Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

High-Frequency Market Manipulation Detection with a Markov-modulated Hawkes process

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

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

Pith's one-line read The paper shows that a regime-switching self-exciting point process, estimated by a new EM algorithm, flags rare bursts of zero-price-return trades as suspicious and detects them out-of-sample on a cryptocurrency exchange.

desk verdict Solid estimation paper with a credible unsupervised burst detector; the wash-trading label outruns the evidence, but the authors say so themselves. read the letter →

arxiv 2502.04027 v1 pith:VBKBKSWT submitted 2025-02-06 stat.ME q-fin.STq-fin.TR

Timothée Fabre , Ioane Muni Toke This is my paper · ORCID
classification stat.MEq-fin.STq-fin.TR MSC 60G5562M0562P0562F10
keywords Markov-modulatedHawkesprocessexpectation-maximizationregimeswitchingpointwashtradingmarketmanipulationcryptocurrencyViterbialgorithm
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 proposes a point process whose self-exciting Hawkes intensity switches between unobserved regimes through a hidden Markov chain, and makes the intensity piecewise constant between events with a discretization step $\delta$. The authors develop an expectation-maximization algorithm that estimates both the Hawkes parameters and the Markov transition rates, and they validate its convergence on simulations. Their central claim is that the resulting MMHP-$\delta$ model, applied to zero-price-return trades on a top centralized cryptocurrency exchange, isolates extreme bursts of same-price trades that look like wash trading or pinging. They report better out-of-sample fit than a Markov-modulated Poisson process and a more conservative detection of suspicious volume. If the claim holds, exchanges could flag suspicious bursts in near real time without trader identities.

What carries the argument

The central object is the MMHP-$\delta$ process: a Hawkes-like intensity whose kernel is held constant on subintervals of length $\delta$ between events, so that the decay is piecewise constant rather than continuous. This specification makes the forward and backward transition matrices products of matrix exponentials of the matrices $Q-\Lambda_t$, and the EM E-step reduces the expected occupation times, transition counts, and integrated intensities to upper-right blocks of exponentials of block matrices. The same matrix-exponential machinery drives the goodness-of-fit residuals and the historical and online Viterbi algorithms.

What would settle it

Run the identical detector on a cryptocurrency asset for which the exchange has provided trader identifiers or confirmed wash-trade flags; if the state-3 bursts are not enriched in confirmed wash trades relative to equally intense random bursts, the detection claim is falsified.

Watch

Extended reading notes

Core claim

The central claim is that a Markov-modulated Hawkes process with a $\delta$-piecewise constant kernel is identifiable in practice and useful for market surveillance. The paper obtains closed-form forward and backward transition matrices as products of matrix exponentials, which turns the EM expectation step into matrix-exponential computations, and a Viterbi algorithm reconstructs the hidden state sequence. On the zero-price-return trade subsample restricted to 5:00–12:00 UTC, the fitted three-state model spends 90.31% of calendar time in normal activity, 9.55% in high activity, and 0.14% in extreme bursts, yet the extreme-burst state accounts for 24.20% of buy volume and 21.39% of sell volume, about $216M over the test period. The paper further shows that entering this burst state breaks the usual liquidity-imbalance relation and is followed by significant mid-price moves, which it interprets as evidence of potential manipulation.

Load-bearing premise

The load-bearing premise is that the zero-price-return trade arrivals in the 5:00–12:00 UTC window are generated by a three-state MMHP-$\delta$; if the model is misspecified, the recovered states and the 'suspicious' label need not correspond to actual manipulative trading.

Editorial extensions

If this is right

  • The same EM procedure applies to any non-negative decreasing kernel, not only the exponential kernels used in the numerical experiments.
  • As $\delta \to 0$, the piecewise-constant model converges to the continuous Markov-modulated Hawkes process, and residual tests on simulated data support this approximation.
  • On the real data, MMHP-$\delta$ with $\delta=1$ second fits out-of-sample zero-return trades better than the Markov-modulated Poisson process and flags a smaller, more conservative share of volume as suspicious.
  • Both historical and online Viterbi state estimation are provided, so the same fitted model can produce a live feed of regime classifications.
  • The extreme-burst state is associated with an inverted liquidity-imbalance distribution and subsequent mid-price moves, a pattern that could allow a manipulative agent to move prices while seeming to trade at a standstill.

