3.1. Notations

We define some notations, similar to those in Pechon et al. (Reference Pechon, Denuit and Trufin2019). Let $\mathcal{H}$ be the set of households, and let us introduce the following claim count variables:

1. — $N_{h(i), t}^{TPL}$ : the number of claims of policyholder i from household h that triggered only TPL during year t;

2. — $N_{h(i), t}^{MD}$ : the number of claims of policyholder i from household h that triggered only MD during year t;

3. — $N_{h(i), t}^{MD:TPL}$ : the number of claims of policyholder i from household h that triggered both TPL and MD simultaneously during year t;

4. — $N_{h, t}^{Home}$ : the number of claims of household h that triggered Home insurance during year t.

Note that to simplify some expressions, sometimes we will write $N_{h(i), t}^{Home}$ when we actually mean $N_{h, t}^{Home}$ (e.g. in the first assumption later in this section). Clearly, $N_{h(i), t}^{Home} = N_{h, t}^{Home}$ for every policyholder i belonging to household h.

As in Pechon et al. (Reference Pechon, Denuit and Trufin2019), this implies that the number of claims for policyholder i from household h that triggered TPL (respectively, MD) during year t is $N_{h(i), t}^{TPL} + N_{h(i), t}^{MD:TPL}$ (respectively, $N_{h(i), t}^{MD} + N_{h(i), t}^{MD:TPL}$ ).

We denote the corresponding a priori expected claim frequencies as $\lambda_{h(i), t}^{TPL} = \text{E}\left[ N_{h(i), t}^{TPL} \right]$ , $\lambda_{h(i), t}^{MD} = \text{E}\left[ N_{h(i), t}^{MD} \right]$ , $\lambda_{h(i), t}^{MD:TPL} = \text{E}\left[ N_{h(i), t}^{MD:TPL} \right]$ and $\lambda_{h, t}^{Home} = \text{E}\left[ N_{h, t}^{Home} \right]$ .

We also introduce the aggregated number of claims over the considered time horizon $t=1, \dots, T$ , namely

1. — $N_{h(i), \bullet}^{TPL} = \sum_{t=1}^{T} N_{h(i), t}^{TPL}$ ;

2. — $N_{h(i), \bullet}^{MD} = \sum_{t=1}^{T} N_{h(i), t}^{MD}$ ;

3. — $N_{h(i), \bullet}^{MD:TPL} = \sum_{t=1}^{T} N_{h(i), t}^{MD:TPL}$ ;

4. — $N_{h, \bullet}^{Home} = \sum_{t=1}^{T} N_{h, t}^{Home}$ .

The corresponding a priori expected claim frequencies are denoted $\lambda_{h(i), \bullet}^{TPL} = \text{E}\left[ N_{h(i), \bullet}^{TPL} \right]$ , $\lambda_{h(i), \bullet}^{MD} = \text{E}\left[ N_{h(i), \bullet}^{MD} \right]$ , $\lambda_{h(i), \bullet}^{MD:TPL} = \text{E}\left[ N_{h(i), \bullet}^{MD:TPL} \right]$ and $\lambda_{h, \bullet}^{Home} = \text{E}\left[ N_{h, \bullet}^{Home} \right]$ . In our study, $T = 3$ . This means that the policyholders have been observed for at most 3 years. In fact, a variable exposure expresses the amount of time the policyholders were covered during the T years.

Let $\mathcal{G}_{Motor} = \{TPL, MD, MD\,{:}\,TPL\}$ be the set of types of claims in Motor insurance and $\mathcal{G} = \mathcal{G}_{Motor} \cup \{Home\}$ be the set of all claim types considered in this paper.

3.2. Correlation structure

We introduce a Poisson multivariate mixture to allow for dependence between the latent effects which represents all the unobserved risk factors affecting the claim frequency. Let us make the following assumptions:

1. 1. $\forall h \in \mathcal{H}$ , $i = 1, 2 $ , $\forall g \in \mathcal{G}$ , given $\Theta_{h(i)}^{g} = \theta$ , the random variables $N_{h(i), 1}^{g}, \dots, N_{h(i), T}^{g}$ are independent.

2. 2. $\forall h \in \mathcal{H}$ , $i,j = 1, 2 $ , $\forall g_i, g_j \in \mathcal{G}_{Motor}$ such that $i\ne j $ or $g_i \ne g_j$ , given $(\Theta_{h(i)}^{g_i}, \Theta_{h(j)}^{g_j}) = (\theta_i, \theta_j)$ , the sequences of random variables $N_{h(i), 1}^{g_i}, N_{h(i), 2}^{g_i}, \dots, N_{h(i), T}^{g_i}$ and $N_{h(j), 1}^{g_j}, N_{h(j), 2}^{g_j}, \dots, N_{h(j), T}^{g_j}$ are independent.

3. 3. $\forall h \in \mathcal{H}$ , $i = 1,2$ , $\forall g_i \in \mathcal{G}_{Motor}$ , given $(\Theta_{h(i)}^{g_i}, \Theta_{h}^{Home}) = (\theta_i, \theta)$ , the sequences of random variables $N_{h(i), 1}^{g_i}, N_{h(i), 2}^{g_i}, \dots, N_{h(i), T}^{g_i}$ and $N_{h, 1}^{Home}, N_{h, 2}^{Home}, \dots, N_{h, T}^{Home}$ are independent.

4. 4. For a household with two Motor insurance policies and one Home insurance policy, let $\boldsymbol\Theta_{\textbf{\textit{h}}} = (\Theta_h^{Home}, \Theta_{h(1)}^{TPL}, \Theta_{h(1)}^{MD}, \Theta_{h(1)}^{MD:TPL}, \Theta_{h(2)}^{TPL}, \Theta_{h(2)}^{MD}, \Theta_{h(2)}^{MD:TPL})$ be the vector of all random effects. We assume that $\text{E}({\boldsymbol\Theta}_h) = \textbf{1}$ . Furthermore, we assume that the random effects can be decomposed using the hierarchical structure described in Figure 2 (with nested random effects):

   1. — $\forall h \in \mathcal{H}, \Theta_h^{Home} = \exp\left(\mathcal{E}^{Household}_h + \mathcal{E}_h^{Home} \right)$ ;

   2. — $\forall h \in \mathcal{H},\forall i\in h, \forall g \in \mathcal{G}_{Motor}, \Theta_{h(i)}^{g} = \exp\left(\mathcal{E}^{Household}_h + \mathcal{E}_h^{Motor} + \mathcal{E}_{h(i)}^{PolId} + \mathcal{E}_{h(i)}^{g} \right)$ .

   We will assume that $\mathcal{E}_h^{Household} \sim N\Big({-}\frac{\varsigma_{Household}^2}{2}, \varsigma_{Household}^2\Big)$ , $\mathcal{E}_h^{Home} \sim N\Big({-}\frac{\varsigma_{Home}^2}{2}, \varsigma_{Home}^2\Big)$ , $\mathcal{E}_h^{Motor} \sim N\Big({-}\frac{\varsigma_{Motor}^2}{2}, \varsigma_{Motor}^2\Big)$ , $\mathcal{E}_{h(i)}^{PolId} \sim N\Big({-}\frac{\varsigma_{PolId}^2}{2}, \varsigma_{PolId}^2\Big)$ , $\forall g \in \mathcal{G}_{Motor}\,\,\, \mathcal{E}_{h(i)}^{g} \sim N\Big({-}\frac{\varsigma_{g}^2}{2}, \varsigma_{g}^2\Big)$ . The random vector $\boldsymbol{\mathcal{E}}_{\textbf{\textit{h}}} = (\mathcal{E}^{Household}_h, \mathcal{E}_h^{Home}, \mathcal{E}_h^{Motor}, \mathcal{E}_{h(i)}^{PolId}, \mathcal{E}_{h(i)}^{TPL}, \mathcal{E}_{h(i)}^{MD}, \mathcal{E}_{h(i)}^{MD:TPL})$ is such that $\text{E}[\exp \boldsymbol{\mathcal{E}}_{\textbf{\textit{h}}}] = \textbf{1}$ . The variance–covariance matrix of $\boldsymbol{\mathcal{E}}_{\textbf{\textit{h}}}$ is given by

   \begin{equation*}\text{V}\left[ \boldsymbol{\mathcal{E}}_{\textbf{\textit{h}}} \right] = diag(\varsigma_{Household}^2, \varsigma_{Home}^2, \varsigma_{Motor}^2, \varsigma_{PolId}^2, \varsigma_{TPL}^2, \varsigma_{MD}^2, \varsigma_{MD:TPL}^2)\end{equation*}
   The diagonal structure of the matrix $\text{V}\left[ \boldsymbol{\mathcal{E}}_{\textbf{\textit{h}}} \right]$ means that the additive random effects are independent.

