Insurance: Mathematics and Economics 99 (2021) 174–189

Contents lists available at ScienceDirect

Insurance: Mathematics and Economics

www.elsevier.com/locate/ime

Joint and survivor annuity valuation with a bivariate reinforced urn

process
Luis A. Souto Arias a,∗ , Pasquale Cirillo b
a
Centrum Wiskunde & Informatica and Utrecht University, the Netherlands
b
Institute of Business Information Technology, Zurich University of Applied Sciences, Switzerland

a r t i c l e i n f o a b s t r a c t

Article history: We introduce a novel way of modeling the dependence of coupled lifetimes, for the pricing of joint and
Received July 2020 survivor annuities. Using a well-known Canadian data set, our results are analyzed and compared with
Received in revised form February 2021 the existing literature, mainly relying on copulas. Based on urn processes and a one-factor construction,
Accepted 6 April 2021
the proposed model is able to improve its performances over time, in line with the machine learning
Available online 20 April 2021
paradigm, and it also allows for the use of experts’ judgements, to complement the empirical data.
JEL classification: © 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license
G22 (http://creativecommons.org/licenses/by/4.0/).
C11
E17

Keywords:
Annuity
Bivariate survival function
Reinforced urn process
Bayesian nonparametrics
Right-censoring

1. Introduction to a particular class of models (Muliere et al., 2000; Walker

and Muliere, 1997) with the extremely useful ability of com-
We propose a novel nonparametric approach to the modeling bining some a priori knowledge–possibly referring to experts’
of joint and survivor annuities, particularly useful in the case of judgements–with the information coming from actual data, per-
right-censored observations. We build on previous results by Bulla fectly in line with the Bayesian paradigm. This possibility allows
et al. (2007); Muliere et al. (2000); Walker and Muliere (1997). The for the incorporation of trends, tail events or other aspects that
goal is to propose an alternative to copulas and other parametric can be rarely observed in a data set–hence invisible to standard
constructions already used for the modeling of coupled lifetimes in machine learning approaches, yet possible and with dramatic con-
insurance, for example in Frees et al. (1996), Carriere (2000) and sequences (Taleb, 2007). For instance, if an expert thinks that their
Luciano et al. (2008), to cite a few important contributions. data under-represent a given phenomenon, like for example some
The basic ingredients of our approach are Pólya urns (Mah- unusual lifetime combinations in the modeling of survivor annu-
moud, 2008), probabilistic objects with the ability of intuitively ities, they could solve the problem by eliciting an a priori assigning
representing the idea of learning via reinforcement. As observed a higher mass to those combinations, thus obliging the model to
in Cirillo et al. (2013) and Cheng and Cirillo (2018), urns are able always take into account such a possibility, even if rarely (or not)
to learn hidden patterns in the data, discovering previously ignored present in the data. In a sense, a clever use of the priors can thus
features. The model here presented can actually be seen as a non- be an elegant way of dealing with sampling and historical bias
conventional machine learning algorithm (Murphy, 2012). (Derbyshire, 2017; Shackle, 1955).
Differently from the great majority of machine learning al- Clearly, nothing guarantees the ability of eliciting a sound and
gorithms, our Bivariate Reinforced Urn Process (B-RUP) belongs reliable a priori: experts could naturally be wrong. The answer
to such a relevant observation–we shall see–is that the bivariate
urn model learns over time, at every interaction with actual data.
A sufficient amount of data can therefore compensate for judge-
*
Corresponding author.
ments one does not trust completely, for unrealistic beliefs, or for
E-mail addresses: [email protected] (L.A. Souto Arias),
[email protected] (P. Cirillo). no belief at all. In fact, as it shall be clearer later, the elicitation

https://doi.org/10.1016/j.insmatheco.2021.04.004
0167-6687/© 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

of an a priori is not compulsory, it is only a relevant plus, and a We can then look at the probability that at least one of the two
B-RUP can also be used in an “objective” empirically-based way. annuitants survives another k years, i.e.
Another interesting feature of reinforced urn processes, when (x, y )
seen through statistical learning glasses, is that they are somehow P L S (k) := P [( X > x + k) ∪ (Y > y + k)| X > x, Y > y ]
immune to the “black box” argument (Knight, 2017) commonly (x, y ) (x, y ) (x, y )
= SX (k) + S Y (k) − S X Y (k, k), (2)
used to criticize machine/deep learning. In fact, differently from
other approaches, a B-RUP can be controlled and studied in its where the subscript L S is the acronym of “last survivor”, with
(x, y )
probabilistic details, controlling its features and allowing for fine S Y (k) = P (Y > y + k| X > x, Y > y ) indicating the marginal sur-
tuning. vival probability of the second annuitant given the present situa-
(x, y )
The paper is organized as follows. In Section 2 we briefly re- tion of the couple, and similarly S X (k) for the first annuitant.
call joint and last-survivor annuities, while in Section 3 we shortly It is important to stress that, in Equation (2), both marginal prob-
deal with the problem of right-censoring. In Section 4, we describe abilities are conditioned on the ages of both annuitants. This is
the B-RUP model, focusing on its ability of dealing with censored because, in the general situation where the lifetimes X and Y
(x, y ) (x,0)
observations. In Section 5, we discuss our results, using artificial share some kind of dependence, we have that S X (k) = S X (k)
and real data, and providing suggestions on how to use the B-RUP (Youn et al., 2002).
in practice, in terms of prior elicitation and parameter tuning. The If we assume a constant interest rate r, the basic one-unit pric-
real data are the same used by Frees et al. (1996), that we thank ing formula for a 100% joint and last-survivor annuity is given by
for sharing them with us, and that will be a benchmark for us. In-
∞ (x, y )

terestingly, as we shall see, the B-RUP is able to obtain the same P LS(k)
C (x, y ) = . (3)
results of Frees et al. (1996), while guaranteeing more flexibility. (1 + r )k
k =0
Finally, Section 6 concludes the work and looks at future research.
An appendix extends some aspects of Section 5. In the 100p% situation, with 0 < p < 1, this formula is easily ad-
justed (Winklevoss, 1977).
2. Joint and survivor annuities Equations (2) and (3) imply that, in order to evaluate joint and
survivor annuities, one needs to specify the dependence between
An annuity is a financial contract between two counterparties: the lifetimes of the annuitants. If we assume that their lifetimes
(x, y ) (x) ( y)
an individual, called the annuitant, and a financial institution, like are independent, then S X Y (k, k) = S X (k) S Y (k), and the estima-
an insurance company. In such a contract, the former pays a given tion of the marginal probabilities of survival is all that is needed.
amount of money (in full or periodically) to the latter and, in re- Since the estimation of univariate survival functions is a widely
turn, they receive periodic payments, usually starting from a given studied and well-known topic (e.g. Cox and Oakes (1984)), pricing
time point in the future, and commonly until death. The goal annuities under independence makes both modeling and compu-
of annuities is therefore to provide the annuitant with a steady tations considerably easier (Winklevoss, 1977), but not necessarily
stream of income, typically during retirement, and to offer in- realistic.
surance against lifetime uncertainty. It goes without saying that That of independence is indeed always a strong assumption. As
annuities are strictly connected to private pensions and other sim- Frees et al. (1996); Carriere (2000); Luciano et al. (2008); Sanders
ilar products (Sheshinski, 2007). and Melenberd (2016) and others have shown, the lifetimes of
A simple and quite popular extension of the single annuitant several couples in annuity contracts present a non-negligible pos-
case we have just described is given by annuities involving more itive dependence. Moreover, some studies suggest that annuitants
annuitants, most of the times just two. These are often the mem- tend to have higher individual survival probabilities with respect
bers of a married couple or of a partnership, but parent-child and to people who do not buy annuities (Mitchell et al., 1999). Such a
relative-relative combinations are also common (Frees et al., 1996; result opens to the possibility that also the dependence structure
Brown and Poterba, 2000). An example is offered by joint and last- among couples that buy annuities, and couples that do not, dif-
survivor annuities, in which the insurance company pays as long fers. Although there is no conclusive study about differences in the
as at least one of the two annuitants is alive. In this kind of con- coupled lifetimes of annuitants versus non-annuitants, Sanders and
tracts, the death of one of the annuitants may have an impact on Melenberd (2016) observe that the magnitude of the positive de-
the stream of payments that the survivor receives. For instance, for pendence for a random subset of 50, 000 couples–not necessarily
p ∈ [0, 1], in a 100p% (e.g. 50%) joint and survivor annuity contract, annuitants–sampled from the whole Dutch population is signifi-
payments are made in full while both annuitants are alive, but if cantly smaller than what Frees et al. (1996) and Luciano et al.
one of the two dies, only 100p% (e.g. 50%) of the original amount is (2008) find for about 15000 couples of annuitants in Canada.1
paid to the survivor (Sheshinski, 2007; Winklevoss, 1977). A pop- Regardless of its magnitude, there are several reasons that ex-
ular version is the contingent joint and survivor annuity, which plain, at least qualitatively, why the dependence in the lifetimes
pays the full benefit as long as the first annuitant (plan member) of couples (especially married couples and partnerships) is posi-
tive. Some of them are rather intuitive, like the possibility of both
is alive, and then changes the benefit to a portion which is usually
individuals dying at the same time because of an accident or a
one-half or two-thirds.
contagious disease, or the common habits they might share as a
From an actuarial point of view, to evaluate a joint and sur-
consequence of living together, like the food they eat or the en-
vivor annuity, one needs the marginal and joint survival functions
vironment they live in. Others are more linked to emotional, psy-
of both annuitants. Let X and Y be two random variables rep-
chological and medical aspects, like the “broken heart syndrome”,
resenting the lifetimes of two annuitants in a joint and survivor a well-known phenomenon in medicine, in which the sudden pass-
annuity. Assume that today the age (say in years) of the first an- ing of one spouse greatly increases (at least temporarily) the prob-
nuitant is observed to be x, while for the second we have y. The ability of death of the surviving one because of bereavement and
joint probability of survival for another k years, given the current sadness (Jagger and Sutton, 1991).
ages, is
(x, y )
S X Y (k, k) := P ( X > x + k, Y > y + k| X > x, Y > y ), (1) 1
Naturally an analysis taking into consideration the differences in the two coun-
tries and the related markets would be necessary to clean the effects, but the idea
where k = 0, 1, ..., K . of some selection mechanism does not seem ludicrous.