Reading between the lines

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

  • A testable extension is to run the fitted detector on other illiquid newly listed tokens and check whether the state-3 volume share stays in the 20–25% range; if it does not, the reported thresholds are specific to this asset.
  • Because state-3 bursts occur when the book is imbalanced the wrong way, the model may be capturing high-frequency ping or quote-stuffing strategies generally, not only wash trading; separating those cases would need order-cancellation or account-level data.
  • Combining the online Viterbi feed with a second-stage rule on trade size or queue consumption would give exchanges a near-real-time alerting system, a step the paper mentions but does not build.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 introduces a Markov-modulated Hawkes process (MMHP) with a δ-piecewise constant intensity kernel, derives an EM algorithm for parameter estimation, provides a Viterbi algorithm for state decoding in historical and online settings, and validates the estimators on simulated data. In the application, the model is fit to zero-price-return trades from the SEI-USD market on Coinbase during a restricted intraday window, and the Viterbi state with the largest excitation parameter is interpreted as an extreme-burst "suspicious" regime. The authors report better out-of-sample goodness-of-fit than a Markov-modulated Poisson process and claim the model can detect suspicious trading activity and potential market manipulation.

Significance. The methodological contribution is substantial and carefully developed. The δ-piecewise constant specification makes the forward and backward transition matrices explicit, the E-step is derived analytically using Van Loan-type matrix-exponential integrals, and the simulation study (Section 3.5) convincingly demonstrates convergence with respect to EM steps, sample size, and δ. This is a useful extension of the MMHPSD of Wang (2010) and the analytic E-step is a nontrivial contribution. The out-of-sample goodness-of-fit comparison with the MMPP is also clearly presented and shows an advantage for the MMHP-δ. The empirical detection claim, however, is not supported by ground-truth validation: the "suspicious" state is identified solely as the high-excitation Viterbi state, with no labels, account identities, or independent confirmation that these bursts are manipulative rather than ordinary clustered trading. The title and abstract overstate the detection aspect relative to what the evidence supports.

major comments (3)
  1. [Section 4.2.2] The central application claim that state 3 corresponds to "suspicious trading activity" is not validated against any ground truth. The Viterbi state with the largest α is, by construction, the state with the strongest self-excitation, and any M=3 Hawkes-type model will allocate an upper-tail state to bursty periods. Without labeled data, account identifiers, or independent cross-checks, the paper cannot distinguish manipulation from legitimate clustered trading, market-making activity, or iceberg-order execution. The conclusion's caveat that the method is indirect and "prone to false positives" is appropriate, but the abstract, title, and Section 4.2.2 text currently assert detection relevance without this qualification. Please either (i) add external validation (e.g., known manipulation episodes or comparison with direct wash-trading detection from other data), or (ii) reframe the empirical contribution as anomaly/burst detection and substantially soften the manipulation language throughout.
  2. [Section 4.2.2 and Section 2] The zero-price-return filter used to construct the subsample {τ_k} is not justified with respect to the MMHP model class. Selecting trades with zero price return depends on the order book and price process, which are outside the modeled filtration; the resulting thinned process is not guaranteed to remain an MMHP. Consequently, the excellent out-of-sample QQ plots in Figure 7 may only show that the MMHP-δ family is flexible enough to fit the filtered event stream, not that the estimated states reveal manipulation. The authors should either prove or simulate that zero-price-return thinning preserves the MMHP structure under reasonable order-book assumptions, or explicitly model the thinning as a marked point process and discuss how this affects interpretation.
  3. [Section 4.2.2 and Table 1] The model selection step is inconsistent with the subsequent analysis. Table 1 shows that AIC favors M=4 with δ=1, yet the paper fixes M=3 for all empirical results, citing computational cost. The choice of M=3 materially affects which events are assigned to the extreme-burst state, so the detection statistics (e.g., 24.20% of buy volume classified as suspicious) may be sensitive to this arbitrary choice. Please report the sensitivity of the main detection results to M (and to δ) or provide a principled justification for the selected configuration beyond AIC ranks.
minor comments (5)
  1. [Appendix B] Figure 13 caption contains "Janurary" and Figure 14 contains "3th"; these should be corrected to "January" and "3rd".
  2. [Appendix A.5] The heading "Proof of Proportion 3.1" should read "Proof of Proposition 3.1".
  3. [Section 2.2] The definition of ω(t) uses the notation "Nu − Nu− = 1", which is non-standard for a jump indicator; please state explicitly that this is a jump at time u and align the notation with that used for t_n.
  4. [Section 3.1] The sentence "Note that we do not consider that the process N jumps at time 0" is confusing given the convention t_0 = 0; please clarify that the counting process has no event at time 0.
  5. [Table 1] The statement that δ = 100s "could be considered as a proxy of the MMHPSD process" needs elaboration, since the MMHP-δ differs from the MMHPSD even for large δ; please explain the intended sense of approximation.

