Pith. sign in

REVIEW 3 major objections 4 minor 30 references

Under continuous monitoring, drift detectors accumulate false alarms at rates far above their nominal per-test rates, with PSI triggering on nearly all days when batch sizes fall below about 200 samples.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 18:17 UTC pith:QTOYYJDD

load-bearing objection The paper's central Bonferroni claim is contradicted by its own tables; the PSI batch-size observation is plausible but the evidence as written doesn't support the conclusions. the 3 major comments →

arxiv 2607.17336 v1 pith:QTOYYJDD submitted 2026-07-19 cs.LG

When Drift Detectors cry Wolf: False Alarm Rates in continuous ML Monitoring

Raj Shekhar Singh This is my paper
classification cs.LG
keywords drift detectionfalse alarm ratescontinuous monitoringpopulation stability indexKolmogorov-Smirnov testMMDBonferroni correctionbatch size behavior
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Continuous machine-learning monitoring repeatedly tests the same features over time, so even low per-test false-positive rates pile up into frequent alarms. This paper runs a 30-day no-drift simulation and a gradual-drift simulation on a 14-feature tabular benchmark, and measures alarm days for PSI, KS, MMD, LSDD, and adversarial validation. It finds that PSI alarms on nearly all 30 days at batch sizes of 50–100, then stabilizes only after batches exceed roughly 200 samples, while the other detectors show low but non-trivial false-alarm variability. Applying a Bonferroni correction across features drives false alarms down for the statistical detectors, but also lowers true-positive rates on small drift. The upshot is that batch size and multiple-testing corrections, not just detector choice, determine how often a drift monitor 'cries wolf'.

Core claim

In the no-drift condition, PSI with standard thresholds produced false alarms on 30/30 days at batch size 50 and 29/30 at batch size 100, dropping to about 12/30 at 150 and reaching near-zero only at 200 or more samples. KS, MMD, and LSDD stayed between 0 and about 0.6 false-alarm days across all batch sizes. Under injected age drift, KS achieved the best sensitivity–stability trade-off, MMD detected essentially nothing with default kernels, and PSI flagged drift but with the same false-alarm load as before. Bonferroni correction reduced false alarms for the statistical detectors but cut true-positive rates, illustrating the stability–sensitivity tension in continuous monitoring.

What carries the argument

The central mechanism is the continuous monitoring protocol: a per-day, per-feature hypothesis test repeated over 30 days and 14 features, where a daily alarm is raised if any test (or some aggregation of tests) crosses its threshold. This repeated-testing structure is what converts a nominal per-test false-positive rate into accumulated alarm days, and it is the object the paper manipulates by varying batch size and applying Bonferroni corrections. The named detectors—PSI, KS, MMD, LSDD, adversarial validation—are the instruments; the protocol is the load-bearing setup.

Load-bearing premise

The load-bearing premise is that the simulation—independent daily batches drawn from the same distribution, drift isolated to one univariate feature starting after day 30, and an unstated rule for combining 14 per-feature tests into one daily alarm—is a faithful model of continuous production monitoring; if real data are autocorrelated, multi-feature, or aggregated differently, the reported false-alarm counts and the 200-sample guideline will not transfer.

What would settle it

Re-run the no-drift experiment with autocorrelated daily batches (e.g., each day's batch is the previous day's batch plus independent noise), over the same 30-day window and batch sizes. If PSI's false alarms remain near 30/30 at all batch sizes, or if a different cliff appears, the paper's central practical guideline—use batch sizes above 200—would fail for a realistic production data stream.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • PSI should only be used when batches contain at least about 200 samples; below that it is effectively always alarming under the tested protocol.
  • KS is the most reliable default for tabular monitoring in the tested setup, with low false alarms and high detection for moderate drift.
  • MMD with default Gaussian kernels and fixed permutation settings can be nearly blind to simple univariate drift, so kernel configuration matters more than detector family.
  • Bonferroni correction across features reduces false alarms for the statistical detectors, but the paper also reports that it lowers true-positive rates, illustrating a stability–sensitivity trade-off.
  • Detector rankings from single-shot benchmark evaluations do not carry over directly to continuous monitoring; calibration must be done on the actual monitoring schedule.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's simulated protocol samples daily batches independently from the reference distribution, which eliminates temporal autocorrelation; real production streams are often autocorrelated, so the ~200-sample threshold may shift or vanish under correlated data. A testable extension is to rerun the sweep with an autoregressive batch generator.
  • The paper never states the rule that maps 14 per-feature test outcomes to a daily alarm; different aggregation rules (any-feature vs. majority vs. mean statistic) will change false-alarm counts, so the absolute numbers are protocol-dependent.
  • A natural next step is to correct for the 30-day repeated-testing horizon as well as the 14 features, using a sequential or family-wise correction; this would target alarm fatigue directly rather than only per-feature error rates.
  • The PSI result is consistent with its histogram-based definition: small batches make bin counts noisy, so a closed-form bound on false-alarm probability vs. bin count could predict the cliff without simulation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The manuscript reports a simulated continuous-monitoring study of five drift detectors (PSI, KS, MMD, LSDD, adversarial validation) on the Adult Income dataset, with 30-day monitoring cycles, batch sizes 50–500, and no-drift/drift scenarios. The main claims are that PSI produces near-daily false alarms below roughly 200 samples and stabilizes above that size; KS, MMD, and LSDD are comparatively stable but with persistent batch-size fluctuations; a Bonferroni correction substantially reduces false alarms at a cost in sensitivity; and adversarial validation is conservative. The paper draws practical deployment guidelines from these results.

