Scandinavian Actuarial Journal

ISSN: 0346-1238 (Print) 1651-2030 (Online) Journal homepage: www.tandfonline.com/journals/sact20

Robust optimal excess-of-loss reinsurance and

investment strategy for an insurer in a model with
jumps

Danping Li, Yan Zeng & Hailiang Yang

To cite this article: Danping Li, Yan Zeng & Hailiang Yang (2018) Robust optimal excess-of-
loss reinsurance and investment strategy for an insurer in a model with jumps, Scandinavian
Actuarial Journal, 2018:2, 145-171, DOI: 10.1080/03461238.2017.1309679

To link to this article: https://doi.org/10.1080/03461238.2017.1309679

Published online: 04 Apr 2017.

Submit your article to this journal

Article views: 805

View related articles

View Crossmark data

Citing articles: 35 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=sact20
SCANDINAVIAN ACTUARIAL JOURNAL, 2018
VOL. 2018, NO. 2, 145–171
https://doi.org/10.1080/03461238.2017.1309679

Robust optimal excess-of-loss reinsurance and investment

strategy for an insurer in a model with jumps
Danping Lia , Yan Zengb and Hailiang Yangc
a Department of Statistics and Actuarial Science, University of Waterloo, Waterloo, Canada; b Lingnan (University)
College, Sun Yat-sen University, Guangzhou, P.R. China; c Department of Statistics and Actuarial Science, The
University of Hong Kong, Hong Kong, P.R. China

ABSTRACT ARTICLE HISTORY

This paper considers a robust optimal excess-of-loss reinsurance- Received 5 August 2016
investment problem in a model with jumps for an ambiguity-averse insurer Accepted 19 March 2017
(AAI), who worries about ambiguity and aims to develop a robust optimal KEYWORDS
reinsurance-investment strategy. The AAI’s surplus process is assumed to Robust optimal control;
follow a diffusion model, which is an approximation of the classical risk excess-of-loss reinsurance
model. The AAI is allowed to purchase excess-of-loss reinsurance and invest and investment;
her surplus in a risk-free asset and a risky asset whose price is described by jump-diffusion model; utility
a jump-diffusion model. Under the criterion for maximizing the expected maximization;
exponential utility of terminal wealth, optimal strategy and optimal value ambiguity-averse insurer
function are derived by applying the stochastic dynamic programming
approach. Our model and results extend some of the existing results in
the literature, and the economic implications of our findings are illustrated.
Numerical examples show that considering ambiguity and reinsurance
brings utility enhancements.

1. Introduction
Reinsurance is an effective risk-spreading approach, while investment is an increasingly important
way of using insurers’ surplus, and both are popular in the insurance industry. In recent years,
the problem of optimal reinsurance-investment has attracted significant attention in the literature.
A number of scholars have considered the problem of maximizing the expected utility of terminal
wealth. For example, Yang & Zhang (2005) studies the optimal investment problem for an insurer in a
jump-diffusion risk model. Lin & Li (2011) discusses the optimal reinsurance-investment problem in a
jump-diffusion insurance risk model where the dynamics of the risky asset are governed by a constant
elasticity of variance (CEV) model. Liang & Yuen (2016) derives the optimal reinsurance strategy for
maximizing the expected exponential utility of terminal wealth in a risk model with dependent risks.
Likewise, other optimization objectives are considered in the literature. We refer readers to Schmidli
(2002) and Jang & Kim (2015) for the criterion of ruin probability minimization, and Pressacco
et al. (2011) and Bi et al. (2013) for the mean-variance criterion.
Although the problem of optimal reinsurance-investment has been widely investigated by many
scholars, two aspects merit further exploration. The majority of the above-mentioned literature
ignores ambiguity. However, it is a notorious fact that the return of risky assets is difficult to estimate
with precision. Thus, some scholars have advocated and investigated the effect of ambiguity on
portfolio selection, noting that in many cases, the parametric models used in theory, such as the Black-
Scholes model, contain significant uncertainties in parameter estimates. Take drift parameter as an

CONTACT Yan Zeng [email protected]

© 2017 Informa UK Limited, trading as Taylor & Francis Group
146 D. LI ET AL.

example. As the expected return of a risky asset is not known in a priori with any adequate precision,
the investor must typically account for a significant level of error in drift parameter estimates.
Compared with making ad-hoc decisions about how much error is contained in the estimates for the
parameters of risky assets, investors may consider alternative models that are close to the estimated
model. This method has been well accepted in quantitative finance to deal with portfolio selection
and asset pricing in case of ambiguity or misspecification. For instance, Anderson et al. (1999)
introduces ambiguity-aversion into the Lucas model and formulates alternative models. Uppal &
Wang (2003) extends the model in Anderson et al. (1999) and develops a framework that allows
investors to consider the level of ambiguity. Anderson et al. (2003) studies the continuous-time
asset pricing model in which the investor takes the model misspecification into account. Maenhout
(2004) optimizes an inter-temporal consumption problem with ambiguity, and derives the closed-
form expressions of the optimal strategies under ‘homothetic robustness’. Maenhout (2006) obtains
the optimal portfolio choice to maximize the expected power utility of the terminal wealth under
ambiguity and stochastic premium. Flor & Larsen (2014) considers an investor who is ambiguous
about the interest rate and stock returns models. For an insurer who manages her risk by purchasing
reinsurance and investing her surplus in a financial market, ambiguity situation is identical to that
of the above-mentioned investors. Moreover, the accurate estimation of an insurer’s surplus process
can also be called into question. An ambiguity-averse insurer (AAI) would hope for a systematic
and quantitative way to take ambiguity into account. For example, Korn et al. (2012) investigates
the optimal reinsurance problem and the optimal reinsurance-investment problem with ambiguity
by using the stochastic differential game approach. Yi et al. (2013) studies the problem of robust
optimal reinsurance-investment under the Heston model for an AAI. Yi et al. (2015) obtains the
robust optimal reinsurance-investment strategy under the benchmark and mean-variance criteria.
Pun & Wong (2015) considers the problem of robust optimal reinsurance-investment with multi-
scale stochastic volatility using a general concave utility function.
Although research on the robust optimal investment problem has been rapidly increasing in recent
years, very few of these contributions deals with the problem in relation to ambiguity with jumps,
which has a significant effect on the optimal strategy. Branger & Larsen (2013) analyzes the optimal
portfolio selection problem for an ambiguity-averse investor who invests in a risky asset following a
jump-diffusion process using the criterion of maximizing the expected power utility of the terminal
wealth. Aït-Sahalia & Matthys (2014) considers the optimal consumption-portfolio selection problem
in the presence of ambiguity where the price of the risky asset satisfies a Lévy process. Both Branger &
Larsen (2013) and Aït-Sahalia & Matthys (2014) point out that the risks related to the uncertainty of
the drift and the probability of jumps are fundamentally different in the portfolio selection problem,
such that ignoring ambiguity with respect to (w.r.t.) the jump risk may result in large losses in
the financial market. For robust optimal reinsurance-investment problem with jumps, Zeng et al.
(2016) and Zheng et al. (2016) study the optimal proportional reinsurance-investment problem with
ambiguity under criteria of mean-variance and expected utility maximization, respectively.
Unlike Yi et al. (2013), Yi et al. (2015), Pun & Wong (2015), Zeng et al. (2016) and Zheng
et al. (2016), we are interested in the excess-of-loss reinsurance, which is preferred than proportional
reinsurance in most situations (see Asmussen et al. 2000). Recently, more and more scholars focus
on the optimal excess-of-loss reinsurance-investment problems, such as Gu et al. (2012), Zhao
et al. (2013), and so on. To the best of our knowledge, this paper is a prior research on the robust
optimal excess-of-loss reinsurance-investment problem with jumps for an AAI. In our model, the
insurer’s surplus process is assumed to be a Brownian motion with drift that can be considered
as an approximation of the classical insurance risk models, and the insurer is allowed to purchase
excess-of-loss reinsurance and invest her surplus in a risk-free asset and a risky asset whose price
process is described by a jump-diffusion model. Given that the market (true model) may deviate
from the estimated model (reference model) in reality, we incorporate ambiguity into our study, and
assume that the insurer is ambiguity-averse to diffusion and jump risks. Following Maenhout (2004;
2006), the ambiguity level is chosen as inversely proportional to the optimal value function. Moreover,
 SCANDINAVIAN ACTUARIAL JOURNAL 147

depending on the available information, the AAI may exhibit different levels of ambiguity to diffusion
and jump risks. The infrequent nature of jumps in the price process for the risky asset makes it hard
to estimate the intensity of jump risk, which indicates that the AAI is more ambiguity averse to the
jump risk than to the diffusion risk, making it seem natural to have different levels of ambiguity
aversion to diffusion and jump risks. Based on the above setup, we formulate a robust optimization
problem with alternative models, and derive the explicit expressions of the robust optimal excess-
of-loss reinsurance-investment strategy to maximize the expected exponential utility of terminal
wealth. Some special cases of our model and results are provided, and the economic implications of
our findings and utility enhancements from considering ambiguity and reinsurance are analyzed
using numerical examples. The main contributions of this paper are as follows: (i) Ambiguity
with jumps is introduced into the optimal excess-of-loss reinsurance-investment framework; (ii)
utility enhancements from considering ambiguity and reinsurance are presented, which reveals that
ambiguity and reinsurance should not be ignored; and (iii) some special cases of our model, such
as the cases of investment-only, ambiguity-neutral insurer (ANI) and no jump, are provided, which
demonstrates that our model are more general and can reduce to many special cases considered in
the literature.
The remainder of this paper is organized as follows. Section 2 describes the formulation of
the model. Section 3 derives the explicit expressions of the robust optimal reinsurance-investment
strategy and the corresponding optimal value function. Section 4 provides some special cases of our
model. Section 5 presents some numerical examples and sensitivity analysis of utility enhancements
to illustrate our results. Section 6 concludes the paper.