Circularity Check

1 steps flagged · score 6.0 of 10

The MMHP-δ estimation methodology is self-contained and validated on simulations; however, the empirical 'suspicious trading' detection reduces to renaming the fitted maximum-excitation Viterbi state as suspicious, without external ground truth.

  1. renaming known result [Section 4.2.2, 'Suspicious trading detection with the MMHP-δ model']
    "The states are sorted from 1 to 3 with respect to the excitation parameter α of the exponential kernel form ϕ(t) = αe^{−βt}. The larger the hidden state, the larger this parameter. ... In our setting, trades classified in state 3 are natural candidates for suspicious trading activity."

    State 3 is, by construction, the fitted state with the largest Hawkes excitation parameter α. Declaring state 3 trades 'suspicious' therefore equates the detection output with a relabeling of the model's own high-excitation state. A 3-state MMHP fitted to any bursty zero-price-return trade sequence will have a maximum-α state, so the 'detection' is guaranteed by the model specification rather than established by independent evidence of manipulation. The paper provides no labeled trades, account IDs, or external benchmark; the reported 24.20%/21.39% suspicious volume is just the volume in state 3.

full rationale

The mathematical core of the paper is not circular. The MMHP-δ likelihood, forward/backward ODEs, EM updates, and Viterbi recursions are derived explicitly from the model definition; the simulation experiments (Section 3.5) estimate parameters on data generated from a known MMHP and show convergence, which is an external check on the estimator. The application also contains a genuine out-of-sample component: models are estimated on one weekday and evaluated on the same weekday of the following week (Section 4.1), and the goodness-of-fit Q-Q plots compare MMHP-δ against an MMPP benchmark. No load-bearing self-citation chain was found; the citations to the authors' own prior work are background references, not used to force the model choice. The circularity is confined to the empirical detection claim. The paper defines the suspicious class as the Viterbi state with maximal excitation parameter, so the finding that the model 'detects' extreme bursts of zero-price-return trades is a restatement of how state 3 was constructed. The additional market-characterization results (imbalance skew, price response) are independent descriptions, but they do not validate the manipulation label, and the conclusion explicitly limits the method to indirect detection with false positives. Because the central empirical claim partially reduces to a renaming of the fitted state, while the estimation methodology and out-of-sample fit are self-contained, a moderate circularity score is appropriate.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claims rest on standard point process assumptions plus the structural assumption that the zero-price-return trade flow is well described by an MMHP-delta with exponential kernels. The model parameters, delta, and M are fitted or chosen by hand; no new physical or latent entities are introduced beyond the hidden Markov states, which are standard.

