λ e−λt −∞ λeλs Xs ds − Xt dt

= −λYt dt

So that we have:
dYt = dHt − dXt
(14)
= −λYt dt− dXt
By Ito’s lemma applied to Xt , we obtain:

dSt 1 dSt2
dXt = −
St 2 St2
σt2
!
= µ− dt + σt dWt
2
By substituting this expression in the equation (14), we obtain:

1
  
dYt + λYt dt = − µ − f (Yt )2 dt + f (Yt ) dWt (15)
2

29
5.1 Problem formulation 5 STOCHASTIC VOLATILITY

5.1 Problem formulation

• State process:

The model is driven by two differential equations:

dωt = rωt dt + θt [f (Yt ) dWt + (µ − r) dt] − ct dt,

(16)
1
 
dYt = −λYt dt − µ − f (Yt )2 dt − f (Yt ) dWt
2
where µ, r, λ are constants.
!
ωt
We set the state process to be the vector Zt := . It is the unique solution of
Yt
the following equation:
 
 rωt + θt (µ − r) − ct  θt f (Yt ) 0 dB1t
" #" #
dZt =  1  dt +
 (17)
−λYt − µ + f (Yt )2  −f (Yt ) 0 dB2t
2
 
where B1t , B2t is a two-dimensional standard Brownian motion.

Thus, (Zt ) is a diffusion process with drift and volatility matrix:

" # " #
ωr + θ (µ − r) − c θf (Y ) 0
µ(ω, Y , θ, c) := and σ (ω, Y , θ, c) := respectively.
−λY − µ + 21 f (Y )2 −f (Y ) 0

• Control process:
" #
c
The control process is set to be the vector νt := t .
θt

• Objective functional:

The objective functional is the expected discounted utility derived from con-
sumption:
"Z ∞ #
−ρs
V (t, ω, y) = sup E e U (c(s))ds | ωt = ω, Yt = y
(θ,c)∈A(ω,y) t

ω1−R
where the utility function is the CRRA function U (ω) := .
1−R

30
5 STOCHASTIC VOLATILITY 5.2 HJB equation

5.2 HJB equation

The associated HJB equation is given by:

ρV (ω, Y ) − H(ω, Y , ∇V (ω, Y ), D 2 V (ω, Y )) = 0, ω, Y ∈ R (18)

where:
H(ω, Y , ∇V (ω, Y ), D 2 V (ω, Y )) := supθ≥0,c≥0 [µ(ω, Y , θ, c) ∇V (ω, Y )+

+ 12 tr(σ σ T (ω, Y , θ, c)D 2 V (ω, Y )) + U (c)

(19)
Let’s compute:

" #T " #
ωr + θ (µ − r) − c Vω
µ(ω, Y , k, c)T ∇V (ω, Y ) =
−λY − µ + 21 f (Y )2 VY
 
= (ωr + θ (µ − r) − c) Vω + −λY − µ + 12 f (Y )2 VY

and

" #" #T " #

  θf (Y ) 0 θf (Y ) 0 V ωω V ωY 
tr σ σ T (ω, Y , θ, c)D 2 V (ω, Y ) = tr 

 

f (Y ) 0 f (Y ) 0 VωY VY Y 
 
= f (Y )2 θ 2 Vωω − 2θVωY + VY Y

By substituting these expressions in the equation (19), we obtain:

1
 
H(ω, Y , ∇V (ω, r), D 2 V (ω, Y )) = supθ≥0,c≥0 [(ωr + θ (µ − r) − c) Vω + −λY − µ + f (Y )2 VY +
2
 
+ 12 f (Y )2 θ 2 Vωω − 2θVωY + VY Y + U (c)]

1
 
2
= ωrVω + −λY − µ + f (Y ) VY + 21 f (Y )2 VY Y +
2
h i
+ supθ≥0 θ (µ − r) Vω + 21 f (Y )2 θ 2 Vωω − f (Y )2 θVωY

+ supc≥0 [−cVω + U (c)]

(20)
By the same arguments as in sections (4.2) and 4.3, the value function can be
written as:

ω1−R
V (ω, Y ) = U (ω)g(Y ) = g(Y )
1−R

31
5.3 Optimal control parameters 5 STOCHASTIC VOLATILITY

5.3 Optimal control parameters

Here we derive the optimal values of consumption and the wealth invested in the
stock by solving the optimisation problems:

1
 
sup [−cVω + U (c)] and sup θ (µ − r) Vω + f (Y )2 θ 2 Vωω − f (Y )2 θVωY
c≥0 θ≥0 2

- Optimal c∗ :

By the same assumption as the previous section, the first order condition can
be written as:

∂c (−cVω + U (c)) = −ω−R g(Y ) + c−R = 0

i.e. the optimal proportion of wealth consumed is:

c∗ 1
q∗ = = (g(Y ))− R
ω

- Optimal θ ∗ :

The function ω 7→ θ (µ − r) Vω + 21 f (Y )2 θ 2 Vωω − f (Y )2 θVωY is polynomial of

second degree. Assuming Vωω < 0, we can deduce that it has a maximum at-
tained at:

−f (Y )2 ω−R g 0 (Y ) + (µ − r)g(Y )ω−R

θ∗ =
Rf (Y )2 g(Y )ω−R−1

And the optimal proportion of wealth invested in the risky asset is:

−f (Y )2 g 0 (Y ) + (µ − r)g(Y )
π∗ = 2
" (Y ) g(Y ) #
Rf
1 µ−r 1 0
= − g
R f (Y )2 g

