2.1. Logit-weighted reduced mixture of experts

Let ${\textit{\textbf{x}}}_{i} = (x_{i0}, x_{i1}, \dots, x_{iP})^{T}$ denote the $(P+1)$-dimensional covariate vector for policyholder i ($i=1, 2, \dots, n$) with intercept $x_{i0} = 1$, and ${\textit{\textbf{y}}}_{i} = (y_{i1}, y_{i2}, \dots, y_{iD})^{T}$ denote the D-dimensional vector of their response variables, which can be either claim frequency or severity. Let ${\textit{\textbf{X}}} = ({\textit{\textbf{x}}}_{1}, {\textit{\textbf{x}}}_{2}, \dots, {\textit{\textbf{x}}}_{n})^{T}$ and ${\textit{\textbf{Y}}} = ({\textit{\textbf{y}}}_{1}, {\textit{\textbf{y}}}_{2}, \dots, {\textit{\textbf{y}}}_{n})^{T}$ denote all covariates and responses for a group of n policyholders.

Based on the covariates, policyholder i is classified into one of the g latent risk classes by a logit gating function

(1) \begin{equation} \pi_{j}({\textit{\textbf{x}}}_{i}; {\boldsymbol{\alpha}}_{j}) = \frac{exp({\boldsymbol{\alpha}}^{T}_{j}{\textit{\textbf{x}}}_{i})}{\sum_{j^{\prime}=1}^{g} exp({\boldsymbol{\alpha}}^{T}_{j^{\prime}}{\textit{\textbf{x}}}_{i}) }, \quad j = 1, 2, \dots, g\end{equation}

where ${\boldsymbol{\alpha}}_{j} = (\alpha_{j0}, \alpha_{j1}, \dots, \alpha_{jP})^{T}$ is a vector of regression coefficients for latent class j.

Given the assignment of latent class $j \in \{ 1, 2, \dots, g \}$, the response variables ${\textit{\textbf{y}}}_{i} = (y_{i1}, y_{i2}, \dots, y_{iD})^{T}$ are conditionally independent. For $d = 1, \dots, D$ and $i=1,\ldots,n$, the dth dimension probability density function (or probability mass function) of $y_{id}$ is given by

(2) \begin{equation} g_{jd}(y_{id}; \delta_{jd}, {\boldsymbol{\psi}}_{jd}) = \delta_{jd}\boldsymbol{1}\{ y_{id}=0\} + (1-\delta_{jd})f_{jd}(y_{id}; {\boldsymbol{\psi}}_{jd})\end{equation}

where $\delta_{jd}\in [0,1]$ is a parameter representing the probability of assigning an observation into a zero-inflated component (i.e. a probability mass at zero), $\boldsymbol{1}\{ y_{id}=0\}$ is the indicator function, and $f_{jd}(y_{id}; {\boldsymbol{\psi}}_{jd})$ is the density of some commonly used distribution with parameters ${\boldsymbol{\psi}}_{jd}$ for modelling insurance losses. Table 1 gives a list of parametric distributions supported by the LRMoE package. Note that a probability mass $\delta_{jd}$ at zero allows for more realistic modelling of insurance data which usually exhibit excess zeros. As a naming convention, we will refer to $g_{jd}$ as a component distribution/expert function, and $f_{jd}$ as its positive part, although claim frequency distributions (e.g. Poisson) also have some probability mass at zero.

Table 1.

Distributions supported by LRMoE.

Under LRMoE, the conditional probability density function (or probability mass fuction) of ${\textit{\textbf{y}}}_i$ given covariates ${\textit{\textbf{x}}}_i$ is

