Presented by Tom Keller, FSA, MAAA Craig Magnuson, FSA, MAAA Moderator Kathy Thomas
Predictive Modeling >>> Forecasting Claims Costs
Proposed: First Law of Predictive Modeling
Predictive models can be…
• Quantitative • Non-quantitative • Some of both
It’s hard to think of anything we do that’s not predictive modeling.
that purpose will probably be…
Better Financial Results
. predictive modeling is just another answer in search of a question
Unless you work for a charitable organizations.Proposed: Second Law of Predictive Modeling
Without a purpose.
Proposed: Third Law of Predictive Modeling
To be worthwhile.
Steps in Building a Decision Metric Predictive Model
Identify the decision to be made
Define what you’d like to know to make the decision
Identify the predictive models needed to calculate the decision metric
. a quantitative predictive model must be part of calculating a metric used to choose between a universe of possible decisions.
Let’s look at the Small Group renewal rating decision…
. directionally correct quantitative predictive model is usually better than no model at all.Proposed: Fourth Law of Predictive Modeling
P of Lapse) ¾ Formula for the Expected Financial Results
• P of Lapse X Financial Impact of Lapse.P of Lapse) X Financial Impact of Retention
. plus • (1.Defining the Decision Metric
¾ Small Group regulations restrict range of possible decisions ¾ For any possible decision. there are only two possible outcomes
• Takes the increase and stays with you (Retain) • Doesn’t take the increase and leaves (Lapse)
¾ The financial consequences of the two outcomes would be different ¾ The probabilities of the two outcomes would most likely be different. but they must add up to 100%
Actuarial Science 101A
For any possible decision…
¾ P of Retention = (1.
minus • Marginal expenses
¾ Must predict Gross Margin impact if group lapses and if it doesn’t. minus • Claims.
.Which decision would you make?
a) b) c) d) e) Stop talking to actuaries Get out of the Small Group business Ask the Sales guys what to do Increase my donation to the PAC fund the one with the best Expected Financial Result
One Definition of Best Expected Financial Result
¾ the one with the largest Expected Gross Margin ¾ where Gross Margin is defined as
. Milliman Syed Mehmud. Milliman
. Actuarial Assistant. FSA. Principal & Consulting Actuary.a subject worthy of its own workshop
“A Comparative Analysis of Claims-Based Tools for Health Risk Assessment”
by Ross Winkelman.And that means we need to be able to predict…
¾ Premiums ¾ Claims ¾ Marginal expenses
Predictive Models for Claims
You’re probably familiar with a few of them…
¾ Your rate manual ¾ Your underwriting debit system ¾ Last year’s loss ratio
• large claims pooling • credibility factors • pooling accident claims
¾ Underwriter judgment
Any input that improves predictive accuracy is fair game… ¾ ¾ ¾ ¾ ¾ diagnosis codes procedure codes Rx history prior period claims experience your CEO’s mother’s birthday (Yes. he has one!)
P of Lapse
.The Shortcoming of Most of Them
… for the Small Group Renewal Rating Decision
¾ predictive accuracy is highest where the payoff is lowest ¾ vice versa
Predicting the Financial Impact of Lapse and Retention is only half the battle…
…what’s the other half?
no one is going to find it ¾ An imperfect quantitative model that is directionally correct can provide useful and surprising insights (Fourth Law) ¾ Everyone in this room has experiences and intuition that could be used to build a directionally correct predictor of lapse behavior
What We All Know
¾ Some groups are going to lapse no matter what rate you offer (P of It Really Doesn’t Matter) ¾ Groups that don’t shop for another offer won’t lapse if they weren’t going to lapse no matter what
• • • • • • • • • There is some rate increase threshold below which a case is unlikely to shop The larger the rate increase. switching becomes a virtual certainty
¾ Not all cases that shop get a better offer
¾ Not all cases that get a better offer switch carriers
. the more likely the case will shop At some point. the more likely the group will switch At some point.A Few Observations About P of Lapse
¾ Whether you realize it or not. renewal rating practices incorporate someone’s beliefs about lapse behavior ¾ If there is a perfect quantitative model. shopping becomes a virtual certainty Groups that don’t get a better offer won’t switch unless you’ve done something to antagonize them Healthy groups are more likely to get a better offer than unhealthy groups Some groups are going to go right to the top of every carrier’s rate corridor There is some threshold below which a group will find it too much trouble to switch The greater the difference between your offer and theirs.
P of It Really Doesn’t Matter)]
So for the Small Group Renewal Decision Metric Predictive Model. we need predictive models for…
¾ Components of Gross Margin Under Retention and Lapse
• Premiums (with rate increase if retained. without if lapsed) • Claims (probably what you thought we were going to talk about) • Marginal expenses (Some easy.Actuarial Science 101B
Turning what we all know into a lapse predictive model…
P of Lapse= (P of It Really Doesn’t Matter) + [(P of Shopping) X (P of Getting Better Deal) X (P of Accepting Better Deal) X (1. some hard)
¾ Components of a Directionally Correct Lapse Model
• • • • • • • Baseline lapse rate (P of It Doesn’t Matter) Threshold of Shopping Shape of (P of Shopping) above Threshold (P of Getting Better Offer) (good market intelligence) Magnitude of Better Offer (good market intelligence) Reluctance to Change Threshold Shape of (P of Accepting Better Deal) above Threshold
• They all boil down to structured trial and error. you have a system of equations with
• • • • • some constants some constraints one independent variable .the rate increase several intermediate dependent variable the dependent variable you are trying to use the rate increase to optimize – the Expected Financial Result
• There are a variety of optimization tools you can try. what would you do with it?
• In math terms.
Proposed: Fifth Law of Predictive Modeling
Always start with the most simple approach
.Suppose you had all that.
The Simple Approach
¾ Determine range of possible rate changes ¾ Divide range into a fixed number intervals ¾ Financial Consequences of Lapse same for all possible rate changes ¾ Calculate Expected Financial Result for each interval ¾ Pick rate change that produces largest Expected Financial Result
Case by Case Assumptions
Minimum Increase Test
Maximum Increase Test
Optimal Increase Test
Results for All Cases
Remember when the Sales guy said that a big rate increase will only drive away the healthy groups?
there was something to that
(No salesmen will call.Keller@MagnumActuary. send us an email:
Tom.It’s harder to make a profit with a tight rate corridor than a broad rate corridor…
you knew that
To obtain a free copy of the model used in this presentation.)
ignores impact on new business
Shortcomings of the Simple Approach
¾ It’s not that simple ¾ Completely deterministic.Proposed: Sixth Law of Predictive Modeling
There’s always a better way. little insight on risk ¾ Independent decisions may not produce optimal result ¾ Focused solely on in force business.
Tools to Deal with Shortcomings
¾ Old-fashioned trial and error sensitivity testing ¾ Monte Carlo simulations (Crystal Ball. @Risk)
• Assign probability distributions to predictive model intermediate outputs • Produces probability distribution for Financial Results • Can provide useful information concerning where accuracy is most important
¾ Deterministic optimization with constraints (Excel Solver) ¾ Optimization under uncertainty with constraints (OptQuest)