We can notice that the optimal consumption and investment proportions of

wealth don’t depend on th wealth ω.
By substituting c∗ and θ ∗ in (30), we obtain the second order ODE:
n o2
(µ − r)g − f 2 g 0 1 1
Rg 1−1/R −ρg +r(1−R)g +(1−R) 2 Rg
+ f 2 g 00 −(λY +µ− f 2 )g 0 = 0 (21)
2f 2 2

32
5 STOCHASTIC VOLATILITY 5.4 Solving the ODE

5.4 Solving the ODE

By reordering the terms of the equation (21), we obtain:
(µ − r)2 (1 − R)
" # " #
(µ − r)(1 − R) 1 2 ∂g
r(1 − R) − ρ + g + − − µ + f − λY
2f 2 R R 2 ∂Y
(22)
2
!2
f 1 ∂g 1 ∂2 g
+(1 − R) + f 2 2 + Rg 1−1/R = 0
2R g ∂Y 2 ∂Y
This equation is not linear, to linearise it [43] we introduce the function h such that:
g(Y ) = h(Y )R .
The derivatives of g are:
∂g ∂h
= Rh(Y )R−1
∂Y ∂Y
 !2 
∂2 g ∂h ∂ 2h 
= Rh(Y )R−2 (R − 1)

+ h(Y ) 2 
 
∂Y 2 ∂Y ∂Y

So that: !2 !2
1 ∂g ∂h
= R2 h(Y )−R
g ∂Y ∂Y
 !2 
∂2 g ∂h ∂2h 
= Rh(Y )R−2 (R − 1)

+ h(Y ) 2 
 
∂Y 2 ∂Y ∂Y
Thus !2
f 2 1 ∂g 1 ∂2 g 1 ∂2 h
(1 − R) + f 2 2 = f 2 Rh(Y )R−1 2
2R g ∂Y 2 ∂Y 2 ∂Y

Plugging this formula into equation (22), we obtain the following linear second-
order ODE:

1 − R ρ (µ − r)2 (1 − R)
" #
1−R 1 2 ∂h
 
r − + h + −(µ − r) − µ + f − λY
R R 2f 2 R2 R 2 ∂Y
(23)
1 ∂2 h
+ f 2 2 +1 = 0
2 ∂Y
This equation doesn’t admit a closed-form solution in general.
Inspired by the paper [40], we will find a solution of the ODE by considering first a
finite time horizon T using a boundary condition.
ρ
By considering: h(Y , t) := e− R t h(Y ), the equation (23) has the form:

∂h ∂h 1 ∂2 h
(Y , t) + µ̂(Y , t) (Y , t) + σ̂ 2 (Y , t) 2 (Y , t) − φ̂(Y , t)h(Y , t) + F̂(Y , t) = 0
∂t ∂Y 2 ∂Y

33
5.4 Solving the ODE 5 STOCHASTIC VOLATILITY

defined for all Y and t ∈ [0, T ], with terminal condition

h(Y , T ) = ψ(Y )

where:
1−R 1
µ̂(Y , t) −(µ − r) − µ + f 2 (Y ) − λY
R 2
2
σ̂ 2 (Ŷ , t) f (Y )
(µ − r)2 (1 − R)
φ̂(Y , t) −r(1 − R) −
2f 2 (Y )R
ρ
F̂(Y , t) e− R t

Table 2: Coefficients of PDE of h

The Feynman-Kac theorem states that the solution can be written as a condi-
tional expectation:
"Z T R #
r RT
− t φ̂(Ŷτ ,τ )dτ − t φ̂(Ŷτ ,τ )dτ
   
h(Y , t) = E e F̂ Ŷr , r dr + e ψ ŶT | Ŷt = Y (24)
t

where Ŷ is an Itô process verifying the equation:

d Ŷ = µ̂(Ŷ , t)dt + σ̂ (Ŷ , t)dW (25)

such that W is a Brownian motion and the initial condition for Ŷ (t) is Ŷ (t) = Y .

Adding the transversality condition ([1] and [18], [42]):

 RT 
− t φ̂(Ŷτ ,τ )dτ
 
lim E e ψ ŶT | Ft = 0,
T →∞

and taking the limit T → ∞ in equation (24), we obtain :

"Z ∞ R #
s
− t φ̂(Ŷτ ,τ )dτ − Rρ s
h(Y , t) = E e e ds | Ŷt = Y
t

To find the value of this function, we will use Monte Carlo (MC) simulation.

34
5 STOCHASTIC VOLATILITY 5.5 Numerical solution

5.5 Numerical solution

In this section, we will implement numerically the optimal solution of the consumption-
investment problem assuming a Hobson-Rogers SV model.

The stock dynamics are:

dSt = µSt dt + σt St dWt

where
Rt
σt = f (Yt ) ; Yt = −∞
λeλ(s−t) (Xs − Xt ) ds ; Xt = ln(St )

Recall that the optimal controls are:

c∗ 1
q∗ = = (g(Y ))− R
ω
and " #
1 µ−r 1 0
π∗ = − g
R f (Y )2 g
and the value function is:
ω1−R
V (t, ω, y) = e−ρt U (ω)g(y) = e−ρt g(y)
1−R
where
g(y) = h(y)R
"Z ∞ Z s ! #
ρ
   
h(y) = E exp − φ̂ Ŷτ , τ dτ exp − s ds | Ŷ0 = y
0 0 R

d Ŷ = µ̂(Ŷ , t)dt + σ̂ (Ŷ , t)dW

The functions µ̂(Ŷ , t), σ̂ (Ŷ , t) and φ̂(Ŷ , t) are given in table (2).