(3) \begin{equation} h({\textit{\textbf{y}}}_i; {\textit{\textbf{x}}}_i, {\boldsymbol{\alpha}}, {\boldsymbol{\delta}}, {\boldsymbol{\Psi}}) = \sum_{j=1}^{g} \left[ \pi_{j}({\textit{\textbf{x}}}_{i}; {\boldsymbol{\alpha}}_{j}) \times \prod_{d=1}^{D} g_{jd}(y_{id}; \delta_{jd}, {\boldsymbol{\psi}}_{jd}) \right]\end{equation}

where ${\boldsymbol{\alpha}} = ({\boldsymbol{\alpha}}_{1}, {\boldsymbol{\alpha}}_{2}, \dots, {\boldsymbol{\alpha}}_{g})^{T}$ is a $g \times (P+1)$-matrix of mixing weights with ${\boldsymbol{\alpha}}_{g} = (0, 0, \dots, 0)^{T}$ representing the default class, ${\boldsymbol{\delta}} = (\delta_{jd})_{1 \leq j \leq g, 1 \leq d \leq D}$ is a $(g \times D)$-matrix of probability masses at zero by component and by dimension, and ${\boldsymbol{\Psi}} = \{ {\boldsymbol{\psi}}_{jd}: 1 \leq j \leq g, 1 \leq d \leq D \}$ is a list of parameters for the positive part $f_{jd}$ by component and by dimension. Note that ${\boldsymbol{\alpha}}_{g} = (0, 0, \dots, 0)^{T}$ ensures that the model is identifiable, i.e. there is a one-to-one mapping between the regression distributions and the parameters (see Jiang & Tanner Reference Jiang and Tanner1999 and Fung et al. Reference Fung, Badescu and Lin2019a). Note that the total number of parameters of the model is given by $(g-1)\times (P+1)+g\times D+\sum_{j=1}^g\sum_{d=1}^{D}\#\{{\boldsymbol{\psi}}_{jd}\}$, where $\#\{{\boldsymbol{\psi}}_{jd}\}$ is the length of ${\boldsymbol{\psi}}_{jd}$.

For a group of n policyholders, the likelihood function is given by

(4) \begin{equation} L({\boldsymbol{\alpha}}, {\boldsymbol{\delta}}, {\boldsymbol{\Psi}}; {\textit{\textbf{X}}}, {\textit{\textbf{Y}}}) = \prod_{i=1}^{n} h({\textit{\textbf{y}}}_i; {\textit{\textbf{x}}}_i, {\boldsymbol{\alpha}}, {\boldsymbol{\delta}}, {\boldsymbol{\Psi}}) = \prod_{i=1}^{n}\left\{ \sum_{j=1}^{g} \left[ \pi_{j}({\textit{\textbf{x}}}_{i}; {\boldsymbol{\alpha}}_{j}) \times \prod_{d=1}^{D} g_{jd}(y_{id}; \delta_{jd}, {\boldsymbol{\psi}}_{jd}) \right] \right\}\end{equation}

2.2. Parameter estimation

For parameter estimation in finite mixture models, the Expectation–Maximisation (EM) algorithm is commonly used (see, e.g. Dempster et al. Reference Dempster, Laird and Rubin1977 and McLachlan & Peel Reference McLachlan and Peel2004). However, for the LRMoE model, the M-step requires maximisation of a non-concave function over all elements of ${\boldsymbol{\alpha}}$. Fung et al. (Reference Fung, Badescu and Lin2019a) thus use the ECM algorithm (Meng & Rubin Reference Meng and Rubin1993), which breaks the M-step into several substeps. The ECM algorithm implemented in the LRMoE package is described as follows.

Denote ${\boldsymbol{\Phi}} = ({\boldsymbol{\alpha}}, {\boldsymbol{\delta}}, {\boldsymbol{\Psi}})$ as the parameters to estimate. For $i = 1, 2, \dots, n$, we introduce a latent random vector ${\textit{\textbf{Z}}}_{i} = (Z_{i1}, Z_{i2}, \dots, Z_{ig})^{T}$, where $Z_{ij} = 1$ if ${\textit{\textbf{y}}}_{i}$ comes from the jth component distribution and $Z_{ij} = 0$ otherwise. For $d = 1, 2, \dots, D$, we further write $Z_{ij} = Z_{ijd0} + Z_{ijd1}$, where $Z_{ijd0}= 0$ and $Z_{ijd1}= 1$ if the dth dimension of ${\textit{\textbf{y}}}_{i}$ comes from the positive part $f_{jd}$ of the jth component, and $Z_{ijd0}= 1$ and $Z_{ijd1}= 0$ if it comes from the zero inflation $\delta_{jd}$.

