{"id":"3e2945a3-da59-4624-ab2c-d08e7bb64a0c","arxiv_id":"2502.04027","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"The paper develops and estimates a Markov-modulated Hawkes process with piecewise constant decay and uses it to flag extreme trade bursts on a cryptocurrency exchange.","lead":"A new statistical model of trade arrivals alternates between hidden market states and flags extreme bursts of trades that may indicate wash trading on a crypto exchange. The authors extend the Markov-modulated Hawkes process to let the trade intensity decay between trades, estimate it with an EM algorithm, and test it on Coinbase SEI-USD data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The EM methodology is credible, but the empirical detection claim is not: state 3 is defined by high Hawkes excitation, with no ground-truth validation that these bursts are wash trading rather than ordinary clustered trading.","rationale":"The reader's weakest assumption and my stress-test identify the same load-bearing point: the empirical 'suspicious trading' conclusion rests on unvalidated labels. The methodological contribution is genuinely useful and independently supported by extensive simulation evidence. The empirical application, however, demonstrates only that the model finds extreme bursts of zero-price-return trades, not that those bursts are manipulative. The paper's concluding caveats already acknowledge this limitation, which is exactly why the conditional verdict is appropriate. A stronger verdict such as ACCEPT would overstate the evidence, while REJECT would ignore the solid methodological core. The proposed synthetic labeled benchmark would directly test whether the Viterbi state-3 flags recover known injected bursts, and would settle whether the application claim has real detection power or merely re-discovers ordinary clustering.","tokens_in":27164,"tokens_out":4654,"duration_ms":59793,"concrete_test":"Build a synthetic labeled benchmark from the SEI-USD data: estimate the MMHP-delta model (M=3, delta=1s) on the zero-price-return trade times of a training week; then simulate a control week from the fitted model using only states 1 and 2, and a treatment week in which known wash-trading bursts (e.g., 50-500 same-side trades executed against the same price within 1-2 seconds, at random times) are injected. Run the full pipeline, including Viterbi Algorithm 2, on both weeks, and compute precision, recall, and the false-positive rate of state-3 flags against the injected burst labels. If state-3 flags do not align with injected bursts, or if the control week yields many state-3 flags, then the paper's detection claim is unsupported; if the flags align with high precision and recall, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim has two parts. The first, that the MMHP-delta model can be estimated by the proposed EM algorithm, is well supported by the simulation studies of Section 3.5: the estimators converge with respect to EM steps, sample size, and the discretization step, and the goodness-of-fit residuals behave as expected as delta tends to zero. The second, load-bearing part, is the application claim in Section 4.2.2 that the model detects suspicious trading activity. This claim depends entirely on equating the Viterbi state with the largest excitation parameter alpha with 'suspicious/wash trading'. No labeled data, account identities, or independent confirmation is provided. The zero-price-return filter is motivated by wash trading, but it also selects exactly the trades that occur during normal price-stable periods, during iceberg or large-order execution, and during market-making activity. A three-state exponential Hawkes process will typically allocate one state to the upper tail of bursty activity, so the Viterbi algorithm will always find an extreme-burst state; that does not make the state manipulative. The paper's own conclusion concedes that the method is an indirect detection methodology and is 'prone to false positives', and that the bursts may not be due to manipulative behavior. Furthermore, thinning the trade process to zero-price-return events does not preserve the MMHP structure, so even an excellent out-of-sample fit (Figure 7) only establishes that the model describes the filtered event times, not that its hidden states correspond to actual market manipulation. The detection claim therefore lacks the validation needed to support a live-ready surveillance tool.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":27477,"tokens_out":3733,"duration_ms":43492,"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":[{"comment":"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.","section":"Section 4.2.2"},{"comment":"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.","section":"Section 4.2.2 and Section 2"},{"comment":"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.","section":"Section 4.2.2 and Table 1"}],"minor_comments":[{"comment":"Figure 13 caption contains \"Janurary\" and Figure 14 contains \"3th\"; these should be corrected to \"January\" and \"3rd\".","section":"Appendix B"},{"comment":"The heading \"Proof of Proportion 3.1\" should read \"Proof of Proposition 3.1\".","section":"Appendix A.5"},{"comment":"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.","section":"Section 2.2"},{"comment":"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.","section":"Section 3.1"},{"comment":"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.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The methodological core of this paper is solid and likely publishable in a statistics or quantitative-finance journal. The main obstacle is the mismatch between the empirical claim of detecting suspicious trading and the evidence provided. Given the authors' own caveats in the conclusion, I believe a major revision that reframes the application as anomaly or burst detection, or provides external validation, would bring the manuscript into acceptable scope. The title and abstract should then be adjusted accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The methodological core is the real contribution: an MMHP with a piecewise-constant kernel that decays between events, plus an EM algorithm with analytic E-step for general decreasing kernels. That is a genuine extension of Wang's stepwise-decay setup, and the simulation study is careful and convincing. The estimators converge with respect to EM steps, sample size, and discretization step; the residual diagnostics behave as they should as delta shrinks. I buy the method. The paper deserves credit for working through the matrix-exponential integrals cleanly and for spelling out the Viterbi and online variants.