Introduction
Predicting outstanding insurance claims requires both a forecast of future payments and a quantification of the uncertainty in this forecast. Mack’s [11 ] chain-ladder (CL) method is one of the early derivations that quantified the prediction uncertainty of the CL reserving method in terms of the conditional mean squared error of prediction (MSEP). Mack’s [11 ] uncertainty estimate was derived in a distribution-free stochastic version of the CL method, and still today it is the state-of-the-art in claims reserving and its risk management implications. The broader literature considers alternative distributional assumptions, estimation procedures and predictive models; see, e.g., England–Verrall [8 ] and Wüthrich–Merz [20 ] . Within this literature, an important modeling choice is how information is shared across accident years when estimating a development pattern. The classical CL specification uses a common expected development factor for each development period, which can be seen as a mean-stationarity property. In practice, however, this mean-stationarity property is often violated, as external factors and portfolio changes may lead to trends in the development pattern. To cope with such trends, practitioners often use a rolling window estimation approach, meaning that instead of considering all past observations, they only use the more recent ones (in a given time window). This is illustrated in Figure 1 : the left-hand side (lhs) of that figure considers all past observations indicated by the orange trapezoid, whereas a rolling window approach only considers a recent diagonal slab, as shown on the right-hand side (rhs) of Figure 1 .
Figure 1. (lhs) Available information at time I I I , (rhs) observations selected for model fitting.
This rolling window approach seems problematic for two reasons. First, it neglects the observations in the upper-left part of the observed claims array. Second, the need for considering an estimation window indicates that the model assumptions are violated, thus, one works in an inappropriate stochastic reserving setting. Below, we consider a full stochastic model that results in an exponential seniority weighting approach, see (4.10), that is, the optimal CL factors are in our stochastic model given by
f ^ j ( I ) = ∑ i = 1 I − j − 1 α i , j C i , j + 1 C i , j with α i , j = q I − j − 1 − i C i , j ∑ k = 1 I − j − 1 q I − j − 1 − k C k , j . (1) \hat{f}_{j}^{(I)} = \sum_{i=1}^{I-j-1} \alpha_{i,j} \frac{C_{i,j+1}}{C_{i,j}}
\qquad\text{with}\qquad
\alpha_{i,j} = \frac{q^{I-j-1-i} C_{i,j}}{\sum_{k=1}^{I-j-1} q^{I-j-1-k} C_{k,j}}. \tag*{(1)} f ^ j ( I ) = i = 1 ∑ I − j − 1 α i , j C i , j C i , j + 1 with α i , j = ∑ k = 1 I − j − 1 q I − j − 1 − k C k , j q I − j − 1 − i C i , j . (1)
The crucial part in this estimation formula is the exponential weighting q I − j − 1 − i q^{I-j-1-i} q I − j − 1 − i for a given parameter q ∈ ( 0 , 1 ] q \in(0,1] q ∈ ( 0 , 1 ] . For the latest observation, we have i = I − j − 1 i = I-j-1 i = I − j − 1 , thus, q I − j − 1 − i = 1 q^{I-j-1-i} = 1 q I − j − 1 − i = 1 ; for the second latest observation i = I − j − 2 i = I-j-2 i = I − j − 2 , thus, q I − j − 1 − i = q q^{I-j-1-i} = q q I − j − 1 − i = q ; etc. For q = 1 q = 1 q = 1 , this yields the classical CL factors of Mack [11 ] , and for q < 1 q < 1 q < 1 , the observations receive an exponential seniority weighting. A related estimation approach has been considered in Marais–Grove [13, formula (4)], however, this proposal is not based on a fully-fledged stochastic reserving model, but it manually imposes this structure by formulating a corresponding estimation objective. We introduce a full stochastic model that allows us to analytically compute the predictors resulting precisely in such an exponential seniority weighting (1.1), i.e., this exponential estimation procedure is a natural consequence of our model assumptions. This full stochastic model provides us with the advantage of being able to explicitly simulate the conditional MSEP in a correctly specified stochastic reserving environment, and we receive explicit closed-form results for the one-year claims development uncertainty view. In particular, to implement a weighting approach (1.1) in a state-space model requires dependence along the accident year axis, which is neither the case in Mack [12 ] , Marais–Grove [13 ] nor in Sloma [15 ] . This literature considers weighting, but the weighting is added manually (somehow retrospectively) and these weights are not the natural product of a full stochastic model.
Furthermore, Marais–Grove [13, Section 4.2] add regularization with a prior development pattern. Our stochastic reserving model admits a similar regularization version because we work in an observation-driven state-space model where regularization is a natural implication from the latent state-space process. A unregularized version is then obtained by letting the latent state-space process be non-informative; this is rather similar to the Bayesian considerations in Wüthrich–Merz [21 Section 2.2] .
Early versions of state-space models for claims reserving were considered by de Jong–Zehnwirth [10 ] and Verrall [17 , 18 ] ; a nice historical overview is given in Chukhrova–Johannssen [6 ] . In a multivariate log-normal context there is full modeling flexibility of the dependence structure along all time directions; see Wüthrich–Merz [21 Chapter 4] . Most of these state-space proposals have in common that they do not use a linear scaling in the conditional variance as proposed and introduced by Mack [11 ] , thus, they do not directly lift Mack’s [11 ] distribution-free CL model to a version that allows for seniority weighting. Chukhrova–Johannssen [6 Model Assumptions 2] lift Mack’s [11 ] CL model to a distribution-free state-space model, which is then solved by applying the Kalman filter. However, their model does not allow for a seniority weighting in the accident year direction because the state-space dynamics evolves along the development year axis, whereas our proposal will move along the accident year axis. The way we are going to process information is similar to Cipra–Hendrych [7 ] , but in contrast to that reference, we process cumulative claims and we ensure that Mack’s [11 ] linear conditional variance assumption is fulfilled, so that Mack’s [11 ] method can be seen as a boundary case of our exponential seniority weighting claims reserving model.
More explicitly, we combine the gamma-gamma Bayes’ CL model of Wüthrich–Merz [21 ] with the observation-driven state-space model of Ahn et al. [2 ] . This yields a full stochastic model that allows us to compute all relevant quantities because we can express the likelihood of the observations in closed form in this model. More specifically, the gamma-gamma Bayes’ CL model considers conjugate priors that allow for explicit computations of posteriors, and they also allow us to incorporate Mack’s linear conditional variance assumption. This model can be marginalized by integrating out the latent states resulting in prime-beta distributions. We then lift this model to a state-space version using the observation-driven state-space dynamics of Ahn et al. [2 ] , which was inspired by the econometrics models of Smith–Miller [16 ] and Harvey–Fernandes [9 ] . This observation-driven state-space model retains analytical tractability, and this precisely makes it possible to derive our results.
Our contributions are the following. We introduce a full stochastic claims reserving model that is analytically tractable. This model allows us to determine the claims reserves, the conditional MSEP of the total run-off and the conditional MSEP of the one-year claims development result. This model has the nice property that it allows us to select the hyper-parameters such that it yields the CL claims reserves with an exponential seniority weighting as a natural result from the model assumptions. Its boundary (special) case are the classical CL reserves, and the resulting uncertainty formulas are lower bounded by Mack’s [11 ] MSEP formula and by the Merz–Wüthrich [14 ] one-year formula; this is in complete analogy to the considerations in Wüthrich–Merz [21 Section 2.2] . To the best of our knowledge, this is the first model where all these quantities can be fully derived including an exponential seniority weighting in the CL factors, i.e., yielding consistent results within a unified stochastic model where the seniority weighting is a natural consequence of the model assumptions, and not a manually enforced one.
Organization. In the next section, we introduce the claims development model, and we discuss basic properties. Section 3 presents the main properties of the model – mean-stationarity of the latent state-space process – and it derives different versions of the CL factors. Section 4 discusses ultimate claims predictions and the special case of exponential seniority weighting in the CL factors. These computations are feasible because we have tractable likelihoods with explicit factorizations. Sections 5 and 6 give the results for the total prediction uncertainty and the one-year uncertainty MSEPs, respectively. Section 7 gives a numerical example on a standard dataset, and in Section 8 we conclude.
The model
Claims prediction and the claims reserving problem
Let C i , j C_{i,j} C i , j denote the cumulative claims in accident year i ∈ { 1 , … , I } i \in\{1,\ldots,I\} i ∈ { 1 , … , I } after development year j ∈ { 0 , … , J } j \in\{0,\ldots,J\} j ∈ { 0 , … , J } , and assume that these cumulative claims are strictly positive. We generally assume I > J I > J I > J . The total claim of accident year i ∈ { 1 , … , I } i \in\{1,\ldots,I\} i ∈ { 1 , … , I } is given by the ultimate claim C i , J C_{i,J} C i , J . This ultimate claim is the main object of interest that we aim to forecast based on the available information set at time I I I , given by the upper trapezoid
D I = σ { C i , j ; i + j ≤ I , 1 ≤ i ≤ I , 0 ≤ j ≤ J } . \mathcal{D}_I = \sigma\{C_{i,j};\ i+j \le I,\ 1 \le i \le I,\ 0 \le j \le J\}. D I = σ { C i , j ; i + j ≤ I , 1 ≤ i ≤ I , 0 ≤ j ≤ J } .
This information set specifies all observations in the complete rows i ≤ I − J i \le I-J i ≤ I − J , and the observed ones in the incomplete rows i > I − J i > I-J i > I − J , i.e., given up to the claims development index
d i = min { J , I − i } , (2) d_i = \min\{J,I-i\}, \tag*{(2)} d i = min { J , I − i } , (2)
this is illustrated in Figure 1 (lhs). Our goal is to compute the best-estimates of the ultimate claims at time I I I for accident years i + J > I i+J>I i + J > I , given by
E [ C i , J ∣ D I ] . \mathbb{E}[C_{i,J} \mid\mathcal{D}_I]. E [ C i , J ∣ D I ] .
The accident years i ≤ I − J i \le I-J i ≤ I − J do not need predictions because they are fully observed at time I I I (d i = J d_i=J d i = J ). Our goal is to compute these best-estimates under given model assumptions. If the cumulative claims C i , j C_{i,j} C i , j correspond to cumulative payments, then the claims reserves for the outstanding loss liabilities of accident years i + J > I i+J>I i + J > I are given by
R i = E [ C i , J ∣ D I ] − C i , d i . R_i = \mathbb{E}[C_{i,J} \mid\mathcal{D}_I] - C_{i,d_i}. R i = E [ C i , J ∣ D I ] − C i , d i .
The starting point to compute the best-estimates and the claims reserves is a Bayesian version of Mack’s CL model [11 ] . Namely, we consider an accident-year dynamic version of the gamma-gamma Bayes CL model of Wüthrich–Merz [21 ] . The accident-year dynamic part uses the observation-driven state-space model ideas of Ahn et al. [2 ] involving latent state-space processes Θ 1 : I , j \Theta_{1:I,j} Θ 1 : I , j for j ∈ { 0 , … , J − 1 } j \in\{0,\ldots,J-1\} j ∈ { 0 , … , J − 1 } , i.e., these latent processes evolve on the accident year axis for fixed development period j j j . Before doing so, we introduce the relevant information sets.
Information sets
We introduce the following notation. Generically, we write for vectors C 1 : i , j = ( C 1 , j , … , C i , j ) C_{1:i,j}=(C_{1,j},\ldots,C_{i,j}) C 1 : i , j = ( C 1 , j , … , C i , j ) and C i , 0 : j = ( C i , 0 , … , C i , j ) C_{i,0:j}=(C_{i,0},\ldots,C_{i,j}) C i , 0 : j = ( C i , 0 , … , C i , j ) , and for 2D tensors C 1 : i , 0 : j = ( C l , k ) l = 1 , … , i ; k = 0 , … , j C_{1:i,0:j}=(C_{l,k})_{l=1,\ldots,i;\,k=0,\ldots,j} C 1 : i , 0 : j = ( C l , k ) l = 1 , … , i ; k = 0 , … , j . An empty sequence C 1 : 0 , j C_{1:0,j} C 1 : 0 , j or an empty 2D tensor C 1 : 0 , 0 : j C_{1:0,0:j} C 1 : 0 , 0 : j generates the trivial σ \sigma σ -field.
We introduce the relevant information sets. We define
B j = σ ( C 1 : I , 0 : j ) for j ∈ { 0 , … , J } , \mathcal{B}_j = \sigma\left(C_{1:I,0:j}\right) \qquad\text{for } j \in\{0,\ldots,J\}, B j = σ ( C 1 : I , 0 : j ) for j ∈ { 0 , … , J } ,
this is the complete rectangle up to development period j j j , and
H i , j = B j ∨ σ ( C 1 : i , j + 1 ) for j ∈ { 0 , … , J − 1 } and i ∈ { 0 , … , I } , (3) \mathcal{H}_{i,j} = \mathcal{B}_j \vee\sigma\left(C_{1:i,j+1}\right) \qquad\text{for } j \in\{0,\ldots,J-1\} \text{ and } i \in\{0,\ldots,I\}, \tag*{(3)} H i , j = B j ∨ σ ( C 1 : i , j + 1 ) for j ∈ { 0 , … , J − 1 } and i ∈ { 0 , … , I } , (3)
this information set is illustrated in Figure 2 . We interpret H i − 1 , j \mathcal{H}_{i-1,j} H i − 1 , j as the history before C i , j + 1 C_{i,j+1} C i , j + 1 has been observed, and H i , j \mathcal{H}_{i,j} H i , j is the history thereafter, i.e., C i , j + 1 C_{i,j+1} C i , j + 1 is H i , j \mathcal{H}_{i,j} H i , j -measurable. These σ \sigma σ -fields do not respect the trapezoidal structure of D I D_I D I , see Figure 2 , but they are necessary to define our claims reserving model. These definitions yield the filtration
Figure 2. Information set H i , j \mathcal{H}_{i,j} H i , j with B j \mathcal{B}_j B j in blue color and σ ( C 1 : i , j + 1 ) \sigma(C_{1:i,j+1}) σ ( C 1 : i , j + 1 ) in yellow color.
⋯ ⊆ H I , j − 1 = B j = H 0 , j ⊆ H 1 , j ⊆ ⋯ ⊆ H I − 1 , j ⊆ H I , j = B j + 1 = H 0 , j + 1 ⊆ ⋯ . \cdots\subseteq\mathcal{H}_{I,j-1} = \mathcal{B}_j = \mathcal{H}_{0,j} \subseteq\mathcal{H}_{1,j} \subseteq\cdots\subseteq\mathcal{H}_{I-1,j} \subseteq\mathcal{H}_{I,j} = \mathcal{B}_{j+1} = \mathcal{H}_{0,j+1} \subseteq\cdots. ⋯ ⊆ H I , j − 1 = B j = H 0 , j ⊆ H 1 , j ⊆ ⋯ ⊆ H I − 1 , j ⊆ H I , j = B j + 1 = H 0 , j + 1 ⊆ ⋯ .
The filtration defined by (2.2) moves column-wise through all the rows starting with column j = 0 j=0 j = 0 and ending with column j = J − 1 j=J-1 j = J − 1 (outputting column J J J ), that is, we complete the yellow column j + 1 j+1 j + 1 in Figure 2 , before moving to the next column by using the identity H I , j = B j + 1 = H 0 , j + 1 \mathcal{H}_{I,j} = \mathcal{B}_{j+1} = \mathcal{H}_{0,j+1} H I , j = B j + 1 = H 0 , j + 1 .
For the latent state-space processes Θ 1 : I , j \Theta_{1:I,j} Θ 1 : I , j , we consider the analogous flow of information
G i , j = σ ( Θ 1 : I , 0 : j − 1 ) ∨ σ ( Θ 1 : i , j ) . \mathcal{G}_{i,j} = \sigma\left(\Theta_{1:I,0:j-1}\right) \vee\sigma\left(\Theta_{1:i,j}\right). G i , j = σ ( Θ 1 : I , 0 : j − 1 ) ∨ σ ( Θ 1 : i , j ) .
The σ \sigma σ -fields H i , j \mathcal{H}_{i,j} H i , j consider the observable cumulative claims, the σ \sigma σ -fields G i , j \mathcal{G}_{i,j} G i , j are auxiliary latent objects (unobservable) that are needed to formulate our observation-driven state-space model. This model is going to be based on gamma distributed random variables X ∼ Γ ( γ , c ) X \sim\Gamma(\gamma,c) X ∼ Γ ( γ , c ) with shape parameter γ > 0 \gamma> 0 γ > 0 and rate parameter c > 0 c > 0 c > 0 resulting in mean E [ X ] = γ / c \mathbb{E}[X] = \gamma/c E [ X ] = γ / c . Moreover, we mostly consider reciprocal gamma random variables with mean E [ X − 1 ] = c / ( γ − 1 ) \mathbb{E}[X^{-1}] = c/(\gamma-1) E [ X − 1 ] = c / ( γ − 1 ) for γ > 1 \gamma> 1 γ > 1 .
Observation-driven gamma-gamma state-space CL model
We introduce a state-space model for every development period j + 1 ∈ { 1 , … , J } j+1 \in\{1,\ldots,J\} j + 1 ∈ { 1 , … , J } . This state-space model is based on a minimal set of assumptions. There will be an extended discussion after these assumptions.