The complete-data log-likelihood function is given by

(5) \begin{equation}\begin{aligned} l^{\textrm{com}}\left({\boldsymbol{\Phi}}; {\textit{\textbf{X}}}, {\textit{\textbf{Y}}}, {\textit{\textbf{Z}}}\right) & = \sum_{i=1}^{n}\sum_{j=1}^{g} Z_{ij} \left\{ \log \pi_{j}\left({\textit{\textbf{x}}}_{i}; {\boldsymbol{\alpha}}_{j}\right) + \sum_{d=1}^{D}\log g_{jd}(y_{id}; \delta_{jd}, {\boldsymbol{\psi}}_{jd}) \right\} \\ & = \sum_{i=1}^{n}\sum_{j=1}^{g} Z_{ij} \log \pi_{j}({\textit{\textbf{x}}}_{i}; {\boldsymbol{\alpha}}_{j}) + \sum_{i=1}^{n}\sum_{j=1}^{g}\sum_{d=1}^{D} \left\{ Z_{ijd0}\log \delta_{jd} + Z_{ijd1}\log (1- \delta_{jd}) \right\} \\ & + \sum_{i=1}^{n}\sum_{j=1}^{g}\sum_{d=1}^{D} Z_{ijd1}\log f_{jd}\left(y_{id}; {\boldsymbol{\psi}}_{jd}\right)\end{aligned}\end{equation}

E-Step:

For each $i=1, 2, \dots, n$, the random vector ${\textit{\textbf{Z}}}_{i}$ follows a multinomial distribution with count 1 and probabilities $(\pi_{1}({\textit{\textbf{x}}}_{i}; {\boldsymbol{\alpha}}_{1}), \pi_{2}({\textit{\textbf{x}}}_{i}; {\boldsymbol{\alpha}}_{2}), \dots, \pi_{g}({\textit{\textbf{x}}}_{i}; {\boldsymbol{\alpha}}_{g}))$. Given $Z_{ij}=1$, the conditional distribution of $Z_{ijd0}$ is Bernoulli with probability $\delta_{jd}$. Hence, at the tth iteration, the posterior expectations of $Z_{ij}$, $Z_{ijd0}$ and $Z_{ijd1}$ are

(6) \begin{equation}\begin{aligned} z^{(t)}_{ij} & = \texttt{E}\{ Z_{ij} | {\boldsymbol{\Phi}}^{(t-1)}, {\textit{\textbf{X}}}, {\textit{\textbf{Y}}} \} = \texttt{P}\{ Z_{ij} = 1 | {\boldsymbol{\Phi}}^{(t-1)}, {\textit{\textbf{X}}}, {\textit{\textbf{Y}}} \} \\ & = \frac{ \pi_{j}\left({\textit{\textbf{x}}}_{i}; {\boldsymbol{\alpha}}_{j}^{(t-1)}\right) \times \prod_{d=1}^{D} g_{jd}\left(y_{id}; \delta_{jd}^{(t-1)}, {\boldsymbol{\psi}}_{jd}^{(t-1)}\right) }{\sum_{j^{\prime}=1}^{g} \pi_{j^{\prime}}\left({\textit{\textbf{x}}}_{i}; {\boldsymbol{\alpha}}_{j^{\prime}}^{(t-1)}\right) \times \prod_{d=1}^{D} g_{j^{\prime}d}\left(y_{id}; \delta_{j^{\prime}d}^{(t-1)}, {\boldsymbol{\psi}}_{j^{\prime}d}^{(t-1)}\right)},\end{aligned}\end{equation}

(7) \begin{equation}\begin{aligned} z^{(t)}_{ijd0} & = \texttt{E}\{ Z_{ijd0} | {\boldsymbol{\Phi}}^{(t-1)}, {\textit{\textbf{X}}}, {\textit{\textbf{Y}}} \} \\ & = \texttt{P}\{ Z_{ijd0} = 1 | {\boldsymbol{\Phi}}^{(t-1)}, {\textit{\textbf{X}}}, {\textit{\textbf{Y}}}, Z_{ij} = 1 \} \times \texttt{P}\{ Z_{ij} = 1 | {\boldsymbol{\Phi}}^{(t-1)}, {\textit{\textbf{X}}}, {\textit{\textbf{Y}}} \} \\ & = \frac{\delta_{jd}^{(t-1)} \boldsymbol{1}\{y_{id}=0\} }{\delta_{jd}^{(t-1)} \boldsymbol{1}\{y_{id}=0\} + \left(1-\delta_{jd}^{(t-1)}\right)F_{jd}\left(0; {\boldsymbol{\psi}}_{jd}^{(t-1)}\right)} \times z^{(t)}_{ij}\end{aligned}\end{equation}