2. General formulation
Let (, F, {Ft }t∈[0,T] , P) be a filtered complete probability space satisfying the usual condition, where
T > 0 is a finite constant representing the investment time horizon, Ft stands for the information
available until time t, and P is a reference probability.
Suppose that an insurer’s surplus process follows a diffusion model. To understand better that
the diffusion model can be considered as an approximation of the classical insurance risk model, we
start with the classical Cramér–Lundberg (C–L) model. In the C–L model, without reinsurance and
investment, the surplus process of an insurer is described by


N1 (t)
R(t) = x0 + pt − Zi ,
i=1

 1 (t)
where x0  0 is the initial surplus; p is the premium rate; N i=1 Zi is a compound Poisson process,
representing the cumulative claims up to time t; {N1 (t)}t∈[0,T] is a homogeneous Poisson process
with intensity λ1 > 0; and the claim sizes Z1 , Z2 , . . . are assumed to be independent and identically
distributed (i.i.d.) positive random variables with finite first moment E[Zi ] = μZ and second moment
E[Zi2 ] = σZ2 . Z1 , Z2 , . . . are further assumed to be independent of N1 (t) with common distribution
F(z). Denote by D = sup{z : F(z)  1} < +∞, then F(0) = 0, 0 < F(z) < 1 for 0 < z < D and
F(z) = 1 for z  D. Suppose that the premium rate p is calculated according to the expected value
principle, i.e. p = (1 + η)λ1 μZ , where η > 0 is the safety loading of the insurer.
To disperse the underlying insurance business risk, the insurer is allowed to purchase excess-of-
loss reinsurance. Let a be a (fixed) excess-of-loss retention level and Zi(a) = min{Zi , a} denote the
part of the claims held by the insurer. Then, the surplus process of the insurer becomes


N1 (t)
R̄ (a) (a)
(t) = x0 + p t − Zi(a) ,
i=1
148 D. LI ET AL.

where the premium rate

(a) (a)
p(a) = (1 + η)λ1 μZ − (1 + θ )λ1 (μZ − E[Zi ]) = (η − θ )λ1 μZ + λ1 (1 + θ )E[Zi ],

in which θ denotes the safety loading of the reinsurer. Suppose that θ > η, which implies that
the reinsurance is not cheap. According to Grandell (1991), the surplus process R̄(a) (t) can be
approximated by the following diffusion model

(a) (a)
dR̄(a) (t) = (p(a) − λ1 E[Zi ])dt + λ1 E[(Zi )2 ]dB1 (t)
√
= λ1 [θ μ̄(a) + (η − θ )μZ ]dt + λ1 σ̄ (a)dB1 (t),

where {B1 (t)}t∈[0,T] is a standard Brownian motion and

 a  a
(a)
μ̄(a) = E[Zi ] = (1 − F(s))ds = F̄(s)ds,
0 0 
a a
(a)
(σ̄ (a))2 = E[(Zi )2 ] = 2s(1 − F(s))ds = 2sF̄(s)ds.
0 0

Moreover, we assume that the insurer is allowed to invest in a financial market comprising a
risk-free asset and a risky asset. The price process of the risk-free asset is described by

dS0 (t) = r0 S0 (t)dt, (2.1)

where r0 > 0 represents the risk-free interest rate. The price process of the risky asset follows a
jump-diffusion process
⎡ ⎤

N2 (t)
dS(t) = S(t − ) ⎣μdt + σ dB2 (t) + d Yi ⎦ , (2.2)
i=1

where μ and σ are constant; {B2 (t)}t∈[0,T] is a standard Brownian motion; {N2 (t)}t∈[0,T] , representing
the number of the risky asset price’s jumps that occur during time interval [0, t], is a homogeneous
Poisson process with intensity λ2 ; Yi is the ith jump amplitude of the risky asset price; and Yi ,
i = 1, 2, . . . are i.i.d. random variables with distribution function G(y), finite first moment E[Yi ] = μY
and second moment E[Yi2 ] = σY2 . Similar to Yi et al. (2013) and Pun & Wong (2015), we assume that
N2 (t)
{B1 (t)}t∈[0,T] , {B2 (t)}t∈[0,T] and i=1 Yi are independent, and that P{Yi  −1 for all i  1} = 1
to ensure that the risky asset price remains positive. Generally, the expect return of the risky asset
is larger than the risk-free interest rate, so we assume that μ + λ2 μY > r0 . In Equation (2.2), the
diffusion term captures normal market movements, and the jumps describe sudden and unusually
disastrous events. Next, we use a Poisson random measure N(·, ·) on  × [0, T] × [−1, ∞) to denote
2 (t)
the compound Poisson process N i=1 Yi as


N2 (t)  t ∞
Yi = yN(ds, dy), ∀t ∈ [0, T].
i=1 0 −1

Denote by ν(dt, dy) = λ2 dtdG(y), then

⎡ ⎤

N2 (t)  t ∞
E⎣ Yi ⎦ = yν(ds, dy), ∀t ∈ [0, T],
i=1 0 −1
 SCANDINAVIAN ACTUARIAL JOURNAL 149

where ν(·, ·) is the compensator of the random measure N(·, ·). Thus, the compensated measure
Ñ(·, ·) = N(·, ·) − ν(·, ·) is related to the compound Poisson process as follows

⎡ ⎤
 t ∞ 
N2 (t) 
N2 (t)
y Ñ(ds, dy) = Yi − E ⎣ Yi ⎦ , ∀t ∈ [0, T].
0 −1 i=1 i=1

Let u := {u(t) := (a(t), π(t))}t∈[0,T] be the reinsurance and investment strategy,

π(t) is the amount of money invested in the risky asset at time t, the remainder X u (t) − π(t)
is invested in the risk-free asset, X u (t) is the wealth at time t associated with strategy u. Then, the
wealth process {X u (t)}t∈[0,T] can be described by

dS0 (t) dS(t)

dX u (t) = dR̄(a) (t) + (X u (t) − π(t)) + π(t)
S0 (t) S(t − )
= r0 X u (t) + λ1 [θ μ̄(a) + (η − θ )μZ ] + π(t) μ − r0 dt (2.3)
√ ∞
+ λ1 σ̄ (a)dB1 (t) + π(t)σ dB2 (t) + π(t)yN(dt, dy).
−1

In the traditional reinsurance-investment model, the insurer is assumed to be ambiguity-neutral

with objective function, as follows

sup Et,x [U(X u (T))] = sup E[U(X u (T))|X u (t) = x], (2.4)
u∈ ˜ u∈ ˜

where ˜ is the set of all admissible strategies u in a given market. The utility function U(x) is
typically increasing and concave (U  (x) < 0). However, it is reasonable to assume that the insurer
is ambiguity-averse and thus wants to guard herself against worst-case scenarios. We assume that
the knowledge about ambiguity for the AAI is described by probability P, namely, the reference
probability (or model). However, she is sceptical about this reference model, and hopes to consider
alternative models. Following Anderson et al. (1999), the AAI recognizes that the model under
probability P is an approximation of the true model, thus she notes alternative models, broadly
defined here as a class of probabilities that are equivalent to P as follows:

Q := {Q|Q ∼ P}.

Definition 2.1: [Admissible strategy] A strategy u = {u(t) := (a(t), π(t))}t∈[0,T] is said to be

admissible if
(i) ∀t ∈ [0, T], a(t) ∈ [0, D];  
∗ T
(ii) u is predictable w.r.t. {Ft }t∈[0,T] , and EQ 0 (a(t)) 2 + (π(t))2 dt < ∞;

(iii) ∀(t, x) ∈ [0, T] × R, Equation (2.3) has a pathwise unique solution {X u (t)}t∈[0,T] with
Q∗  
Et,x U(X u (T)) < +∞, where Q∗ is the chosen model to describe the worst case and will be
shown later.
Let be the set of all admissible strategies.
It is obvious that is not empty, since at least it contains deterministic controls. For such a class
of controls, existence and path uniqueness of the solution to Equation (2.3) is proved in Øksendal &
Sulem (2007).
150 D. LI ET AL.

By Girsanov’s Theorem, ∀Q ∈ Q, there exists := {φ(t) := (φ1 (t), φ2 (t), φ3 (t))}t∈[0,T] 1 such
that
dQ
= (T),
dP
where
 t   t 
1 t 1 t
(t) = exp φ1 (s)dB1 (s) − (φ1 (s))2 ds + φ2 (s)dB2 (s) − (φ2 (s))2 ds
 t ∞
0 2 0  t ∞ 0 2 0 (2.5)
+ ln φ3 (s)N(ds, dy) + (1 − φ3 (s))ν(ds, dy)
0 −1 0 −1

is a P-martingale. Karatzas & Shreve (1988) can be consulted for this theorem. By Girsanov’s theorem,
under probability Q,
dB1Q (t) = dB1 (t) − φ1 (t)dt,
and
dB2Q (t) = dB2 (t) − φ2 (t)dt
are Brownian motions. Following Branger & Larsen (2013), for tractability and ease of interpretation,
the distribution of the claim Yi is assumed to be known and is restricted to be identical under
probabilities P and Q, i.e. EQ [h(y)] = EP [h(y)], where h( · ) is a function of y. Under Q, the random
measure N Q (dt, dy) has compensator measure given by λ2 φ3 (t)G(dy)dt. Thus, the dynamic of the
wealth process under probability Q is
√
dX u (t) = r0 X u (t) + λ1 [θ μ̄(a) + (η − θ )μZ ] +π(t)(μ − r0 ) + λ1 σ̄ (a)φ1 (t) + π(t)σ φ2 (t) dt
√ ∞
+ λ1 σ̄ (a)dB1Q (t) + π(t)σ dB2Q (t) + π(t)yN Q (dt, dy).
−1
(2.6)
Suppose that the insurer tries to seek a robust optimal control which is the best choice in some
worst-case models. Inspired by Maenhout (2004) and Branger & Larsen (2013), we formulate a robust
control problem to modify problem (2.4) as follows
 
T
Q
sup inf Et,x (s, X u (s), φ(s))ds + U(X u (T)) , (2.7)
u∈ ∈ 0

where
(φ1 (t))2 (φ2 (t))2 λ2 (φ3 (t) ln φ3 (t) − φ3 (t) + 1)
(t, X u (t), φ(t)) = + + ,
2ϕ 1 (t)
B 2ϕ 2 (t)
B ϕ J (t)
and the expectation is calculated under the alternative probability Q; ϕ B1 (t), ϕ B2 (t) and ϕ J (t) are
strictly positive deterministic functions and represent the preference parameters for ambiguity-
aversion, which measure the degree of confidence to the reference probability P at time t; and
deviations from the reference model are penalized by the first three terms in the expectation, which
depends on the relative entropy arising from the diffusion and jump risks. In Appendix 1, we show
that the increase in relative entropy from t to t + dt equals
 
1 1
(φ1 (t))2 + (φ2 (t))2 + λ2 (φ3 (t) ln φ3 (t) − φ3 (t) + 1) dt. (2.8)
2 2
1 := {φ(t) := (φ1 (t), φ2 (t), φ3 (t))}t∈[0,T ] satisfies three conditions: (i) φ1 and φ2 are Ft -adapted, and φ3 is Ft -predictable; (ii)
    
φ3 (t) > 0, for a.a. (t, ω) ∈ [0, T ] × ; and (iii) E exp 12 0T (φ12 (t) + φ22 (t))dt + λ2 0T φ3 (t) ln φ3 (t) − φ3 (t) + 1 dt < ∞.
We denote for the space of all such processes . Condition (iii) can be referred to Branger & Larsen (2013) and Zheng et al.
(2016)
 SCANDINAVIAN ACTUARIAL JOURNAL 151

According to Maenhout (2004), we know that the larger ϕ B1 (t), ϕ B2 (t) and ϕ J (t) are, the less the
deviations from the reference model are penalized. Furthermore, the AAI has less faith in the reference
model, and she is more likely to consider alternative models. Hence, the AAI’s ambiguity aversion is
increasing w.r.t. ϕ B1 (t), ϕ B2 (t) and ϕ J (t).
To solve problem (2.7), we define the optimal value function V (t, x) as
  
T  u
V (t, x) = sup inf E Q (s, X (s), φ(s))ds + U(X (T))Xt = x .
u u (2.9)
u∈ ∈ t

Let C 1,2 ([0, T] × R) denote a class of functions that are continuously differentiable w.r.t. t on [0, T],
and twice continuously differentiable w.r.t. x on R. Similar to Maenhout (2006) and Branger & Larsen
(2013), for any V (t, x) ∈ C 1,2 ([0, T] × R), according to the principle of dynamic programming, we
can derive the HJB equation for problem (2.9):

sup inf Aφ,u V (t, x) + (t, x, φ) = 0, (2.10)

