Glossary
Model drift
Degradation over time as the world stops resembling the training data.
Drift is gradual and silent. Fraud patterns move, products change, lighting on a production line shifts with the seasons.
Monitoring input distributions and output rates catches it before the business notices as a performance problem.
Monitoring the inputs is often more practical than monitoring the outputs, because input distribution shifts are visible immediately while performance degradation only becomes measurable once ground truth arrives, which can be weeks later.
The frequency of retraining should follow the rate of change in the domain, not the calendar. Fraud models may need monthly attention; a document classifier over stable forms may run for years. Setting an arbitrary quarterly cycle wastes effort in one case and is far too slow in the other.
Set the alert on a business metric as well as a statistical one. Input distribution shifts are useful early warnings, but the number that gets attention is the one someone is accountable for.
Related terms, in context
The concepts you almost always meet alongside model drift.
- MLOps
- The practices that keep models deployable, monitorable and reversible in production.
- Evaluation
- A labelled test set that turns 'it seems good' into a number you can regress against.
- Predictive analytics
- Using historical data to forecast outcomes, demand, churn, risk, failure.
Where this shows up in our work
Model drift is not an abstraction for us. It is a decision we make on live projects. It shows up most directly in ai infrastructure & mlops, predictive analytics & forecasting, where getting it wrong has a cost someone can measure.
If you are evaluating a vendor on this, the useful question is not whether they can define the term. It is what they measure, what they would refuse to do, and what happens in their system when the assumption behind model drift stops holding.
Questions
What is Model drift?
Degradation over time as the world stops resembling the training data.
Does Orqent Labs build this?
Yes, AI Infrastructure & MLOps and Predictive Analytics & Forecasting. We work across India, covering all 19,238 PIN codes remotely.
Building something that involves model drift?
We will tell you honestly whether it is the right approach for your problem.
Or email bd@dtrasglobal.com · call +91 74118 77878