Assumptions 2.1 (Observation-driven gamma-gamma state-space CL model). Fix the following set of deterministic parameters: γ j > 0 \gamma_j > 0 γ j > 0 , f j > 0 f_j > 0 f j > 0 , σ j 2 > 0 \sigma_j^2 > 0 σ j 2 > 0 , p i , j ∈ [ 0 , 1 ] p_{i,j} \in[0,1] p i , j ∈ [ 0 , 1 ] and q i , j > 0 q_{i,j} > 0 q i , j > 0 for i ∈ { 1 , … , I − 1 } i \in\{1,\ldots,I-1\} i ∈ { 1 , … , I − 1 } and j ∈ { 0 , … , J − 1 } j \in\{0,\ldots,J-1\} j ∈ { 0 , … , J − 1 } .
(1) State-space equation. For j ∈ { 0 , … , J − 1 } j \in\{0,\ldots,J-1\} j ∈ { 0 , … , J − 1 } and i ∈ { 1 , … , I } i \in\{1,\ldots,I\} i ∈ { 1 , … , I } , assume
Θ i , j ∣ H i − 1 , j ∼ Γ ( 1 + a i , j − , b i , j − ) (4) \Theta_{i,j} \mid\mathcal{H}_{i-1,j} \sim\Gamma\left(1+a_{i,j}^{-},b_{i,j}^{-}\right) \tag*{(4)} Θ i , j ∣ H i − 1 , j ∼ Γ ( 1 + a i , j − , b i , j − ) (4) with initialization a 1 , j − = γ j > 0 a_{1,j}^{-}=\gamma_j>0 a 1 , j − = γ j > 0 and b 1 , j − = f j γ j > 0 b_{1,j}^{-}=f_j\gamma_j>0 b 1 , j − = f j γ j > 0 , and affine updates for i ∈ { 1 , … , I − 1 } i \in\{1,\ldots,I-1\} i ∈ { 1 , … , I − 1 }
a i + 1 , j − = ( p i , j + q i , j ) a i , j + (5) a_{i+1,j}^{-}=(p_{i,j}+q_{i,j})a_{i,j}^{+} \tag*{(5)} a i + 1 , j − = ( p i , j + q i , j ) a i , j + (5) b i + 1 , j − = p i , j a i , j + f j + q i , j b i , j + , (6) b_{i+1,j}^{-}=p_{i,j}a_{i,j}^{+}f_j+q_{i,j}b_{i,j}^{+}, \tag*{(6)} b i + 1 , j − = p i , j a i , j + f j + q i , j b i , j + , (6) where we defined for i ∈ { 1 , … , I } i \in\{1,\ldots,I\} i ∈ { 1 , … , I }
a i , j + = a i , j − + C i , j σ j 2 > 0 and b i , j + = b i , j − + C i , j + 1 σ j 2 > 0. (7) a_{i,j}^{+}=a_{i,j}^{-}+\frac{C_{i,j}}{\sigma_j^2}>0
\qquad\text{and}\qquad
b_{i,j}^{+}=b_{i,j}^{-}+\frac{C_{i,j+1}}{\sigma_j^2}>0. \tag*{(7)} a i , j + = a i , j − + σ j 2 C i , j > 0 and b i , j + = b i , j − + σ j 2 C i , j + 1 > 0. (7) (2) Observation equation. For i ∈ { 1 , … , I } i \in\{1,\ldots,I\} i ∈ { 1 , … , I } and j ∈ { 0 , … , J − 1 } j \in\{0,\ldots,J-1\} j ∈ { 0 , … , J − 1 } , we assume for the cumulative claims
C i , j + 1 ∣ H i − 1 , j , G i , j = ( d ) C i , j + 1 ∣ C i , j , Θ i , j ∼ Γ ( C i , j σ j 2 , Θ i , j σ j 2 ) . (8) C_{i,j+1}\mid\mathcal{H}_{i-1,j},\mathcal{G}_{i,j}
\overset{\mathrm{(d)}}{=}
C_{i,j+1}\mid C_{i,j},\Theta_{i,j}
\sim\Gamma\left(\frac{C_{i,j}}{\sigma_j^2},\frac{\Theta_{i,j}}{\sigma_j^2}\right). \tag*{(8)} C i , j + 1 ∣ H i − 1 , j , G i , j = ( d ) C i , j + 1 ∣ C i , j , Θ i , j ∼ Γ ( σ j 2 C i , j , σ j 2 Θ i , j ) . (8) Moreover, the initial column C 1 : I , 0 C_{1:I,0} C 1 : I , 0 is always assumed to be fixed and strictly positive.
We discuss Assumptions 2.1:
The conditional law of the cumulative claim C i , j + 1 C_{i,j+1} C i , j + 1 in (8) is fully determined by the latent state Θ i , j \Theta_{i,j} Θ i , j and the previous cumulative claim C i , j C_{i,j} C i , j . It has conditional mean
E [ C i , j + 1 ∣ H i − 1 , j , G i , j ] = E [ C i , j + 1 ∣ C i , j , Θ i , j ] = Θ i , j − 1 C i , j . (9) \mathbb{E}[C_{i,j+1}\mid\mathcal{H}_{i-1,j},\mathcal{G}_{i,j}]
=
\mathbb{E}[C_{i,j+1}\mid C_{i,j},\Theta_{i,j}]
=
\Theta_{i,j}^{-1}C_{i,j}. \tag*{(9)} E [ C i , j + 1 ∣ H i − 1 , j , G i , j ] = E [ C i , j + 1 ∣ C i , j , Θ i , j ] = Θ i , j − 1 C i , j . (9)
Thus, the reciprocal latent state Θ i , j − 1 \Theta_{i,j}^{-1} Θ i , j − 1 plays the role of an individual CL factor . Moreover, we obtain Mack’s [11 ] linear conditional variance assumption for given individual CL factors
Var ( C i , j + 1 ∣ H i − 1 , j , G i , j ) = Var ( C i , j + 1 ∣ C i , j , Θ i , j ) = σ j 2 Θ i , j − 2 C i , j , (10) \operatorname{Var}(C_{i,j+1}\mid\mathcal{H}_{i-1,j},\mathcal{G}_{i,j})
=
\operatorname{Var}(C_{i,j+1}\mid C_{i,j},\Theta_{i,j})
=
\sigma_j^2\Theta_{i,j}^{-2}C_{i,j}, \tag*{(10)} Var ( C i , j + 1 ∣ H i − 1 , j , G i , j ) = Var ( C i , j + 1 ∣ C i , j , Θ i , j ) = σ j 2 Θ i , j − 2 C i , j , (10)
with variance parameter σ j 2 Θ i , j − 2 > 0 \sigma_j^2\Theta_{i,j}^{-2}>0 σ j 2 Θ i , j − 2 > 0 .
In Lemma 3.1 , below, we see that Θ 1 : I , j − 1 \Theta_{1:I,j}^{-1} Θ 1 : I , j − 1 is mean-stationary in accident year i i i with E [ Θ i , j − 1 ] = f j \mathbb{E}[\Theta_{i,j}^{-1}]=f_j E [ Θ i , j − 1 ] = f j for all i ∈ { 1 , … , I } i \in\{1,\ldots,I\} i ∈ { 1 , … , I } . Thus, each individual CL factor in (9) has baseline mean f j > 0 f_j>0 f j > 0 . We are going to study below how this baseline assessment changes through Bayesian inference.
As highlighted in Ahn et al. [2 Remark 2] , there exist infinitely many stochastic models that comply with Assumptions 2.1, e.g., if one uses the gamma-thinning construction presented in Ahn et al. [1 ] , one receives a full stochastic model that has the affine parameter structure (5) –(7) under the additional assumption of p i , j q i , j ∈ ( 0 , 1 ) p_{i,j}q_{i,j}\in(0,1) p i , j q i , j ∈ ( 0 , 1 ) . We emphasize that Assumptions 2.1 do not uniquely define a model, i.e., the state-space update (4) does not uniquely define the state-space dynamics, but rather a large model class thereof. In particular, it allows for different covariance structures in the reciprocal latent state-space processes Θ 1 : I , j − 1 \Theta_{1:I,j}^{-1} Θ 1 : I , j − 1 ; see Ahn et al. [2 , 3 ] . This is further discussed below.
Filtering, credibility, and mean-stationarity
We discuss the different types of CL factors that result from Assumptions 2.1. For this, we introduce the following notation
f i , j − = b i , j − a i , j − , f i , j + = b i , j + a i , j + , ω i , j = C i , j / σ j 2 a i , j − + C i , j / σ j 2 ∈ ( 0 , 1 ) , and Δ i , j = q i , j p i , j + q i , j ∈ ( 0 , 1 ] . f^{-}_{i,j} = \frac{b^{-}_{i,j}}{a^{-}_{i,j}}, \quad f^{+}_{i,j} = \frac{b^{+}_{i,j}}{a^{+}_{i,j}}, \quad\omega_{i,j} = \frac{C_{i,j}/\sigma_j^2}{a^{-}_{i,j} + C_{i,j}/\sigma_j^2} \in(0,1), \quad\text{and} \quad\Delta_{i,j} = \frac{q_{i,j}}{p_{i,j} + q_{i,j}} \in(0,1]. f i , j − = a i , j − b i , j − , f i , j + = a i , j + b i , j + , ω i , j = a i , j − + C i , j / σ j 2 C i , j / σ j 2 ∈ ( 0 , 1 ) , and Δ i , j = p i , j + q i , j q i , j ∈ ( 0 , 1 ] .
Bayesian update and a first credibility formula
Based on the observation equation (8) , having observed C i , j + 1 C_{i,j+1} C i , j + 1 , we apply a Bayesian inference step. Note that H i , j = H i − 1 , j ∨ σ ( C i , j + 1 ) \mathcal{H}_{i,j} = \mathcal{H}_{i-1,j} \vee\sigma(C_{i,j+1}) H i , j = H i − 1 , j ∨ σ ( C i , j + 1 ) , and applying Bayes’ rule to (8) yields
Θ i , j ∣ H i , j ∼ Γ ( 1 + a i , j + , b i , j + ) . (11) \Theta_{i,j} \mid\mathcal{H}_{i,j} \sim\Gamma\left(1 + a^{+}_{i,j}, b^{+}_{i,j}\right). \tag*{(11)} Θ i , j ∣ H i , j ∼ Γ ( 1 + a i , j + , b i , j + ) . (11)
The state-space equation (4) gives us the predictive prior mean , based on history H i − 1 , j \mathcal{H}_{i-1,j} H i − 1 , j ,
f i , j − = E [ Θ i , j − 1 ∣ H i − 1 , j ] = b i , j − a i , j − , (12) f^{-}_{i,j} = \mathbb{E}\left[\Theta^{-1}_{i,j} \mid\mathcal{H}_{i-1,j}\right] = \frac{b^{-}_{i,j}}{a^{-}_{i,j}}, \tag*{(12)} f i , j − = E [ Θ i , j − 1 ∣ H i − 1 , j ] = a i , j − b i , j − , (12)
and the Bayesian inference step (11) turns this into the posterior mean
f i , j + = E [ Θ i , j − 1 ∣ H i , j ] = b i , j + a i , j + = ω i , j C i , j + 1 C i , j + ( 1 − ω i , j ) f i , j − . (13) f^{+}_{i,j} = \mathbb{E}\left[\Theta^{-1}_{i,j} \mid\mathcal{H}_{i,j}\right] = \frac{b^{+}_{i,j}}{a^{+}_{i,j}} = \omega_{i,j}\frac{C_{i,j+1}}{C_{i,j}} + (1-\omega_{i,j})f^{-}_{i,j}. \tag*{(13)} f i , j + = E [ Θ i , j − 1 ∣ H i , j ] = a i , j + b i , j + = ω i , j C i , j C i , j + 1 + ( 1 − ω i , j ) f i , j − . (13)
Thus, the Bayesian inference step yields the Bühlmann–Straub [5 ] credibility form for the individual CL factor Θ i , j − 1 \Theta^{-1}_{i,j} Θ i , j − 1 , conditional on the observations H i , j \mathcal{H}_{i,j} H i , j , with predictive prior mean f i , j − f^{-}_{i,j} f i , j − , observation update C i , j + 1 / C i , j C_{i,j+1}/C_{i,j} C i , j + 1 / C i , j and credibility weight ω i , j ∈ ( 0 , 1 ) \omega_{i,j} \in(0,1) ω i , j ∈ ( 0 , 1 ) .
State-space update and a second credibility formula
Based on (11) , we consider the state-space update (4) for i → i + 1 i \to i+1 i → i + 1 . That is,
Θ i , j ∣ H i , j ∼ Γ ( 1 + a i , j + , b i , j + ) ⟶ Θ i + 1 , j ∣ H i , j ∼ Γ ( 1 + a i + 1 , j − , b i + 1 , j − ) . (14) \Theta_{i,j} \mid\mathcal{H}_{i,j} \sim\Gamma\left(1 + a^{+}_{i,j}, b^{+}_{i,j}\right) \quad\longrightarrow\quad\Theta_{i+1,j} \mid\mathcal{H}_{i,j} \sim\Gamma\left(1 + a^{-}_{i+1,j}, b^{-}_{i+1,j}\right). \tag*{(14)} Θ i , j ∣ H i , j ∼ Γ ( 1 + a i , j + , b i , j + ) ⟶ Θ i + 1 , j ∣ H i , j ∼ Γ ( 1 + a i + 1 , j − , b i + 1 , j − ) . (14)
This is precisely the step that lacks uniqueness in the model definition of Assumptions 2.1. Both conditional distributions of Θ i , j \Theta_{i,j} Θ i , j and Θ i + 1 , j \Theta_{i+1,j} Θ i + 1 , j are conditioned on the same information set H i , j \mathcal{H}_{i,j} H i , j , i.e., we consider the conditional joint distribution of
Θ i , j , Θ i + 1 , j ∣ H i , j , (15) \Theta_{i,j}, \Theta_{i+1,j} \mid\mathcal{H}_{i,j}, \tag*{(15)} Θ i , j , Θ i + 1 , j ∣ H i , j , (15)
in the state-space update (14) that maps between the two conditional gamma distributions. Obviously, any copula can be selected in (15) to link these two conditional distributions. Therefore, we face the problem of non-uniqueness. For the computation of ultimate claim predictions and the conditional MSEP however, this non-identifiability is irrelevant (this is seen below), it only starts to matter if we try to understand the risk drivers in the conditional MSEP.
We compute the conditional mean of the state-space update, i.e., the new predictive prior mean, see (12) . This uses the assumptions (2.4)-(2.6) to obtain
f i + 1 , j − = E [ Θ i + 1 , j − 1 ∣ H i , j ] = b ˉ i + 1 , j a ˉ i + 1 , j = Δ i , j f i , j + + ( 1 − Δ i , j ) f j = Δ i , j ( ω i , j C i , j + 1 C i , j + ( 1 − ω i , j ) f i , j − ) + ( 1 − Δ i , j ) f j . (16) \begin{aligned}
f^{-}_{i+1,j} &= \mathbb{E}\left[\Theta^{-1}_{i+1,j}\mid\mathcal{H}_{i,j}\right] = \frac{\bar{b}_{i+1,j}}{\bar{a}_{i+1,j}} \\
&= \Delta_{i,j}f^{+}_{i,j}+(1-\Delta_{i,j})f_j \\
&= \Delta_{i,j}\left(\omega_{i,j}\frac{C_{i,j+1}}{C_{i,j}}+(1-\omega_{i,j})f_{i,j}^{-}\right)+(1-\Delta_{i,j})f_j. \tag*{(16)}
\end{aligned} f i + 1 , j − = E [ Θ i + 1 , j − 1 ∣ H i , j ] = a ˉ i + 1 , j b ˉ i + 1 , j = Δ i , j f i , j + + ( 1 − Δ i , j ) f j = Δ i , j ( ω i , j C i , j C i , j + 1 + ( 1 − ω i , j ) f i , j − ) + ( 1 − Δ i , j ) f j . (16)
The nature of the first credibility formula (3.3) is different from the second one (3.6). The former is a proper Bühlmann credibility form from a Bayesian inference step, whereas the latter is an enforced one by the affine parameter assumptions (2.4)-(2.6) in the state-space update (2.3).
Mean-stationarity
A main purpose of setting up the affine mean structure in the state-space update (2.3) is to receive the following mean-stationarity of the latent process. In particular, a priori, before having any observations, we cannot differentiate the expected CL factors between the different accident years i i i . The proof of the following lemma is given in the appendix.
Lemma 3.1 The above model satisfies mean-stationarity E [ Θ i , j − 1 ∣ B j ] = f j \mathbb{E}[\Theta^{-1}_{i,j}\mid\mathcal{B}_j]=f_j E [ Θ i , j − 1 ∣ B j ] = f j for all i ∈ { 1 , … , I } i\in\{1,\ldots,I\} i ∈ { 1 , … , I } .
