Hostname: page-component-586b7cd67f-2brh9 Total loading time: 0 Render date: 2024-11-23T19:17:12.591Z Has data issue: false hasContentIssue false

GEMAct: a Python package for non-life (re)insurance modeling

Published online by Cambridge University Press:  14 February 2024

Gabriele Pittarello*
Affiliation:
Università ‘La Sapienza’, Rome, Italy
Edoardo Luini
Affiliation:
Università Cattolica del Sacro Cuore, Milan, Italy
Manfred Marvin Marchione
Affiliation:
Università ‘La Sapienza’, Rome, Italy
*
Corresponding author: Gabriele Pittarello; Email: [email protected]
Rights & Permissions [Opens in a new window]

Abstract

This paper introduces gemact, a Python package for actuarial modeling based on the collective risk model. The library supports applications to risk costing and risk transfer, loss aggregation, and loss reserving. We add new probability distributions to those available in scipy, including the (a, b, 0) and (a, b, 1) discrete distributions, copulas of the Archimedean family, the Gaussian, the Student t and the Fundamental copulas. We provide an implementation of the AEP algorithm for calculating the cumulative distribution function of the sum of dependent, nonnegative random variables, given their dependency structure specified with a copula. The theoretical framework is introduced at the beginning of each section to give the reader with a sufficient understanding of the underlying actuarial models.

Type
Actuarial Software
Creative Commons
Creative Common License - CCCreative Common License - BY
This is an Open Access article, distributed under the terms of the Creative Commons Attribution licence (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted re-use, distribution and reproduction, provided the original article is properly cited.
Copyright
© The Author(s), 2024. Published by Cambridge University Press on behalf of Institute and Faculty of Actuaries

1. Introduction

In non-life insurance, the accurate representation and quantification of future losses is a foundational task, central to several areas ranging from pricing and reserving to risk management. Indeed, the actuarial literature is rich in models that are relevant in such applications. Among those, the collective risk model has been widely studied as it is mathematically tractable, it requires little and general information, and it can be efficiently implemented (Embrechts & Frei, Reference Embrechts and Frei2009; Klugman et al., Reference Klugman, Panjer and Willmot2012; Parodi, Reference Parodi2014). In particular, by knowing the frequency and severity distributions of the losses, it is possible to compute the distribution of the aggregate (total) losses. The collective risk model represents the common thread of this work, and we developed gemact to provide a collection of tools for (re)insurance modeling under a unified formal framework.

After a brief discussion on how to install the software in Section 2, we introduce the statistical framework of the collective risk model in Section 3. There, we define an aggregate loss distribution as a random sum of i.i.d. random variables, which can be computed using the recursive formula (Panjer, Reference Panjer1981), the discrete Fourier transform (DFT) (Bühlmann, Reference Bühlmann1984; Grübel & Hermesmeier, Reference Grübel and Hermesmeier1999; Wang, 1998), and a Monte Carlo (MC) simulation approach (Klugman et al., Reference Klugman, Panjer and Willmot2012, p. 467). Once the aggregate loss distribution is available, its expected value can be used for costing purposes. In this respect, the package supports typical coverage modifiers like (individual and aggregate) deductibles, limits, and reinstatements (see Sundt, Reference Sundt1990). Also, we consider different methods for the discretization of continuous distributions (Gerber, Reference Gerber1982).

Often, it is necessary to model the sum of a fixed number of dependent random variables. In order to do so, in Section 4, we introduce the AEP algorithm (Arbenz et al., Reference Arbenz, Embrechts and Puccetti2011) and a MC simulation approach for evaluating the cumulative distribution function of a sum of random variables with a given dependency structure. The dependency structure can be specified with the copulas we implemented. These are listed in Section A and include copulas of the Archimedean family, the Gaussian, the Student t, and the Fundamental copulas (Nelsen, Reference Nelsen2007).

Lastly, assuming a collective risk model holds for the cells of a loss development triangle, it is possible to define the stochastic claims reserving model in Ricotta & Clemente (Reference Ricotta and Clemente2016), Clemente et al. (Reference Clemente, Savelli and Zappa2019). In this case, the user obtains information on the frequency and severity parameters of the cells from the Fisher–Lange method (Fisher & Lange, Reference Fisher and Lange1973). Both these approaches are described in Section 5.

1.1 Context, scope, and contributions

In the recent years, programming languages and statistical computing environments, such as Python (Van Rossum & Drake, Reference Van Rossum and Drake2009) and R (R Core Team, 2017), have become increasingly popular (Ozgur et al., Reference Ozgur, Colliau, Rogers and Hughes2022). Currently, coding skills form part of the body of knowledge of actuaries and actuarial science researchers. In R, an extensive implementation for aggregate loss modeling based on the collective risk theory is the actuar package (Dutang et al., Reference Dutang, Goulet and Pigeon2008, Reference Dutang, Goulet and Langevin2022). An available library in Python is aggregate, which implements the computation of compound probability distributions via fast Fourier transform (FFT) convolution algorithm (Mildenhall, Reference Mildenhall2022). This package employs a specific grammar for the user to define insurance policy declarations and distribution calculation features. Direct access to its objects and their components is also possible.

With regard to claims reserving, chainladder offers in Python standard aggregate reserving techniques, like deterministic and stochastic chain-ladder methods, the Bornhuetter–Ferguson model, and the Cape Cod model (Bogaardt, Reference Bogaardt2022). This package is available in R and Python (Gesmann et al., Reference Gesmann, Murphy, Zhang, Carrato, Wüthrich, Concina and Dal Moro2022). Furthermore, apc provides the family of age-period-cohort approaches for reserving. This package is also available in both the above-mentioned programming languages (Nielsen, Reference Nielsen2015).

When it comes to the topic of dependence modeling via copulas, in Python one can use the copulas and copulae packages (Bok, Reference Bok2022; Lab, Reference Lab2022). Similarly, copula features in R are implemented in copula; see the package and its extensions in Jun Yan (Reference Yan2007), Kojadinovic & Yan (Reference Kojadinovic and Yan2010), and Hofert & Mächler (Reference Hofert and Mächler2011).

In this manuscript, we present an open-source Python package that extends the existing software tools available to the actuarial community. Our work is primarily aimed at the academic audience, who can benefit from our implementation for research and teaching purposes. Nonetheless, our package can also support non-life actuarial professionals in prototypes modeling, benchmarking and comparative analyses, and ad hoc business studies.

From the perspective of the package design, gemact adopts an explicit, direct, and integrated object-oriented programming (OOP) paradigm. In summary, our goal is to provide:

  • A computational software environment that gives users control over mathematical aspects and actuarial features, enabling the creation of models tailored to specific needs and requirements.

  • An object-oriented system whose elements (i.e., objects, methods, and attributes) can be accessed and managed via our API in such a way as to be interactive and suitable for users familiar with OOP designs and with the underlying modeling framework.

  • A collection of extensible libraries to make gemact survive over time. Our package is designed in an attempt to be easily extended or integrated with new functionalities and modules and to respect the attributes that qualify extensible software (Johansson & Löfgren, Reference Johansson and Löfgren2009). Namely, a modification to the functionalities should involve the least number of changes to the least number of possible elements (modifiability, Bass et al., Reference Bass, Clements and Kazman2003, p. 137), the addition of new requirements should not raise new errors (maintainability, Sommerville, Reference Sommerville2011, p. 25), and the system should be able to expand in a chosen dimension without major modifications to its architecture (scalability, Bondi, 2000).

From the perspective of actuarial advancements and developments, gemact provides an implementation to established algorithms and methodologies. In particular, our package:

2. Installation

The production version of the package is available on the Python Package Index (PyPi). Users can install gemact via pip, using the following command in the operating system command line interface.

pip install gemact==1.2.1

Examples on how to get started and utilize objects and functionalities of our package will be shown below. This work refers to production version 1.2.1.

Furthermore, the developer version of gemact can be found on GitHub at:

https://github.com/gpitt71/gemact-code

Additional resources on our project, including installation guidelines, API reference, technical documentations, and illustrative examples, can be found at:

https://gem-analytics.github.io/gemact/

3. Loss model

Within the framework of the collective risk model (Embrechts & Frei, Reference Embrechts and Frei2009), all random variables are defined on some fixed probability space $(\Omega, \mathcal{F}, P)$ . Let

  • $N$ be a random variable taking values in $\mathbb{N}_0$ representing the claim frequency.

  • $\left \{ Z_i\right \}_{i \in \mathbb{N}}$ be a sequence of i.i.d nonnegative random variables independent of $N$ ; $Z$ is the random variable representing the individual (claim) loss.

The aggregate loss $X$ , also referred to as aggregate claim cost, is

(1) \begin{equation} X=\sum _{i=1}^{N} Z_i, \end{equation}

with $\sum _{i=1}^{0} Z_i=0$ . Details on the distribution functions of $N$ , $Z$ , and $X$ are discussed in the next sections. Equation (1) is often referred to as the frequency–severity loss model representation. This can encompass common coverage modifiers present in (re)insurance contracts (Parodi, Reference Parodi2014, p. 50). More specifically, let us consider:

  • For $a \in [0, 1]$ , the function $Q_a$ apportioning the aggregate loss amount:

    (2) \begin{equation} Q_a (X)= a X. \end{equation}
  • For $c,d \geq 0$ , the function $L_{c, d}$ applied to the individual claim loss:

    (3) \begin{equation} L_{c, d} (Z_i) = \min \left \{\max \left \{0, Z_i-d\right \}, c\right \}. \end{equation}
    Herein, for each and every loss, the excess to a deductible $d$ (sometimes referred to as priority) is considered up to a cover or limit $c$ . In line with Albrecher et al. (Reference Albrecher, Beirlant and Teugels2017, p. 34), we denote $[d, d+c]$ as layer. An analogous notation is found in Ladoucette & Teugels (Reference Ladoucette and Teugels2006) and Parodi (Reference Parodi2014, p. 46). Similarly to the individual loss $Z_i$ , Equation (3) can be applied to the aggregate loss $X$ .

Computing the aggregate loss distribution is relevant for several actuarial applications (Parodi, Reference Parodi2014, p. 93). The gemact package provides various methods for calculating the distribution of the loss model in Equation (1) that allow the transformations of Equations (2) and (3) and their combinations to be included.

3.1 Risk costing

In this section, we describe an application of the collective risk model of Equation (1). The expected value of the aggregate loss of a portfolio constitutes the building block of an insurance tariff. This expected amount is called pure premium or loss cost, and its calculation is referred as risk costing (Parodi, Reference Parodi2014, p. 282). Insurers frequently cede parts of their losses to reinsurers, and risk costing takes this transfers into account. Listed below are some examples of basic reinsurance contracts whose pure premium can be computed using gemact.

  • The Quota Share (QS), where a share $a$ of the aggregate loss ceded to the reinsurance (along with the respective premium) and the remaining part is retained:

    (4) \begin{equation} \text{P}^{QS} = \mathbb{E}\left [ Q_a \left ( X \right)\right]. \end{equation}
  • The Excess-of-loss (XL), where the insurer cedes to the reinsurer each and every loss exceeding a deductible $d$ , up to an agreed limit or cover $c$ , with $c,d \geq 0$ :

    (5) \begin{equation} \text{P}^{XL} = \mathbb{E}\left [ \sum _{i=1}^{N} L_{c,d} (Z_i) \right]. \end{equation}
  • The Stop Loss (SL), where the reinsurer covers the aggregate loss exceedance of a (aggregate) deductible $v$ , up to a (aggregate) limit or cover $u$ , with $u,v \geq 0$ :

    (6) \begin{equation} \text{P}^{SL} = \mathbb{E}\left [ L_{u, v} (X) \right]. \end{equation}

The model introduced by Equation (1) and implemented in gemact can be used for costing contracts like the XL with Reinstatements (RS) in Sundt (Reference Sundt1990). Assuming the aggregate cover $u$ is equal to $ (K + 1) c$ , with $K \in \mathbb{Z}^+$ :

(7) \begin{equation} \text{P}^{RS} = \frac{\mathbb{E}\left [ L_{u, v} (X) \right]}{1+\frac{1}{c} \sum _{k=1}^K l_k \mathbb{E}\left [ L_{c, (k-1)c+v}(X) \right]}, \end{equation}

where $K$ is the number of reinstatement layers and $l_k \in [0, 1]$ is the reinstatement premium percentage, with $k=1, \ldots, K$ . When $l_k = 0$ , the $k$ -th resinstatement is said to be free. In detail, the logic we implemented implies that whenever a layer is used the cedent pays a reinstatement premium, that is, $l_k \text{P}^{RS}$ , and the cover $c$ is thus reinstated. The reinstatement premium will usually be paid in proportion to the amount that needs to be reinstated (Parodi, Reference Parodi2014, p. 52). In practice, the reinstatement premium percentage $l_k$ is a contractual element, given a priori as a percentage of the premium paid for the initial layer. In fact, in gemact $l_k$ is provided by the user. The mathematics behind the derivation of $\text{P}^{RS}$ is beyond the scope of this manuscript; the interested reader can refer to Sundt (Reference Sundt1990), Parodi (Reference Parodi2014, p. 325), and Antal (Reference Antal2009, p. 57).