175
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

A question that naturally arises, when we consider the several copulas, but, at least within the realm of positive dependence, it
possibilities for the modeling of joint mortality, is how these affect also allows for a more flexible modeling, as it improves its per-
the annuity price in Equation (3). This question is well-addressed formances over time, and it allows for the exploitation of experts’
in Frees et al. (1996), where they compute the ratio between the judgements.3
annuity price under positive dependence and the price under inde- To describe the model, we start by defining its main building
pendence (we shall call it annuity ration from now on). To model blocks: the beta-Stacy process of Walker and Muliere (1997), and
dependence, they make use of copula models with Gompertz and the Reinforced Urn Process (RUP) of Muliere et al. (2000). Then
Weibull marginals, and they find out that the annuity ratio varies we show an intuitive representation of RUPs via urns. Finally we
between approximately 0.95 and 1.05, depending on the initial discuss how several RUPs can be easily combined into a one-factor
ages of the annuitants at contract initiation. This means that, if construction leading to the B-RUP.
one assumes independence to simplify modeling, they may end
up with the underestimation or the overestimation of the annu- 4.1. Reinforced Urn Processes and beta-Stacy processes
ity price by as much as a 5%. Compatible results are more recently
obtained by Sanders and Melenberd (2016), even if the error they The Reinforced Urn Process was first described in Walker and
find is smaller than 5% on average, in absolute terms, for different Muliere (1997), where the results of Blackwell and MacQueen
types of contracts. Furthermore, the same authors underline that (1973) and Ferguson (1973) are extended to right-censored data,
the error can be particularly larger and relevant for specific types thus defining the so-called beta-Stacy process, a neutral-to-the-
of annuities, in which the joint survival distribution plays a major right (Doksum, 1974) generalization of the Dirichlet process (Fer-
role, like for instance joint annuities. guson, 1973).
The RUP is a combinatorial stochastic process, and it can be
3. Right-censoring seen as a reinforced random walk over a state space of urns. De-
pending on how its parameters are specified, it can generate a
In survival analysis, right-censoring is a well-known problem large number of interesting models. Essential references on the
in estimation (Kaplan and Meier, 1958). This usually occurs when topic are Muliere et al. (2000, 2003) and Fortini and Petrone
individuals drop from the study before the event of interest has (2012). Examples of applications can be found for example in Cir-
happened. If X is our age variable, when right-censoring occurs, illo et al. (2013) and Peluso et al. (2015).
the minimum X ∗ = min( X , T X ) between X and a random censor- In this paper we specify a reinforced urn process able to gener-
ing time T X is observed.2 Every potentially censored observation ate a discrete beta-Stacy process, a particular random distribution
is therefore represented by the couple ( X ∗ , δ), where the indica- over the space of discrete distributions.
tor δ tells if a given observation is censored (δ = 0) or not (δ = 1).
Left-censoring can also be defined, but it is less relevant here. For Definition 4.1 (Walker and Muliere (1997)). A random distribution
more details see Klein and Moeschberger (2003). function F is a discrete beta-Stacy process with jumps at j ∈ N
While univariate censoring has been extensively studied in the and parameters {β j , ω j ∈ R+ , j ∈ N}, if there exist mutually in-
literature, e.g. in Kaplan and Meier (1958); Elandt-Johnson and dependent random variables { V j } j ∈N , each beta distributed with
Johnson (1980) or Cox and Oakes (1984), far less progress has been parameters (β j , ω j ), such that therandom mass assigned by F to
made in the bivariate case. One of the first works to study this { j }, written F ({ j }), is given by V j i < j (1 − V i ).
problem is that of Dabrowska (1988), in which the Kaplan-Meier
estimator is extended to the two-dimensional framework. How- Following Definition 4.1, we introduce couples {β j , ω j } ∈
ever, this estimator is known to produce negative masses when R+ × R + , with j ∈ N , such that β , ω ≥ 0, β + ω > 0, and
j j j j
a large amount of censoring is present (Pruitt, 1991). In Wang and n ω
limn→∞ j =0 β +jω = 0. Then, given a discrete beta-Stacy process
j j
Wells (1997) or Gribkova and Lopez (2015), the censoring condi-
F with parameters {β j , ω j ∈ R+ , j ∈ N}, and a sample ( X n∗ , δn ) of
tions are relaxed to obtain better estimators. For example, in the
context of coupled lifetimes, Gribkova and Lopez (2015) assume exchangeable4 and potentially censored observations, with X n∗ =
that the age difference between individuals is known and related { Xn∗ , n ≥ 1}, the sequence X n = { Xn , n ≥ 1} is a RUP if
to the censoring variables; in Lin (1993), conversely, the censoring
Ŝ X (x) = P ( X n+1 > x| X n∗ = xn∗ , δn = dn )
variable is taken to be the same for both lifetimes. Other recent  
articles on nonparametric estimators under bivariate censoring are 
x
β j + m∗j (xn∗ , dn )
= 1− , (4)
those of Lopez (2012) and Shen and Yan (2008). In all these works, β j + ω j + s j (xn∗ )
censoring is assumed to be independent from the variables of in- j =0
n
terest, an assumption we will also make in our construction. where m∗j (xn , dn ) = i =1 1{xi = j ,di =1} is the number of exact ob-
In line with the univariate case, in the presence of bivariate n
servations at x = j and s j (xn ) = i =1 1{xi ≥ j } is the number of
censoring one plays with the quartet ( X ∗ , δ, Y ∗ ,  ), where  is the
observations at x ≥ j. In the context of survival analysis, Equation
censoring variable associated to Y , and Y ∗ = min(Y , T Y ). The vari-
(4) gives the probability of survival of a new individual ( X n+1 ),
ables ( X , Y ) are assumed independent from ( T X , T Y ).
conditioned on the lifetime observations of previous individuals
( X n∗ ).
4. The Bivariate Reinforced Urn Process
Notice that, by defining β ∗j = β j + m∗j (xn , dn ) and ω∗j = ω j +
The Bivariate Reinforced Urn Process (B-RUP) is our proposal for s j (xn ) − m∗j (xn , dn ), we can obtain a new beta-Stacy process F ∗
the modeling of joint and survivor annuities. Originally introduced
by Bulla (2005); Bulla et al. (2007), it is here extended and adapted
3
to the annuity framework. It is worth noticing that copulas can also deal with negative dependence, so
they are more flexible in that sense. However, when positive dependence can be
We will show that this model is not only able to replicate
safely assumed, the B-RUP needs less and weaker hypotheses about the shape of
the results one can obtain with other existing methodologies, like the joint distribution.
4
Exchangeability is a common assumption in Bayesian statistics (de Finetti,
2017), and it represents a relaxation of the stronger hypothesis of independence.
2
A common assumption (Gribkova and Lopez, 2015) is that the support of the A sequence of random variables is exchangeable if their joint distribution is im-
distribution of X is included in the support of the censoring variable T X . mune to permutations in the order of appearance of the variables.

