Definition
Model Drift Detection Rate (MDDR) measures the rate at which a deployed prediction model's accuracy degrades over time, expressed as the slope of prediction accuracy over rolling evaluation windows. A negative slope means the model is getting worse. The steeper the slope, the faster the degradation.
The Measurement Gap
A power generation company deployed ML-based predictions for gas turbine hot-section component failures in 2023. The model performed well in its first year. Prediction accuracy held steady around 82%, confidence calibration was reasonable, and the maintenance team was scheduling hot-section interventions 3 to 4 weeks ahead of failure instead of reacting to performance loss. The investment was paying off.
By mid-2024, the fleet started running differently. Carbon pricing incentives shifted the dispatch profile. Turbines that had previously run at steady base load were now cycling more frequently, ramping up and down as renewable generation fluctuated. The thermal cycling patterns were fundamentally different from the steady-state conditions the model was trained on.
Nobody updated the model. Nobody was watching for the need to.
Over the next 8 months, prediction accuracy eroded from 82% to 61%. But it didn't drop suddenly. It lost about 2.5 percentage points per month, slowly enough that no individual prediction event triggered alarm. The monthly PdM review showed "predictions active, system operational". MTBF held roughly stable because the turbines were newer and hadn't accumulated enough cycles for the new failure patterns to dominate. OEE was fluctuating, but everyone attributed that to the cycling dispatch pattern, not to maintenance prediction failures.
Then, in early 2025, two hot-section failures occurred within three weeks on turbines the model had scored as "low risk". The combined cost of emergency repairs, replacement power procurement, and the insurance deductible exceeded $2.8 million. The post-incident review traced the failures to thermal fatigue patterns the model had never seen in its training data. The model was still running. It was still producing predictions. It just wasn't right anymore.
The existing metrics had no mechanism to detect this. PdM Hit Rate was reported annually and hadn't been recalculated since deployment. MTBF captured the turbine failures but as isolated data points, not as evidence of model degradation. Availability dropped when the failures occurred but couldn't trace the root cause to a prediction system that had silently stopped working months earlier.
There is no metric that tells you whether a deployed prediction model's accuracy is degrading over time, which means organizations continue relying on models that have silently stopped working.
Formula
Model Drift Detection Rate (MDDR) measures the rate at which a deployed prediction model's accuracy degrades over time, expressed as the slope of prediction accuracy over rolling evaluation windows.
In plain language:
Is my model getting worse, and if so, how fast?
The Formula
MDDR = dA / dt = (A_t - A_t-n) / n
Where A_t is prediction accuracy in the current evaluation window, A_t-n is accuracy n windows ago, and n is the number of windows in the measurement period.
Worked Example
A fleet operator tracks monthly prediction accuracy for their engine health model across 12 months:
| Month | Accuracy |
| Jan | 0.84 |
| Feb | 0.83 |
| Mar | 0.82 |
| Apr | 0.81 |
| May | 0.80 |
| Jun | 0.78 |
| Jul | 0.77 |
| Aug | 0.75 |
| Sep | 0.73 |
| Oct | 0.72 |
| Nov | 0.70 |
| Dec | 0.68 |
MDDR = (0.68 - 0.84) / 12 = -0.16 / 12 = -0.0133 per month
The model is losing 1.33 percentage points of accuracy per month. At this rate, a model that started at 84% accuracy will cross 70% (proposed investigation threshold) in about 10.5 months from its start, and cross 60% (proposed suspension threshold) in about 18 months. The turbine operator in the problem scenario was experiencing roughly double this rate (-0.025/month) due to the more dramatic operating context shift.
What matters isn't just the current MDDR value. It's the trajectory. If MDDR was -0.005/month for the first six months and then shifted to -0.025/month in the last three, something changed in the operating environment that accelerated degradation. That inflection point is where you investigate.
Sign Convention
MDDR of 0 means stable model. Negative values mean degradation. Positive values mean improvement, which is also worth knowing: why is the model getting better? More negative means faster degradation. Thresholds are proposed, not proven.
Edge Cases
Seasonal accuracy patterns (models that are better in summer than winter due to thermal effects) will show cyclical MDDR. Use at least 12 months of data before declaring a trend. Sudden accuracy drops (a single month dropping 10+ points) should trigger immediate investigation independent of MDDR. They suggest an acute event rather than gradual drift.
Cross-References
MDDR is the trajectory companion to the Prediction Confidence Calibration Index snapshot. That index tells you "the model is miscalibrated". MDDR tells you "the model is getting worse at a rate of X per month, and at this rate will cross your action threshold in Y months".
Upstream: Prediction Confidence Calibration Index feeds MDDR conceptually. That index measured at multiple time points gives you the data to compute MDDR. But MDDR can also be computed on raw accuracy without calibration analysis, making it independently computable.
Downstream: Autonomous WO Signal-to-Noise Ratio degrades as models drift. A drifting model generates increasingly noisy work orders. Degradation Projection Validity Rate declines as the model driving projections loses accuracy. Strategy Override Safety Index becomes more concerning because strategy recommendations from a drifting model inherit its degradation.
The most dangerous failure mode is a model that drifts slowly enough that no individual prediction raises alarm, but fast enough that within a year the model has crossed from reliable to harmful. MDDR makes this visible by measuring the slope, not just the current value.
Proposed formula, pending validation pilots.
At a Glance
Decision Supported
Is this model still reliable, or has it silently degraded to the point where it needs retraining, supplementation, or retirement?