Pith. sign in

REVIEW 3 major objections 5 minor 28 references

Velocity- and Regime-Aware Detection of Intraday Options Market Manipulation, with Explainable Attribution

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read One minute-level velocity signature of pump-and-crash manipulation recovers all ten regulator-identified options-manipulation days and transfers to U.S. equities.

desk verdict Careful time-partitioned evaluation and honest negative results, but the central 'velocity' claim is not backed by the paper's own SHAP numbers or an ablation. read the letter →

arxiv 2608.05373 v1 pith:6PPX5575 submitted 2026-08-05 q-fin.TR cs.LGq-fin.ST

classification q-fin.TRcs.LGq-fin.ST
keywords marketmanipulationdetectionanomalyDeltavelocitypump-and-reversalSHAPattributionhiddenMarkovmodeloptionsmarketsenforcementlabels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that intraday pump-and-crash manipulation leaves a dynamic signature in the velocity of market state—a sharp rise immediately followed by collapse—rather than in price levels or volatility magnitude. On a locked out-of-sample year for Indian BANKNIFTY index options, a minute-level autoencoder using smoothed $\Delta$ velocity recovers all ten regulator-identified manipulation days, with precision near 25 percent that the authors argue is a floor caused by incomplete enforcement labels rather than detector failure. The same shape-based signature ranks alleged manipulation days in thinly traded U.S. equities with AUC 0.91 and 0.81, and exact SHAP attribution shows unconfirmed alerts share the confirmed days' attribution profile. If right, the paper gives regulators a leak-safe, explainable surveillance screen and isolates the quantity that transfers across markets: the shape of the move, not its size.

What carries the argument

The load-bearing object is the smoothed state-velocity feature $v_t = |\bar{\Delta}^{(15)}_t - \bar{\Delta}^{(15)}_{t-1}|/(\tau_t - \tau_{t-1})/\text{minute}$, the absolute one-minute rate of change of the trailing 15-observation mean of option $\Delta$, blanked when quote gaps fall outside 0.5–2 minutes and grouped by day to avoid overnight jumps. This feature writes the minute-scale pump-then-crash sequence into every row, so even a row-level autoencoder can see the dynamic. For equities, the same shape is captured by a prespecified pump-reversal score (price rises over ten minutes, gives back within five, weighted by volume participation), and exact SHAP attribution (enumerating all $2^3$ feature coalitions) explains each alert in units of reconstruction error.

What would settle it

Audit the 31 unconfirmed BANKNIFTY alert days (41 alerts minus the 10 regulator-identified days) using participant-attributed order data or later adjudication: if none shows manipulative trading, or if the SHAP profile of those alerts matches ordinary crisis days as closely as it matches the confirmed days, the precision floor is a detector failure and the practical-utility claim is weakened.

Watch

Extended reading notes

Core claim

The central discovery is that a specific dynamic pattern, called the "Mountain," separates manipulation from ordinary volatility: option $\Delta$ is forced up sharply over minutes and crashes back, producing two large velocity spikes. The paper builds a feature that measures this rate of change directly, $\Delta$ velocity, and shows that a plain autoencoder trained on it recovers 10 of 10 regulator-identified manipulation days in the sequestered 2024 BANKNIFTY test, while a pump-reversal shape score ranks SEC-alleged days at AUC 0.91 (ARQQ) and 0.81 (ACY). The paper also reports a negative result: conditioning detection on hidden-Markov volatility regimes trades recall for precision and does not resolve the manipulation-volatility overlap. Attribution over every alert shows unconfirmed alerts share the regulator-identified days' attribution profile (cosine similarity 0.9894), which the authors read as consistent with incomplete enforcement labels.

Load-bearing premise

The claim that near-25 percent precision is a label-completeness floor rather than detector failure rests on the assumption that the unconfirmed alerts sharing the regulator-identified days' attribution profile are themselves manipulation; the paper concedes those days could simply be extreme-but-legitimate volatility.