and

(8) \begin{equation}\begin{aligned} z^{(t)}_{ijd1} & = z^{(t)}_{ij} - z^{(t)}_{ijd0}\end{aligned}\end{equation}

where $F_{jd}$ is the cumulative distribution function of the positive part $f_{jd}$.

CM-Step:

In the CM-step, we aim to maximise $Q({\boldsymbol{\Phi}}; {\boldsymbol{\Phi}}^{(t-1)}, {\textit{\textbf{X}}}, {\textit{\textbf{Y}}})$, which can be decomposed into three parts as

(9) \begin{equation} Q({\boldsymbol{\Phi}}; {\boldsymbol{\Phi}}^{(t-1)}, {\textit{\textbf{X}}}, {\textit{\textbf{Y}}}) = Q^{(t)}_{{\boldsymbol{\alpha}}} + Q^{(t)}_{{\boldsymbol{\delta}}} + Q^{(t)}_{{\boldsymbol{\Psi}}}\end{equation}

where

(10) \begin{equation} Q^{(t)}_{{\boldsymbol{\alpha}}} = \sum_{i=1}^{n}\sum_{j=1}^{g} z^{(t)}_{ij} \log \pi_{j}({\textit{\textbf{x}}}_{i}; {\boldsymbol{\alpha}}_{j}),\end{equation}

(11) \begin{equation} Q^{(t)}_{{\boldsymbol{\delta}}} = \sum_{i=1}^{n}\sum_{j=1}^{g}\sum_{d=1}^{D} \left\{ z^{(t)}_{ijd0}\log \delta_{jd} + z^{(t)}_{ijd1}\log (1- \delta_{jd}) \right\}\end{equation}

and

(12) \begin{equation} Q^{(t)}_{{\boldsymbol{\Psi}}} = \sum_{i=1}^{n}\sum_{j=1}^{g}\sum_{d=1}^{D} z^{(t)}_{ijd1}\log f_{jd}\left(y_{id}; {\boldsymbol{\psi}}_{jd}\right)\vspace*{-4pt}\end{equation}

To maximise $Q^{(t)}_{{\boldsymbol{\alpha}}}$, we use the same conditional maximisation as described in Fung et al. (Reference Fung, Badescu and Lin2019a). We first maximise it with respect to ${\boldsymbol{\alpha}}_{1}$ with ${\boldsymbol{\alpha}}_{j}$ fixed at ${\boldsymbol{\alpha}}_{j}^{(t-1)}$ for $j = 2, 3, \dots, g-1$. The next step is to maximise with respect to ${\boldsymbol{\alpha}}_{2}$ with updated ${\boldsymbol{\alpha}}_{1}^{(t)}$ and other ${\boldsymbol{\alpha}}_{j}$ fixed at ${\boldsymbol{\alpha}}_{j}^{(t-1)}$ for $j = 3, 4, \dots,\break g-1$. The process continues until all ${\boldsymbol{\alpha}}$’s have been updated. For obtaining each ${\boldsymbol{\alpha}}_{j}^{(t)}$, the Iteratively Reweighted Least Square (IRLS) approach (Jordan & Jacobs Reference Jordan and Jacobs1994) is used until convergence.