Significance. If the empirical claims were reliable, the paper would provide useful operational guidance—notably the ~200-sample threshold for PSI and the explicit stability–sensitivity trade-off under multiple-comparison correction—and would help bridge the gap between research evaluations of drift detectors and production monitoring. The study’s strengths include using standard default thresholds rather than fitted constants, covering five detector families, and reporting false-positive days, TPR, and TTD. However, as written the central quantitative claims are internally inconsistent: the Bonferroni results in Appendix B exactly duplicate the uncorrected tables, the drift-timing protocol makes the TPR/TTD results impossible, and the daily-alarm aggregation rule is never specified. Without code or raw data, these contradictions cannot be resolved from the manuscript, so the practical guidelines are not supported by the reported evidence.

major comments (3)
  1. [Appendix B, Tables 6–7; §3.4; Fig. 3] The central Bonferroni claim is contradicted by the paper's own tables. For KS, LSDD, and MMD, every Standard and Bonferroni entry is numerically identical at all batch sizes (e.g., KS: 0.40±0.49 at batch 50, 0.20±0.40 at 100–200; LSDD and MMD likewise). With α reduced from 0.05 to ≈0.007, p-values in (0.007, 0.05] should convert some standard alarms to non-alarms, so exact equality across all seeds and batch sizes is implausible. Figure 3 nevertheless claims a 'substantial reduction.' The paper provides no code/data to disambiguate, so the main quantitative conclusion is unsupported as written.
  2. [Appendix A.2, Table 3; Table 8] The drift-span protocol is internally impossible. Appendix A.2 specifies a 30-day monitoring duration and states 'Drift begins after day 30; gradual drift over 10 days.' If monitoring ends at day 30, no drift days exist, so the TPR and TTD values in Table 8 (e.g., KS TPR up to 0.90 with TTD 2 days; PSI TPR 0.84) cannot be computed. The drift window must extend beyond 30 days or begin earlier; as written, the sensitivity side of the stability–sensitivity trade-off rests on an undefined or contradictory scenario.
  3. [§3.2–3.4; Appendix A.4] The definition of a daily alarm is never given. The paper uses 14 features and reports per-feature tests for KS/PSI, and the α=0.05/num_features Bonferroni values imply 14 simultaneous tests per day, but the aggregation from per-feature p-values to a daily alarm (any-feature alarm, maximum statistic, average p-value, etc.) is not stated. Consequently the false-positive-day counts for statistical detectors cannot be reproduced, and the batch-size comparisons in Figure 1 are not operationally interpretable. This is a load-bearing methodological gap rather than a presentation issue.
minor comments (4)
  1. [Appendix B, Table 8] Table 8 reports TPR and TTD without stating the batch size for the drift experiment; if it is the batch-size sweep or a fixed batch size, this should be explicit.
  2. [References] References [1] and [18] are duplicated (identical 'A survey on concept drift adaptation' by Gama et al.); [18] is incorrectly attributed to Fanaee-T and Gama. Please correct.
  3. [Figure 3] The caption says 'substantially reduces false alarms' but the underlying tables show identical counts; either the figure or the tables are wrong, and this must be resolved before publication.
  4. [Appendix A.4] The PSI threshold of 0.2 and binning details are given, but the number of PSI features aggregated per day and feature-wise vs multivariate operation are not described; please clarify.