Editorial extensions

If this is right

  • A public-data velocity screen can rank trading days for forensic review: in the BANKNIFTY case it flags all ten regulator-identified days and 31 additional unconfirmed days that share their attribution profile.
  • Regime-aware detection does not separate manipulation from legitimate volatility; the hidden-Markov conditioning only reduces alert volume at the cost of recall, so the bottleneck is label scope, not architecture.
  • The shape of the signature transfers across instrument types, but its magnitude does not, so velocity-based detectors should be designed around shape, not thresholds on raw magnitude.
  • Days attributed to marking the close (rather than pump-and-reversal) are poorly visible to velocity screens, indicating that suppression-style manipulation requires participant-attributed data rather than public price bars.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the label-incompleteness reading is correct, the same detector should flag additional days when future enforcement orders or adjudications expand the labeled set; this is a testable prediction the paper does not itself make.
  • The cosine-similarity attribution profile could be used as a triage score in practice, but the paper's own caveat means a shared profile is not a fingerprint; it needs validation against independently adjudicated normal days.
  • A natural extension is to test whether the velocity signature detects pump-and-dump episodes in other thin markets such as crypto or small-cap equities, where enforcement labels exist but are scarce.
  • The marking-the-close negative result suggests a hybrid screen: public velocity shape for pump-reversals, and participant-attributed volume share for suppression, mirroring the two enforcement mechanisms.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a minute-level detection pipeline for intraday market manipulation based on a "pump-and-crash" signature in the velocity of market state: option-Delta velocity for BANKNIFTY index options and price-velocity shape scores for U.S. equities. The protocol is strictly time-partitioned: models are trained on 2022, thresholds are calibrated on 2023 and locked, and 2024 is held out. The plain autoencoder is reported to recover 10 of 10 regulator-identified manipulation days on the locked 2024 test, at roughly 24% closed-world precision. A hidden-Markov-regime conditioning experiment yields a negative result, and exact SHAP attribution over alert days is used to argue that unconfirmed alerts share the attribution profile of regulator-identified days. Cross-market validation on SEC v. Patel equities alleges transfer of the shape signature with AUC 0.91 for ARQQ and 0.81 for ACY.

Significance. If the central velocity claim holds, the paper provides a rare enforcement-linked, leakage-safe evaluation of manipulation detection at minute resolution, with a clean time-partitioned protocol, locked thresholds, and exact SHAP attribution. These are real strengths: the protocol prevents label leakage, the SHAP additivity error is machine-checked, and the claimed transfer to a different market and instrument is an unusual falsifiable result. However, the significance depends on the velocity feature actually being the operative signal, and the current evidence for that is indirect: no ablation or feature-subset rerun is reported, and the SHAP attribution assigns Delta_Velocity only 20.5% of average absolute attribution. The interpretation of the roughly 25% precision as a floor due to incomplete enforcement labels is plausible but not demonstrated, and the label set itself was selected by a profitability screen. The paper is honest about many of these limitations, but the load-bearing claims need additional verification before the contribution can be accepted as stated.