This mean-stationarity is based on the following observations used in the proof of Lemma 3.1 in the appendix:
E [ C i , j + 1 ∣ H i − 1 , j ] = E [ E [ C i , j + 1 ∣ H i − 1 , j , G i , j ] ∣ H i − 1 , j ] = C i , j E [ Θ i , j − 1 ∣ H i − 1 , j ] = C i , j f i , j − . (17) \begin{aligned}
\mathbb{E}[C_{i,j+1}\mid\mathcal{H}_{i-1,j}]
=
\mathbb{E}\left[\mathbb{E}[C_{i,j+1}\mid\mathcal{H}_{i-1,j},G_{i,j}]\mid\mathcal{H}_{i-1,j}\right] \\
&=
C_{i,j}\mathbb{E}\left[\Theta^{-1}_{i,j}\mid\mathcal{H}_{i-1,j}\right]
=
C_{i,j}f^{-}_{i,j}. \tag*{(17)}
\end{aligned} E [ C i , j + 1 ∣ H i − 1 , j ] = E [ E [ C i , j + 1 ∣ H i − 1 , j , G i , j ] ∣ H i − 1 , j ] = C i , j E [ Θ i , j − 1 ∣ H i − 1 , j ] = C i , j f i , j − . (17)
the right-hand side considers the predictive prior mean f i , j − f^{-}_{i,j} f i , j − of Θ i , j − 1 \Theta^{-1}_{i,j} Θ i , j − 1 , given H i − 1 , j \mathcal{H}_{i-1,j} H i − 1 , j , see (12) . Moreover, the tower property yields the calibration for the posterior mean, see (3.3),
E [ f i , j + ∣ H i − 1 , j ] = f i , j − . (18) \mathbb{E}\left[f^{+}_{i,j}\mid\mathcal{H}_{i-1,j}\right]=f^{-}_{i,j}. \tag*{(18)} E [ f i , j + ∣ H i − 1 , j ] = f i , j − . (18)
This can be iterated which gives the following corollary.
Corollary 3.2 For h ∈ { 0 , … , i − 1 } h\in\{0,\ldots,i-1\} h ∈ { 0 , … , i − 1 }
E [ f i , j − ∣ H h , j ] = E [ f i , j + ∣ H h , j ] = ( ∏ k = h + 1 i − 1 Δ k , j ) f h + 1 , j − + ( 1 − ∏ k = h + 1 i − 1 Δ k , j ) f j , \mathbb{E}\left[f^{-}_{i,j}\mid\mathcal{H}_{h,j}\right]
=
\mathbb{E}\left[f^{+}_{i,j}\mid\mathcal{H}_{h,j}\right]
=
\left(\prod_{k=h+1}^{i-1}\Delta_{k,j}\right)f^{-}_{h+1,j}
+
\left(1-\prod_{k=h+1}^{i-1}\Delta_{k,j}\right)f_j, E [ f i , j − ∣ H h , j ] = E [ f i , j + ∣ H h , j ] = ( k = h + 1 ∏ i − 1 Δ k , j ) f h + 1 , j − + ( 1 − k = h + 1 ∏ i − 1 Δ k , j ) f j , where an empty product is set equal to one, and with initialization f 1 , j − = f j f^{-}_{1,j}=f_j f 1 , j − = f j .
Table 6.
Ultimate claim predictions
We translate the above considerations to ultimate claim predictions of C i , J C_{i,J} C i , J , i + J > I i+J>I i + J > I , under trapezoidal observations D I D_I D I
C ^ i , J = E [ C i , J ∣ D I ] , \widehat{C}_{i,J}=\mathbb{E}[C_{i,J}\mid D_I], C i , J = E [ C i , J ∣ D I ] ,
we refer to Figure 1 (lhs). Thus, we consider all accident years i i i with incomplete row observations.
Data likelihood
The available data D I D_I D I is either a triangle for I = J + 1 I=J+1 I = J + 1 or a trapezoid for I > J + 1 I>J+1 I > J + 1 . The data likelihood factorizes along the development year axis j j j , and it can be written in an elegant way along the accident year axis i i i .
We start by introducing the full data distribution over the entire development rectangle. We therefore introduce the (scaled) prime-beta densities on the positive support x > 0 x>0 x > 0
K ( x ∣ c , a , b ) = Γ ( c + 1 + a ) Γ ( c ) Γ ( 1 + a ) b 1 + a x c − 1 ( b + x ) c + 1 + a , K(x\mid c,a,b)=\frac{\Gamma(c+1+a)}{\Gamma(c)\Gamma(1+a)}\frac{b^{1+a}x^{c-1}}{(b+x)^{c+1+a}}, K ( x ∣ c , a , b ) = Γ ( c ) Γ ( 1 + a ) Γ ( c + 1 + a ) ( b + x ) c + 1 + a b 1 + a x c − 1 ,
for positive parameters a , b , c > 0 a,b,c>0 a , b , c > 0 . The prime-beta random variables C ∼ K ( ⋅ ∣ c , a , b ) C\sim K(\cdot\mid c,a,b) C ∼ K ( ⋅ ∣ c , a , b ) can be simulated from independent random variables U ∼ Γ ( c , 1 ) U\sim\Gamma(c,1) U ∼ Γ ( c , 1 ) and V ∼ Γ ( 1 + a , b ) V\sim\Gamma(1+a,b) V ∼ Γ ( 1 + a , b ) by setting C = U / V C=U/V C = U / V .
Lemma 4.1 The conditional density of the model in Assumptions 2.1, given the first column C 1 : I , 0 C_{1:I,0} C 1 : I , 0 , factorizes as
f ( ( C i , j ) i = 1 , … , I , j = 1 , … , J ∣ C 1 : I , 0 ) = ∏ j = 0 J − 1 ∏ i = 1 I K ( C i , j + 1 ∣ C i , j / σ j 2 , a i , j − , σ j 2 b i , j − ) . f\left((C_{i,j})_{i=1,\ldots,I,\ j=1,\ldots,J}\mid C_{1:I,0}\right)=\prod_{j=0}^{J-1}\prod_{i=1}^{I}K\left(C_{i,j+1}\mid C_{i,j}/\sigma_j^2,a_{i,j}^{-},\sigma_j^2b_{i,j}^{-}\right). f ( ( C i , j ) i = 1 , … , I , j = 1 , … , J ∣ C 1 : I , 0 ) = j = 0 ∏ J − 1 i = 1 ∏ I K ( C i , j + 1 ∣ C i , j / σ j 2 , a i , j − , σ j 2 b i , j − ) .
The proof of this lemma is provided in the appendix. We relate this lemma to the data likelihood of D I D_I D I . There are two important observations. First, the trapezoidal structure respects parenting, meaning that C i , j + 1 C_{i,j+1} C i , j + 1 only depends on C 1 : i , j C_{1:i,j} C 1 : i , j and C 1 : i − 1 , j + 1 C_{1:i-1,j+1} C 1 : i − 1 , j + 1 . Second, related to the first item, the prime-beta densities only depend on these parents, but not on the entire history H i − 1 , j H_{i-1,j} H i − 1 , j , that is,
K ( C i , j + 1 ∣ C i , j / σ j 2 , a i , j − , σ j 2 b i , j − ) = f ( C i , j + 1 ∣ C 1 : i , j , C 1 : i − 1 , j + 1 ) . (19) K\left(C_{i,j+1}\mid C_{i,j}/\sigma_j^2,a_{i,j}^{-},\sigma_j^2b_{i,j}^{-}\right)=f\left(C_{i,j+1}\mid C_{1:i,j},C_{1:i-1,j+1}\right). \tag*{(19)} K ( C i , j + 1 ∣ C i , j / σ j 2 , a i , j − , σ j 2 b i , j − ) = f ( C i , j + 1 ∣ C 1 : i , j , C 1 : i − 1 , j + 1 ) . (19)
These two items allow us to adapt the density of Lemma 4.1 to the trapezoidal shape of data D I D_I D I by first integrating out C r J + 1 : I , J C_{r_J+1:I,J} C r J + 1 : I , J , then C r J − 1 + 1 : I , J − 1 C_{r_{J-1}+1:I,J-1} C r J − 1 + 1 : I , J − 1 , etc. This results in Corollary 4.2, below, but we first introduce some notation.
The index d i ∈ { 0 , … , J } d_i\in\{0,\ldots,J\} d i ∈ { 0 , … , J } in (2.1) defines the latest observed cumulative claims C i , d i C_{i,d_i} C i , d i at time I I I for all accident years i ∈ { 1 , … , I } i\in\{1,\ldots,I\} i ∈ { 1 , … , I } . Equivalently, we can define the latest observations C r j , j C_{r_j,j} C r j , j for all columns j ∈ { 0 , … , J } j\in\{0,\ldots,J\} j ∈ { 0 , … , J } by the accident year indices
r j = I − j ∈ { I − J , … , I } . r_j=I-j\in\{I-J,\ldots,I\}. r j = I − j ∈ { I − J , … , I } .
These two indices are illustrated in Figure 3 . This provides the following corollary from Lemma 4.1, for convenience, we shift the development parameter j j j by one unit.
Figure 3. Latest observations C i , d i C_{i,d_i} C i , d i and C r j , j C_{r_j,j} C r j , j at time I I I for i ∈ { 1 , … , I } i \in\{1,\ldots,I\} i ∈ { 1 , … , I } and j ∈ { 0 , … , J } j \in\{0,\ldots,J\} j ∈ { 0 , … , J } .
Corollary 4.2. The conditional data likelihood of D I D_I D I , given the first column C 1 : I , 0 C_{1:I,0} C 1 : I , 0 , factorizes as
f ( ( C i , j ) i + j ≤ I , j ≥ 1 ∣ C 1 : I , 0 ) = ∏ j = 1 J ∏ i = 1 r j K ( C i , j ∣ C i , j − 1 / σ j − 1 2 , a i , j − 1 − , σ j − 1 2 b i , j − 1 − ) . f\left((C_{i,j})_{i+j\leq I,j\geq1}\mid C_{1:I,0}\right)=\prod_{j=1}^{J}\prod_{i=1}^{r_j}K\left(C_{i,j}\mid C_{i,j-1}/\sigma_{j-1}^{2},a^{-}_{i,j-1},\sigma_{j-1}^{2}b^{-}_{i,j-1}\right). f ( ( C i , j ) i + j ≤ I , j ≥ 1 ∣ C 1 : I , 0 ) = j = 1 ∏ J i = 1 ∏ r j K ( C i , j ∣ C i , j − 1 / σ j − 1 2 , a i , j − 1 − , σ j − 1 2 b i , j − 1 − ) .
The likelihoods of Lemma 4.1 and Corollary 4.2 give nice insights. We are able to express the data likelihood explicitly (fully analytically), and the non-identifiability of the copula in (3.5) does not have any impact. Thus, Assumptions 2.1 are sufficient to compute the ultimate claim predictions and the prediction uncertainty, given the data D I D_I D I . We can interpret the latent state-space processes as auxiliary tools that help us to shape our understanding and interpretation, and to arrive at the likelihoods of Lemma 4.1 and Corollary 4.2.
Another noteworthy remark is that the accident year factorization in Lemma 4.1 and Corollary 4.2 is not an independent one as a i , j a_{i,j} a i , j and b i , j b_{i,j} b i , j collect the observed histories C 1 : i − 1 , j C_{1:i-1,j} C 1 : i − 1 , j and C 1 : i − 1 , j + 1 C_{1:i-1,j+1} C 1 : i − 1 , j + 1 .
Observable chain-ladder factors and ultimate claim prediction
To compute the ultimate claim predictions, we need to be able to determine the CL factors, given D I D_I D I . Such a computation has already been presented in Corollary 3.2, however, the information sets D I D_I D I and H r j − 1 , j \mathcal{H}_{r_j-1,j} H r j − 1 , j do not directly align, see Figure 4 .
Figure 4. (lhs) Information set D I D_I D I and (rhs) information set H r j − 1 , j = H r j + 1 , j \mathcal{H}_{r_j-1,j}=\mathcal{H}_{r_{j+1},j} H r j − 1 , j = H r j + 1 , j .
We compute for a fixed j ∈ { 0 , … , J − 1 } j \in\{0,\ldots,J-1\} j ∈ { 0 , … , J − 1 }
E [ f i , j − ∣ D I ] for i = r j , … , I , \mathbb{E}\left[f^{-}_{i,j}\mid D_I\right]\qquad\text{for }i=r_j,\ldots,I, E [ f i , j − ∣ D I ] for i = r j , … , I ,
this corresponds to the cumulative claims C i , j + 1 C_{i,j+1} C i , j + 1 , see (17) , which are illustrated by the continuation of the yellow column in Figure 4 (rhs). Fortunately, this can be done, and the result is provided in the next proposition; the proof is given in the appendix.
Proposition 4.3 Consider i ≥ r j = I − j i \ge r_j = I-j i ≥ r j = I − j . We have observable CL factors at time I I I
f ^ i , j ( I ) = E [ f i , j ∣ D I ] = E [ f i , j − ∣ H r j − 1 , j ] = ( ∏ k = r j i − 1 Δ k , j ) f r j , j − + ( 1 − ∏ k = r j i − 1 Δ k , j ) f j . \widehat{f}_{i,j}^{(I)} = \mathbb{E}\left[f_{i,j}\mid\mathcal{D}_I\right] = \mathbb{E}\left[f_{i,j}^{-}\mid\mathcal{H}_{r_j-1,j}\right] = \left(\prod_{k=r_j}^{i-1}\Delta_{k,j}\right)f_{r_j,j}^{-}+\left(1-\prod_{k=r_j}^{i-1}\Delta_{k,j}\right)f_j. f i , j ( I ) = E [ f i , j ∣ D I ] = E [ f i , j − ∣ H r j − 1 , j ] = k = r j ∏ i − 1 Δ k , j f r j , j − + 1 − k = r j ∏ i − 1 Δ k , j f j . Using (3.6) and the conditional independent decoupling argument of the previous proof, we also get
f ^ i , j ( I ) = E [ f i , j − ∣ D I ] = E [ f i , j − ∣ H r j − 1 , j ] = E [ Θ i , j − 1 ∣ H r j − 1 , j ] . (20) \widehat{f}_{i,j}^{(I)}=\mathbb{E}\left[f_{i,j}^{-}\mid\mathcal{D}_I\right]=\mathbb{E}\left[f_{i,j}^{-}\mid\mathcal{H}_{r_j-1,j}\right]=\mathbb{E}\left[\Theta_{i,j}^{-1}\mid\mathcal{H}_{r_j-1,j}\right]. \tag*{(20)} f i , j ( I ) = E [ f i , j − ∣ D I ] = E [ f i , j − ∣ H r j − 1 , j ] = E [ Θ i , j − 1 ∣ H r j − 1 , j ] . (20)
Thus, f ^ i , j ( I ) \widehat{f}_{i,j}^{(I)} f i , j ( I ) is the observable CL factor at time I I I . Proposition 4.3 now allows us to compute the ultimate claim predictions, given observations D I \mathcal{D}_I D I . We have the following theorem.
Theorem 4.4 For accident years i + J > I i+J>I i + J > I we have ultimate claim predictor
C ^ i , J = E [ C i , J ∣ D I ] = C i , d i ∏ j = d i J − 1 f ^ i , j ( I ) . \widehat{C}_{i,J}=\mathbb{E}[C_{i,J}\mid\mathcal{D}_I]=C_{i,d_i}\prod_{j=d_i}^{J-1}\widehat{f}_{i,j}^{(I)}. C i , J = E [ C i , J ∣ D I ] = C i , d i j = d i ∏ J − 1 f i , j ( I ) .
The proof of this theorem is given in the appendix. For accident years i + J > I i+J>I i + J > I , Theorem 4.4 requires computation of the observable CL factors, see Proposition 4.3 ,
f ^ i , j ( I ) = ( ∏ k = r j i − 1 Δ k , j ) f r j , j − + ( 1 − ∏ k = r j i − 1 Δ k , j ) f j . (21) \widehat{f}_{i,j}^{(I)}=\left(\prod_{k=r_j}^{i-1}\Delta_{k,j}\right)f_{r_j,j}^{-}+\left(1-\prod_{k=r_j}^{i-1}\Delta_{k,j}\right)f_j. \tag*{(21)} f i , j ( I ) = k = r j ∏ i − 1 Δ k , j f r j , j − + 1 − k = r j ∏ i − 1 Δ k , j f j . (21)
These observable CL factors involve the deterministic weights Δ k , j = q k , j / ( p k , j + q k , j ) ∈ ( 0 , 1 ] \Delta_{k,j}=q_{k,j}/(p_{k,j}+q_{k,j})\in(0,1] Δ k , j = q k , j / ( p k , j + q k , j ) ∈ ( 0 , 1 ] , the fixed baseline mean f j f_j f j , and the observable predictive prior mean at time I I I given by
f r j , j − = Δ r j − 1 , j ( ω r j − 1 , j C r j − 1 , j + 1 C r j − 1 , j + ( 1 − ω r j − 1 , j ) f r j − 1 , j − ) + ( 1 − Δ r j − 1 , j ) f j . (22) f_{r_j,j}^{-}=\Delta_{r_j-1,j}\left(\omega_{r_j-1,j}\frac{C_{r_j-1,j+1}}{C_{r_j-1,j}}+(1-\omega_{r_j-1,j})f_{r_j-1,j}^{-}\right)+(1-\Delta_{r_j-1,j})f_j. \tag*{(22)} f r j , j − = Δ r j − 1 , j ( ω r j − 1 , j C r j − 1 , j C r j − 1 , j + 1 + ( 1 − ω r j − 1 , j ) f r j − 1 , j − ) + ( 1 − Δ r j − 1 , j ) f j . (22)
with weights
ω r j − 1 , j = C r j − 1 , j / σ j 2 a r j − 1 , j − + C r j − 1 , j / σ j 2 = C r j − 1 , j ∑ k = 1 r j − 1 C k , j ∏ ℓ = k r j − 2 ( p ℓ , j + q ℓ , j ) + σ j 2 γ j ∏ ℓ = 1 r j − 2 ( p ℓ , j + q ℓ , j ) ∈ ( 0 , 1 ) . \begin{aligned}
\omega_{r_j-1,j}=\frac{C_{r_j-1,j}/\sigma_j^2}{a_{r_j-1,j}^{-}+C_{r_j-1,j}/\sigma_j^2} \\
&=\frac{C_{r_j-1,j}}{\sum_{k=1}^{r_j-1}C_{k,j}\prod_{\ell=k}^{r_j-2}(p_{\ell,j}+q_{\ell,j})+\sigma_j^2\gamma_j\prod_{\ell=1}^{r_j-2}(p_{\ell,j}+q_{\ell,j})}\in(0,1).