Considering Figure 2, $\mathcal{E}_{Household}$ can be understood as the Household effect. Indeed, all the shared risk factors across the policyholders from the household and across both products are included in this random effect. Thanks to this random effect, both products are correlated. In Motor insurance, some effects may be common to policyholders from the same household, in addition to the previously cited Household effect. Indeed, the random effect $\mathcal{E}_{Motor}$ induces additional covariance between the claim frequencies in Motor insurance. Finally, some policyholder-related latent factors may exist and influence the claim frequencies; these effects are modelled thanks to $\mathcal{E}_{PolId}$ .

Note that the assumptions above that relate to Motor insurance are similar to those made in Pechon et al. (Reference Pechon, Denuit and Trufin2019). The assumption of normality means that the vector of random effects $\boldsymbol\Theta_{\textbf{\textit{h}}}$ has a multivariate LogNormal distribution with unit mean. In Pechon et al. (Reference Pechon, Trufin and Denuit2018), a Poisson mixture was used to introduce dependence between policyholders from the same household in TPL insurance. The authors compared a Poisson–Gamma, where the Gamma-distributed random effects were correlated using a Gaussian copula, with a Poisson–LogNormal model and concluded that the Poisson–LogNormal model outperformed the Poisson–Gamma model. The model involves a hierarchical structure of random effects representing the latent effects. The variance–covariance matrix of $\log \boldsymbol\Theta_{\textbf{\textit{h}}}$ (where the log is taken on each component) is given by

\begin{equation*}\boldsymbol{\Sigma}_{\log \boldsymbol\Theta_{\textbf{\textit{h}}}} = \left(\begin{array}{c | c@{\quad}c@{\quad}c | c@{\quad}c@{\quad}c}\sigma_{Home}^2 & \sigma_{HM} & \sigma_{HM} & \sigma_{HM} & \sigma_{HM} & \sigma_{HM} & \sigma_{HM} \\ \hline\sigma_{HM} & \sigma_{TPL}^2 & \sigma_{PolId} & \sigma_{PolId} & & & \\\sigma_{HM} & \sigma_{PolId} & \sigma_{MD}^2 & \sigma_{PolId} & & \sigma_{Motor} & \\\sigma_{HM} & \sigma_{PolId} & \sigma_{PolId} & \sigma_{MD:TPL}^2 & & & \\ \hline\sigma_{HM} & & & & \sigma_{TPL}^2 & \sigma_{PolId} & \sigma_{PolId} \\\sigma_{HM} & & \sigma_{Motor} & & \sigma_{PolId} & \sigma_{MD}^2 & \sigma_{PolId} \\\sigma_{HM} & & & & \sigma_{PolId} & \sigma_{PolId} & \sigma_{MD:TPL}^2\end{array}\right) \end{equation*}

where

\begin{align*}\sigma_{Home}^2 & = \varsigma_{Household}^2 + \varsigma_{Home}^2 & \\\sigma_{TPL}^2 & = \varsigma_{Household}^2 + \varsigma_{Motor}^2 + \varsigma_{PolId}^2 + \varsigma_{TPL}^2 & \displaybreak\\\sigma_{MD}^2 & = \varsigma_{Household}^2 + \varsigma_{Motor}^2 + \varsigma_{PolId}^2 + \varsigma_{MD}^2 & \\\sigma_{MD:TPL}^2 & = \varsigma_{Household}^2 + \varsigma_{Motor}^2 + \varsigma_{PolId}^2 + \varsigma_{MD:TPL}^2 & \\\sigma_{HM} & = \varsigma_{Household}^2 & \text{(Household effect)} \\\sigma_{Motor} & = \varsigma_{Household}^2 + \varsigma_{Motor}^2 & \text{(Inter-policyholder)} \\\sigma_{PolId} & =\varsigma_{Household}^2 + \varsigma_{Motor}^2 + \varsigma_{PolId}^2 & \text{(Intra-policyholder)}\end{align*}

3.3. Estimation

Since the support of each random effect is $\mathbb{R}_+$ , we can write the likelihood as

(1) \begin{align} L(\boldsymbol{\Sigma_{\log\Theta}}) &= \prod_{h \in \mathcal{H}} \int_{\mathbb{R}_+^7} \left[ \prod_{i = 1}^2 \prod_{g \in \mathcal{G}} \exp\left({-}\lambda_{h(i), \bullet}^{g}\theta_{h(i)}^{g} \right) \frac{\left(\lambda_{h(i), \bullet}^{g}\theta_{h(i)}^{g}\right)^{n_{h(i), \bullet}^{g}}}{n_{h(i), \bullet}^{g}!} \right] f_{\boldsymbol\Theta_{\textbf{\textit{h}}}} (\boldsymbol\theta_{\textbf{\textit{h}}}) d\boldsymbol\theta_{\textbf{\textit{h}}} \end{align}

Note that if for some household the length of the vector $\boldsymbol\Theta_{\textbf{\textit{h}}}$ is smaller than 7 (i.e. no Home insurance policy or less than two Motor insurance policies), by adopting the convention that $0^0 = 1$ and by setting the corresponding a priori expected claim frequency $\lambda$ to zero yield the correct likelihood. In such a case, the corresponding integral vanishes.

The computations have been realised in the statistical software R. The estimation was done in a two-part approach.

First, the a priori expected claim frequencies have been estimated with a Poisson regression using GAMs. It has to be noted that no covariates from Motor insurance have been used in the Home insurance regression and vice versa. One reason for this is that some households have only Motor (respectively, Home) insurance, and the covariates related to the missing product would therefore result in missing values, which the GAMs do not handle. This does not prevent us from using common factors, such as the length of time the policyholder or household has been with the company. Such a covariate can be a part of each model as long as the parameters are functionally independent. Four different claims are considered: one in Home insurance and three in Motor insurance. As explained in Section 2, a pool of multiple guarantees in Home insurance is considered, whereas in Motor insurance, three types of claims are considered. The GAMs were chosen to model the claim frequencies using a Poisson regression with a $\log$ link function. See Wood (Reference Wood2017) for an introduction to GAMs. In the present study, the mgcv package available in R was used.

The models used in Motor insurance are the same as in Pechon et al. (Reference Pechon, Denuit and Trufin2019). More details about the a priori analyses can be found in Section 6.

In the second part, the a priori expected claim frequencies are fixed, and the variance–covariance matrix of the random effects is estimated by maximising the objective function (1) by holding the a priori expected claim frequencies $\lambda$ fixed.

Although this two-part approach will be used in this paper, we provide an algorithm in Section6 that resembles an expectation/conditional maximisation (ECM) procedure. The algorithm proposed in this paper differs from the one proposed in Meng & Rubin (Reference Meng and Rubin1993) with the fact that we maximise two different objective functions. We stress out that further research is necessary, however, to assess the validity of the proposed algorithm.

