Paper: arXiv 2502.04027

Authors: Timothée Fabre, Ioane Muni Toke

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.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 7.5/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced mathematics, including a Markov-modulated Hawkes process, expectation-maximization algorithms, and ODE solutions, scoring high in math complexity. It also demonstrates empirical rigor by validating estimators on simulations and applying the model to real high-frequency cryptocurrency data for detection of suspicious trading activities.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>High-Frequency Market Manipulation Detection"] --> B["Methodology:<br>Markov-modulated Hawkes Process<br>Hawkes excitation kernels"]

  subgraph Inputs
      direction LR
      B1["Simulated Data<br>for Convergence Tests"] --> C
      B2["High-Frequency Crypto Data<br>Trade Events"] --> C
  end

  B --> C["Statistical Inference:<br>Expectation-Maximization Algorithm"]

  C --> D["Computational Process:<br>Estimation of Parameters<br>Intensities & Transition Probabilities"]

  D --> E{"Benchmarking<br>vs. Markov-modulated Poisson"}

  E --> F["Key Findings:<br>1. Convergence validated on simulation<br>2. Superior Goodness-of-Fit<br>3. Detection of anomalous trade bursts"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a novel self‑exciting point process with piecewise‑constant Hawkes intensity between events, estimated via an EM algorithm and a Viterbi procedure for state inference.

Source sentence
We introduce a novel self-exciting point process whose intensity is piecewise constant between event times, with a discretization step parameter δ.

Assumes 6

  • The underlying hidden Markov chain is M‑state, irreducible and aperiodic. model
    Source sentence
    parameters varying according to an unobservable M‑state irreducible and aperiodic continuous‑time Markov chain (St)t≥0.
  • The Hawkes kernel in each state is non‑negative and decreasing. model
    Source sentence
    Inspired by Wang (2010), we develop an estimation procedure for such processes with a general non‑negative, decreasing kernel function.
  • Intensity is piecewise constant between consecutive events (discretization step δ). model
    Source sentence
    We introduce a novel self‑exciting point process whose intensity is piecewise constant between event times, with a discretization step parameter δ.
  • Trade timestamps have microsecond precision and reflect the matching‑engine processing time. data
    Source sentence
    With every trade comes a timestamp with microsecond precision. This timestamp corresponds to the time at which the matching engine of the market processed the transaction.
Show 2 more
  • Seasonality of trade intensity is removed by normalising over 5‑minute bins and using average daily patterns. data
    Source sentence
    over 5‑minutes bins each day. Then, we compute the average of these intensities per bin over the time period, with a 95% confidence interval.
  • No trader identification numbers are available, so detection must rely on indirect methods. execution
    Source sentence
    traders’ IDs – even anonymous – are not provided by centralized venues, which makes wash trading detection extremely challenging.

Data

  • SEI‑USD trades · Coinbase crypto-onchain proprietary
    2023‑12‑01 to 2024‑03‑31 · SEI token
    Source sentence
    The dataset is composed of trades executed for SEI‑USD between December 1st, 2023 and March 31st, 2024 on Coinbase.

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • The MMHP‑δ model provides a superior fit to the data compared with the Markov‑modulated Poisson process, as measured by AIC. statistical
    Source sentence
    Firstly, the MMHP‑δ demonstrates a superior fit compared to the MMPP.
  • Reducing the discretization step δ improves the information‑criterion and goodness‑of‑fit, while larger δ worsens the QQ plots. statistical
    Source sentence
    Additionally, for each regime configuration, a reduction in the discretization step δ appears to improve the criterion.
  • Calibration of the MMHP‑δ on a single day takes about 15 minutes with δ = 100 s and around ten hours with δ = 1 s. other
    Source sentence
    As an example, for the estimation of the MMHP‑δ over a single day, the computation time is approximately 15 minutes with δ = 100s and around ten hours with δ = 1s.
  • The Viterbi‑derived state 3 trades are identified as natural candidates for suspicious wash‑trading activity. performance
    Source sentence
    In our setting, trades classified in state 3 are natural candidates for suspicious trading activity.
  • Out‑of‑sample QQ plots show that MMHP‑δ with small δ (1 s or 10 s) outperforms the MMPP in goodness‑of‑fit. statistical
    Source sentence
    We observe the MMHP‑δ with small δ (1 or 10) provides better results than the MMPP in terms of out‑of‑sample goodness‑of‑fit.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 47 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 2 papers in the archive (2 by independent authors)

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.