3.2 Computational methods for the aggregate loss distribution

The cumulative distribution function (cdf) of the aggregate loss in Equation (1) is

(8) \begin{align} F_X(x)=P[X \leq x]=\sum _{k=0}^{\infty } p_k F_Z^{* k}(x) \end{align}

where $p_k=P[N=k]$ , $F_Z(x)=P\left [ Z \leq x \right]$ and $F_Z^{* k}(x)=P\left [ Z_1 + \ldots + Z_k \leq x \right]$ .

Moreover, the characteristic function of the aggregate loss $\phi _{X}\;:\; \mathbb{R} \rightarrow \mathbb{C}$ can be expressed in the form:

(9) \begin{equation} \phi _{X}(t)=\mathcal{P}_{N}\left (\phi _{Z}(t)\right), \end{equation}

where $\mathcal{P}_{N}(t)=\mathrm{E}\left [t^N\right]$ is the probability generating function of the frequency $N$ and $\phi _{Z}(t)$ is the characteristic function of the severity $Z$ (Klugman et al., Reference Klugman, Panjer and Willmot2012, p. 153).

The distribution in Equation (8), except in a few cases, cannot be computed analytically, and its direct calculation is numerically expensive (Parodi, Reference Parodi2014, p. 239). For this reason, different approaches have been analyzed to approximate the distribution function of the aggregate loss, including parametric and numerical quasi-exact methods (for a detailed treatment refer to Shevchenko, Reference Shevchenko2010). Among the latter, gemact implements MC simulation (Klugman et al., Reference Klugman, Panjer and Willmot2012, p. 467), DFT (Bühlmann, Reference Bühlmann1984; Grübel & Hermesmeier, Reference Grübel and Hermesmeier1999; Wang, 1998), and the so-called recursive formula (Panjer, Reference Panjer1981). A brief comparison of accuracy, flexibility, and speed of these methods can be found in Parodi (Reference Parodi2014, p. 260) and Wüthrich (Reference Wüthrich2023, p. 127). This section details these last two computational methods based on discrete mathematics to approximate the aggregate loss distribution.

Henceforth, let us consider, for $j=0,1,2,\ldots,m-1$ and $h\gt 0$ , an arithmetic severity distribution with probability sequence:

\begin{align*} \textbf{\{f\}}=\{f_0, f_1, \ldots, f_{m-1}\}, \end{align*}

where $f_j=P[Z=j\cdot h]$ . The discrete version of Equation (8) becomes

\begin{equation*} g_s=\sum _{k=0}^{\infty } p_k f_s^{* k}, \end{equation*}

where $g_s = P[X=s]$ and

\begin{equation*} f_s^{* j}\;:\!=\; \begin {cases}1 & \text { if } k=0 \text { and } s=0 \\[5pt] 0 & \text { if } k=0 \text { and } s \in \mathbb {N} \\[5pt] \sum _{i=0}^s f_{s-i}^{*(k-1)} f_i & \text { if } k \gt 0.\end {cases} \end{equation*}

3.2.1 Discrete fourier transform

The DFT of the severity $\textbf{\{f\}}$ is, for $k=0,\ldots,m-1$ , the sequence

\begin{equation*} \{\hat{\textbf{f}}\}= \{\hat{f}_0, \hat {f}_1, \ldots, \hat{f}_{m-1} \},\end{equation*}

where

(10) \begin{equation} \widehat{f}_k=\sum _{j=0}^{m-1}f_j e^{\frac{2\pi i k j}{m}}. \end{equation}

The original sequence can be reconstructed with the inverse DFT:

\begin{equation*}f_j=\frac {1}{m} \sum _{k=0}^{m-1} \widehat {f_k} e^{-\frac {2 \pi i k j}{ m}}.\end{equation*}

The sequence of probabilities $\textbf{\{g\}}=\{g_0, g_1, \ldots, g_{m-1}\}$ can be approximated taking the inverse DFT of

(11) \begin{equation} \{ \hat{\textbf{g}}\}\;:\!=\; \mathcal{P}_{N}\left (\{\hat{\textbf{f}}\}\right). \end{equation}

The original sequence can be computed efficiently with a FFT algorithm, when $m$ is a power of $2$ (Embrechts & Frei, Reference Embrechts and Frei2009).

3.2.2 Recursive formula

Assume that the frequency distribution belongs to the $(a, b, 0)$ class, that is, for $k \geq 1$ and $a, b\in \mathbb{R}$ :

(12) \begin{equation} p_k=\left (a+\frac{b}{k}\right)p_{k-1}. \end{equation}

Here, $p_0$ is an additional parameter of the distribution (Klugman et al., Reference Klugman, Panjer and Willmot2012, p. 505). The $(a, b, 0)$ class can be generalized to the $(a, b, 1)$ class assuming that the recursion in Equation (12) holds for $k=2,3,\ldots$

The recursive formula was developed to compute the distribution of the aggregate loss when the frequency distribution belongs to the $(a,b,0)$ or the $(a,b,1)$ class. The sequence of probabilities $\textbf{{g}}$ can obtained recursively using the following formula:

(13) \begin{equation} g_s=\frac{\left [p_1-(a+b) p_0\right] f_s+\sum _{j=1}^s(a+b j/ s) f_j g_{s-j}}{1-a f_0}, \end{equation}

with $1 \leq s \leq m-1$ and given the initial condition $g_0=\mathcal{P}_{N}\left (f_0\right)$ .

3.3 Severity discretization

The calculation of the aggregate loss with DFT or with the recursive formula requires an arithmetic severity distribution (Embrechts & Frei, Reference Embrechts and Frei2009). Conversely, the severity distribution in Equation (1) is often calibrated on a continuous support. In general, one needs to choose a discretization approach to arithmetize the original distribution. This section illustrates the methods for the discretization of a continuous severity distribution available in the gemact package.

Let $F_z\;:\; \mathbb{R^+} \rightarrow [0, 1]$ be the cdf of the distribution to be discretized. For a bandwidth, or discretization step, $h\gt 0$ and an integer $m$ , a probability $f_j$ is assigned to each point $h j$ , with $j= 0, \ldots,m-1$ . Four alternative methods for determining the values for $f_j$ are implemented in gemact.

  1. 1. The method of mass dispersal:

    \begin{align*} f_j= \begin{cases}F_Z \left ( \frac{h}{2} \right) & j=0 \\[5pt] F_Z\left ( hj+ \frac{h}{2} \right)-F_Z\left (hj-\frac{h}{2}\right) & j=1,\ldots,m-2\\[5pt] 1-F_Z\left (hj-\frac{h}{2}\right) & j=m-1. \end{cases} \end{align*}
  2. 2. The method of the upper discretization:

    \begin{align*} f_j= \begin{cases}F_Z\left ( hj+ h \right)-F_Z\left (hj\right) & j=0,\ldots,m-2\\[5pt] 1-F_Z\left (hj\right) & j=m-1. \end{cases} \end{align*}
  3. 3. The method of the lower discretization:

    \begin{align*} f_j= \begin{cases}F_Z \left ( 0 \right) & j=0 \\[5pt] F_Z\left ( hj\right)-F_Z\left (hj-h\right) & j=1,\ldots,m-1.\\[5pt] \end{cases} \end{align*}
  4. 4. The method of local moment matching:

    \begin{align*} f_j= \begin{cases}1-\frac{\mathbb{E}[Z\wedge h] }{h} & j=0 \\[5pt] \frac{2 \mathbb{E}[Z \wedge h j]-\mathbb{E}[Z \wedge h (j-1)]-\mathbb{E}[Z \wedge h (j+1)]}{h} & j=1,\ldots,m-1\end{cases} \end{align*}
    where $\mathbb{E}[Z\wedge h] = \int _{-\infty }^h t \; dF_z(t)+h[1-F_z(h)]$ .

Fig. 1 illustrates and graphically contrasts the four different discretization techniques.

Figure 1 Illustration of the discretization methods applied to a gamma( $\texttt{a}=5$ ) severity. The graphs compare the original cdf (blue line) and the discretized (red line) cdf for mass dispersal (top left), upper discretization (top right), lower discretization (bottom left), and local moment matching (bottom right) methods. No coverage modifiers are present.

The default discretization method in gemact is the method of the mass dispersal, in which each $h j$ point is assigned with the probability mass of the $h$ -span interval containing it (Fig. 1 top left). The upper discretization and lower discretization methods generate, respectively, pointwise upper and lower bounds to the true cdf. Hence, these can be used to provide a range where the original $F_z$ is contained (see Fig. 1, top right and bottom left graphs). The method of local moment matching (Fig. 1 bottom right) allows the moments of the original distribution $F_z$ to be preserved in the arithmetic distribution. A more general definition of this approach can be found in Gerber (Reference Gerber1982). We limited this approach to the first moment as, for higher moments, it is not well defined and it can potentially lead to a negative probability mass on certain lattice points (Embrechts & Frei, Reference Embrechts and Frei2009). The method of local moment matching shall be preferred for large bandwidths. However, considering that there is no established analytical procedure to determine the optimal step, we remark that the choice of the discretization step is an educated guess and it really depends on the problem at hand. In general, $h$ should be chosen such that it is neither too small nor too large relative to the severity losses. In the first case, the $hj$ points are not sufficient to capture the full range of the loss amount and the probability in the tail of the distribution exceeding the last discretization node $h(m-1)$ is too large. In the second case, the granularity of the severity distribution is not sufficient, and small losses are over-approximated. Additional rules of thumb and guidelines for the choice of discretization parameters can be found in Parodi (Reference Parodi2014, p. 248). For example, one option is to perform the calculation with decreasing values of $h$ and check, graphically or according to a predefined criterion, whether the aggregate distribution changes substantially (Embrechts & Frei, Reference Embrechts and Frei2009). The reader should refer to Klugman et al. (Reference Klugman, Panjer and Willmot2012, p. 179) and Embrechts & Frei (Reference Embrechts and Frei2009) for a more detailed treatment and discussion of discretization methods.

The above-mentioned discretization methods are modified accordingly to reflect the cases where the transformation of Equation (3) is applied to the severity (Klugman et al., Reference Klugman, Panjer and Willmot2012, p. 517). Below, we illustrate how to perform severity discretization in gemact.

Once a continuous distribution is selected from those supported in our package (see Section A), the severity distribution is defined via the Severity class.

>>> from gemact.lossmodel import Severity

>>> severity = Severity(dist='gamma', par={'a': 5})

The dist argument contains the name of the distribution, and the par argument specifies, as a dictionary, the distribution parameters. In the latter, each item key–value pair represents a distribution parameter name and its value. Refer to distributions module for a list of the distribution names and their parameter specifications.

The discretize method of the Severity class produces the discrete severity probability sequence according to the approaches described above. Below, we provide an example for mass dispersal.

>>> massdispersal = severity.discretize(

discr_method='massdispersal',

n_discr_nodes = 50000,

discr_step = .01,

deductible = 0

)

In order to perform the discretization, the following arguments are needed:

  • The chosen discretization method via discr_method.

  • The number of nodes ( $m$ ) set in the n_discr_nodes argument.

  • The severity discretization step ( $h$ ) is in the discr_step argument.

  • If necessary, a deductible specifying where the discretization begins. The default value is zero.

After the discretization is achieved, the mean of the discretized distribution can be calculated.