To find q∗ , π∗ and V (t, ω, y), we will compute h(y) using Ns MC.

To derive π∗ , we need to compute an approximation of g 0 (Y ) by the finite-difference
method:
∂g(y) g(y + ∆) − g(y)
≈
∂y ∆

Steps:
We can summarize the steps of the approximation as the following:

1. Discretize the time onto a grid of Nt equally-spaced values : 0 = t0 ≤ · · · ≤ tN

for a large value of tN ;

35
5.5 Numerical solution 5 STOCHASTIC VOLATILITY

2. Simulate Ns paths of the process Y starting at y : (Yti )0≤i≤tN using MC simula-

tions:
r
1−R 1 2 TN σ (1 + Yti ) TN
 
Yti+1 = Yti + −(µ − r) − µ + f (Yti ) − λYti + q Z
R 2 N N i+1
2 + Yt2i

where Zi+1 are independent standard normal variables.

Z∞ Zs !
ρ
 
3. Approximate exp − φ (Yτ , τ) dτ exp − s ds by Riemann sums:
0 0 R
 
Nt i
tN  ρ tN
X  X  
h̃ = exp − φ(Ytj )  exp − ti ;


 Nt  R Nt
i=0 j=0

4. Take the average of these values and obtain h(y) then g(y):

s N
1 X
ĥ(y) = h̃i ; ĝ(y) = ĥ(y)R
Ns
i=1

∂g(y)
5. Compute approximations of :
∂y

ĝ(y + ∆) − ĝ(y)
gˆ0 (y) = ;
∆

6. Compute π∗ , q∗ and V .

For this numerical part, we will consider the following parameters:

Parameter Value
µ 0.15
r 0.05
R 2
λ 0.1
ρ 0.02
1 + y2
f(y) σ
2 + y2
σ 0.35
tN 25
Ns 1000
Nt 10000
∆ 10−1

Table 3: Parameter values of the numerical solution of SV model

36
5 STOCHASTIC VOLATILITY 5.5 Numerical solution

Plot of f:

In Hobson and Rogers stochastic volatility model, the volatility function is assumed
to be Lipshitz. The choice made for f guarantees this condition. Furthermore, f is
an even function.

1 + y2
Figure 2: Function f : y 7→ σ
2 + y2

Computation of g(5):

Figure 3: Ns Path simulations of Ŷt Figure 4: Ns simulations of ĥ(5) with

with Y0 = 5 and Ns = 1000 Ns = 1000

The figure (3) represents Ns MC paths simulations of the process Ŷt with the dynam-
ics given by equation (25). The figure (4) shows the simulated values of ĥ(5). These
values fluctuate around the mean with a small variance, which makes the algorithm
stable and reliable. We obtain ĥ(5) ≈ 9.75 and ĝ(5) ≈ 95.

37
5.5 Numerical solution 5 STOCHASTIC VOLATILITY

Plots and analysis of g, g 0 , q∗ and π∗

Let’s plot the functions g, g 0 and the optimal consumption and investment propor-
tions q∗ and π∗ :

Figure 5: The numerical solution of Figure 6: The numerical derivative of

the function y 7→ ĝ(y) the function y 7→ ĝ 0 (y)

We can notice that the function g has the same shape as the function f . The
plot of g is smooth and the one of g 0 is less smoother, this might be due to the step
∆ used to approximate the derivative.

Figure 7: The optimal consumption Figure 8: The optimal investment

proportion of wealth c∗ as a function proportion of wealth π∗ as a function
of y of y

We can see in figure 8 that the investment rate is surprisingly greater than 1
for some values of the offset. This might be due to the computational errors.
The consumption rate in figure 7 is high for low offsets, and it gets lower for
higher offsets y. This behaviour is similar to Merton’s problem, where the con-
sumption proportion is inversely proportional to the volatility when R > 1 which
corresponds to a risk averse investor.
The same behaviour can be seen for the investment proportion. In both models,
the investor invests less in stocks with high volatility.

38
5 STOCHASTIC VOLATILITY 5.5 Numerical solution

Analysis of parameters effect:

In this section, we will analyze the effects of varying the parameters of risk aversion
R, volatility σ of the function f and the discounting rate of past information λ on
the consumption and investment proportions.

Effect of risk aversion parameter R:

Figure 9: Effect of varying R on Figure 10: Effect of varying R on

the optimal consumption proportion the optimal investment proportion of
of wealth c∗ evaluated at y = 1 wealth π∗ evaluated at y = 1

The figures 9 and 10 show that the optimal consumption and investment pro-
portions decrease as the factor R increases, i.e as the investor becomes more risk-
averse. This can be guessed intuitively as a risk-averse investor would take less risk,
and thus would consume and invest less in the risky asset.

Effect of the volatility parameter σ :

Figure 11: Effect of varying σ on Figure 12: Effect of varying σ on

the optimal consumption proportion the optimal investment proportion of
of wealth c∗ evaluated at y = 1 wealth π∗ evaluated at y = 1

The plots 11 and 12 suggest that the consumption should be lower and the

39
5.5 Numerical solution 5 STOCHASTIC VOLATILITY

investment should be higher as the parameter σ increases. The first part is as ex-
pected, but the second part concerning the investment proportion is surprising since
intuitively, the higher is a volatility parameter the less should the investor take risk
and invest in the risky asset. The parameter σ can be interpreted as the limit of the
volatility of the asset prices as y → ∞.