free parameters (6)
  • delta (discretization step) = 1 s (10 s and 100 s also tested)
    Chosen by hand to trade off approximation accuracy and computational cost; affects the likelihood computation and the granularity of state detection.
  • M (number of hidden states) = 3 (2 and 4 also tested)
    Selected using AIC; the subsequent 'suspicious' classification depends on this choice.
  • mu_i (baseline intensities) = estimated daily, not tabulated
    Fitted via EM for each state; part of the intensity specification.
  • alpha_i (excitation amplitudes) = estimated daily, not tabulated
    Fitted via EM; states are sorted by alpha and the highest-alpha state is labeled suspicious.
  • beta_i (exponential decay rates) = estimated daily, not tabulated
    Fitted via EM; controls how quickly past events excite future arrivals.
  • q_ij (Markov transition rates) = stationary distribution roughly 90.31%, 9.55%, 0.14%
    Fitted via EM; determines the frequency and persistence of the extreme burst state.
assumptions (5)
  • domain assumption The observed point process N has intensity given by Eqs. (3)-(5), a Markov-modulated Hawkes process with state-dependent baseline and kernel.
    The entire likelihood construction starts from this specification; if the true data are not generated by such a process, the estimates and state classification lack meaning.
  • standard math The hidden Markov chain S is irreducible, aperiodic, and has finite state space with generator Q.
    Used for the forward-backward transition matrices and the existence of a stationary distribution.
  • domain assumption The likelihood factorization in Proposition 3.1, adapted from Wang (2010), holds for the MMHP.
    Separates the Markov chain term from the Hawkes term; the EM updates rely on this decomposition being exact.
  • domain assumption The delta-piecewise constant kernel preserves the Markov property on the grid, allowing explicit matrix exponentials (Proposition 2.3).
    This is the key modeling trick; it is an approximation to a continuous kernel and its validity is demonstrated only numerically.
  • domain assumption Residual transformed durations tau_n are i.i.d. Exp(1) under the fitted model (Section 3.4).
    Used for goodness-of-fit via Q-Q plots; relies on correct model specification and plug-in estimation of the intensity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of High-Frequency Market Manipulation Detection with a Markov-modulated Hawkes process." pith.science (2026). https://pith.science/paper/VBKBKSWT

@misc{pith2026250204027,
  author       = {Pith},
  title        = {Pith review of: High-Frequency Market Manipulation Detection with a Markov-modulated Hawkes process},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VBKBKSWT}},
  note         = {Machine review of arXiv:2502.04027}
}
read the original abstract

This work focuses on a self-exciting point process defined by a Hawkes-like intensity and a switching mechanism based on a hidden Markov chain. Previous works in such a setting assume constant intensities between consecutive events. We extend the model to general Hawkes excitation kernels that are piecewise constant between events. We develop an expectation-maximization algorithm for the statistical inference of the Hawkes intensities parameters as well as the state transition probabilities. The numerical convergence of the estimators is extensively tested on simulated data. Using high-frequency cryptocurrency data on a top centralized exchange, we apply the model to the detection of anomalous bursts of trades. We benchmark the goodness-of-fit of the model with the Markov-modulated Poisson process and demonstrate the relevance of the model in detecting suspicious activities.

Figures

Figures reproduced from arXiv: 2502.04027 by the authors.