>>> import numpy as np

>>> discrete_mean = np.sum(massdispersal['nodes'] * massdispersal['fj'])

>>> print('Mean of the discretised distribution:', discrete_mean)

Mean of the discretised distribution: 5.000000000000079

Additionally, the arithmetic distribution obtained via the severity discretization can be visually examined using the plot_discretized_severity_cdf. This method is based on the pyplot interface to matplotlib (Hunter, Reference Hunter2007). Hence, plot_discretized_severity_cdf can be used together with pyplot functions and can receive pyplot.plot arguments to change its output. In the following code blocks, we adopt the plot_discretized_severity_cdf method in conjunction with the plot function from matplotlib.pyplot to compare the cdf of a gamma distribution with mean and variance equal to $5$ , with the arithmetic distribution obtained with the method of mass dispersal above. We first import the gamma distribution from the distributions module and compute the true cdf.

>>> from gemact import distributions

>>> dist = distributions.Gamma(a = 5)

>>> nodes = np.arange(0, 20)

>>> true_cdf = dist.cdf(nodes)

Next, we plot the discrete severity using the plot_discr_sev_cdf method.

>>> import matplotlib.pyplot as plt

>>> severity.plot_discr_sev_cdf(

discr_method='massdispersal',

n_discr_nodes = 20,

discr_step = 1,

deductible = 0,

color='#a71429'

)

>>> plt.plot(nodes, true_cdf, color='#4169E1')

>>> plt.title('Method of Mass Dispersal')

>>> plt.xlabel('z')

>>> plt.show()

The arguments discr_method, n_discr_nodes, discr_step, and deductible can be used in the same manner as those described in the discretize method. The argument color is from the matplotlib.pyplot.plot method. The methods title and xlabel were also exported from matplotlib.pyplot.plot to add custom labels for the title and the x-axis (Hunter, Reference Hunter2007).

The output of the previous code is shown in the top left graph of Fig. 1. To obtain the other graphs, simply set discr_method to the desired approach, that is, ‘upper_discretisation,’ ‘lower_discretisation’, or ‘localmoments.’

3.4 Supported distributions

The gemact package makes for the first time the $(a, b, 0)$ and $(a, b, 1)$ distribution classes (Klugman et al., Reference Klugman, Panjer and Willmot2012, p. 81) available in Python. In the following code block, we show how to use our implementation of the zero-truncated Poisson from the distributions module.

>>> ztpois = distributions.ZTPoisson(mu = 2)

Each distribution supported in gemact has various methods and can be used in a similar fashion to any scipy distribution. Next, we show how to compute the approximated mean via MC simulation, with the random generator method for the ZTPoisson class.

>>> random_variates = ztpois.rvs(10**5, random_state = 1)

>>> print('Simulated Mean: ', np.mean(random_variates))

Simulated Mean: 2.3095

>>> print('Exact Mean: ', ztpois.mean())

Exact Mean: 2.3130352854993315

Furthermore, supported copula functions can be accessed via the copulas module. Below, we compute the cdf of a two-dimensional Gumbel copula.

>>> from gemact import copulas

>>> gumbel_copula = copulas.GumbelCopula(par = 1.2, dim = 2)

>>> values = np.array([ [.5, .5]])

>>> print(' Gumbel copula cdf: ', gumbel_copula.cdf(values)[0])

Gumbel copula cdf: 0.2908208406483879

In the above example, it is noted that the copula parameter and dimension are defined by means of the par and the dim arguments, respectively. The argument of the cdf method must be a numpy array whose dimensions meet the following requirements. Its first component is the number of points where the function shall be evaluated and its second component equals the copula dimension (values.shape of the example is in fact (1,2)).

The complete list of the distributions and copulas supported by gemact is available in Section A. We remark that the implementation of some distributions is available in both gemact and scipy.stats. However, the objects of the distributions module include additional methods that are specific to their use in actuarial science. Examples are lev and censored_moment methods, which allow the calculation of the limited expected value and censored moments of continuous distributions. Furthermore, the choice of providing a Severity class is in order to have a dedicated object that includes functionalities relevant only for the calculation of a loss model and not for distribution modeling in general. An example of this is the discretize method. A similar reasoning applies to the Frequency class.

3.5 Illustration lossmodel

The following are examples of how to get started and use lossmodel module and its classes for costing purposes. As an overview, Fig. 2 schematizes the class diagram of the lossmodel module, highlighting its structure and the dependencies of the LossModel class.

The Frequency and the Severity classes represent, respectively, the frequency and the severity components of a loss model. In these, dist identifies the name of the distribution and par specifies its parameters as a dictionary, in which each item key–value pair corresponds to a distribution parameter name and value. Please refer to distributions module for the full list of the distribution names and their parameter specifications. The code block below shows how to initiate a frequency model.

Figure 2 Class diagram of the lossmodel module. A rectangle represents a class; an arrow connecting two classes indicates that the target class employs the origin class as an attribute. In this case, a LossModel object entails Frequency, Severity, and PolicyStructure class instances. These correspond to the frequency model, the severity model, and the policy structure, respectively. The latter, in particular, is in turn specified via one or more Layer objects, which include coverage modifiers of each separate policy component.

>>> from gemact.lossmodel import Frequency

>>> frequency = Frequency(

dist='poisson',

par={'mu': 4},

threshold = 0

)

In practice, losses are reported only above a certain threshold (the reporting threshold) and the frequency model can be estimated only above another, higher threshold called the analysis threshold (Parodi, Reference Parodi2014, p. 323). This can be specified in Frequency with the optional parameter threshold, whose default value is 0 (i.e., the analysis threshold equals the reporting threshold). Severity models in gemact always refer to the reporting threshold.

A loss model is defined and computed through the LossModel class. Specifically, Frequency and Severity objects are assigned to frequency and severity arguments of LossModel to set the parametric assumptions of the frequency and the severity components. Below, we use the severity object we instanced in Subsection 3.3.

>>> from gemact.lossmodel import LossModel

>>> lm_mc = LossModel(

frequency=frequency,

severity=severity,

aggr_loss_dist_method='mc',

n_sim = 10**5,

random_state = 1

)

INFO:lossmodel|Approximating aggregate loss distribution via Monte Carlo simulation

INFO:lossmodel|MC simulation completed

In the previous example, in more detail, lm_mc object adopts the MC simulation for the calculation of the aggregate loss distribution. This approach is set via the aggr_loss_dist_method equal to ‘mc.’ The additional parameters required for the simulation are as follows:

  • the number of simulations n_sim,

  • the (pseudo)random number generator initializer random_state.

The cdf of the aggregate loss distribution can be displayed with the plot_dist_cdf method. Moreover, a recap of the computation specifications can be printed with the print_aggr_loss_specs method.

The aggregate loss mean, standard deviation, and skewness can be accessed with the mean, std, and skewness methods, respectively. The code below shows how to use these methods.

>>> lm_mc.mean(use_dist=True)

19.963155575580227

>>> lm_mc.mean(use_dist=False)

20.0

>>> lm_mc.coeff_variation(use_dist=True)

0.5496773171375182

>>> lm_mc.coeff_variation(use_dist=False)

0.5477225575051661

>>> lm_mc.skewness(use_dist=True)

0.6410913579225725

>>> lm_mc.skewness(use_dist=False)

0.6390096504226938

When the use_dist argument is set to True (default value), the quantity is derived from the approximated aggregate loss distribution (dist property). Conversely, when it is False, the calculation relies on the closed-form formulas of the moments of the aggregate loss random variable. These can be obtained directly from the closed-form moments of the frequency and the severity model transformed according to the coverage modifiers (Parodi, Reference Parodi2014, p. 322). This option is available for mean, std, var (i.e., the variance), coeff_variation (i.e., the coefficient of variation), and skewness methods. It should be noted that the calculation with use_dist=False is not viable when aggregate coverage modifiers are present. In such circumstance, the method must necessarily be based on the approximated aggregate loss distribution. In any situations, it is possible to get the moments of the approximated aggregate loss distribution via the moment method. The central and n arguments specify, respectively, whether the moment is central and the order of the moment.

>>> lm_mc.moment(central=False, n = 1)

19.963155575580227

Furthermore, for the aggregate loss distribution, the user can simulate random variates via the rvs method. The quantile and the cdf functions can be computed via the ppf and the cdf methods. Below is an example of the ppf method returning the 0.80- and 0.70-level quantiles.

>>> lm_mc.ppf(q=[0.80, 0.70])

array([28.81343738, 24.85497983])

The following code block shows the costing of a "20 excess 5" XL reinsurance contract. Coverage modifiers are set in a PolicyStructure object.

>>>from gemact.lossmodel import PolicyStructure, Layer

>>> policystructure = PolicyStructure(

layers=Layer(

cover = 20,

deductible = 5

))

More precisely, in the Layer class, the contract cover and the deductible are provided. Once the model assumptions are set and the policy structure is specified, the aggregate loss distribution can be computed.

>>> lm_XL = LossModel(

frequency=frequency,

severity=severity,

policystructure=policystructure,

aggr_loss_dist_method='fft',

sev_discr_method='massdispersal',

n_aggr_dist_nodes = 2**17

)

INFO:lossmodel|Approximating aggregate loss distribution via FFT

INFO:lossmodel|FFT completed

It can be noted that, in the previous code block, we determined the aggregate loss distribution with ‘fft’ as aggr_loss_dist_method. In such case, LossModel requires additional arguments for defining the computation process, namely:

  • The number of nodes of the aggregate loss distribution n_aggr_dist_nodes.

  • The method to discretize the severity distribution sev_discr_method. Above, we opted for the method of mass dispersal ("massdispersal").

  • The number of nodes of the discretized severity n_sev_discr_nodes (optional).

  • The discretization step sev_discr_step (optional). When a cover is present, gemact automatically adjusts the discretization step parameter to have the correct number of nodes in the transformed severity support.

The same arguments shall be specified while computing the aggregate loss distribution with the recursive formula, that is, aggr_loss_dist_method set to "recursive."

The costing specifications of a LossModel object can be accessed with the method print_costing_specs().

The previous output exhibits a summary of the contract structure (cover, deductible, aggregate cover, and aggregate deductible) and details about the costing results. It is noted that the default value of the share participation equals 1, that is, $a$ is equal to $1$ in Equation (2).

Similar to the previous example, user can access moments of the aggregate loss calculated from the approximated distribution and from the closed-from solution.

>>> lm_XL.mean(use_dist=True)

3.509346100359707

>>> lm_XL.mean(use_dist=False)

3.50934614394912

>>> lm_XL.coeff_variation(use_dist=True)

1.0001481880266856

>>> lm_XL.coeff_variation(use_dist=False)

1.0001481667319252

>>> lm_XL.skewness(use_dist=True)

1.3814094240544392

>>> lm_XL.skewness(use_dist=False)

1.3814094309741256

The next example illustrates the costing of an XL with RS. The PolicyStructure object is set as follows.

>>> policystructure_RS = PolicyStructure(

layers=Layer(

cover = 100,

deductible = 0,

aggr_deductible = 100,

reinst_percentage = 1,

n_reinst = 2

))

The relevant parameters are as follows:

  • the aggregate deductible parameter aggr_deductible,

  • the number of reinstatements n_reinst,

  • the reinstatement percentage reinst_percentage.

Below, we compute the pure premium of Equation (7), given the parametric assumptions on the frequency and the severity of the loss model.

>>> lm_RS = LossModel(

frequency=Frequency(

dist='poisson',

par={'mu': .5}

),

severity=Severity(

dist='pareto2',

par={'scale': 100, 'shape': 1.2}

),

policystructure = policystructure_RS,

aggr_loss_dist_method='fft',

sev_discr_method='massdispersal',

n_aggr_dist_nodes = 2**17

)

>>> print(Pure premium (RS):, lm_RS.pure_premium_dist[0])

Pure premium (RS): 4.319350355177216

A PolicyStructure object can handle multiple independent layers simultaneously. These can overlap and do not need to be contiguous. The length property indicates the number of layers. The code block below deals with three Layer objects, the first without aggregate coverage modifiers and (participation) share of 0.5, the second with reinstatements, and the third with an aggregate cover.