For $Q^{(t)}_{{\boldsymbol{\delta}}}$, each $\delta_{jd}$ can be updated using the following closed-form solution

(13) \begin{equation} \delta_{jd}^{(t)} = \frac{\sum_{i=1}^{n} z^{(t)}_{ijd0}}{\sum_{i=1}^{n} \left(z^{(t)}_{ijd0} + z^{(t)}_{ijd1}\right)}\end{equation}

The maximisation of $Q^{(t)}_{{\boldsymbol{\Psi}}}$ can also be divided into smaller problems by component j and by dimension d. For each $j=1,\ldots,g$ and $d=1,\ldots,D$, the problem is reduced to maximise a weighted log-likelihood of an expert function where the weights of each observations are given by $z^{(t)}_{ijd1}$. For updating each ${\boldsymbol{\psi}}^{(t)}_{jd}$, therefore, closed-form solutions are only available for very special distributions (e.g. Poisson, Lognormal). Numerical optimisation is used in most cases, especially when the observation ${\textit{\textbf{y}}}_{i}$’s are not observed exactly (see section 2.3).

As discussed in McLachlan & Peel (Reference McLachlan and Peel2004), mixture of severity distributions may have unbounded likelihood, which leads to spurious models with extremely large or small parameter values. In the LRMoE package, we adopt the same maximum a posteriori (MAP) approach in Fung et al. (forthcoming), which uses appropriate prior distributions to penalise the magnitude of fitted parameters (see section 3). The rationale of including penalty functions is to avoid obtaining spurious model due to unbounded nature of log-likelihood function. The penalty itself should be small enough so that it results to negligible impacts on the fitted model. On the other hand, the penalty functions avoid parameters diverging to unreasonable values so that the fitted model would become more robust. For more details regarding to the rationales and executions, readers are recommended to refer to section 4 of Fung et al. (forthcoming).

2.3. LRMoE with censoring and truncation

Censoring and truncation are common in insurance data sets and need to be dealt with. For example, when a policy limit is applied, loss amounts above the limit will be recorded as the limit only, which creates right censoring of the complete loss data; when a policy deductible is applied, loss amounts below the deductible are not reported to the insurer, thus leading to left truncation.

Fung et al. (Reference Fung, Badescu and Lin2020a) have discussed the LRMoE model with censoring and truncation, where all component distributions are Gamma. For parameter estimation with data censoring and truncation, the ECM algorithm in section 2.2 is slightly modified, with an additional E-step to remove the uncertainties arising from censoring and truncation. Since the main purpose of this paper is to demonstrate the application of LRMoE package, we will omit the details and refer interested readers to the cited paper.

For all distributions included in it, the LRMoE package can perform parameter estimation in the presence of data truncation and censoring. Consequently, the user’s input is slightly different from existing packages for mixture models. A detailed example on model fitting in our package is given in section 3.

2.4. Ratemaking and reserving in LRMoE

The model structure of LRMoE allows for easy computation of quantities relevant to actuarial ratemaking and reserving (see Fung et al. Reference Fung, Badescu and Lin2019b). At the policyholder level, the moments and common measures of dependence (e.g. Kendall’s tau and Spearman’s rho) of ${\textit{\textbf{y}}}_i$ can be computed in simple forms. The value-at-risk (VaR) and conditional tail expectation (CTE) can also be numerically solved without much difficulty. Various premium principles can be applied to price insurance contracts, including pure premium, standard deviation (SD) premium, limited expected value (LEV) and stop-loss (SL) premium. Risk measures can also be calculated for each individual policyholder (e.g. 99% VaR). At the portfolio level, simulation can be conducted to obtain the distribution of the aggregate loss of all policyholders, which is useful for calculating the total loss reserve and premium calculation. The simulation process is facilitated by a data simulator included in our package (see section 4.4).

4.1. Dataset and computation time

Throughout this section, we will use a French auto insurance claims dataset freMTPLfreq and freMTPLsev included in the R package CASdatasets (Dutang & Charpentier, Reference Dutang and Charpentier2019). Exploratory analysis and data cleaning procedures can be found on the accompanying website of our package. The cleaned dataset has 412,609 observations. The claim amount has high zero inflation, as less than 4% of policyholders have filed at least one claim. For positive claim amounts, the distribution is right-skewed, multimodal and heavy-tailed. The covariates used for modelling are described in Table 6.

