According to Gartner, only 20 percent of analytic solutions actually deliver business outcomes. Similarly, a report from VentureBeat found that 87 percent of data analytics projects never even make it into production. 

There are likely many drivers and unique factors behind these poor success rates, but model drift is undoubtedly one of the most significant. Changes in the data and the relationships between data variables can lead to model performance degradation, or “drift,” which negatively affects the accuracy of any resulting insights. 

The primary cause for model drift is a change in business. Mergers and acquisitions activity, the introduction of new products, entry into a new market, or the emergence of new laws and regulations could all lead to model drift. Because businesses are constantly evolving, it’s important that data leads monitor these changes and course-correct as necessary to get the model back to an acceptable level of performance. 

To properly account for and problem-solve model drift, consider implementing the following strategies: 

  1. Effective Data Governance 

Effective data governance practices are one way that companies can manage data drift. First, identify the variables in the hypothesis, then define the data quality KPIs, and from there, set targets and thresholds. Finally, continually track these KPIs to stay abreast of any changes in data quality. 

  1. Periodically Review Model Relevance 

Assessing business dynamics and reviewing the relevancy of the existing models is another critical step in combatting model drift, which is why it’s important to engage with key stakeholders and get their feedback on the approach. Topics to consider here include:

  • Why do we want to have these insights? 
  • What is the value of knowing and not knowing this information?
  • Who owns the insights coming out of these models? 
  • What are the relevant data attributes required for the model to yield accurate insights? 
  1. Integrate ModelOps and DataOps 

Integrating ModelOps and DataOps enables companies to replace the deployed analytics model quickly and ethically if business circumstances change. These methodologies help get analytics models from the lab to production faster, helping analytics keep pace with evolving organizational strategies and avoiding much, if not all, model drift.

Check out this recent Forbes article for more on the above and other considerations in correcting model drift.