Effect of the discounting rate of past information parameter λ:

Figure 13: Effect of varying λ on Figure 14: Effect of varying λ on

the optimal consumption proportion the optimal investment proportion of
of wealth c∗ evaluated at y = 1 wealth π∗ evaluated at y = 1

As λ increases, the investor relies more on the past stock prices. The optimal
proportions of the figures 13 and 14 suggest that the investor should consume and
invest more in the risky asset, this result might be surprising at first sight since the
investor won’t be relying on recent information; but at the same time, this might
be expected as for such stochastic volatility model, the investor believes that the
volatility is persistent and clustered. Thus, there is a trade-off between relying on
recent information and taking historic past values into account. We can therefore
expect these plots to be increasing for small values of λ and then decreasing for very
high values.

40
6 STOCHASTIC INTEREST RATE

6 Stochastic interest rate

In the classical Merton’s problem, the agent invests in a risky asset and a risk-free
asset assuming a deterministic fixed interest rate r. In this section, we relax this
assumption and consider the following stochastic differential equation for interest
rate :
drt = µ(t, rt )dt + σ (t, rt )dWt
where Wt is a standard Brownian Motion.

The choice of µ(t, rt ) and σ (t, rt ) differs from a model to another. The main
interest rate models [8] are:

- Vasicek (1977) [54]:

drt = k (θ − rt ) dt + σ dWt , α = (k, θ, σ )

- Cox-Ingersoll-Ross (CIR, 1985) [16]:

√
drt = k (θ − rt ) dt + σ rt dWt , α = (k, θ, σ ), 2kθ > σ 2

- Dothan / Rendleman and Bartter [47]:

 
(a− 21 σ 2 )t+σ Wt
drt = art dt + σ rt dWt , rt = x0 e , α = (a, σ )

- Exponential Vasicek:

rt = exp (zt ) , dzt = k (θ − zt ) dt + σ dWt , α = (k, θ, σ )

Each model has its advantages and drawbacks.

The Vasicek model assumes a linear equation that can be solved explicitly
which makes it tractable. The expectation of rt converges to θ when t → ∞, which
means that the process rt is mean reverting. However, this model allows rt to take
negative values with positive probability. This assumption used to be a drawback
before the financial crisis of 2008, but is legitimate since then.

The CIR model assumes a mean reverting non negative process rt and is usu-
ally closer to market implied distributions than Vasicek model, but rt follows a Chi-
squared distribution under this model and thus it is less tractable.

The third one is a Black and Scholes equation. Although rt has a log-normal
distribution and is tractable, this model is not mean reverting; in fact, the expecta-
tion of rt can converge to ∞ if a > 0 which is not realistic since interest rates are
controlled by the Central Bank.

41
6.1 Problem formulation 6 STOCHASTIC INTEREST RATE

Finally, the fourth equation models a log-normal, non negative and mean re-
verting process but is not tractable.

We will assume a Vasicek model for interest rate rt to derive the formulas. In
this model, the parameter θ can be interpreted as the long term mean, k as the speed
of convergence to the long term mean and σ is the volatility. The variance var(rt )
σ2
converges to as t 7→ ∞, which means that increasing the speed k or decreasing
2k
the volatility σ leads to decrease the uncertainty and hinders the variance to explode.

6.1 Problem formulation

• State process:

The model is driven by two differential equations:

dωt = rt ωt dt + θt (σ dWt + (µ − rt ) dt) − ct dt,

drt = β (r̄ − rt ) dt + σr dBt , (26)

dhW , Bit = η dt,

where µ, σ , β, r̄, σr are constants and the correlation parameter η lies in [−1, 1].
" #
ω
We set the state process to be the vector Zt := t . It is the unique solution of
rt
the following equation:
" # " # " 1#
rt ωt + θt (µ − rt ) − ct σ θt 0 dBt
dZt = dt + (27)
dB2t
p
β (r̄ − rt ) σr η σr 1 − η 2
 
where B1t , B2t is a two-dimensional standard Brownian motion.

Thus, (Zt ) is a diffusion process with drift and volatility matrix:

" # " #
ωr + θ (µ − r) − c σθ 0
µ(ω, r, θ, c) := and σ (ω, r, θ, c) := p respectively.
β (r̄ − r) σr η σr 1 − η 2

• Control process:

The control processes are the same as in Merton’s problem: the consumption
process ct and θ"t representing
# the value of the holding asset, ν is thus set to be
c
the vector νt := t .
θt

42
6 STOCHASTIC INTEREST RATE 6.2 HJB equation

• Objective functional:

The objective functional is the expected discounted utility derived from con-
sumption:
"Z ∞ #
−ρt
J(Z(·), ν(·)) = E e U (c(t))dt | ω(0) = ω, r(0) = r
0

ω1−R
where the utility function is the CRRA function U (ω) := .
1−R

Our goal is to find the supremum of the above functional at every time t:
"Z ∞ #
−ρs
V (t, ω, r) = sup E e U (cs ) ds | ω(t) = ω, r(t) = r
(θ,c)∈A(t,ω,r) t

Again the scaling property gives the following decomposition of the value function:

V (t, ω, r) = e−ρt U (ω)f (r)

6.2 HJB equation

The associated HJB equation for this infinite horizon problem is given by:

ρV (ω, r) − H(ω, r, ∇V (ω, r), D 2 V (ω, r)) = 0, ω, r ∈ R (28)

where:
1
 
2 T 2
H(ω, r, ∇V (ω, r), D V (ω, r)) := sup µ(ω, r, θ, c) ∇V (ω, r) + tr(σ σ (ω, r, θ, c)D V (ω, r)) + U (c)
θ≥0,c≥0 2

(29)