Before proceeding, we remark on the computational time of our package. Compared with the demo dataset in section 3, analysing freMTPLfreq and freMTPLsev resembles a realistic actuarial modelling problem with many covariates and observations. In such case, the julia programming language is significantly more advantageous compared with traditional statistical languages such as R. Some standard benchmarks can be found on https://julialang.org/benchmarks/. Our experience confirms julia’s better performance: it takes around 20 hours in R (with optimisation in Rcpp) to fit a model in this section, while julia takes less than 5 hours to complete the same procedure. For the corresponding R packages which we have developed, we refer readers to Tseung et al. (Reference Tseung, Badescu, Fung and Lin2020).

4.2. Parameter initialisation

Since the fitting procedure of LRMoE involves multivariate optimisation, a good initialisation of parameters will often lead to faster convergence, compared with a non-informative guess. Gui et al. (Reference Gui, Huang and Lin2018) have proposed an initialisation procedure for fitting mixture of Erlang distributions, which involves k-means clustering and clusterised method of moments (CMM). This has been used in Fung et al. (forthcoming) and offers reasonably good starting values of parameters.

Our package contains an initialisation function which applies the CMM method to all component distributions. Some preliminary analysis is needed to determine the number of clusters (components) to use. Since the positive part of all distributions included in our package is unimodal, a heuristic starting point is to examine the empirical histogram of data and count the number of peaks (see Figure 3).

As an example, the procedure to initialise a three-component LRMoE is given as follows. The user needs to input response Y and covariate X, as used in the fitting function. In addition, the third argument 3 corresponds to the number of components, while the last argument [“continuous”] indicates that the response Y is one-dimensional and continuous. For the demo dataset used in section 3, the input would be [“discrete” “continuous”].

The cmm_init function will return a list of parameter initialisation. For the logit regression coefficients, we assume a non-informative guess on covariate influence, resulting in zero coefficients on all covariates. The user is free to incorporate prior knowledge by modification afterwards. The relative size of each cluster is reflected by the intercept terms. In the following initialisation, the size of three latent clusters is proportional to $(e^{-1.41667}, e^{-2.8687}, e^{0.0})$, i.e. the proportion of these clusters within the entire dataset is given by $(0.3842, 0.0330,\break 0.5827)$.

As for the parameters of component distributions, for each dimension of Y, we will apply the CMM method to obtain initialisation for all types of expert functions. A summary of all initialisation is given in Table 7. The initialisation function cmm_init also returns summary statistics (mean, coefficient of variation, skewness and kurtosis), which helps with choosing what combination of expert functions to use. Initialisation with extremely large or small parameter values could result in a spurious model or bad fit, thus should be avoided.

Table 7.

Sample initialisation of parameters.

Note: Since the moments cannot be written in closed form, these parameters are solved by maximising log-likelihood based on the observation in clusters.

The cmm_init function also returns two suggested models based on the highest log-likelihood (ll_best) and the best Kolmogorov–Smirnov test (ks_best). For example, the model initialisation with the highest log-likelihood is given by the following.

4.3. Fitting results and model selection

For illustration purposes, we fit only a selected number of LRMoEs to the entire French auto insurance dataset. The fitted log-likelihood, AIC and BIC are summarised in Table 8. In most cases, adding more components will increase the fitted log-likelihood (e.g. consider models iwl, will and wllil). The models selected by AIC and BIC have six and five components, respectively. In this particular setting, BIC heavily penalises models with more components, since the sample size is large, and adding one component roughly increases the number of parameters by 30 (the number of logit regression coefficients).

Table 8.

Fitting results of French auto insurance data.

Note 1: Component distributions are represented by first letter, for example, l for Lognormal, b for Burr, etc. All components are zero-inflated. The number in brackets is the total number of parameters.

Note 2: Stopping criterion is $<0.05$ improvement in log-likelihood, or 500 iterations.

Note 3: For each g, log-likelihood values are in decreasing order. The overall optimal values are bolded.

Apart from AIC and BIC, cross-validation (CV) is an alternative model selection criterion which avoids overfitting with too many latent components. For example, Gui et al. (Reference Gui, Huang and Lin2018) consider a 10-fold CV for fitting mixture of Erlangs, where the averaged log-likelihood on the test sets is used as a score function to select the optimal number of components. CV can be implemented with the help of parallel computing in julia (see also section 4.5).

