Predictive Analytics in the Insurance Industry

Hongyuan Wang, PhD & MAS, Director of Predictive Analytics, Data and Analytics, Employers

Hongyuan Wang, PhD & MAS, Director of Predictive Analytics, Data and Analytics, Employers

Predictive Analytics (PA) is a process to translate data into business decisions and then turn it into profit. PA can be approached by using traditional statistical predictive models or advanced machine learning models, which are actively used in all major industries. PA is a complicated process starting from original business ideas and progressing through data preparation, creation of model dataset, development and validation of the predictive model, implementation of the model, and monitoring the model results. A very good fitted model which cannot be applied correctly or fails to implement due to various reasons is not a successful PA application. This article will focus on the PA applications and some challenges of using them in the insurance industry.

PA can help insurance companies to reduce the Loss and Expense Ratio and grow organically through different applications. For example, predicting the individual policy’s claim frequency and severity, targeting the right policy, quantifying the potential lifetime value of a policy, and smoothing the operational processes.

Business Requirements and Understanding the Data

A PA application always comes from business needs. Understanding what business is required and the related historical data are most important. For example, a sales team likes to know how the retention ratio could impact both growth and profit, two important goals of the insurance company.

Building a consistent and high-quality PA process requires a complete understanding of business, data, modeling, and tools for the project leader

An underwriting team likes to know how the company should price risks for each renewal. Based on questions such as these, an analytics team must research and understand the requirements and available data, both internal and external data, to establish a PA. Determining the final reliable and consistent data to be used for the PA is always a challenge due to so many data resources and noisy information out there. At the same time analytics tools and model methodologies to be used also need to be determined.These are critical and will directly impact the model development and implementation process.

Data Preparation and Development of Predictive Model

There are a variety of analytics tools which can be used for data preparation and predictive model development. For example, SAS, R, or Python. Which tool to use depends on the company’s IT infrastructure and the analyst’s coding and modeling knowledge. But one thing is for sure, the tool should be able to connect to all internal databases, be able to easily import available external data, be able to efficiently handle large data manipulation and run smoothly and fast during model development. Some external data may be very complicated, with different formats or structures which need to borrow other specific tools to grab the useful data.

After the data preparation is done, analysts may continue to use the tool to do further data transformations and start model development. Other modeling tools can be used to automate the model tuning and variable selection from many different model methodologies with a very high computational speed, such as a black-box modeling tool. One of the challenges is to determine which model methodology (for example, GLM vs GBM) should be used. Business requirements, regulations or IT Infrastructure may restrict you from using many different advanced model methodologies, however, if very high quality and relevant data is available for the PA, different methodologies may not make a big difference.

Implementation of the Model

It is crucial to create an implementation system or API to automatically pull the data from different data resources, run the model scoring process, and integrate with the other business operation systems to make the final business decisions based on the model scores. It can be a real-time process or can be scheduled for a daily, weekly or monthly run. The bigger challenge here is creating an implementation system extensive enough to integrate future IT infrastructure changes (Data Warehouse changes, Operation System changes, or Cloud Environment changes, etc.).

Conclusion

A successful application of PA must be satisfied at every step of the process. Building a consistent and high-quality PA process requires a complete understanding of business, data, modeling, and tools for the project leader. It is a collaboration of different teams which may include Business Operation, Data and Analytics, and IT.