>>> policystructure=PolicyStructure(

layers=[

Layer(cover = 100, deductible = 100, share = 0.5),

Layer(cover = 200, deductible = 100, n_reinst = 2,

reinst_percentage = 0.6),

Layer(cover = 100, deductible = 100, aggr_cover = 200)

])

>>> lossmodel_multiple = LossModel(

frequency=Frequency(

dist='poisson',

par={'mu': .5}

),

severity=Severity(

dist='genpareto',

par={'loc': 0, 'scale': 83.34, 'c': 0.834}

),

policystructure=policystructure

)

WARNING:lossmodel|Aggregate loss distribution calculation is omitted as

aggr_loss_dist_method is missing

WARNING:lossmodel|Layer 2: costing is omitted as aggr_loss_dist_method is missing

WARNING:lossmodel|Layer 3: costing is omitted as aggr_loss_dist_method is missing

As outlined by the warning messages, since the instantiation of lossmodel_multiple lacks of aggr_loss_dist_method, the calculation of the aggregate loss distribution is omitted. Therefore, costing results are accessible solely for the first Layer, since this is the only one without aggregate coverage modifiers. This fact is reflected in pure_premium and pure_premium_dist properties, containing the premiums derived from the closed-form means and the approximated aggregate loss distribution means, respectively. The latter are indeed not available.

>>> lossmodel_multiple.pure_premium

[8.479087307840043, None, None]

>>> lossmodel_multiple.pure_premium_dist

[None, None, None]

Contrarily, once the aggregate loss distribution is determined (via dist_calculate method), all layer premiums in pure_premium_dist are available from the costing (costing method). As expected, pure_premium content remains unaffected.

>>> lossmodel_multiple.dist_calculate(

aggr_loss_dist_method='fft',

sev_discr_method='massdispersal',

n_aggr_dist_nodes = 2**17

)

INFO:lossmodel|Computation of layers started

INFO:lossmodel|Computing layer: 1

INFO:lossmodel|Approximating aggregate loss distribution via FFT

INFO:lossmodel|FFT completed

INFO:lossmodel|Computing layer: 2

INFO:lossmodel|Approximating aggregate loss distribution via FFT

INFO:lossmodel|FFT completed

INFO:lossmodel|Computing layer: 3

INFO:lossmodel|Approximating aggregate loss distribution via FFT

INFO:lossmodel|FFT completed

INFO:lossmodel|Computation of layers completed

>>> lossmodel_multiple.costing()

>>> lossmodel_multiple.pure_premium_dist

[8.479087307062226, 25.99131088702302, 16.88704720494799]

>>>lossmodel_multiple.pure_premium

[8.479087307840043, None, None]

At last, it is remarked that each Layer in PolicyStructure is associated with an index idx, starting from 0, based on the layer order of the instantiation of the PolicyStructure object. This is of help when the user needs to retrieve particular information and features, or to apply methods to one specific layer. All the methods that include the idx argument have 0 as default value, meaning that they are applied to the first (or only) layer unless otherwise specified. For example, the print_policy_layer_specs method produces a table of recap of the features of the layer indicated by the idx argument. Below idx equals 1, namely the second Layer in policystructure.

Likewise, the code block below returns the aggregate loss distribution mean of the third Layer, as idx is set to 2.

>>> lossmodel_multiple.mean(idx = 2)

16.88704720494799

3.6 Comparison of the methods for computing the aggregate loss distribution

In this section, we analyze accuracy and speed of the computation of the aggregate loss distribution using FFT, recursive formula (recursion), and MC simulation approaches, as the number of nodes, the discretization step, and the number of simulations vary.

For this purpose, a costing example was chosen such that the analytical solutions of the moments of the aggregate loss distributions are known and can be compared to those obtained from the approximated aggregate loss distribution. The values of the parameters of the severity and frequency models are taken from the illustration in Parodi (Reference Parodi2014, p. 262). Specifically, these and the policy structure specifications are as follows.

  • Severity: lognormal distribution with parameters shape $=1.3$ and scale $=36315.49$ , hence whose mean and standard deviation are $84541.68$ and $177728.30$ , respectively.

  • Frequency: Poisson distribution with parameter $\mu = 3$ , with analysis threshold $d$ . It belongs to the $(a, b, 0)$ family described in Subsection 3.2.2 with parameters $a=0$ , $b=3$ and $p_0=e^{-3}$ .

  • Policy structure: contract with deductible $d=10000$ .

In particular, the accuracy in the approximation of the aggregate loss distribution has been assessed using the relative error in the estimate of the mean, coefficient of variation (CoV), and skewness, with respect to their reference values (i.e., error = estimate/reference - 1). The latter are obtained using the following closed-form expressions (Bean, Reference Bean2000, p. 382):

\begin{align*} & \mathbb{E}\left [X\right] = \mu \mathbb{E}\left [ L_{d, \infty } (Z) \right],\\[5pt] & \text{CoV}\left [X\right] = \dfrac{ \mathbb{E}\left [ L_{d, \infty } (Z)^2 \right]^{1/2}}{\mu ^{1/2} \mathbb{E}\left [ L_{d, \infty } (Z) \right]},\\[5pt] & \text{Skewness}\left [X\right] = \dfrac{ \mathbb{E}\left [ L_{d, \infty } (Z)^3 \right]}{\mu ^{1/2} \mathbb{E}\left [ L_{d, \infty } (Z)^2 \right]^{3/2}}. \end{align*}

Their values for the example are in the top part of Table 1. The test of the speed of our implementation has been carried out by measuring the execution time of the approximation of the aggregate loss distribution function with the built-in timeit library. In line with best practice (Martelli et al., Reference Martelli, Ravenscroft and Ascher2005, Chapter 18), the observed minimum execution time of independent repetitions of the function call was adopted.

The results of the analysis are reported in Table 1. We considered for FFT and recursion the values 50, 100, 200, and 400 for the discretization step $h$ , and $2^{14}, 2^{16}, 2^{18}$ for the number of nodes. It can be noted that FFT and recursion produce similar figures in terms of accuracy, but the former is drastically faster. This is expected as FFT takes essentially $\mathcal{O}(m \log (m))$ operations, compared to the $\mathcal{O}(m^2)$ operations for recursion (Embrechts & Frei, Reference Embrechts and Frei2009). Furthermore, for this example, in both FFT and recursion, when $h$ increases, the error reduces. Finally, MC approach lies in between the two other alternatives, when it comes to computing time.

3.7 Comparison with aggregate FFT implementation

The aggregate package (Mildenhall, Reference Mildenhall2022) allows to compute the aggregate loss distribution using FFT. In this section, we compare our implementation with aggregate (version 0.9.3) implementation to show that the two provide similar results. We adopt the same underlying frequency and severity assumptions of Subsection 3.6 and contracts with different combinations of individual and aggregate coverage modifiers. In particular, we first consider no reinsurance, then an XL, with individual-only coverage modifiers, a SL, and finally an XL with individual and aggregate coverage modifiers (XL w/agg.). In line with the example in Parodi (Reference Parodi2014, p. 262), individual coverage modifiers are $c= 1 000 000$ and $d= 10 000$ , and aggregate coverage modifiers are $u= 1 000 000$ and $v= 50 000$ . The number of nodes $m$ is set to $2^{22}$ in all the calculations.

Table 1. Accuracy and speed of the approximation of the aggregate loss distribution using fast Fourier transform (FFT), the recursive formula (recursion), and the Monte Carlo (MC) simulation when varying the number of nodes ( $m$ ), the discretization step ( $h$ ), and number of simulations. The upper table contains the reference values obtained from the closed-form solutions. The lower table reports the execution times in second and the relative errors with respect to the reference values

The comparison of the speed of the two implementations has been carried out by means of the built-in timeit library. In particular, we measured the execution time of both the initialization of the main computational object and the calculation of the aggregate loss distribution in order to make the comparison consistent and adequate. In line with best practice (Martelli et al., Reference Martelli, Ravenscroft and Ascher2005, Chapter 18), the observed minimum execution time of independent repetitions of the function call was adopted.

As reported by Table 2, the two implementations generate consistent results; their estimates for mean, CoV, and skewness tend to coincide for all contracts and are close to the reference values when these are available. When it comes to computing time, gemact takes a similar time, just under one second, for all the examples considered. For the case without reinsurance and the XL, aggregate performs slightly better. Conversely, for SL and XL w/agg., gemact is more than twice as fast.

Table 2. Comparison for different contracts of aggregate and gemact implementation of the aggregate loss distribution computation via FFT. When there are no aggregate coverage modifiers, reference values are given in addition to estimated ones. For the XL and the XL w/agg. contracts, individual conditions are $c= 1 000 000$ and $d= 10 000$ ; for the SL and the XL w/agg. contracts, aggregate coverage modifiers are $u= 1 000 000$ and $v= 50 000$ . Execution times are expressed in seconds

4. Loss aggregation

In insurance and finance, the study of the sum of dependent random variables is a central topic. A notable example is risk management, where the distribution of the sum of certain risks needs to be approximated and analyzed for solvency purposes (Wilhelmy, Reference Wilhelmy2010). Another application is the pricing of financial and (re)insurance contracts where the payout depends on the aggregation of two or more dependent outcomes (see, e.g., Cummins et al., Reference Cummins, Lewis, Phillips and Froot1999; Wang, Reference Wang2013). In this section, in contrast with the collective risk theory in Section 3, we model the sum of a given number of random variables $d \gt 1$ that are neither independent nor necessarily identically distributed.

More specifically, consider now the random vector:

\begin{equation*}\left (X_1, \ldots, X_d\right)\;:\; \Omega \rightarrow \mathbb {R}^d,\end{equation*}

whose joint cdf

(14) \begin{equation} H\left (x_1,\ldots,x_d\right)=P\left [X_1\le x_1,\ldots,X_d\le x_d\right] \end{equation}

is known analytically or can be numerically evaluated in an efficient way. For a real threshold $s$ , the gemact package implements the AEP algorithm and a MC simulation approach to model:

(15) \begin{equation} P\left [X_1 + \ldots + X_d \leq s\right], \end{equation}

given a set of parametric assumptions on the one-dimensional marginals $X_1, \ldots, X_d$ and their copula.

More specifically, the AEP algorithm is designed to approximate Equation (14) through a geometric procedure, without relying on simulations or numerical integration of a density. In Section B, a brief description of the algorithm is given. For a complete mathematical treatment of the subject, the reader should refer to Arbenz et al. (Reference Arbenz, Embrechts and Puccetti2011).

4.1 Illustration lossaggregation

Below are some examples of how to use the LossAggregation class. This belongs to the lossaggregation module, whose class diagram is depicted in Fig. 3. The main class is LossAggregation, which is the computation object of the random variable sum. This depends on the classes Margins and Copula. Evidently, the former represents the marginal distributions and the latter describes the dependency structure, that is, the copula.

Figure 3 Class diagram of the lossaggregation module. A rectangle represents a class; an arrow connecting two classes indicates that the target class employs the origin class as an attribute. In this case, a LossAggregation object entails Margins and Copula class instances.

Consistent with the gemact framework, the specifications needed to instantiate Margins and Copula are akin to those of the Frequency and Severity classes in lossmodel. Copula objects are specified by:

  • dist: the copula distribution name as a string (str),

  • par: the parameters of the copula, as a dictionary.

Likewise, Margins objects are defined by:

  • dist: the list of marginal distribution names as strings (str),

  • par: the list of parameters of the marginal distributions, each list item is a dictionary.

Please refer to Table A.1 and Table A.2 of Section A for the complete list of the supported distributions and copulas.

>>> from gemact import LossAggregation, Copula, Margins

>>> lossaggregation = LossAggregation(

margins=Margins(

dist=['genpareto', 'lognormal'],

par=[{'loc': 0, 'scale': 1/.9, 'c': 1/.9}, {'loc': 0, 'scale': 10,

'shape': 1.5}],

),

copula=Copula(

dist='frank',

par={'par': 1.2, 'dim': 2}

),

n_sim = 500000,

random_state = 10,

n_iter = 8

)

Besides marginal and copula assumptions, the instantiation of LossAggregation accepts the arguments n_sim and random_state to set the number of simulation and the (pseudo)random number generator initializer of the MC simulation. If n_sim and the random_state are omitted, the execution is bypassed. Nonetheless, the user can perform it at a later point through the dist_calculate method. Also, n_iter controls the number of iterations of the AEP algorithm. This parameter is optional (default value is 7) and can be specified either at the creation of class or directly in methods that include the use of this approach.

In the following code block, we show how to calculate the cdf of the sum of a generalized Pareto and a lognormal-dependent random variables, using the AEP algorithm ("aep") and the MC simulation approach ("mc"). The underlying dependency structure is a Frank copula.