+
u∈[0,D]×R φ∈R×R×R

with the boundary condition V (T, x) = U(x), where u = (a, π ), φ = (φ1 , φ2 , φ3 ) denote the values
that u and take, and
 √ 
Aφ,u V (t, x) = Vt + Vx r0 x + λ1 (θ μ̄(a) + (η − θ )μZ ) + π(μ − r0 ) + λ1 σ̄ (a)φ1 + π σ φ2
1
+ Vxx [λ1 (σ̄ (a))2 + π 2 σ 2 ] + λ2 φ3 EQ [V (t, x + π y) − V (t, x)],
2
here, Vt , Vx and Vxx represent the partial derivatives of V (t, x) w.r.t. the corresponding variables.
Proposition 2.2: If there exist a function W(t, x) ∈ C 1,2 ([0, T] × R) and a Markovian control
∗ ∗
( ∗ , u∗ ) ∈ × , ∗ (t) = ∗ (t, X u (t)), u∗ (t) = u∗ (t, X u (t)) such that
∗
(i) for any φ ∈ R × R × R+ , Aφ,u W(t, x) + (t, x, φ)  0;
∗
(ii) for any u ∈ [0, D] × R, Aφ ,u W(t, x) + (t, x, φ ∗ )  0;
∗ ∗
(iii) Aφ ,u W(t, x) + (t, x, φ ∗ ) = 0;
(iv) for all ( , u) ∈ × , limt→T− W(t, X u (t)) = U(X u (T));
(v) {W(τ , X u (τ ))}τ ∈T and {(τ , X u (τ ), φ(τ ))}τ ∈T are uniformly integrable, where T denotes the
set of stopping times τ  T.
Then W(t, x) = V (t, x) and ( ∗ , u∗ ) is an optimal control.
Proof: The proof is similar to the proof of Theorem 3.2 in Mataramvura & Øksendal (2008).

3. Explicit robust control: exponential utility

To derive explicit results, we need to make some assumptions regarding the AAI’s utility. Suppose
that the AAI has an exponential utility, i.e.

1 −mx
U(x) = − e , (3.1)
m
where m > 0 is a constant representing the absolute risk aversion coefficient. As we know, the
exponential utility function plays an important role in insurance mathematics and actuarial practice.
It is the only utility function under the principle of ‘zero utility’ giving a fair premium that is
independent of the level of insurers’ wealth (see Gerber 1979).
Following Maenhout (2004), we assume that ϕ B1 (t), ϕ B2 (t) and ϕ J (t) are state-dependent as

γ B1 γ B2 γJ
ϕ B1 (t) = − , ϕ B2 (t) = − , ϕ J (t) = − , (3.2)
mV mV mV
152 D. LI ET AL.

where the ambiguity aversion coefficients γ B1 , γ B2 and γ J are nonnegative and describe the insurer’s
attitudes to the model uncertainty. From Equation (3.2), we find that ϕ B1 (t), ϕ B2 (t) and ϕ J (t) are
increasing w.r.t. parameters γ B1 , γ B2 and γ J , respectively.
Then, the robust optimal reinsurance-investment strategy and the corresponding optimal value
function can be derived and summarized in the following theorem.
Theorem 3.1: For problem (2.7) with exponential utility function (3.1) and assumption (3.2), the
robust optimal reinsurance-investment strategy u∗ = {(a∗ (t), π ∗ (t))}t∈[0,T] is given by

a∗ (t)
⎧ −r0 (T−t) −r0 (T−t)
⎪ θe !  , θe
⎪  !  + D · 1 
⎨ γ B1 + m · 1 l
⎪ θe−r0 (T−t)
max{l(D),0} l(D)max l θe−r0 (T−t)
,0 γ B1 + m
∈ [0, D],
= γ B1 +m γ B1 +m
⎪
⎪ θe−r0 (T−t)
⎪
⎩ D · 1{l(D)0} , ∈ [D, +∞),
γ B1 + m
(3.3)

 ∗ r (T−t)

∗ e−r0 (T−t) μ − r0 λ2 EQ [ye−mπ (t)ye 0 ] γJ
EQ [e−mπ
∗ (t)yer0 (T−t)
−1]
π (t) = B + e m , (3.4)
γ 2 +m σ2 σ2

and the optimal value function is given by

1 −m[er0 (T−t) x−f¯ (t)]

V (t, x) = − e , (3.5)
m

where

1
l(a) = −λ1 θ μ̄(a)Vmer0 (T−t) + λ1 m(γ B1 + m)(σ̄ (a))2 V e2r0 (T−t) , (3.6)
2
 T  T
λ (η − θ )μ
f¯ (t) = (T−t)
1 Z r
(1 − e 0 )+ l̄1 (ω)dω − l̄2 (ω)dω, (3.7)
r0 t t
 a∗ (ω)
l̄1 (ω) = er0 (T−ω) [λ1 (γ B1 + m)ser0 (T−ω) − λ1 θ ]F̄(s)ds, (3.8)
0
1
l̄2 (ω) = π ∗ (ω)(μ − r0 )er0 (T−ω) − (γ B2 + m)σ 2 (π ∗ (ω))2 e2r0 (T−ω)
2 !
λ2 γ J Q −mπ ∗ (ω)yer0 (T−ω)
E [e −1]
+ J 1−e m . (3.9)
γ

Proof: See Appendix 2.

Proposition 3.2: Equation (3.4) has a unique positive root, i.e. there exists a unique π ∗ (t) ∈ [0, +∞)
that satisfies Equation (3.4).
Proof: To proof the existence-uniqueness of π ∗ (t), Equation (3.4) can be transformed into

∗ yer0 (T−t) γJ Q [e−mπ ∗ yer0 (T−t) −1]

(γ B2 + m)π ∗ σ 2 er0 (T−t) = μ − r0 + λ2 EQ [ye−mπ ]e m E .

Suppose that

r (T−t) γJ Q [e−mπ yer0 (T−t) −1]

h(π ) = μ − r0 + λ2 EQ [ye−mπ ye 0 ]e m E − (γ B2 + m)π σ 2 er0 (T−t) .
 SCANDINAVIAN ACTUARIAL JOURNAL 153

We obtain

r (T−t) γJ Q −mπ yer0 (T−t)

h (π ) = −λ2 mer0 (T−t) EQ [y 2 e−mπ ye 0 ]e m E [e −1]
" r (T−t)
#2 γ J
Q −mπ ye 0 (T−t) −1]
r
− λ2 γ J er0 (T−t) EQ [ye−mπ ye 0 ] e m E [e − (γ B2 + m)σ 2 er0 (T−t) < 0,

which implies that h(π ) is a decreasing function w.r.t. π . Furthermore, we have h(0) = μ − r0 +
μ−r0 +λ2 EQ [y]
λ2 EQ [y] > 0. Also, we can find that if π > (γ B2 +m)σ 2 er0 (T−t) > 0, we have h(π ) < 0. Therefore,
Equation (3.4) has a unique positive root.
Remark 3.3: If the distribution of claim size satisfies
⎧
⎨ θ μ̄(D) − (γ + m)e (σ̄ (D)) > 0,
B1 r0 T 2
⎪
2 (3.10)
⎪
⎩
min {ψ(T), ψ(0)} > 0,

" −r (T−t) # r0 (T−t)

" " −r (T−t) ##2 r0 (T−t)
where ψ(t) = θ μ̄ θeγ B1 +m − (γ +m)e − θ μ̄(D) + (γ +m)e 2 (σ̄ (D)) ,
B1 B1 2
σ̄ θeγ B1 +m
0 0
2
for example, exponential distribution and uniform distribution with certain parameters, etc., the
robust optimal reinsurance-investment strategy becomes
θ
(i) if D > γ B1 +m
, the robust optimal reinsurance-investment strategy is given by

θ e−r0 (T−t)
a∗ (t) = , 0  t  T, (3.11)
γ B1 + m

 ∗ r (T−t)

e−r0 (T−t) μ − r0 λ2 EQ [ye−mπ (t)ye 0 ] γJ ∗ r (T−t)
EQ [e−mπ (t)ye 0
π ∗ (t) = B + e m −1]
,
γ 2 +m σ2 σ2
0  t  T, (3.12)

and the optimal value function is given by

1 −m[er0 (T−t) x−f1 (t)]

V (t, x) = − e , 0  t  T; (3.13)
m
θ
(ii) if D  γ B1 +m
, the robust optimal reinsurance-investment strategy is given by

⎧ −r0 (T−t)
⎪ θe 1 D(γ B1 + m)
⎪
⎨ γ B1 + m , 0  t  T + r0 ln
⎪
θ
,
∗
a (t) = (3.14)
⎪
⎪ D(γ B1 + m)
⎪
⎩ D, 1
T + ln < t  T,
r0 θ

 ∗ r (T−t)

∗ e−r0 (T−t) μ − r0 λ2 EQ [ye−mπ (t)ye 0 ] γJ
EQ [e−mπ
∗ (t)yer0 (T−t)
−1]
π (t) = B + e m ,
γ 2 +m σ2 σ2
0  t  T, (3.15)
154 D. LI ET AL.

and the optimal value function is given by

⎧
⎪
⎪ 1 r0 (T−t) x−f (t)] 1 D(γ B1 + m)
⎪ − e−m[e
⎨
2 , 0  t  T + ln ,
m r0 θ
V (t, x) = (3.16)
⎪
⎪ 1 −m[er0 (T−t) x−f (t)] D(γ B1 + m)
⎪
⎩− e 3
1
, T + ln < t  T,
m r0 θ

where
 T  T
λ1 (η − θ )μZ
f1 (t) = (1 − er0 (T−t) ) + l1 (ω)dω − l2 (ω)dω, (3.17)
r0 t t
 k  T
λ1 (η − θ )μZ r0 (T−k)
f2 (t) = (e − er0 (T−t) ) + l1 (ω)dω − l2 (ω)dω
r0 t t
λ1 ημZ λ1 σZ2 (γ B1 + m)
+ (1 − er0 (T−k) ) − (1 − e2r0 (T−k) ),
r0 4r0
1 D(γ B1 + m)
k = T + ln , (3.18)
r0 θ
 T
λ1 ημZ λ1 σZ2 (γ B1 + m)
f3 (t) = (1 − er0 (T−t) ) − (1 − e2r0 (T−t) ) − l2 (ω)dω, (3.19)
r0 4r0 t
 θe−r0 (T−ω)
r0 (T−ω) γ B1 +m
l1 (ω) = e [λ1 (γ B1 + m)ser0 (T−ω) − λ1 θ ]F̄(s)ds, (3.20)
0
1
l2 (ω) = π ∗ (ω)(μ − r0 )er0 (T − ω) − (γ B2 + m)σ 2 (π ∗ (ω))2 e2r0 (T−ω)
2 !
λ2 γ J Q −mπ ∗ (ω)yer0 (T−ω)
+ J 1 − e m E [e −1]
. (3.21)
γ

Proof: See Appendix 2.

