Casualty Actuarial Society Public Loss Simulator

The Casualty Actuarial Society Loss Development Simulator generates synthetic loss histories from frequency, severity, lag, and case reserve distributions specified by the user

Background:

In principle, historical loss experience used for loss reserving could be extremely detailed, including all correspondence and reports in the insurer’s claim files.  In fact, such detailed information is used at one stage in the reserving process, when an estimated case reserve value is assigned to each claim.  At an aggregate level, however, insurers typically collapse the vast amount of information in their files into certain numeric quantities, all derived from the date and amount of each event such as a payment or a change in case reserves.  This aggregate data may be as detailed as a complete listing of all such policy and claim transactions.  More commonly, it consists of one or more “triangles” such as matrix of paid losses by month incurred versus month paid.

Insurers have developed numerous methods to estimate reserves from aggregate data.  Explicitly or implicitly, such methods amount to models of the loss process and estimators of the parameters of such models.  The appropriateness of a particular reserve method to particular loss data depends on how well the underlying model fits the data and often – because individual data sets may be small – on how well the model fits other data sets believed to be similar, such as those involving the same line of insurance over approximately the same time period.

Since suitable similar data sets may be unavailable, insurers need other ways of evaluating their reserving methods.  One way that has proven useful is Monte Carlo simulation.  If the insurer can generate numerous synthetic data sets with characteristics believed similar to those of the case at hand, it can apply several candidate estimators to these data sets and see which estimators are most efficient at forecasting the runoff of losses from a particular date, and what bias if any they have in their predictions.

It has been the task of the LSMWP to produce a software simulator that might aid insurers and researchers in investigating such questions.  Our goals are for the simulator to be directly applicable to most cases and easily customizable for others.  To achieve these goals the simulator must be available in both executable and source code format, be able to run on multiple platforms, and be written in one or more languages that are accessible, free or inexpensive, and widely understood in the actuarial community.

After extensive discussion we have constructed a prototype simulator that confirms the feasibility of the project and offers guidance as to one possible set of features and user interface.  This prototype simulator is written in the language APL.  While APL is an excellent language for actuarial work and is particularly well suited to prototyping, it is not as easily accessible nor as inexpensive as many other languages, it uses a special character set that makes printing, transmission, and communication of the source code difficult, and its user community is not as large as those of Visual Basic, R, J or similar recent languages.  Therefore we wish to create a simulator with capabilities equal to or greater than those of the prototype, but not necessarily a direct copy of it, in a language better suited for distribution of open source software than APL.

We recognize that the italicized portion of the preceding sentence implies some vagueness in our requirements.  In this document we explain what we consider most important in this project and therefore what will be given greatest weight in evaluating proposals.  There are several documents on the CAS web site, HUwww.casact.org/research/lsmwpUH, that give further insight into our goals and that should be consulted to obtain a full understanding of this project.