176
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

Further, the exchangeability of the observations is implied by the

evolution of the Pólya urns (Mahmoud, 2008).
As said, the urn construction just analyzed generates a RUP
without right-censoring. The possibility of censored observations
can however be easily introduced: in case of a right-censored value
at Urn j, we add extra green balls from Urn 0 to Urn j (included)
and no red balls. Then we start again from Urn 0.
At every cycle, which for us represents the lifetime of an in-
dividual, the RUP learns and keeps memory of what happened
before, thanks to the Pólya reinforcement, combining the informa-
tion from samplings with the initial compositions of the urns (the
a priori). In this way what we learn from a group of people can be
used to make inference about other people.
In Muliere et al. (2000), the properties of the urn construction,
including the conditions for recurrence, are studied in detail.
Fig. 1. Representation of the RUP as a series of Pólya urns. After each sampling the
urns are updated in a way that reinforces the probability of that sampling, giving
the RUP its name. (For interpretation of the colors in the figure(s), the reader is 4.1.2. Defining a prior
referred to the web version of this article.) One of the characteristics of the beta-Stacy process, inherited
from the Dirichlet process, is that its trajectories are centered
with parameters {β ∗j , ω∗j ∈ R+ , j ∈ N}. The beta-Stacy process is around a certain probability distribution G (·), that is E[ F ({ j })] =
thus conjugate and neutral-to-the-right (Doksum, 1974). The im- G ({ j }), which—in Bayesian terms—plays the role of the prior. As
portance of conjugacy is to be stressed, as it allows for continuous shown in Walker and Muliere (1997), a necessary condition for
updating of the process over time, every time new observations this is that
become available. In other words, the posterior one can obtain as βj G ( j ) − G ( j − 1)
the result of a first cycle of Bayesian updates can then represent = , j∈N (5)
the a priori of a new cycle, without the necessity of restarting ev- βj + ωj 1 − G ( j − 1)
erything from scratch (Muliere et al., 2000). where G ( j ) = PG ( X ≤ j ) is the probability that X is at most j
under the centering measure.
4.1.1. Urn representation The prior distribution G can also be used to define the initial
Following Muliere et al. (2000), we can show that Equation (4) compositions in a RUP. Following Walker and Muliere (1997) and
can be obtained via a sequence of two-color Pólya urns. This intu-
Muliere et al. (2000), it is sufficient to set
itive representation makes the beta-Stacy process easier to grasp.
To further simplify the treatment, we momentarily ignore censor-
β j = c j G ({ j }), ω j = c j (1 − G ( j )), c j ∈ R+ , j ∈ N, (6)
ing.
Imagine we have M + 1 urns containing balls of two colors (red with c j denoting the so-called strength of belief in the prior
and green), as in Fig. 1a (hence M = 4). Urn 0 only contains green knowledge, and G ({ j }) = PG ( X = j ). Notice that, given Equation
balls. To generate a RUP, we move along the urns according to the (6), the necessary conditions for the couples {β j , ω j } to properly
following rules: define a beta-Stacy process are automatically verified, if G (·) is a
well-defined probability distribution.
• The process starts from Urn 0. By giving high values to the strength of belief c j , larger
• For each urn, the probability that a given color is selected sim- amounts of data are required in order for the posterior distribu-
ply depends on the urn composition at the time of sampling. tion, obtained via sampling and given in Equation (4), to move
• Every time we pick a ball, we note the color, and we reinforce away from the a priori. On the contrary, when c j → 0, Equation
the urn, that is we put the ball back together with an extra (4) tends to the Kaplan-Meier (KM) estimator of Kaplan and Meier
ball of the same color. This increases the probability of picking (1958). As observed in Walker and Muliere (1997), setting c j = c,
that color again in the future. ∀ j, leads to a Dirichlet prior.
• If the color of the ball is green, we move forward to the next
urn, which we sample to continue our walk. Conversely, if the
4.2. Building dependence
color is red we must go back to Urn 0, and the process starts
anew. Every time we restart from Urn 0, we define a new cycle
We have seen how the RUP model can be used to model life-
for the RUP.
times of single individuals. However, the ultimate goal of this work
is the modeling of coupled lifetimes, and for that reason we need
In Fig. 1b we see an example of a possible trajectory (•, •, •) and
to extend the RUP so that it can deal with this type of problems.
the resulting urn composition after each sampling. Notice that the
We can do so by using the one-factor construction introduced in
probability of observing the same path in the next cycle has in-
Bulla (2005): the Bivariate Reinforced Urn Process (B-RUP).
creased, thanks to reinforcement. Please notice that from Urn 0 we
Assume we observe  n couples of (possibly) censored lifetime
necessarily move to Urn 1, while from the all other urns, we can
data of the form ( X n∗ , δn ), (Y n∗ ,  n ) , where X n∗ and Y n∗ are n-
move forward or jump back to Urn 0.
dimensional vectors of observations, and where δ n and  n are the
If we define the age X of death of an individual as the Urn X
corresponding vectors of indicators for right-censoring, as defined
where we sample a red ball, and every urn corresponds to one year
in Section 3.
of life, it is easy to see that the survival probability given by the
urn construction above and that of Equation (4) coincide, under Let A n , B n and C n be three independent RUPs, defined exactly
no censoring. The pairs {β j , ω j } would correspond to the initial as in Section 4.1 and with their respective parameters (β jA , ω Aj ),
numbers of balls of each color for the j-th urn, and the functions (β Bj , ω Bj ) and (β Cj , ω Cj ) for j ∈ N . One can easily create a bivariate
s j (·) and m∗j (·, ·) to the extra numbers of balls due to sampling. model of the form

177
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

Xi = Ai + B i P [ B n = xn − an , δn = dn | B n−1 = X n∗−1 − An[k−−11] , δn−1 ] =

(7)
Y i = Ai + Ci , 1 ≤ i ≤ n. P [ B n ≥ xn − an | B n−1 = X n∗−1 − An[k−−11] , δn−1 ] −
[k−1]
The lifetimes of X i and Y i are thus composed of a common quan- 1{δn =1} P [ B n > xn − an | B n−1 = X n∗−1 − An−1 , δn−1 ],
tity A i and two personal elements, B i and C i . In this way, the
dependence between X i and Y i relies entirely on A i , and thus, and with the conditional distribution of C n defined in an anal-
conditioned on A i , X i and Y i are independent. ogous way.
[k]
A straightforward calculation yields Since An is an exchangeable process, Equation (10) also ap-
[k] [k−1]
plies for any element of A n by simply changing A n−1 with
Cov( X n+1 , Y n+1 | A n , B n , C n ) = Var( A n+1 | A n ) > 0, n ≥ 1, (8) [k−1]
A −i = { An[k] , 1 ≤ n < i , or An[k−1] , n > i ≥ 1}. With this in
thus implying that the B-RUP construction can only model positive [k] [k−1]
mind, start simulating A 1 using the values of A −1 and re-
dependence. For the application at hand, that is the joint modeling [k]
peat this process until a new A n has been obtained. Once
of lifetimes of coupled annuitants, this is not a problem, since pre- [k] [k]
this is done, compute B n = X n − A n and C n = Y n∗ − A n . Note
∗
vious studies have shown a strictly positive dependence (Frees et
that, defined in this way, right-censoring is only applied to B n
al., 1996). However, it is important to stress that the B-RUP model [k]
should not be used when negative dependence is also possible. and C n , and not to A n . As already underlined in Frees et al.
In Bulla (2005), the sequence {( X n , Y n ), n ≥ 1} is shown to be (1996) and Carriere (2000), such an assumption appears ac-
exchangeable (given that the RUPS A n , B n and C n are exchangeable ceptable from an empirical point of view.
by construction), and therefore, by the de Finetti representation 2. Create a new combination ( A k , B k , C k ) from the respective
theorem (de Finetti, 2017), there exists a joint random distribu- conditional marginal distributions, as per Equation (4).
tion function F X Y conditionally on which the elements of ( X n , Y n ) 3. Compute X k = A k + B k and Y k = A k + C k . Set k = k + 1.
are independent and identically distributed according to F X Y . The 4. Repeat steps 1-3 until k reaches the maximum number of it-
properties of F X Y have been studied in detail in Bulla et al. (2007). erations N.
Let F X and F Y be the marginal distributions of X and Y , re-
Notice that the previous algorithm requires, besides the (possibly)
spectively. Clearly, we have
censored observations of ( X , Y ), an initial sample for A, that is,
[0]
F X = F A × FB, A n . Since A is not observed in practice, a sensible approach is to
[0]
run the MCMC several times for different samples A n .
FY = F A × FC ,

with × denoting the convolution operation, so that both F X and 5. Empirical results
F Y are convolutions of beta-Stacy processes. Furthermore, if P is
the probability function corresponding to F , one has In this section we apply the B-RUP model to calculate the de-
pendence of coupled lifetimes, in order to price joint and survivor
annuities.

min(x, y )
P X Y (x, y ) = P A (a) P B (x − a) P C ( y − a), ∀x, y ∈ N02 . We first consider an analytical example, where we sample pos-
a =0
sibly censored observations from a known distribution. Knowing
the original distribution will not only allow us to estimate the
(9)
difference between the original distribution and the posterior com-
Although the joint posterior distribution of Equation (9) can in puted through MCMC, but also to study how these differences
principle be computed analytically5 (Bulla et al., 2007), its com- affect the final price of the annuity.
plexity grows quickly with the number of observations, making In the second part of the section, the B-RUP is tested on a well-
it already unfeasible for a relatively small sample. A convenient known Canadian data set originally used by Frees et al. (1996),
way of by-passing the problem is therefore to use a Markov Chain which we will also use as a benchmark.
Monte Carlo (MCMC) approach, also known as Gibbs sampler, to
obtain the joint probability distribution from the posterior distri- 5.1. Analytical example
butions of A n , B n and C n , using the one-factor construction.
The MCMC method consists of the following steps: Assume that X and Y are linked through the one-factor model
of Equation (7). Assume that A follows a Poisson distribution with
[k]
1. Generate a realization of A n , with the superscript referring to intensity parameter λ equal to 25 (from now on we use the no-
the iteration number, at the k-th iteration conditioned on the tation A ∼ Poi(25)), and that B ∼ Poi(35) and C ∼ Poi(40). This
[k] implies that the marginals of X and Y are Poisson distributions
available data of the previous iteration. That is, sample A n
from its conditional distribution, i.e. with parameters 60 and 65, respectively; and that Cov( X , Y ) =
Var( A ) = 25. The correlation between X and Y is therefore ap-
P [ An[k] = an | An[k−−11] , X n∗ , δn , Y n∗ ,  n ] ∝ (10) proximately 0.4.
[k−1] Moreover, while not necessary true in general, we will make
P [ B n = xn − an , δn = dn | B n−1 = X n∗−1 − A n −1 , δ n −1 ] some assumptions about the dependence between the censoring
[k−1]
P [C n = yn − an , n = en |C n−1 = Y n∗−1 − A n −1 ,  n −1 ] variables, T X and T Y (see Section 3), so that we can generate
censored observations and better study the properties of our con-
P [ An = an | An[k−−11] ],
[k]
struction. We start from the naive observation that the two an-
nuitants of a joint and survivor annuity enter the contract at the
where, using the fact that P [ X ≥ k] − P [ X > k] = P [ X = k],
same time; say that X 0 is the age of the first annuitant at the date
one has
of the signature, while Y 0 is that of the second person. If we de-
note the observation period by , it is clear that T X = X 0 +  and
5
Also observe that the conjugacy of the constituent RUPs { An }, { B n } and {C n } T Y = Y 0 + . Moreover, X 0 = Y 0 + θ , where θ is the age difference
allows for the continuous updating of the B-RUP as well. between the annuitants, which can be negative if Y 0 > X 0 .