Figure 1
Figure 1. An illustration of the dynamics of MMHP- [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Convergence with respect to the number of EM steps — Box plots of the parameters with respect to the number of EM steps. Whiskers represent 5% and 95% quantiles. The plain line represents the true parameter. However, note that the estimation procedure holds for any general positive kernel. The parameters used in the experiment are M = 2, µ = (1.0, 1.0)′ , α = (1.0, 4.0)′ , β = (2.0, 10.0)′ , Q =  −1.0 1.0 1.0 −1.0 … view at source ↗
Figure 3
Figure 3. Convergence with respect to the number of events — Box plots of the parameters with respect to the number of events. Whiskers represent 5% and 95% quantiles. The plain line represents the true parameter. 102 103 Number of events 10−1 100 101 σ K K−0.5 K−1 (a) µ1 102 103 Number of events 10−1 100 101 σ K K−0.5 K−1 (b) µ2 102 103 Number of events 10−2 10−1 100 σ K K−0.5 K−1 (c) α1 102 103 Number of events 10−1 100 101… view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Convergence with respect to the number of events — Standard deviation of the esti￾mated parameters with respect to the number of events. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Convergence towards the continuous MMHP — Box plots of the p-values of the Komogorov-Smirnov test for three values of δ, and average QQ-plots of the simulation experi￾ments with respect to the parameter δ. For the box plots, the dashed line represents the 5% threshold …
Figure 6
Figure 6. Figure 6: Intraday seasonality — Intensity of marketable orders expressed in number of events per second, computed over the day for the weekday seasonality, and over bins of 5 minutes for the intraday seasonality. The intensity is averaged over the 4 months. The error bar of the…
Figure 7
Figure 7. Figure 7: Goodness-of-fit — Median QQ plots and their 95% confidence intervals, computed on a daily basis over the out-of-sample data from December 8th, 2023 to March 31st, 2024. Confidence intervals are computed via bootstrapping. 1 to 3 with respect to the excitation parameter…
Figure 8
Figure 8. Figure 8: Suspicious trading activity — An example of “suspicious” behavior identified by the bid MMHP-δ as the extreme burst regime (state 3), anomalies detected on December 26th, 2023 UTC time zone. We show a typical example of suspicious trading activity that is spotted by th…
Figure 9
Figure 9. Figure 9: Suspicious trading activity — Market characterization of the trading activity regimes identified by the MMHP-δ model. The price response is expressed in basis points and the horizon in number of mid price moves. for the intensity of buy trades). But this measure does n…
Figure 10
Figure 10. Figure 10: Convergence towards the continuous MMHP — Box plots of the parameters with respect to the parameter δ. Whiskers represent 5% and 95% quantiles. The plain line represents the true parameter used for the simulation of the MMHP with continuous kernel. and Z ∆n 0 e (Q−Λtn…
Figure 11
Figure 11. Figure 11: Convergence with respect to δ — Box plots of the parameters with respect to the parameter δ. Whiskers represent 5% and 95% quantiles. The plain line represents the true parameter. 08:51:56 08:52:02 08:52:09 t 0.73850 0.73875 0.73900 0.73925 0.73950 0.73975 best bid pr…
Figure 12
Figure 12. Figure 12: Suspicious trading activity — An example of “suspicious” behavior identified by the bid MMHP-δ as the extreme burst regime (state 3), anomalies detected on March 6th, 2024 UTC time zone. 33 [PITH_FULL_IMAGE:figures/full_fig_p033_12.png]
Figure 13
Figure 13. Figure 13: Suspicious trading activity — An example of “suspicious” behavior identified by the ask MMHP-δ as the extreme burst regime (state 3), anomalies detected on Janurary 15th, 2024 UTC time zone. 11:31:33 11:31:34 11:31:36 t 0.8356 0.8358 0.8360 0.8362 0.8364 best bid pric…
Figure 14
Figure 14. Figure 14: Suspicious trading activity — An example of “suspicious” behavior identified by the ask MMHP-δ as the extreme burst regime (state 3), anomalies detected on March 3th, 2024 UTC time zone. 34 [PITH_FULL_IMAGE:figures/full_fig_p034_14.png]
Figure 15
Figure 15. Figure 15: Suspicious trading activity — Market characterization of the trading activity regimes identified by the MMHP-δ model, when the model is trained on December 1st, 2023, without taking into account the intraday seasonality. 35 [PITH_FULL_IMAGE:figures/full_fig_p035_15.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Clustering-Based Framework for Identifying Suspicious Trading Patterns in Capital Market

    cs.AI 2026-07 conditional novelty 3.0 of 10

    K-Means++ plus percentile and price-change heuristics flag 2.02% of ~1M DSE trades as suspicious and assign mostly spoofing or unclassified labels, with only a 0.561 silhouette score as validation.