\end{aligned} ω r j − 1 , j = a r j − 1 , j − + C r j − 1 , j / σ j 2 C r j − 1 , j / σ j 2 = ∑ k = 1 r j − 1 C k , j ∏ ℓ = k r j − 2 ( p ℓ , j + q ℓ , j ) + σ j 2 γ j ∏ ℓ = 1 r j − 2 ( p ℓ , j + q ℓ , j ) C r j − 1 , j ∈ ( 0 , 1 ) .
The latter uses (2.4) and (2.6) to unfold a i , j − > 0 a_{i,j}^{-}>0 a i , j − > 0 as follows
a i , j − = ( p i − 1 , j + q i − 1 , j ) C i − 1 , j σ j 2 + ( p i − 1 , j + q i − 1 , j ) a i − 1 , j − = ∑ k = 1 i − 1 C k , j σ j 2 ∏ ℓ = k i − 1 ( p ℓ , j + q ℓ , j ) + γ j ∏ ℓ = 1 i − 1 ( p ℓ , j + q ℓ , j ) . (23) \begin{aligned}
a_{i,j}^{-}=(p_{i-1,j}+q_{i-1,j})\frac{C_{i-1,j}}{\sigma_j^2}+(p_{i-1,j}+q_{i-1,j})a_{i-1,j}^{-} \\
&=\sum_{k=1}^{i-1}\frac{C_{k,j}}{\sigma_j^2}\prod_{\ell=k}^{i-1}(p_{\ell,j}+q_{\ell,j})+\gamma_j\prod_{\ell=1}^{i-1}(p_{\ell,j}+q_{\ell,j}). \tag*{(23)}
\end{aligned} a i , j − = ( p i − 1 , j + q i − 1 , j ) σ j 2 C i − 1 , j + ( p i − 1 , j + q i − 1 , j ) a i − 1 , j − = k = 1 ∑ i − 1 σ j 2 C k , j ℓ = k ∏ i − 1 ( p ℓ , j + q ℓ , j ) + γ j ℓ = 1 ∏ i − 1 ( p ℓ , j + q ℓ , j ) . (23)
Formula (4.4) gives the full updating formula of the observable CL factor across the accident year indices i = 2 , … , r j i=2,\ldots,r_j i = 2 , … , r j , this is the part that is covered by observations D I \mathcal{D}_I D I . Formula (4.3) extrapolates the accident year index updates i = r j + 1 , … , I i=r_j+1,\ldots,I i = r j + 1 , … , I into the lower triangle, where no observations are available. The structure is the same, but the missing observations are integrated out.
Special case: Exponential seniority weighting in the CL factors
In the applications, our main interest lies in the special case p i , j ≡ 0 p_{i,j} \equiv0 p i , j ≡ 0 and q i , j ≡ q j ≤ 1 q_{i,j} \equiv q_j \le1 q i , j ≡ q j ≤ 1 . The first assumption gives us Δ i , j ≡ 1 \Delta_{i,j} \equiv1 Δ i , j ≡ 1 which simplifies (22) as follows, note that these observable CL factors no longer have an accident year i i i dependence,
f ^ j ( I ) : = f ‾ r j , j = ω r j − 1 , j C r j − 1 , j + 1 C r j − 1 , j + ( 1 − ω r j − 1 , j ) f ‾ r j − 1 , j = ∑ k = 1 r j − 1 [ ∏ ℓ = k + 1 r j − 1 ( 1 − ω ℓ , j ) ] ω k , j C k , j + 1 C k , j + ∏ ℓ = 1 r j − 1 ( 1 − ω ℓ , j ) f j . (24) \begin{aligned}
\widehat{f}^{(I)}_j := \overline{f}_{r_j,j} = \omega_{r_j-1,j}\frac{C_{r_j-1,j+1}}{C_{r_j-1,j}} + (1-\omega_{r_j-1,j})\overline{f}_{r_j-1,j} \\
&= \sum_{k=1}^{r_j-1}\left[\prod_{\ell=k+1}^{r_j-1}(1-\omega_{\ell,j})\right]\omega_{k,j}\frac{C_{k,j+1}}{C_{k,j}}+\prod_{\ell=1}^{r_j-1}(1-\omega_{\ell,j})f_j. \tag*{(24)}
\end{aligned} f j ( I ) := f r j , j = ω r j − 1 , j C r j − 1 , j C r j − 1 , j + 1 + ( 1 − ω r j − 1 , j ) f r j − 1 , j = k = 1 ∑ r j − 1 [ ℓ = k + 1 ∏ r j − 1 ( 1 − ω ℓ , j ) ] ω k , j C k , j C k , j + 1 + ℓ = 1 ∏ r j − 1 ( 1 − ω ℓ , j ) f j . (24)
This is a classical credibility formula that stems from a time-series consideration, and the credibility weights simplify under q i , j ≡ q j ≤ 1 q_{i,j} \equiv q_j \le1 q i , j ≡ q j ≤ 1 to
ω i , j = C i , j ∑ k = 1 i q j i − k C k , j + q j i − 1 σ j 2 γ j ∈ ( 0 , 1 ) . (25) \omega_{i,j}=\frac{C_{i,j}}{\sum_{k=1}^{i}q_j^{i-k}C_{k,j}+q_j^{i-1}\sigma_j^2\gamma_j}\in(0,1). \tag*{(25)} ω i , j = ∑ k = 1 i q j i − k C k , j + q j i − 1 σ j 2 γ j C i , j ∈ ( 0 , 1 ) . (25)
If we furthermore consider the non-informative prior limits γ j ↓ 0 \gamma_j\downarrow0 γ j ↓ 0 , we obtain
ω i , j = lim γ j ↓ 0 C i , j ∑ k = 1 i q j i − k C k , j + q j i − 1 σ j 2 γ j = C i , j ∑ k = 1 i q j i − k C k , j ∈ ( 0 , 1 ] . (26) \omega_{i,j}=\lim_{\gamma_j\downarrow0}\frac{C_{i,j}}{\sum_{k=1}^{i}q_j^{i-k}C_{k,j}+q_j^{i-1}\sigma_j^2\gamma_j}=\frac{C_{i,j}}{\sum_{k=1}^{i}q_j^{i-k}C_{k,j}}\in(0,1]. \tag*{(26)} ω i , j = γ j ↓ 0 lim ∑ k = 1 i q j i − k C k , j + q j i − 1 σ j 2 γ j C i , j = ∑ k = 1 i q j i − k C k , j C i , j ∈ ( 0 , 1 ] . (26)
Compared to the classical CL factors f ^ j C L \widehat{f}^{\mathrm{CL}}_j f j CL of Mack [11 ] , we obtain an exponential seniority weighting q j i − k q_j^{i-k} q j i − k in past observations; the classical CL factors of Mack [11 ] are given by
f ^ j C L = ∑ k = 1 r j − 1 C k , j + 1 ∑ k = 1 r j − 1 C k , j = ∑ i = 1 r j − 1 [ C i , j ∑ k = 1 r j − 1 C k , j ] C i , j + 1 C i , j . (27) \widehat{f}^{\mathrm{CL}}_j=\frac{\sum_{k=1}^{r_j-1}C_{k,j+1}}{\sum_{k=1}^{r_j-1}C_{k,j}}=\sum_{i=1}^{r_j-1}\left[\frac{C_{i,j}}{\sum_{k=1}^{r_j-1}C_{k,j}}\right]\frac{C_{i,j+1}}{C_{i,j}}. \tag*{(27)} f j CL = ∑ k = 1 r j − 1 C k , j ∑ k = 1 r j − 1 C k , j + 1 = i = 1 ∑ r j − 1 [ ∑ k = 1 r j − 1 C k , j C i , j ] C i , j C i , j + 1 . (27)
Moreover, in this non-informative prior limit (26) , also the credibility weight on the baseline f j f_j f j in (24) will vanish. In summary, this gives us the following theorem.
Theorem 4.5 (Exponential seniority weighted CL factors). Under Assumptions 2.1, let p i , j ≡ 0 p_{i,j}\equiv0 p i , j ≡ 0 , q i , j ≡ q j ≤ 1 q_{i,j}\equiv q_j\le1 q i , j ≡ q j ≤ 1 and γ j → 0 \gamma_j\to0 γ j → 0 . For accident years i ≥ I − J + 1 = r J − 1 i\ge I-J+1=r_{J-1} i ≥ I − J + 1 = r J − 1 , we have the ultimate claims predictors
C ^ i , J C L ( q ) : = lim γ ↓ 0 E γ [ C i , J ∣ D I ] = C i , d i ∏ j = d i J − 1 f ^ j ( I ) , \widehat{C}^{\mathrm{CL}(q)}_{i,J}:=\lim_{\gamma\downarrow0}\mathbb{E}_{\gamma}[C_{i,J}\mid D_I]=C_{i,d_i}\prod_{j=d_i}^{J-1}\widehat{f}^{(I)}_j, C i , J CL ( q ) := γ ↓ 0 lim E γ [ C i , J ∣ D I ] = C i , d i j = d i ∏ J − 1 f j ( I ) , with CL factors and credibility weights, respectively, for j ∈ { 0 , … , J − 1 } j\in\{0,\ldots,J-1\} j ∈ { 0 , … , J − 1 }
f ^ j ( I ) = ∑ i = 1 r j − 1 α i , j C i , j + 1 C i , j and α i , j = q j r j − 1 − i C i , j ∑ k = 1 r j − 1 q j r j − 1 − k C k , j ∈ ( 0 , 1 ] . (28) \widehat{f}^{(I)}_j=\sum_{i=1}^{r_j-1}\alpha_{i,j}\frac{C_{i,j+1}}{C_{i,j}}\qquad\text{and}\qquad\alpha_{i,j}=\frac{q_j^{r_j-1-i}C_{i,j}}{\sum_{k=1}^{r_j-1}q_j^{r_j-1-k}C_{k,j}}\in(0,1]. \tag*{(28)} f j ( I ) = i = 1 ∑ r j − 1 α i , j C i , j C i , j + 1 and α i , j = ∑ k = 1 r j − 1 q j r j − 1 − k C k , j q j r j − 1 − i C i , j ∈ ( 0 , 1 ] . (28)
Theorem 4.5 allows for an exponential seniority weighting in the computation of the CL factors, this is illustrated in Figure 5 (rhs). Moreover, the choice q j ≡ 1 q_j\equiv1 q j ≡ 1 , yields the classical CL predictors of Mack [11 ] .
Figure 5. (lhs) Information set D I D_I D I and (rhs) exponential seniority weighting of past information according to Theorem 4.5 .
Total prediction uncertainty
We introduce the following abbreviations for the accident year aggregated versions
C ∙ = ∑ i = 1 I C i , J , and C ^ ∙ = ∑ i = 1 I C ^ i , J = E [ C ∙ ∣ D I ] . C_{\bullet} = \sum_{i=1}^{I} C_{i,J}, \qquad\text{and} \qquad\widehat{C}_{\bullet} = \sum_{i=1}^{I} \widehat{C}_{i,J} = \mathbb{E}[C_{\bullet} \mid D_I]. C ∙ = i = 1 ∑ I C i , J , and C ∙ = i = 1 ∑ I C i , J = E [ C ∙ ∣ D I ] .
We measure prediction uncertainty in the classical actuarial view by computing the conditional MSEP given by
msep C ∙ ∣ D I ( C ^ ∙ ) = E [ ( C ∙ − C ^ ∙ ) 2 ∣ D I ] . \operatorname{msep}_{C_{\bullet}\mid D_I}\left(\widehat{C}_{\bullet}\right) = \mathbb{E}\left[\left(C_{\bullet}-\widehat{C}_{\bullet}\right)^2 \mid D_I\right]. msep C ∙ ∣ D I ( C ∙ ) = E [ ( C ∙ − C ∙ ) 2 ∣ D I ] .
Because we use the conditional expectations C ^ i , J = E [ C i , J ∣ D I ] \widehat{C}_{i,J} = \mathbb{E}[C_{i,J}\mid D_I] C i , J = E [ C i , J ∣ D I ] as predictors, we obtain
msep C ∙ ∣ D I ( C ^ ∙ ) = Var ( C ∙ ∣ D I ) . (29) \operatorname{msep}_{C_{\bullet}\mid D_I}\left(\widehat{C}_{\bullet}\right) = \operatorname{Var}(C_{\bullet}\mid D_I). \tag*{(29)} msep C ∙ ∣ D I ( C ∙ ) = Var ( C ∙ ∣ D I ) . (29)
Coming back to the non-identifiability of the unspecified copula in (15) , we have seen in § 4.1 that the unspecified copula does not impact the computation of the conditional MSEP. However, it can affect the analysis of the two terms of average process uncertainty and parameter uncertainty, respectively,
msep C ∙ ∣ D I ( C ^ ∙ ) = E [ Var ( C ∙ ∣ Θ , D I ) ∣ D I ] + Var ( E [ C ∙ ∣ Θ , D I ] ∣ D I ) . \operatorname{msep}_{C_{\bullet}\mid D_I}\left(\widehat{C}_{\bullet}\right) = \mathbb{E}\left[\operatorname{Var}(C_{\bullet}\mid\Theta,D_I)\mid D_I\right] + \operatorname{Var}\left(\mathbb{E}[C_{\bullet}\mid\Theta,D_I]\mid D_I\right). msep C ∙ ∣ D I ( C ∙ ) = E [ Var ( C ∙ ∣ Θ , D I ) ∣ D I ] + Var ( E [ C ∙ ∣ Θ , D I ] ∣ D I ) .
here Θ = Θ 1 : I , 0 : J − 1 \Theta= \Theta_{1:I,0:J-1} Θ = Θ 1 : I , 0 : J − 1 . Our focus is on (29) , which means that we do not need to specify the copula.
Unfortunately, we cannot compute (29) in closed-form, except for special cases like in the gamma-gamma Bayes’ CL model of Wüthrich–Merz [21 ] . Fortunately, the fact that we have a full likelihood from Lemma 4.1 permits a Monte Carlo (MC) solution because we can simulate the lower triangle, given D I D_I D I . That is, we can recursively simulate from
f ( ( C i , j ) i + j > I ∣ D I ) = ∏ j = 1 J ∏ i = r j + 1 I K ( C i , j ∣ C i , j − 1 / σ j − 1 2 , a i , j − 1 − , σ j − 1 2 b i , j − 1 − ) , f\left((C_{i,j})_{i+j>I}\mid D_I\right) = \prod_{j=1}^{J}\prod_{i=r_j+1}^{I} K\left(C_{i,j}\mid C_{i,j-1}/\sigma_{j-1}^{2},a_{i,j-1}^{-},\sigma_{j-1}^{2}b_{i,j-1}^{-}\right), f ( ( C i , j ) i + j > I ∣ D I ) = j = 1 ∏ J i = r j + 1 ∏ I K ( C i , j ∣ C i , j − 1 / σ j − 1 2 , a i , j − 1 − , σ j − 1 2 b i , j − 1 − ) ,
and using the recursive structure (2.4)–(2.6) with initialization a 1 , j − = γ j > 0 a_{1,j}^{-}=\gamma_j>0 a 1 , j − = γ j > 0 and b 1 , j − = f j γ j > 0 b_{1,j}^{-}=f_j\gamma_j>0 b 1 , j − = f j γ j > 0 , and recall r j = I − j r_j=I-j r j = I − j .
Algorithm 1 Conditional MC simulation of the lower triangle and MSEP estimation
Require: Observations { C i , j : i + j ≤ I } \{C_{i,j}:i+j\le I\} { C i , j : i + j ≤ I } ; for j ∈ { 1 , … , J } j\in\{1,\ldots,J\} j ∈ { 1 , … , J } , the parameters of Assumptions 2.1, a r j + 1 , j − 1 − a_{r_{j+1},j-1}^{-} a r j + 1 , j − 1 − and b r j + 1 , j − 1 − b_{r_{j+1},j-1}^{-} b r j + 1 , j − 1 − ; number of simulations M > 1 M>1 M > 1 ; predictor C ^ ∙ \widehat{C}_{\bullet} C ∙ .
