Predictive Modeling Capability - No Longer an Analytics “Gold Standard”
February 20, 2019| Von Upendra Belhe
P/C General Industry
Region: North America
As a company leader, you no doubt value predictive models as an important asset or aspiration for your company. Many insurers have developed predictive models already, perhaps for underwriting, pricing and/or claims. They know it’s a competitive advantage and worry that they should be doing more. But do carriers have the correct expectations? Are predictive models the right choice?
Analytics is a journey and not a destination. Maturity is characterized by stages; each stage a necessary building block for the next. The ability to model evolves with extreme data familiarity and using it to question correlations that impact the business. The actionable insights generated through this evolution may have a profound impact on how you conduct your business. As trends, combined ratios and renewal success rates evolve, so will - or should - your models. Data from your business will help you observe trends and validate business insights - and continually test whether or not your model is performing as intended.
No matter where a carrier is in their journey, the expectation of 100% accuracy is flawed. When the stakes are low, it may be fine to use complex predictive models and base decisions on those results. However, if there is a need to use data as the basis for strategic corporate decisions, or, for decisions regarding an entire portfolio, traditional predictive models may not be the answer.
Predictive modeling, which predicts a value or outcome, uses statistical or machine learning methods. Machine Learning (ML), in fact, can be used more broadly to efficiently recognize patterns in data. Not only can it be used to identify clusters of data that may benefit from specialized actionable insights, but the algorithms learn from the data, which help to identify changing patterns and allow executives to alter strategy as needed. ML contemplates that one strategy does not fit all.
Traditional analytics rely heavily on having clean data sets before developing models. However, it is actually better to work with large data sets, even if a little messy (e.g., words spelled incorrectly or abbreviated differently in a document), versus curating small amounts of clean data. ML allows us to obtain insights from large sets of “not so clean” data, and thus gives us a better understanding of the business. Imagine the possibilities ML provides! For instance, what if you were able to leverage profiles, actions, behaviors, and information about your customers across all channels, and continuously learn from it as the market changes over time. It can be very powerful.
You may be thinking, “How does ML work?” Machine learning algorithms use computational methods to “learn” directly from data without relying on a predetermined equation as a model. The algorithms adaptively improve their performance as the number of samples available for learning increases. A myriad of ML techniques are available through open source software. Although there are more sophisticated methods, we like to use easy to interpret methods, for example, Decision Trees.
A Decision Tree allows our executives to decide if they want to focus on their strengths, weaknesses or both when formulating a strategy. More than its sophistication, ML’s objectivity towards data and independence from preconceived notions of statistical significance, helps us to use algorithms to better understand potential revenue/profitable opportunities. It also allows us to find the combination of attributes (both internal and external) that lead to a particular result.
Here are some examples of where we used a Decision Tree’s hierarchical tree-like structure to visualize the rules that lead to a particular outcome.
- Of the many factors available within a given market segment, which have the highest impact on our ability to bind a risk?
- Given a context, what are the words within claims documents that are associated with greater severity?
- Within a specific market, what range of premium has performed better from an underwriting metric point of view than others?
Using our business knowledge, we utilize this information to think about what can be done differently to improve results. More than answering questions, we use ML to gain insights that drive action - especially insights that makes us rethink of our current strategies and pushes us in a new direction. Maximizing actionable insights, from internal and external data on our data platform, is important to our data-driven success.
Predictive modeling, which used to be something that carriers strove for, isn’t the destination anymore. From the matured carrier’s perspective, as well as our own, we are observing that analytics driven success is about actionable insights using all available internal and external data. Impactful data driven business insights are not obtained by perfecting small data sets to predict a tactical point estimate, but rather by utilizing as much data as possible to identify a direction to address strategic priorities for the carrier’s business.
Learn more about our viewpoint and experience in discovering insights through machine learning by contacting Gen Re’s Business Data and Analytics team through your Gen Re account executive.