A new approach treats machine learning model degradation not as a binary "working or broken" state, but as a time-to-failure problem. This perspective, drawing from survival analysis and reliability engineering, quantifies how long an ML model remains dependable after deployment.

Survival analysis offers tools like survival curves and hazard functions to understand model lifecycles. These concepts, originally from medicine and industry, help answer questions about expected model lifespan and how risk evolves as data drift accumulates. The Weibull distribution is presented as a flexible model for failure-time analysis, capable of representing decreasing, constant, or increasing hazard rates with its shape (β) and scale (η) parameters.

Applying these methods means defining a clear "time-to-failure" threshold based on business needs. The analysis accounts for censored data—models that don't fail within the observation period—which is common in real-world deployments. This structured approach moves beyond ad hoc thresholds, enabling more principled, data-driven decisions for retraining schedules and alerting policies.