Ensure: MC estimate M S E P ^ M \widehat{\mathrm{MSEP}}_{M} MSEP M and MC error C ^ ∙ − C ‾ ∙ \widehat{C}_{\bullet}-\overline{C}_{\bullet} C ∙ − C ∙ .
for m = 1 , … , M m=1,\ldots,M m = 1 , … , M do
Set C i , d i ( m ) ← C i , d i C_{i,d_i}^{(m)}\leftarrow C_{i,d_i} C i , d i ( m ) ← C i , d i for every observed cell i ∈ { 1 , … , I } i\in\{1,\ldots,I\} i ∈ { 1 , … , I } .
for j = 1 , … , J j=1,\ldots,J j = 1 , … , J do
for i = r j + 1 , … , I i=r_j+1,\ldots,I i = r j + 1 , … , I do
Draw C i , j ( m ) ∼ K ( ⋅ ∣ C i , j − 1 ( m ) / σ j − 1 2 , a i , j − 1 − , σ j − 1 2 b i , j − 1 − ) C_{i,j}^{(m)}\sim K\left(\,\cdot\mid C_{i,j-1}^{(m)}/\sigma_{j-1}^{2},a_{i,j-1}^{-},\sigma_{j-1}^{2}b_{i,j-1}^{-}\right) C i , j ( m ) ∼ K ( ⋅ ∣ C i , j − 1 ( m ) / σ j − 1 2 , a i , j − 1 − , σ j − 1 2 b i , j − 1 − ) .
if i < I i<I i < I then
a i , j − 1 + ← a i , j − 1 − + C i , j − 1 ( m ) / σ j − 1 2 a_{i,j-1}^{+}\leftarrow a_{i,j-1}^{-}+C_{i,j-1}^{(m)}/\sigma_{j-1}^{2} a i , j − 1 + ← a i , j − 1 − + C i , j − 1 ( m ) / σ j − 1 2 .
b i , j − 1 + ← b i , j − 1 − + C i , j ( m ) / σ j − 1 2 b_{i,j-1}^{+}\leftarrow b_{i,j-1}^{-}+C_{i,j}^{(m)}/\sigma_{j-1}^{2} b i , j − 1 + ← b i , j − 1 − + C i , j ( m ) / σ j − 1 2 .
a i + 1 , j − 1 − ← ( p i , j − 1 + q i , j − 1 ) a i , j − 1 + a_{i+1,j-1}^{-}\leftarrow(p_{i,j-1}+q_{i,j-1})a_{i,j-1}^{+} a i + 1 , j − 1 − ← ( p i , j − 1 + q i , j − 1 ) a i , j − 1 + .
b i + 1 , j − 1 − ← p i , j − 1 a i , j − 1 + f j − 1 + q i , j − 1 b i , j − 1 + b_{i+1,j-1}^{-}\leftarrow p_{i,j-1}a_{i,j-1}^{+}f_{j-1}+q_{i,j-1}b_{i,j-1}^{+} b i + 1 , j − 1 − ← p i , j − 1 a i , j − 1 + f j − 1 + q i , j − 1 b i , j − 1 + .
end if
end for
end for
Compute C ∙ ( m ) ← ∑ i = 1 I C i , J ( m ) C_{\bullet}^{(m)}\leftarrow\sum_{i=1}^{I}C_{i,J}^{(m)} C ∙ ( m ) ← ∑ i = 1 I C i , J ( m ) .
end for
Compute C ‾ ∙ ← 1 M ∑ m = 1 M C ∙ ( m ) \overline{C}_{\bullet}\leftarrow\frac{1}{M}\sum_{m=1}^{M}C_{\bullet}^{(m)} C ∙ ← M 1 ∑ m = 1 M C ∙ ( m ) .
Compute M S E P ^ M ← 1 M − 1 ∑ m = 1 M ( C ∙ ( m ) − C ‾ ∙ ) 2 \widehat{\mathrm{MSEP}}_{M}\leftarrow\frac{1}{M-1}\sum_{m=1}^{M}(C_{\bullet}^{(m)}-\overline{C}_{\bullet})^{2} MSEP M ← M − 1 1 ∑ m = 1 M ( C ∙ ( m ) − C ∙ ) 2 .
return M S E P ^ M \widehat{\mathrm{MSEP}}_{M} MSEP M and C ^ ∙ − C ‾ ∙ \widehat{C}_{\bullet}-\overline{C}_{\bullet} C ∙ − C ∙ .
The resulting algorithm is given in Algorithm 1. If E [ C ∙ 2 ∣ D I ] < ∞ \mathbb{E}[C_{\bullet}^{2}\mid D_I]<\infty E [ C ∙ 2 ∣ D I ] < ∞ , the MC simulation of Algorithm 1 converges to Var ( C ∙ ∣ D I ) \operatorname{Var}(C_{\bullet}\mid D_I) Var ( C ∙ ∣ D I ) , a.s. Checking finite second moments is a bit tricky, we give a sufficient condition. For j ∈ { 0 , … , J − 1 } j\in\{0,\ldots,J-1\} j ∈ { 0 , … , J − 1 } , consider A r j , j = a r j , j A_{r_j,j}=a_{r_j,j} A r j , j = a r j , j , and iterate A i + 1 , j = ( p i , j + q i , j ) A i , j A_{i+1,j}=(p_{i,j}+q_{i,j})A_{i,j} A i + 1 , j = ( p i , j + q i , j ) A i , j for i ≥ r j i\ge r_j i ≥ r j . Clearly a i , j ≥ A i , j a_{i,j}\ge A_{i,j} a i , j ≥ A i , j for every simulated history in the lower triangle. We then require
ε : = min i + j ≥ I ( A i , j − 1 ) > 1. \varepsilon:=\min_{i+j\ge I}(A_{i,j}-1)>1. ε := i + j ≥ I min ( A i , j − 1 ) > 1.
This implies
E [ ( f i + 1 , j − ) 2 ∣ H i − 1 , j ] = ( Δ i , j f i , j − + ( 1 − Δ i , j ) f j ) 2 + Δ i , j 2 ω i , j ( f i , j − ) 2 a i , j − 1 ≤ ( 1 + ε − 1 ) ( f i , j − ) 2 + f j 2 . \mathbb{E}\left[\left(f_{i+1,j}^{-}\right)^2\mid\mathcal{H}_{i-1,j}\right]
=\left(\Delta_{i,j}f_{i,j}^{-}+(1-\Delta_{i,j})f_j\right)^2+\frac{\Delta_{i,j}^{2}\omega_{i,j}(f_{i,j}^{-})^2}{a_{i,j}-1}
\le(1+\varepsilon^{-1})(f_{i,j}^{-})^2+f_j^2. E [ ( f i + 1 , j − ) 2 ∣ H i − 1 , j ] = ( Δ i , j f i , j − + ( 1 − Δ i , j ) f j ) 2 + a i , j − 1 Δ i , j 2 ω i , j ( f i , j − ) 2 ≤ ( 1 + ε − 1 ) ( f i , j − ) 2 + f j 2 .
This can be iterated and implies E [ C ∙ 2 ∣ D I ] < ∞ \mathbb{E}[C_{\bullet}^{2}\mid D_I]<\infty E [ C ∙ 2 ∣ D I ] < ∞ .
We emphasize that the parameters γ j \gamma_j γ j , f j f_j f j , σ j 2 \sigma_j^2 σ j 2 , p i , j p_{i,j} p i , j and q i , j q_{i,j} q i , j are deterministic and fixed throughout.
Claims development result
This section studies the claims development result (CDR), which describes the one-year update of the ultimate claims view used in solvency considerations; see Merz–Wüthrich [14 ] . Introduce the available information at time I + 1 I+1 I + 1
D I + 1 = σ { C i , j ; i + j ≤ I + 1 , 1 ≤ i ≤ I , 0 ≤ j ≤ J } . D_{I+1}=\sigma\left\{C_{i,j};\ i+j\le I+1,\ 1\le i\le I,\ 0\le j\le J\right\}. D I + 1 = σ { C i , j ; i + j ≤ I + 1 , 1 ≤ i ≤ I , 0 ≤ j ≤ J } .
This is the run-off view, we retain the accident years 1 , … , I 1,\ldots,I 1 , … , I , and add the observations on the calendar year diagonal I + 1 I+1 I + 1 , i.e., no new accident year is introduced. The CDRs are then given by
C D R i = E [ C i , J ∣ D I ] − E [ C i , J ∣ D I + 1 ] and C D R = ∑ i = 1 I C D R i . \mathrm{CDR}_i=\mathbb{E}\left[C_{i,J}\mid D_I\right]-\mathbb{E}\left[C_{i,J}\mid D_{I+1}\right]
\qquad\text{and}\qquad
\mathrm{CDR}=\sum_{i=1}^{I}\mathrm{CDR}_i. CDR i = E [ C i , J ∣ D I ] − E [ C i , J ∣ D I + 1 ] and CDR = i = 1 ∑ I CDR i .
By the tower property, E [ C D R ∣ D I ] = 0 \mathbb{E}[\mathrm{CDR}\mid D_I]=0 E [ CDR ∣ D I ] = 0 . We investigate the conditional MSEP of the CDR
msep C D R ∣ D I ( 0 ) = E [ C D R 2 ∣ D I ] = Var ( C D R ∣ D I ) = Var ( ∑ i = 1 I E [ C i , J ∣ D I + 1 ] ∣ D I ) . (30) \operatorname{msep}_{\mathrm{CDR}\mid D_I}(0)=\mathbb{E}\left[\mathrm{CDR}^2\mid D_I\right]=\operatorname{Var}\left(\mathrm{CDR}\mid D_I\right)=\operatorname{Var}\left(\left.\sum_{i=1}^{I}\mathbb{E}\left[C_{i,J}\mid D_{I+1}\right]\right|D_I\right). \tag*{(30)} msep CDR ∣ D I ( 0 ) = E [ CDR 2 ∣ D I ] = Var ( CDR ∣ D I ) = Var ( i = 1 ∑ I E [ C i , J ∣ D I + 1 ] D I ) . (30)
Our goal is to compute this last expression.
Prediction uncertainty in the one-year update
We start by introducing the relevant notation. For i ∈ { I − J + 1 , … , I } i\in\{I-J+1,\ldots,I\} i ∈ { I − J + 1 , … , I } and j ∈ { d i + 1 , … , J − 1 } j\in\{d_i+1,\ldots,J-1\} j ∈ { d i + 1 , … , J − 1 } , we abbreviate
δ i , j = ∏ k = r j i − 1 Δ k , j and w j = ω r j , j = C r j , j / σ j 2 α ‾ r j , j + C r j , j / σ j 2 ∈ ( 0 , 1 ) . \delta_{i,j}=\prod_{k=r_j}^{i-1}\Delta_{k,j}
\qquad\text{and}\qquad
w_j=\omega_{r_j,j}=\frac{C_{r_j,j}/\sigma_j^2}{\overline{\alpha}_{r_j,j}+C_{r_j,j}/\sigma_j^2}\in(0,1). δ i , j = k = r j ∏ i − 1 Δ k , j and w j = ω r j , j = α r j , j + C r j , j / σ j 2 C r j , j / σ j 2 ∈ ( 0 , 1 ) .
This allows us to define the following D I D_I D I -measurable variables for j > d i j>d_i j > d i
ν i , j = δ i , j ( 1 − w j ) f ‾ r j , j + ( 1 − δ i , j ) f j and κ i , j = δ i , j w j C r j , j . \nu_{i,j}=\delta_{i,j}(1-w_j)\overline{f}_{r_j,j}+(1-\delta_{i,j})f_j
\qquad\text{and}\qquad
\kappa_{i,j}=\delta_{i,j}\frac{w_j}{C_{r_j,j}}. ν i , j = δ i , j ( 1 − w j ) f r j , j + ( 1 − δ i , j ) f j and κ i , j = δ i , j C r j , j w j .
Moreover, we set the boundary definitions
( ν i , j , κ i , j ) = { ( 0 , 1 ) for j = d i , ( 1 , 0 ) for j < d i . (\nu_{i,j},\kappa_{i,j})=
\begin{cases}
(0,1) & \text{for }j=d_i,\\
(1,0) & \text{for }j<d_i.
\end{cases} ( ν i , j , κ i , j ) = { ( 0 , 1 ) ( 1 , 0 ) for j = d i , for j < d i .
Under Assumptions 2.1, and in complete analogy to (4.3)-(4.4), the ( I + 1 ) (I+1) ( I + 1 ) -observable CL factors, for i ≥ I − J + 1 = r J − 1 i\ge I-J+1=r_{J-1} i ≥ I − J + 1 = r J − 1 and j > d i j>d_i j > d i , are given by
f ^ i , j ( I + 1 ) = ( ∏ k = r j + 1 i − 1 Δ k , j ) f ‾ r j + 1 , j + ( 1 − ∏ k = r j + 1 i − 1 Δ k , j ) f j = ν i , j + κ i , j C r j , j + 1 . \widehat{f}^{(I+1)}_{i,j}
=
\left(\prod_{k=r_j+1}^{i-1}\Delta_{k,j}\right)\overline{f}_{r_j+1,j}
+
\left(1-\prod_{k=r_j+1}^{i-1}\Delta_{k,j}\right)f_j
=
\nu_{i,j}+\kappa_{i,j}C_{r_j,j+1}. f i , j ( I + 1 ) = k = r j + 1 ∏ i − 1 Δ k , j f r j + 1 , j + 1 − k = r j + 1 ∏ i − 1 Δ k , j f j = ν i , j + κ i , j C r j , j + 1 .
This allows us to rewrite the ultimate claim predictors at time I + 1 I+1 I + 1 for i ≥ I − J + 1 i\ge I-J+1 i ≥ I − J + 1 by
E [ C i , J ∣ D I + 1 ] = C i , d i + 1 ∏ j = d i + 1 J − 1 f ^ i , j ( I + 1 ) = ∏ j = 0 J − 1 ( ν i , j + κ i , j C r j , j + 1 ) . (31) \mathbb{E}\left[C_{i,J}\mid D_{I+1}\right]
=
C_{i,d_i+1}\prod_{j=d_i+1}^{J-1}\widehat{f}^{(I+1)}_{i,j}
=
\prod_{j=0}^{J-1}\left(\nu_{i,j}+\kappa_{i,j}C_{r_j,j+1}\right). \tag*{(31)} E [ C i , J ∣ D I + 1 ] = C i , d i + 1 j = d i + 1 ∏ J − 1 f i , j ( I + 1 ) = j = 0 ∏ J − 1 ( ν i , j + κ i , j C r j , j + 1 ) . (31)
This one-year update I → I + 1 I \to I+1 I → I + 1 considers the diagonal observations ( C r j + 1 , j ) j = 1 J (C_{r_{j+1},j})_{j=1}^{J} ( C r j + 1 , j ) j = 1 J , and these are conditionally independent, given D I D_I D I ,
f ( ( C r j + 1 , j ) j = 1 , … , J ∣ D I ) = ∏ j = 1 J K ( C r j + 1 , j ∣ C r j + 1 , j − 1 / σ j − 1 2 , a r j + 1 , j − 1 − , σ j − 1 2 b r j + 1 , j − 1 − ) . (32) f\left((C_{r_{j+1},j})_{j=1,\ldots,J}\mid D_I\right)=\prod_{j=1}^{J}K\left(C_{r_{j+1},j}\mid C_{r_{j+1},j-1}/\sigma_{j-1}^{2},a^{-}_{r_{j+1},j-1},\sigma_{j-1}^{2}b^{-}_{r_{j+1},j-1}\right). \tag*{(32)} f ( ( C r j + 1 , j ) j = 1 , … , J ∣ D I ) = j = 1 ∏ J K ( C r j + 1 , j ∣ C r j + 1 , j − 1 / σ j − 1 2 , a r j + 1 , j − 1 − , σ j − 1 2 b r j + 1 , j − 1 − ) . (32)
This follows directly from Lemma 4.1. This conditional independence property is the key for the closed-form solution of the conditional MSEP of the CDR.
Using similar arguments as in (17) yields
μ j : = E [ C r j , j + 1 ∣ D I ] = f r j , j − C r j , j , (33) \mu_j:=\mathbb{E}\left[C_{r_j,j+1}\mid D_I\right]=f^{-}_{r_j,j}C_{r_j,j}, \tag*{(33)} μ j := E [ C r j , j + 1 ∣ D I ] = f r j , j − C r j , j , (33)
τ j 2 : = E [ C r j , j + 1 2 ∣ D I ] = ( f r j , j − ) 2 a r j , j − a r j , j − − 1 ( C r j , j 2 + σ j 2 C r j , j ) , (34) \tau_j^2:=\mathbb{E}\left[C_{r_j,j+1}^2\mid D_I\right]=\left(f^{-}_{r_j,j}\right)^2\frac{a^{-}_{r_j,j}}{a^{-}_{r_j,j}-1}\left(C_{r_j,j}^2+\sigma_j^2C_{r_j,j}\right), \tag*{(34)} τ j 2 := E [ C r j , j + 1 2 ∣ D I ] = ( f r j , j − ) 2 a r j , j − − 1 a r j , j − ( C r j , j 2 + σ j 2 C r j , j ) , (34)
where the second line requires a r j , j − > 1 a^{-}_{r_j,j}>1 a r j , j − > 1 .