178
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

Fig. 2. Comparison of the marginal distributions of X and Y . In green we show the prior distribution, the Kaplan-Meier estimators in red, the B-RUP solution in blue and the
original distribution in black. All distributions are conditioned on the minimum value of the uncensored observations. For this sample those were Xmin = 34 and Y min = 39.

Table 1 so that, for this particular example, we generate 104 samples from
Proposed strength of belief scenarios. Notice that, since we are considering a data a Poi(20).
set of size 104 , even under the high belief scenario, the prior will not be able to
In Fig. 2 we show the recovered marginal distribution using the
actually affect those areas where the observations are more concentrated, but it
will definitely influence the areas with fewer data points. B-RUP via MCMC, as well as the KM estimators and the original
distribution. Notice how, due to the lack of exact samples on the
Low belief High belief
right tail of the distribution, the KM estimator is not properly de-
c A , c B , cC 10−6 102 fined on the whole support of the original distribution, but only up
to the maximum value of the uncensored observations. The B-RUP
model, on the contrary, is able to recover the right tail of the dis-
Therefore, to obtain right-censored observations, we have to tribution, although the fitting is clearly worse in that part of the
choose the distributions of T X ,  and θ . In what follows we take curve, because of the evident lack of observations. If available, a
T X ∼ Poi(50),  ∼ Poi(2) and θ = θ ∗ − θ0 , with θ ∗ ∼ Poi(7) and better a priori could here be used to improve tail fitting.
θ0 = 5. We then sample n = 104 couples of observations, of which We also present the joint distribution in Figs. 3a and 3b, for the
almost 90% turn out to be at least partially censored (i.e. at least low and high belief scenarios, respectively. Notice how, in Fig. 3a,
one of the annuitants in the couple is right-censored). the fitting is worse in the upper-right corner than in the lower-left
Following the procedure defined in Section 4, the first step to corner, in accordance with the right-censoring effect. For the high
use the B-RUP is to define prior distributions for the target vari- belief scenario, however, the prior distribution dominates in those
ables according to Equation (6), as well as the strength of belief areas of few observations and we recover a smoother surface, as
per Fig. 3b.
parameters. The priors we choose are G A = Poi(20), G B = Poi(20)
A more quantitative comparison can be found in Table 2, where
and G C = Poi(20), respectively. For the strengths of belief, ck with
we show the means, the variances and the correlation for both
k ∈ { A , B , C }, we consider two different scenarios: one where the
B-RUP scenarios, as well as for the KM estimator, the theoretical
strengths of belief are very small, so that the posterior distribution
analytical solution, and the data. In this last case we provide the
is strongly affected by the incoming data, as if we did not trust our
sample estimates under two different points of view: 1) correctly
a priori; and another scenario where the strengths of belief are big
using only the truly uncensored observations (“Uncensored”), and
enough to keep memory of the a priori—which we trust—and, in
2) taking into consideration the entire data set, ignoring the pres-
particular, to influence those areas in the data with less observa-
ence of censoring (“Whole”). While this second approach is not
tions. As said, the a priori can indeed be used to complement the
correct (Klein and Moeschberger, 2003), it can be heuristically use-
data with information about rarely observed events and trends. In
ful to better understand the impact of right-censoring in the data.
Table 1 the values of the strengths of belief for both scenarios6 are
For example we can clearly see that not considering censoring can
provided. In the following we will refer to them as the “low be-
lead to a serious overestimation of the dependence in the data.
lief” and the “high belief” scenarios, or with the subscripts l and h, Always in Table 2, observe how the B-RUP is able to recover the
when referring to the results obtained with the B-RUP model. For correlation, under both scenarios,7 something not possible when
example, when we write B-RUPl , we refer to the B-RUP estimators using the KM approach. Moreover, observe how the B-RUP, in par-
in the low belief scenario. B-RUPh is defined analogously for the ticular under the low strength of belief, better captures the vari-
high belief scenario. ability of both X and Y .
In order to initialize the Gibbs sampler, we also need to define In Table 3 we show permutation tests performed on the means
an initial sample for A, which is not observable in practice. There- and the variances. Under the null hypothesis of no difference be-
fore, this sample will be “artificial”, and we need a reasonable way tween the obtained estimates and the true analytical values of
of generating it, taking into consideration the characteristics of A, Table 2, the tests never reject the null for the B-RUP estimators
like for example its positiveness. Moreover, the resulting posterior under the low strength of belief scenario, at a standard 5% signif-
distribution will be affected by this artificial sample, and thus in icance level. These permutation tests where performed simulating
practice it is recommended to compute the posterior for several samples of size 103 from both the B-RUP (low and high strength
appropriate initial samples. In our case, we will sample observa- of belief) and the KM marginals. For the mean, the test statistic
tions from the same distribution that we chose for the prior of A, used is the absolute difference between the means, while for the
variance Good’s test (Good, 1994) is employed. Looking at Table 3,

6
Note that we could go as far as to define different strengths of belief for each
7
urn within a RUP (Muliere et al., 2000). In this work, however, we will always as- Although both scenarios give reasonable results, the low belief scenario returns
sume the same strength of belief for each RUP. values closer to the actual ones, being more data-driven, as expected.

179
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

Fig. 3. Contour plot comparing the analytical distribution with the one obtained through the B-RUP model for the low (a) and high (b) belief scenarios. All distributions are
conditioned on the minimum value of the uncensored observations: X min = 34 and Y min = 39.

Fig. 4. Ratio between the annuity values of the joint and independent survival models. The left plot (a) shows the surface obtained using the B-RUP construction with low
strength of belief, while the right plot (b) shows the price ratio using the actual distribution (via the known analytics). Note how the ratio differs more between the two
plots for high initial ages, since, because of the censored observations, the B-RUP fitting is worse in that area. The result is nevertheless quite positive.

Table 2 ticularly sensitive to discrepancies in the tails, and we already said

Comparison of the means, the variances and the correlation of X and Y using the that the B-RUPh puts a non-negligible mass on the right tail, which
uncensored data, the whole data set (wrongly assuming no censoring), the KM es-
is never really updated by the data. Once again we want to stress
timator, the B-RUP estimator and the original analytical solution, respectively.
that this is actually a sign of the power and flexibility of the B-
Uncensored Whole KM B-RUPl B-RUPh Analytical RUP, in case of a credible a priori (clearly not the case in this very
Mean(X) 52.877 50.684 59.811 59.811 59.341 60.005 simple example).
Mean(Y) 56.690 53.088 64.691 64.691 64.227 65.007
Once we have the joint distribution of the pair ( X , Y ), we can
Var(X) 35.498 39.524 53.486 55.755 50.886 59.905
Var(Y) 35.466 46.689 54.775 62.528 56.071 64.840
price annuities using Equation (3). Following Frees et al. (1996), we
Corr(X,Y) 0.470 0.883 – 0.421 0.361 0.400 are interested in the annuity ratio between the price when assum-
ing dependence and the price when assuming independence. For
the independent scenario we simply use the marginals obtained
Table 3 in Fig. 2, while for the dependent scenario we use the joint dis-
Permutation test for the means and the variances of X and Y , using samples of 103 tribution of Fig. 3a. Moreover, according to Equation (3), once we
observations with 104 permutations. The samples are taken from the KM estimator know the joint distribution, the annuity price depends on three
and the B-RUP with low and high strength of belief. In bold we underline the situa-
tions for which the null hypothesis of no difference between the estimates and the
quantities: the ages of the two annuitants at the beginning of the
true values (from the analytical model) is rejected, for a 5% significance level. contract, which we call entry/initial ages, and the interest rate. As
in Frees et al. (1996), we first show the dependence of the annuity
KM B-RUPl B-RUPh
ratio for several initial ages with a fixed interest rate, and then we
mean( X) 0.6612 0.4561 0.3254
vary the interest rate and assume that both annuitants have the
mean(Y ) 0.6256 0.1354 0.5106
var( X) 0.0881 0.1934 0.0193 same age, so that we can still represent the results in a 3D plot.
var(Y ) 0.0239 0.4859 0.0217 In Figs. 4a and 4b, we show the annuity ratio obtained with
the B-RUP construction and the true (analytical) distribution for
an interest rate of r = 0.05. Notice how, in this example, when
we can conclude that the B-RUP with low strength of belief even considering dependence, the price can be up to 7% higher than
beats the KM, which is not able to correctly recover the variance in the independent scenario for some initial ages, especially when
of Y . If we also take into account that the KM approach cannot es- the age difference is large. For smaller age differences, conversely,
timate the correlation, as already discussed for Table 2, the B-RUPl the annuity ratio takes values below the unit. For low initial ages
performance is even more appreciable. for both annuitants, the death event is still far away with high
Further, observe that, since the a priori is considerably differ- probability, and thus the annuity ratio barely moves from the unity
ent from the analytical solution, it is no surprise that, for the high in that area.
strength of belief case, the B-RUPh performances worsen for the Looking at Figs. 4a and 4b, it is precisely for high annuity ra-
variances of both X and Y . As known, the variance is in fact par- tios that the B-RUP estimators differ more from the analytical ones.

