Paper: arXiv 2505.21662

Authors: Mateusz Wilinski, Anubha Goel, Alexandros Iosifidis, Juho Kanniainen

Abstract

The rapid development of sophisticated machine learning methods, together with the increased availability of financial data, has the potential to transform financial research, but also poses a challenge in terms of validation and interpretation. A good case study is the task of classifying financial investors based on their behavioral patterns. Not only do we have access to both classification and clustering tools for high-dimensional data, but also data identifying individual investors is finally available. The problem, however, is that we do not have access to ground truth when working with real-world data. This, together with often limited interpretability of modern machine learning methods, makes it difficult to fully utilize the available research potential. In order to deal with this challenge we propose to use a realistic agent-based model as a way to generate synthetic data. This way one has access to ground truth, large replicable data, and limitless research scenarios. Using this approach we show how, even when classifying trading agents in a supervised manner is relatively easy, a more realistic task of unsupervised clustering may give incorrect or even misleading results. We complete the results with investigating the details of how supervised techniques were able to successfully distinguish between different trading behaviors.

Complexity vs Empirical Score

  • Math Complexity: 4.0/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper uses advanced machine learning methods but focuses heavily on empirical validation via a synthetic agent-based limit order book environment, featuring reproducible code and detailed implementation of trading simulations.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Classify & Cluster<br>Trading Agents"] --> B["Methodology<br>Agent-Based Model<br>for Synthetic Data"]
  B --> C["Data Generation<br>Ground Truth Labels<br>with Behavioral Patterns"]
  C --> D["Computational Process<br>Supervised Classification<br>e.g., Random Forest, SVM"]
  C --> E["Computational Process<br>Unsupervised Clustering<br>e.g., K-Means, GMM"]
  D --> F["Key Findings<br>Supervised: Highly Accurate"]
  E --> G["Key Findings<br>Unsupervised: Error/Misleading"]

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 using a realistic agent‑based model to generate synthetic data with known ground truth for classifying and clustering trading agents.

Source sentence
In order to deal with this challenge we propose to use a realistic agent‑based model as a way to generate synthetic data.

Assumes 5

  • Market makers assume that best‑ask and best‑bid orders are simply the best available offers (At = at and Bt = bt). model
    Source sentence
    In our simulations market makers assume that best‑ask and best‑bid orders are simply the best available offers, which means that At = at and Bt = bt.
  • Market takers split a large order into smaller Gaussian‑distributed orders and send them at normally distributed time intervals. model
    Source sentence
    In our simulation they simply cut the large order into smaller ones and then execute them in consecutive time intervals. The size of the large order is fixed for each agent, but the smaller orders have a Gaussian distribution with fixed mean and standard deviation. The time intervals between sending the smaller orders
  • The simulation is calibrated to match stylized facts of price return distributions and autocorrelation structures observed in real markets. statistical
    Source sentence
    While the goal of this article is not to create the most realistic model, we wanted to test at least the trademarks of true financial data. As a result, we focused on the stylized facts related to the distribution of price returns and the autocorrelation structure of price returns.
  • Features are computed per agent over all of their orders, trades, or cancellations. data
    Source sentence
    We compute these features for each agent separately over all of their orders, trades, or cancellations.
Show 1 more
  • Hierarchical clustering hyper‑parameters are optimized on the noise‑free dataset with all 18 features. model
    Source sentence
    The clustering hyper‑parameters, including the linkage method and the number of clusters, are optimized using the noise‑free dataset with all 18 features.

Data

  • Synthetic data generated by the realistic agent‑based model synthetic unknown
    Source sentence
    we propose to use a realistic agent‑based model as a way to generate synthetic data.

Costs

Not applicable The paper reports no trading or execution result.

Finds 4

  • Supervised classification of trading agents achieves very high accuracy (up to 99%) even with substantial label noise. predictive-accuracy
    Source sentence
    (a) Scenario with 18 features and no noise. Overall accuracy 0.99.
  • Market makers are classified correctly regardless of noise level or feature set. performance
    Source sentence
    The first observation is that the market makers, in the current form, are easily classified regardless of the size of the noise and the features used.
  • Unsupervised clustering can produce misleading results, often merging fundamentalists and chartists. negative
    Source sentence
    While working on the problem, we noticed that not only is clustering unable to fully reproduce the dependencies we find using supervised approach, it is also much more limited in terms of the
  • Standard features struggle to distinguish investors with sophisticated strategies; trend‑based features improve discrimination. statistical
    Source sentence
    In summary, we have shown that standard features may have difficulties in distinguishing investors with more sophisticated strategies. Additionally, we suggested targeted features to deal with that problem, i.e. trend‑based features allowed to distinguish chartists from other agents.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 25 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 1 paper in the archive

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.