\n\nThe soft spot is exactly where the reader and stress-test put it: the empirical detection claim. Equating Viterbi state 3, defined as largest Hawkes excitation alpha, with 'suspicious trading' is not validated. The zero-price-return filter selects trades during price-stable periods, which also includes ordinary market-making and iceberg execution; a three-state Hawkes model will always find an extreme-burst state. The out-of-sample QQ plots only show that the filtered event stream is well described by the model, not that the hidden state is manipulation. The paper's own conclusion concedes this -- it calls the method indirect and 'prone to false positives'. So the body is more honest than the title and abstract. The imbalance and price-response analyses are suggestive and interesting, but they are descriptive, not confirmatory.\n\nI do not think the central methodological argument is flawed. The missing ground truth is an empirical validation gap, not a mathematical one. A referee should ask the authors to either temper the detection language throughout or add a synthetic labeled-burst experiment to show the Viterbi state actually tracks injected wash-like activity. The paper would also be stronger if it acknowledged that thinning trades to zero-price-return events does not preserve the MMHP structure, so the fitted model is a model of the filtered events, not of the original trade process.\n\nWho is this for? Anyone working on regime-switching point processes or on unsupervised burst detection in high-frequency data. It is a serious methodological paper with an overreaching application framing. I would send it to a statistical methodology journal and let the referee push for a more measured empirical section. Worth engaging with; the estimation machinery is likely to be reused.","headline":"Solid estimation paper with a credible unsupervised burst detector; the wash-trading label outruns the evidence, but the authors say so themselves.","tokens_in":28020,"tokens_out":1164,"would_cite":true,"duration_ms":15775,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G55","62M05","62P05","62F10"],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["Markov-modulated Hawkes process","expectation-maximization","regime switching","point process","wash trading","market manipulation","cryptocurrency","Viterbi algorithm"],"falsifier":"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.","tokens_in":26973,"feed_emoji":"🪙","tokens_out":8920,"duration_ms":89720,"temperature":0.7,"pith_summary":"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.","feed_headline":"Hidden-regime model flags crypto wash-trade bursts","feed_subtitle":"The model's rare extreme-burst state covers 0.14% of time yet about 20–25% of volume, and precedes price moves.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the stepwise-decay MMHP and the EM strategy that the paper generalizes to piecewise-constant kernels.","marker":"Wang (2010)"},{"why":"Formalizes the Markov-modulated Hawkes process with stepwise decay, the baseline model the paper extends.","marker":"Wang et al. (2012)"},{"why":"Provides the EM algorithm for Markov-modulated Poisson processes whose likelihood decomposition the paper adapts.","marker":"Rydén (1996)"},{"why":"Gives the filtered and smoothed probability recursions and the matrix-integral notation used in the E-step.","marker":"Roberts et al. (2006)"},{"why":"Provides the block-matrix-exponential method used to compute the expected sufficient statistics analytically.","marker":"Van Loan (1978)"},{"why":"Defines the price-response framework adapted to measure the impact of regime transitions.","marker":"Bouchaud et al. (2003)"},{"why":"Documents crypto wash trading and motivates the indirect detection problem the paper targets.","marker":"Cong et al. (2023)"},{"why":"Shows limitations of indirect volume-based detection, motivating a point-process alternative.","marker":"Falk et al. (2023)"}],"fun_headline_variants":["Rare burst state covers 0.14% of time, 20% of volume","Extreme-burst state flags 20% of volume, precedes price moves","Model finds rare state driving 20% of crypto volume","Hidden Markov-Hawkes model catches suspicious trade bursts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Rare burst state covers 0.14% of time, 20% of volume","Extreme-burst state flags 20% of volume, precedes price moves","Model finds rare state driving 20% of crypto volume","Hidden Markov-Hawkes model catches suspicious trade bursts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000893,"raw_usage":{"total_tokens":3811,"prompt_tokens":870,"completion_tokens":2941,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":2864}},"tokens_in":486,"tokens_out":2941,"duration_ms":18927,"temperature":1.0,"reasoning_tokens":2864,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T23:47:24.567823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the stepwise-decay MMHP and the EM strategy that the paper generalizes to piecewise-constant kernels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Formalizes the Markov-modulated Hawkes process with stepwise decay, the baseline model the paper extends."},{"cited_title":"J., Ephraim, Y., and Dieguez, E","cited_arxiv_id":null,"evidence_quote":"Gives the filtered and smoothed probability recursions and the matrix-integral notation used in the E-step."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the block-matrix-exponential method used to compute the expected sufficient statistics analytically."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the price-response framework adapted to measure the impact of regime transitions."},{"cited_title":"W., Li, X., Tang, K., and Yang, Y","cited_arxiv_id":null,"evidence_quote":"Documents crypto wash trading and motivates the indirect detection problem the paper targets."},{"cited_title":"Can AI Detect Wash Trading? Evidence from NFTs","cited_arxiv_id":"2311.18717","evidence_quote":"Shows limitations of indirect volume-based detection, motivating a point-process alternative."}],"review_version":1}