Casualty Loss Reserve Seminar (CLRS)
September 15-17, 2013
Boston Marriott Copley Place
Workshop 1: Loss Reserving Boot Camp
What better way to jump start or refresh your loss reserving skills than in this one day loss reserving boot camp. Participants will be part of a loss reserving team evaluating and estimating reserves from data organization to selecting the best estimate. Complete with hands on learning, groups will be divided into small teams of five to seven professionals who will work with experienced loss reserving professionals and use realistic loss and expense claims data.
Part I – Basics of Loss Reserving
The basics of loss reserving begins with the "CAS Statement of Principles Regarding Loss and Loss Adjustment Expense Reserves," including the definitions and considerations that guide the actuary. Following the discussion for the "Statement of Principles," participants will work through, step-by-step, data collection and segregation, the organization of data for loss reserving, and the calculation and selection of loss development factors. Guided by experienced loss reserving actuaries, each participant will work independently and within a small group in Excel with realistic sets of data.
Part II – Basic Loss Reserving Methods
Using the data and selections from Part 1, participants will learn 3 basic loss reserving methods - loss reserve development, expected loss ratio, and the Bornhuetter-Ferguson method. These methods will then be used with multiple data sets to estimate the ultimate losses. The strengths and weaknesses of each method in a variety of situations be discussed. Additionally, methods of estimating reserves for defense and cost containment expenses, adjusting and all other expenses, and salvage and subrogation will also be presented. Participants will continue to work through the hands on examples with guidance from experienced actuaries to gain a working knowledge of these methods.
Part III – Comparison of Methods and Selections
Using the results of the methods in Part 2, the methods will be compared side by side. The participants will be exposed to various diagnostics that can be used to aid in understanding the results. Within the small group, participants will discuss the results of the methods, the indications and observation in the diagnostics and make their selections. The range of reasonably possible outcomes will be discussed and participants will make low, high and central selections.
Part IV – Final Answer
Working in the small group, the team will summarize their selections from Part III, calculate IBNR and total reserves and make their final selections. Each group will present a summary of their analysis and the groups will compare the differences in estimation between the groups.
A laptop, with Excel 2007 or newer will be required for the workshop. Wireless connectivity may be helpful. This Workshop will be limited to a maximum of 35 participants in order to optimize audience participation and learning experience.
Workshop 2: "R" is for Reserving – IntroductionPart I - Beginning R
The workshop will begin with participants learning about the R programming language. The pros and cons of R versus excel will be discussed. Participants will be asked to have their laptops loaded with R so that they can experiment. Simple examples with data sets will be used.
Part II - "R" is for Reserving - Introduction
In this session of the workshop, the basics of "R" will be discussed and experimented with.
Part III - Using the Chain Ladder in R
In the afternoon sessions of the workshop, participants will apply the knowledge of R to reserving methods. Our speaker will show how R can be used in conjunction with Chain Ladder methods.
Part IV - Multivariate Regression Models for Reserving - An R Package for Loss Reserving
This afternoon workshop session will explore the treatment of loss reserving as a multivariate regression problem. We will discuss how to properly structure the data, use of statistical diagnostics to assess the goodness of fit and model selection and how to use the model to predict future realizations. The model is presented using R. All of the underlying code is publicly available for use. Examples using publicly available data will be shown.
Workshop 3: Stochastic Loss Reserving with Bayesian MCMC Models
Part I - Bayesian Concepts, Computation, and Software
This first session briefly reviews fundamental concepts of the Bayesian data analysis paradigm and then introduces the MCMC (Markov Chain Monte Carlo) simulation technique. Participants will solve a simple motivating example first using only R, then by using the JAGS (Just Another Gibbs Sampler) software.Part II - Bayesian Actuarial Case Studies
Participants will be guided through a sequence of case studies including Bayesian versions of regression, GLM, loss distribution analysis, exponential trend analysis, and the chain ladder model. Hierarchical model structure and the Bayesian hierarchical growth curve model may also be introduced. Participants will practice data analysis using JAGS to prepare for the afternoon sessions but will also learn how to "think Bayesian" in practical actuarial work.Part III - Validating Loss Reserving Models
Modern Bayesian data analysis offers unprecedented flexibility in specifying models appropriate for the data-generating process. However, model criticism and validation are needed to ensure that the model captures salient features of the data and predicts well out-of-sample. This session outlines a model validation methodology and discusses model features that can improve predictive accuracy. Participants will test the performance of the popular Mack and Bootstrap Overdispersed Poisson models on actual data from the CAS Loss Reserve Database.Part IV - Stochastic Loss Reserving Using Bayesian MCMC Analysis
In this final session, participants will explore various stochastic loss reserving models including:
- A Bayesian version of the chain ladder reserving model.
- The correlated chain ladder (CCL) model that allows for correlations between observations from successive accident years.
- The correlated incremental trend (CIT) model that employs a skewed distribution with support over the entire real line to accommodate the possibility of negative incremental paid observations, and allows for payment year trends on paid data.
- Hierarchical model structure and the Bayesian hierarchical growth curve loss reserving model.
The CCL and CIT models will be tested using actual data for several insurers from the CAS Loss Reserve Database.
Basic familiarity with R is a prerequisite. Participants must bring their own laptops with R, RStudio and JAGS installed. Within R, the "rjags" and "ChainLadder" packages should be installed. Attendance will be limited to the first 25 registrants.