Theorem 6.1 Under Assumptions 2.1 and a r j , j − > 1 a^{-}_{r_j,j}>1 a r j , j − > 1 for all j ∈ { 0 , … , J − 1 } j\in\{0,\ldots,J-1\} j ∈ { 0 , … , J − 1 }
msep C D R ∣ D I ( 0 ) = ∑ i = I − J + 1 I ∑ k = I − J + 1 I ( ∏ j = 0 J − 1 T i , k , j − C ^ i , J C ^ k , J ) , \operatorname{msep}_{\mathrm{CDR}\mid D_I}(0)=\sum_{i=I-J+1}^{I}\sum_{k=I-J+1}^{I}\left(\prod_{j=0}^{J-1}T_{i,k,j}-\widehat{C}_{i,J}\widehat{C}_{k,J}\right), msep CDR ∣ D I ( 0 ) = i = I − J + 1 ∑ I k = I − J + 1 ∑ I ( j = 0 ∏ J − 1 T i , k , j − C i , J C k , J ) , where the ultimate claims predictions are computed from Theorem 4.4 and with
T i , k , j = ν i , j ν k , j + ( ν i , j κ k , j + ν k , j κ i , j ) μ j + κ i , j κ k , j τ j 2 . T_{i,k,j}=\nu_{i,j}\nu_{k,j}+(\nu_{i,j}\kappa_{k,j}+\nu_{k,j}\kappa_{i,j})\mu_j+\kappa_{i,j}\kappa_{k,j}\tau_j^2. T i , k , j = ν i , j ν k , j + ( ν i , j κ k , j + ν k , j κ i , j ) μ j + κ i , j κ k , j τ j 2 .
The key difference between the one-year MSEP and the total MSEP (5.1) is that in the former we obtain the conditional independence decoupling (6.3) which allows for closed-from solutions. This does independent decoupling, given D I D_I D I , does not hold in the total MSEP case.
The exponential seniority weighting case
In the exponential seniority weighting case of § 4.3 , we receive an interpretable version of Theorem 6.1 that is in the spirit of [21 Theorem 2.15] . That is, consider the special case p i , j ≡ 0 p_{i,j}\equiv0 p i , j ≡ 0 and q i , j ≡ q j ≤ 1 q_{i,j}\equiv q_j\leq1 q i , j ≡ q j ≤ 1 . This implies, see (23) ,
A j : = a r j , j − = ∑ k = 1 r j − 1 q j r j − k C k , j σ j 2 + q j r j − 1 γ j , A_j:=a^{-}_{r_j,j}=\sum_{k=1}^{r_j-1}q_j^{r_j-k}\frac{C_{k,j}}{\sigma_j^2}+q_j^{r_j-1}\gamma_j, A j := a r j , j − = k = 1 ∑ r j − 1 q j r j − k σ j 2 C k , j + q j r j − 1 γ j ,
B j : = b r j , j − = ∑ k = 1 r j − 1 q j r j − k C k , j + 1 σ j 2 + q j r j − 1 γ j f j , B_j:=b^{-}_{r_j,j}=\sum_{k=1}^{r_j-1}q_j^{r_j-k}\frac{C_{k,j+1}}{\sigma_j^2}+q_j^{r_j-1}\gamma_j f_j, B j := b r j , j − = k = 1 ∑ r j − 1 q j r j − k σ j 2 C k , j + 1 + q j r j − 1 γ j f j ,
with f ^ j ( I ) = B j / A j \widehat{f}^{(I)}_j=B_j/A_j f j ( I ) = B j / A j . Introduce for A j > 1 A_j>1 A j > 1 the D I D_I D I -measurable variables
Ψ j = 1 A j − 1 = σ j 2 ∑ k = 1 r j − 1 q j r j − k C k , j + σ j 2 ( q j r j − 1 γ j − 1 ) , \Psi_j=\frac{1}{A_j-1}=\frac{\sigma_j^2}{\sum_{k=1}^{r_j-1}q_j^{r_j-k}C_{k,j}+\sigma_j^2\left(q_j^{r_j-1}\gamma_j-1\right)}, Ψ j = A j − 1 1 = ∑ k = 1 r j − 1 q j r j − k C k , j + σ j 2 ( q j r j − 1 γ j − 1 ) σ j 2 ,
ω j ⋆ = C r j , j ∑ ℓ = 1 r j q j r j − ℓ C ℓ , j + q j r j − 1 σ j 2 γ j = C r j , j σ j 2 A j + C r j , j , \omega_j^\star=\frac{C_{r_j,j}}{\sum_{\ell=1}^{r_j}q_j^{r_j-\ell}C_{\ell,j}+q_j^{r_j-1}\sigma_j^2\gamma_j}=\frac{C_{r_j,j}}{\sigma_j^2A_j+C_{r_j,j}}, ω j ⋆ = ∑ ℓ = 1 r j q j r j − ℓ C ℓ , j + q j r j − 1 σ j 2 γ j C r j , j = σ j 2 A j + C r j , j C r j , j ,
the latter stems from weight ω r j , j \omega_{r_j,j} ω r j , j given in (4.7).
Theorem 6.2 Consider the special case p i , j ≡ 0 p_{i,j} \equiv0 p i , j ≡ 0 and q i , j ≡ q j ≤ 1 q_{i,j} \equiv q_j \le1 q i , j ≡ q j ≤ 1 . For accident years i + J > I i+J>I i + J > I
msep C D R i ∣ D I ( 0 ) = ( C ^ i , J ) 2 [ ( 1 + Ψ d i ω d i ⋆ ) ∏ j = d i + 1 J − 1 ( 1 + ω j ⋆ Ψ j ) − 1 ] , \operatorname{msep}_{\mathrm{CDR}_i \mid\mathcal{D}_I}(0) = \left(\widehat{C}_{i,J}\right)^2 \left[\left(1+\frac{\Psi_{d_i}}{\omega_{d_i}^{\star}}\right)\prod_{j=d_i+1}^{J-1}\left(1+\omega_j^{\star}\Psi_j\right)-1\right], msep CDR i ∣ D I ( 0 ) = ( C i , J ) 2 ( 1 + ω d i ⋆ Ψ d i ) j = d i + 1 ∏ J − 1 ( 1 + ω j ⋆ Ψ j ) − 1 , where we assume that A j > 1 A_j>1 A j > 1 for all d i ≤ j ≤ J − 1 d_i\le j\le J-1 d i ≤ j ≤ J − 1 , otherwise the corresponding conditional MSEP is infinite, and for aggregated accident years
msep C D R ∣ D I ( 0 ) = ∑ i msep C D R i ∣ D I ( 0 ) + 2 ∑ i < m C ^ i , J C ^ m , J [ ( 1 + Ψ d i ) ∏ j = d i + 1 J − 1 ( 1 + ω j ⋆ Ψ j ) − 1 ] , \operatorname{msep}_{\mathrm{CDR}\mid\mathcal{D}_I}(0)=\sum_i\operatorname{msep}_{\mathrm{CDR}_i\mid\mathcal{D}_I}(0)+2\sum_{i<m}\widehat{C}_{i,J}\widehat{C}_{m,J}\left[\left(1+\Psi_{d_i}\right)\prod_{j=d_i+1}^{J-1}\left(1+\omega_j^{\star}\Psi_j\right)-1\right], msep CDR ∣ D I ( 0 ) = i ∑ msep CDR i ∣ D I ( 0 ) + 2 i < m ∑ C i , J C m , J ( 1 + Ψ d i ) j = d i + 1 ∏ J − 1 ( 1 + ω j ⋆ Ψ j ) − 1 , where the summations run over I − J + 1 ≤ i ≤ I I-J+1\le i\le I I − J + 1 ≤ i ≤ I and I − J + 1 ≤ i < m ≤ I I-J+1\le i<m\le I I − J + 1 ≤ i < m ≤ I , respectively.
This theorem generalizes Wüthrich–Merz [21 Theorem 2.15] to the exponential seniority weighting case, and we can also consider the non-informative prior limit γ j ↓ 0 \gamma_j\downarrow0 γ j ↓ 0 to get the analogue to Wüthrich–Merz [21 formula (2.25)] for general q j ≤ 1 q_j\le1 q j ≤ 1 .
Example
We select the triangular data example for D I \mathcal{D}_I D I , I = J + 1 = 10 I=J+1=10 I = J + 1 = 10 , from Wüthrich–Merz [21 Table 1.4] ; the data are provided in Table 5 in the appendix below. For this dataset, we compute Mack’s [11 ] CL reserves, Mack’s conditional MSEP estimate, and the non-informative prior gamma-gamma Bayes’ MSEP estimate; see Wüthrich–Merz [21 Table 2.4] . We estimate the variance parameters σ j 2 \sigma_j^2 σ j 2 as in Wüthrich–Merz [21 Table 2.3] , and we keep these estimates fixed for all computations; the square roots σ ^ j \widehat{\sigma}_j σ j is given in Table 5 in the appendix.
Table 5. Observed cumulative claims C i , j C_{i,j} C i , j , i + j ≤ I i+j\le I i + j ≤ I , the classical Mack’s CL factors f ^ j ( I ) \hat{f}^{(I)}_j f ^ j ( I ) given in (27) , and the (rooted) variance estimates taken from Wüthrich–Merz [21 Table 2.3] .
In this example we consider p i , j ≡ 0 p_{i,j}\equiv0 p i , j ≡ 0 , q i , j ≡ q ∈ ( 0 , 1 ] q_{i,j}\equiv q\in(0,1] q i , j ≡ q ∈ ( 0 , 1 ] , and the non-informative prior limits γ j → 0 \gamma_j\to0 γ j → 0 . The reserves are then for accident years i + J > I i+J>I i + J > I given by
R ^ i C L ( q ) = C ^ i , J C L ( q ) − C i , d i = lim γ ↓ 0 E γ [ C i , J ∣ D I ] − C i , d i , \widehat{R}_i^{\mathrm{CL}(q)}=\widehat{C}_{i,J}^{\mathrm{CL}(q)}-C_{i,d_i}=\lim_{\gamma\downarrow0}\mathbb{E}_{\gamma}\left[C_{i,J}\mid\mathcal{D}_I\right]-C_{i,d_i}, R i CL ( q ) = C i , J CL ( q ) − C i , d i = γ ↓ 0 lim E γ [ C i , J ∣ D I ] − C i , d i ,
and aggregated over all accident years, we obtain R ^ ∙ C L ( q ) = ∑ i = I − J + 1 I R ^ i C L ( q ) \widehat{R}_{\bullet}^{\mathrm{CL}(q)}=\sum_{i=I-J+1}^{I}\widehat{R}_i^{\mathrm{CL}(q)} R ∙ CL ( q ) = ∑ i = I − J + 1 I R i CL ( q ) . For q = 1 q=1 q = 1 , we obtain Mack’s [11 ] CL reserves. The results are presented in Table 1 .
Table 1. Mack’s [11 ] CL reserves and RMSEPs, non-informative prior gamma-gamma Bayes’ CL RMSEPs of Wüthrich–Merz [21 Section 2.2] , Monte Carlo RMSEP approximations from Algorithm 1 and Monte Carlo errors C ^ i − C i \widehat{C}_i-C_i C i − C i in reserves using one million Monte Carlo samples.
The first two numerical columns of Table 1 show Mack’s [11 ] CL reserves R ^ i C L ( 1 ) \widehat{R}_i^{\mathrm{CL}(1)} R i CL ( 1 ) and the resulting rooted mean squared error of prediction (RMSEP), and the third column gives the RMSEP of the non-informative prior gamma-gamma Bayes’ CL model. This latter model provides the identical CL reserves, and its RMSEP is lower bounded by Mack’s RMSEP; see Wüthrich–Merz [21 Section 2.2.3] . The numbers in Table 1 verify this lower bound property, as the third column dominates the second one.
For q = 1 q=1 q = 1 , Algorithm 1 should reproduce the non-informative gamma-gamma Bayes’ CL results, up to a Monte Carlo simulation error. This is verified by the last two columns of Table 1 which are based on one million Monte Carlo samples; note that the Monte Carlo errors can directly be compared to the reserves R ^ i C L ( 1 ) \widehat{R}_i^{\mathrm{CL}(1)} R i CL ( 1 ) , because the payments C i , d i C_{i,d_i} C i , d i are known at time I I I .
The last two columns of Table 2 give a non-informative prior example with constant exponential seniority weighting q i , j ≡ q = 0.9 q_{i,j}\equiv q=0.9 q i , j ≡ q = 0.9 (and p i , j ≡ 0 p_{i,j}\equiv0 p i , j ≡ 0 ), i.e., this applies an exponential seniority weighting as illustrated in Figure 5 (rhs). The resulting reserves R ^ ∙ C L ( q ) \widehat{R}^{\mathrm{CL}(q)}_{\bullet} R ∙ CL ( q ) are smaller than Mack’s CL reserves R ^ ∙ C L ( 1 ) \widehat{R}^{\mathrm{CL}(1)}_{\bullet} R ∙ CL ( 1 ) . This comes from the fact that the observations C i , j + 1 / C i , j C_{i,j+1}/C_{i,j} C i , j + 1 / C i , j have a slightly decreasing trend in the left-lower triangle of the observations D I D_I D I , see Table 3 , which leads to smaller CL factor estimates through the exponential seniority weighting.
Table 2. Mack’s [11 ] CL reserves and RMSEPs, non-informative prior gamma-gamma Bayes’ CL RMSEPs of Wüthrich–Merz [21 Section 2.2] ; exponential seniority weighted CL reserves with q = 0.9 q=0.9 q = 0.9 and resulting RMSEP from Algorithm 1 using one million Monte Carlo samples.
Table 3. Individual observations C i , j + 1 / C i , j C_{i,j+1}/C_{i,j} C i , j + 1 / C i , j in D I D_I D I .
Interestingly, the conditional RMSEPs in the last column of Table 2 increase for q = 0.9 q=0.9 q = 0.9 compared to the CL case q = 1 q=1 q = 1 . Intuitively, this is explained by the fact that an exponential seniority weighting assigns less weight to older claims, hence, relying more on the most recent claims. This can be interpreted as a decrease of the amount of past information used, which leads to a bigger uncertainty in the CL factor estimation. For small q q q , we almost exclusively rely on the latest observed diagonal i + j = I i+j=I i + j = I , i.e., in that case, the CL factor f ^ j ( I ) \widehat{f}^{(I)}_j f j ( I ) is estimated almost entirely from a single observation C r j − 1 , j + 1 / C r j − 1 , j C_{r_{j-1},j+1}/C_{r_{j-1},j} C r j − 1 , j + 1 / C r j − 1 , j , see (28) . Figure 6 (rhs) shows the resulting exponential seniority weights using q ∈ { 0.70 , 0.75 , … , 0.95 } q \in\{0.70,0.75,\ldots,0.95\} q ∈ { 0.70 , 0.75 , … , 0.95 } . We observe that for q = 0.70 q=0.70 q = 0.70 , observations with a time lag of more than 4 periods to the last observed diagonal receive weights of less than 20%.
Figure 6. (lhs) Total RMSEPs as a function of q ∈ { 0.70 , 0.75 , … , 1.00 } q \in\{0.70, 0.75, \ldots, 1.00\} q ∈ { 0.70 , 0.75 , … , 1.00 } , (rhs) exponential seniority weights q k q^k q k , for k ∈ { 0 , … , 9 } k \in\{0, \ldots, 9\} k ∈ { 0 , … , 9 } and q ∈ { 0.70 , 0.75 , … , 0.95 } q \in\{0.70, 0.75, \ldots, 0.95\} q ∈ { 0.70 , 0.75 , … , 0.95 } .
Figure 6 (lhs) shows the total RMSEPs for different seniority weighting q ∈ { 0.70 , 0.75 , … , 1.00 } q \in\{0.70,0.75,\ldots,1.00\} q ∈ { 0.70 , 0.75 , … , 1.00 } : the smaller the weight q q q , the less weight is assigned to older observations, and the bigger the RMSEP becomes, because older observations are less informative. If we use the interpretation of the gamma-thinning construction presented in [1 ] , this is justified by the fact that the state-space update is thinning past information and it is adding new noise to the latent (memory) process Θ 1 : I , j \Theta_{1:I,j} Θ 1 : I , j .
In view of the above results, there is the question of an optimal choice of the parameter q ∈ ( 0 , 1 ] q \in(0, 1] q ∈ ( 0 , 1 ] . So far, it has been treated as a hyper-parameter selected by the modeler. An elegant approach is the re-reserving and model validation method by [4 ] . The idea is to work through the triangles/trapezoids
D t = σ { C i , j ; i + j ≤ t , 1 ≤ i ≤ I , 0 ≤ j ≤ J } for t ∈ { 1 , … , I − 1 } , D_t = \sigma\{C_{i,j};\ i+j \le t,\ 1 \le i \le I,\ 0 \le j \le J\}\qquad\text{for }t \in\{1,\ldots,I-1\}, D t = σ { C i , j ; i + j ≤ t , 1 ≤ i ≤ I , 0 ≤ j ≤ J } for t ∈ { 1 , … , I − 1 } ,
by making predictions based on D t D_t D t and validating them against D t + 1 D_{t+1} D t + 1 , that is, by successively adding new diagonals to the prediction and validation process (note that there are not necessarily observations for all development periods j j j ). This can be done for different values of q q q , and the one that has the best validation performance is the one selected for extrapolation beyond time I I I . In the same way, the band in Figure 1 (rhs) can be selected. Note that this is an out-of-sample validation, because one validates against new observations in D t + 1 ∖ D t D_{t+1} \setminus D_t D t + 1 ∖ D t and the resulting predictions at time t + 1 t+1 t + 1 .