major comments (3)
  1. [Section 7, Figure 6, Table 1] The central claim of a "velocity-based" detector is not supported by the attribution evidence. Section 7 reports mean absolute SHAP attribution of Option_Price 40.5%, Implied_Vol 39.0%, and Delta_Velocity 20.5% across the 41 alert days, so the model's decisions are dominated by static price and volatility features rather than by the Eq. (1) velocity feature. No experiment removes Delta_Velocity from the feature set or reruns the locked-evaluation protocol with price and implied volatility only. Without such an ablation or a two-feature rerun, the 10/10 recall could be driven by generic price/volatility outliers that coincide with the labeled days; the title's velocity mechanism would then be a narrative overlay rather than the operative signal. This is load-bearing because the cross-market transfer in Section 6 uses a different shape score and therefore cannot validate the BANKNIFTY autoencoder's mechanism. Please add a locked-threshold ablation or a two-feature rerun, and report whether recall and precision change.
  2. [Section 8, "Label scope" and Section 5] The interpretation of the roughly 25% precision as a floor caused by incomplete enforcement labels is an unverified assumption, not a demonstrated result. Section 8 states that the enforcement order formed its labeled set by selecting the 30 most profitable days for minute-level review, and that unlisted days are unverified. The paper's own Section 7 concession that "a shared profile could also characterize extreme days in general rather than manipulation in particular" acknowledges the alternative: the unconfirmed alerts may be ordinary extreme-volatility days, in which case the low precision reflects genuine detector failure. The matched comparison in Section 7 is suggestive but does not resolve this, because it only shows that identified days differ from calm non-alert days, not that unconfirmed alerts are manipulation-like. Please either provide a falsifiable test that distinguishes these hypotheses (for example, using contract-level or event-context evidence on a sample of unconfirmed alerts) or materially weaken the claim that the precision ceiling is consistent with label incompleteness rather than detector failure.
  3. [Section 6, "prespecified pump-reversal shape score"] The cross-market validation relies on a shape-score parameterization described as "prespecified," but no evidence is provided that the 10-minute rise and 5-minute give-back windows, or the volume participation weighting, were locked before the SEC v. Patel data were examined. This matters because the ACY result is weaker (AUC 0.81), and the paper reports that an unweighted variant outperforms the prespecified primary on ACY (0.87 versus 0.81), which suggests sensitivity to the scoring rule. Please report the source and date of prespecification, or provide a parameter-sensitivity analysis showing that the AUC conclusions are robust to reasonable perturbations of the windows. Without this, the cross-market transfer claim is not fully established.
minor comments (5)
  1. [Equation (1)] In Eq. (1), the denominator is written as "(tau_t - tau_{t-1})/minute"; please clarify the intended units and state explicitly that on uninterrupted one-minute data the velocity is expressed in Delta change per minute.
  2. [Section 7] The SHAP "exactness" is exact only for the deterministic 64-row sample per alert day, not for the full day's record; the paper says this, but the abstract and contribution list state "exact SHAP attribution over every alert" without that qualification. Please make the sampling explicit in the contribution statement.
  3. [Figure 6] The caption states that rows are not normalized to sum to one and that one color scale applies to both panels; this makes it difficult to compare attribution profiles across rows. Consider adding normalized profiles in a supplementary figure.
  4. [Reference [8]] The primary data source is a Kaggle dataset without a persistent DOI or versioned identifier; a frozen working copy is mentioned, but please include a checksum or a more permanent archival location to support reproducibility.
  5. [Section 5] The sentence "A detector built for the pump-and-reversal signature reported, accurately, which days contain that signature" is more definitive than the evidence supports; the detector was built with a three-feature set including price and volatility, not velocity alone, so "accurately" should be qualified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the held-out test recall is a genuine out-of-sample result under locked thresholds.

full rationale

The paper's central claims are supported by a strictly time-partitioned protocol: the autoencoder is trained on 2022, thresholds are calibrated on 2023, and the 2024 test set is sequestered until after thresholds are locked. The 10/10 recall on regulator-identified days is therefore not forced by construction: the test labels were not used to set the detection thresholds, and the model could have failed on those days. The Delta-velocity feature is indeed motivated by the documented pump-and-crash signature, but that is standard feature engineering informed by the known phenomenon; the evaluation still requires the feature to generalize to held-out days, which it does according to the reported results. The cross-market shape score is likewise applied to complaint-labeled ticker-days, but the labels are legal allegations, not a re-encoding of the score's 10-minute/5-minute parameters, so the AUC results remain empirical. The paper also reports honest negative results (regime conditioning trades recall for precision), discloses that Delta_Velocity carries only 20.5% of SHAP attribution, and explicitly concedes that a shared attribution profile could characterize extreme days generally rather than manipulation specifically. No load-bearing self-citation or uniqueness theorem is invoked. The only concerns—prespecification evidence for the shape-score parameters and the absence of a velocity-ablation experiment—are reproducibility and evidentiary issues, not circularity. Under the stated criteria, no derivation step reduces by definition to its inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central results depend on a small number of hand-chosen hyperparameters and domain assumptions about the labels and data quality. No new physical or theoretical entities are introduced. The most consequential assumption is the label-incompleteness interpretation used to explain the low precision.

