We will be using a dataset of personal auto insurance claims which include an outcome label of Fraud / Not Fraud. This dataset consists of 15,420 claims from between January 1994 and December 1996. The data consist of 6% fraudulent claims and 94% legitimate claims. It contains 6 ordinal and 25 categorical features in addition to the outcome variable. Features are as listed below.
| Column Name | Description |
|---|---|
| Month | Month of year in which the claim occurred |
| WeekOfMonth | Week within the month on which the claim occurred |
| DayOfWeek | Day of the week on which the claim occurred |
| Make | Make of the vehicle in the claim |
| AccidentArea | Urban classification of area where the accident occurred (e.g., urban, suburban, rural) |
| DayOfWeekClaimed | Day of the week on which the claim was reported |
| MonthClaimed | Month of year in which the claim was reported |
| WeekOfMonthClaimed | Week number within the month in which the claim was reported |
| Sex | Sex of the claimant |
| MaritalStatus | Marital status of the claimant |
| Age | Age of the claimant |
| Fault | Who was found to be at fault in the accident |
| PolicyType | Type of policy the policyholder has |
| VehicleCategory | Type of vehicle involved in the accident |
| VehiclePrice | Original cost new of the vehicle involved in the accident |
| FraudFound | Whether the claim was determined to be fraudulent or not |
| PolicyNumber | Unique insurance policy identifier |
| RepNumber | Unique identifier of the claims adjuster who handled each claim |
| Deductible | Amount the insured must pay before the insurer begins paying for repairs |
| DriverRating | Unknown |
| Days:Policy-Accident | Days between policy inception and the reported date of accident |
| Days:Policy-Claim | Days between policy inception and the date the claim was reported |
| PastNumberOfClaims | Number of prior claims on the insured's policy |
| AgeOfVehicle | Age in years of the vehicle involved in the accident |
| AgeOfPolicyHolder | Age group of the policyholder |
| PoliceReportFiled | Whether a police report was filed at the time of the accident |
| WitnessPresent | Whether any witnesses were present at the time of the accident |
| AgentType | External |
| NumberOfSuppliments | Number of additional payments issued on the claim |
| AddressChange-Claim | Time between most recent policyholder change of address and date the claim was reported |
| NumberOfCars | Number of vehicles on the policy |
| Year | Year of vehicle manufacture |
| BasePolicy | Base policy type (liability-only or full coverage) |
- Claim: A report of an accident to an insurance company for the purpose of payment for damages
- Claimant:
- Policyholder:
- Coverage:
- Liability Coverage:
- Comprehensive/Collision Coverage:
- Deductible:
- Loss Cost:
- Adjuster:
- Loss Date:
- Reported Date:
The goal of this project is to minimize overall loss costs to the insurer incurred by fraudulent claims. For our purposes, we will assume the following:
| Predicted Fraud | Predicted Not Fraud | |
|---|---|---|
| Fraud | $400 | $2500 |
| Not Fraud | $3500 | $2000 |
This confusion matrix should be used to assess the relative business value of any proposed model.