180
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

Fig. 5. Ratio between the annuity values of the joint and independent survival models. The left plot (a) shows the surface obtained using the B-RUP construction with low
strength of belief, while the right plot (b) shows the price ratio using the true (analytical) distribution. A visual comparison suggests that the B-RUP definitely performs well.

Table 4 the observation period, thus excluding a very limited number of

Absolute value differences in the annuity ratios obtained with the B-RUP model and parent-child annuities. All in all we have ended up with a data set
the analytical distribution for several initial ages of the annuitants (Fist Annuitant -
of 11,421 male-female couples.
FA, Second Annuitant - SA) with an interest rate of 0.05. Given the small differences,
we can say that the B-RUP is able to recover the “truth” from the data. For each couple, the following information is available: dates of
Age SA
birth, dates of death or ages at the end of observation period, and
20 30 40 50 60 date of contract initiation. Once the observation period ends, we
Age FA
20 0.0003 0.0000 0.0001 0.0001 0.0002 no longer have any information about the individuals, and thus for
30 0.0000 0.0005 0.0001 0.0001 0.0005 those who lived longer than December 31, 1993, the lifetimes are
40 0.0001 0.0000 0.0009 0.0000 0.0008 right-censored.
50 0.0001 0.0003 0.0001 0.0013 0.0011
Only a small portion of all the lifetimes in our data is actu-
60 0.0005 0.0010 0.0024 0.0034 0.0002
ally uncensored, i.e. fully observed. In particular: 10.87% of the
male lifetimes, 3.93% of the female lifetimes, and 1.71% of the
This, again, is a consequence of the censored observations, that do joint (couple) lifetimes. The last 1.71% corresponds to 195 couples,
not allow for a proper estimation due to the lack of information of which 24 died within a one-day period, 59 within one month,
about the right-tail of the distribution.8 In Table 4 we present the and 92 within one year. Therefore, a considerable portion (47%) of
absolute value differences between the B-RUP and true estimators the uncensored couple deaths is either likely due to an accident–
for several initial ages and a 5% interest rate. The results observed which would explain both spouses’ dying within a day–or to some
are clearly in line with what we have just said in terms of age other impactful reason like a highly contagious disease, or a strong
differences. The performances of the B-RUP are definitely satisfac- broken heart syndrome (Jagger and Sutton, 1991). In these situ-
tory. ations, it is expected that the correlation between the annuitants’
Finally, in Figs. 5a and 5b, one can find the annuity ratio as lifetimes is strongly positive, and it actually is. However, as we will
a function of the interest rate for the same initial age for both see shortly, taking only into account the uncensored observations
annuitants. In accordance with the previous results, the annuity highly overestimates said dependence for all the other couples.
ratio is above one for large initial ages, and below one for low To model the data with the B-RUP, we follow the exact proce-
initial ages. Furthermore, this effect seems to be more pronounced dure as before. We start by identifying the target variables, define
for small interest rates, since the curves are steeper than for high their respective priors, and then apply the Gibbs sampler with pre-
interest rates. This tendency is naturally the same for both the B- defined strengths of belief to merge the a priori knowledge with
RUP estimator and the original distribution. the data. The targets are, once again, A, B and C , and the priors we
chose are G A = Poi(60), G B = Poi(15) and G C = Poi(15). With this
5.2. Canadian data set of Frees et al. (1996) choice, the marginals for X and Y are both Poisson distributions
with parameter μ = 75, and the correlation is approximately 0.8.
Let us now consider a real-world application, using a well- This prior setting (from now on, the Poisson scenario) takes into
known data set, initially presented in Frees et al. (1996), and later account the sample moments and correlation one can estimate di-
also studied in the relevant works of Youn and Shemaykin (1999), rectly from the uncensored observations, and that are shown in
Carriere (2000), Youn and Shemaykin (2001) and Luciano et al. the first column of Table 5. For the strengths of belief we will still
(2008), among others. use the scenarios in Table 1. As before, for initializing the Gibbs
The data set contains 14,497 contracts from a large Canadian sampler we will generate observations from the prior we chose for
insurer, and the period of observation runs from December 29, A.
1988, until December 31, 1993. To simplify the interpretation and The decision of choosing distributions whose moments are in
the comparability of our results with the previous works of Frees line with (a part of) the empirical data is to show a possible way
et al. (1996) and Luciano et al. (2008), we have removed same of eliciting the a priori. Clearly, in case of experts’ judgements or
sex contracts and, for every couple, we have kept only one con- other sources of knowledge, one can also elicit completely subjec-
tract, leaving a total of 11,454 contracts. Further, we have ignored tive priors. In Appendix A, we show the impact of changing the a
couples were the annuitants age was less than 40 at the end of priori.
The rest of the section is structured as follows: in Subsection
8
5.2.1 we show the B-RUP estimation of the joint survival function,
Note that this should be seen as an indicator of the importance of choosing a
and in Subsection 5.2.2 the relative annuity calculations; then in
proper prior distribution. Since the data barely contains any information about the
right tail of the distribution, it is up to the expert to give a reasonable insight about Subsection 5.2.3, we compare our results with the copula approach
tail events. one can find for example in Frees et al. (1996).

181
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

Fig. 6. Comparison of the marginals of X and Y using a Poisson prior. In green we show the prior distribution, the Kaplan-Meier estimators in red and the B-RUP solution in
blue. The upper row shows the results for the low belief scenario and the lower row the same for the high belief scenario. All distributions are conditioned on the minimum
value of the uncensored observations. For this example those were X min = 51 and Y min = 46.

Table 5 females are three years younger than males, when the observation
Comparison of the means, variances and correlation of X and Y using: the uncen- period starts. This combination of younger entry age plus higher
sored data, the whole data set (wrongly assuming no censoring), the KM estimator
life expectancy could be the reason why censoring affects more
and the B-RUP estimator for both low and high strength of belief, and the copula
model, respectively. the distribution of females than that of males for this particular
data set.
Unc. Whole KM Poil Poih
Some additional numbers can give insight about the lack of
Mean(X) 74.514 71.710 86.769 85.183 85.801 data in the tails. Only 48 females have indeed survived until at
Mean(Y) 74.011 68.871 89.625 85.509 87.108
Var(X) 52.049 42.821 75.553 57.556 74.305
least 90 years. Conversely, the number of males observed in the
Var(Y) 61.707 54.118 56.316 37.034 54.947 same range is 77. While this difference may not seem significant,
Corr(X,Y) 0.820 0.776 – 0.477 0.421 notice that the life expectancy of males, according to the results in
Table 5, is 3 years lower than that of women. Therefore one would
expect to observe more women at high ages than men, whereas
5.2.1. B-RUP results for the joint survival function the opposite happens in our data. This could also be explained by
Fig. 6 shows the B-RUP marginal distributions for both the low the lower entry ages of females with respect to males.
and high strengths of belief. We observe that the marginal distri- Table 5 presents the means and the variances of X and Y , and
bution for the first annuitants (males) is nicely recovered. However, their correlation, for the raw uncensored data, the KM estimator
the differences are clearly bigger9 for the marginal of Y –the sec- (notice that in this case correlation is not obtained because inde-
ond annuitant (females)–since censoring seems to affect more the pendence is assumed), and the B-RUP. For completeness, we also
right-tail. provide all quantities when taking all observations as uncensored
The lack of information about the right-tail of the distribution (“Whole”), as we did in the analytical example.
of female lifetimes is so large that the KM cumulative probability First notice that using only uncensored observations leads to an
at the age of 97 is around 0.8, while for the age of 98 the cu- underestimation of the average lifetimes and to a likely overesti-
mulative distribution function (cdf) reaches its maximum value of mation of dependence. But this is no surprise, as it is known that
1. Taken literally, this would mean that there is a 20% probabil- ignoring censoring is highly misleading in survival analysis (Klein
ity for females of dying between 97 and 98 years, something we and Moeschberger, 2003).
know is not true. This unrealistic behavior could be due to the fact In Figs. 7a and 7b we show the joint distributions obtained
that, as per Table 5, female annuitants have a slightly longer life with the B-RUP model for the low and high belief scenarios, re-
expectancy than males, while, from the initial ages at which they spectively, as well as the prior distribution for comparison pur-
enter the contract, we see that on average (Luciano et al., 2008) poses. Notice how, although their marginals are very similar–see
Fig. 6–the shape of the contours is considerably different around
the areas of lower probability. Conversely the joints are more com-
9
We observe that the differences are slightly smaller for the high strength of
patible around the inner contours, for the bulk of the distribution,
belief. This could be due to the fact that a strong a priori helps filling the gaps in sign that the strength of belief is not large enough to affect the ar-
the right tail of the marginals, which, because of censoring, are highly unrealistic. eas where data are concentrated. In particular, for high strengths of

182
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

Fig. 7. Contour plot showing the joint distribution obtained via the B-RUP model for the low (a) and high (b) belief scenarios and their respective densities via bivariate
kernel estimation ((c) and (d)) using a Poisson prior. All distributions are conditioned on the minimum value of the uncensored observations, i.e. Xmin = 51 and Y min = 46.