>>> s = 300 # arbitrary value

>>> p_aep = lossaggregation.cdf(x = s, method=aep)

>>> print('P(X1+X2 <= s) = ', p_aep)

P(X1+X2 <= s) = 0.9811620158197308

>>> p_mc = lossaggregation.cdf(x = s, method='mc')

>>> print('P(X1+X2 <= s) = ', p_mc)

P(X1+X2 <= s) = 0.98126

LossAggregation includes other functionalities like the survival function sf, the quantile function ppf, and the generator of random variates rvs. Furthermore, for the MC simulation approach, it is possible to derive empirical statistics and moments using methods such as for example moment, mean, var, skewness, and censored_moment. To conclude, the last code block illustrates the calculation of the quantile function via the ppf method.

>>> lossaggregation.ppf(q = p_aep, method=aep)

300.0000003207744

>>> lossaggregation.ppf(q = p_mc, method=mc)

299.9929982860278

4.2 Comparison of the methods for computing the cdf

In this section, we compared our implementation of the AEP algorithm with the alternative solution based on MC simulation in terms of speed and accuracy. Accuracy in the calculation of the cdf has been assessed by means of the relative error with respect to the reference value for a chosen set of quantiles (i.e., error = estimate/reference - 1). We replicate the experiment in Arbenz et al. (Reference Arbenz, Embrechts and Puccetti2011) and take the reference values that the authors computed in the original manuscript. The analysis considers four different Clayton–Pareto models, for $d=2,3,4,5$ , with the following parametric assumptions. In the two-dimensional case ( $d=2$ ), the tail parameters $\gamma$ of the marginal distributions are $0.9$ and $1.8$ ; the Clayton copula has parameter $\theta =1.2$ . In three dimensions ( $d=3$ ), the additional marginal has parameter $\gamma =2.6$ , and the copula has $\theta =0.4$ . For the four-dimensional ( $d=4$ ) and five-dimensional ( $d=5$ ) cases, the extra-marginal component has parameter equal to $3.3$ and $4$ , and the Clayton copula has parameter 0.2 and 0.3, respectively. The cdf has been evaluated at the quantiles $s=\{10^0, 10^2, 10^4, 10^6\}$ for d = 2 and d = 3, and $s=\{10^1, 10^2, 10^3, 10^4\}$ , when d = 4 and d = 5.

The test of the speed of our implementation has been carried out by measuring the execution time of the cdf function for a single quantile with the built-in timeit library. In line with best practice (Martelli et al., Reference Martelli, Ravenscroft and Ascher2005, Chapter 18), the observed minimum execution time of independent repetitions of the function call was adopted.

Table 3 shows the results of our comparison of the two methodologies for the calculation of the cdf.

Table 3. Accuracy and speed of the cdf calculation using the AEP algorithm and the Monte Carlo (MC) simulation approach for the sum of Pareto random variables coupled with a Clayton copula for different dimensions and quantiles. The upper table contains the reference values from Arbenz et al. (Reference Arbenz, Embrechts and Puccetti2011). The lower table reports the execution times in seconds and the relative errors with respect to the reference values. The time column on the left is about the execution times of gemact, while the time column on the right (labeled with a *) lists those of the original manuscript

It can be noted that our implementation of the AEP algorithm shows a high accuracy in the calculation of the cdf, for all quantiles and dimensions. Its precision also remains valid for the five-dimensional case. The figures are in line with the results of the original study. In general, in cases considered, the AEP algorithm is closer to the reference values than the MC simulation approach. Nevertheless, the latter shows contained errors whose order of magnitude is $10^{-2}$ at most, when the number of simulation is set to the lowest value. On the other hand, the AEP algorithm is outperformed by the MC simulation approach in terms of execution speed. The largest gaps have been observed, especially when the number of iterations of the AEP is higher and the dimension is $4$ and $5$ . However, it should be noted that the computational times for the AEP algorithm remains acceptable, in most cases below one second even in high dimensions. For the sake of completeness and as a reference, the computational time figures of the first implementation, reported in the study of the original manuscript, are also given.

To conclude, we perform a sensitivity study of the two above-mentioned approaches for different dependency structures and number of dimensions. Fig. 4 shows how the cdf value calculated with the AEP algorithm changes as the underlying degree of dependency varies, in the cases of a bivariate Gaussian copula and a three-dimensional Clayton copula. The solid black lines of the graphs represent the cdf evaluated at four quantiles $s$ for increasing values of $\rho$ (the non-diagonal entry of the correlation matrix) and $\theta$ parameters, for the Gaussian copula and the Clayton copula, respectively. The cdf values were also compared with those obtained by the MC simulation approach. The average absolute difference between the results of the two methods, across the four quantiles, is highlighted by the dotted red line. It can be seen that this remains stable at low values, regardless of the underlying dependency structure. In all cases, the results produced by the two methods almost coincide.

Figure 4 Sensitivity analysis of the cdf at four quantiles $s$ calculated using the AEP algorithm (n_iter = 7) for different copula models, dimensionality, and underlying degree of dependency. The values of $s$ are 1.25, 1.5, 1.75, and 1.95 for the bivariate Gaussian copula (left plot) and 2, 2.3, 2.65, and 2.85 for the three-dimensional Clayton (right plot). Each solid black line indicates the values of the cdf for a given $s$ , as the respective parameters $\rho$ and $\theta$ change. The results of the AEP algorithm correspond to those of the Monte Carlo (MC) simulation approach, using $10^7$ number of simulations. The dashed red line represents the average absolute difference between the two method cdf values, calculated across the four quantiles.

5. Loss reserve

In non-life insurance, contracts do not settle when insured events occur. At the accident date, the insured event triggers a claim that will generate payments in the future. The task of predicting these liabilities is called claims reserving and it assesses claim outstanding loss liabilities (see, e.g., Wüthrich & Merz, Reference Wüthrich and Merz2015, p. 11). In the present work, we refer to the total outstanding loss liabilities of past claims as the loss reserve or claims reserve. Fig. 5 sketches an example of the timeline evolution of an individual claim. The insured event occurs within the insured period, but the claim settles after several years. In particular, after the claim is reported, the insurance company makes an initial quantification of the claim payment size, the so-called case estimate. Two payments occur thereafter. These payments are not known at the evaluation date, and they require to be estimated. In between the payments, the case estimate is updated. Until a claim settle, the insurance company refers to it as an open claim. In certain circumstances, settled claims can be reopened (Friedland, Reference Friedland2010, p. 431).

Figure 5 Example of events of a non-life insurance claim.

In this section, we first define the development triangles, the data structure commonly used by actuarial departments for claims reserving (Friedland, Reference Friedland2010, p. 51). These are aggregate representations of the individual claim data. Then, we present the collective risk model for claims reserving in Ricotta & Clemente (Reference Ricotta and Clemente2016) and Clemente et al. (Reference Clemente, Savelli and Zappa2019), which allows to estimate the variability of the reserve. This model requires extra-inputs from a deterministic model to be implemented. In this manuscript, we rely on the Fisher–Lange model (Fisher & Lange, Reference Fisher and Lange1973) as proposed in Savelli & Clemente (Reference Savelli and Clemente2014). Further details can be found in Section C.

5.1 Problem framework

Let the index $i=0,\ldots,\mathcal{J}$ denotes the claim accident period, and let the index $j=0,\ldots,\mathcal{J}$ represents the claim development period, over the time horizon $\mathcal{J}\gt 0$ . The so-called development triangle is the set:

\begin{equation*} \mathcal {T}=\{(i, j)\;: \; i=0, \ldots, \mathcal {J}, \; j=0, \ldots, \mathcal {J}; \; i+j \leq \mathcal {J}\}, \quad \mathcal {J}\gt 0. \end{equation*}

Below is the list of the development triangles used in this section.

  • The triangle of incremental paid claims:

    \begin{equation*}X^{(\mathcal {T})} =\left \{x_{i, j}\;:\; (i, j) \in \mathcal {T} \right \}, \end{equation*}
    with $x_{i, j}$ being the total payments from the insurance company for claims occurred at accident period $i$ and paid in period $i+j$ .
  • The triangle of the number of paid claims:

    \begin{equation*}N^{(\mathcal {T})} =\left \{n_{i, j}\;:\; (i, j) \in \mathcal {T} \right \}, \end{equation*}
    with $n_{i, j}$ being the number of claim payments occurred in accident period $i$ and paid in period $i+j$ . The triangle of average claim cost can be derived from the incremental paid claims triangle and the number of paid claims. Indeed, we define
    \begin{equation*}m_{i,j}=\frac {x_{i,j}}{n_{i, j}},\end{equation*}
    where $m_{i,j}$ is the average claim cost for accident period $i$ and development period $j$ .
  • The triangle of incremental amounts at reserve:

    \begin{equation*}R^{(\mathcal {T})} =\left \{r_{i, j}\;:\; (i, j) \in \mathcal {T} \right \}, \end{equation*}
    with $r_{i, j}$ being the amount booked in period $i+j$ for claims occurred in accident period $i$ .
  • The triangle of the number of open claims:

    \begin{equation*}O^{(\mathcal {T})} =\left \{o_{i, j}\;:\; (i, j) \in \mathcal {T} \right \},\end{equation*}
    with $o_{i, j}$ being the number of claims that are open in period $i+j$ for claims occurred in accident period $i$ .
  • The triangle of the number of reported claims:

    \begin{equation*}D^{(\mathcal {T})} =\left \{d_{i, j}\;:\; (i, j) \in \mathcal {T} \right \}, \end{equation*}
    with $d_{i, j}$ being the claims reported in period $j$ and belonging to accident period $i$ . Often, the number of reported claims is aggregated by accident period $i$ .

We implemented the model of Ricotta & Clemente (Reference Ricotta and Clemente2016) and Clemente et al. (Reference Clemente, Savelli and Zappa2019), hereafter referred to as CRMR, which connects claims reserving with aggregate distributions. Indeed, the claims reserve is the sum of the (future) payments:

(16) \begin{equation} R = \sum _{i+j\gt \mathcal{J}}X_{i,j}. \end{equation}

The authors represent the incremental payments in each cell of the triangle of incremental payments as a compound mixed Poisson-gamma distribution under the assumptions in Ricotta & Clemente (Reference Ricotta and Clemente2016). For $i,j=0,\ldots,\mathcal{J},$

(17) \begin{equation} X_{i,j}=\sum _{h=1}^{N_{i,j}} \psi Z_{h;i,j}, \end{equation}

where the index $h$ is referred to the individual claim severity. The random variable $\psi$ follows a gamma distribution, see Table 4, and introduces dependence between claim-sizes of different cells.

Table 4. Parametric assumptions of the CRMR. In the upper table, we show the parameters for $Z_{h;i,j}, \psi,$ and $q$ that are gamma-distributed. The parameters of the structure variables are specified from the user starting from the variance of $\psi$ and $q$ , indeed Ricotta and Clemente (Reference Ricotta and Clemente2016) assume $\mathbb{E}[\psi]=\mathbb{E}[q]=1$ . The estimator for the average cost of the individual payments is derived with the Fisher–Lange. The variability of the individual payments is instead obtained from the company database. In the lower table, we show the parameter for the claim payment number, that is, a mixed Poisson-gamma distribution with $\hat n_{i,j}$ derived from the Fisher–Lange

5.1.1 Predicting the claims reserve

In order to estimate the value for the claims reserve $R$ , this approach requires additional parametric assumptions on $N_{i,j}$ and $Z_{h;i,j}$ when $i+j \gt \mathcal{T}$ . In this section, we illustrate how to use the results from the Fisher–Lange to determine the parameters of $N_{i,j}$ and $Z_{h;i,j}$ . The Fisher–Lange is a deterministic average cost method for claims reserving (Institute and Faculty of Actuaries, 1997, Section H). The interested reader can refer to Section C for a discussion on the Fisher–Lange and an implementation of the model using gemact.