Table 3.

Estimates at each step of the optimisation process.

In order to compute the likelihood in equation (1), we need to rely on numerical integration. We rely on the multivariate Gauss Hermite quadrature, which is available in R thanks to the package MultiGHQuad contributed by Kroeze (Reference Kroeze2016). The integrand is written in C++ so as to fasten the computation. The C++ integrand can be included in an R function using the Rcpp package (see Eddelbuettel & François Reference Eddelbuettel and François2011). We have a step-by-step approach to estimate the parameters, which we shall describe below. The step-by-step approach allows to break down the estimation in different steps of smaller dimension (i.e. which require less computations) which in turn give us good initial values for the final step that incorporates all the parameters. Having good initial values for this final step allows to have less iterations in the optimisation process and consequently reduces drastically the computation time.

First, we estimate the parameters related to Motor insurance at a policyholder level (the variances $\varsigma^2_{PolId}, \varsigma^2_{TPL}, \varsigma^2_{MD}, \varsigma^2_{MD:TPL}$ ), the other variances being fixed temporary at zero. As initial values, we rely on the marginal estimates of the variance. To obtain initial values for the parameters $ \varsigma^2_{TPL}, \varsigma^2_{MD}, \varsigma^2_{MD:TPL}$ , we can estimate the three variances of three different univariate Poisson–LogNormal models. A initial value for $\varsigma^2_{PolId}$ can be given by optimising a trivariate Poisson–LogNormal model with the three previous variances fixed and the covariance (i.e. $\varsigma^2_{PolId}$ ) being estimated. Since these models are of low dimension and each of them involves only one parameter to optimise, convergence is fast.

In the second step, the variance $\varsigma^2_{Motor}$ is estimated. In fact, the parameters $\sigma^2_{TPL}, \sigma^2_{MD}, \sigma^2_{MD:TPL}$ are kept constant in this step, and an increase of the parameter $\varsigma^2_{Motor}$ (inducing the dependence between random effects from the same household) is compensated by a decrease of the parameter $\varsigma^2_{PolId}$ (inducing dependence between random effects from the same policyholder). Note that the difference at this point with Pechon et al. (Reference Pechon, Denuit and Trufin2019) comes from the dependence structure. The covariance (induced by $\Theta_{Motor}$ ) is estimated and is assumed to be the same for any pair of different random effects belonging to the same policyholder, whereas in Pechon et al. (Reference Pechon, Denuit and Trufin2019), each pair of random effects related to the same policyholder had its own covariance parameter.

In the third step, Home insurance is then included. However, the hierarchical structure imposes that the correlations (respectively, covariances) are positive. As opposed to Motor insurance, for which we could rely on the results found in Pechon et al. (Reference Pechon, Denuit and Trufin2019) establishing that the correlation is positive, we first need to assess whether the dependence between Home and Motor insurance is positive before using the hierarchical structure. For this preliminary assessment, we first introduce a single random effect $\Theta^{Home}$ and introduce dependence with respect to Motor insurance thanks to a unique covariance parameter between the random variable $\Theta^{Home}$ and the random variables $\Theta^{TPL}, \Theta^{MD}, \Theta^{MD:TPL}$ . The variance (i.e. the residual heterogeneity in Home insurance) is estimated marginally and then is fixed at its value to estimate the covariance. We find a covariance equals to 0.1202 (s.e. 0.0174).

Given that the covariance between Home and Motor insurance’s random effects is significantly positive, we can rely on the hierarchical structure which imposes a positive dependence and estimate all the parameters simultaneously in a third step by maximising the objective function (3.3) holding the a priori expected claim frequencies $\lambda$ fixed.

We display all the estimates at each step in Table 3. We note that at each step there is a transfer that occurs: the estimated $\widehat{\varsigma}_{PolId}^2$ from step 1 is split into $\widehat{\varsigma}_{PolId}^2$ and $\widehat{\varsigma}_{Motor}^2$ in step 2. This breakdown can be understood in the following way. The introduction of the variance $\widehat{\varsigma}_{Motor}^2$ induces covariance between policyholders from the same household in Motor insurance, while in step 1, only the random effects of guarantees related to the same policyholder had a strictly covariance. The increase in $\widehat{\varsigma}_{Motor}^2$ is, however, compensated by the decrease in $\widehat{\varsigma}_{PolId}^2$ . The covariance for guarantees related to the same policyholder does not change, while the policyholders from the same household in Motor insurance become correlated. Similarly, from step 2 to step 3, the introduction of the random effect related to the Home induces a decrease of $\widehat{\varsigma}_{Motor}^2$ that is compensated by an increase of $\widehat{\varsigma}_{Household}^2$ . The covariances in Motor insurance remain the same than in step 2, however, due to the common variance $\widehat{\varsigma}_{Household}^2$ , Home and Motor insurance become correlated, with a covariance equal to $\widehat{\varsigma}_{Household}^2$ .

Finally, we observe that the policyholder effect (i.e. $\varsigma^2_{PolId}$ ) does not seem to be significantly different from zero. This appears to be coherent with the results found in Pechon et al. (Reference Pechon, Denuit and Trufin2019), where the estimates of the analogue parameters appear to have overlapping confidence intervals. This model is then re-estimated by forcing $\varsigma^2_{PolId} = 0$ . The estimates can be found in Table 4.

Table 4.

Final estimates.

A likelihood ratio test is conducted to assess whether the policyholder effect is significant (i.e. $\varsigma^2_{PolId}>0$ ). The value of the statistic is $t = 0.1866$ . Since the test involves a value on the boundary of the domain, Self & Liang (Reference Self and Liang1987) and Agresti (Reference Agresti2003) suggest to compute the p-value using $\frac{1}{2} \Pr\left(\chi^2_1 > t\right)$ . We obtain a p-value of $0.3329$ , which confirms that the policyholder effect is not significant. Notice that because of the two-step estimation procedure, this p-value must be considered with caution, but its relatively high value is taken as an evidence supporting the model choice ${\varsigma}_{PolId}^2 = 0$ retained here.

The final model used in the remainder of this paper, without the policyholder effect, amounts to a hierarchical structure as depicted in Figure 3.

Figure 3

Hierarchical structure of final model.

Note that given the parametrisation using a hierarchical structure, any value within the confidence intervals will yield an appropriate (i.e. positive-definite) variance–covariance matrix. This is one of the advantages of the hierarchical parametrisation.

Some sensitivity analysis with respect to the number of nodes per dimension used in the Gauss–Hermite quadrature has been conducted. The results thereof can be found in the Supplementary Material in this paper. It appears that starting at seven nodes per dimension, the estimates stabilise. By choosing $m=7$ nodes per dimension, for the households with two Motor insurance policies and one Home insurance policy, the number of nodes used to compute the contribution of this household to the likelihood amounts to $m^7=7^7 =$ 823,543. It has to be noted, though, that not all households have these three policies. The contribution to the likelihood of these households therefore implies the evaluation of less integrals.

3.4. Implied dependence structure

Now that the model has been estimated, we can compute the implied dependence structure between these latent factors. We start with the variances, which represent the strength of these latent factors (i.e. the amount of residual heterogeneity in each risk class constructed with the a priori model).

We start with the implied dependence structure of the underlying multivariate normal distribution. Afterwards, the implied dependence structure on the LogNormal scale (i.e. the dependence structure of $\boldsymbol\Theta_{\textbf{\textit{h}}}$ ) will be deduced.

For $i = 1,2$ , we can compute the residual heterogeneity:

\begin{align*} \text{V}\left[\log \Theta^{Home}_h \right] & = \varsigma_{Household}^2 + \varsigma_{Home}^2 & \text{estimated to } & 0.3110\,\, \\ \text{V}\left[\log \Theta_{h(i)}^{TPL}\right] & = \varsigma_{Household}^2 + \varsigma_{Motor}^2 + \varsigma_{TPL}^2 & \text{estimated to } & 0.5592\,\, \\ \text{V}\left[\log \Theta_{h(i)}^{MD}\right] & = \varsigma_{Household}^2 + \varsigma_{Motor}^2 + \varsigma_{MD}^2 & \text{estimated to } & 0.3651\,\, \\ \text{V}\left[\log \Theta_{h(i)}^{MD:TPL}\right] & = \varsigma_{Household}^2 + \varsigma_{Motor}^2 + \varsigma_{MD:TPL}^2 & \text{estimated to } & 0.3715\end{align*}

As stated above, the hierarchical structure induces positive covariances between the random effects. We can now explicitly compute these. Indeed, we have that for $i,j= 1, 2$

\begin{align*} \mathbb{C}\mbox{ov}\left[\log \Theta_{h}^{Home}, \log \Theta_{h(i)}^{TPL} \right] &= \mathbb{C}\mbox{ov}\left[\log \Theta_{h}^{Home}, \log \Theta_{h(i)}^{MD} \right]\\ &= \mathbb{C}\mbox{ov}\left[\log \Theta_{h}^{Home}, \log \Theta_{h(i)}^{MD:TPL} \right]\\ &= \varsigma_{Household}^2 \text{ estimated to } 0.1238\\ \mathbb{C}\mbox{ov}\left[\log \Theta_{h(i)}^{TPL}, \log \Theta_{h(j)}^{MD} \right] &=\mathbb{C}\mbox{ov}\left[\log \Theta_{h(i)}^{TPL}, \log \Theta_{h(j)}^{MD:TPL} \right]\\ &=\mathbb{C}\mbox{ov}\left[\log \Theta_{h(i)}^{MD}, \log \Theta_{h(j)}^{MD:TPL} \right] \\ &= \varsigma_{Household}^2 + \varsigma_{Motor}^2 \text{ estimated to } 0.2379\\ \text{and for } i\ne j, \mathbb{C}\mbox{ov}\left[\log \Theta_{h(i)}^{TPL}, \log \Theta_{h(j)}^{TPL} \right]&=\mathbb{C}\mbox{ov}\left[\log \Theta_{h(i)}^{MD}, \log \Theta_{h(j)}^{MD} \right] \\ &= \mathbb{C}\mbox{ov}\left[\log \Theta_{h(i)}^{MD:TPL}, \log \Theta_{h(j)}^{MD:TPL} \right]\\ &= \varsigma_{Household}^2 + \varsigma_{Motor}^2 \text{ estimated to } 0.2379\end{align*}

We can compute the correlations between the $\log$ of the random effects. The estimated correlations are shown in Table 5. These correlations are related to the underlying normal distribution, which is on the score scale (i.e. the $\log$ scale).

Table 5.

Estimated correlation structure between the random effects on the log scale (i.e. at the score scale).

Now that the dependence structure of the multivariate normal random vector is computed, we can deduce the correlation structure between the LogNormal random effects. Let us compute the variances as well as the correlations between the LogNormal distributed random effects

\begin{equation*} \boldsymbol\Theta_{\textbf{\textit{h}}} = \left(\Theta_h^{Home}, \Theta_{h(1)}^{TPL}, \Theta_{h(1)}^{MD}, \Theta_{h(1)}^{MD:TPL}, \Theta_{h(2)}^{TPL}, \Theta_{h(2)}^{MD}, \Theta_{h(2)}^{MD:TPL}\right)\end{equation*}

We show the estimated variances in Table 6 and the estimated correlations in Table 7.

Table 6.

Estimated variances of random effects for $i=1,2$ .

Table 7.

Estimated correlation structure between the LogNormal random effects.

Let us provide some interpretations for the values displayed in Tables 6 and 7.

The heterogeneity appears to be the smallest in Home insurance. This is very intuitive because claims in Motor insurance are strongly related to the way of driving as well as the risk aversion of the policyholder, whereas claims in Home insurance can be more often the result of a external circumstances, such as bad location (e.g. flood zone), or to a badly constructed building (e.g. faulting electrical connections).

The dependence between both Home and Motor insurance materialises in a correlation of about 30%. It appears, however, to be weaker than within guarantees related to Motor insurance. Again, this is intuitive, as less choices are made by the policyholder in Home insurance than in Motor insurance. The correlations in Motor insurance appear to be of similar order of magnitude than those given in Pechon et al. (Reference Pechon, Denuit and Trufin2019), although a different pair appears to have the strongest correlation. By looking at the standard errors of the estimates, we can assume that this could be related to randomness.

Consequently, there appears to be some unobserved risk factor that is affecting the whole household and that is correlating the claim frequencies in both Motor insurance guarantees as well as in Home insurance. This could be a very local geographic effect that the a priori model could not identify using a bivariate function of the latitude and longitude of the place of residence, or even the socio-economic status, as the latitude and longitude variables available in the data set are in fact the centre of each district. Therefore, socio-economic status, often shared in neighbourhoods, could not be integrated in the a priori model and could be part of the unobserved risk factors. Nevertheless, one can only take guesses, since this information is not available and consequently cannot be added to the a priori model.

The fact, however, that the correlation between Motor insurance guarantees is the highest suggests that some other unobserved factor representing the behaviour and risk averseness of the policyholder is hidden in this residual heterogeneity and is also possibly shared among policyholders from the same household.

4.1. A posteriori corrections

As time passes by and the insurer observes the number of claims, we can compute corrections to apply to the a priori expected claim frequency conditional to the observed number of claims thanks to credibility theory. Due to the dependence between the random effects related to the same household, the number of claims of any policyholder from the household as well as of any guarantee is relevant.

Let a household h consists of two policyholders, h(1) and h(2), and let them both have TPL and MD insurance. Moreover, the household also has a Home insurance policy at the same insurance company. We will assume for simplicity that all the policies start at the same date. At any time, the insurer computes for each guarantee the cumulated (i.e. aggregated) number of claims since the contracts inception as well as the cumulated (i.e. aggregated) a priori expected claim frequencies. For instance, at the end of year T, the insurer can then compute the following corrections for the household in Home insurance:

\begin{align*}&\text{E}\left[\Theta_h^{Home} | \textbf{\textit{N}}_{\textbf{\textit{h}}} = \textbf{\textit{n}}_{\textbf{\textit{h}}}\right]\\&\quad= \frac{1}{\Pr\left[\textbf{\textit{N}}_{\textbf{\textit{h}}} = \textbf{\textit{n}}_{\textbf{\textit{h}}} \right]}\int_{\mathbb{R}_+^7} \theta^{Home} \Pr\left[\textbf{\textit{N}}_{\textbf{\textit{h}}} = \textbf{\textit{n}}_{\textbf{\textit{h}}} |{\boldsymbol\Theta} =\boldsymbol{\theta} \right] f_{{\boldsymbol\Theta}}({\boldsymbol\Theta}) d{\boldsymbol\Theta}\\&\quad= \frac{1}{\Pr\left[\textbf{\textit{N}}_{\textbf{\textit{h}}} = \textbf{\textit{n}}_{\textbf{\textit{h}}} \right]}\frac{1+n^{Home}_{h,\bullet}}{\lambda^{Home}_{h,\bullet}}\int_{\mathbb{R}^7_+} \Pr\left[\textbf{\textit{N}}_{\textbf{\textit{h}}}^{({-}Home)} = \textbf{\textit{n}}_{\textbf{\textit{h}}}^{({-}Home)},\right.\\ &\qquad\left. N_{h,\bullet}^{Home} = n_{h,\bullet}^{Home}+1 |{\boldsymbol\Theta} =\boldsymbol{\theta} \right] f_{{\boldsymbol\Theta}}({\boldsymbol\Theta}) d{\boldsymbol\Theta}\\&\quad=\frac{\Pr\left[\textbf{\textit{N}}_{\textbf{\textit{h}}}^{({-}Home)} = \textbf{\textit{n}}_{\textbf{\textit{h}}}^{({-}Home)}, N_{h,\bullet}^{Home} = n_{h,\bullet}^{Home}+1 \right]}{\Pr\left[\textbf{\textit{N}}_{\textbf{\textit{h}}} = \textbf{\textit{n}}_{\textbf{\textit{h}}} \right]}\frac{1+n^{Home}_{h,\bullet}}{\lambda^{Home}_{h,\bullet}}\end{align*}

