Handbook on loss reserving (Q299762): Difference between revisions

This book presents more or less self-contained articles on the basic aspects of actuarial loss reserving. Besides the traditional ``univariate'' methods, related to the analysis of a single risky portfolio, more recent ``multivariate'' methods are also discussed, which aim at simultaneous prediction of losses for several portfolios and take dependencies between them into account. The articles of the Handbook can be grouped into four categories. The first, presented in Table (T1), contains those which introduce the basic notions related to loss reserving (arrows indicate some connections between the articles and outline the structure of the handbook). \[ \begin{matrix} \text{marginal sums} & & \text{development patterns} & & \text{loss ratios} \\ \text{method} & & \text{(estimation)} & & \\ \uparrow & & \uparrow & \nearrow &\\ \text{multiplicative models} & \leftarrow & \text{development patterns} & \rightarrow & \text{tail estimation} \\ & & \text{(basics)} & & \\ \downarrow & & \downarrow & & \\ \text{Poisson model} & & \text{run-off triangles} & \rightarrow & \text{run-off data} \end{matrix}\tag{T1} \] These articles introduce development patterns, present the general multiplicative model for cumulative/incremental losses, the general approach to the estimation of the parameters in this model, and point out that in the Poisson model -- i.e., where the incremental losses are independent and have Poisson distribution -- the marginal sums method and the maximum likelihood estimator provide the same estimates for the ultimate loss and incremental quota parameters. Best practices to prepare loss data for analysis are also discussed, emphasizing the importance of actuarial expert judgment in selecting the most appropriate modeling approach and parameter estimation method for a specific portfolio. The second group of articles, collected in Table (T2), presents the basic methods of loss reserving, starting off from the Bornhuetter-Ferguson Principle, which provides a general framework for a unified presentation and further extension of these methods. \[ \begin{matrix} & & \text{Bornhuetter-Ferguson} & & \\ & & \text{principle} & & \\ & \swarrow & \downarrow & \searrow &\\ \text{loss development} & \leftarrow & \text{Bornhuetter-Ferguson} & \rightarrow & \text{Cape Cod} \\ \text{method} & & \text{method} & & \text{method} \\ \downarrow & & \downarrow & & \downarrow \\ \text{chain ladder method} & & \text{panning method} & & \text{additive method} \\ \text{(basics)} & & & & \end{matrix}\tag{T2} \] Let \(S_{i,k}\) denote the cumulative loss up to the \(k\)th year for the accident year \(i\), and assume that \(S_{i,k}\) is observable if and only if \(i+k \leq n\). Furthermore, assume that there exist \(\gamma_{0},\dots,\gamma_{n}\) with \(\gamma_{n}=1\) such that \(\mathbb{E}[S_{i,k}]/\mathbb{E}[S_{i,n}]=\gamma_{k}\) \((i,k\leq n)\). Let \(\hat \gamma_{0},\dots,\hat \gamma_{n}\) denote the estimates for \(\gamma_{k}\) \((k \leq n)\), and let \(\hat \alpha_{k}\) denote the estimate for the ultimate loss for accident year \(k\) \((k \leq n)\). Then the Bornhuetter-Ferguson predictors are \(S^{BF}_{i,k} = S_{i,n-i}+(\hat \gamma_{k} - \hat \gamma_{n-i})\hat \alpha_{i}\) \((i+k>n)\); and the Bornhuetter-Ferguson principle reads as follows: {\parindent=8mm \begin{itemize}\item[--] Analytical part: The predictors for \(S_{i,k}\) \((i+k > n)\) are \(S^{BF}_{i,k}\) for appropriate \(\hat \gamma_{k}, \hat \alpha_{k}\) \((k \leq n)\). \item[--] Synthetic part: A comparative analysis of loss projection methods based on different Bornhuetter-Ferguson predictors provides information on the impact of using various estimation methods and data sources. \end{itemize}} The loss development, chain ladder, additive, Cape Cod, and panning methods are introduced as applications of the Bornhuetter-Ferguson principle. Numerical examples are given to illustrate the application of these methods in practice, and the statistical properties of the estimators (bias, standard error of prediction, sensitivity to outliers, etc.) are briefly discussed. In the third category of articles, presented in Table (T3), loss projection methods are complemented with modeling assumptions which give a theoretical backing for the estimators, and also outline their domain of applicability. \[ \begin{matrix} \text{Munich chain ladder} & & \text{multivariate} & & \text{paid \& incurred} \\ \text{method} & & \text{methods} & & \text{problem} \\ \downarrow & \swarrow & \downarrow & \searrow &\\ \text{chain ladder method} & & \text{panning method} & & \text{additive method} \\ \text{(models)} & & \text{} & & \text{} \\ & \nwarrow & \uparrow & \nearrow & \\ \text{credibility models} & & \text{linear models} & & \text{log-normal log-linear} \\ \text{(loss reserving)} & & \text{(loss reserving)} & & \text{models (loss reserving)} \\ \uparrow & & \uparrow & & \uparrow \\ \text{credibility models} & \leftarrow & \text{linear models} & \rightarrow & \text{log-normal log-linear} \\ \text{(basics)} & & \text{(basics)} & & \text{models (basics)} \end{matrix}\tag{T3} \] The two families of models considered are the linear models, for which unbiased and optimal estimators -- the so-called Gauss-Markov estimator -- are obtained under standard assumptions on invertable covariance matrices; and the log-normal log-linear models in which the incremental losses are assumed to follow a multivariate log-normal distribution. Then concrete applications of these modeling approaches for loss reserving are presented, for example: {\parindent=8mm \begin{itemize}\item[--] The chain ladder models of Mack and Schnauss are introduced, and their corresponding predictors are shown to be optimal in the sense that they minimize the squared prediction error. \item[--] The paid \& incurred problem, namely that estimates for ultimate losses using data on paid losses vs.\ data on incurred losses often differ from each other, is discussed, and the bi-variate versions of the additive and panning models are introduced. The parameter estimators for the Munich (bi-variate) chain ladder model are also given. \item[--] The general multivariate versions of the additive, panning, and chain ladder methods are stated as possible approaches to estimate losses for several portfolios -- or lines of business -- simultaneously, in a consistent way. \item[--] The credibility models of Mack, Witting, and Hesslanger and Witting are also mentioned, in which the multivariate distribution of incremental losses -- conditional on deterministic or stochastic latent risk parameters -- is specified up to the first two moments, or is postulated to be multinomial. \end{itemize}} The fourth set of articles -- including controlling, solvency II, reinsurance -- provide a broader context for loss reserving as part of the insurance companies' business, portfolio, and risk management processes, as well as capital planning and regulatory compliance. Each article concludes with a bibliography, directing the interested reader to recent publications where further details can be found on the topics presented. The authors also invite the reader to consult ``A bibliography on loss reserving'', available at \url{https://www.math.tu-dresden.de/sto/schmidt/dsvm/reserve.pdf}, for a literature overview.