The Jumper Dilemma – Why is Claims Modeling So Hard?

Claims Modeling CLCM

Claims Modeling CLCMClaims modeling is gaining traction right now.  Attend a conference, or talk to some data science / modeling staff, and you’ll likely hear about some current or impending efforts to build a claims model.

“What kind of claims model are you building?” is a natural line of questioning.  If you’re talking to the same groups of people that I am, you’ll hear three general answers:

  1. A jumper model
  2. A triage model
  3. A reserving model

I’d like to discuss each of these, in turn, in the context of an observation that’s becoming more and more clear to me:  The number one mistake being made in claims modeling is that modelers are with shocking frequency attempting to answer the wrong question.

The Jumper Claim Model

Let’s start with the Jumper model.  This model attempts to answer the question: “Which claims are most likely to jump by more than $50,000 from the initial case reserve estimate at 30 days?”  To build this model, the analyst assembles claim information and examines the case incurred amounts for each claim at 30 days and at some future date, with the target variable being a binary Yes/No if the claim met the jumper definition.  There are several big potential pitfalls with this approach:

  1. The jumper criteria (ie $50K increase from 30 days to ult) must be determined before modeling begins
  2. The jumper criteria is almost certainly not optimal
  3. If case reserving methods change (say as a result of the findings of the modeling), this can invalidate the model’s predictive accuracy
  4. There is no prescriptive value in this model; merely identifying claims likely to be jumpers does not say anything about what to do with those claims to change the future. 

The Claims Triage Model

With triage models, the modeler is attempting to answer the question: “Which claims are likely to be more complicated or higher severity, and should be assigned to more experienced adjusters?”  Triage models are prescriptive models, in that they attempt to prescribe a future action to help mitigate or reduce future payments.  In that light, triage models can also be built to indicate when and where particular loss control or settlement actions should be performed.  In my opinion, this approach has a good probability of being successful at mitigating claim costs, but there are still a few potential pitfalls:

  1. For many carriers, there is no coding of past loss control procedures, so there’s just nothing there to model on.  Carriers need to start coding their loss control efforts in a regimented way for some time to gather the data needed to model the effectiveness of those actions
  2. In implementation, many triage models are used primarily to assign complex / high severity claims to more senior claim adjusters.  This is certainly a smart move, but it’s not scalable.  What is that senior-level, experienced claim adjuster going to do that a junior adjuster wouldn’t do?  Wouldn’t it be great if the model could tell us that?  (see also the first point)
  3. As with the jumper approach, if the claims adjustment process changes as a result of the triage model, and the triage model is based on the case reserves, this can effectively break the model when the claims department starts changing behavior

The Claims Reserve Model

Reserving models are in the minority.  Very little of the claims modeling efforts are being invested in building more a more accurate reserves picture.  With reserving models, the modeler is attempting to answer the question: “What is the likely future ultimate value of a reported claim?”  This is a much simpler question, with quite a few ready-made applications.  It would be hard to argue that the modeler is asking the wrong question here.  Instead, the biggest potential pitfalls are in using case reserves as a model input:

  1. Again, what if the case reserving process changes, particularly in reaction to the model?  This breaks the model.
  2. Typically, these models reveal that one of the most important predictors is the case reserve itself.  How can this be executed?  Does this mean we should fire the modelers and hire/train better Claims staff?

The Ideal Claims Model

This is not to say that attempting to build a claims model is an exercise in futility.  Ideally, claims models should

  • Be based on objective information (this does not include case reserves)
  • Include all available information – exposure detail, claims detail, transactional (time series) data, free-form text, external data, etc.
  • Be flexible enough to answer multiple questions
  • Provide a springboard to enable new actuarial and analytics projects

CLCM Exemplifies the Ideal Claims Model

For the past several years, we’ve been using a different approach at claims modeling that incorporates these ideals:  the Claim Life Cycle Model (CLCM).  Over the course of the next 12 weeks, I’m going to be interrogating the Claim Life Cycle Model process from a number of different angles in an attempt to explain its strengths, capabilities, and limitations.  I’ll compare and contrast with the three more common claims modeling approaches I introduced above.  My aim in this is to provide some support for analysts using other claims modeling approaches to avoid some of the pitfalls commonly encountered in claims modeling efforts, and ultimately to convince a few of you that using a Claim Life Cycle Model approach may ultimately be the best way forward to help you achieve your goals in claims modeling.   


To learn more about the Claim Life Cycle Model approach, and how you can employ it to build better claims models for your organization, contact me at Bret.Shroyer@cgconsult.com.