Remark 3.4: From Equation (3.11), we find that the robust optimal reinsurance strategy decreases
w.r.t. the ambiguity aversion coefficient γ B1 , which provides the same insight as the intuition that an
AAI with a higher ambiguity aversion level is prone to purchasing more reinsurance. This property
is also shown in Yi et al. (2013). Analogously, we can derive the effects of ambiguity aversion
coefficients on the robust optimal investment strategy, and we further analyze the effects using
numerical examples. Equation (3.12) illustrates that the first term of the robust optimal investment
strategy is the speculative demand, which depends on the expected excess return. The second term
arises from the jump risk in the risky asset price process, and it is not present when the risky asset’s
price process has a continuous sample paths. In addition, we find that the robust optimal reinsurance
strategy is a function of the current time t, and is independent of the parameters of the risky asset,
while the robust optimal investment strategy is independent of parameters of the insurance business.
Special cases of our results in Theorem 3.1 can be found in the literature. If γ B1 = γ B2 = γ J = 0,
problem (2.7) reduces to problem (2.4), and the optimal excess-of-loss reinsurance strategy reduces
to that in Bai & Guo (2010). If λ2 = 0, there is no jump in the price process of the risky asset and
the optimal investment strategy is similar to that in Maenhout (2004), which considers the robust
portfolio selection maximizing a power utility.
Remark 3.5: (Proportional reinsurance case). If the AAI can purchase proportional reinsurance
or acquire new business (by acting as a reinsurer for other insurers, for example) to manage her
insurance business risk, the reinsurance level at any time t is associated with the value 1 − p(t), where
p(t) ∈ [0, +∞) can be regarded as the value of the risk exposure. Then, up := {(p(t), π(t))}t∈[0,T] is
 SCANDINAVIAN ACTUARIAL JOURNAL 155

Table 1. Optimal value functions with excess-of-loss and proportional reinsurance.

t 0 1 2 3 4 5 6 7
V(t, x) −0.1488 −0.1754 −0.2066 −0.2431 −0.2857 −0.3353 −0.3932 −0.4606
V p (t, x) −0.1673 −0.1946 −0.2262 −0.2627 −0.3049 −0.3537 −0.4099 −0.4749

the reinsurance-investment strategy, and the wealth process of the AAI under probability Q is
√
dX up (t) = r0 X up (t) + λ1 (η − θ + θ p(t))μZ + π(t)(μ
 ∞− r0 ) + λ1 p(t)σZ φ1 (t) + π(t)σ φ2 (t) dt
√
+ λ1 p(t)σZ dB1Q (t) + π(t)σ dB2Q (t) + π(t)yN Q (dt, dy).
−1
(3.22)
Similarly, we can obtain the robust optimal proportional reinsurance-investment strategy u∗p :=
{(p∗ (t), π ∗ (t))}t∈[0,T] as

θ μZ e−r0 (T−t)
p∗ (t) = , 0  t  T, (3.23)
(γ B1 + m)σZ2
 ∗ r (T−t)

∗ e−r0 (T−t) μ − r0 λ2 EQ [ye−mπ (t)ye 0 ] γ J EQ [e−mπ ∗ (t)yer0 (T−t) −1]
π (t) = B + em , 0  t  T,
γ 2 +m σ2 σ2
(3.24)

and the optimal value function V p (t, x) as

1 −m[er0 (T−t) x−f4 (t)]

V p (t, x) = − e , 0  t  T, (3.25)
m

where

λ1 (η − θ )μZ λ1 θ 2 μ2Z T
f4 (t) = (1 − er0 (T−t) ) − (T − t) − l2 (ω)dω (3.26)
r0 2(γ B1 + m)σZ2 t

with l2 (ω) given by Equation (3.21).

In addition, from Equations (3.23) and (3.24), we find that the effects of the model parameters on
the robust optimal proportional reinsurance-investment strategy are similar to those in an excess-
of-loss reinsurance case. However, according to the results of the optimal value functions with
excess-of-loss and proportional reinsurance given in Table 1 with the values of the parameters given
in Table 2, we find that V (t, x)  V p (t, x), which implies that excess-of-loss reinsurance is preferred
than proportional reinsurance in our model.

4. Special cases
This section provides some special cases of our model: investment-only, ANI and no jump. We can
derive the corresponding robust optimal strategies and optimal value functions for these cases similar
to the expression of Theorem 3.1, however, to make the robust optimal strategies simple and intuitive,
we consider these special cases under condition (3.10).
156 D. LI ET AL.

4.1. Investment-only case

If there is no reinsurance in our model, i.e. a(t) ≡ D, the wealth process of an AAI under the
probability Q reduces to
√
dX u (t) = r0 X u (t) + λ1 ημZ + π(t)(μ − r0 ) + λ1 σZ φ1 (t) + π(t)σ φ2 (t) dt
√ ∞
(4.1)
+ λ1 σZ dB1Q (t) + π(t)σ dB2Q (t) + π(t)yN Q (dt, dy),
−1

and the corresponding HJB equation becomes

 
sup inf Ṽt + Ṽx r0 x + λ1 ημZ
+
π ∈R (φ1 ,φ2 ,φ3 )∈R×R×R 
√
+π(μ − r0 ) + λ1 σZ φ1 + π σ φ2
1 (4.2)
+ Ṽxx [λ1 σZ2 + π 2 σ 2 ] + λ2 φ3 EQ [Ṽ (t, x + π y) − Ṽ (t, x)]
2 
φ12 φ22 λ2 (φ3 ln φ3 − φ3 + 1)
+ B + + = 0,
2ϕ 1 (t) 2ϕ B2 (t) ϕ J (t)

where Ṽ is a short notation for Ṽ (t, x) representing the optimal value function of the investment-only
problem with the boundary condition Ṽ (T, x) = U(x). Similar to the derivations of the reinsurance-
investment case, we have the robust optimal investment strategy and the corresponding optimal value
function as follows.
Proposition 4.1: For the investment-only problem, i.e. a(t) ≡ D, under assumptions (3.1) and (3.2),
the robust optimal investment strategy is given by
 ∗ r (T−t)

∗ e−r0 (T−t) μ − r0 λ2 EQ [ye−mπ (t)ye 0 ] γJ
EQ [e−mπ
∗ (t)yer0 (T−t)
−1]
π (t) = B + e m , 0  t  T,
γ 2 +m σ2 σ2
(4.3)
and the optimal value function is given by

1 −m[er0 (T−t) x−f3 (t)]

Ṽ (t, x) = − e , 0  t  T, (4.4)
m
where f3 (t) is given by Equation (3.19).
In addition, if we do not consider the insurance business, i.e. λ1 = 0, then the robust optimal
investment strategy is given by Equation (4.3), and the optimal value function is

1 −m[er0 (T−t) x−f5 (t)]

V̆ (t, x) = − e , 0  t  T, (4.5)
m
with
 T
f5 (t) = − l2 (ω)dω, (4.6)
t

where l2 (ω) is given by Equation (3.21).

4.2. ANI case

If all of the ambiguity aversion coefficients equal 0, i.e, γ B1 = γ B2 = γ J = 0, our model reduces to an
optimization problem for an ANI. Then, the ANI’s wealth process under probability P is described
 SCANDINAVIAN ACTUARIAL JOURNAL 157

by Equation (2.3) and the objective function is given by Equation (2.4). Denote the optimal value
function by

V̂ (t, x) = sup E[U(X û (T))], (4.7)

û∈ ˜

where û = {(â(t), π̂ (t))}t∈[0,T] . The corresponding HJB equation is


 
sup V̂t + V̂x r0 x + λ1 (θ μ̄(â) + (η − θ )μZ ) + π̂ (μ − r0 )
(â,π̂)∈[0,D]×R  (4.8)
1
+ V̂xx [λ1 (σ̄ (â)) + π̂ σ ] + λ2 E[V̂ (t, x + π̂y) − V̂ (t, x)] = 0,
2 2 2
2

where V̂ is a short notation for V̂ (t, x) with V̂ (T, x) = U(x).

Proposition 4.2: For problem (4.7) of an ANI who ignores ambiguity with utility (3.1),
θ
(i) if D > m, the optimal reinsurance-investment strategy is given by

θ e−r0 (T−t)
â∗ (t) = , 0  t  T, (4.9)
m

 ∗ r (T−t)

∗ e−r0 (T−t) μ − r0 λ2 E[ye−mπ̂ (t)ye 0 ]
π̂ (t) = + , 0  t  T, (4.10)
m σ2 σ2

and the optimal value function V̂ (t, x) is given by

1 −m[er0 (T−t) x−f6 (t)]

V̂ (t, x) = − e , 0  t  T; (4.11)
m
θ
(ii) if D  m, the optimal reinsurance-investment strategy is given by
⎧ −r (T−t)
⎪
⎪ θe 0 1 Dm
⎪
⎨ , 0  t  T + ln ,
m r θ
â∗ (t) =
0
(4.12)
⎪
⎪ 1 Dm
⎪
⎩ D, T + ln < t  T,
r0 θ

 ∗ r (T−t)

e−r0 (T−t) μ − r0 λ2 E[ye−mπ̂ (t)ye 0 ]
π̂ ∗ (t) = + , 0  t  T, (4.13)
m σ2 σ2

and the optimal value function V̂ (t, x) is given by

⎧ 1 1 Dm
⎪ −m[er0 (T−t) x−f7 (t)] ,
⎨−me
⎪ 0tT+
r0
ln
θ
,
V̂ (t, x) = (4.14)
⎪ 1
⎪ 1 Dm
⎩ − e−m[er0 (T−t) x−f8 (t)] , T + ln < t  T,
m r0 θ
158 D. LI ET AL.

where
 T  T
λ1 (η − θ )μZ
f6 (t) = (1 − er0 (T−t) ) + l3 (ω)dω − l4 (ω)dω, (4.15)
r0 t t
 k1  T
λ1 (η − θ )μZ r0 (T−k1 )
f7 (t) = (e − er0 (T−t) ) + l3 (ω)dω − l4 (ω)dω
r0 t t
λ1 ημZ λ1 σZ2 m
+ (1 − er0 (T−k1 ) ) − (1 − e2r0 (T−k1 ) ),
r0 4r0
1 Dm
k1 = T + ln , (4.16)
r0 θ
 T
λ1 ημZ λ1 σZ2 m
f8 (t) = (1 − er0 (T−t) ) − (1 − e2r0 (T−t) ) − l4 (ω)dω, (4.17)
r0 4r0 t
 θe−r0 (T−ω)
m
r0 (T−ω)
l3 (ω) = e [λ1 mser0 (T−ω) − λ1 θ ]F̄(s)ds, (4.18)
0
1
l4 (ω) = π̂ ∗ (ω)(μ − r0 )er0 (T−ω) − mσ 2 (π̂ ∗ (ω))2 e2r0 (T−ω)
2
λ2 ∗ r0 (T−ω)
− E[e−mπ̂ (ω)ye − 1]. (4.19)
m
Proposition 4.2 shows that the optimal reinsurance strategy in Equation (4.9) becomes the result
in Bai & Guo (2010), i.e. our model extends the optimal excess-of-loss reinsurance strategy in Bai &
Guo (2010) to the case of robust optimal formulation.
4.3. No jump case
If the intensity of the jump in the price process of the risky asset equals 0, i.e. λ2 = 0, our model
reduces to a case without jumps. Then, the wealth process of the AAI under probability Q with the
strategy ū = {(ā(t), π̄ (t))}t∈[0,T] is
√
dX ū (t) = r0 X ū (t) + λ1 [θ μ̄(ā) + (η − θ )μZ ] + π̄ (t)(μ − r0 ) + λ1 σ̄ (ā)φ1 (t) + π̄ (t)σ φ2 (t) dt
√
+ λ1 σ̄ (ā)dB1Q (t) + π̄ (t)σ dB2Q (t),
(4.20)
and the corresponding HJB equation is
 
sup inf V̄t + V̄x r0 x + λ1 (θ μ̄(ā) + (η − θ )μZ ) + π̄ (μ − r0 )
(ā,π̄ )∈[0,D]×R (φ1 ,φ2 )∈R×R 
√
+ λ1 σ̄ (ā)φ1 + π̄ σ φ2 (4.21)