where

\begin{align*}\textbf{\textit{N}}_{\textbf{\textit{h}}} &= \left(N_{h,\bullet}^{Home}, N_{h(1),\bullet}^{TPL}, N_{h(1),\bullet}^{MD}, N_{h(1),\bullet}^{MD:TPL},N_{h(2),\bullet}^{TPL}, N_{h(2),\bullet}^{MD}, N_{h(2),\bullet}^{MD:TPL}\right)\\\textbf{\textit{n}}_{\textbf{\textit{h}}} &= \left(n_{h,\bullet}^{Home}, n_{h(1),\bullet}^{TPL}, n_{h(1),\bullet}^{MD}, n_{h(1),\bullet}^{MD:TPL},n_{h(2),\bullet}^{TPL}, n_{h(2),\bullet}^{MD}, n_{h(2),\bullet}^{MD:TPL}\right)\end{align*}

and where $\textbf{\textit{N}}_{\textbf{\textit{h}}}^{({-}Home)}$ (respectively, $\textbf{\textit{n}}_{\textbf{\textit{h}}}^{({-}Home)}$ ) is the vector $\textbf{\textit{N}}_{\textbf{\textit{h}}}$ (respectively, $\textbf{\textit{n}}_{\textbf{\textit{h}}}$ ) without the item related to the Home guarantee.

In Motor insurance, we can compute the conditional expectation of any random effect:

\begin{align*} \forall g\in\mathcal{G}_{Motor},\,\, &\text{E}\left[\Theta_{h(i)}^{g} | \textbf{\textit{N}}_{\textbf{\textit{h}}} = \textbf{\textit{n}}_{\textbf{\textit{h}}}\right] = \frac{\Pr\left[\textbf{\textit{N}}_{\textbf{\textit{h}(\textit{i})}}^{({-}g)} = \textbf{\textit{n}}_{\textbf{\textit{h}(\textit{i})}}^{({-}g)}, N_{h(i),\bullet}^{g} = n_{h(i),\bullet}^{g}+1 \right]}{\Pr\left[\textbf{\textit{N}}_{\textbf{\textit{h}}} = \textbf{\textit{n}}_{\textbf{\textit{h}}} \right]}\frac{1+n^{g}_{h(i),\bullet}}{\lambda^{g}_{h(i),\bullet}}\end{align*}

where

\begin{equation*}\Pr\left[\textbf{\textit{N}}_{\textbf{\textit{h}}} = \textbf{\textit{n}}_{\textbf{\textit{h}}} \right] = \displaystyle\int_{\mathbb{R}^7_+} \prod_{i \in h} \prod_{g \in \mathcal{G}} \exp\left({-}\lambda_{h(i), \bullet}^{g}\theta_{h(i)}^{g}\right)\frac{\left(\lambda_{h(i), \bullet}^{g}\theta_{h(i)}^{g}\right)^{n_{h(i), \bullet}^{g}}}{n_{h(i), \bullet}^{g}!}f_{\boldsymbol\Theta_{\textbf{\textit{h}}}} (\boldsymbol\theta_{\textbf{\textit{h}}}) d\boldsymbol\theta_{\textbf{\textit{h}}}\end{equation*}

can be computed numerically, for instance with the Gauss–Hermite quadrature, and where $\textbf{\textit{N}}_{\textbf{\textit{h}(\textit{i})}}^{({-}g)}$ (respectively, $\textbf{\textit{n}}_{\textbf{\textit{h}(\textit{i})}}^{({-}g)}$ ) is the vector $\textbf{\textit{N}}_{\textbf{\textit{h}}}$ (respectively, $\textbf{\textit{n}}_{\textbf{\textit{h}}}$ ) without the item related to the guarantee g of policyholder h(i).

Although the modelling in Motor insurance includes three count variables and three random effects to model two guarantees (in order to capture the events that trigger both guarantees at the same time), we can in fact calculate the correction that is applied in TPL and in MD by considering the following linear combinations:

(2) \begin{equation}\begin{aligned}&\frac{1}{\lambda_{h(i), \bullet}^{TPL} +\lambda_{h(i), \bullet}^{MD:TPL}} \left( \lambda_{h(i), \bullet}^{TPL} \text{E}\left[\Theta_{h(i)}^{TPL} | \textbf{\textit{N}}_{\textbf{\textit{h}}} = \textbf{\textit{n}}_{\textbf{\textit{h}}}\right] + \lambda_{h(i), \bullet}^{MD:TPL} \text{E}\left[\Theta_{h(i)}^{MD:TPL} | \textbf{\textit{N}}_{\textbf{\textit{h}}} = \textbf{\textit{n}}_{\textbf{\textit{h}}}\right] \right)\\&\frac{1}{\lambda_{h(i), \bullet}^{MD} +\lambda_{h(i), \bullet}^{MD:TPL}} \left( \lambda_{h(i), \bullet}^{MD} \text{E}\left[\Theta_{h(i)}^{MD} | \textbf{\textit{N}}_{\textbf{\textit{h}}} = \textbf{\textit{n}}_{\textbf{\textit{h}}}\right] + \lambda_{h(i), \bullet}^{MD:TPL} \text{E}\left[\Theta_{h(i)}^{MD:TPL} | \textbf{\textit{N}}_{\textbf{\textit{h}}} = \textbf{\textit{n}}_{\textbf{\textit{h}}}\right] \right)\end{aligned}\end{equation}

In order to numerically compute these corrections, we need to determine the a priori risk profiles of both policyholders in TPL and in MD as well as the household’s risk profile in Home insurance. For this matter, we bin the predicted a priori expected claim frequencies obtained with the GAMs into three categories, using the quantiles $2/6$ and $4/6$ : low-, medium- and high-risk profiles. The three risk profiles are then associated with the numerical values given by the quantiles $1/6$ , $3/6$ and $5/6$ (i.e. the corresponding median value of estimated expected claim frequency of each risk class). In the following numerical examples, we will only consider the cases in which the same a priori risk profile is shared across all the policyholders and all the guarantees. Of course, other combinations are possible.

4.1.1. A posteriori corrections in the claim-free case

Let us first illustrate the example in which no claim was reported in any guarantee. We show in Figure 4 the corrections to apply in Home and Motor insurance when no claim occurred as time passes. As expected, the correction factor drops as time passes and no claim is reported. The riskier profiles have a stronger correction, in line with the idea that claim-free years are more expected from a safer policyholder than from a riskier policyholder. Note that the corrections are similar in TPL and in MD, although they appear to be slightly stronger in MD. This comes from the fact that the average claim frequency is higher in MD than in TPL, although this effect is, in these examples, partly compensated by the fact that the heterogeneity is greater in TPL than in MD. Also in Home insurance, a correction factor below 90% is observed after 5 claim-free years, meaning that in average, claim-free households (in Motor and Home insurance) with a low-risk profile have about 10% less claims in Home insurance than the claim frequency given by their a priori risk class.

Figure 4

Corrections to apply in TPL (left), MD (middle) and Home insurance (right) when no claim occurred in the household.

