Model Drift
Model drift (also called concept drift) is the gradual decline in a machine learning model's predictive performance over time, occurring because the real-world relationship between inputs and the target outcome has changed since the model…
Definition
Model drift (also called concept drift) is the gradual decline in a machine learning model's predictive performance over time, occurring because the real-world relationship between inputs and the target outcome has changed since the model was trained.
Overview
A model is trained to capture a statistical relationship that existed in its training data at a particular point in time. Model drift happens when that underlying relationship itself shifts — customer preferences change, a competitor enters the market, a fraud pattern evolves to evade detection — so the patterns the model learned no longer accurately predict outcomes, even if the input data still looks superficially similar to what the model was trained on. This is a distinct phenomenon from data drift, which describes a change in the distribution of the input data itself (for example, a new customer segment starting to use a product), even if the relationship between inputs and outputs stays the same. In practice the two often occur together and can be hard to disentangle, but the distinction matters for diagnosis: data drift can sometimes be addressed by better input handling, while model drift usually requires retraining on newer data that reflects the changed relationship. Detecting model drift requires ongoing monitoring in production — tracking metrics like accuracy, precision and recall, or business KPIs against a baseline, and watching for gradual or sudden degradation. Because ground-truth labels are often delayed or unavailable in production (you may not know if a loan actually defaulted for months), teams often use proxy signals like prediction confidence distributions or input feature statistics to detect drift earlier. Addressing model drift typically means retraining the model on recent data, and a model registry makes it straightforward to compare the new model's performance against the old one before promoting it, a workflow central to MLOps & Model Deployment.
Key Concepts
- Describes performance decline caused by a real change in the input-to-output relationship
- Distinct from data drift, which is a change in input data distribution, not the underlying relationship
- Often detected via ongoing production monitoring of accuracy or business KPIs
- Complicated by delayed or missing ground-truth labels in many production settings
- Typically addressed by retraining the model on more recent data
- Supported by model registries that let teams compare retrained models against current production models