1 φ 2 φ 2
+ V̄xx [λ1 (σ̄ (ā))2 + π̄ 2 σ 2 ] + B 1 + B 2 = 0,
2 2ϕ 1 (t) 2ϕ 2 (t)

where V̄ is a short notation for V̄ (t, x) representing the optimal value function of the no jump
case with the boundary condition V̄ (T, x) = U(x). Similarly, we can derive the robust optimal
reinsurance-investment strategy and the corresponding optimal value function, explicitly.
Proposition 4.3: If the risky asset price does not have jumps, under assumptions (3.1) and (3.2), the
robust optimal reinsurance-investment strategy and the optimal value function reduce to the follows:
θ
(1) if D > γ B1 +m
, the robust optimal reinsurance-investment strategy is given by

θ e−r0 (T−t)
ā∗ (t) = , 0  t  T, (4.22)
γ B1 + m
 SCANDINAVIAN ACTUARIAL JOURNAL 159

(μ − r0 )e−r0 (T−t)
π̄ ∗ (t) = , 0  t  T, (4.23)
σ 2 (γ B2 + m)
and the optimal value function is given by

1 −m[er0 (T−t) x−f9 (t)]

V̄ (t, x) = − e , 0  t  T; (4.24)
m
θ
(2) if D  γ B1 +m
, the robust optimal reinsurance-investment strategy is given by
⎧ −r0 (T−t)
⎨ θe
⎪
, 0  t  k,
∗
ā (t) = γ B1 + m (4.25)
⎪
⎩
D, k < t  T,

(μ − r0 )e−r0 (T−t)
π̄ ∗ (t) = , 0  t  T, (4.26)
σ 2 (γ B2 + m)
and the optimal value function is given by
⎧
⎪ 1 −m[er0 (T−t) x−f10 (t)]
⎪
⎨−me , 0  t  k,
V̄ (t, x) = (4.27)
⎪
⎪
⎩ − 1 e−m[er0 (T−t) x−f11 (t)] , k < t  T,
m
with
 T  T
λ1 (η − θ )μZ
f9 (t) = (1 − er0 (T−t) ) + l1 (ω)dω − l5 (ω)dω, (4.28)
r0 t t
 k  T
λ1 (η − θ )μZ r0 (T−k)
f10 (t) = (e − er0 (T−t) ) + l1 (ω)dω − l5 (ω)dω
r0 t t
λ1 ημZ λ1 σZ2 (γ B1 + m)
+ (1 − er0 (T−k) ) − (1 − e2r0 (T−k) ), (4.29)
r0 4r0
 T
λ1 ημZ r0 (T−t) λ1 σZ2 (γ B1 + m) 2r0 (T−t)
f11 (t) = (1 − e )− (1 − e )− l5 (ω)dω, (4.30)
r0 4r0 t
1
l5 (ω) = π̄ ∗ (ω)(μ − r0 )er0 (T−ω) − (γ B2 + m)σ 2 (π̄ ∗ (ω))2 e2r0 (T−ω) , (4.31)
2
where l1 (ω) is given in Equation (3.20).
Proposition 4.3 illustrates that if there is no jump in our model, the robust optimal investment
strategy is similar to that in Maenhout (2004), which considers the robust portfolio selection max-
imizing a power utility. Furthermore, if we do not consider the insurance business and jumps, i.e.
λ1 = λ2 = 0, then the robust optimal investment strategy is given by Equation (4.23), and the optimal
value function is given by

1 −m[er0 (T−t) x−f12 (t)]

V́ (t, x) = − e , 0  t  T, (4.32)
m
with  T
f12 (t) = − l5 (ω)dω (4.33)
t
where l5 (ω) is given by Equation (4.31).
160 D. LI ET AL.

Table 2. Parameters of model in numerical examples.

r0 μ σ η θ m γ B1 γ B2 γJ λ1 λ2 λZ λY T t
0.03 0.08 0.25 0.10 0.30 0.2 0.2 0.5 0.7 1 2 2 2 10 0

Table 3. Partial derivatives of a∗ (t) in the case of D > θ .

γ B1 +m

∂a∗ (t) ∂a∗ (t) ∂a∗ (t) ∂a∗ (t) ∂a∗ (t)
derivatives ∂t ∂r0 ∂m ∂θ ∂γ B1
value >0 <0 <0 >0 <0

5. Numerical examples and utility enhancements

This section provides some numerical examples to illustrate the effects of model parameters on the
robust optimal reinsurance-investment strategy and utility enhancements under condition (3.10).
We take the case D > γ B1θ+m as an example, and the analysis for the case of D  γ B1θ+m is similar.
Moreover, suppose that both claim size Zi and jump size Yi follow exponential distributions with
parameters λZ and λY , respectively, i.e. the density functions of claim size Zi and jump size Yi are
f (z) = λZ e−λZ z , z  0 and g(y) = λY e−λY (y+1) , y  −1. Throughout numerical analysis, unless
otherwise stated, the basic parameters are given in the following table.

5.1. Effects of model parameters on the robust optimal reinsurance-investment strategy

From Equation (3.11), we can obtain partial derivatives of the robust optimal reinsurance strategy
a∗ (t) w.r.t. different parameters. Table 3 shows that: (i) a∗ (t) increases as a function of t and the
reinsurer’s safety loading θ . As the remaining time decreases, the AAI becomes less risk averse and
undertakes more risks by herself, whereas when θ increases, to decrease the expensive payment
to reinsurance, the AAI prefers to take more insurance business and raise the retention level
of reinsurance; (ii) a∗ (t) decreases w.r.t. the risk-free interest rate r0 , the insurer’s risk aversion
coefficient m and the ambiguity aversion coefficient γ B1 . The main reason for this is that with the
increase of r0 , the risk-free asset is more attractive, so the AAI is more likely to invest more wealth
in the risk-free asset instead of purchasing more reinsurance. Note that the larger m is, the more risk
averse the AAI is, so she purchases more reinsurance to spread risk. Moreover, an AAI with a higher
ambiguity aversion level γ B1 is prone to purchasing more reinsurance to disperse the underlying
insurance business risks.
Equation (3.12) reveals that the robust optimal investment strategy π ∗ (t) is implicit and depends
on different parameters in a very complicated way. It is difficult to analyze the effects of different
parameters on π ∗ (t) analytically through the partial derivatives of π ∗ (t). We perform the sensitivity
analysis of π ∗ (t) w.r.t. different parameters through numerical simulations.
Figure 1 shows the effects of the ambiguity aversion coefficients γ B2 and γ J , the jump intensity
of the risky asset λ2 and the parameter λY of distribution function G(y) on the robust optimal
investment strategy π ∗ (t). We find that π ∗ (t) is a decreasing function of γ B2 and γ J ; that is, the
more ambiguity aversion the AAI is, the less risky asset she purchases. In addition, the robust optimal
investment strategy π ∗ (t) is more sensitive to γ B2 and γ J when γ B2 and γ J are small than that
when they are large, which means that the marginal effect of increasing the amount invested in the
risky asset is declining. The robust optimal investment strategy π ∗ (t) also decreases w.r.t. λ2 and λY ,
because that as the jump intensity λ2 increases, the risky asset becomes more risky and less attractive.
Moreover, a larger λY implies that the expectation and variance of Yi decrease, prompting the AAI
to invest more wealth in the risky asset.
 SCANDINAVIAN ACTUARIAL JOURNAL 161

Figure 1. Effects of parameters γ B2 , γ J , λ2 and λY on π ∗ (t).

5.2. Effects of model parameters on the utility enhancements

This subsection discusses the AAI’s utility enhancements by numerical illustration. Without loss of
θ
generality, we consider the case of D > m > γ B1θ+m , and the analyzes of other cases are similar.
First, we study the effect of ambiguity aversion on utility enhancement. We show that an insurer
who suffers from ambiguity aversion follows a suboptimal strategy. Suppose that an AAI does not
take the optimal strategy u∗ = {(a∗ (t), π ∗ (t))}t∈[0,T] given in Theorem 3.1, but instead makes her
decisions as if she were an ANI, i.e. the AAI follows the strategy û∗ = {(â∗ (t), π̂ ∗ (t))}t∈[0,T] given in
Proposition 4.2. The optimal value function for the AAI following strategy û∗ is defined by

V̌ (t, x)  
 
(φ1 (s))2
T (φ2 (s))2 λ2 (φ3 (s) ln φ3 (s) − φ3 (s) + 1) ∗
= inf EQ + + ds + U(X (T)) .
û
∈ t 2ϕ B1 (s) 2ϕ B2 (s) ϕ J (s)
(5.1)
Note that φ1 (t), φ2 (t) and φ3 (t), which describe the alternative model, are determined endogenously
and depend on the reinsurance-investment strategy. Unlike in the optimal case, the reinsurance-
investment strategy is now pre-specified. The functions ϕ B1 (t), ϕ B2 (t) and ϕ J (t) are defined analo-
gously to Equation (3.2).
It is obvious that value function V̌ (t, x) defined in Equation (5.1) is smaller than V (t, x) in Equation
(2.9). Bases on V̌ (t, x), we define the utility enhancement from considering ambiguity as follows

V (t, x)
UE1 (t) := 1 − = 1 − em(f1 (t)−f0 (t)) , (5.2)
V̌ (t, x)

where f1 (t) and f0 (t) are given by Equation (3.17) and

 T  T
λ1 (η − θ )μZ r0 (T−t)
f0 (t) = (1 − e )+ l6 (ω)dω − l7 (ω)dω, (5.3)
r0 t t
 θe−r0 (T−ω)
m
r0 (T−ω)
l6 (ω) = e [λ1 (γ B1 + m)ser0 (T−ω) − λ1 θ ]F̄(s)ds, (5.4)
0
1
l7 (ω) = π̂ ∗ (ω)(μ − r0 )er0 (T−ω) − (γ B2 + m)σ 2 (π̂ ∗ (ω))2 e2r0 .T−!/
2
162 D. LI ET AL.

Table 4. Effects of γ B2 and γ J on the utility enhancement UE1 (t).

γ B2 γ J 0 1 2 3 4 5 6 7
0 0.0775 0.1153 0.1932 0.2719 0.3370 0.3876 0.4262 0.4558
1 0.3028 0.3912 0.5021 0.5786 0.6312 0.6678 0.6941 0.7134
2 0.5289 0.6387 0.7199 0.7704 0.8030 0.8250 0.8404 0.8515
3 0.6867 0.7981 0.8484 0.8781 0.8967 0.9091 0.9176 0.9237
4 0.7928 0.8907 0.9196 0.9362 0.9464 0.9531 0.9577 0.9610
5 0.8633 0.9420 0.9579 0.9669 0.9724 0.9760 0.9784 0.9801
6 0.9099 0.9696 0.9782 0.9830 0.9859 0.9877 0.9890 0.9899
7 0.9407 0.9842 0.9888 0.9913 0.9928 0.9938 0.9944 0.9949

Figure 2. Effects of T − t, γ B1 , γ B2 and γ J on the utility enhancement UE1 (t).