Circularity Check

0 steps flagged

No circularity found: the paper is an empirical benchmark study with standard thresholds; its central claims are not derived from fitted inputs or self-citations.

full rationale

The paper makes no derivation-from-first-principles claims and does not fit parameters to outcomes. Detector thresholds (PSI 0.2, KS/MMD/LSDD alpha = 0.05, Bonferroni alpha = 0.05/14, adversarial AUC >= 0.6) are standard defaults chosen before the experiments, not tuned to the reported false-alarm counts. The 'predictions' are empirical measurements of a simulation protocol, so they are not equivalent to their inputs by construction. There are no author self-citations that carry the argument: all references are to external literature, and no uniqueness theorem or prior work by the same author is invoked to force a conclusion. Potential concerns raised in review, such as the apparent internal inconsistency between the narrative claim that Bonferroni reduces false alarms and Appendix B Tables 6-7 showing identical counts for standard and Bonferroni variants, and the protocol's statement that drift begins after day 30 while the monitoring window is 30 days, are matters of experimental reproducibility and internal consistency, not circularity. They do not involve a derivation reducing to its inputs or a fitted parameter being renamed as a prediction. Accordingly, the appropriate circularity finding is no significant circularity, score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

All central numbers are produced under hand-chosen detector thresholds and simulation choices; none are fitted to the target claim, but the claims are only about these configurations. No new theoretical entities are introduced.

free parameters (6)
  • PSI threshold = 0.2
    Standard industry setting chosen in A.4; determines when PSI alarms and hence the false-alarm counts.
  • PSI binning = 10 equal-frequency bins
    Histogram discretization in A.4; PSI's small-sample behavior is governed by bin counts.
  • Per-test significance alpha = alpha = 0.05; Bonferroni reported ~0.007
    Set in A.4; note 0.05/14 is approximately 0.0036, not 0.007, so the stated Bonferroni value is internally inconsistent.
  • MMD/LSDD kernel configuration = Gaussian kernel default; reference subset 2000; permutations 100
    A.4; kernel choice determines MMD's near-zero TPR, a central result.
  • Adversarial alarm threshold = ROC-AUC >= 0.6
    A.4; controls the adversarial detector's sensitivity/specificity trade-off.
  • Batch-size grid and seeds = 50-500 by 50; 5 seeds
    A.3; the claimed ~200 stabilization boundary is an artifact of this grid, and a coarser grid could locate a different transition.
axioms (5)
  • standard math Bonferroni correction controls family-wise error rate when applied to the per-feature tests.
    Relied on in Section 3.2/3.4 to interpret corrected thresholds; textbook, but the paper's arithmetic for alpha is wrong (0.05/14 is about 0.0036, not 0.007).
  • domain assumption In the no-drift experiment, daily batches are independent random samples from the fixed reference distribution.
    Appendix A.2; this defines the false-positive experiment and is the basis for all no-drift alarm counts.
  • ad hoc to paper The Adult Income dataset and a single shifted age feature are representative enough to support general production monitoring guidance.
    No external validation or multiple datasets; the conclusion section itself calls for future work on other datasets.
  • ad hoc to paper Default kernel settings for MMD/LSDD and thresholds for PSI/adversarial detectors are the appropriate operational configuration.
    A.4 lists Gaussian kernel defaults and thresholds without sensitivity analysis; all detector-family conclusions are conditional on these choices.
  • ad hoc to paper A daily alarm is a well-defined aggregation of the 14 per-feature test results.
    Never stated in the paper; required to compute false-positive days and to reconcile alpha with observed counts.

pith-pipeline@v1.3.0-alltime-deepseek · 6661 in / 17301 out tokens · 167730 ms · 2026-08-01T18:17:29.180101+00:00 · methodology