For any cell ( $i,j$ ), with $i+j \gt \mathcal{T}$ , the Fisher–Lange can be used to determine the expected number of future payments $\hat n_{i,j}$ and the future average claim cost $\hat m_{i,j}$ . The original concept behind the CRMR is found in Savelli & Clemente (2009), where the authors assumed that $N_{i,j}$ is Poisson distributed with mean $\hat{n}_{i,j}$ and $Z_{h;i,j}$ is gamma distributed with mean $\hat m_{i,j}$ . The CoV of $Z_{h;i,j}$ is $\hat c_{z_j}$ , the relative variation of the individual payments in development period $j$ , independent from the accident period. The values for $\hat c_{z_j}$ are estimated using the individual claims data available to the insurer at the evaluation date. The CRMR is extended in Ricotta & Clemente (Reference Ricotta and Clemente2016) to take into account of the variability of the severity parameter estimation (Estimation Variance), in addition to the random fluctuations of the underlying process (Process Variance) (Wüthrich & Merz, Reference Wüthrich and Merz2015, p. 28). This is achieved by considering two structure variables ( $q$ and $\psi$ ), on claim count and average cost, to describe parameter uncertainty on $N_{i,j}$ and $Z_{h;i,j}$ . The parametric assumptions of the model are summarized in Table 4. For $i,j=0,\ldots,\mathcal{J}$ , $\hat m_{i,j}$ and $\hat n_{ij}$ are obtained from the computation of the Fisher–Lange.

5.2 Illustration lossreserve

In this section, we show an example of the CRMR using the lossreserve module. In this respect, we simulated the claims reserving datasets using the individual claim simulator in Avanzi et al. (Reference Avanzi, Taylor, Wang and Wong2021) to generate the upper triangles $X^{\left (\mathcal{T}\right)}, N^{\left (\mathcal{T}\right)}, O^{\left (\mathcal{T}\right)}$ and $D^{\left (\mathcal{T}\right)}$ . The $R^{\left (\mathcal{T}\right)}$ upper triangle has been simulated using the simulator of Avanzi et al. (Reference Avanzi, Taylor and Wang2023). The data are simulated using the same assumptions (on an yearly basis) of the Simple, short tail claims environment shown in Al-Mudafer et al. (Reference Al-Mudafer, Avanzi, Taylor and Wong2022).

No inflation is assumed to simplify the forecast of the future average costs. Estimating and extrapolating a calendar period effect in claims reserving is a delicate and complex subject, and a more detailed discussions can be found in Kuang et al. (Reference Kuang, Nielsen and Nielsen2008a, Reference Kuang, Nielsen and Nielsenb), Pittarello et al. (Reference Pittarello, Hiabu and Villegas2023). Furthermore, in practice, insurers might require a specific knowledge of the environment in which the agents operate to set a value for the inflation (Avanzi et al., Reference Avanzi, Taylor and Wang2023). Here, we want to limit the assumptions for this synthetic scenario.

Since the entire claim history is available from the simulation and we know the (true) value for the future payments, we can use this information to back-test the performance of our model (Gabrielli & Wüthrich, Reference Gabrielli and Wüthrich2019). In particular, we refer to the actual amount that the insurer will pay in future calendar years as the actual reserve, that is, the amount that the insurer should set aside to cover exactly the future obligations. The CRMR is compared with the chain-ladder method of Mack (Reference Mack1993), hereafter indicated as CHL, from the R package ChainLadder (Gesmann et al., Reference Gesmann, Murphy, Zhang, Carrato, Wüthrich, Concina and Dal Moro2022). These data, together with the datasets from Savelli & Clemente (Reference Savelli and Clemente2014), have been saved in the gemdata module for reproducibility.

>>> from gemact import gemdata

>>> ip = gemdata.incremental_payments_sim

>>> pnb = gemdata.payments_number_sim

>>> cp = gemdata.cased_payments_sim

>>> opn = gemdata.open_number_sim

>>> reported = gemdata.reported_claims_sim

>>> czj = gemdata.czj_sim

The lossreserve class diagram is illustrated in Fig. 6: the triangular datasets are stored in the AggregateData class, and the model parameters are contained in the ReservingModel class. The actual computation of the loss reserve is performed with the LossReserve class that takes as inputs an AggregateData object, a ReservingModel object, and the parameters of the claims reserving model computation.

Figure 6 Class diagram of the lossreserve module. A rectangle represents a class; an arrow connecting two classes indicates that the target class employs the origin class as an attribute. In this case, a LossReserve object entails ReserveModel and AggregateData class instances.

>>> from gemact import AggregateData

>>> from gemact import ReservingModel

>>> import numpy as np

>>> ad = AggregateData(

incremental_payments=ip,

cased_payments=cp,

open_claims_number=opn,

reported_claims=reported,

payments_number=pnb

)

>>> resmodel_crm = ReservingModel(

tail=False,

reserving_method='crm',

claims_inflation=np.array([1]),

mixing_fq_par = .01,

mixing_sev_par = .01,

czj=czj

)

In more detail, the ReservingModel class arguments are as follows:

  • The tail parameter (boolean) specifying whether the triangle tail is to be modeled.

  • The reserving_method parameter, "crm" in this case.

  • The claims_inflation parameter indicating the vector of claims inflation. In this case, as no claims inflation is present, we simply set it to one in all periods.

  • The mixing frequency and severity parameters mixing_fq_par and mixing_sev_par. In Ricotta & Clemente (Reference Ricotta and Clemente2016), the authors discuss the calibration of the structure variable. Without having a context of a real-world example, we simply set the structure variables to gamma random variables with mean 1 and standard deviation $0.01$ , as it is a medium–low risk value in the authors’ examples.

  • The coefficients of variation of the individual claim severity computed for each development period. In particular, for each development period, the insurer can compute the CoV from the individual observations and save them in the vector czj.

  • The vector czj of the coefficients of variation of the individual claim severity, for each development period. The insurer can compute the CoV from the individual observations, for each development period.

The computation of the loss reserve occurs within the LossReserve class:

>>> from gemact import LossReserve

>>> lr = LossReserve(

data=ad,

reservingmodel=resmodel_crm,

ntr_sim = 1000,

random_state = 42

)

To instantiate the LossReserve class, AggregateData and ReservingModel objects need to be passed as arguments. The additional arguments required for the computation are as follows:

  • the number of simulated triangles ntr_sim,

  • the (pseudo)random number generator initializer random_state.

The mean reserve estimate for the CRMR can be extracted from the reserve attribute. In a similarly way to LossModel and LossAggregation, the distribution of the reserve can be accessed from the dist property. The reserve quantiles can be obtained with the ppf method. Output figures are expressed in millions to simplify reading.

>>> lr.reserve

8245996481.515498

>>> lr.ppf(q = np.array([.25, .5, .75, .995, .9995]))/10**6

array([8156.79994049, 8244.66099324, 8335.94438221, 8600.42263791, 8676.84206101])

Table 5. Reserve and mean squared error of prediction (MSEP) by accident period for the CRMR and the CHL. The actual reserve and its process error (PE) by accident period are also indicated. Amounts are shown in millions

Table 6. Total reserve estimates, their relative value, as a fraction of the actual value (8599.04), and their coefficients of variation (CoV), for the CRMR and the CHL. Absolute amounts are reported in millions

5.2.1 Comparison with the chain-ladder

This section compares the CRMR and the CHL with the actual reserve we know from the simulation. For clarity, the reserve of accident period $i$ is defined as $R_i=\sum ^{\mathcal{T}}_{j=\mathcal{T}-i+1} X_{i, j}$ . Table 5 reports the main results. We can see that the CHL reserve estimate is closer than the CRMR reserve estimate to the actual reserve (the “Reserve” column of each block). In addition to the reserve estimate, we provide the mean squared error of prediction (MSEP) (Wüthrich & Merz, Reference Wüthrich and Merz2015, p. 268) of the reserve estimator for each accident period and for the total reserve. In the aforementioned manuscript, the author introduces the MSEP as a measure of risk under the notion in Wüthrich & Merz (Reference Wüthrich and Merz2015, p. 169) conditional on the information available in the upper triangle. The MSEP can be decomposed into two components, the irreducible risk arising from the data generation process (process error, PE) and a risk component arising from the model used to calculate the reserve (model error, ME). We can provide a reference value for the true PE by running multiple simulations from the simulator described in the previous section and calculating the standard error of the reserves by accident period. To obtain the PE, we simulated $100$ triangles. Net of the ME component, the results in Table 5 appear to show that the CHL underestimates the forecast error in the early accident periods and overestimates the forecast error in the later accident periods.

To conclude, Table 6 shows the total reserves of the CRMR and the CHL. In particular, it highlights the reserve amount as a percentage of the actual reserve (8599.04) and the CoV, that is, the MSEP divided by the reserve. The total relative variability referred to the PE component gives a CoV of 0.76 %.

6. Conclusions

This paper introduces gemact, a Python package for non-life actuarial modeling based on the collective risk theory.

Our library expands the reach of actuarial sciences within the growing community of Python programming language. It provides new functionalities and tools for loss modeling, loss aggregation, and loss reserving, such as the $(a, b, 0)$ and $(a, b, 1)$ classes of distributions, the AEP algorithm and the collective risk model for claims reserving. Hence, it can be applied in different areas of actuarial sciences, including pricing, reserving, and risk management. The package has been designed primarily for the academic environment. Nevertheless, its use as support for insurance business specialists in prototypes modeling, business studies, and analyses is not to be excluded.

The structure of our package aims to ensure modifiability, maintainability, and scalability, as we thought of gemact as an evolving and growing project in terms of introducing features, integrating functionalities, enhancing methodologies, and expanding its scopes. Possible future enhancements could involve the introduction of new probability distribution families, the implementation of supplementary methodologies for the approximation of quantiles of the sum of random variables, and the addition of costing procedures for exotic and nontraditional reinsurance solutions.

Device specifications

All experiments and analyses were run on a computer with an Intel® Core™ i7-1065G7 CPU processor with 16 GB RAM, running at 1.30 GHz, on Windows 11 Home edition.

Acknowledgments

Previous versions of gemact were presented at the Mathematical and Statistical Methods for Actuarial Sciences and Finance 2022, and at the Actuarial Colloquia 2022, in the ASTIN section. We would like to thank all the people who gave us feedback and suggestions about the project.

Data availability statement

The data and code supporting the findings of this study are openly available in GitHub at gpitt71/gemact-code. Supplementary material not included as code blocks in the manuscript can be found in the subfolder vignette. The results contained in the manuscript are reproducible, excluding environment-specific numerical errors. These discrepancies do not affect the overall validity of the results. The gpitt71/gemact-code folder is registered with the unique Zenodo DOI reference number 10.5281/zenodo.10117505.

Funding statement

This work received no specific grant from any funding agency, commercial, or not-for-profit sectors.

Competing interests

The author(s) declare none.

A. List of the supported distributions

Table A.1 gives an overview of the distributions available in gemact. In particular, the table presents the distribution name (column one) and its name within gemact apparatus (column two). Moreover, it shows the distribution support (column three) and whether the distribution is supported in scipy (column four).

Table A.1. List of distributions supported by gemact

Lastly, the list of available copulas is provided in Table A.2.

Table A.2. List of copulas supported by gemact

B. The AEP algorithm

The AEP algorithm is a numerical procedure, based on a geometric approach, to calculate the joint cdf of the sum of dependent random variables. The graphical interpretation of the basic idea of this algorithm is straightforward, especially in the two-dimensional case. Therefore, to facilitate the reader’s understanding, in this section we limit ourselves to the case where $d = 2$ and focus only on its first two iterations. The underlying logic can then be extended to $d \geq 2$ and to any number of iterations.

Let us first define, for $b_1,b_2 \in \mathbb R$ , a simplex as:

\begin{equation*} \mathcal {S}((b_1, b_2), h)= \begin {cases}\left \{x_1,x_2 \in \mathbb {R}\;:\; x_1-b_1\gt 0, x_2-b_2\gt 0, \text { and } \sum _{k=1}^2\left (x_k-b_k\right) \leq h\right \} & \text { if } h\gt 0, \\[5pt] \left \{x_1,x_2 \in \mathbb {R}\;:\; x_1-b_1 \leq 0, x_2-b_2 \leq 0, \text { and } \sum _{k=1}^2\left (x_k-b_k\right)\gt h\right \} & \text { if } h\lt 0,\end {cases} \end{equation*}

and a square as:

\begin{equation*} \mathcal {Q}((b_1,b_2), h)= \begin {cases}(b_1, b_1+h] \times (b_2, b_2+h] & \text { if } h\gt 0, \\[5pt] (b_1+h, b_1] \times (b_2+h, b_2] & \text { if } h\lt 0 .\end {cases} \end{equation*}

The $H$ -measure of the square $V_H\left (\mathcal{Q}(b_1,b_2, h)\right)$ is computed as follows (Nelsen, Reference Nelsen2007, p. 8):

\begin{equation*} V_H\left (\mathcal {Q}((b_1,b_2), h)\right) = H\left (b_1+h, b_2+h\right)-H\left (b_1, b_2+h\right)-H\left (b_1+h, b_2\right)+H\left (b_1, b_2\right), \end{equation*}

where $H$ the joint cdf in Equation (15). In general, the algorithm is based on the observation that a simplex approximated by a square generates three smaller simplexes, each of which can in turn be approximated by a square that generates three new, even smaller simplexes, and so on. By repeating this iterative scheme with an increasing number of iterations, the quality of the approximation improves and the error tends to 0. It can be noted that some simplexes generated by the process lies outside the original simplex. The measure of those needs to be subtracted instead of being added.

