Web Service Demonstration of BayesFraud

This page contains a demonstration of HUGIN's BayesFraud solution implemented as a simple web service allowing the user to interact with the system by entering different pieces of information on a nonlife insurance claim. The demonstration is based on the fraud model described in the HUGIN FDM Technical White Paper.

Model Usage

Claim Type

Type of claim

Claim Data

Time of event

Difference between damage and policy start

Difference between claim date and damage date

Policy Data

Gender

Claims history

Residence

Recent policy increase

Claims Handler Observations

Does claimant appear nervous

Pressure for immediate settlement

Model Adjustment

Adjust time of event impact

Weight of Time of event for normal claims

Weight of Time of event for unusual claims

Adjust Gender impact

Weight of Gender for normal claims

Weight of Gender for unusual claims

As an example consider a claim where the observations are that a male person has made a car claim where the event has occurred during night. That is, select car, male and night for the variables Type of claim, Gender and Time of event under the Main tab. For this claim, the system computes the probability of the claim being unsual to be 0.2759. Now it is possible to adjust the model such that this number becomes smaller or larger. This can be done, for instance, by moving the slider associated with Weight of Time of event for unusual cases. If the weight associated with Night is changed to 0.9 (from 0.3), then the probability is computed as 0.5333 and the traffic light shows the color yellow. Increasing the weight of Time of event for unusual cases specifies that a unusual cases almost always occurs duing night. This means that observering an event to change place during night increases the probability of an unusual claim.


Probability of Fraud and Traffic Light

Using the web service interface above we can efficiently and in real time compute the probability of fraud with partial information on the claim. The probability of fraud is displayed both as a probability and as a traffic light. The probability of fraud changes with each new observation. The color of the traffic light may change as information is being entered.

Try to enter the information that a male person from Copenhagen with a history of 2 claims reports a claim of type car where the event happened during the night (the traffic light turns yellow). Notice how the probability of fraud changes with each observation whereas the trafic light only changes when the probability of fraud reaches a certain threshold. Subsequently, enter the information that the policy has been increased resently. This turns the traffic light red.

Press reset to remove the information on the claim made to initialise the web form to consider an different claim.

The Value of Information button identifies the two most informative observations with respect to the risk of fraud. The score is the normalized mutual information between the observation variable and the fraud variable. The scores are normalised with respect to the entropy of the fraud variable.

The Explanation button identifies the at most two single observations that best explains the risk of fraud, if any. The score is Bayes factor of the hypothesis of fraud compared to non-fraud. Bayes factor is the relative evidence in the observation, i.e., the posterior odds are equal to the product of the prior odds and Bayes factor.

The web service is only supposed to illustrate how the HUGIN FDM solution can be used to identify suspicious insurance claims using the model described in the HUGIN FDM Technical White Paper. In a real application, the HUGIN FDM should be customised to the needs and requirements of the insurance company.