Reference graph

Works this paper leans on

36 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [1]

    and Li, J

    Aloosh, A. and Li, J. (2024). Direct evidence of bitcoin wash trading. Management Science

    work page 2024

  2. [2]

    and Muzy, J.-F

    Bacry, E. and Muzy, J.-F. (2016). First-and second-order statistics characterization of H awkes processes and non-parametric estimation. IEEE Transactions on Information Theory , 62(4):2184--2202

    work page 2016

  3. [3]

    Bergault, P., Bertucci, L., Bouba, D., Gu \'e ant, O., and Guilbert, J. (2024). Price-aware automated market makers: Models beyond brownian prices and static liquidity. arXiv preprint arXiv:2405.03496

    arXiv 2024

  4. [4]

    Bouchaud, J.-P., Gefen, Y., Potters, M., and Wyart, M. (2003). Fluctuations and response in financial markets: the subtle nature of random price changes. Quantitative finance , 4(2):176

    work page 2003

  5. [5]

    Cao, Y., Li, Y., Coleman, S., Belatreche, A., and McGinnity, T. M. (2015). Detecting wash trade in financial market using digraphs and dynamic programming. IEEE transactions on neural networks and learning systems , 27(11):2351--2363

    work page 2015

  6. [6]

    Chen, S., Chen, J., Yu, J., Luo, X., Wang, Y., and Zheng, Z. (2023). The dark side of NFT s: A large-scale empirical study of wash trading. arXiv preprint arXiv:2312.12544

    work page Pith review arXiv 2023

  7. [7]

    W., Li, X., Tang, K., and Yang, Y

    Cong, L. W., Li, X., Tang, K., and Yang, Y. (2023). Crypto wash trading. Management Science , 69(11):6427--6454

    work page 2023

  8. [8]

    J., Vere-Jones, D., et al

    Daley, D. J., Vere-Jones, D., et al. (2003). An introduction to the theory of point processes: volume I: elementary theory and methods . Springer

    work page 2003

Show all 36 references
  1. [9]

    Davison, A., Ramesh, N., Barnett, V., and Turkman, K. (1993). A stochastic model for times of exposures to air pollution from a point source. Statistics for the Environment , pages 123--138

    1993

  2. [10]

    and Muni Toke , I

    Fabre, T. and Muni Toke , I. (2024). Neural H awkes: Non-parametric estimation in high dimension and causality analysis in cryptocurrency markets. arXiv preprint arXiv:2401.09361

    2024

  3. [11]

    H., Tsoukalas, G., and Zhang, N

    Falk, B. H., Tsoukalas, G., and Zhang, N. (2023). Crypto wash trading: Direct vs. indirect estimation. arXiv preprint arXiv:2311.18717

    2023 arXiv

  4. [12]

    Fosset, A., Bouchaud, J.-P., and Benzaquen, M. (2022). Non-parametric estimation of quadratic H awkes processes for order book events. The European Journal of Finance , 28(7):663--678

    2022

  5. [13]

    and Tas, B

    Imisiker, S. and Tas, B. K. O. (2018). Wash trades as a stock market manipulation tool. Journal of behavioral and experimental finance , 20:92--98

    2018

  6. [14]

    Jayant, A. (2023). The economics of wash trading. Available at SSRN

    2023

  7. [15]

    and Saito, H

    Kawashima, K. and Saito, H. (1990). Teletraffic issues in atm networks. Computer Networks and ISDN Systems , 20(1-5):369--375

    1990

  8. [16]

    Krolzig, H.-M. (2013). Markov-switching vector autoregressions: Modelling, statistical inference, and application to business cycle analysis , volume 454. Springer Science & Business Media

    2013

  9. [17]

    and Abergel, F

    Lu, X. and Abergel, F. (2018). High-dimensional H awkes processes for limit order books: modelling, empirical analysis and numerical calibration. Quantitative Finance , 18(2):249--264

    2018

  10. [18]