Fig. 8. Surface plot showing the relative difference between the joint survival distribution and the product of the marginal survival distributions (independence) for the low
(a) and high (b) belief scenarios.

belief the contours are somewhat smoother, a phenomenon which symmetric than those of the low belief kernel estimate, in accor-
we already observed in the analytical example. This makes the dis- dance with the results of Figs. 7a and 7b.
tribution of Fig. 7b easier to interpret. Nevertheless, notice that Notice that, by increasing the strength of belief, the areas with
this smoothing induced by the strength of belief is a secondary fewer observations are dominated by the prior, and this causes a
role, while the main purpose of a reliable prior is to embed behav- decline in the correlation, which goes from 0.477 to 0.421, consis-
iors that are not captured by the data. Moreover, whereas further tently with the findings in Table 5.
increasing the strength of belief would surely result in smoother As observed in Carriere (2000) for this very same data set–
distributions, we could end up giving a larger weight to the prior and more recently in Sanders and Melenberd (2016) for a data set
distribution than originally intended. concerning the whole Dutch population–the positive dependence
A way to improve our previous results, obtaining a more between the lifetimes of married couples increases with the ages
tractable joint distribution, while maintaining the desired strength of its members. In other words, people in their forties are expected
of belief, is to use a kernel bivariate density estimation. For ex- to present a smaller positive dependence than a couple in their
ample, we can employ a bivariate normal kernel, with bandwidth eighties. To check for this phenomenon, as in Sanders and Melen-
parameters consistent with Sylverman’s rule of thumb (Sylverman, berd (2016) we compute the relative difference between the joint
1986). The obtained distributions are presented in Figs. 7c and 7d, survival function and the product of the marginal survival distri-
for low and high strengths of belief, respectively. Note that the butions. The results are in Figs. 8a and 8b for the low and high
contours of the high belief kernel estimate are considerably more belief scenarios, respectively, and they are clearly in line with an

183
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

Fig. 9. Ratio between the annuity values of the joint and independent survival models using the B-RUP estimators. The left plot (a) shows the dependency with the initial
age of the annuitants an interest rate of 0.05 for low strength of belief. The right plot (b) shows similar results for high strength of belief.

Fig. 10. Ratio between the annuity values of the joint and independent survival models using the B-RUP with different strengths of belief ((a) and (b)).

increase of dependence with age. Moreover, from the same figures more relevant in the near future than in the long term—according
we can conclude that the relative difference is bigger under the to the meaning of “high” entry ages, there is no long term survival
low belief scenario, which would explain the slight increase in the to be expected in the first place—and the annuity ratio keeps on
overall correlation (Table 5). decreasing below the unit. All these findings are in line with the
results obtained in Frees et al. (1996) for the same data set with
5.2.2. Annuity pricing with the B-RUP copula models.
Now that we have estimated the joint probability distribution, Also notice the similarity of these surfaces with those of Fig. 4:
we can calculate the annuity ratio. We still assume a fixed interest qualitatively the behavior is mostly the same. This is because in
rate of 0.05 and show the annuity ratio for several initial ages of both cases we have positive dependence, and thus all the reason-
the annuitants. Next, when we want to compute the annuity ratio ing developed in this section also applies for the analytical exam-
as a function of the interest rate, we will hypothesize that annui- ple. Moreover, since the correlation levels are very similar (around
tants enter the contract at the same age to simplify the exposition, 0.5), we obtain annuity ratios of comparable orders.
so that we can easily plot the results as a function of the interest
The results of the annuity ratio as a function of the interest
rate only.
rate can be found in Fig. 10. Please notice how the overall shape is
Fig. 9 shows the results for the low and high strengths of be-
in accordance with what we already observed in Fig. 5. For small
lief, and the fixed interest rate. In line with expectations, even for
interest rates and entry ages, the annuity ratio is below one, mean-
the annuity ratio, the high belief scenario gives again a smoother
ing that the annuity price with independent mortality is overesti-
surface.
mated. As we increase the entry age, the annuity ratio decreases
Notice that, in both cases, the annuity ratio increases with the
up to a minimum (0.9527 for and interest rate equal to 0 and an
age difference between the annuitants. When one of the annui-
tants dies at a very old age, the probability for the surviving spouse entry age of 71 years), from where it starts to increase again, also
to also die at an old age is larger under positive correlation than reaching values above the unit.
under independence. This effect seems to be bigger when the fe- We would like to stress that, while annuity ratios of 1.02 or 0.98
male is the surviving annuitant, probably because of the larger life may seem close enough to 1, and this could lead to the conclusion
expectancy of females with respect to males (Carriere, 2000). that it is ok to assume independence, the nominal value for an
When both spouses are around the same age, the annuity ra- annuity contract is usually considerably large, and thus the final
tio takes values below one, specially for very high entry ages. If difference in the prices may not be negligible in monetary terms.
one annuitant dies at a young age, the probability of the remain- In Table 6 we show some values of the annuity ratio for particular
ing annuitant to also die in the near future is higher when positive entry ages.
dependence is assumed, but in the long term the surviving annu- The results above are in line with what Frees et al. (1996) and
itant will count as being independent from the deceased spouse, Luciano et al. (2008) found on the same data, using different ap-
and the annuity ratio is closer to 1. Conversely, for high entry ages, proaches like copulas, suggesting that the B-RUP can be safely used
the impact of the deceased spouse upon the surviving annuitant is as an alternative model.

184
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

Fig. 11. Comparison of the marginals of X and Y using a Poisson prior. In red we show the Kaplan-Meier estimators and the copula distribution in black. All distributions
are conditioned on the minimum value of the uncensored observations. For this example those were Xmin = 51 and Y min = 46.

Table 6 Table 7
Annuity ratio obtained with the B-RUP model with low strength of belief parame- Calibration of the copula model via MLE. The subscripts X , Y refer to the marginals
ters for several initial ages of the annuitants with an interest rate of 0.05. Assuming of the male and female annuitants, respectively. Notice that α is negative and away
independence can definitely be a bad choice in pricing. from 0, so according to Equation (12) there is positive dependence.
Y μX σX μY σY α
40 50 60 70 80
X
88.783 5.927 90.118 5.145 −4.144
40 0.993 0.996 0.999 1.001 1.007
50 0.996 0.989 0.992 1.000 1.012
60 0.999 0.993 0.979 0.988 1.019
70 1.001 1.001 0.990 0.966 1.009 Table 8
80 1.007 1.013 1.021 1.015 0.977 Comparison of the means, variances and correlation of X and Y using: the uncen-
sored data, the whole data set (wrongly assuming no censoring), the KM estimator
and the copula model, respectively.
In the next Subsection we compare the B-RUP and a copula Unc. Whole KM Copula
approach in some more detail.
Mean(X) 74.514 71.710 86.769 85.427
Mean(Y) 74.011 68.871 89.625 87.167
5.2.3. What about copulas? Var(X) 52.049 42.821 75.553 55.199
We finally compare the performance of the B-RUP with copu- Var(Y) 61.707 54.118 56.316 43.090
las. In particular, we choose the Frank copula model of Frees et al. Corr(X,Y) 0.820 0.776 – 0.512

(1996), which proved to give good results for the annuity problem.
In Frees et al. (1996), Gompertz distributions are used for the
individual lifetimes, and the Frank copula to model the depen-
dence. The Gompertz distribution is given by
μ x
Gomp(x; μ, σ ) = 1 − exp e − σ (1 − e σ ) , (11)

where μ, σ are the location and scale parameter, respectively.