4.1.2. A posteriori corrections in Motor insurance after a claim

In the following, we will first consider that a claim occurred (either in Home or in Motor insurance) and discuss the corrections to apply to the guarantees in Motor insurance. The corrections to apply in Home insurance will be considered later. Note that, even though two policyholders are considered in the household, we can consider in these examples without loss of generalisation that h(1) experiences the claims in Motor insurance and h(2) remains claim-free. Also, note that for Home insurance, it is neither h(1) nor h(2) that experiences the claim, but rather the whole household h.

In Figure 5, the corrections to apply in TPL to both policyholders are depicted. The figures show that a claim in any guarantee will increase the correction factors in all the guarantees and for all the policyholders from the household. We see, however, that the biggest increase is for h(1) in TPL (i.e. the policyholder at fault in the triggered guarantee). We also note that both h(1) and h(2) have the same corrections after a claim of h(1) in MD only. This comes from the fact that the model assumes the same covariance between the different random effects in Motor insurance (regardless if they are related to the same policyholder from the household). This is not the case for the corrections after a claim triggering both TPL and MD, as the correction in TPL explicitly combines both kind of claims: those triggering only TPL and those triggering TPL and MD (see equation (2)).

Figure 5

Correction to apply to TPL insurance a priori expected claim frequencies when one claim was reported.

Similarly, in Figure 6, the corrections to apply in MD to both policyholders are shown. By comparing the corrections in TPL after a claim in TPL (Figure 5) and the corrections in MD after a claim in MD (Figure 6), it appears that we have a stronger correction factor in TPL when a claim is reported in TPL than in MD when a claim in MD is reported. This comes from the greater heterogeneity in TPL (see Table 6) and the lower claim frequencies in TPL which means that a claim is less expected in TPL than in MD. Moreover, the decrease of these correction factors is greater in MD than in TPL, in line with the greater claim frequencies in MD than in TPL.

Figure 6

Correction to apply to MD insurance a priori expected claim frequencies when one claim was reported.

Furthermore, it appears that a claim triggering both guarantees at the same time has less consequences on the MD a posteriori expected claim frequency than on the TPL a posteriori expected claim frequency. One explanation is that in average the claim frequency related to MD:TPL claims, $\lambda^{MD:TPL}$ , only plays for about 28% of the total claim frequency in MD (i.e. $\lambda^{MD}+ \lambda^{MD:TPL}$ ), whereas it is about 55% of the total claim frequency in TPL (i.e. $\lambda^{TPL}+ \lambda^{MD:TPL}$ ). Hence, the correction arising from the conditional expectation of $\Theta^{MD:TPL}$ is less weighted (i.e. in the convex combination (2)) in the MD case than in the TPL case.

Finally, let us analyse the correction factors when a claim was reported in Home insurance (Figures 5 and 6) and compare them to those when no claim was reported (Figure 4). Both in TPL and in MD, we observe that after 1 year, there is a difference of around 10%. As time passes, the difference shrinks to around 5% after 5 years. Consequently, we see that taking into account the reported claims in Home insurance to compute a posteriori expected claim frequencies is pertinent.

4.1.3. A posteriori corrections in Home insurance after a claim

We can also compute the corrections to apply to the a priori expected claim frequency in Home insurance as a function of the number of reported claims in both Home and Motor insurance. As before, we assume a household with two policies in Motor insurance as well as one Home insurance. The correction to apply to Home insurance is depicted in Figure 7. Note that, on the contrary to the Motor insurance case, only one correction applies, as the policyholder is assumed to be the whole household in this case.

Figure 7

Correction to apply to Home insurance a priori expected claim frequency when one claim was reported.

Symmetrically to the Motor insurance case, a claim in Home insurance appears to increase the most the correction factor to apply to the a priori expected claim frequency in Home insurance. We note, however, that in the long run, differences between lower- and higher-risk profiles appear to be large (about 10–15%). Only a claim in Home insurance yields a correction factor larger than 1 after 5 years. Moreover, note that the correction factors arising from a claim in any guarantee in Motor insurance are all equal, regardless of which guarantee was triggered. The reason of this comes from the common covariance between the Home insurance guarantee and the three random effects related to the three count variables in Motor insurance.

4.2. Impact of dependence between Home and Motor insurance on corrections

Let us assess on an example in what way the dependence between Motor and Home insurance changes the corrections to apply.

We will consider a household with two policyholders, each holding a policy in Motor insurance, covering both TPL and MD. In addition, the household also holds a Home insurance policy. To ease the presentation, we will consider a unique risk profile for each guarantee and each of the policyholders. More specifically, we will consider the median expected a priori claim frequency for each of the four count variables (i.e. medium-risk profile in the previous example).

In order to measure the impact of the dependence between Home and Motor insurance, we will distinguish two cases:

1. 1. The correlation structure of the random effects is given in Table 7;

2. 2. The same variance–covariance matrix as above, except for the off-diagonal terms of the first row (respectively, column), is set to 0 (i.e. independence between Home and Motor insurance is assumed). The dependence in Motor insurance for policyholders from the same household is therefore kept at there levels given in Table 7.

Six examples are considered: (i) a claim-free case (meaning that in all three policies no claim was reported), (ii) one claim in Home insurance (during the first semester), (iii) one claim from policyholder h(1) in Motor TPL, (iv) one claim from policyholder h(1) in Motor MD, (v) one claim from policyholder h(1) that triggered both TPL and MD simultaneously and (vi) two claims from policyholder h(1): one triggering TPL and the other triggering MD.

4.2.1. Impact on Home insurance

Let us first show the corrections to apply in Home insurance, by discussing the six examples detailed above. The corrections in Motor insurance will be presented later. The corrections are depicted in Figure 8.

Figure 8

Impact of the dependence of the random effects related to Home and Motor insurance measured by the corrections in Home insurance. Six cases are considered, whether a claim occurred or not, and if so, which guarantees were triggered by the claim.

In the claim-free case, we see that both Motor policies help decrease the estimate by an extra 10% compared to the independence case. In the second example, where one claim was reported in Home insurance, the correction is above 1, but the dependence allows this correction to be weaker at first. As time goes by, the correction decreases faster than in the independence case thanks to the claim-free years in Motor insurance. In the next three examples, in which one claim was reported in Motor insurance, we observe similar corrections to be applied in Home insurance. We see that even though a claim was reported in Motor insurance, the correction factor in Home insurance falls below 1 after about 2 years. Note that the difference with the first example (i.e. no claims in any guarantee) is about 5% after 5 years when considering the dependence. So, correcting the Home insurance a priori expected claim frequency without considering the Motor insurance experience may yield too advantageous corrections in the considered examples. In the last example, we note the importance of distinguishing when a single event is triggering two guarantees (i.e. only one claim) compared to two events each triggering a different guarantee (i.e. two claims). Indeed, in the former case, the correction factor in Home insurance is 5–10% lower than in the latter case. These differences also exhibit the importance of identifying whether two claim counts are in fact related to a single event (i.e. one claim triggering both guarantees) or whether they are unrelated and should be considered as two separate claims.

Finally, note that as could be expected, the corrections in the independence case are the same in five of the six examples (i.e. when no claim was reported in Home insurance).