Having an analytical full data likelihood, see Corollary 4.2 , one can alternatively perform maximum likelihood estimation (MLE). To keep the estimation complexity under control, we use the CL estimates for the variance parameters σ j 2 \sigma_j^2 σ j 2 , given in Table 5 . This is an ad-hoc solution because these parameter estimates are based on the choice q = 1 q=1 q = 1 , and alternatively one could employ an iterative estimation procedure. For simplicity, we do not do this here. Thus, we fix σ j 2 = σ ^ j 2 \sigma_j^2=\widehat{\sigma}_j^2 σ j 2 = σ j 2 , we set p i , j ≡ 0 p_{i,j}\equiv0 p i , j ≡ 0 , we consider the non-informative prior limit γ j ↓ 0 \gamma_j\downarrow0 γ j ↓ 0 , and we optimize the resulting data log-likelihood over q i , j ≡ q ∈ ( 0 , 1 ] q_{i,j}\equiv q\in(0,1] q i , j ≡ q ∈ ( 0 , 1 ] . It is given by
q ↦ ℓ D I ( q ) = ∑ j = 1 J ∑ i = 2 r j log K ( C i , j | C i , j − 1 / σ ^ j 2 , a i , j − 1 − , σ ^ j 2 b i , j − 1 − ) . (35) q \mapsto\ell_{D_I}(q)=\sum_{j=1}^{J}\sum_{i=2}^{r_j}\log K\left(C_{i,j}\middle|C_{i,j-1}/\widehat{\sigma}_j^2,\ a_{i,j-1}^{-},\ \widehat{\sigma}_j^2 b_{i,j-1}^{-}\right). \tag*{(35)} q ↦ ℓ D I ( q ) = j = 1 ∑ J i = 2 ∑ r j log K ( C i , j C i , j − 1 / σ j 2 , a i , j − 1 − , σ j 2 b i , j − 1 − ) . (35)
with a 2 , j − = q C 1 , j / σ ^ j 2 a_{2,j}^{-}=qC_{1,j}/\widehat{\sigma}_j^2 a 2 , j − = q C 1 , j / σ j 2 and b 2 , j − = q C 1 , j + 1 / σ ^ j 2 b_{2,j}^{-}=qC_{1,j+1}/\widehat{\sigma}_j^2 b 2 , j − = q C 1 , j + 1 / σ j 2 , and further updates, see (5) –(7) ,
a i + 1 , j − = q ( a i , j − + C i , j / σ ^ j 2 ) and b i + 1 , j − = q ( b i , j − + C i , j + 1 / σ ^ j 2 ) . a_{i+1,j}^{-}=q\left(a_{i,j}^{-}+C_{i,j}/\widehat{\sigma}_j^2\right)
\qquad\text{and}\qquad
b_{i+1,j}^{-}=q\left(b_{i,j}^{-}+C_{i,j+1}/\widehat{\sigma}_j^2\right). a i + 1 , j − = q ( a i , j − + C i , j / σ j 2 ) and b i + 1 , j − = q ( b i , j − + C i , j + 1 / σ j 2 ) .
This corresponds to the non-informative prior limit γ j ↓ 0 \gamma_j\downarrow0 γ j ↓ 0 , which requires us to drop accident year i = 1 i=1 i = 1 from the log-likelihood (35) .
Numerical maximization provides q ^ M L E = 0.717 \widehat{q}^{\mathrm{MLE}}=0.717 q MLE = 0.717 , with a log-likelihood value of − 447.36 -447.36 − 447.36 , the choice q = 0.9 q=0.9 q = 0.9 gives a log-likelihood value of − 448.05 -448.05 − 448.05 , and q = 1.0 q=1.0 q = 1.0 a value of − 448.87 -448.87 − 448.87 . Thus, we should select a q q q -value that is close to the red curve q = 0.70 q=0.70 q = 0.70 in Figure 6 (rhs).
Next, we turn our attention to the one-year CDR analysis. The results are presented in Table 4 , and they are essentially the same as for the total RMSEP. Namely, the Bayes’ version with q = 1 q=1 q = 1 dominates the Merz–Wüthrich (MW) CDR formula [14 ] , decreasing q q q increases the CDR RMSEP, and generally, we have
Table 4. Claims reserves and CDR RMSEP: Merz–Wüthrich (MW) formula [14 ] , gamma-gamma Bayes’ CL CDR uncertainty Wüthrich–Merz [21 Theorem 2.15] , this corresponds to q = 1 q=1 q = 1 , and CDR RMSEPs for q = 0.9 q=0.9 q = 0.9 using Theorem 6.2.
rmsep C ∙ ∣ D I ( C ^ ∙ ) ≥ rmsep C D R ∣ D I ( 0 ) , \operatorname{rmsep}_{C_\bullet\mid D_I}\left(\widehat{C}_\bullet\right)\geq\operatorname{rmsep}_{\mathrm{CDR}\mid D_I}(0), rmsep C ∙ ∣ D I ( C ∙ ) ≥ rmsep CDR ∣ D I ( 0 ) ,
which follows from a martingale argument; see Wüthrich–Merz [21, Corollary 1.2]. The gap between the two RMSEPs is relatively small in our example because the data shows a fast run-off, see Table 5 . This concludes the example.
Conclusion
We presented an observation-driven state-space version of the gamma-gamma Bayes’ CL model of Wüthrich–Merz [21 ] . This observation-driven state-space dynamics considers a gamma latent stochastic process as proposed in Ahn et al. [2 ] . This leads to a full analytical representation of the data likelihood, which allows us to compute the resulting claims reserves, the conditional mean squared error of prediction of the ultimate claims can be determined by a simple Monte Carlo simulation, and the one-year claims development result uncertainty is available in closed form.
This observation-driven gamma-gamma state-space model has the properties that it can replicate the chain-ladder reserves under non-informative priors and under specific hyper-parameter choices. These hyper-parameters also admit an exponential seniority weighting in past information entering the chain-ladder computation, and the tractable data likelihood even allows one to estimate the optimal seniority weights with maximum likelihood estimation.
To the best of our knowledge, this is the first fully tractable model that allows for seniority weighting and at the same time it obeys the crucial assumptions of Mack’s [11 ] distribution-free chain-ladder model, in particular, the proportional linear conditional variance behavior.
There are similar results on the chain-ladder factor estimation side, see Marais–Grove [13 ] , but these results are not obtained in a full stochastic reserving model, they are rather enforced manually. Therefore, they do not allow for a consistent computation of the prediction uncertainty. All our results are a natural consequence of the stated model assumptions.
AI declaration. The author developed the concept of the paper, all mathematical results, and he has fully written the first version of the paper. ChatGPT-6 was used to improve the manuscript, to check the mathematical results and to implement the examples.
Proofs
In this appendix, we give all mathematical proofs.
Proof of Lemma 3.1. We have B j ⊂ H i , j \mathcal{B}_j \subset\mathcal{H}_{i,j} B j ⊂ H i , j , in particular, all cumulative claims C 1 : I , j C_{1:I,j} C 1 : I , j are given to model the next ones C i , j + 1 C_{i,j+1} C i , j + 1 for i ∈ { 1 , … , I } i \in\{1,\ldots,I\} i ∈ { 1 , … , I } . From (16) , we obtain for i ∈ { 1 , … , I − 1 } i \in\{1,\ldots,I-1\} i ∈ { 1 , … , I − 1 } , since ω i , j \omega_{i,j} ω i , j are B j \mathcal{B}_j B j -measurable,
E [ Θ i + 1 , j − 1 ∣ B j ] = Δ i , j ( ω i , j E [ C i , j + 1 C i , j ∣ B j ] + ( 1 − ω i , j ) E [ f i , j − ∣ B j ] ) + ( 1 − Δ i , j ) f j = Δ i , j ( ω i , j E [ 1 C i , j E [ C i , j + 1 ∣ H i − 1 , j , G i , j ] ∣ B j ] + ( 1 − ω i , j ) E [ Θ i , j − 1 ∣ B j ] ) + ( 1 − Δ i , j ) f j = Δ i , j E [ Θ i , j − 1 ∣ B j ] + ( 1 − Δ i , j ) f j . \begin{aligned}
\mathbb{E}\left[\Theta^{-1}_{i+1,j}\mid\mathcal{B}_j\right]
= \Delta_{i,j}\left(\omega_{i,j}\mathbb{E}\left[\left.\frac{C_{i,j+1}}{C_{i,j}}\right|\mathcal{B}_j\right]+(1-\omega_{i,j})\mathbb{E}\left[\left.f^{-}_{i,j}\right|\mathcal{B}_j\right]\right)+(1-\Delta_{i,j})f_j \\
&= \Delta_{i,j}\left(\omega_{i,j}\mathbb{E}\left[\left.\frac{1}{C_{i,j}}\mathbb{E}\left[C_{i,j+1}\mid\mathcal{H}_{i-1,j},\mathcal{G}_{i,j}\right]\right|\mathcal{B}_j\right]+(1-\omega_{i,j})\mathbb{E}\left[\left.\Theta^{-1}_{i,j}\right|\mathcal{B}_j\right]\right)+(1-\Delta_{i,j})f_j \\
&= \Delta_{i,j}\mathbb{E}\left[\left.\Theta^{-1}_{i,j}\right|\mathcal{B}_j\right]+(1-\Delta_{i,j})f_j.
\end{aligned} E [ Θ i + 1 , j − 1 ∣ B j ] = Δ i , j ( ω i , j E [ C i , j C i , j + 1 B j ] + ( 1 − ω i , j ) E [ f i , j − B j ] ) + ( 1 − Δ i , j ) f j = Δ i , j ( ω i , j E [ C i , j 1 E [ C i , j + 1 ∣ H i − 1 , j , G i , j ] B j ] + ( 1 − ω i , j ) E [ Θ i , j − 1 B j ] ) + ( 1 − Δ i , j ) f j = Δ i , j E [ Θ i , j − 1 B j ] + ( 1 − Δ i , j ) f j .
The proof then follows by induction because E [ Θ 1 , j − 1 ∣ B j ] = E [ Θ 1 , j − 1 ∣ H 0 , j ] = f j \mathbb{E}[\Theta^{-1}_{1,j}\mid\mathcal{B}_j]=\mathbb{E}[\Theta^{-1}_{1,j}\mid\mathcal{H}_{0,j}]=f_j E [ Θ 1 , j − 1 ∣ B j ] = E [ Θ 1 , j − 1 ∣ H 0 , j ] = f j .
Proof of Lemma 4.1. Select j ∈ { 0 , … , J − 1 } j \in\{0,\ldots,J-1\} j ∈ { 0 , … , J − 1 } . We begin by considering the density of C 1 : I , j + 1 C_{1:I,j+1} C 1 : I , j + 1 , given B j \mathcal{B}_j B j ,
f ( C 1 : I , j + 1 ∣ B j ) = f ( C 1 : I , j + 1 ∣ H I − 1 , j ) f ( C 1 : I − 1 , j + 1 ∣ B j ) = ∏ i = 1 I f ( C i , j + 1 ∣ H i − 1 , j ) . (36) f\left(C_{1:I,j+1}\mid\mathcal{B}_j\right)=f\left(C_{1:I,j+1}\mid\mathcal{H}_{I-1,j}\right)f\left(C_{1:I-1,j+1}\mid\mathcal{B}_j\right)=\prod_{i=1}^{I}f\left(C_{i,j+1}\mid\mathcal{H}_{i-1,j}\right). \tag*{(36)} f ( C 1 : I , j + 1 ∣ B j ) = f ( C 1 : I , j + 1 ∣ H I − 1 , j ) f ( C 1 : I − 1 , j + 1 ∣ B j ) = i = 1 ∏ I f ( C i , j + 1 ∣ H i − 1 , j ) . (36)
Next, we compute each term in the product of (A.1)
f ( C i , j + 1 ∣ H i − 1 , j ) = ∫ 0 ∞ f ( C i , j + 1 ∣ H i − 1 , j , Θ i , j ) f ( Θ i , j ∣ H i − 1 , j ) d Θ i , j = ∫ 0 ∞ ( Θ i , j / σ j 2 ) C i , j / σ j 2 Γ ( C i , j / σ j 2 ) C i , j + 1 C i , j / σ j 2 − 1 e − Θ i , j / σ j 2 C i , j + 1 ( b i , j − ) 1 + a i , j − Γ ( 1 + a i , j − ) Θ i , j a i , j − e − b i , j − Θ i , j d Θ i , j = K ( C i , j + 1 ∣ C i , j / σ j 2 , a i , j − , σ j 2 b i , j − ) . \begin{aligned}
f\left(C_{i,j+1}\mid\mathcal{H}_{i-1,j}\right)=\int_{0}^{\infty}f\left(C_{i,j+1}\mid\mathcal{H}_{i-1,j},\Theta_{i,j}\right)f\left(\Theta_{i,j}\mid\mathcal{H}_{i-1,j}\right)\,\mathrm{d}\Theta_{i,j} \\
&=\int_{0}^{\infty}\frac{\left(\Theta_{i,j}/\sigma_j^2\right)^{C_{i,j}/\sigma_j^2}}{\Gamma(C_{i,j}/\sigma_j^2)}C_{i,j+1}^{C_{i,j}/\sigma_j^2-1}e^{-\Theta_{i,j}/\sigma_j^2 C_{i,j+1}}\frac{(b_{i,j}^{-})^{1+a_{i,j}^{-}}}{\Gamma(1+a_{i,j}^{-})}\Theta_{i,j}^{a_{i,j}^{-}}e^{-b_{i,j}^{-}\Theta_{i,j}}\,\mathrm{d}\Theta_{i,j} \\
&=K\left(C_{i,j+1}\mid C_{i,j}/\sigma_j^2,a_{i,j}^{-},\sigma_j^2b_{i,j}^{-}\right).
\end{aligned} f ( C i , j + 1 ∣ H i − 1 , j ) = ∫ 0 ∞ f ( C i , j + 1 ∣ H i − 1 , j , Θ i , j ) f ( Θ i , j ∣ H i − 1 , j ) d Θ i , j = ∫ 0 ∞ Γ ( C i , j / σ j 2 ) ( Θ i , j / σ j 2 ) C i , j / σ j 2 C i , j + 1 C i , j / σ j 2 − 1 e − Θ i , j / σ j 2 C i , j + 1 Γ ( 1 + a i , j − ) ( b i , j − ) 1 + a i , j − Θ i , j a i , j − e − b i , j − Θ i , j d Θ i , j = K ( C i , j + 1 ∣ C i , j / σ j 2 , a i , j − , σ j 2 b i , j − ) .
The development year factorization follows from the definition. This completes the proof.
Proof of Proposition 4.3. We need to align the information set D I \mathcal{D}_I D I and H r j − 1 , j \mathcal{H}_{r_{j-1},j} H r j − 1 , j , see Figure 4 . We start from Corollary 4.2. The information set H r j − 1 , j \mathcal{H}_{r_{j-1},j} H r j − 1 , j contains the observations C 1 : r j − 1 , j + 1 = C 1 : r j + 1 , j + 1 C_{1:r_j-1,j+1}=C_{1:r_{j+1},j+1} C 1 : r j − 1 , j + 1 = C 1 : r j + 1 , j + 1 , this is the yellow column in Figure 4 (rhs). We condition on these observations in the data likelihood
f ( ( C i , k ) i + k ≤ I , k ≥ 1 , k ≠ j + 1 ∣ C 1 : I , 0 , C 1 : r j + 1 , j + 1 ) ∝ ∏ k = 1 j + 1 ∏ i = 1 r k K ( C i , k ∣ C i , k − 1 / σ k − 1 2 , a i , k − 1 − , σ k − 1 2 b i , k − 1 − ) f\left((C_{i,k})_{i+k\leq I,k\geq1,k\ne j+1}\mid C_{1:I,0},C_{1:r_{j+1},j+1}\right)\propto\prod_{k=1}^{j+1}\prod_{i=1}^{r_k}K\left(C_{i,k}\mid C_{i,k-1}/\sigma_{k-1}^2,a_{i,k-1}^{-},\sigma_{k-1}^2b_{i,k-1}^{-}\right) f ( ( C i , k ) i + k ≤ I , k ≥ 1 , k = j + 1 ∣ C 1 : I , 0 , C 1 : r j + 1 , j + 1 ) ∝ k = 1 ∏ j + 1 i = 1 ∏ r k K ( C i , k ∣ C i , k − 1 / σ k − 1 2 , a i , k − 1 − , σ k − 1 2 b i , k − 1 − )
× ∏ k = j + 2 J ∏ i = 1 r k K ( C i , k ∣ C i , k − 1 / σ k − 1 2 , a i , k − 1 − , σ k − 1 2 b i , k − 1 − ) . \qquad\qquad\qquad\qquad\qquad\qquad\qquad\times\prod_{k=j+2}^{J}\prod_{i=1}^{r_k}K\left(C_{i,k}\mid C_{i,k-1}/\sigma_{k-1}^2,a_{i,k-1}^{-},\sigma_{k-1}^2b_{i,k-1}^{-}\right). × k = j + 2 ∏ J i = 1 ∏ r k K ( C i , k ∣ C i , k − 1 / σ k − 1 2 , a i , k − 1 − , σ k − 1 2 b i , k − 1 − ) .
