Developed a two stage model using dcrypt to determine litigation propensity by leveraging structured and unstructured data
The Business Challenge
A leading insurer wanted to bring improvements to litigation management by making data-driven decisions. Specifically, they wanted to get an in-depth view into the drivers of litigation, what types of litigation were driving up costs and expenses, in what lines of business, and in what types of claims. Their objective was to identify claims most likely to be litigated by third-parties or the insured, and to predict the estimated litigation spend.
This information helped bring improvements in litigation management by appropriately allocating skilled adjustors on claims with high litigation propensity and spend.
Fractal Analytics used dCrypt to develop a two-stage solution with application of machine learning and text mining techniques to more accurately predict litigation propensity. This included:
- Leveraging a predictive model using text mining (RF and SVM) on unstructured data.
- Deploying a predictive model using probabiliites from text mining as an independent variable along with structured variables to arrives at the final ensemble model.
- Estimating litigation spending using business rules identified from cross-validated decision tree analysis.
- Identifying litigation propensity and spending as early as 15 days after the first notification of loss..
Fractal built advanced machine learning models for text mining using our IP dCrypt solution to yield potential savings of $0.97M to $1.93M annually by actively pursuing approximately 1,200 claims.
- Claims with low litigation spend can potentially be considered for early settlement avoiding litigation cost and expenses
- Claims with high litigation spend can potentially be assigned to appropriate counsel thus minimizing legal spend and claim payout
- Claimants not at fault are more likely to litigate as they may have a stronger case.
- Claims with high subrogation propensity have a high change of entering into litigation.
- As the severity of accident increases, the likelihood of litigation increases.
- A total loss claim is less likely to be litigated.
- The higher the age of the claimant, the higher the average litigation spend.
- The litigation propensity model is able to concentrate 56% litigations at 20% population and 69% litigations at 30% population.
- Additional expense incurred for reviewing and actively pursuing false positive cases are also featured in the benefits calculation.
- Fractal harnessed the power of unstructured data with structured data elements to build the final solution for predicting litigation propensity using our IP dCrypt solution.
- Using unstructured data, Fractal used predicted probability from the text mining model as one of the independent variables along with 10 other structures variables to determine the final ensemble model.