0 comments
read the original abstract

Drift detection is a core component of production machine learning monitoring systems, where detectors are used to compare incoming data with a reference distribution and trigger alerts when changes occur. However, these detectors are often evaluated in research settings that emphasize detection accuracy under synthetic shifts, while overlooking false alarms under continuous monitoring. In production environments, models are monitored repeatedly over time and across many features, and even small false positive rates can accumulate into frequent alerts, leading to alarm fatigue. We empirically analyze false positive behavior across five commonly used drift detectors: PSI, KS, MMD, LSDD, and adversarial validation. Consistent with existing literature, PSI exhibits strong sensitivity to batch size, producing frequent false alarms at small sample sizes; however, we further observe that its behavior stabilizes and improves substantially once batch sizes exceed approximately 200 samples. In contrast, KS, MMD, and LSDD display persistent fluctuations across batch sizes, while remaining comparatively more reliable than PSI in low-data regimes. Applying a Bonferroni correction reduces false positive rates, but often at the cost of reduced true positive sensitivity, reinforcing the well-known stability - sensitivity trade-off in drift detection. This work provides a systematic comparison of false positive behavior across multiple drift detectors under continuous monitoring conditions. We identify tradeoffs across detector families and provide practical guidelines for selecting and calibrating drift detectors in production ML systems.

Figures

Figures reproduced from arXiv: 2607.17336 by Raj Shekhar Singh.

Figure 1
Figure 1. Figure 1: False alarm rates as a function of batch size. Panel A (left) — Industry-style detectors (PSI and adversarial validation). PSI exhibits extreme false alarm rates at small batch sizes, approaching alarms on nearly all 30 days below 200 samples, followed by a sharp stability transition. Panel B (right) — Statistical detectors (KS, MMD, LSDD) at α = 0.05. These methods maintain relatively low false alarm rate… view at source ↗
Figure 2
Figure 2. Figure 2: Sensitivity–stability trade-off at batch size 200. Each detector is plotted by its false alarm rate (x￾axis) against its true positive rate for a strong drift of 20 years (y-axis). Detectors in the upper-left quadrant (low false alarms, high TPR) are most desirable. KS achieves the best overall balance, while MMD shows near-zero sensitivity despite low false alarms [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 2 linked inside Pith

  1. [1]

    A survey on concept drift adaptation.ACM Computing Surveys, 2014

    João Gama, Indr˙e Žliobait˙e, Albert Bifet, Mykola Pechenizkiy, and Abdelhamid Bouchachia. A survey on concept drift adaptation.ACM Computing Surveys, 2014

    2014

  2. [2]

    MIT Press, 2009

    Joaquin Quionero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil Lawrence.Dataset shift in machine learning. MIT Press, 2009

    2009

  3. [3]

    Learning under concept drift: A review.IEEE Transactions on Knowledge and Data Engineering, 2018

    Jie Lu, Anjin Liu, Feng Dong, Feng Gu, João Gama, and Guangquan Zhang. Learning under concept drift: A review.IEEE Transactions on Knowledge and Data Engineering, 2018

    2018

  4. [4]

    Failing loudly: An empirical study of methods for detecting dataset shift.NeurIPS, 2019

    Stephan Rabanser, Stephan Günnemann, and Zachary Lipton. Failing loudly: An empirical study of methods for detecting dataset shift.NeurIPS, 2019. 5 When Drift Detectors Cry Wolf: False Alarm Rates in Continuous ML Monitoring

    2019

  5. [5]

    Adaptive drift detection methods for data streams.Machine Learning, 2018

    Jacob Montiel et al. Adaptive drift detection methods for data streams.Machine Learning, 2018

    2018

  6. [6]

    Sculley et al

    D. Sculley et al. Hidden technical debt in machine learning systems. InNeurIPS, 2015

    2015

  7. [7]

    MIT Press, 2018

    Albert Bifet and Ricard Gavaldà.Machine learning for data streams. MIT Press, 2018

    2018

  8. [8]

    Population stability index: A practical guide.Towards Data Science, 2021

    Harpreet Sidhu. Population stability index: A practical guide.Towards Data Science, 2021

    2021

  9. [9]

    Oxford University Press, 2015

    Bart Baesens et al.Credit Risk Analytics. Oxford University Press, 2015

    2015

  10. [10]

    Controlling the false discovery rate.Journal of the Royal Statistical Society, 1995

    Yoav Benjamini and Yosef Hochberg. Controlling the false discovery rate.Journal of the Royal Statistical Society, 1995

    1995

  11. [11]

    Multiple comparisons among means.Journal of the American Statistical Association, 1961

    Olive Jean Dunn. Multiple comparisons among means.Journal of the American Statistical Association, 1961

    1961

  12. [12]

    A. N. Kolmogorov. On the empirical determination of a distribution law.Giornale dell’Istituto Italiano degli Attuari, 1933

    1933

  13. [13]

    Lewis.An Introduction to Credit Scoring

    Edward M. Lewis.An Introduction to Credit Scoring. Athena Press, 1994

    1994

  14. [14]