!
λ2 γJ
EQ [e−mπ̂
∗ (ω)yer0 (T−ω)
−1]
+ J 1−e m . (5.5)
γ

Derivation of suboptimal value function are given in Appendix C.

Table 4 shows the effects of ambiguity aversion coefficients for diffusion and jump risks on the
utility enhancement UE1 (t). We find that UE1 (t) increases w.r.t γ B2 and γ J , and the change trends
are also clearly showed in Figure 2. Particularly, the first column (γ J = 0) is the case of ambiguity
without jumps. We find that the value of UE1 (t) with ambiguity for jump-diffusion risks (γ J = 0)
is much larger than that without jumps. Thus, the utility enhancement from ambiguity for the jump
risk cannot be ignored.
Figure 2 discloses the utility enhancement UE1 (t) as an increasing function of γ B1 , γ B2 , γ J and the
remaining time T − t. UE1 (t) is higher for the AAI with less information about the model P (higher
γ B1 , γ B2 and γ J ) than that for the AAI with more information (smaller γ B1 , γ B2 and γ J ). Moreover,
UE1 (t) has a remarkable upward trend as T − t increases. Figure 3 reveals that UE1 (t) increases w.r.t.
λ2 and λY . A larger λ2 or λY implies more uncertainties for the price of the risky asset, such that
taking robust optimal strategy for an AAI brings larger utility enhancement.
Next, we analyze the utility enhancement from considering reinsurance. As stated in some
researches about optimal reinsurance-investment problem, reinsurance can optimize the utility. This
is also true in the optimal reinsurance-investment problem with ambiguity. Bases on Ṽ (t, x) (the
optimal value function without reinsurance), we define the utility enhancement from considering
reinsurance as follows
 SCANDINAVIAN ACTUARIAL JOURNAL 163

Figure 3. Effects of λ2 and λY on the utility enhancement UE1 (t).

Figure 4. Effects of T − t, γ B1 , θ and m on the utility enhancement UE2 (t).

V (t, x)
UE2 (t) := 1 − = 1 − em(f1 (t)−f3 (t)) , (5.6)
Ṽ (t, x)
where f1 (t) and f3 (t) are given in Equations (3.17) and (3.19).
Figure 4 illustrates the effects of the remaining time T − t, the ambiguity aversion coefficient γ B1 ,
the reinsurer’s safety loading θ and the insurer’s risk aversion coefficient m on the utility enhancement
UE2 (t). From Figure 4, we find that UE2 (t) increases w.r.t. T − t. A reasonable explanation is that
the AAI faces more uncertainty when T − t is longer. In addition, UE2 (t) is increasing w.r.t. γ B1 ,
which means that the AAI with a higher ambiguity aversion level is prone to purchasing more
reinsurance to disperse the insurance business risk. Then, the utility enhancement from considering
reinsurance increases. As shown in Figure 4, UE2 (t) decreases w.r.t. θ , which can be explained by
the fact that a higher θ increases the cost of the AAI’s purchase of reinsurance, so she cedes less risk
to the reinsurer. Intuitively, an extremely high reinsurance premium provides a favored position for
acquiring business, causing utility enhancement if new business acquisition is prohibited. Moreover,
Figure 4 shows that UE2 (t) increases w.r.t. m, which indicates that the AAI with a higher risk aversion
coefficient is more likely to purchase more reinsurance for risk-spreading.
164 D. LI ET AL.

Figure 5. Effects of λ1 and λZ on the utility enhancement UE2 (t).

As shown in Figure 5, the utility enhancement UE2 (t) increases w.r.t. the jump intensity of claim
size λ1 . A possible reason for this is that a higher λ1 results in a more severe fluctuation in the AAI’s
surplus process, making her more likely to cede more reinsurance to the reinsurer for dispersing
risks. Consequently, the utility enhancement from considering reinsurance increases. In addition,
from Figure 5, we find that the effect of λz on UE2 (t) is small.

6. Conclusion
In this paper, we consider a robust optimal excess-of-loss reinsurance and investment problem with
jumps for an AAI, who worries about ambiguity and aims to develop robust optimal strategies. The
insurer’s surplus is assumed to follow a Brownian motion with drift, and the insurer is allowed
to purchase excess-of-loss reinsurance and invest her surplus in a risk-free asset and a risky asset
whose price dynamics is described by a jump-diffusion model. Meanwhile, the insurer may lack full
confidence on the model describing the economy, in which case we formulate a systematic analysis
of the robust reinsurance-investment problem. By applying the stochastic dynamic programming
approach, explicit expressions for the robust optimal reinsurance-investment strategy to maximize
the expected exponential utility of terminal wealth and the corresponding optimal value function are
obtained. Some special cases of our model and results are provided, and the economic implications
of our findings and utility enhancements from considering ambiguity and reinsurance are analyzed
using numerical examples. We find that (i) the effect of ambiguity by jump risk on the robust optimal
investment strategy and the optimal value function is significant, and utility enhancement from
considering ambiguity in the case of jump-diffusion risks is much larger than that without jumps; (ii)
the robust optimal reinsurance-investment strategy for the AAI is affected by the attitude towards
ambiguity, so the AAI facing model uncertainty has a smaller optimal strategy than an ANI; (iii)
reinsurance brings large utility enhancement for the AAI, which implies that reinsurance is very
important in risk management; and (iv) for the robust optimal reinsurance-investment problem,
the optimal value function with excess-of-loss reinsurance is preferred than that with proportional
reinsurance.
In future research, more complicated models, such as stochastic volatility with jumps, can be taken
into account, although doing so may make it difficult to obtain the closed-form solution. Thus, other
methods, such as asymptotic, or other practical methods may be introduced to deal with the robust
optimal reinsurance-investment problem.
 SCANDINAVIAN ACTUARIAL JOURNAL 165

Disclosure statement
No potential conflict of interest was reported by the authors.

Funding
This research was supported by the National Natural Science Foundation of China [grant number 71571195]; Research
Grants Council of the Hong Kong Special Administrative Region [grant number HKU 17330816]; For Ying Tung
Eduction Foundation for Young Teachers in the Higher Education Institutions of China [grant number 151081];
Guangdong Natural Science Funds for Distinguished Young Scholar [grant number 2015A030306040]; Natural Science
Foundation of Guangdong Province of China [grant number 2014A030310195]; Science and Technology Planning
Project of Guangdong Province [grant number 2016A070705024]. The authors are also very grateful to Jieming Zhou
and Pei Wang for their helpful suggestions.

References
Aït-Sahalia, Y. & Matthys, F. H. A., (2014). Robust portfolio optimization with jumps, Working Paper, http://scholar.
princeton.edu/sites/default/files/fmatthys/files/robustpfoptwithjumps_main_v7.pdf
Anderson, E. W., Hansen, L. P. & Sargent, T. J., (1999). Robustness detection and the price of risk, Working Paper.
https://files.nyu.edu/ts43/public/research/.svn/text-base/ahs3.pdf.svn-base
Anderson, E. W., Hansen, L. P. & Sargent, T. J. (2003). A quartet of semi-groups for model specification, robustness,
prices of risk, and model detection. Journal of the European Economic Association 1(1), 68–123.
Asmussen, S., Højgaard, B. & Taksar, M. (2000). Optimal risk control and dividend distribution policies: example of
excess-of-loss reinsurance for an insurance corporation. Finance and Stochastics 4(3), 299–324.
Bai, L. & Guo, J. (2010). Optimal dynamic excess-of-loss reinsurance and multidimensional portfolio selection. Science
China Mathematics 53(7), 1787–1804.
Bi, J., Meng, Q. & Zhang, Y. (2013). Dynamic mean-variance and optimal reinsurance problems under the no-
bankruptcy constraint for an insurer. Annals of Operations Research 212(1), 43–59.
Branger, N. & Larsen, L. S. (2013). Robust portfolio choice with uncertainty about jump and diffusion risk. Journal of
Banking and Finance 37(12), 5036–5047.
Fleming, W. & Soner, M. (2006). Controlled Markov Processes and Viscosity Solutions, 2nd ed., New York: Springer-
Verlag.
Flor, C. R. & Larsen, L. S. (2014). Robust portfolio choice with stochastic interest rates. Annals of Finance 10(2),
243–265.
Gerber, H. (1979). An introduction to mathematical risk theory. Series No. 8 of S.S. Huebner Foundation Monograph,
Homewood, IL: Irwin.
Grandell, J. (1991). Aspects of Risk Theory. New York: Springer-Verlag.
Gu, A., Guo, X., Li, Z. & Zeng, Y. (2012). Optimal control of excess-of-loss reinsurance and investment for insurers
under a CEV model. Insurance: Mathematics and Economics 51(3), 674–684.
Jang, B. G. & Kim, K. T. (2015). Optimal reinsurance and asset allocation under regime switching. Journal of Banking
and Finance 56, 37–47.
Karatzas, I. & Shreve, S. (1988). Brownian Motion and Stochastic Calculus. New York: Springer-Verlag.
Korn, R., Menkens, O. & Steffensen, M. (2012). Worst-case-optimal dynamic reinsurance for large claims. European
Actuarial Journal 2(1), 21–48.
Kraft, H. (2004). Optimal Porfolios with Stochastic Interest Rates and Defaultable Assets. Berlin: Springer-Verlag.
Liang, Z. & Yuen, K. C. (2016). Optimal dynamic reinsurance with dependent risks: variance premium principle.
Scandinavian Actuarial Journal 2016(1), 18–36.
Lin, X. & Li, Y. (2011). Optimal reinsurance and investment for a jump diffusion risk process under the CEV model.
North American Actuarial Journal 15(3), 417–431.
Maenhout, P. J. (2004). Robust portfolio rules and asset pricing. Review of Financial Studies 17(4), 951–983.
Maenhout, P. J. (2006). Robust portfolio rules and detection-error probabilities for a mean-reverting risk premium.
Journal of Economic Theory 128(1), 136–163.
Mataramvura, S. & Øksendal, B. (2008). Risk minimizing and HJBI equations for stochastic differential games. Stochastic
80, 317–337.
Øksendal, B. & Sulem, A. (2007). Applied stochastic control of jump diffusions, 2nd ed., Berlin: Springer.
Pressacco, F., Serafini, P. & Ziani, L. (2011). Mean-variance efficient strategies in proportional reinsurance under group
correlation in a gaussian framework. European Actuarial Journal 1(2), 433–454.
Pun, C. S. & Wong, H. Y. (2015). Robust investment-reinsurance optimization with multiscale stochastic volatility.
Insurance: Mathematics and Economics 62, 245–256.
Schmidli, H. (2002). On minimizing the ruin probability by investment and reinsurance. Annals of Applied Probability
12(3), 890–907.
166 D. LI ET AL.