Thus, conditioning on C 1 : r j + 1 , j + 1 C_{1:r_{j+1},j+1} C 1 : r j + 1 , j + 1 leads to an independent decoupling of B j \mathcal{B}_j B j and the upper right trapezoid D I ; j + 2 : J = ( C i , k ) i + k ≤ I , j + 2 ≤ k ≤ J \mathcal{D}_{I;j+2:J}=(C_{i,k})_{i+k\leq I,\ j+2\leq k\leq J} D I ; j + 2 : J = ( C i , k ) i + k ≤ I , j + 2 ≤ k ≤ J . The same argument holds for B j ∨ σ ( C r j + 1 + 1 , j + 1 ) \mathcal{B}_j\vee\sigma(C_{r_{j+1}+1,j+1}) B j ∨ σ ( C r j + 1 + 1 , j + 1 ) being conditionally independent of the upper right trapezoid D I ; j + 2 : J \mathcal{D}_{I;j+2:J} D I ; j + 2 : J , given C 1 : r j + 1 , j + 1 C_{1:r_{j+1},j+1} C 1 : r j + 1 , j + 1 . Therefore, this upper right trapezoid does not impact the computation of f i , j − f^{-}_{i,j} f i , j − , given H r j + 1 , j \mathcal{H}_{r_{j+1},j} H r j + 1 , j . This conditional independent decoupling implies for i ≥ r j = r j + 1 + 1 i\geq r_j=r_{j+1}+1 i ≥ r j = r j + 1 + 1
E [ f i , j ∣ D I ] = E [ E [ f i , j ∣ H r j + 1 , j , D I ] ∣ D I ] = E [ E [ f i , j ∣ H r j + 1 , j ] ∣ D I ] = E [ ( ∏ k = r j + 1 + 1 i − 1 Δ k , j ) f r j + 1 + 1 , j − + ( 1 − ∏ k = r j + 1 + 1 i − 1 Δ k , j ) f j | D I ] , \begin{aligned}
\mathbb{E}\left[f_{i,j}\mid\mathcal{D}_I\right]=\mathbb{E}\left[\mathbb{E}\left[f_{i,j}\mid\mathcal{H}_{r_{j+1},j},\mathcal{D}_I\right]\mid\mathcal{D}_I\right]=\mathbb{E}\left[\mathbb{E}\left[f_{i,j}\mid\mathcal{H}_{r_{j+1},j}\right]\mid\mathcal{D}_I\right] \\
&=\mathbb{E}\left[\left(\prod_{k=r_{j+1}+1}^{i-1}\Delta_{k,j}\right)f^{-}_{r_{j+1}+1,j}+\left(1-\prod_{k=r_{j+1}+1}^{i-1}\Delta_{k,j}\right)f_j\middle|\mathcal{D}_I\right],
\end{aligned} E [ f i , j ∣ D I ] = E [ E [ f i , j ∣ H r j + 1 , j , D I ] ∣ D I ] = E [ E [ f i , j ∣ H r j + 1 , j ] ∣ D I ] = E k = r j + 1 + 1 ∏ i − 1 Δ k , j f r j + 1 + 1 , j − + 1 − k = r j + 1 + 1 ∏ i − 1 Δ k , j f j D I ,
the last step uses Corollary 3.2. The claim then follows because r j + 1 + 1 = r j r_{j+1}+1=r_j r j + 1 + 1 = r j and because all terms in the last expression are D I \mathcal{D}_I D I -measurable, i.e., in particular, f r j , j − f^{-}_{r_j,j} f r j , j − .
Proof of Theorem 4.4. We use the conditional independent property used in Proposition 4.3 to operate with information sets H i , j \mathcal{H}_{i,j} H i , j , i ≥ r j i\geq r_j i ≥ r j , and dropping the upper right trapezoid of D I \mathcal{D}_I D I for the development periods j + 2 , … , J j+2,\ldots,J j + 2 , … , J . For i ≥ I − J + 1 = r J − 1 i\geq I-J+1=r_{J-1} i ≥ I − J + 1 = r J − 1 , we first apply the CL step
C ^ i , J = E [ C i , J ∣ D I ] = E [ E [ C i , J ∣ H i − 1 , J − 1 , D I ] ∣ D I ] = E [ C i , J − 1 f i , J − 1 − ∣ D I ] . \widehat{C}_{i,J}=\mathbb{E}\left[C_{i,J}\mid\mathcal{D}_I\right]=\mathbb{E}\left[\mathbb{E}\left[C_{i,J}\mid\mathcal{H}_{i-1,J-1},\mathcal{D}_I\right]\mid\mathcal{D}_I\right]=\mathbb{E}\left[C_{i,J-1}f^{-}_{i,J-1}\mid\mathcal{D}_I\right]. C i , J = E [ C i , J ∣ D I ] = E [ E [ C i , J ∣ H i − 1 , J − 1 , D I ] ∣ D I ] = E [ C i , J − 1 f i , J − 1 − ∣ D I ] .
In the second step, we compute the D I D_I D I -measurable CL factor
C ^ i , J = E [ C i , J ∣ D I ] = E [ C i , J − 1 f ~ i , J − 1 ∣ D I ] = E [ E [ C i , J − 1 f ~ i , J − 1 ∣ H r J − 1 , J − 1 , D I ] ∣ D I ] = E [ C i , J − 1 E [ f ~ i , J − 1 ∣ H r J − 1 , J − 1 , D I ] ∣ D I ] = f ^ i , J − 1 ( I ) E [ C i , J − 1 ∣ D I ] , \begin{aligned}
\widehat{C}_{i,J} = \mathbb{E}[C_{i,J}\mid D_I] = \mathbb{E}[C_{i,J-1}\tilde{f}_{i,J-1}\mid D_I] \\
&= \mathbb{E}\left[\mathbb{E}[C_{i,J-1}\tilde{f}_{i,J-1}\mid\mathcal{H}_{r_{J-1},J-1},D_I]\mid D_I\right] \\
&= \mathbb{E}\left[C_{i,J-1}\mathbb{E}[\tilde{f}_{i,J-1}\mid\mathcal{H}_{r_{J-1},J-1},D_I]\mid D_I\right]
= \widehat{f}^{(I)}_{i,J-1}\mathbb{E}[C_{i,J-1}\mid D_I],
\end{aligned} C i , J = E [ C i , J ∣ D I ] = E [ C i , J − 1 f ~ i , J − 1 ∣ D I ] = E [ E [ C i , J − 1 f ~ i , J − 1 ∣ H r J − 1 , J − 1 , D I ] ∣ D I ] = E [ C i , J − 1 E [ f ~ i , J − 1 ∣ H r J − 1 , J − 1 , D I ] ∣ D I ] = f i , J − 1 ( I ) E [ C i , J − 1 ∣ D I ] ,
the last step used Proposition 4.3. This is recursively iterated until development period j = d i j=d_i j = d i .
Proof of Theorem 6.1. From (30) and (31) , we have
msep C D R ∣ D I ( 0 ) = Var ( ∑ i = 1 I E [ C i , J ∣ D I + 1 ] ∣ D I ) = Var ( ∑ i = I − J + 1 I ∏ j = 0 J − 1 ( ν i , j + κ i , j C r j , j + 1 ) ∣ D I ) = ∑ i = I − J + 1 I ∑ k = I − J + 1 I Cov ( ∏ j = 0 J − 1 ( ν i , j + κ i , j C r j , j + 1 ) , ∏ j = 0 J − 1 ( ν k , j + κ k , j C r j , j + 1 ) ∣ D I ) . \begin{aligned}
\operatorname{msep}_{\mathrm{CDR}\mid D_I}(0)=\operatorname{Var}\left(\left.\sum_{i=1}^{I}\mathbb{E}[C_{i,J}\mid D_{I+1}]\right|D_I\right) \\
&=\operatorname{Var}\left(\left.\sum_{i=I-J+1}^{I}\prod_{j=0}^{J-1}(\nu_{i,j}+\kappa_{i,j}C_{r_j,j+1})\right|D_I\right) \\
&=\sum_{i=I-J+1}^{I}\sum_{k=I-J+1}^{I}\operatorname{Cov}\left(\left.\prod_{j=0}^{J-1}(\nu_{i,j}+\kappa_{i,j}C_{r_j,j+1}),\prod_{j=0}^{J-1}(\nu_{k,j}+\kappa_{k,j}C_{r_j,j+1})\right|D_I\right).
\end{aligned} msep CDR ∣ D I ( 0 ) = Var ( i = 1 ∑ I E [ C i , J ∣ D I + 1 ] D I ) = Var ( i = I − J + 1 ∑ I j = 0 ∏ J − 1 ( ν i , j + κ i , j C r j , j + 1 ) D I ) = i = I − J + 1 ∑ I k = I − J + 1 ∑ I Cov ( j = 0 ∏ J − 1 ( ν i , j + κ i , j C r j , j + 1 ) , j = 0 ∏ J − 1 ( ν k , j + κ k , j C r j , j + 1 ) D I ) .
Using the standard representation of covariances and the conditionally independent decoupling (32) along the diagonal yields
Cov ( ∏ j = 0 J − 1 ( ν i , j + κ i , j C r j , j + 1 ) , ∏ j = 0 J − 1 ( ν k , j + κ k , j C r j , j + 1 ) ∣ D I ) = E [ ∏ j = 0 J − 1 ( ν i , j + κ i , j C r j , j + 1 ) ( ν k , j + κ k , j C r j , j + 1 ) ∣ D I ] − C ^ i , J C ^ k , J = ∏ j = 0 J − 1 E [ ( ν i , j + κ i , j C r j , j + 1 ) ( ν k , j + κ k , j C r j , j + 1 ) ∣ D I ] − C ^ i , J C ^ k , J = ∏ j = 0 J − 1 [ ν i , j ν k , j + ( ν i , j κ k , j + ν k , j κ i , j ) μ j + κ i , j κ k , j τ j 2 ] − C ^ i , J C ^ k , J . \begin{aligned}
\operatorname{Cov}\left(\left.\prod_{j=0}^{J-1}(\nu_{i,j}+\kappa_{i,j}C_{r_j,j+1}),\prod_{j=0}^{J-1}(\nu_{k,j}+\kappa_{k,j}C_{r_j,j+1})\right|D_I\right) \\
&=\mathbb{E}\left[\left.\prod_{j=0}^{J-1}(\nu_{i,j}+\kappa_{i,j}C_{r_j,j+1})(\nu_{k,j}+\kappa_{k,j}C_{r_j,j+1})\right|D_I\right]-\widehat{C}_{i,J}\widehat{C}_{k,J} \\
&=\prod_{j=0}^{J-1}\mathbb{E}\left[(\nu_{i,j}+\kappa_{i,j}C_{r_j,j+1})(\nu_{k,j}+\kappa_{k,j}C_{r_j,j+1})\mid D_I\right]-\widehat{C}_{i,J}\widehat{C}_{k,J} \\
&=\prod_{j=0}^{J-1}\left[\nu_{i,j}\nu_{k,j}+(\nu_{i,j}\kappa_{k,j}+\nu_{k,j}\kappa_{i,j})\mu_j+\kappa_{i,j}\kappa_{k,j}\tau_j^2\right]-\widehat{C}_{i,J}\widehat{C}_{k,J}.
\end{aligned} Cov ( j = 0 ∏ J − 1 ( ν i , j + κ i , j C r j , j + 1 ) , j = 0 ∏ J − 1 ( ν k , j + κ k , j C r j , j + 1 ) D I ) = E [ j = 0 ∏ J − 1 ( ν i , j + κ i , j C r j , j + 1 ) ( ν k , j + κ k , j C r j , j + 1 ) D I ] − C i , J C k , J = j = 0 ∏ J − 1 E [ ( ν i , j + κ i , j C r j , j + 1 ) ( ν k , j + κ k , j C r j , j + 1 ) ∣ D I ] − C i , J C k , J = j = 0 ∏ J − 1 [ ν i , j ν k , j + ( ν i , j κ k , j + ν k , j κ i , j ) μ j + κ i , j κ k , j τ j 2 ] − C i , J C k , J .
This completes the proof.
Proof of Theorem 6.2. Introduce the normalized quantities
Z j = C r j , j + 1 f ^ j ( I ) C r j , j and H j = 1 + ω j ⋆ ( Z j − 1 ) . Z_j=\frac{C_{r_j,j+1}}{\widehat{f}^{(I)}_j C_{r_j,j}}
\qquad\text{and}\qquad
H_j=1+\omega_j^\star(Z_j-1). Z j = f j ( I ) C r j , j C r j , j + 1 and H j = 1 + ω j ⋆ ( Z j − 1 ) .
These are precisely the variables that contribute to the randomness in (31) . We have conditional means
E [ Z j ∣ D I ] = E [ H j ∣ D I ] = 1 , \mathbb{E}[Z_j\mid D_I]=\mathbb{E}[H_j\mid D_I]=1, E [ Z j ∣ D I ] = E [ H j ∣ D I ] = 1 ,
this follows from (33) . Using (34) we have
Var ( Z j ∣ D I ) = Var ( C r j , j + 1 ∣ D I ) ( f ^ j ( I ) C r j , j ) 2 = C r j , j + σ j 2 A j ( A j − 1 ) C r j , j = Ψ j ω j ⋆ . \operatorname{Var}(Z_j\mid D_I)=\frac{\operatorname{Var}(C_{r_j,j+1}\mid D_I)}{(\widehat{f}^{(I)}_jC_{r_j,j})^2}
=\frac{C_{r_j,j}+\sigma_j^2A_j}{(A_j-1)C_{r_j,j}}
=\frac{\Psi_j}{\omega_j^\star}. Var ( Z j ∣ D I ) = ( f j ( I ) C r j , j ) 2 Var ( C r j , j + 1 ∣ D I ) = ( A j − 1 ) C r j , j C r j , j + σ j 2 A j = ω j ⋆ Ψ j .
This implies
E [ Z j 2 ∣ D I ] = 1 + Ψ j ω j ⋆ , E [ H j 2 ∣ D I ] = 1 + ω j ⋆ Ψ j and E [ Z j H j ∣ D I ] = 1 + Ψ j . (37) \mathbb{E}[Z_j^2\mid D_I]=1+\frac{\Psi_j}{\omega_j^\star},
\qquad
\mathbb{E}[H_j^2\mid D_I]=1+\omega_j^\star\Psi_j
\qquad\text{and}\qquad
\mathbb{E}[Z_jH_j\mid D_I]=1+\Psi_j. \tag*{(37)} E [ Z j 2 ∣ D I ] = 1 + ω j ⋆ Ψ j , E [ H j 2 ∣ D I ] = 1 + ω j ⋆ Ψ j and E [ Z j H j ∣ D I ] = 1 + Ψ j . (37)
Observe that we can write the one-year update of the CL factor as
f ^ j ( I + 1 ) = ω j ⋆ C r j , j + 1 C r j , j + ( 1 − ω j ⋆ ) f ^ j ( I ) = f ^ j ( I ) ( ω j ⋆ Z j + ( 1 − ω j ⋆ ) ) = f ^ j ( I ) H j . \widehat{f}^{(I+1)}_j=\omega_j^\star\frac{C_{r_j,j+1}}{C_{r_j,j}}+(1-\omega_j^\star)\widehat{f}^{(I)}_j
=\widehat{f}^{(I)}_j\left(\omega_j^\star Z_j+(1-\omega_j^\star)\right)
=\widehat{f}^{(I)}_jH_j. f j ( I + 1 ) = ω j ⋆ C r j , j C r j , j + 1 + ( 1 − ω j ⋆ ) f j ( I ) = f j ( I ) ( ω j ⋆ Z j + ( 1 − ω j ⋆ ) ) = f j ( I ) H j .
Thus, we can rewrite (31) as
E [ C i , J ∣ D I + 1 ] = C i , d i + 1 ∏ j = d i + 1 + 1 J − 1 f ^ i , j ( I + 1 ) = C ^ i , J ( Z d i ∏ j = d i + 1 + 1 J − 1 H j ) . \mathbb{E}\left[C_{i,J}\mid D_{I+1}\right] = C_{i,d_{i+1}} \prod_{j=d_{i+1}+1}^{J-1} \widehat{f}^{(I+1)}_{i,j} = \widehat{C}_{i,J}\left(Z_{d_i}\prod_{j=d_{i+1}+1}^{J-1} H_j\right). E [ C i , J ∣ D I + 1 ] = C i , d i + 1 j = d i + 1 + 1 ∏ J − 1 f i , j ( I + 1 ) = C i , J Z d i j = d i + 1 + 1 ∏ J − 1 H j .
The claim for the individual accident years i i i now immediately follows from the conditionally independent decoupling (32) , given D I D_I D I , and from (A.2).
The aggregated version follows completely analogously by considering the conditional covariances for i < m i < m i < m , which gives d i > d m d_i > d_m d i > d m ,
Cov ( C D R i , C D R m ∣ D I ) = C ^ i , J C ^ m , J ( Z d i ∏ j = d i + 1 + 1 J − 1 H j ) ( Z d m ∏ j = d m + 1 + 1 J − 1 H j ) . \operatorname{Cov}\left(\mathrm{CDR}_i,\mathrm{CDR}_m\mid D_I\right) = \widehat{C}_{i,J}\widehat{C}_{m,J}\left(Z_{d_i}\prod_{j=d_{i+1}+1}^{J-1} H_j\right)\left(Z_{d_m}\prod_{j=d_{m+1}+1}^{J-1} H_j\right). Cov ( CDR i , CDR m ∣ D I ) = C i , J C m , J Z d i j = d i + 1 + 1 ∏ J − 1 H j Z d m j = d m + 1 + 1 ∏ J − 1 H j .
with the difference that the term Z d i 2 Z_{d_i}^{2} Z d i 2 needs to be replaced by Z d i H d i Z_{d_i}H_{d_i} Z d i H d i . This completes the proof.