    Gretton, K

    A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Schölkopf, and A. J. Smola. A kernel two-sample test.Journal of Machine Learning Research, 13:723–773, 2012

    2012

  15. [15]

    Cambridge University Press, 2012

    Masashi Sugiyama et al.Density Ratio Estimation in Machine Learning. Cambridge University Press, 2012

    2012

  16. [16]

    Lopez-Paz and M

    D. Lopez-Paz and M. Oquab. Revisiting classifier two-sample tests.arXiv preprint arXiv:1610.06545, 2016

    Pith/arXiv arXiv 2016

  17. [17]

    Detecting and correcting for label shift with black box predictors

    Zachary Lipton et al. Detecting and correcting for label shift with black box predictors. InICML, 2018

    2018

  18. [18]

    A survey on concept drift adaptation.ACM Computing Surveys, 2014

    Hadi Fanaee-T and João Gama. A survey on concept drift adaptation.ACM Computing Surveys, 2014

    2014

  19. [19]

    Uci machine learning repository: Adult data set

    Dheeru Dua and Casey Graff. Uci machine learning repository: Adult data set. https://archive.ics. uci.edu/ml/datasets/Adult, 2019. Accessed: 2026-02-11

    2019

  20. [20]

    Scaling up the accuracy of naive-bayes classifiers.KDD, 1996

    Ron Kohavi. Scaling up the accuracy of naive-bayes classifiers.KDD, 1996

    1996

  21. [21]

    Nikiforov

    Michèle Basseville and Igor V . Nikiforov. Detection of abrupt changes: Theory and application.Prentice Hall, 1993

    1993

  22. [22]

    Learning with drift detection

    João Gama, Pedro Medas, Gladys Castillo, and Pedro Rodrigues. Learning with drift detection. InBrazilian Symposium on Artificial Intelligence, pages 286–295, 2004

    2004

  23. [23]

    Model monitoring and robustness of in-use machine learn- ing models: Quantifying data distribution shifts using population stability index.arXiv preprint arXiv:2302.00775, 2023

    Aria Khademi, Michael Hopka, and Devesh Upadhyay. Model monitoring and robustness of in-use machine learn- ing models: Quantifying data distribution shifts using population stability index.arXiv preprint arXiv:2302.00775, 2023

    Pith/arXiv arXiv 2023

  24. [24]

    Statistical properties of the population stability index.Journal of Risk Model Validation, 14(4):89–100, 2018

    Bilal Yurdakul and Joshua Naranjo. Statistical properties of the population stability index.Journal of Risk Model Validation, 14(4):89–100, 2018

    2018

  25. [25]

    Tonguz and Federico Taschin

    Ozan K. Tonguz and Federico Taschin. Using kolmogorov-smirnov distance for measuring distribution shift in machine learning.arXiv preprint arXiv:2510.15996, 2025

    arXiv 2025

  26. [26]

    An incremental change detection test based on density difference estimation.Machine Learning (preprint), 2017

    Xia Bu et al. An incremental change detection test based on density difference estimation.Machine Learning (preprint), 2017

    2017

  27. [27]

    X. Zhou, L. Tian, C. Peng, and Y . Lei. Dual: Learning diverse kernels for aggregated two-sample and independence tests.arXiv preprint arXiv:2510.11140, 2025

    arXiv 2025

  28. [28]

    Mmd aggregated two-sample test

    Antonin Schrab et al. Mmd aggregated two-sample test. InProceedings of the 2023 Conference on Learning Theory, 2023

    2023

  29. [29]

    A. S. Palli, J. Jaafar, H. M. Gomes, M. A. Hashmani, and A. R. Gilal. An experimental analysis of drift detection methods on multi-class imbalanced data streams.Applied Sciences, 12(22):11688, 2022

    2022

  30. [30]

    Concept drift detection delay index.IEEE Transactions on Knowledge and Data Engineering, 35(5):4585–4597, 2022

    Anjin Liu, Jie Lu, Yiliao Song, Junyu Xuan, and Guangquan Zhang. Concept drift detection delay index.IEEE Transactions on Knowledge and Data Engineering, 35(5):4585–4597, 2022. 6 When Drift Detectors Cry Wolf: False Alarm Rates in Continuous ML Monitoring A Appendix A: Experimental Details A.1 Dataset Preprocessing Table 2: Dataset preprocessing summary. ...

    2022