Cited References (61)

  1. An agent-based model of intra-day financial markets dynamics (2021)
  2. Statistical theory of the continuous double auction (2003)
  3. Market making and mean reversion (2011)
  4. Welfare effects of market making in continuous double auctions (2017)
  5. Learning to classify and imitate trading agents in continuous double auction markets (2021)
  6. Statistical properties of stock order books: empirical results and models (2002)
  7. Econophysics review: I. Empirical facts (2011)
  8. More statistical properties of order books and price impact (2003)
and 53 more
  1. The predictive power of zero intelligence in financial markets (2005)
  2. The information content of an open limit-order book (2009)
  3. An empirical analysis of the limit order book and the order flow in the Paris Bourse (1995)
  4. Get real: Realism metrics for robust limit order book market simulations (2020)
  5. Analysis and modeling of client order flow in limit order markets (2023)
  6. The flash crash: High-frequency trading in an electronic market (2017)
  7. The Flash Crash: A cautionary tale about highly fragmented markets (2019)
  8. Forecasting jump arrivals in stock prices: new attention-based network architecture using limit order book data (2019)
  9. Empirical properties of asset returns: stylized facts and statistical issues (2001)
  10. An introduction to variable and feature selection (2003)
  11. The challenges of data quality and data quality assessment in the big data era (2015)
  12. The variation of certain speculative prices (1963)
  13. Benchmark dataset for mid-price forecasting of limit order book data with machine learning methods (2018)
  14. Ward’s hierarchical agglomerative clustering method: which algorithms implement Ward’s criterion? (2014)
  15. The economy needs agent-based modelling (2009)
  16. Machine learning in finance (2020)
  17. Machine learning in finance: the case of deep learning for option pricing (2017)
  18. Credit risk assessment model for Jordanian commercial banks: Neural scoring approach (2014)
  19. Predicting the direction of stock market prices using tree-based classifiers (2019)
  20. A reinforcement learning extension to the Almgren-Chriss framework for optimal trade execution (2014)
  21. Generative ai for end-to-end limit order book modelling: A token-level autoregressive generative model of message flow using a deep state space network — → 2309.00638
  22. Temporal Data Meets LLM–Explainable Financial Time Series Forecasting — → 2306.11025
  23. Explainable Artificial Intelligence (XAI) 2.0: A manifesto of open challenges and interdisciplinary research directions (2024)
  24. Synergistic integration between machine learning and agent-based modeling: A multidisciplinary review (2021)
  25. A simple learning agent interacting with an agent-based market model (2024)
  26. Towards multi-agent reinforcement learning-driven over-the-counter market simulations (2024)
  27. Calibrating agent-based models to microdata with graph neural networks (2022) — link
  28. Price-time priority and pro rata matching in an order book model of financial markets (2011)
  29. ABIDES: Towards high-fidelity multi-agent market simulation (2020)
  30. Fast Agent-Based Simulation Framework with Applications to Reinforcement Learning and the Study of Trading Latency Effects (2021)
  31. Limit order books (2013)
  32. The survival of noise traders in financial markets (1988)
  33. Bid-ask spreads, trading volume and volatility: Intra-day evidence from the London Stock Exchange (1997)
  34. Layer-wise relevance propagation: an overview (2019)
  35. A survey of deep neural network architectures and their applications (2017)
  36. Selecting the number of clusters k with a stability trade-off: an internal validation criterion (2023)
  37. Tune: A research platform for distributed model selection and training (2018) — link
  38. Evaluation: From Precision, Recall and F-Measure to ROC, Informedness, Markedness & Correlation (2011)
  39. Liquidity and Market Structure (1988)
  40. High-Frequency Trading and Price Discovery (2014)
  41. High frequency trading and the new market makers (2013)
  42. Optimal liquidity trading (2005)
  43. Detecting liquidity traders (2009)
  44. Foundations of Technical Analysis (2000)
  45. Deep Learning for Price Prediction in Financial Markets (2018)
  46. The noise trader approach to finance (1990)
  47. Noise Trader Risk in Financial Markets (1990)
  48. Identification of information networks in stock markets (2021)
  49. Statistically validated lead-lag networks and inventory prediction in the foreign exchange market (2018)
  50. The diversity of high-frequency traders (2013)
  51. High-frequency trading around large institutional orders (2019)
  52. https://github.com/mateuszwilinski/cluster-investors/ (2025) — link
  53. https://github.com/mateuszwilinski/lob-abm-simulator/ (2025) — link

2 of 61 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.