The Frank copula is defined as

1 (e α u − 1)(e α v − 1)
C (u , v ; α ) = log 1 + , (12)
α eα − 1
where u , v are the marginal distributions for the male and female
annuitant, respectively, and α is the parameter controlling the de- Fig. 12. Joint lifetime distribution for the copula construction of Frees et al. (1996).
pendence. A negative value of α indicates positive dependence,
while α = 0 means that we have independence (Nelsen, 2006).
It was already shown in Frees et al. (1996) that this model can earlier, for this particular data set there is barely any information
be calibrated via maximum likelihood estimation (MLE), and we about the right tail of the females’ distribution, and so there is no
refer to the original paper for all the details regarding the estima- surprise in models yielding slightly different results.
tion of the parameters. In Table 7 we present the results obtained In Fig. 12 we show the joint distribution given by the copula,
using MLE, where (μ X , σ X ) are the Gompertz estimates for the with the B-RUP results already presented in Fig. 7. As expected,
male annuitant, and (μY , σY ) the estimates for the female annui- the parametric copula produces a less “cloudy” and elliptical dis-
tant. Since the value of α is highly negative, we expect a strong tribution, with less mass on “peripheral” couples like (70, 85). The
positive dependence. It can be seen in Table 8 that this corre- copula also gives a slightly stronger correlation (0.51 versus 0.48),
sponds to a Pearson correlation of 0.51. but we can consider this difference as negligible in practice.
In Figs. 11a and 11b we present the marginal distributions for Finally, we check if the small discrepancies in the survival func-
this model, with also the KM estimator for comparison. If we com- tions of the B-RUP and the copula model are also observed once
pare these plots with those of Fig. 6, we see that both the B-RUP we compute the annuity ratio. The annuity results for the Frank
and the copula model return very similar curves. We can also com- copula are in Figs. 13a and 13b. Apart from the areas where both
pare the moments using Tables 5 and 8. For the male annuitants, annuitants are considerably old (> 75) at the time of contract ini-
both the mean and the variance are remarkably similar, while tiation, we see that the B-RUP and copula models tend to give very
there are some discrepancies for female annuitants. As discussed similar results.

185
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

Fig. 13. Ratio between the annuity values of the joint and independent survival models using the copula construction. The left plot (a) shows the annuity ratio for a fixed
interest rate r = 0.05. Similar results when the couples have the same age are presented on the right plot (b).

All in all the B-RUP and the Frank copula of Frees et al. (1996) In terms of performances, the model provides results in a very
show similar performances–but this holds true also for the other reasonable time, spanning from a few minutes to a couple of
models of Frees et al. (1996) and Luciano et al. (2008)–so why hours, depending on the size of the data set and the amount of
should one prefer the B-RUP for annuity pricing? predictions required. In particular, for the Canadian data set con-
While working nicely on data, copulas do not have the same sidered and 104 iterations of the Gibbs sampler, the algorithm took
flexibility of the B-RUP. In particular when a priori knowledge can approximately 1 minute of computing time in a C++ environment
be used to improve fitting, or to deal with specific characteris- and an Intel(R) Core(TM) i7-7700HQ CPU @ 2.80 GHz processor.
tics of the data—like a severe lack of observations in the tails—the In the future, it would be extremely useful to go deeper into
B-RUP clearly has an edge over the other commonly used ap- the study of credible and reliable a prioris for annuity modeling,
proaches. And this without taking into consideration the model by gathering opinions and recommendations from experts. From
risk involved in every parametric choice behind the use of copulas a computational point of view, conversely, it could be meaning-
(Hull, 2015; Mikosch, 2006). This is why we believe that the non- ful to find ways of introducing parallelization in the simulations of
parametric approach the B-RUP proposes, which combines a priori the B-RUP. While not immediately relevant to the application de-
knowledge (when meaningful) and the ability of learning from the scribed in this paper, such a possibility could dramatically expand
data, represents a viable and powerful alternative to the existing the applicability and the performances of reinforced urn models
approaches to the modeling of coupled lifetimes. The fact that one in insurance and finance (Amerio et al., 2004; Cheng and Cirillo,
can get the same results, but also extend them with experts’ judge- 2018, 2019; Peluso et al., 2015).
ments (eliciting a different a priori as in Appendix A) is a plus we
deem worth using. Declaration of competing interest

6. Conclusions There is no competing interest.

In this paper we have proposed the bivariate reinforced urn Acknowledgements

process (B-RUP) as a way of modeling dependent mortality, to
price joint and survivor annuities. First of all, the authors would like to thank Prof. Edward W.
The main advantages of the B-RUP lie in its intuitive con- Frees for sharing his Canadian insurance data with us.
struction, in the possibility of combining experts’ judgements with The help of two anonymous referees and the editor in improv-
empirical evidence, in the ability of the model to learn and im- ing the first version of the manuscript is also highly appreciated.
prove its performances over time, like in many machine learning Finally, this research has been financed by the European Union,
approaches, but without “black boxes” (Knight, 2017), and in the under the H2020-EU.1.3.1. MSCA-ITN-2018 scheme, thanks to Grant
successful treatment of right-censored observations, very common 813261.
in annuity modeling.
In the absence of a credible a priori, which would give a strong Appendix A. Changing the prior
competitive edge to its use, the B-RUP is nevertheless able to repli-
cate the performances of other commonly used approaches (e.g. In this section we include additional results about the impact of
copulas as in Frees et al. (1996) or Luciano et al. (2008)), thus different priors on the joint lifetime distribution for the Canadian
showing an interesting flexibility. Differently from other models, data set of Section 5.2. These results are meant to be compared
however, the B-RUP can be used on a continuous basis, as it auto- with the Poisson prior and the copula model of Section 5.2.
matically updates its parameters whenever new data become avail- We consider two alternative settings:
able, without the necessity of re-estimating the model entirely.
This can be extremely useful in an online learning environment. • G A = Gomp(50, 6), G B = Gomp(30, 2) and G C = Gomp(30, 3),
Finally, being nonparametric, the B-RUP is less subject to model where Gomp(μ, σ ) denotes a Gompertz distribution (see
risk than copulas or similar approaches (Hull, 2015), especially if a Equation (11)). We call this the Gompertz scenario.
clever use of priors can also account for data problems. • G A = Uni(85), G B = Uni(20) and G C = Uni(20), where Uni(μ)
Using artificial data and a well-known Canadian data set of an- denotes a discrete uniform distribution with support [0, μ].
nuities (Frees et al., 1996), we have discussed the performances of This is the Uniform scenario.
the B-RUP, which appear definitely satisfactory. In analyzing those
performances, we have also discussed how a sufficient number of The results obtained using the Gompertz scenario are in Fig. 14,
data can correct a wrong a priori, but also how strong priors may and they are very similar to the Poisson case. On the other hand,
try to correct for historical bias and deal with tail events. the results obtained with the Uniform scenario, shown in Fig. 15,

186
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

Fig. 14. Comparison of the marginals of X and Y using a Gompertz prior. In green we show the prior distribution, the Kaplan-Meier estimators in red and the B-RUP solution
in blue. The upper row shows the results for the low belief scenario and the lower row the same for the high belief scenario. All distributions are conditioned on the
minimum value of the uncensored observations. For this example, X min = 51 and Y min = 46.

Fig. 15. Comparison of the marginals of X and Y using an Uniform prior. In green we show the prior distribution, the Kaplan-Meier estimators in red and the B-RUP solution
in blue. The upper row shows the results for the low belief scenario and the lower row the same for the high belief scenario. All distributions are conditioned on the
minimum value of the uncensored observations. Here X min = 51 and Y min = 46.