Fig. C.1a shows the simplex $\mathcal{S}_1=\mathcal{S}((0, 0), s)$ , where $s \in \mathbb{R}^+$ is the value at which the joint cdf is calculated in Equation (15). For the first iteration (Fig. C.1b), we adopt the square $\mathcal{Q}_1 = \mathcal{Q}((0,0), \frac 2 3 s)$ . Arbenz et al. (Reference Arbenz, Embrechts and Puccetti2011) explains that the choice of the $2/3$ factor provides fastest convergence when $d=2$ . This factor is set automatically in our implementation. Hence, at the end of the first iteration, we have

\begin{equation*} P\left [X_1+X_2 \leq s \right] \approx P_1 \left (s\right), \end{equation*}

where $P_1 \left (s\right)=V_H\left (\mathcal{Q}_1\right)$ . For example, if we consider uniform marginals and a Gaussian copula with correlation $0.7$ , we would obtain the following.

>>> from gemact import Margins, Copula

>>> margins = Margins(

dist=['uniform', 'uniform'],

par=[{'a': 0, 'b': 1}, {'a': 0, 'b': 1}]

)

>>> copula = Copula(

dist='gaussian',

par={'corr': [ [1, 0.7], [0.7, 1]]}

)

>>> la = LossAggregation(

copula=copula,

margins=margins

)

>>> la.cdf(x=1, n_iter=1, method='aep')

0.55188934403716

In the second iteration of the algorithm, shown in Fig. 7c, we use again the same logic and approximate the simplexes in Fig. 7b:

\begin{align*} &\mathcal{S}_2 = \mathcal{S} \left(\left(\frac 2 3 s, 0\right), \frac 1 3 s\right) \\[5pt] &\mathcal{S}_3=\mathcal{S}\left(\left(\frac 2 3 s, \frac 2 3 s\right),\frac{-1}{3} s\right)\\[5pt] &\mathcal{S}_4=\mathcal{S}\left(\left(0,\frac 2 3 s\right), \frac 1 3 s\right) \end{align*}

with the squares:

\begin{align*} &\mathcal{Q}_2 = \mathcal{Q} \left(\left(\frac 2 3 s, 0\right), \frac 1 3 s\right) \\[5pt] &\mathcal{Q}_3=\mathcal{Q}\left(\left(\frac 2 3 s, \frac 2 3 s\right),\frac{-1}{3} s\right)\\[5pt] &\mathcal{Q}_4=\mathcal{Q}\left(\left(0,\frac 2 3 s\right), \frac 1 3 s\right). \end{align*}

Note that this time the $H$ -measure of $\mathcal{Q}_3$ is subtracted. Similarly to the first iterations, the $\frac{1}{3}$ factors are chosen accordingly to the guidelines of the original manuscript to guarantee the fastest convergence. We obtain, at the second iteration:

\begin{equation*} P\left [X_1+X_2 \leq s \right] \approx P_2 \left (s\right), \end{equation*}

with $P_2 \left (s\right)= P_1\left (s\right)+V_H\left (\mathcal{Q}_2\right)-V_H\left (\mathcal{Q}_3\right)+V_H\left (\mathcal{Q}_4\right).$

Figure C.1 Sketch of the first two iterations of the AEP algorithm in the two-dimensional case.

To conclude, continuing with the previous example, the results for the first two iterations is given in the code block below.

>>> la.cdf(x=1, n_iter=2, method='aep')

0.4934418427652146

C. Claims reserving with the Fisher–Lange

This section briefly introduces the Fisher–Lange approach. The claims reserve is the sum of the (future) payments, forecast as the product between the predicted future average cost and the predicted number of future payments, for each cell. In formula:

(C1) \begin{equation} R = \sum _{i+j\gt \mathcal{J}}\hat m_{i,j} \hat n_{i,j}. \end{equation}

The average claim cost in the lower triangle is forecast as the projection of the inflated average claim cost:

(C2) \begin{equation} \widehat{m}_{i, j}=m_{\mathcal{J}-j, j} \prod _{h=\mathcal{J}+1}^{i+j}\left (1+\delta _h\right), \end{equation}

where $\delta _h$ represents the claims inflation for calendar period $h$ .

As far the number of claims are concerned, this method assumes that the future number of paid claims is related to the percentage of open claims at the evaluation date and to the claims settlement speed. Indeed, at the evaluation date, the lower triangle is estimated as:

(C3) \begin{equation} \hat{n}_{i, j}=o_{i, \mathcal{J}-i} \cdot \alpha _{\mathcal{J}-i} \cdot v_j^{(i)}, \end{equation}

where $\left [\alpha _j\right]$ , $j=0,\ldots, \mathcal{J}-1$ , is the vector of open claims given by $\alpha _j = \mathbb{E}\left [ \tau _{i, j}\right]$ , and

(C4) \begin{equation} \tau _{i, j}=\frac{\sum _{h=j+1}^{\mathcal{J}-i} n_{i, h}+o_{i, \mathcal{J}-i}}{o_{i, j}}, \end{equation}

for $i=0, \ldots, \mathcal{J}-1$ and $j=0, \ldots, \mathcal{J}-i-1$ . It is assumed that $\alpha _{\mathcal{J}}=1$ . The claim settlement speed is then computed for each accident year. The settlement speed for accident period $\mathcal{J}$ is

(C5) \begin{equation} v_j^{(\mathcal{J})}=\frac{n_{\mathcal{J}-j, j} \cdot \frac{d_{\mathcal{J}}}{d_{\mathcal{J}-j}}}{\sum _{j=1}^{\mathcal{J}} n_{\mathcal{J}-j, j} \cdot \frac{d_{\mathcal{J}}}{d_{\mathcal{J}-j}}}, \end{equation}

where $d_i$ represents the number of reported claims for accident period $i$ , with $i=0,\ldots,\mathcal{J}$ . The formula is corrected for other accident years following the approach in Savelli & Clemente (Reference Savelli and Clemente2014, p. 141).

Similarly to the CRMR described in Section 5, the results for the Fisher–Lange can be computed with the gemact package. Below, we show an example using the simulated datasets from Section 5.

>>> from gemact import gemdata

>>> ip = gemdata.incremental_payments_sim

>>> pnb = gemdata.payments_number_sim

>>> cp = gemdata.cased_payments_sim

>>> opn = gemdata.open_number_sim

>>> reported = gemdata.reported_claims_sim

>>> czj = gemdata.czj_sim

>>> claims_inflation = np.array([1])

The data are represented in the AggregateData class.

>>> from gemact.lossreserve import AggregateData

>>> ad = AggregateData(

incremental_payments=ip,

cased_payments=cp,

open_claims_number=opn,

reported_claims=reported,

payments_number=pnb)

Afterward, we specify the ReservingModel. In this example, we fix the parameter tail to True to obtain an estimate of the tail.

Table C.1. Reserves by accident period for the CRMR and the Fisher–Lange. We also report the actual reserve. Amounts are shown in millions

>>> resmodel = ReservingModel(

tail=False,

reserving_method='fisher_lange',

claims_inflation=claims_inflation)

Thereafter, the actual computation of the loss reserve is performed within the LossReserve class:

>>> from gemact.lossreserve import LossReserve

>>> lossreserve = LossReserve(data=ad, reservingmodel=resmodel)

The LossReserve class comes with a summary view of the estimated reserve per each accident period, in a similar way to the LossModel class, the print_loss_reserve method. In Table C.1, we report the CRMR results from Table 5 and we add the results for the Fisher–Lange.

As expected, being the Fisher–Lange the underlying methodology to the CRMR, the results for the claims reserve provided from the two approaches are consistent.

On top of this, insights on the behavior of Fisher–Lange $\left [\alpha _j\right]$ and settlement speed $\left [v^{(i)}_j\right]$ , for $j=0,\ldots, \mathcal{J}-1$ , can be inspected with the plot_alpha_fl and plot_ss_fl methods.