    Meier-Hellstern, K. (1984). A statistical procedure for fitting M arkov modulated P oisson processes. Technical report, Univ. of Delaware

    1984

  11. [19]

    Meier-Hellstern, K. S. (1987). A fitting algorithm for M arkov-modulated P oisson processes having two arrival rates. European Journal of Operational Research , 29(3):370--377

    1987

  12. [20]

    and Pakkanen, M

    Morariu-Patrichi, M. and Pakkanen, M. S. (2022). State-dependent H awkes processes and their application to limit order book modelling. Quantitative Finance , 22(3):563--583

    2022

  13. [21]

    and Yoshida, N

    Muni Toke, I. and Yoshida, N. (2017). Modelling intensities of order flows in a limit order book. Quantitative Finance , 17(5):683--701

    2017

  14. [22]

    and Yoshida, N

    Muni Toke, I. and Yoshida, N. (2020). Analyzing order flows in limit order books with ratios of C ox-type intensities. Quantitative Finance , 20(1):81--98

    2020

  15. [23]

    and Neugebauer, T

    Nekrasova, E. and Neugebauer, T. (2023). Does wash trading impact information dissemination in security markets? Available at SSRN 4644462

    2023

  16. [24]

    and Spagnolo, N

    Psaradakis, Z. and Spagnolo, N. (2003). On the determination of the number of regimes in markov-switching autoregressive models. Journal of time series analysis , 24(2):237--252

    2003

  17. [25]

    Rambaldi, M., Bacry, E., and Lillo, F. (2017). The role of volume in order book dynamics: a multivariate H awkes process analysis. Quantitative Finance , 17(7):999--1020

    2017

  18. [26]

    J., Ephraim, Y., and Dieguez, E

    Roberts, W. J., Ephraim, Y., and Dieguez, E. (2006). On R ydén's EM algorithm for estimating MMPP s. IEEE Signal Processing Letters , 13(6):373--376

    2006

  19. [27]

    Ryd \'e n, T. (1994). Parameter estimation for M arkov modulated P oisson processes. Stochastic Models , 10(4):795--829

    1994

  20. [28]

    Ryd \'e n, T. (1996). An EM algorithm for estimation in M arkov-modulated P oisson processes. Computational Statistics & Data Analysis , 21(4):431--447

    1996

  21. [29]

    and Muni Toke , I

    Sfendourakis, E. and Muni Toke , I. (2023). L O B modeling using H awkes processes with a state-dependent factor. Market Microstructure and Liquidity

    2023

  22. [30]

    To s i \'c , A., Hrovatin, N., and Vi c i c , J. (2023). Beyond the surface: Advanced wash trading detection in decentralized NFT markets. arXiv preprint arXiv:2312.16603

    2023 arXiv

  23. [31]

    Van Loan, C. (1978). Computing integrals involving the matrix exponential. IEEE transactions on automatic control , 23(3):395--404

    1978

  24. [32]

    and Weintraud, A

    Victor, F. and Weintraud, A. M. (2021). Detecting and quantifying wash trading on decentralized cryptocurrency exchanges. In Proceedings of the Web Conference 2021 , pages 23--32

    2021

  25. [33]

    Wang, T. (2010). Statistical models for earthquakes incorporating ancillary data . PhD thesis, Massey University

    2010

  26. [34]

    Wang, T., Bebbington, M., and Harte, D. (2012). M arkov-modulated H awkes process with stepwise decay. Annals of the Institute of Statistical Mathematics , 64:521--544

    2012

  27. [35]

    G., Curley, J., and Zheng, T

    Wu, J., Ward, O. G., Curley, J., and Zheng, T. (2022). M arkov-modulated H awkes processes for modeling sporadic and bursty event occurrences in social interactions. The Annals of Applied Statistics , 16(2):1171--1190

    2022

  28. [36]

    Wu, P., Rambaldi, M., Muzy, J.-F., and Bacry, E. (2023). Queue-reactive H awkes models for the order flow. Market microstructure and liquidity

    2023

Pith tools

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