4.4. Pricing and reserving functions

Our package contains a collection of functions related to actuarial pricing, reserving and risk management, including calculation of mean, variance, VaR, CTE, LEV $E[(Y \wedge u)]$ and SL premium $E[(Y - d)_+]$ of the response variable. These functions start with root predict_, followed by appropriate quantities of interest (mean, var, VaR, CTE, limit, excess) and corresponding function arguments.

As an illustration, we consider three policyholders in Table 9, where A has no claim history, B has a medium-sized claim and C has a large claim. Below is the sample code to calculate different quantities of interest.

Actuarial pricing calculation can be done based on either individual loss distributions, or the aggregated loss distribution of the entire portfolio. Both can be achieved using built-in functions in our package.

Table 9.

Selected policyholders in French auto insurance dataset.

On the individual level, the functions above can be called directly to calculate the pure premium E[Y], as well as the premium value in the presence of a policy deductible ($E[(Y - d)_+]$) or policy limit ($E[(Y \wedge u)]$). The first part of Table 10 summarises the premium calculation for Policyholders A, B and C, based on individual-level premium principles.

On the portfolio level, the sim_dataset function (see also section 4.6) will simulate one response value (i.e. claim amount) for each individual policyholder, which can be summed up to the aggregated portfolio loss under one possible scenario. Repeated simulation of the entire portfolio will produce the empirical distribution for the aggregated loss. The VaR and CTE of the aggregated loss can be obtained from the simulated sample, which are useful for setting the insurer’s total reserve. In addition, the VaR and CTE can be allocated back to policyholders as a loaded premium, according to some weighting scheme which reflects policyholders’ relative riskiness (e.g. relative magnitude of their pure premium). The second part of Table 10 summarises the premium calculation for Policyholders A, B and C, based on portfolio-level premium principles.

Table 10.

Pricing calculation for selected policyholders in French Auto Insurance Dataset. For portfolio-level premium principles, the allocation is based on the relative size of pure premium, where the weight for policyholder i is $w_i = E[Y_i]/\sum_{j}E[Y_j]$. For each policyholder and premium principle, the calculation is done based on both prior probability (left column) and posterior probability (right column).

4.5. Parameter uncertainty

In addition to obtaining point estimates of model parameters, it is crucial to also calculate their confidence intervals in order to identify significant parameters. Given that the log-likelihood in equation (4) becomes much more complicated when data are truncated and/or censored, analytically deriving the variance and confidence intervals of parameter estimates would not be feasible, and could be subject to numerical issues in implementation.

Instead, we can use bootstrapping methods in Grün & Leisch (Reference Leisch2004). A straightforward nonparametric bootstrapping algorithm is described as follows:

1. 1. Fit an LRMoE using the original dataset, and obtain an estimate $\hat{{\boldsymbol{\Phi}}}$.

2. 2. For a fixed number of total iterations, say 200, sample with replacement from the original dataset, on which to fit an LRMoE again with the same component distributions.

3. 3. The estimates $\hat{{\boldsymbol{\Phi}}}_1, \hat{{\boldsymbol{\Phi}}}_2, \dots, \hat{{\boldsymbol{\Phi}}}_{200}$ from the bootstrapped samples will provide an estimate of the variance of parameters.

The algorithm above can be easily implemented in parallel in julia as follows. We have included the estimated confidence intervals in the Appendix, calculated based on 200 bootstrapped samples.

4.6. Model simulation

In the LRMoE setting, the loss distributions of policyholders are mixtures of the same expert functions, but with potentially different mixing weights. Consequently, the distribution of the aggregate loss, as a sum of individual losses, usually do not admit a simple form.

Our package contains a dataset simulator, which helps with analysing the distribution of aggregated loss. Given a portfolio of policyholders X and a model specification alpha and comp_dist, the simulator will return one set of random realisation of losses for each policyholder.

With multiple calls to sim_dataset, an approximation of the aggregated loss can be obtained. This has applied in section 4.4 to calculate premium based on portfolio-level prinmium principles.

4.7. Model visualisation

After fitting and choosing an appropriate LRMoE model, the user can visualise it with in-package plotting functions, or create more customised plots using generic simulation functions combined with plotting utilities or other dedicated packages such as Plots.jl and StatsPlots.jl. In this subsection, we will use the six-component llllll model for demonstration.

