Technical Summary: When Drift Detectors Cry Wolf: False Alarm Rates in Continuous ML Monitoring

Problem Statement

While drift detection is a fundamental component of production machine learning (ML) monitoring systems, existing research predominantly evaluates detectors in controlled, single-shot settings using synthetic shifts. These evaluations prioritize detection accuracy but often overlook the behavior of detectors under continuous monitoring conditions. In production environments, models are monitored repeatedly over time across numerous features. Under these conditions, even detectors with nominal false positive rates can accumulate frequent alerts due to random variation. This accumulation leads to "alarm fatigue," where engineers may ignore or disable monitoring signals, undermining system reliability. The paper addresses the lack of systematic empirical analysis regarding how often drift detectors "cry wolf" (generate false alarms) in realistic, continuous scenarios and investigates the trade-offs between sensitivity and stability.

Methodology

The authors conducted controlled experiments to quantify false positive behavior across five widely used drift detectors:

1. Population Stability Index (PSI)
2. Kolmogorov–Smirnov (KS) test
3. Maximum Mean Discrepancy (MMD)
4. Least-Squares Density Difference (LSDD)
5. Adversarial Validation

Experimental Setup:

• Dataset: The Adult Income dataset, chosen for its mixed numerical and categorical features.
• Scenarios:
  • No-Drift: Daily batches sampled from the reference distribution over a 30-day cycle.
  • Gradual-Drift: Controlled shifts injected into the "age" feature with magnitudes of 5, 10, 15, and 20 years.
• Variables:
  • Batch Size: Swept from 50 to 500 samples (increments of 50).
  • Thresholds: Evaluated both standard thresholds and those adjusted via Bonferroni correction to address multiple testing issues.
• Metrics:
  • False Positive Days: Number of days with alarms in the no-drift scenario (out of 30).
  • True Positive Rate (TPR): Fraction of drift-period days with correct detection.
  • Time to Detection (TTD): Days between drift onset and first detection.

Key Results

The study reveals significant variations in detector behavior depending on batch size and the specific method used:

• PSI (Population Stability Index): Exhibits extreme sensitivity to small batch sizes. In the no-drift setting, it triggered alarms on nearly all days (≈30/30) when batch sizes were between 50 and 100 samples. However, its behavior stabilizes significantly once batch sizes exceed approximately 200 samples, after which false alarm rates drop to near zero.
• KS and LSDD: These statistical detectors demonstrate more consistent performance in low-data regimes compared to PSI. However, they still exhibit persistent, non-trivial fluctuations in false alarm rates across the full range of batch sizes under standard thresholds. KS showed a favorable balance, achieving a True Positive Rate (TPR) up to 0.90 as drift magnitude increased.
• MMD (Maximum Mean Discrepancy): Displayed very low false alarm rates but suffered from near-zero detection rates under the tested univariate drift scenarios, suggesting default kernel settings may miss localized drift.
• Adversarial Validation: Showed stable behavior with false alarms decreasing rapidly as batch size increased. However, it was conservative, detecting only larger, strongly separable distribution changes (TPR up to 0.46).
• Bonferroni Correction: Applying this correction substantially reduced false positive rates (often to near-zero) for statistical detectors. However, this came at the cost of reduced sensitivity to small or gradual drifts, reinforcing the inherent stability–sensitivity trade-off.

Key Contributions

1. Empirical Quantification of False Alarms: The paper provides a systematic comparison of false positive behavior across multiple detector families under continuous monitoring conditions, moving beyond single-shot synthetic evaluations.
2. Batch Size Analysis: It identifies a critical threshold for PSI, demonstrating that its reliability is heavily dependent on batch sizes exceeding ~200 samples, whereas other detectors behave differently in low-data regimes.
3. Trade-off Characterization: The work explicitly highlights the trade-off between stability (low false alarms) and sensitivity (detecting subtle drift), particularly when applying multiple testing corrections like Bonferroni.
4. Practical Guidelines: The authors offer specific deployment guidelines for selecting and calibrating drift detectors in production ML systems based on batch size constraints and the tolerance for false alarms versus missed detections.

Significance and Claims

The paper claims that false alarms are not merely a theoretical concern but a practical operational issue that can degrade the effectiveness of production ML monitoring. By demonstrating that standard thresholds designed for single-shot experiments are often inadequate for continuous monitoring, the work underscores the need for careful configuration.

The authors conclude that:

• PSI should only be used when batch sizes are sufficiently large (>200 samples); otherwise, it produces near-constant alarms.
• KS serves as a reliable default for tabular monitoring with small to moderate batch sizes due to its balance of stability and sensitivity.
• Bonferroni corrections are effective for prioritizing operational stability but require practitioners to accept reduced sensitivity to subtle drifts.

Ultimately, the paper argues that configuration choices materially affect monitoring reliability, and practitioners must explicitly manage the stability–sensitivity trade-off rather than relying on default settings derived from research benchmarks.