TL;DR
- What it is: Model drift is the drop in a machine learning model’s predictive performance that happens as the data or the world it scores changes, without anyone touching the model itself.
- The core mechanic: Drift shows up in three places: the input data, the relationship between inputs and outcomes, or the pipeline feeding the model. Each needs a different fix.
- Who it affects: Any model that scores something continuously, not just once at launch. Lending, collections, insurance, and fraud models are all exposed because the population they score keeps moving.
- Key fact: The RBI’s June 2026 draft guidance names this pattern “time-suitability,” one of three sources of model risk it identifies.
- One caveat: This RBI guidance is a draft, with public comments closed 24 July 2026, not yet in force, and it doesn’t prescribe how often to check for drift. That’s a practitioner decision, covered in the deeper monitoring guide.
A model doesn’t need a bug to stop working. It just needs the world to change underneath it.
Model Drift Is a Performance Problem, Not a Code Problem
Model drift is the general decline in a machine learning model’s predictive performance that happens over time, as the data it scores or the relationships in that data move away from what the model was trained on. Also called model decay, it’s the reason a model that scored well at launch can quietly get worse without a single line of its code changing.
That definition undersells the problem in one way: drift doesn’t announce itself. A drifting model keeps returning a score for every borrower, every claim, every account. The score just means less than it used to, and it keeps meaning less until someone checks.
Three Ways a Model Actually Drifts
“Model drift” is the umbrella term. Underneath it, three distinct things can be moving.
Data drift is a shift in the input data itself. The features a model scores start looking different from what it was trained on, for example a shift in customer demographics or a change in which products are available, even though the relationship between those features and the outcome hasn’t changed.
Concept drift is a shift in the relationship between inputs and outcomes. The inputs can look statistically the same, but what they mean has changed. This can happen suddenly, the way default behavior across an entire portfolio shifted during the pandemic, or gradually, the way fraud patterns evolve as bad actors adapt.
Upstream data change is a shift introduced further back in the pipeline, before the model ever sees the data: a currency conversion, a change in how a unit is recorded, a new data source feeding an old field. This one is easy to miss because the model’s own inputs can look fine while the meaning underneath them has already changed.
Knowing which one is happening decides what fixes it. Data drift is usually a retraining problem. Concept drift is usually a redesign problem, because the assumption the model was built on no longer holds. Upstream change is a data-engineering problem hiding inside what looks like a model problem.
What Actually Causes a Lending or Collections Model to Drift
In a regulated lending book, drift usually traces back to something real changing outside the model. Borrower behavior shifts with the economic cycle: a repayment pattern that held in a stable year doesn’t hold the same way once interest rates move or employment conditions shift. A regulatory or policy change can move the ground under a model just as fast, since a shift in permitted contact windows or consent requirements changes who a collections model can act on, which changes the population its outcomes get measured against even though the model’s own math never changed.
None of this means the model was built wrong. It means the model was built for a world that has since moved, which is exactly what makes drift hard to catch without deliberately checking for it.
Why Undetected Drift Is a Regulatory Problem, Not Just an Accuracy One
The Reserve Bank of India’s Draft Guidance on Regulatory Principles for Model Risk Management, released 24 June 2026, names three sources of model risk: model error, misapplication, and time-suitability, where a model becomes less fit for purpose as conditions change even though nothing about the model itself broke. Drift is what time-suitability looks like in a live model.
This RBI guidance is a draft under public consultation, not yet in force, with comments closed 24 July 2026. It doesn’t set a numeric schedule for how often a model should be checked for drift. What it points to is the underlying expectation: a model that was sound at validation can stop being sound before its next scheduled review, and an institution needs a way to know that’s happening rather than finding out when the numbers don’t add up.
How Drift Actually Gets Caught
Drift is measured, not felt. Statistical checks compare a model’s current inputs, outputs, or performance against what they looked like at training time: the Population Stability Index for distribution shifts, the Kolmogorov-Smirnov statistic for whether the model still separates outcomes as well as it used to. Tracked on a schedule, or continuously, these checks turn “the model feels off” into a number crossing a threshold.
That’s a deeper topic than this article covers. Model Drift: Catching Decay Before It Reaches the Regulator walks through the specific metrics, the standard thresholds practitioners use, and why a fixed calendar review misses drift that a continuous check would catch.
Building Models That Show Their Own Decay
Catching drift early is a monitoring design problem, not a modeling problem. A governed platform tracks a model’s stability metrics continuously against live performance, instead of waiting for a scheduled review, and flags a shift the moment it crosses a threshold. Every prediction stays traceable to an explanation, so when something moves, the cause is visible immediately instead of reconstructed after the fact.
A leading NBFC in India used this approach to rank borrowers by risk and focus collections on the highest-risk segment, seeing a 116% improvement in collections and 86% predictive accuracy, deployed in two weeks. Catching drift before it compounds was part of what kept that accuracy holding after launch, not just at it.
This sits inside the wider model risk management framework, which covers the governance, inventory, and validation that work alongside drift monitoring to oversee an institution’s full model portfolio.
Sources
- IBM, “What is model drift?,” published 16 July 2024. Source for the core definition, the data drift / concept drift distinction, the upstream data change cause, the COVID-19 concept-drift example, and the PSI/KS/Wasserstein detection methods list. General industry source, not India- or RBI-specific.
- Arize AI, “Model Drift: Definition, Causes, Types, and Impact.” Source for the prediction-drift framing and adversarial-input cause. General industry source, an ML observability vendor, not a named competitor in iTuring’s category and not cited by name in body copy.
- Reserve Bank of India, Draft Guidance on Regulatory Principles for Model Risk Management, Press Release 2026-2027/528, issued 24 June 2026; public comments closed 24 July 2026. Reused unchanged from the cluster’s verified facts table (same fact set as articles #1-4, no re-fetch drift). (Secondary summary: corplawupdates.in; verify final wording against the RBI primary document before publish.)
- iTuring case study, “Improve Collections and Optimize Efforts” (Leading NBFC in India): 116% collections improvement, 86% predictive accuracy, two-week deployment. Source: ituring.ai live case study.