Latent Class Probabilities

The logit regression in LRMoE assigns each policyholder into latent risk classes based on covariates. Given a fitted mode and a vector of covariates, the probability of latent classes can be computed and visualised using the built-in predict_class_prior and plot_class_prob functions.

Figure 2.

Prior and posterior latent class probabilities for selected policyholders. Component 6 (Lognormal(6.58, 1.95)) corresponds to the tail, while Component 3 (Lognormal(6.95, 1.05)) corresponds to the largest spike in the dataset.

Consider the policyholders in Table 9. The predict_class_prior function will return both latent class probabilities (prob), as well as the most likely class (max_prob_idx). The following sample code calculates their latent class probabilties, where each row represents a policyholder and each column represents a latent class. The prior probabilities are plotted in Figure 2 based solely on covariates information, there is not a large difference between these policyholders, and all are most likely coming from the first latent class.

For policyholders with known claim history, it may be more informative to consider the posterior latent class probabilities by calling corresponding functions for posterior probabilities. The posterior probability of latent class j is given by $\texttt{P}\{ Z_{ij} = 1 | \hat{{\boldsymbol{\Phi}}}, {\textit{\textbf{X}}}, {\textit{\textbf{Y}}} \}$ for fitted parameters $\hat{{\boldsymbol{\Phi}}}$ which can be computed analogous to equation (6). Readers may also refer to section 6.3.2 of Fung et al. (Reference Fung, Badescu and Lin2019a) for more details.

Consider again the same policyholders in Table 9. The code to calculate posterior probabilities is similar as above. The posterior probabilities are also plotted in Figure 2. Given that Policyholders B and C have non-zero claims, a lot of latent class probability is shifted to Classes 3 and 6, which correspond to the middle and tail parts of the loss distribution (see also the largest spike in Figure 3). Meanwhile, Policyholder A has no claim, resulting in little change to the latent class probabilities since all components have very similar zero inflation.

Figure 3.

Overall goodness of fit of positive claims in the French auto insurance dataset.

The posterior probabilities are also helpful for adjusting premium rates. In Table 10, we contrast premium calculation based on both prior and posterior probabilities. It is clear that Policyholder A is rewarded by a decrease in premium for having no claim history, while B and C have significant increase in premium rates.

Overall Goodness of Fit

The overall goodness of fit can be examined by contrasting the empirical histogram against fitted density curve and the Q–Q plot of empirical and fitted quantiles. These can be produced with the sim_dataset function included in our package, combined with basic plotting functions in julia. The corresponding plots are shown in Figure 3.

Covariate Influence

The partial dependence plot is commonly used in machine learning to investigate the influence of a particular covariate on the response, assuming independence amongst covariates (Friedman Reference Friedman2001). For example, the marginal effect of a covariate on the mean claim amount can be obtained using the following steps:

1. 1. Fix the covariate at a particular value (say, car brand = “F”) for all policyholders, while other covariates remain unchanged;

2. 2. Use predict_mean_prior function to compute mean response values by policyholder (see section 4.4);

3. 3. Compute the grand mean of all values in step (2), which averages out the effect of other covariates and yields the mean response value when the car brand is of type “F”;

4. 4. Repeat steps (1)–(3) for a range of values of the same covariate of interest, for example, varying car brand from “F” to “VAS”.

The procedure above can be generalised to continuous covariates and other quantities of interest, such as the quantile of the response variable. The covariate influence plot graphically illustrates how the characteristics of a policyholder are associated with a particular measure of the response variable. For example, plotting the mean (or VaR/CTE) of the response variable against car brand reflects how a policyholder’s overall riskiness (or tail risk) is related to the car brand.

Figure 4 shows the influence of covariates Brand, Region and Car Age, which may be interpreted as follows. For car brand, Mercedes, Chrysler or BMW (MCB), Opel, General Motors or Ford (OGF) and Volkswagen, Audi, Skoda or Seat (VAS) are generally riskier. Also, compared with other regions, policies issued in regions IF (Ile-de-France) and L (Limousin) may be considered riskier. In addition, older cars are generally associated with lower risks.