4.2.2. Impact on Motor insurance

Let us now assess the corrections to apply in Motor insurance (TPL and MD), by discussing the same six examples as previously. Even though the modelling in Motor insurance involves three count variables, we can compute two correction factors, one for TPL and one for MD, using the formulas given in (2). The corrections are shown in Figure 9 for TPL and in Figure 10 for MD. Note that only the corrections for policyholder h(1) (who is the only policyholder among the two policyholders in the household that experiences claims in Motor insurance) are shown. By looking at Figure 9 (respectively, Figure 10) which displays the corrections to apply to h(1) in TPL (respectively, MD), we note that the dependence between Home and Motor insurance does not seem to considerably change the corrections. While the independence case considers that there is independence between Home and Motor insurance, no independence between guarantees in Motor insurance for different policyholders from the same household is assumed. This means that the observed number of claims of any policyholder in Motor insurance remains relevant in both cases displayed in Figures 9 and 10. Let us explain why the additional information related to Home insurance does not change a lot the correction factors. Compared to the aggregated a priori expected claim frequencies for two policies in Motor insurance (TPL and MD), the a priori expected claim frequency in Home insurance is actually small. Actually, it is comparable to the one of one policyholder in Motor TPL, so that it amounts to around $1/7$ of the total expected claim frequency in the household. This means that a claim-free year in Home insurance will have a little impact on the correction factors in TPL and MD. Only when a claim occurs in Home insurance, it appears to have an impact on the correction factors in TPL and MD. In that situation, Figures 9 and 10 show that after 5 years there is a difference of about 5% in the correction factors to apply to TPL and MD due to the dependence with Home insurance.

Figure 9

Impact of the dependence of the random effects related to Home and Motor insurance measured by the corrections in Motor TPL for h(1). Six cases are considered, whether a claim occurred or not, and if so, which guarantees were triggered by the claim.

Figure 10

Impact of the dependence of the random effects related to Home and Motor insurance measured by the corrections in Motor MD for h(1). Six cases are considered, whether a claim occurred or not, and if so, which guarantees were triggered by the claim.

The corrections to apply to h(2) (who does not experience any claims in any of the six cases) exhibit a similar difference between the independence case and the dependence case.

5.1. Poisson deviance

Let us first use the Poisson deviance. Indeed, one way to measure the predictive power of a model with count data is to compute the Poisson deviance on the out-of-sample data. The Poisson deviance has been computed on each fold on the third year’s predicted claim frequencies and observed number of claims.

First, we computed the deviance on Home insurance, Motor TPL and Motor MD separately. In Figure 11, on top, we see that the GAMs alone, as expected, have the worst performance in the three cases, and the multivariate Poisson mixture outperforms the univariate Poisson mixtures. We note, however, that in Home insurance, the difference is small between the univariate and the multivariate credibility models. This difference is, however, the greatest in Motor TPL (which is where the greatest heterogeneity has been found). On the bottom, the three cases are summed to obtain the deviance of the whole portfolio over all insurance products. Using a univariate Poisson mixture allows to take into account part of the heterogeneity and ‘improves therefore the predictability. However, the GAMs along with multivariate credibility appear to outperform both models, on every fold, by exploiting the dependence between the latent risk factors.

Figure 11

Poisson deviance on out-of-sample data. Three models are considered: GAMs, GAMs corrected by a univariate credibility model and GAMs corrected by the multivariate credibility model. Top: the deviance for Home insurance and Motor TPL and MD is illustrated. Bottom: total deviance (sum of the three deviances).

Let us note that small differences in deviances for insurance type count data (i.e. low claim frequencies) can in reality imply large differences in the predicted claim frequencies. Therefore, we will also use a second method to assess the predictive power of our model by using the lift.

5.2. Lift

Another way to compare the predictive performance is to rely on the lift. Define the relativity $r^g_{h(i)}$ as the ratio of the a posteriori expected claim frequency according to the competing model to that of the benchmark model. If the model does not include any random effect (model 1), then the a posteriori expectation reduces to $\widehat{\lambda}_{h(i), T+1}^g$ .

Based on these relativities, it is then possible to bin policyholders into groups of similar exposure. Then, on each bin, the loss ratio, that is, the ratio of the sum of the actual number of claims observed in the bin to the sum of the expected number of claims in the bin given by the benchmark model. A loss ratio of 1 for each group means that the benchmark model is perfectly accurate. The loss ratios for each group can be plotted and, if an upward trend is visible, it implies that the competing model outperforms the benchmark model. The intuitive idea is as follows. Since we have sorted the relativities, from smallest to largest, if we have an upward trend, then it means that for small relativities, the loss ratio is below 1. Consequently, since the relativity is small, the competing model is predicting less claims than the benchmark model, and the loss ratio below 1 implies that there are less actual claims than predicted by the benchmark model. Therefore, it could be that competing model is predicting better than the benchmark model. A similar argument holds for larger relativities. See Henckaerts et al. (Reference Henckaerts, Côté, Antonio and Verbelen2019) for a comprehensive presentation of the lift.

Let us use the lift methodology in two different comparisons. First, we will compare model 3 (comp) with model 1 (bench). Then, we will compare model 3 (comp) with model 2 (bench). In each comparison, we will consider the three different claim frequencies that have been covered in this paper: claim frequencies in Home insurance, in Motor TPL and in Motor MD.

The computations are done in the third year of fold 1, when the model has been estimated on the four remaining folds (i.e. folds 2–5). Since we are using the third year to compute the lift, we can use the two first years of observations to compute a posteriori corrections for the third year. When comparing model 3 with model 1, the relativities can be written as

\begin{align*}\begin{cases} r_{h(i)}^{Home} = \dfrac{\lambda_{h,3}^{Home} \text{E}\left[\Theta_{h}^{Home}| \bf{N_{h, 1-2}} = \bf{n_{h, 1-2}}\right]}{\lambda_{h,3}^{Home}} = \text{E}\left[\Theta_{h}^{Home}|\bf{N_{h, 1-2}} = \bf{n_{h, 1-2}}\right]\\\\[-5pt] r_{h(i)}^{TPL} = \dfrac{\lambda_{h(i),3}^{TPL} \text{E}\left[\Theta_{h(i)}^{TPL}| \bf{N_{h, 1-2}} = \bf{n_{h, 1-2}}\right] + \lambda_{h(i),3}^{MD:TPL} \text{E}\left[\Theta_{h(i)}^{MD:TPL}| \bf{N_{h, 1-2}} = \bf{n_{h, 1-2}}\right]}{\lambda_{h(i),3}^{TPL} + \lambda_{h(i),3}^{MD:TPL}}\\\\[-4pt] r_{h(i)}^{MD} = \dfrac{\lambda_{h(i),3}^{MD} \text{E}\left[\Theta_{h(i)}^{MD}| \bf{N_{h, 1-2}} = \bf{n_{h, 1-2}}\right] + \lambda_{h(i),3}^{MD:TPL} \text{E}\left[\Theta_{h(i)}^{MD:TPL}| \bf{N_{h,1-2}} = \bf{n_{h, 1-2}}\right]}{\lambda_{h(i),3}^{TPL} + \lambda_{h(i),3}^{MD:TPL}}\end{cases}\end{align*}

where $\bf{n_{h, 1-2}}$ is a vector indicating the number of claims observed in every policy of the household during years 1 and 2, that is, in a household h with a Home insurance policy and two Motor insurance policies,

\begin{equation*}\bf{n_{h, 1-2}} = \sum_{t=1}^2 \left(n_{h,t}^{Home}, n_{h(1),t}^{TPL}, n_{h(1),t}^{MD}, n_{h(1),t}^{MD:TPL}, n_{h(2),t}^{TPL}, n_{h(2),t}^{MD}, n_{h(2),t}^{MD:TPL}\right)'\end{equation*}