free parameters (3)
  • Delta smoothing window length = 15 observations
    Chosen by hand; the velocity feature and all downstream detections depend on this window length.
  • Row and daily hit-rate alert thresholds = Calibrated on 2023 validation data
    The 10/10 test recall is conditional on thresholds tuned on the 2023 development year, which includes eight regulator-identified days.
  • Pump-reversal shape score windows = Rise: 10 minutes; give-back: 5 minutes
    The paper calls this prespecified from the complaint's description, but no evidence is given that the parameters were locked before seeing the test data, and multiple variants are reported for ACY.
assumptions (4)
  • domain assumption Black-Scholes Delta is an adequate summary of option market state for detecting the manipulation signature.
    The pipeline computes Delta with Black-Scholes and validates against U.S. vendor Greeks, but not against Indian market Greeks; the detection relies on this metric.
  • domain assumption The public one-minute BANKNIFTY options dataset is accurate and complete enough to compute the velocity features.
    The entire detection pipeline depends on this external Kaggle dataset, which is not redistributed and whose errors are not independently audited.
  • domain assumption Unlabeled days are unverified, not normal, so the closed-world precision is a floor, not a point estimate.
    The evaluation treats every unlisted day as negative for metric bookkeeping, and the paper argues this biases precision downward; if unlabeled days contain manipulation, the true precision could be higher.
  • ad hoc to paper The manipulation signature is best represented by pump-and-reversal dynamics in Delta velocity.
    The detector is built around this mechanism, and the paper uses the strategy split (pump-reversal vs. marking-the-close) to explain detection failures, which could make the detector appear successful by construction on the days it is designed to catch.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Velocity- and Regime-Aware Detection of Intraday Options Market Manipulation, with Explainable Attribution." pith.science (2026). https://pith.science/paper/6PPX5575

@misc{pith2026260805373,
  author       = {Pith},
  title        = {Pith review of: Velocity- and Regime-Aware Detection of Intraday Options Market Manipulation, with Explainable Attribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6PPX5575}},
  note         = {Machine review of arXiv:2608.05373}
}
read the original abstract

Intraday market manipulation is hard to detect because its footprint is brief, buried in millions of quotes, and statistically similar to ordinary volatility. Detectors reach high recall only by flagging so many other days that measured precision collapses, producing alerts no regulator can act on. We show that this manipulation leaves a distinctive dynamic signature: a pump-and-crash pattern visible in the velocity of market state, rather than its level. We build a minute-level detection pipeline, strictly partitioned in time, based on smoothed state velocity: option-Delta velocity for index options and price velocity for equities. We explain every alert with SHAP attribution. We hold the test period strictly out-of-sample and fix all thresholds before evaluation. On the locked Indian BANKNIFTY index-options test, the plain autoencoder recovers 10 of 10 regulator-identified manipulation days. Conditioning detection on market regimes inferred by a hidden Markov model yields an instructive negative result. The regimes are descriptively distinct, but using them trades recall for precision. Under the closed-world assumption that unlabeled days are normal, precision remains near 25%. The same dynamic appears in thinly traded U.S. equities (SEC v. Patel). The shape of the signature survives the transfer; its velocity magnitude does not. A pump-reversal shape score ranks the complaint's alleged manipulation days with AUC 0.91 (ARQQ) and 0.81 (ACY). On the ARQQ worked example, the score peaks inside the complaint's documented minute window. Finally, exact SHAP attribution over every alert shows that unconfirmed alerts share the regulator-identified days' attribution profile (cosine similarity 0.99). The precision ceiling is consistent with incomplete enforcement labels rather than detector failure. What transfers across markets and instrument types is the dynamic signature itself.