187
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

Table 9 Derbyshire, J., 2017. The siren call of probability: dangers associated with using
Comparison of the means, variances and correlation of X and Y using: the uncen- probability for consideration of the future. Futures 88, 43–54. https://doi.org/
sored data, the whole data set (wrongly assuming no censoring), the KM estimator 10.1016/j.futures.2017.03.011.
and the B-RUP estimator with Gompertz and Uniform priors, respectively. The sub- Doksum, K., 1974. Tailfree and neutral random probabilities and their posterior
script l (h) indicates low (high) strength of belief. distributions. Annals of Probability 2 (2), 183–201. https://doi.org/10.1214/aop/
1176996703.
Unc. Whole KM Gompl Gomph Unil Unih
Elandt-Johnson, R.C., Johnson, N.L., 1980. Survival Models and Data Analysis. John
Mean(X) 74.514 71.710 86.769 85.907 85.865 87.457 90.819 Wiley. ISBN 9780471031741.
Mean(Y) 74.011 68.871 89.625 84.950 88.208 89.696 94.295 Ferguson, T.S., 1973. A Bayesian analysis of some nonparametric problems. The An-
Var(X) 52.049 42.821 75.553 65.519 67.886 106.766 208.802 nals of Statistics 1 (2), 209–230. https://doi.org/10.1214/aos/1176342360.
Var(Y) 61.707 54.118 56.316 45.047 64.267 91.114 217.951 Fortini, S., Petrone, S., 2012. Hierarchical reinforced urn processes. Statistics & Prob-
Corr(X,Y) 0.820 0.776 – 0.475 0.414 0.611 0.810 ability Letters 82 (8), 1521–1529. https://doi.org/10.1016/j.spl.2012.04.012.
Frees, E.W., Carriere, J., Valdez, E., 1996. Annuity valuation with dependent mortal-
ity. The Journal of Risk and Insurance 63 (2), 229–261. https://doi.org/10.2307/
253744.
are significantly different, especially for high strengths of belief–
Good, P., 1994. Permutation, Parametric and Bootstrap Tests of Hypotheses. Springer,
see Figs. 15c and 15d–where the B-RUP and KM marginals have New York.
very little in common. Trivially, as one grows older the probability Gribkova, S., Lopez, O., 2015. Non-parametric copula estimation under bivariate cen-
of demise increases, and therefore imposing a uniform distribu- soring. Scandinavian Journal of Statistics 42 (4), 925–946. https://doi.org/10.
tion for the lifetime of an individual is highly unrealistic, so this 1111/sjos.12144.
Hull, J.C., 2015. Risk Management and Financial Institutions, 4th edition. Wiley, New
mismatch was actually to be expected. In this sense, the B-RUP
York. ISBN 9781119448112.
inherits the pros and the cons of the prior setting used, for sure Jagger, C., Sutton, C.J., 1991. Death after marital bereavement—is the risk increased?
under high strengths of belief. Naturally and conversely, for low Statistics in Medicine 10 (3), 395–404. https://doi.org/10.1002/sim.4780100311.
strengths of belief, the B-RUP is able to capture both marginals, Kaplan, E.L., Meier, P., 1958. Nonparametric estimation from incomplete observa-
tions. Journal of the American Statistical Association 53 (282), 457–481. https://
except for the areas with very few observations, so that the model
doi.org/10.2307/2281868.
is able to override a wrong prior given a sufficient amount of ob- Klein, J.P., Moeschberger, M.L., 2003. Survival Analysis Techniques for Censored and
servations. Truncated Data, second edition.
In Table 9 we show the first two moments and the correla- Knight, W., 2017. The dark secret at the heart of AI. Technology Review 120, 54–63.
tion, similar to what we did in Table 5 for the Poisson prior. As Lin, D.Y., 1993. A simple nonparametric estimator of the bivariate survival function
under univariate censoring. Biometrika 80, 573–581. https://doi.org/10.2307/
before, the subscript l (h) denotes low (high) strength of belief.
2337178.
Notice that, for low strengths of belief, the difference in the B-RUP Lopez, O., 2012. A generalization of Kaplan-Meier estimator for analyzing bivariate
results are mainly due to the choice of the initial sample for A in mortality under right-censoring and left-truncation with applications to model-
the Gibbs sampler, which in our case is generated from its prior checking for survival copula models. Insurance. Mathematics & Economics 51
(3), 505–516. https://doi.org/10.1016/j.insmatheco.2012.07.009.
distribution. Consistently with what we have already said, the uni-
Luciano, E., Spreeuw, J., Vigna, E., 2008. Modelling stochastic mortality for depen-
form scenario with a high strength of belief generates the most dent lives. Insurance. Mathematics & Economics 43, 234–244. https://doi.org/10.
radical results, including a correlation close to the one (0.8) we 1016/j.insmatheco.2008.06.005.
can estimate from the uncensored observations. All the other B- Mahmoud, H., 2008. Polya Urn Models. Chapman & Hall/CRC, Boca Raton.
RUPs tend to suggest a less extreme correlation, in the vicinity of ISBN 9781420059830.
Mikosch, T., 2006. Copulas: tales and facts. Extremes 9, 3–20. https://doi.org/10.
0.5, in line with Frees et al. (1996) and Luciano et al. (2008). In
1007/s10687-006-0015-x.
any case, a correlation of 0.5 should definitely not be ignored by Mitchell, O., Poterba, J.M., Warshawsky, M., Brown, J., 1999. New evidence on the
assuming independence, as per the KM approach. money’s worth of individual annuities. The American Economic Review 89 (5),
1299–1318. http://www.jstor.org/stable/117059.
Muliere, P., Secchi, P., Walker, S.G., 2000. Urn schemes and reinforced random walks.
References
Stochastic Processes and Their Applications 88 (1), 59–78. https://doi.org/10.
1016/S0304-4149(99)00119-2. ISSN 03044149.
Amerio, E., Muliere, P., Secchi, P., 2004. Reinforced urn processes for modeling credit Muliere, P., Secchi, P., Walker, S.G., 2003. Reinforced random processes in continuous
default distributions. International Journal of Theoretical and Applied Finance 07 time. Stochastic Processes and Their Applications 104 (1), 117–130. https://doi.
(04), 407–423. https://doi.org/10.1142/S0219024904002505. org/10.1016/S0304-4149(02)00234-X.
Blackwell, D., MacQueen, J.B., 1973. Ferguson distributions via Pólya urn schemes. Murphy, K.P., 2012. Machine Learning: A Probabilistic Perspective. The MIT Press,
The Annals of Statistics 1 (2), 353–355. https://doi.org/10.1214/aos/1176342372. Cambridge MA. ISBN 0262018020. 9780262018029.
Brown, J.R., Poterba, J.M., 2000. Joint life annuities and annuity demand by married Nelsen, R.B., 2006. An Introduction to Copulas. Springer, New York. http://www.
couples. The Journal of Risk and Insurance 74 (4), 527–553. https://doi.org/10. worldcat.org/search?qt=worldcat_org_all&q=0387286594.
2307/253849. Peluso, S., Mira, A., Muliere, P., 2015. Reinforced urn processes for credit risk models.
Bulla, P., 2005. Application of Reinforced Urn Processes to Survival Analysis. PhD Journal of Econometrics 184 (1), 1–12. https://doi.org/10.1016/j.jeconom.2014.
thesis. Bocconi University. 08.003.
Bulla, P., Muliere, P., Walker, S., 2007. Bayesian nonparametric estimation of a bi- Pruitt, R.C., 1991. On negative mass assigned by the bivariate Kaplan-Meier
variate survival function. Statistica Sinica 17 (2), 427–444. http://www.jstor.org/ estimator. The Annals of Statistics 19, 443–453. https://doi.org/10.1214/aos/
stable/24307726. 1176347992.
Carriere, J.F., 2000. Bivariate survival models for coupled lives. Scandinavian Actuar- Sanders, L., Melenberd, B., 2016. Estimating the joint survival probabilities of mar-
ial Journal 1, 17–32. https://doi.org/10.1080/034612300750066700. ried individuals. Insurance. Mathematics & Economics 67, 88–106. https://doi.
Cheng, D., Cirillo, P., 2018. A reinforced urn process modeling of recovery rates and org/10.1016/j.insmatheco.2015.12.006.
recovery times. Journal of Banking & Finance (ISSN 0378-4266) 96, 1–17. http:// Shackle, G.L.S., 1955. Uncertainty in Economics and Other Reflections. Cambridge
www.sciencedirect.com/science/article/pii/S037842661830178X. University Press, Cambridge. ISBN 9780521153317.
Cheng, D., Cirillo, P., 2019. A urn-based nonparametric modeling of the dependence Shen, P.-S., Yan, Y.-F., 2008. Nonparametric estimation of the bivariate survival func-
between PD and LGD with an application to mortgages. https://doi.org/10.2139/ tion with left-truncated and right-censored data. Journal of Statistical Planning
ssrn.3360531. and Inference 138 (12), 4041–4054. https://doi.org/10.1016/j.jspi.2008.02.007.
Cirillo, P., Hüsler, J., Muliere, P., 2013. Alarm systems and catastrophes from a di- Sheshinski, E., 2007. The Economic Theory of Annuities. Economics Books, vol. 8536.
verse point of view. Methodology and Computing in Applied Probability 15 (4), Princeton University Press. ISBN 9780691133058.
821–839. Sylverman, B.W., 1986. Density Estimation for Statistics and Data Analysis. Chapman
Cox, D.R., Oakes, D., 1984. Analysis of Survival Data. Chapman & Hall. & Hall/CRC. ISBN 9780412246203.
ISBN 9780412244902. Taleb, N.N., 2007. The Black Swan: The Impact of the Highly Improbable. Random
Dabrowska, D.M., 1988. Kaplan-Meier estimate on the plane. The Annals of Statis- House. ISBN 978-0812973815.
tics 16 (4), 1475–1489. https://doi.org/10.1214/aos/1176351049. Walker, S., Muliere, P., 1997. Beta-Stacy processes and a generalization of the Pólya-
de Finetti, B., 2017. Theory of Probability: A Critical Introductory Treatment. Wiley urn scheme. The Annals of Statistics 25 (4), 1762–1780. https://doi.org/10.1214/
Series in Probability and Statistics. Wiley. ISBN 9781119286370. aos/1031594741.

188
L.A. Souto Arias and P. Cirillo Insurance: Mathematics and Economics 99 (2021) 174–189

Wang, W., Wells, M.T., 1997. Nonparametric estimators of the bivariate survival func- STATISTICAL-ASPECTS-OF-JOINT-LIFE-INSURANCE-PRICING-Youn-Shemyakin/
tion under simplified censoring conditions. Biometrika 84, 863–880. https:// a77552c3400d4d49cd97777bea63524f54299b64.
doi.org/10.1093/biomet/84.4.863. Youn, H., Shemaykin, A., 2001. Pricing practices for joint last survivor
Winklevoss, H., 1977. Pension Mathematics with Numerical Illustrations. insurance. In: Actuarial Research Clearing House, 2001.1. https://
Pension Research Council, Wharton School, University of Pennsylvania. www.soa.org/globalassets/assets/library/research/actuarial-research-clearing-
ISBN 9780812231960. house/2000-09/2001/arch-1/arch01v19.pdf.
Youn, H., Shemaykin, A., 1999. Statistical aspects of joint life insurance pricing. Youn, H., Shemyakin, A., Herman, E., 2002. A re-examination of the joint mortality
In: 1999 Proceedings of the Business and Statistics Section of the Ameri- functions. North American Actuarial Journal 6 (1), 166–170. https://doi.org/10.
can Statistical Association, pp. 34–38. https://www.semanticscholar.org/paper/ 1080/10920277.2002.10596035.

189