First, we compute the quantities:

" #T " #
ωr + θ (µ − r) − c Vω
µ(ω, r, k, c)T ∇V (ω, r) =
β (r̄ − r) Vr

= (ωr + θ (µ − r) − c) Vω + β (r̄ − r) Vr
and

43
6.3 Optimal control parameters 6 STOCHASTIC INTEREST RATE

" #" #T " #


T 2
  σ θ 0 σ θ 0 V ωω V ωr

tr σ σ (ω, r, θ, c)D V (ω, r) = tr  p p 
σr η σ r 1 − η 2 σr η σ r 1 − η 2 Vωr Vrr 

= σ 2 θ 2 Vωω + 2σ σr ηθVωr + σr2 Vrr

By substituting these expressions in the equation (6.2), we obtain:

H(ω, r, ∇V (ω, r), D 2 V (ω, r)) = supθ≥0,c≥0 [(ωr + θ (µ − r) − c) Vω + β (r̄ − r) Vr +

+ 12 σ 2 θ 2 Vωω + σ σr ηθVωr + 12 σr2 Vrr + U (c)]

= ωrVω + β (r̄ − r) Vr + 12 σr2 Vrr + supc≥0 [−cVω + U (c)]

h i
+ supθ≥0 θ (µ − r) Vω + 21 σ 2 θ 2 Vωω + σ σr ηθVωr
(30)
where:

ω1−R
V (ω, r) = U (ω)f (r) = f (r)
1−R

6.3 Optimal control parameters

Here we derive the optimal values of consumption and the wealth invested in the
stock by solving the optimisation problems:

1
 
sup [−cVω + U (c)] and sup θ (µ − r) Vω + σ 2 θ 2 Vωω + σ σr ηθVωr
c≥0 θ≥0 2

- Optimal c∗ :

c1−R
Since R > 0, the function c 7→ −cVω + is concave and thus it admits a
1−R
supremum.
The first order condition can be written as:

∂c (−cVω + U (c)) = −ω−R f (r) + c−R = 0

i.e. the optimal proportion of wealth consumed is:

c∗ 1
q∗ = = (f (r))− R
ω

44
6 STOCHASTIC INTEREST RATE 6.4 Solving the ODE

- Optimal θ ∗ :

The function ω 7→ θ (µ − r) Vω + 21 σ 2 θ 2 Vωω + σ σr ηθVωr is polynomial of second

degree. Assuming Vωω < 0, we can deduce that it has a maximum attained at:

σ σr ηω−R f 0 (r) + (µ − r)f (r)ω−R

θ∗ =
Rσ 2 f (r)ω−R−1

And the optimal proportion of wealth invested in the risky asset is:

θ ∗ σ σr ηf 0 (r) + (µ − r)f (r)

s∗ = =
ω Rσ 2 f (r)

By substituting c∗ and θ ∗ in (30), we obtain the second order ODE:

2
(µ − r)f + σ σr ηf 0

1−1/R 1
Rf − ρf + r(1 − R)f + (1 − R) + σr2 f 00 + β(r̄ − r)f 0 = 0 (31)
2σ 2 Rf 2

6.4 Solving the ODE

The ODE (31) doesn’t admit a closed-form solution.

However, we can notice that this ODE looks like the ODE (21). The main differ-
ence is the term of σ σr ηf 0 . In fact, in the previous problem, we managed to linearise
the ODE and write the solution as an expectation using Feynman Kac theorem. We
can have the same shape of ODE if we take η = 1 and σ = σr , which means that the
interest rate and the stock prices are driven by the same source of randomness and
they also have the same volatility. These assumptions don’t reflect the real dynamics
of stock prices and interest rates.

If η = 0, i.e the stock prices and interest rates are driven by independent Brow-
nian motions, then the ODE becomes:

1−1/R (µ − r)2 1
Rf − ρf + r(1 − R)f + (1 − R) 2
f + σr2 f 00 + β(r̄ − r)f 0 = 0
2σ R 2
Subsolution and supersolution method:

By considering
Z(r) := ln f (r)
Flemming and Pang [24], [45] showed under some conditions the existence of a
subsolution Z and supersolution Z̄ such that:

Z(r) 6 Z(r) 6 Z̄(r), ∀r ∈ R

45
6.4 Solving the ODE 6 STOCHASTIC INTEREST RATE

We state their findings adapted to our framework as the following:

Theorem 5 (Case R > 1 [24])
Define
−2(1 − R)
a1 := , a2 := µ − σ 2 R
3σ 2 R2
Then there exists a constant ā3 > 0 such that for any a3 ≥ ā3
 
Z(r) := −R ln a1 (r − a2 )2 + a3

is a subsolution.
Define

R−1
b1 := , b2 := µ − σ 2 R
2σ 2 R2
h i
2σr2 32 − (1 − R) + (1 − R)η 2 − 2ησ 3 σr R(1 − R) + β (r̄ − b2 )
b3 := b1
2β + β (r̄ − b2 )

If
 
σ 2 (1−R)
ρ ≥ µ(1 − R) + R 2β β (r̄ − b2 ) − 2 +

h i
3
2(1 − R)σr2 2 − (1 − R) + (1 − R)η 2 − 2ησ 3 σr (1 − R)2 R + (1 − R) β (r̄ − b2 )
− h i
2σ 2 R 2β + β (r̄ − b2 )

Then  −R 
2
Z(r) := log b1 (r − b2 ) + b2

is a supersolution .

