How machine learning helps insurers increase profitability
The 'sweet spot' for insurers when using machine learning involves gaining a clearer picture of customer behaviors.
For many U.S insurers, particularly smaller and midsize ones, the standard approach to rate changes is to assess various rate scenarios and select the one that best supports business objectives while heeding regulatory requirements. This approach usually does not incorporate a sophisticated view of how a customer would react to the various scenarios.
Historically, the development of these customer behavior models takes a significant amount of effort to build and maintain. As such, insurance executives rely on anecdotal analysis.
Machine learning can help insurers understand customer behavior. But before insurers embrace this technology, they need to consider two central issues: off-balancing books of business and long-term goals.
Off-balancing books of business
Insurers will do a lot to get a general direction of overall rate adequacy. Most of the time when changing component(s) of the rate, the pricing team must incorporate an “off-balance” factor so that the proposed rate balances back to the overall adequacy.
Traditional approaches use historical data to produce the off-balance factor. The problem with this approach is that it doesn’t consider the future shape of the book of business.
The underlying assumption is that whatever rate was offered would be accepted by the customer. Customers that may cancel will be replaced with a new business that looks exactly like the customer who just left. Both assumptions are unlikely to pan out.
This results in the actual change being quite different from the desired change. This problem arises because the pricing team is only considering assumptions that relate to insurance loss and expense costs.
How to find your sweet spot
To better understand the expected impact of a change, the insurer should incorporate customer behavior into the analysis. But what do we mean by customer behavior, and how can it be modeled efficiently and accurately with machine learning methods/? How do we manage the long- and short-term objectives of the overall portfolio?
Simply put, customer behavior is the demand customers have for a carrier’s product. Demand in insurance covers the conversion of new policies (acquisition) and the retention of existing policies. Acquisition is oftentimes further broken out into awareness, consideration and purchase models whereas retention is broken out into renewal notification and mid-term cancellation models.
Short and long-term goals
While an overall rate increase may help achieve immediate profit targets over the short term, how might it affect demand and the business longer term? Suppose you’re considering four different rate scenarios for an upcoming rate filing. By considering the impact of customer demand, you can track the true profit impact of each scenario over multiple periods as the book begins to reshape. Without this insight, you may be left in “no man’s land” when forecasting future performance.
Baseline data
The data used to build these models can be quite extensive and are often categorized as follows:
- Attributes and attitudes: What is the customer like?
- Environmental: What are external influences?
- Influences: What have you done to the customer?
- Status changes and triggers: What has changed and when?
Demand models are built at either a portfolio or key sub-segment level to accurately reflect customer behavior. In addition, unlike most models, demand experience changes at a much faster rate, which requires more frequent model updates. Building and updating these models requires significant work. All of this is possible using traditional modeling tools such as a generalized linear model (GLM). But these traditional tools also require skilled resources to develop and maintain. This is where machine learning techniques are extremely useful as they automate this process.
Modeling customer behavior
Traditional modeling tools like GLMs are often referred to as parametric solutions. This means that the resulting prediction of behavior can be expressed as an equation, like a standard rate order calculation. This form is useful because it is easier to interpret. However, these solutions can be more time-consuming to create and maintain.
Machine learning models are referred to as non-parametric solutions. In this form, the prediction behavior is expressed as a decision tree (or a series of decision trees).
The popularity of these techniques has increased because they generally have greater predictive power than traditional modeling tools, and they incorporate automation more explicitly so that they are faster to build and easier to maintain.
One of the most common pitfalls to watch out for is overfitting, a term used when the model describes the experience data well but does a poor job of predicting future outcomes (i.e., overfitted models are “stuck in the past”). The likelihood of overfitting is extremely high when machine learning models are being used. This is further complicated because it is common to incorporate a layer of automation when updating model results.
Therefore A/B testing is recommended through sampling which establishes the modeling, validation, and testing data sets. The modeling data is then folded to further mitigate the risk of overfitting.
Another challenging pitfall of machine learning model is interpretation. We can look at the tree, but keep in mind we built a complex series of recursive trees. So it is extremely difficult to articulate why a customer is likely to have higher or lower demand for the insurance product. A couple of strategies to manage this are:
- The factor importance output. This allows you to identify which factor is most influential in the model. Care must be taken when interpreting this result because it only is telling which factor is important. It is not saying whether that importance is associated with either high or low demand. A proprietary algorithm that identifies the most important factors and the most important combinations of factors identified by the machine learning algorithm is critical to better understanding the underlying structure.
- Partial dependency plots. This allow the interpreter to explain complex models into more basic statements. This is quite useful when trying to get a sense of what the model is saying, however, it is still fundamentally an approximation.
By using these interpretation techniques along with the machine learning methods, a company can articulate which customers are likely to have a higher demand for their product (the machine learning output) and why the machine learning tool identified those customers (what are key factors, key profiles and how the model could weigh the different factors within a profile).
Agility in the marketplace
Customer behavior is an important assumption that is critical in assessing the impact of the product and pricing. As noted, traditional off-balance approaches may lead to insufficient rates, resulting in future increases that could materially impact the shape of the book. However, unlike cost and expense models, the assumptions in these demand models become out of date rather quickly so they require more maintenance.
The value of machine learning algorithms as a critical tool for behavior analysis represents a step up from standard modeling techniques. Not only do they produce more accurate predictions; they can also be set up for easier maintenance. While these are substantial benefits, there is still much to be concerned about due to the black-box nature of such tools.
The benefit of customer behavior models results in better estimates of rate impact, quantifying the short- and long-term value of each customer and creating a robust framework for assessing existing products as well as identifying the art of the possible in future products.
Insurers can find it difficult to collect data on new business quote acceptance, renewals and policy cancellations. By using machine learning, however, they get a clearer picture of how customers — both existing and new — are likely to respond to rate changes. This can provide greater insight when determining rate changes and achieve greater profitability — a grand slam for their business.
Serhat Guven is the managing director of WTW’s Insurance Consulting & Technology and Global Proposition Leader for Personal Lines Pricing, Product Claims, and Underwriting. He and his team are responsible for the delivery of consulting services and technology solutions that are uniquely designed to help insurers respond to significant industry trends; as well as provide carriers support in core areas that are fundamental for effective business management and profitability.
See also: