When an insurance company is looking to improve its use of analytics, especially in its property & casualty (P&C) business, pricing is always near the top, along with underwriting, claims and marketing. One area that is often overlooked is analytics at the agency or distribution channel level.
In an informal poll at a recent presentation on this topic with clients, less than 10% of companies are using anything more complex than a traditional loss ratio analysis in their P&C sales channels. The lack of advanced analytics in distribution is common. Use of calendar-year loss ratio as part of an agent’s sales goals and bonus structures are ideal to use to rank agencies and offer a good starting point for other agent initiatives such as a review of agent business practices.
So how do we get from simple heuristics to a system that is actually predictive of whether an agent will succeed in the future?
A company could have many objectives when reviewing agency data. When actuaries and data scientists work with underwriters and field executives, the most important point is to ensure the model answers the right question. The best predictive modeler at your company is going to struggle to succeed if the business ask is not clearly understood.
Below are a few examples of requests for agency data that would require models that have subtle but material differences in how they are constructed.
Once you have determined what question you are trying to answer, it is important to continue to collaborate with the business areas that will eventually use your model. Their business insights are invaluable in finding the right variables and interactions to include.
Working closely with them will also build rapport and help to gain buy-in of your final model. This will also allow you to gauge the appetite of using a model that may not be overly interpretable. If they are trusting of the process you may be able to hand over the output only, but typically the “why” will still be important.
Once you start modeling, there are many challenges with reviewing agencies that are quite unique and could require adjustments to your models.
Balancing buy-in and usability from the business side with making the model as sophisticated as possible is key. When starting out, some simple one-way or two-way interactions may be the most appropriate model to build until there is comfort with the methodology.
Once you’ve established comfort, using more sophisticated models such as a Tweedie generalized linear model (GLM) with your target variable being loss ratio or nonparameterized models such as a gradient boosting machine (GBM) or neural net could be even more effective, but with loss of interpretability.
A potential way to get the best of both worlds is with a layered GBM. While a traditional GBM can give you the factor importance from each variable, a layered GBM allows you to understand the gain for each interaction as well. This is appealing if you are using the GBM as the final model or to assist with discovering interactions as part of a GLM model.
Even for the best models, there may be business considerations that trump use of the model as the data scientist intends. For example, if a model is built to reduce binding authority for agents with high projected loss ratios, it may be difficult to get buy-in from field executives and general agents if the producer has had a low loss ratio historically. While this may frustrate the modeler, they must understand their role is to add as much value as possible in a typically overlooked area for analytics, which can be rewarding in and of itself.
Ultimately, the best results will come from a combination of using advanced modeling techniques and collaborating with other areas of the company to ensure the right business insights are included in the final product.