The Big Picture
Insurance companies are organic institutions that evolve over time, especially when it comes to analytics. The industry trailblazers in using predictive analytics faced new challenges and had to find their own solutions, but they experienced the benefit of first-mover status. Because the technology available was simultaneously evolving at a rapid pace, competitive advantages were found in investing in cutting-edge expertise and tools.
However, this situation is changing because predictive analytics is maturing within the insurance industry. Many of the challenges associated with predictive modeling are now well known, but most insurers are meeting these by relying on simple extensions of previous solutions to previous challenges. Now is a good time to step back and reevaluate the end-to-end process of getting value from predictive models, and to ask if we should be doing things differently.
Guidewire believes that efficient and agile predictive analytics requires an end-to-end approach. In the next installments of this series, we will consider the issues of extracting and preparing data, building predictive models, integrating those models into business processes, and monitoring the results.
Note that predictive analytics involves building models to predict future results and is more than looking at historical data to understand past results (often called “descriptive analytics” or “business intelligence”). Predicting the future involves separating random events from predictable patterns and requires well-grounded mathematical techniques.
Issues with Expanding the Current Approach to Analytics
While many insurers struggle with getting value out of the predictive models they have built (among other struggles as well), we will assume here that an insurance company has managed this for multiple models addressing multiple business problems. Typically, this is done by having skilled and experienced individuals who can navigate the process.
Even in this case, where insurers have teams that have successfully implemented multiple models, there are limits to how this approach can scale to meet the increasing demand for analytics. Consider the following:
Increasing investment – expanding predictive analytics to more and more use cases requires scaling up the investment in individuals experienced in predictive modeling.
Increasing impact of legacy systems – solutions for extracting and transforming data that worked for specific predictive models may not scale into an efficient framework for the future, exacerbating the cost required to expand.
Increasing “key-person risk” – as specific, talented individuals build customized models, the risk associated with those people leaving the group (to another company or another part of the insurer) increases accordingly.
Maintaining operationalized models over time – a need exists for making sure that operationalized models remain efficient and are rebuilt when necessary. Every new model put into production comes with an on-going maintenance effort.
New applications of predictive analytics can require new analytical techniques – while this seems like a straight-forward statement, the introduction of new analytical techniques impacts the skills required of those in the predictive analytics group, and more importantly impacts the implementation of these predictive models.
Cost and complexity of deploying various predictive model formats – as new techniques are used, the complexity and variety of model rules expands. Without standardization, the costs associated with meeting each new situation escalates.
Managing model risk (ERM) – one underappreciated issue is the increasing burden of model management. This is a general Enterprise Risk Management (ERM) task – do you know how many models are used in production business processes, including which business processes and which version of the model? Keeping track of this information and auditing it periodically is key to minimizing model risk.
While predictive analytics holds tremendous potential to transform insurer operations, the efficiency of these efforts should not be overlooked. Reconsidering the entire end-to-end process is required. The other parts in this six-part series will explore each of these issues in more depth. Part two will discuss the data needs of predictive analytics.