Uppal, R. & Wang, T. (2003). Model misspecification and underdiversification. The Journal of Finance 58(6), 2465–
2486.
Yang, H. & Zhang, L. (2005). Optimal investment for insurer with jump-diffusion risk process. Insurance: Mathematics
and Economics 37(3), 615–634.
Yi, B., Li, Z., Viens, F. G. & Zeng, Y. (2013). Robust optimal control for an insurer with reinsurance and investment
under Heston’s stochastic volatility model. Insurance: Mathematics and Economics 53(3), 601–614.
Yi, B., Li, Z., Viens, F. G. & Zeng, Y. (2015). Robust optimal strategies for an insurer with reinsurance and investment
under benchmark and mean-variance criteria. Scandinavian Actuarial Journal 2015(8), 725–751.
Zeng, Y., Li, D. & Gu, A. (2016). Robust equilibrium reinsurance-investment strategy for a mean-variance insurer in a
model with jumps. Insurance: Mathematics and Economics 66, 138–152.
Zhao, H., Rong, X. & Zhao, Y. (2013). Optimal excess-of-loss reinsurance and investment problem for an insurer with
jump-diffusion risk process under the Heston model. Insurance: Mathematics and Economics 53(3), 504–514.
Zheng, X., Zhou, J. & Sun, Z. (2016). Robust optimal portfolio and proportional reinsurance for an insurer under a
CEV model. Insurance: Mathematics and Economics 67, 77–87.

Appendix 1.
A.1. Derivation of relative entropy
The relative entropy is defined as the expectation under the alternative probability of the log Radon-Nikodym derivative
defined in Equation (2.5). Using Itô’s formula, we have

d ln = φ1 (t)dB1 (t) + φ2 (t)dB2 (t) + λ2 (1 − φ3 (t))dt (A1)

 ∞
1 1
− (φ1 (t))2 dt − (φ2 (t))2 dt + ln φ3 (t)N(dt, dy). (A2)
2 2 −1

The relative entropy over the interval from t to t + ε is given by

   t+ε  t+ε
(t + ε)
EQ ln = EQ φ1 (s)(dB1Q (s) + φ1 (s)ds) + φ2 (s)(dB2Q (s) + φ2 (s)ds) (A3)
(t) t t
 t+ε 
1 1
+ [λ2 (1 − φ3 (s)) − (φ1 (s))2 − (φ2 (s))2 ]ds (A4)
t 2 2
 t+ε  ∞  t+ε
+ ln φ3 (s)ν(ds, dy) + λ2 φ3 (s) ln φ3 (s)ds (A5)
t −1 t
 t+ε !
1 1
= EQ (φ1 (s))2 + (φ2 (s))2 + λ2 (φ3 (s) ln φ3 (s) − φ3 (s) + 1) . (A6)
t 2 2

Let ε → 0 and we have the continuous-time limit of the relative entropy given by Equation (2.8).

Appendix 2.
B.1. Proof of Theorem 3.1 and Remark 3.3
To solve Equation (2.10), we conjecture that the optimal value function is

1 −m[er0 (T−t) x−f¯ (t)]

V (t, x) = − e (B1)
m

with the boundary condition f¯ (T) = 0. A direct calculation yields

Vt = −Vm[−r0 xer0 (T−t) − f¯t ], Vx = −Vmer0 (T−t) , Vxx = Vm2 e2r0 (T−t) ,
r (T−t) (B2)
V (t, x + π y) − V (t, x) = V (e−mπ ye 0 − 1),

where f¯ = f¯ (t) and f¯t stands for the partial derivative of f¯ (t) w.r.t. time t for short.
 SCANDINAVIAN ACTUARIAL JOURNAL 167

According to the first-order optimality conditions, the functions φ1∗ (t), φ2∗ (t) and φ3∗ (t), which realize the infimum
part of Equation (2.10) are given by
$
φ1∗ (t) = −γ B1 λ1 σ̄ (a)er0 (T−t) , (B3)
φ2∗ (t) = −γ B2 π σ er0 (T−t) , (B4)
γJ Q [e−mπyer0 (T−t) −1]
φ3∗ (t) = e m E . (B5)

Substituting Equations (B2)–(B5) into Equation (2.10), we have


λ1 (σ̄ (a))2 γ B1 Vx2 σ 2 π 2 γ B2 Vx2
sup Vt + r0 xVx + λ1 (θ μ̄(a) + (η − θ)μZ )Vx + π(μ − r0 )Vx + +
(a,π )∈[0,D]×R 2mV 2mV
! (B6)
1 1 2 2 mV λ2 γ J Q −mπyer0 (T−t)
[e −1]
+ λ1 (σ̄ (a)) Vxx + π σ Vxx −
2 1−em E
= 0.
2 2 γJ

The first-order optimality condition gives the optimal reinsurance strategy

dμ̄(a) λ1 ϕ B1 (t)Vx2 d(σ̄ (a))2 λ1 Vxx d(σ̄ (a))2

λ1 θVx − · + · = 0,
da 2 da 2 da

i.e.
F̄(a0∗ )(θVx − ϕ B1 (t)Vx2 a0∗ + Vxx a0∗ ) = 0.

Since 0 < F̄(a0∗ )  1, we have

θVx θe−r0 (T−t)

a0∗ (t) = = B . (B7)
ϕ B1 (t)Vx2 − Vxx γ 1 +m

Furthermore, differentiating Equation (B6) w.r.t. π(t) implies that a nonlinear equation for the robust optimal
investment strategy π ∗ (t) is
 ∗ r (T−t)

∗ e−r0 (T−t) μ − r0 λ2 EQ [ye−mπ (t)ye 0 ] γJ
EQ [e−mπ
∗ (t)yer0 (T−t)
−1]
π (t) = B + e m . (B8)
γ 2 +m σ2 σ2

We first justify that π ∗ given in Equation (B8) derived by the first-order conditions is the optimal investment
strategy. Let

σ 2 π 2 γ B2 Vx2 1 mV λ2 γ J EQ [e−mπyer0 (T−t) −1]

g(π ) = π(μ − r0 )Vx + + π 2 σ 2 Vxx + em ,
2mV 2 γJ

which gather the term of π in the left side of Equation (B6). Furthermore,

σ 2 π γ B2 Vx2 γ J Q −mπyer0 (T−t) r (T−t)

gπ (π ) = (μ − r0 )Vx + + π σ 2 Vxx − mV λ2 er0 (T−t) e m E [e −1] Q
E [ye−mπ ye 0 ],
mV " #
σ 2 γ B2 Vx2 γ J Q −mπyer0 (T−t) r (T−t) 2
gπ π (π ) = + σ 2 Vxx + mγ J V λ2 e2r0 (T−t) e m E [e −1]
EQ [ye−mπ ye 0 ]
mV
γ J Q −mπyer0 (T−t) r0 (T−t)
(T−t) [e −1] Q −mπ
+ m V λ2 e 0
2 2r em E
E [ye ye ].

Since V < 0, Vx > 0 and Vxx < 0, it is obvious that gπ π (π ) < 0 for any admissible π . Therefore, the first-order
optimality condition gives the optimal investment strategy.
Next, similarly, let
λ1 (σ̄ (a))2 γ B1 Vx2 1
l(a) = λ1 θ μ̄(a)Vx + + λ1 (σ̄ (a))2 Vxx ,
2mV 2
which gathers the term of a in the left side of Equation (B6). Since l(a) is a continuous function of a (a ∈ [0, D]), the
−r0 (T−t) −r0 (T−t)
optimal reinsurance strategy a∗ will appears at a0∗ = θ eγ B1 +m such that la ( θ eγ B1 +m ) = 0 or the two end points of the
168 D. LI ET AL.

interval 0 and D. Therefore, the robust optimal reinsurance strategy is given by

⎧ −r0 (T−t)
⎪ θe θe−r0 (T−t)
⎪
⎪ · 1 !  + D · 1  !  , ∈ [0, D],
⎪
⎨ γ B1 + m l θe−r0 (T−t)
max{l(D),0} l(D)max l θe−r0 (T−t)
,0 γ B1 + m
γ B1 +m γ B1 +m
a∗ (t) =
⎪
⎪
⎪
⎪ θe−r0 (T−t)
⎩ D · 1{l(D)0} , ∈ [D, +∞),
γ B1 + m

and the optimal value function is given by

1 −m[er0 (T−t) x−f¯ (t)]

V (t, x) = − e ;
m

where
 T  T
λ1 (η − θ)μZ
f¯ (t) = (1 − er0 (T−t) ) + l̄1 (ω)dω − l̄2 (ω)dω,
r0 t t
 a∗ (ω)
l̄1 (ω) = er0 (T−ω) [λ1 (γ B1 + m)ser0 (T−ω) − λ1 θ]F̄(s)ds,
0
1
l̄2 (ω) = π (ω)(μ − r0 )er0 (T−ω) − (γ B2 + m)σ 2 (π ∗ (ω))2 e2r0 (T−ω)
∗
2 !
λ2 γ J Q −mπ ∗ (ω)yer0 (T−ω)
+ J 1 − e m E [e −1]
.
γ

" −r (T−t) #of claim size satisfies condition (3.10),

In addition, if the distribution
we can derive l θ eγ B1 +m > l(D) > l(0) = 0, ∀t ∈ [0, T].
0

D(γ B1 +m)
Equation (B7) shows that a∗ (t) ∈ [0, D] when t  k := T + r10 ln θ . We try to find the solution to Equation
(B6) in the following cases.

(i) If D > γ B1θ+m , we have k > T. Then, 0  t  T < k, and the robust optimal reinsurance-investment strategy
is shown in Equations (B7) and (B8). Similarly, we assume that the optimal value function is

1 −m[er0 (T−t) x−f1 (t)]

V (t, x) = − e , 0  t  T.
m

Plugging Equations (B7) and (B8) into Equation (B6), we have

 θe−r0 (T−t)
γ B1 +m
f1t − λ1 (η − θ)μZ er0 (T−t) + er0 (T−t) [λ1 (γ B1 + m)ser0 (T−t) − λ1 θ]F̄(s)ds
0 ! (B9)
1 λ2 γJ
EQ [e−mπ
∗ yer0 (T−t)
−1]
− π ∗ (μ − r0 )er0 (T−t) + (γ B2 + m)σ 2 (π ∗ )2 e2r0 .T−t/ − J 1−e m = 0.
2 γ

Considering the boundary condition f1 (T) = 0, the solution to Equation (B9) is

 
λ1 (η − θ)μZ T T
f1 (t) = (1 − er0 (T−t) ) + l1 (ω)dω − l2 (ω)dω, (B10)
r0 t t

where

 θe−r0 (T−ω)
r0 (T−ω) γ B1 +m
l1 (ω) = e [λ1 (γ B1 + m)ser0 (T−ω) − λ1 θ]F̄(s)ds, (B11)
0

1
l2 (ω) = π ∗ (ω)(μ − r0 )er0 (T−ω) − (γ B2 + m)σ 2 (π ∗ (ω))2 e2r0 (T−ω)
2 ! (B12)
λ2 γ J Q −mπ ∗ (ω)yer0 (T−ω)
+ J 1 − e m E [e −1]
.
γ
 SCANDINAVIAN ACTUARIAL JOURNAL 169

(ii) If D  γ B1θ+m , we have k  T. Thus, 0  t < k  T. In the case of 0  t  k, the derivations of the solution
to Equation (B6) are similar to those of case (i), and we assume that the optimal value function is

1 −m[er0 (T−t) x−f2 (t)]

V (t, x) = − e , 0  t  k,
m

where f2 (t) needs to be determined.

In the case of k < t  T, choosing a∗ (t) = D, and Equation (B6) becomes

λ1 σZ2 γ B1 Vx2 σ 2 (π ∗ )2 γ B2 Vx2 1

Vt + r0 xVx + λ1 ημZ Vx + π ∗ (μ − r0 )Vx + + + λ1 σZ2 Vxx
2mV ! 2mV 2 (B13)
1 mV λ2 γ J Q −mπ ∗ yer0 (T−t)
+ (π ∗ )2 σ 2 Vxx − 1 − e m E [e −1]
= 0.
2 γJ

Similarly, we have the solution to Equation (B13) as follows

1 −m[er0 (T−t) x−f3 (t)]

V (t, x) = − e , k < t  T.
m

Taking into account the boundary condition f3 (T) = 0 and the continuity of V (t, x) at time t = k, we have
 k  T
λ1 (η − θ)μZ r0 (T−k)
f2 (t) = (e − er0 (T−t) ) + l1 (ω)dω − l2 (ω)dω
r0 t t (B14)
λ1 ημZ λ1 σZ (γ 1 + m)
2 B
+ (1 − er0 (T−k) ) − (1 − e2r0 (T−k) ),
r0 4r0
 T
λ1 ημZ λ1 σZ2 (γ B1 + m)
f3 (t) = (1 − er0 (T−t) ) − (1 − e2r0 (T−t) ) − l2 (ω)dω, (B15)
r0 4r0 t

where l1 (ω) and l2 (ω) are given in Equations (B11) and (B12).

Summarizing the above analysis, we can derive the robust optimal reinsurance-investment strategy and the optimal
value function explicitly.
Next, conditions (i)–(v) in Proposition 2.2 will be checked. We first give two lemmas.
Lemma B1: The following expectation is finite
%  ! &
T (φ1∗ (t))2 (φ2∗ (t))2 ∗ ∗ ∗
J(T) := E exp + + λ2 (φ3 (t) ln φ3 (t) − φ3 (t) + 1) dt . (B16)
0 2 2

Proof: Substituting Equations (B3)–(B5) into Equation (B16), we have

%  ! &
(φ1∗ (t))2
T (φ2∗ (t))2 ∗ ∗ ∗
E exp + + λ2 (φ3 (t) ln φ3 (t) − φ3 (t) + 1) dt
0 2 2
% 
T 1
= E exp ¯ ∗ ))2 e2r0 (T−t) + 1 γB2 (π ∗ )2 σ 2 e2r0 (T−t)
γ 2 λ1 (σ (a
0 2 B1 2 2
!! 
γJ Q [e−mπ ∗ yer0 (T−t) −1] γ J Q −mπ ∗ yer0 (T−t) γ J Q −mπ ∗ yer0 (T−t)
λ2 e m E E [e − 1] − e m E [e −1]
+1 dt . (B17)
m

Since a∗ and π ∗ are finite, the right side of Equation (B16) is finite.
∗
Lemma B2: The optimal strategy u∗ and the corresponding function W(t, X u (t)) have the following properties:
(a) u∗ is"an admissible strategy; #
∗ ∗
(b) EQ supt∈[0,T] |W(t, X u (t))|4 < ∞;
'  ∗  (
∗  (φ (t))2 (φ ∗ (t))2 λ2 (φ3∗ (t) ln φ3∗ (t) − φ3∗ (t) + 1) 2
(c) EQ supt∈[0,T]  1B + 2B +  < ∞.
2ϕ 1 (t) 2ϕ 2 (t) ϕ J (t)
170 D. LI ET AL.

Proof:

(a) From the process of solving HJB equation, we know condition (i) in Definition 2.1 holds, and the optimal
strategy u∗ is deterministic and state-independent, thus condition (ii) in Definition 2.1 is satisfied. Condition
(iii) in Definition 2.1 can be obtained by property (b).
(b) Substituting Equations (B3)–(B5), (B7)–(B8) into Equation (2.6), we have
 t  t  t
√ ∗ ∗
λ1 σ̄ (a∗ )dB1Q (s) + π ∗ (s)σ dB2Q (s)
∗
X u (t) = x0 er0 t + Ads +
 t ∞ 0 0 0 (B18)
+ π ∗ (s)yN(ds, dy),
0 −1

where A = λ1 (θ μ̄(a∗ ) + (η − θ)μZ ) + (μ − r0 )π ∗ (s) − γ B1 λ1 (σ̄ (a∗ ))2 er0 (T−t) − γ B2 σ 2 (π ∗ (s))2 . Given that
u∗ is deterministic, A is bounded. Inserting Equation (B18) into candidate value function (3.13), we obtain the
following upper boundary with appropriate constants K > 0,
   
 1 r0 (T−t) X u∗ (t)−f¯ (t))   ∗ 
∗
|W(t, X u (t))4 | =  4 e−4m(e  =  1 e−4mer0 (T−t) X u (t)+4mf¯ (t)  (B19)
m   m4 
r0 (T−t) X u∗ (t)
 Ke−4me (B20)
t t √ ∗ t ∗ t ∞
−4mer0 (T−t) (x0 er0 t + 0 Ads+ λ1 σ̄ (a∗ )dB1Q (s)+ ∗ Q
0 π (s)σ dB2 (s)+ π ∗ (s)yN(ds,dy))
= Ke 0 0 −1 (B21)
t √ ∗ t ∗
−4m( λ1 σ̄ (a∗ )dB1Q (s)+ ∗ Q
0 π (s)σ dB2 (s))
 K̄e 0 , (B22)

r0 (T−t)
t t ∞ ∗
where K̄ is a constant satisfying K̄ > Ke−4me (x0 e + 0 Ads+ 0 −1 π (s)yN(ds,dy)) r0 t
. The first inequality
in Equation (B19) is valid, because ¯
f (t) is deterministic and bounded, and the second inequality follows
t t ∞
from the fact that x0 er0 t , er0 (T−t) , 0 Ads and 0 −1 π ∗ (s)yN(ds, dy) are deterministic and bounded. Now, we
t √ ∗
consider the integral e 0 −4m λ1 σ̄ (a∗ )dB1Q (s) .

t √ ∗ t t t √ ∗
−4m λ1 σ̄ (a∗ )dB1Q (s) 8m2 λ1 (σ̄ (a∗ ))2 ds − 8m2 λ1 (σ̄ (a∗ ))2 ds+ −4m λ1 σ̄ (a∗ )dB1Q (s)
e 0 = )e 0
*+ , · )e
0
*+
0
,.
const. martingale

Thus,
 t √ ∗

∗ −4m λ1 σ̄ (a∗ )dB1Q (s)
EQ e 0 < ∞.

 t ∗

Similarly, EQ
∗
e 0 −4mπ ∗ (s)σ dB2Q (s) < ∞. Consequently,

% &
∗ ∗
EQ sup |W(t, X u (t))|4 < ∞.
t∈[0,T]

(φ ∗ (t))2 m (φ ∗ (t))2 m λ (φ ∗ (t) ln φ3∗ (t)−φ3∗ (t)+1)m

(c) Let (t) = 12γ B1 + 22γ B2 + 2 3 γJ
, which is obviously bounded, and according to
Equation (3.2) with W instead of V , we have
'  ∗  (
∗  (φ (t))2 (φ ∗ (t))2 λ2 (φ3∗ (t) ln φ3∗ (t) − φ3∗ (t) + 1) 2
EQ supt∈[0,T]  1B + 2B + 
2ϕ 1 (t) 2ϕ 2 (t) ϕ J (t)
∗
" ∗
#
 EQ supt∈[0,T] |(t)|2 |W(t, X u (t))|2
" ∗ #1 " ∗ ∗
#1
EQ supt∈[0,T] |W(t, X u (t))|4
2 2
 EQ supt∈[0,T] |(t)|4 < ∞.

The first inequality follows from Cauchy–Schwarz inequality, and the second inequality follows from
property (b).
 SCANDINAVIAN ACTUARIAL JOURNAL 171

From the above derivations, it is easy to check that conditions (i)–(iv) in Proposition 2.2 hold for W(t, x). By
Lemma B.2, condition (v) in Proposition 2.2 also holds for W(t, x). Then, W(t, x) is the optimal value function of
problem (2.9), i.e. W(t, x) = V (t, x), and u∗ = {(a∗ (t), π ∗ (t))}t∈[0,T] is the optimal strategy.
The proof can also be referred to Corollary 1.2 in Kraft (2004), Theorem 3.1 in Øksendal & Sulem (2007) and
Theorem 8.1 in Fleming & Soner (2006).

Appendix 3.
C.1. Derivation of suboptimal value function
The optimal value function V̌ (t, x) associated with û∗ solves the infimum problem

 √ 
inf V̌t + V̌x r0 x + λ1 (θ μ̄(â∗ ) + (η − θ)μZ ) + π̂ ∗ (μ − r0 ) + λ1 σ̄ (â∗ )φ1 + π̂ ∗ σ φ2
(φ1 ,φ2 ,φ3 )∈R×R×R+
1
+ V̌xx [λ1 (σ̄ (â∗ ))2 + (π̂ ∗ )2 σ 2 ] + λ2 φ3 EQ [V̌ (t, x + π̂ ∗ y) − V̌ (t, x)] (C1)
2 
φ2 φ2 λ2 (φ3 ln φ3 − φ3 + 1)
+ B1 + B2 + = 0,
2ϕ 1 (t) 2ϕ 2 (t) ϕ J (t)

where V̌ is a short notation for V̌ (t, x) with the boundary condition V̌ (T, x) = U(x). The first-order conditions w.r.t.
φ1 , φ2 and φ3 show that the alternative model is given by φ1∗ , φ2∗ and φ3∗ in Equations (B3)–(B5) with â∗ and π̂ ∗
substituted for a and π , respectively. By inserting Equations (B3)–(B5) into Equation (C1), we have

λ1 (σ̄ (â∗ ))2 γ B1 V̌x2 σ 2 (π̂ ∗ )2 γ B2 V̌x2

V̌t + r0 x V̌x + λ1 (θ μ̄(â∗ ) + (η − θ)μZ )V̌x + π̂ ∗ (μ − r0 )V̌x + +
2mV̌ ! 2mV̌ (C2)
1 1 mV̌ λ2 γJ Q [e−mπ̂ ∗ yer0 (T−t) −1]
+ λ1 (σ̄ (â∗ ))2 V̌xx + (π̂ ∗ )2 σ 2 V̌xx − 1−emE = 0.
2 2 γJ

We try to find the solution to Equation (C2) in the following way

1 −m[er0 (T−t) x−f0 (t)]

V̌ (t, x) = − e . (C3)
m

Plugging the relevant derivatives into Equation (C2), we derive

 
λ1 (η − θ)μZ T T
f0 (t) = (1 − er0 (T−t) ) + l6 (ω)dω − l7 (ω)dω, (C4)
r0 t t

where

 θe−r0 (T−ω)
m
l6 (ω) = er0 (T−ω) [λ1 (γ B1 + m)ser0 (T−ω) − λ1 θ]F̄(s)ds, (C5)
0

1
l7 (ω) = π̂ ∗ (ω)(μ − r0 )er0 (T−ω) − (γ B2 + m)σ 2 (π̂ ∗ (ω))2 e2r0 (T−ω)
2 ! (C6)
λ2 γ J Q −mπ̂ ∗ (ω)yer0 (T−ω)
+ J 1 − e m E [e −1]
.
γ

Then we have the expressions V̌ (t, x).