COVID-19 Has Broken Your Predictive Models

The COVID-19 pandemic has lessened the efficacy of predictive models, regardless of the algorithmic approach that built them, to the point that they may now be counterproductive. COVID-19 has transformed society to an extreme extent and with remarkable speed. It has changed the economic environment, how organizations conduct their work, and has increased morbidity and mortality. Social distancing has impacted all economic sectors; how long it will last is unknown.

Auto lines of business have seen significant reductions in miles driven and lower claims activity. Allstate, GEICO, and American Family have begun to rebate auto premiums. COVID-19 will affect workers’ compensation and disability, especially STD. Employees at risk of furlough or termination may seek to maintain their income by submitting a claim. Coverage may expand to include COVID-19 infections, as is the case in Florida, which recently expanded coverage for frontline workers.

Correlations May No Longer Apply

Statistical modeling and machine learning are based on historical data, and heuristics rest on prior experience. Organizations cannot rely on these methods and information sources to provide useful insights given current circumstances. These analytical approaches depend on identifying correlations in the data that they can associate with future outcomes. Correlations may be broken, business rules may no longer apply, and models may no longer provide valuable insights.

Three Action Steps to Take

Insurers that have operationalized their analytics and embedded models and business rules into their processes will want to take immediate action. Here are steps they can take:

  1. Conduct an analytics inventory to identify all areas where these models and rules are in use. This inventory may require going beyond the primary analytics organization (e.g., data science, actuarial) as lines of business and other functional areas may have implemented their own solutions. Organizations and their use cases include:
     Function   Use Case 
     Sales   Customer profitability, churn, next best action 
     Distribution   Agent/broker performance 
     Underwriting   Risk scores, profitability, premium audit 
     Claims   Severity, fraud, recovery 
     Finance   Forecasts 
  2. Apply business judgment and evaluate the likelihood that the insights and actions that models and rules generate continue to add value.
  3. Develop an action plan where models and rules require revision. In some cases, the action may be to improve data capture, increasing the frequency of data capture to develop a sufficient volume of recent data to build new models. In other cases, organizations may take a completely different approach, e.g., scrapping the churn propensity model and replacing it with a rule that captures the identity of individuals who click “cancel policy” on the website.

Correlations that predictive models identify are always at risk of decay and the need for refinement, even under normal circumstances. Insurers had developed budgets and plans to refresh models. However, the rapid emergence of this pandemic has accelerated the drop-off in correlations; it must not be ignored. Suppress the analytics and conduct more manual decisioning as business volumes fall, and have analytics teams work to gather fresh data sets and build new models.

How can we help?

If you have a question specific to your industry, speak with an expert.  Call us today to learn about the benefits of becoming a client.

Talk to an Expert

Receive email updates relevant to you.  Subscribe to entire practices or to selected topics within
practices.

Get Email Updates