References

Avanzi, B., Taylor, G. & Wang, M. (2023). Splice: a synthetic paid loss and incurred cost experience simulator. Annals of Actuarial Science, 17 (1), 735.CrossRefGoogle Scholar
Al-Mudafer, M. T., Avanzi, B., Taylor, G., & Wong, B. (2022). Stochastic loss reserving with mixture density neural networks. Insurance: Mathematics and Economics, 105, 144174.Google Scholar
Albrecher, H., Beirlant, J., & Teugels, J. L. (2017). Reinsurance: Actuarial and statistical aspects. John Wiley & Sons.CrossRefGoogle Scholar
Antal, P. (2009). Mathematical methods in reinsurance. Swiss Re.Google Scholar
Arbenz, P., Embrechts, P., & Puccetti, G. (2011). The aep algorithm for the fast computation of the distribution of the sum of dependent random variables. Bernoulli, 17(2), 562591.CrossRefGoogle Scholar
Avanzi, B., Taylor, G., Wang, M., & Wong, B. (2021). Synthetic: An individual insurance claim simulator with feature control. Insurance: Mathematics and Economics, 100, 296308.Google Scholar
Bass, L., Clements, P., & Kazman, R. (2003). Software architecture in practice. Addison-Wesley Professional.Google Scholar
Bean, M. A. (2000). Probability: the science of uncertainty with applications to investments, insurance, and engineering. Available online at the address https://api.semanticscholar.org/CorpusID:106862438.Google Scholar
Bogaardt, J. (2022). chainladder package, version 0.8.13 [Computer software manual]. Available online at the address https://chainladder-python.readthedocs.io/en/latest/intro.html.Google Scholar
Bok, D. (2022). copulae package, version 0.7.7 [Computer software manual]. Available online at the address https://copulae.readthedocs.io/en/latest/.Google Scholar
Bondi, A. B. (2000). Characteristics of scalability and their impact on performance. In Proceedings of the 2nd international workshop on Software and performance (pp. 195–203).CrossRefGoogle Scholar
Bühlmann, H. (1984). Numerical evaluation of the compound poisson distribution: recursion or fast fourier transform? Scandinavian Actuarial Journal, 1984(2), 116126.CrossRefGoogle Scholar
Clemente, G. P., Savelli, N., & Zappa, D. (2019). Modelling outstanding claims with mixed compound processes in insurance. International Business Research, 12(3), 123138.CrossRefGoogle Scholar
Cummins, D., Lewis, C. & Phillips, R. (1999) Pricing excess-of-loss reinsurance contracts against catastrophic loss. In: Froot, K. A. (Ed.), The financing of catastrophe risk (pp.93148). University of Chicago Press.Google Scholar
Dutang, C., Goulet, V., & Langevin, N. (2022). Feller-pareto and related distributions: Numerical implementation and actuarial applications. Journal of Statistical Software, 103(6), 122.CrossRefGoogle Scholar
Dutang, C., Goulet, V., & Pigeon, M. (2008). actuar: An r package for actuarial science. Journal of Statistical software, 25, 137.Google Scholar
Embrechts, P., & Frei, M. (2009). Panjer recursion versus fft for compound distributions. Mathematical Methods of Operations Research, 69(3), 497508.CrossRefGoogle Scholar
Fisher, W. H., & Lange, J. T. (1973). Loss reserve testing: a report year approach. Proceedings of the Casualty Actuarial Society, 60, 189207.Google Scholar
Friedland, J. (2010) Estimating unpaid claims using basic techniques. In Casualty actuarial society, vol. 201. Google Scholar
Gabrielli, A., & Wüthrich, M. V. (2019). Back-testing the chain-ladder method. Annals of Actuarial Science, 13(2), 334359.CrossRefGoogle Scholar
Genz, A., & Bretz, F. (1999). Numerical computation of multivariate t-probabilities with application to power calculation of multiple contrasts. Journal of Statistical Computation and Simulation, 63(4), 103117. doi: 10.1080/00949659908811962.CrossRefGoogle Scholar
Genz, A., & Bretz, F. (2002). Comparison of methods for the computation of multivariate t probabilities. Journal of Computational and Graphical Statistics, 11(4), 950971. doi: 10.1198/106186002394. Retrieved from https://doi.org/10.1198/106186002394.CrossRefGoogle Scholar
Gerber, H. U. (1982). On the numerical evaluation of the distribution of aggregate claims and its stop-loss premiums. Insurance: Mathematics and Economics, 1(1), 1318.Google Scholar
Gesmann, M., Murphy, D., Zhang, Y. W., Carrato, A., Wüthrich, M., Concina, F., & Dal Moro, E. (2022). Chainladder: Statistical methods and models for claims reserving in general insurance [Computer software manual]. Received from https://CRAN.R-project.org/package=ChainLadder (R package version 0.2.15).Google Scholar
Grübel, R., & Hermesmeier, R. (1999). Computation of compound distributions i: Aliasing errors and exponential tilting. ASTIN BULLETIN: The Journal of the IAA, 29(2), 197214.CrossRefGoogle Scholar
Hofert, M., & Mächler, M. (2011). Nested archimedean copulas meet R: The nacopula package. Journal of Statistical Software, 39(9), 120. Retrieved from https://www.jstatsoft.org/v39/i09/.CrossRefGoogle Scholar
Hunter, J. D. (2007). Matplotlib: A 2d graphics environment. Computing in Science & Engineering, 9(3), 9095. doi: 10.1109/MCSE.2007.55.CrossRefGoogle Scholar
Institute and Faculty of Actuaries (1997). Claims reserving manual (2nd ed.).Google Scholar
Johansson, N., & Löfgren, A. (2009). Designing for extensibility: An action research study of maximizing extensibility by means of design principles. (B.S. thesis). Univeristy of Gothenburg.Google Scholar
Klugman, S. A., Panjer, H. H., & Willmot, G. E. (2012). Loss models: From data to decisions, vol. 715. John Wiley & Sons.Google Scholar
Kojadinovic, I., & Yan, J. (2010). Modeling multivariate distributions with continuous margins using the copula R package. Journal of Statistical Software, 34(9), 120. Retrieved from https://www.jstatsoft.org/v34/i09/.CrossRefGoogle Scholar
Kuang, D., Nielsen, B., & Nielsen, J. P. (2008a). Forecasting with the age-period-cohort model and the extended chain-ladder model. Biometrika, 95(4), 987991.CrossRefGoogle Scholar
Kuang, D., Nielsen, B., & Nielsen, J. P. (2008b). Identification of the age-period-cohort model and the extended chain-ladder model. Biometrika, 95(4), 979986.CrossRefGoogle Scholar
Lab, M. D. T. A. (2022). Copulas package, version 0.7.0 [Computer software manual]. Retrieved from https://github.com/sdv-dev/Copulas.Google Scholar
Ladoucette, S. A., & Teugels, J. L. (2006). Analysis of risk measures for reinsurance layers. Insurance: Mathematics and Economics, 38(3), 630639.Google Scholar
Mack, T. (1993). Distribution-free calculation of the standard error of chain ladder reserve estimates. ASTIN Bulletin: The Journal of the IAA, 23(2), 213225.CrossRefGoogle Scholar
Martelli, A., Ravenscroft, A., & Ascher, D. (2005). Python cookbook. O’Reilly Media, Inc.Google Scholar
Mildenhall, S. (2022). Aggregate package. [Computer software manual]. Retrieved from https://aggregate.readthedocs.io/en/latest/.Google Scholar
Nelsen, R. B. (2007). An introduction to copulas (2nd ed.). Springer Science & Business Media. doi: 10.1007/0-387-28678-0.Google Scholar
Nielsen, B. (2015). apc: An r package for age-period-cohort analysis. The R Journal, 7(2), 52.CrossRefGoogle Scholar
Ozgur, C., Colliau, T., Rogers, G., & Hughes, Z. (2022). Matlab vs. python vs. r. Journal of Data Science, 15(3), 355372. doi: 10.6339/JDS.201707_15(3).0001.CrossRefGoogle Scholar
Panjer, H. H. (1981). Recursive evaluation of a family of compound distributions. ASTIN Bulletin: The Journal of the IAA, 12(1), 2226.CrossRefGoogle Scholar
Parodi, P. (2014). Pricing in general insurance (1st ed.). CRC Press. doi:10.1201/b17525.CrossRefGoogle Scholar
Pittarello, G., Hiabu, M., & Villegas, A. M. (2023). Replicating and extending chain-ladder via an age-period-cohort structure on the claim development in a run-off triangle. arXiv preprint arXiv:2301.03858.Google Scholar
R Core Team. (2017). R: A language and environment for statistical computing [Computer software manual]. Vienna, Austria. Retrieved from https://www.R-project.org/.Google Scholar
Ricotta, A., & Clemente, G. P. (2016). An extension of collective risk model for stochastic claim reserving. Journal of Applied Finance and Banking, 6(5), 45.Google Scholar
Savelli, N., & Clemente, G. P. (2014). Lezioni di matematica attuariale delle assicurazioni danni. EDUCatt-Ente per il diritto allo studio universitario dell’Università Cattolica. Retrieved from http://hdl.handle.net/10807/67154.Google Scholar
Savelli, N., & Clemente, G. P. (2009). A collective risk model for claims reserve distribution. In Proceedings of “Convegno di Teoria del Rischio”, Campobasso (pp. 59–88).Google Scholar
Shevchenko, P. V. (2010). Calculation of aggregate loss distributions. Journal of Operational Risk, 5(2), 340.CrossRefGoogle Scholar
Sommerville, I. (2011). Software engineering. Pearson Education Inc.Google Scholar
Sundt, B. (1990). On excess of loss reinsurance with reinstatements. Transactions of the ASTIN Colloquium, 12, 73.Google Scholar
Van Rossum, G., & Drake, F. L. (2009). Python 3 reference manual. CreateSpace.Google Scholar
Virtanen, P., Gommers, R., Oliphant, T. E., Haberland, M., Reddy, T., Cournapeau, D.van Mulbregt, P. (2020). SciPy 1.0: fundamental algorithms for scientific computing in python. Nature Methods, 17(3), 261272. doi: 10.1038/s41592-019-0686-2.CrossRefGoogle ScholarPubMed
Wang, P. (2013). Risk modeling of multi-year, multi-line reinsurance using copulas. Journal of Insurance Issues, 36(1), 5881, Retrieved from https://ideas.repec.org/a/wri/journl/v36y2013i1p58-81.html.Google Scholar
Wang, S. (1998). Aggregation of correlated risk portfolios: models and algorithms. In Proceedings of the casualty actuarial society, vol. 85 (pp. 848–939).Google Scholar
Wilhelmy, L. (2010). Economic capital: Applying theory to practice. In: Presentation at the Canadian Institute of Actuaries (CIA) Annual Meeting June 29-30, Vancouver.Google Scholar
Wüthrich, M. V. (2023). Non life insurance: Mathematics & statistics. SSRN Electronic Journal. doi:10.2139/ssrn.2319328.Google Scholar
Wüthrich, M. V., & Merz, M. (2015). Stochastic Claims Reserving Manual: Advances in Dynamic Modeling. Swiss Finance Institute Research Paper No. 15-34. doi:10.2139/ssrn.2649057.CrossRefGoogle Scholar
Yan, J. (2007). Enjoy the joy of copulas: with a package copula. Journal of Statistical Software, 21(4), 121. Retrieved from https://www.jstatsoft.org/v21/i04/.CrossRefGoogle Scholar
Figure 0

Figure 1 Illustration of the discretization methods applied to a gamma($\texttt{a}=5$) severity. The graphs compare the original cdf (blue line) and the discretized (red line) cdf for mass dispersal (top left), upper discretization (top right), lower discretization (bottom left), and local moment matching (bottom right) methods. No coverage modifiers are present.

Figure 1

Figure 2 Class diagram of the lossmodel module. A rectangle represents a class; an arrow connecting two classes indicates that the target class employs the origin class as an attribute. In this case, a LossModel object entails Frequency, Severity, and PolicyStructure class instances. These correspond to the frequency model, the severity model, and the policy structure, respectively. The latter, in particular, is in turn specified via one or more Layer objects, which include coverage modifiers of each separate policy component.

Figure 2

Table 1. Accuracy and speed of the approximation of the aggregate loss distribution using fast Fourier transform (FFT), the recursive formula (recursion), and the Monte Carlo (MC) simulation when varying the number of nodes ($m$), the discretization step ($h$), and number of simulations. The upper table contains the reference values obtained from the closed-form solutions. The lower table reports the execution times in second and the relative errors with respect to the reference values

Figure 3

Table 2. Comparison for different contracts of aggregate and gemact implementation of the aggregate loss distribution computation via FFT. When there are no aggregate coverage modifiers, reference values are given in addition to estimated ones. For the XL and the XL w/agg. contracts, individual conditions are $c= 1 000 000$ and $d= 10 000$; for the SL and the XL w/agg. contracts, aggregate coverage modifiers are $u= 1 000 000$ and $v= 50 000$. Execution times are expressed in seconds

Figure 4

Figure 3 Class diagram of the lossaggregation module. A rectangle represents a class; an arrow connecting two classes indicates that the target class employs the origin class as an attribute. In this case, a LossAggregation object entails Margins and Copula class instances.

Figure 5

Table 3. Accuracy and speed of the cdf calculation using the AEP algorithm and the Monte Carlo (MC) simulation approach for the sum of Pareto random variables coupled with a Clayton copula for different dimensions and quantiles. The upper table contains the reference values from Arbenz et al. (2011). The lower table reports the execution times in seconds and the relative errors with respect to the reference values. The time column on the left is about the execution times of gemact, while the time column on the right (labeled with a *) lists those of the original manuscript

Figure 6

Figure 4 Sensitivity analysis of the cdf at four quantiles $s$ calculated using the AEP algorithm (n_iter = 7) for different copula models, dimensionality, and underlying degree of dependency. The values of $s$ are 1.25, 1.5, 1.75, and 1.95 for the bivariate Gaussian copula (left plot) and 2, 2.3, 2.65, and 2.85 for the three-dimensional Clayton (right plot). Each solid black line indicates the values of the cdf for a given $s$, as the respective parameters $\rho$ and $\theta$ change. The results of the AEP algorithm correspond to those of the Monte Carlo (MC) simulation approach, using $10^7$ number of simulations. The dashed red line represents the average absolute difference between the two method cdf values, calculated across the four quantiles.

Figure 7

Figure 5 Example of events of a non-life insurance claim.

Figure 8

Table 4. Parametric assumptions of the CRMR. In the upper table, we show the parameters for $Z_{h;i,j}, \psi,$ and $q$ that are gamma-distributed. The parameters of the structure variables are specified from the user starting from the variance of $\psi$ and $q$, indeed Ricotta and Clemente (2016) assume $\mathbb{E}[\psi]=\mathbb{E}[q]=1$. The estimator for the average cost of the individual payments is derived with the Fisher–Lange. The variability of the individual payments is instead obtained from the company database. In the lower table, we show the parameter for the claim payment number, that is, a mixed Poisson-gamma distribution with $\hat n_{i,j}$ derived from the Fisher–Lange

Figure 9

Figure 6 Class diagram of the lossreserve module. A rectangle represents a class; an arrow connecting two classes indicates that the target class employs the origin class as an attribute. In this case, a LossReserve object entails ReserveModel and AggregateData class instances.

Figure 10

Table 5. Reserve and mean squared error of prediction (MSEP) by accident period for the CRMR and the CHL. The actual reserve and its process error (PE) by accident period are also indicated. Amounts are shown in millions

Figure 11

Table 6. Total reserve estimates, their relative value, as a fraction of the actual value (8599.04), and their coefficients of variation (CoV), for the CRMR and the CHL. Absolute amounts are reported in millions

Figure 12

Table A.1. List of distributions supported by gemact

Figure 13

Table A.2. List of copulas supported by gemact

Figure 14

Figure C.1 Sketch of the first two iterations of the AEP algorithm in the two-dimensional case.

Figure 15

Table C.1. Reserves by accident period for the CRMR and the Fisher–Lange. We also report the actual reserve. Amounts are shown in millions