Theorem 6 (Case R < 1 [24])

Assume
σ2
ρ > (1 − R)µ − R(1 − R)
2
Define K1 as
K1 := log K̃1
where
σ2
" #
−1 1
K̃1 R = ρ − µ(1 − R) + R(1 − R)
R 2
Then, for any K2 ≤ K1 , K2 is a subsolution.

46
6 STOCHASTIC INTEREST RATE 6.5 Numerical example

Define:
σr2 − 2βησ σr
R1 :=
σ 2 β 2 + σr2 − 2βησ σr
If
max {1, R1 } < R < 1
In addition, define
2
" #
γη
µ1 := −2σ 2 1 +
R
2(1 − R)ησr
µ2 := 2β +
σR
1−R
µ3 := − 2
2σ R
Let a , a be the real roots of µ1 a + µ2 a + µ3 = 0, 0 < a− < a+ .
+ − 2

Then for any a1 ∈ (a− , a+ ) , there exist constants a2 > K1 and C1 (a1 ) , where C1 (·)
is given by:
4λ1 (a1 ) λ3 (a1 ) − λ22 (a1 )
C1 (a1 ) :=
4λ1 (a1 )
2
λ1 (a1 ) := µ1 a1 + µ2 a1 + µ3
" #
2(1 − R)ησr µ(1 − R)
λ2 (a1 ) := − 2β r̄ + a1 + − (1 − R)
σR σ12 R)
(1 − R)µ2
λ3 (a1 ) := −a1 σr2 −
2σ 2 R

such that
Z̄(r) ≡ a1 r 2 + a2
is a supersolution if β > −C1 (a1 ).

6.5 Numerical example

For this numerical part, we will consider the following parameters:

Parameter Value
µ 0.15
r̄ 0.04828
R 2
η 0.45
ρ 0.02
σ 0.35
σr 0.01
β 0.2

Table 4: Parameter values of the numerical solution of stochastic interest rate model

47
6.5 Numerical example 6 STOCHASTIC INTEREST RATE

In this case R > 1, the supersolution and subsolution are given by theorem (5).

Figure 16: Supersolution q̄∗ (r) and

Figure 15: Supersolution f¯(r) and
subsolution q∗ (r) of the consumption
subsolution f (r) of the function f
proportion q∗ (r)

The figure 15 shows the evolution of the supersolution and subsolution of the
function f . The function f is comprised between these two functions. As the con-
1
sumption proportion in this model is q∗ = (f (r))− R , we can find upper and lower
bounds for the optimal consumption proportion given by the figure 16.

Figure 18: Effect of varying the risk

Figure 17: Effect of varying the risk aversion rate R on the subsolution
aversion rate R on the supersolution q∗ (r) of the consumption proportion
f¯(r) of the function f q∗ (r)

The figure 17 shows the effect of varying the parameter R on the supersolution
and subsolution. We can notice that the the supersolution attains a supremum value
that goes to ∞ when R is increasing. This results in a big upper bound and gives
poor information about the optimal value. As for the figure 18, it shows that the
shape of the consumption proportion is steadier. The other parameters r̄, β, η and ρ
don’t have a big effect on the shape of the supersolution/subsolution.
The drawback of this method is that we don’t have any information about the
investment proportion π∗ .

48
7 CONCLUSION

7 Conclusion
In this thesis, we studied utility maximisation problems. First, we assumed a discrete-
time single period model. This mathematically simple setting is an unrealistic rep-
resentation of the financial market, but is an important milestone in the theory of
finance.

We tackled after that the consumption-investment problem under a continuous

time model, where we considered a financial market consisting of a risk-free asset
and risky assets and the objective is to maximize the expected Constant-Relative-
Risk-Aversion utility of consumption under an infinite time horizon. The Dynamic
Programming Principle and the HJB equation were used to derive a colsed-form so-
lution of Merton’s problem: The investor should allocate constant proportions of
wealth to the risky assets and each fraction should be proportional to the Sharpe
Ratio of the asset which is equal to the expected excess return over interest rate ad-
justed by the volatility.

In Merton’s problem, several assumptions are made, for instance the stock
prices volatilities and interest rates are deterministic. We investigated the consis-
tency of these assumptions with the market implied distributions, then we modified
the model to take into account the randomness of these two parameters.

First, we assumed a stochastic volatility model where the dynamics of stock

price volatilities are described by Hobson and Rogers stochastic volatility expressed
as an exponentially-weighted mean of historic log-prices. This model captures the
volatility clustering and persistence and the Leverage Effect observed in the market
and is consistent with the volatility smile and skewness.

Writing down the HJB equation leads to a non linear ODE. Using a change of
variable, the equation is reduced to a linear ODE that can be solved using Feynman-
Kac theorem and expressed as an expectation of an Ito Process. Using MC simula-
tions, we simulated the process paths and computed an estimation of the expectation
and the optimal investment and consumption proportions. This numerical method
is powerful and gives high-standard approximations.

We found some similarities with Merton’s problem: the investor consumes and
invests less in stocks when the offset is high. We have also found that the more the
investor is risk-averse the less should be the consumption and investment propor-
tions. Furthermore, the greater is the discounting rate of past information λ (but
not too large at the same time), the more the investor is comfortable to consume
and invest.