while when comparing model 3 with model 2, the relativities are given by

\begin{align*}\begin{cases}r_{h(i)}^{Home} = \dfrac{\lambda_{h,3}^{Home} \text{E}\left[\Theta_{h}^{Home}| \bf{N_h} = \bf{n_h}\right]}{\lambda_{h,3}^{Home} \text{E}\left[\Theta_{h}^{Home}| N_h^{Home} = n_h^{Home}\right]} \\\\[-5pt]r_{h(i)}^{TPL} = \dfrac{\lambda_{h(i),3}^{TPL} \text{E}\left[\Theta_{h(i)}^{TPL}| \bf{N_h} = \bf{n_h}\right] + \lambda_{h(i),3}^{MD:TPL} \text{E}\left[\Theta_{h(i)}^{MD:TPL}| \bf{N_h} = \bf{n_h}\right]}{\lambda_{h(i),3}^{TPL} \text{E}\left[\Theta_{h(i)}^{TPL}| N_{h, \bullet}^{TPL} = n_h^{TPL}\right] + \lambda_{h(i),3}^{MD:TPL} \text{E}\left[\Theta_{h(i)}^{MD:TPL}| N_{h, \bullet}^{MD:TPL} = n_h^{MD:TPL}\right]}\\\\[-5pt]r_{h(i)}^{MD} = \dfrac{\lambda_{h(i),3}^{MD} \text{E}\left[\Theta_{h(i)}^{MD}| \bf{N_h} = \bf{n_h}\right] + \lambda_{h(i),3}^{MD:TPL} \text{E}\left[\Theta_{h(i)}^{MD:TPL}| \bf{N_h} = \bf{n_h}\right]}{\lambda_{h(i),3}^{MD} \text{E}\left[\Theta_{h(i)}^{MD}| N_{h, \bullet}^{MD} = n_h^{MD}\right] + \lambda_{h(i),3}^{MD:TPL} \text{E}\left[\Theta_{h(i)}^{MD:TPL}| N_{h, \bullet}^{MD:TPL} = n_h^{MD:TPL}\right]}\end{cases}\end{align*}

The policies are then sorted based on their relativity and are binned into 25 risk classes of equal (or almost equal) exposures. On each of these risk classes, the aforementioned loss ratio can be computed, comparing for each risk class the actual number of reported claims with the expected number of reported claims (given by the benchmark model).

The resulting loss ratios are illustrated in Figure 12 in blue. We have used a local polynomial regression to compute the smooth curves. We also computed in dark orange the loss ratios with the competing model (i.e. the multivariate credibility model) in the denominator.

Figure 12

Lift on out-of-sample data. Two comparisons are considered. Left: multivariate credibility is compared to the predictions of the GAMs. Right: multivariate credibility is compared to the predictions of univariate credibility.

First, let us comment on the plots on the left, which compare our multivariate credibility model (model 3) to the GAMs (model 1). The blue curves suggest an increasing trend, in particular, in the higher risk classes. The higher risk classes imply that the multivariate credibility model yields higher claim frequencies than the GAMs, whereas the high loss ratio observed on this risk classes suggests that the GAMs underestimate clearly the claim frequencies. As explained above, if the blue line is not close to 1, then the benchmark model is not perfectly accurate. Moreover, if an uprising trend is visible, then, in fact, the competing model is outperforming the benchmark model. Therefore, one can say that the multivariate credibility model will at least less underestimate the expected claim frequencies, but could in fact overestimate the expected claim frequencies. This is the reason why we also plotted the orange curve which computes the loss ratio between the actual and the expected number of claims (given by the multivariate credibility model). As can be seen, for instance, on the plot on top left, the orange curve is much closer to the horizontal line at level 1. Being close to this red line suggests that the expected number of claims are closer to the actual number of claims (i.e. the model is predicting better on this risk class). It has to be noted that for the orange curve, only the closeness to the horizontal line at level 1 has to be seen (since we are not comparing the predictions with respect to another model). As such, we see that for the three plots on the left, the orange curve is, for most risk classes, closer to the red line than the blue curve is (i.e. the multivariate credibility model outperforms the GAMs on these risk classes). This is especially true for the higher risk classes, where our multivariate credibility model was able to account for the higher latent risk factors.

Let us now compare the multivariate credibility model (model 3) with the univariate credibility model (model 2). In TPL and MD, the blue lines again suggest an increasing trend, especially for the second half of the risk classes. For Home insurance, this upward trend is not as visible. However, when looking at the orange curve, we can see that again, for many classes, the multivariate credibility model is close to predicting the actual number of claims on each risk class (i.e. close to 1). We note that again, there is a big discrepancy in this last risk classes (i.e. the riskiest classes). Indeed, compared to the univariate credibility model, the multivariate credibility is able to better identify the riskiest households by using information related to other products and other policyholders from the household.

6.1. Description of the algorithm

In this paper, model fit is based on a two-step approach. First, the expected a priori claim frequencies are estimated with GAMs using the a priori available information. The expected a priori claim frequencies are then fixed and random effects are introduced. In the second step, the variance–covariance matrix of these random effects is estimated. Therefore, one could say that a limitation of our approach is that parameters in the GAMs could depend on the variance–covariance matrix and could therefore have different estimates, would we have estimated both the parameters of the GAMs and the variance–covariance matrix at the same time.

In this section, we aim to conduct an experiment and show how we can actually cycle both steps. We can summarise the cycling of the two-part approach in the following ways:

1. 1. We compute the conditional expectation of the random effects in the current model. If this is the first time we iterate, these conditional expectations are fixed at 1.

2. 2. The GAMs are estimated, with the conditional expectations found in step 1 put in the offset.

3. 3. The variance–covariance matrix of the random effects is estimated by using the estimated expected a priori claim frequencies from the previous step.

4. 4. Repeat steps 1–3, until convergence is achieved.

The algorithm described resembles to what is known as an ECM algorithm (see Meng & Rubin Reference Meng and Rubin1993). Indeed, the E-step corresponds in our situation to conditional expectation computed in step 1. Then, the conditional maximisation step corresponds to both steps 2 and 3. One key difference, however, relates to the fact that at both M-steps (steps 2 and 3 in the described algorithm above), two different objective functions are maximised, while Meng & Rubin (Reference Meng and Rubin1993) maximise the same objective function at the different M-steps. Indeed, at step 2, the Poisson log-likelihood is maximised, whereas at step 3, the log-likelihood of the multivariate Poisson mixture is maximised.

We stress out that, since this algorithm cannot be considered as an ECM algorithm, further research is necessary to assess its validity. Therefore, this section, along with the practical application of the algorithm that follows, should only be considered as a preliminary experiment and further research could be devoted to this issue.

6.2. Practical application of the algorithm

This algorithm has been implemented on the model presented in this paper. The steps have been repeated five times. We wish now to illustrate the coefficients of the discrete variables in Tables 8–11. The parameters do appear to change a bit after the first iteration and seem to converge rapidly. Note that, for confidentiality reasons, we are not allowed to show the intercept. We can, however, state that the absolute change in the intercept between the first and last iterations was as follows: $0.003$ in TPL, $-0.018$ in MD, $-0.024$ in MD:TPL and $-0.007$ in Home insurance.

Table 8.

Estimate of discrete parameters in TPL at each iteration of the cycling algorithm.

Table 9.

Estimate of discrete parameters in MD at each iteration of the cycling algorithm.

Table 10.

Estimate of discrete parameters in MD:TPL at each iteration of the cycling algorithm.

Table 11.

Estimate of discrete parameters in Home insurance at each iteration of the cycling algorithm.

Figure 13

Effect on score at the first iteration and the last iteration ( $= 5$ ) for the univariate continuous variables.

For the continuous variables, we show, in Figure 13, the estimated effects at iterations 1 and 5 on the log scale. We see that the differences remain small (below 1% in most cases).

Finally, for the geographic effect, Figures 14 and 15 compare the estimated effects at iterations 1 and 5.

Figure 14

Effect on score of the geographic effects estimated at the first iteration and the last iteration ( $= 5$ ) in Motor insurance.

Figure 15

Effect on score of the geographic effect estimated at the first iteration and the last iteration ( $= 5$ ) in Home insurance.

We also comment on the parameters involved in the variance–covariance matrix. The estimates at the different steps are given in Table 12. We note that, similarly to the parameters of the GAMs, the changes occur mainly at the first iteration steps, and convergence is rapid.

To sum up, cycling the different steps produces estimates that rapidly stabilise, with only limited changes between first and last iterations.