Predicting Legal Claims Costs

The Opportunity

The senior general counsel of a company with tens of thousands of employees who sometimes work in hazardous conditions was aware that the company had extensive data on employee injury claims. He was also aware that decision-makers in the Legal department often make decisions that are influenced by bias. They might, for example, respond to a new case based on the outcome of the last one in which they were involved.

He saw an opportunity for the department to make better, more objective decisions about resolving employee injury claims by using data and predictive analytics.

Our Approach

Our machine learning approach focused on the predicted cost of the legal resolution of the claim, either in court or settled out-of-court.

Inputs into the model included basic litigation attributes (personnel involved, court, jurisdiction, outside counsel), and facts about the individual plaintiff (age, severity and type of injury), among other factors.

The model was deployed in a web-based application. In addition to an estimated dollar figure to resolve the case, attorneys are presented with a bar graph of the probability distribution of the estimate. This helps reinforce that the outcome is a matter of statistical probability. Attorneys can also perform “what if” analyses—what would happen to the estimate if the case goes to trial, what if the jurisdiction changes, and so forth. That helps to guide decisions about what might be the most favorable outcome of the case. Users can also query and filter the data to look at comparable cases.

The Impact

Because staff attorneys began to think of cases as having a range of outcomes, this is how we validated our predictions. In addition to the point estimate, results are presented in a set of ten ranges (e.g., between $200k to $350k, $350k to $500k, etc). We held out a sizeable number of recent cases and predicted their outcomes and found that our predictions were either spot-on in the range, or off by one range, 82.7% of the time for direct settlement and 71.7% of the time for litigation.

Adoption by professionals who are not accustomed to using data in their work is a challenge. Use of the system was not mandated but system login records showed wide adoption.

The senior counsel said, “We are already much more data-driven than we were, and that’s a good thing. Our goal is not to base 100% of our decisions on data; our experience is still a valuable guide as well.”