We considered then a model where the interest rate follows a Vasicek model.
There is no closed-form formula of the optimums under the infinite time horizon.
We used the subsolution-supersolution method discussed by Fleming and Pang to
find upper and lower bounds of the solution. This method doesn’t provide a charac-

49
 7 CONCLUSION

terization of the optimal investment proportion, but gives an idea about the range
of the value function and the consumption fraction.

Rogers discussed a numerical method in his book [48] to solve this problem.
He used the policy improvement method to approximate the value function. This
method results in a non-steady algorithm that doesn’t converge for several values of
parameters. Furthermore, this approach doesn’t have a probabilistic interpretation.

Finally, in this thesis we relaxed only two of the non realistic assumptions in
Merton’s problem. We didn’t take into account neither transaction costs nor taxes
nor different utility functions. We also assumed a continuous trading and didn’t
impose any consumption constraints. For a more realistic and accurate portfolio
optimisation, one can use the Reinforcement Learning approach where an agent
is trained and is learning from a dynamic environment and aims to maximize its
reward.

50
REFERENCES REFERENCES

References
[1] Andrew Ang and Jun Liu. “Risk, return, and dividends”. In: Journal of Finan-
cial Economics 85.1 (2007), pp. 1–38.
[2] Kenneth J Arrow. Essays in the theory of risk-bearing. Tech. rep. 1970.
[3] Richard E Bellman and Stuart E Dreyfus. Applied dynamic programming. Prince-
ton university press, 2015.
[4] Alain Bensoussan, Michel Crouhy, and Dan Galai. “Stochastic equity volatility
and the capital structure of the firm”. In: Philosophical Transactions of the
Royal Society of London. Series A: Physical and Engineering Sciences 347.1684
(1994), pp. 531–541.
[5] MJ Best and B Ding. “On the continuity of the minimum in parametric quadratic
programs”. In: Journal of Optimization Theory and Applications 86.1 (1995),
pp. 245–250.
[6] Robert C Blattberg and Nicholas J Gonedes. “A comparison of the stable and
student distributions as statistical models for stock prices”. In: The journal of
business 47.2 (1974), pp. 244–280.
[7] Tim Bollerslev. “Generalized autoregressive conditional heteroskedasticity”.
In: Journal of econometrics 31.3 (1986), pp. 307–327.
[8] Damiano Brigo and Fabio Mercurio. Interest rate models-theory and practice:
with smile, inflation and credit. Springer Science & Business Media, 2007.
[9] Mark Broadie. “Computing efficient frontiers using estimated parameters”. In:
Annals of operations research 45.1 (1993), pp. 21–58.
[10] G Buffon. “Essai d’arithmétique morale”. In: Supplémenta l’Histoire naturelle
4 (1777), p. 1777.
[11] Álvaro Cartea, Sebastian Jaimungal, and José Penalva. Algorithmic and high-
frequency trading. Cambridge University Press, 2015.
[12] George Chacko and Luis M Viceira. “Dynamic consumption and portfolio choice
with stochastic volatility in incomplete markets”. In: The Review of Financial
Studies 18.4 (2005), pp. 1369–1402.
[13] Hao Chang and Kai Chang. “Optimal consumption–investment strategy un-
der the Vasicek model: HARA utility and Legendre transform”. In: Insurance:
Mathematics and Economics 72 (2017), pp. 215–227.
[14] Hao Chang and Xi-min Rong. “An investment and consumption problem with
CIR interest rate and stochastic volatility”. In: Abstract and Applied Analysis.
Vol. 2013. Hindawi. 2013.
[15] Vijay K Chopra and William T Ziemba. “The effect of errors in means, vari-
ances, and covariances on optimal portfolio choice”. In: Handbook of the
Fundamentals of Financial Decision Making: Part I. World Scientific, 2013,
pp. 365–373.

51
REFERENCES REFERENCES

[16] John C Cox, Jonathan E Ingersoll Jr, and Stephen A Ross. “A theory of the term
structure of interest rates”. In: Theory of valuation. World Scientific, 2005,
pp. 129–164.
[17] John C Cox and Stephen A Ross. “The valuation of options for alternative
stochastic processes”. In: Journal of financial economics 3.1-2 (1976), pp. 145–
166.
[18] Darrell Duffie. Dynamic asset pricing theory. Princeton University Press, 2010.
[19] Edwin J Elton and Martin J Gruber. “On the optimality of some multiperiod
portfolio selection criteria”. In: The Journal of Business 47.2 (1974), pp. 231–
243.
[20] Edwin J Elton and Martin J Gruber. “Portfolio theory when investment rel-
atives are lognormally distributed”. In: The Journal of Finance 29.4 (1974),
pp. 1265–1273.
[21] Robert F Engle. “Autoregressive conditional heteroscedasticity with estimates
of the variance of United Kingdom inflation”. In: Econometrica: Journal of the
Econometric Society (1982), pp. 987–1007.
[22] Eugene F Fama and James D MacBeth. “Risk, return, and equilibrium: Empir-
ical tests”. In: Journal of political economy 81.3 (1973), pp. 607–636.
[23] Wendell H Fleming and Daniel Hernández-Hernández. “An optimal consump-
tion model with stochastic volatility”. In: Finance and Stochastics 7.2 (2003),
pp. 245–262.
[24] Wendell H Fleming and Tao Pang. “An application of stochastic control theory
to financial economics”. In: SIAM Journal on Control and Optimization 43.2
(2004), pp. 502–531.
[25] Jean-Pierre Fouque, George Papanicolaou, and K Ronnie Sircar. Derivatives in
financial markets with stochastic volatility. Cambridge University Press, 2000.
[26] Andrea Gamba. “Portfolio analysis with symmetric stable paretian returns”.
In: Current topics in quantitative finance. Springer, 1999, pp. 48–69.
[27] Robert Geske. “The valuation of compound options”. In: Journal of financial
economics 7.1 (1979), pp. 63–81.
[28] Markowitz H. “Portfolio Selection”. In: The Journal of Finance (1952).
[29] David G Hobson and Leonard CG Rogers. “Complete models with stochastic
volatility”. In: Mathematical Finance 8.1 (1998), pp. 27–48.
[30] David Hobson, SL Alex, and Yeqi Zhu. “A multi-asset investment and con-
sumption problem with transaction costs”. In: Finance and Stochastics 23.3
(2019), pp. 641–676.
[31] John Hull. “An analysis of the bias in option pricing caused by a stochastic
volatility”. In: Advances in futures and options research 3 (1988), pp. 29–61.
[32] John Hull and Alan White. “The general Hull–White model and supercalibra-
tion”. In: Financial Analysts Journal 57.6 (2001), pp. 34–43.