Figures

Figures reproduced from arXiv: 2608.05373 by the authors.

Figure 2
Figure 2. The same event in velocity space: the pump and the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. The BIC-selected four-state regime model, profiled [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. ARQQ, October 8, 2021 (one-minute bars). The com [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figures from the paper (1 more)
Figure 6
Figure 6. Figure 6: Per-day mean absolute SHAP attribution profiles for all 41 plain-autoencoder alert days of the locked 2024 test. Rows [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 19 canonical work pages

  1. [1]

    Aggarwal and Guojun Wu

    Rajesh K. Aggarwal and Guojun Wu. 2006. Stock Market Manipulations.The Journal of Business79, 4 (2006), 1915–1953. doi:10.1086/503652

  2. [2]

    Michael Aitken, Douglas Cumming, and Feng Zhan. 2015. High Frequency Trading and End-of-Day Price Dislocation.Journal of Banking & Finance59 (2015), 330–349. doi:10.1016/j.jbankfin.2015.06.011

  3. [3]

    Erdinc Akyildirim, Matteo Gambara, Josef Teichmann, and Syang Zhou. 2022. Ap- plications of Signature Methods to Market Anomaly Detection. arXiv:2201.02441

  4. [4]

    Schrager, Srijan Sood, Jared Vann, and Manuela Veloso

    Tucker Richard Balch, Veronica Mejia Bustamante, Naan Cho, Matthew Howard, Maxime Kawawa-Beaudan, Ganapathy Mani, Ivan Rankenburg, Andrew J. Schrager, Srijan Sood, Jared Vann, and Manuela Veloso. 2026. Method and system for detecting anomalous behavior in stream data. US Patent 12,585,762 B2, assigned to JPMorgan Chase Bank, N.A. Granted March 24, 2026

  5. [5]

    Fischer Black and Myron Scholes. 1973. The Pricing of Options and Corporate Liabilities.Journal of Political Economy81, 3 (1973), 637–654

  6. [6]

    David Diaz, Babis Theodoulidis, and Pedro Sampaio. 2011. Analysis of Stock Market Manipulations Using Knowledge Discovery Techniques Applied to Intra- day Trade Prices.Expert Systems with Applications38, 10 (2011). doi:10.1016/j. eswa.2011.04.066

  7. [7]

    Koosha Golmohammadi and Osmar R. Zaïane. 2015. Time Series Contextual Anomaly Detection for Detecting Market Manipulation in Stock Market. In Proceedings of the IEEE International Conference on Data Science and Advanced Analytics (DSAA). doi:10.1109/DSAA.2015.7344856

  8. [8]

    Ayush Gupta. 2023. Indian Nifty and Banknifty Options Data 2020–2024. Kaggle dataset. https://www.kaggle.com/datasets/ayushsacri/indian-nifty-and- banknifty-options-data-2020-2024. Accessed July 16, 2026. Redistributed data are not included; a frozen working copy is preserved in the project archive

Show all 28 references
  1. [9]

    Hamilton

    James D. Hamilton. 1989. A New Approach to the Economic Analysis of Nonsta- tionary Time Series and the Business Cycle.Econometrica57, 2 (1989), 357–384. doi:10.2307/1912559

  2. [10]

    John C. Hull. 2018.Options, Futures, and Other Derivatives(10 ed.). Pearson

  3. [11]

    Josh Kamps and Bennett Kleinberg. 2018. To the Moon: Defining and Detecting Cryptocurrency Pump-and-Dumps.Crime Science7, 1 (2018). doi:10.1186/s40163- 018-0093-5

  4. [12]

    Maxime Kawawa-Beaudan, Srijan Sood, Soham Palande, Ganapathy Mani, Tucker Balch, and Manuela Veloso. 2024. Ensemble methods for sequence classification with hidden Markov models.arXiv preprint arXiv:2409.07619(2024)

  5. [13]