52
REFERENCES REFERENCES

[33] John Hull and Alan White. “The pricing of options on assets with stochastic
volatilities”. In: The journal of finance 42.2 (1987), pp. 281–300.
[34] Dong Jiuying. “Optimal Investment Consumption Model with Vasicek Interest
Rate”. In: 2007 Chinese Control Conference. IEEE. 2007, pp. 391–394.
[35] JD Jobson. “Confidence regions for the mean-variance efficient set: an alter-
native approach to estimation risk”. In: Review of Quantitative Finance and
Accounting 1.3 (1991), p. 235.
[36] Herb Johnson and David Shanno. “Option pricing when the variance is chang-
ing”. In: Journal of financial and quantitative analysis (1987), pp. 143–151.
[37] Holger Kraft*. “Optimal portfolios and Heston’s stochastic volatility model:
an explicit solution for power utility”. In: Quantitative Finance 5.3 (2005),
pp. 303–313.
[38] Paul Lévy. “Calcul des probabilités”. In: (1925).
[39] Leonard C Maclean and K Laurence Weldon. “Estimating multivariate ran-
dom effects without replication”. In: Communications in Statistics-Theory and
Methods 25.7 (1996), pp. 1447–1469.
[40] Ravi Malhotra. “Portfolio Management in the Presence of Stochastic Volatil-
ity”. PhD thesis. University of London, 2008.
[41] Falcone Maurizio. “Optimal Control and the Dynamic Programming Princi-
ple”. In: Jan. 2014, pp. 1–8. ISBN: 978-1-4471-5102-9. DOI: 10.1007/978-1-
4471-5102-9_209-1.
[42] Robert C Merton. “Lifetime portfolio selection under uncertainty: The continuous-
time case”. In: The review of Economics and Statistics (1969), pp. 247–257.
[43] Marek Musiela and Thaleia Zariphopoulou. “An example of indifference prices
under exponential preferences”. In: Finance and Stochastics 8.2 (2004), pp. 229–
239.
[44] Eun-Jung Noh and Jeong-Hoon Kim. “An optimal portfolio model with stochas-
tic volatility and stochastic interest rate”. In: Journal of Mathematical Analysis
and Applications 375.2 (2011), pp. 510–522.
[45] Tao Pang. “Portfolio optimization models on infinite-time horizon”. In: Journal
of optimization theory and applications 122.3 (2004), pp. 573–597.
[46] Stanley Pliska. Introduction to mathematical finance. Blackwell publishers Ox-
ford, 1997.
[47] Richard J Rendleman. “The pricing of options on debt securities”. In: Journal
of Financial and Quantitative Analysis (1980), pp. 11–24.
[48] Leonard CG Rogers. Optimal investment. Vol. 1007. Springer, 2013.
[49] Mark Rubinstein. “Displaced diffusion option pricing”. In: The Journal of Fi-
nance 38.1 (1983), pp. 213–217.
[50] Louis O Scott. “Option pricing when the variance changes randomly: The-
ory, estimation, and an application”. In: Journal of Financial and Quantitative
analysis (1987), pp. 419–438.

53
REFERENCES REFERENCES

[51] H Mete Soner. Stochastic optimal control in finance. Scuola normale superiore,
2004.
[52] James Tobin. “Liquidity preference as behavior towards risk”. In: The review
of economic studies 25.2 (1958), pp. 65–86.
[53] Jakub Trybuła. “Optimal consumption problem in the Vasicek model”. In:
Opuscula Mathematica 35.4 (2015), pp. 547–560.
[54] Oldrich Vasicek. “An equilibrium characterization of the term structure”. In:
Journal of financial economics 5.2 (1977), pp. 177–188.
[55] James B Wiggins. “Option values under stochastic volatility: Theory and em-
pirical estimates”. In: Journal of financial economics 19.2 (1987), pp. 351–
372.

54
Acronyms Acronyms

Acronyms
CIR Cox-Ingersoll-Ross. 10

CRRA Constant-Relative-Risk-Aversion. 9, 19, 49

DPP Dynamic Programming Principle. 4, 16, 19

HARA Hyperbolic Absolute Risk Aversion. 9, 11

HJB Equation Hamilton Jacobi Bellman Equation. 4, 20

MC Monte Carlo. 34–37, 49

MPOC Martingale Principle of Optimal Control. 20

ODE Ordinary Differential Equation. 33, 45, 49

SV Stochastic Volatility. 28, 35

55