    Massimo La Morgia, Alessandro Mei, Francesco Sassi, and Julinda Stefa. 2023. The Doge of Wall Street: Analysis and Detection of Pump and Dump Cryptocurrency Manipulations.ACM Transactions on Internet Technology23, 1 (2023). doi:10. 1145/3561300

  6. [14]

    Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. 2008. Isolation Forest. In Proceedings of the Eighth IEEE International Conference on Data Mining (ICDM). 413–422. doi:10.1109/ICDM.2008.17

  7. [15]

    Lundberg and Su-In Lee

    Scott M. Lundberg and Su-In Lee. 2017. A Unified Approach to Interpreting Model Predictions. InAdvances in Neural Information Processing Systems 30 (NeurIPS)

  8. [16]

    Karim Mahrous and Roberto Di Pietro. 2025. TSTR for Financial Fraud: Learn- ing to Detect Manipulation Without Real Data. InProceedings of the 6th ACM International Conference on AI in Finance (ICAIF). doi:10.1145/3768292.3770393

  9. [17]

    Stewart Mayhew. 2002. Competition, Market Structure, and Bid-Ask Spreads in Stock Option Markets.Journal of Finance57, 2 (2002), 931–958

  10. [18]

    Robert C. Merton. 1973. Theory of Rational Option Pricing.The Bell Journal of Economics and Management Science4, 1 (1973), 141–183

  11. [19]

    Dmitriy Muravyev and Neil D. Pearson. 2020. Options Trading Costs Are Lower than You Think.The Review of Financial Studies33, 11 (2020), 4973–5014. doi:10. 1093/rfs/hhaa010

  12. [20]

    Sandeep Neela. 2025. AIMM: An AI-Driven Multimodal Framework for Detecting Social-Media-Influenced Stock Market Manipulation. arXiv:2512.16103

  13. [21]

    Astudillo, and David Díaz

    Hugo Núñez Delafuente, César A. Astudillo, and David Díaz. 2024. Ensemble Approach Using k-Partitioned Isolation Forests for the Detection of Stock Market Manipulation.Mathematics12, 9 (2024), 1336. doi:10.3390/math12091336

  14. [22]

    Peter Nystrup, Henrik Madsen, and Erik Lindström. 2018. Dynamic Portfolio Optimization Across Hidden Market Regimes.Quantitative Finance18, 1 (2018), 83–95. doi:10.1080/14697688.2017.1342857

  15. [23]

    Guansong Pang, Chunhua Shen, Longbing Cao, and Anton van den Hengel. 2021. Deep Learning for Anomaly Detection: A Review.Comput. Surveys54, 2 (2021), 1–38. doi:10.1145/3439950

  16. [24]

    Lawrence R. Rabiner. 1989. A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition.Proc. IEEE77, 2 (1989), 257–286. doi:10. 1109/5.18626

  17. [25]

    Why Should I Trust You?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. “Why Should I Trust You?”: Explaining the Predictions of Any Classifier. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 1135–1144. doi:10.1145/2939672.2939778

  18. [26]

    Platt, John Shawe-Taylor, Alex J

    Bernhard Schölkopf, John C. Platt, John Shawe-Taylor, Alex J. Smola, and Robert C. Williamson. 2001. Estimating the Support of a High-Dimensional Distribution.Neural Computation13, 7 (2001), 1443–1471. doi:10.1162/ 089976601750264965

  19. [27]

    Securities and Exchange Board of India. 2025. Interim Order in the Matter of Index Manipulation by Jane Street Group. Order WTM/AN/MRD/MRD-SEC- 3/31013/2025-26, July 3, 2025. https://www.sebi.gov.in/enforcement/orders/jul- 2025/interim-order-in-the-matter-of-index-manipulation...

  20. [28]

    Securities and Exchange Commission

    U.S. Securities and Exchange Commission. 2026. SEC v. Harsh V. Patel, No. 1:26-cv-03203 (S.D.N.Y. filed Apr. 20, 2026). Complaint and Attachment A; SEC Litigation Release No. 26532. https://www.sec.gov/files/litigation/complaints/ 2026/comp26532.pdf

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.