Paper: arXiv 2510.22685

Authors: Ollie Olby, Rory Baggott, Namid Stillman

Abstract

The recent application of deep learning models to financial trading has heightened the need for high fidelity financial time series data. This synthetic data can be used to supplement historical data to train large trading models. The state-of-the-art models for the generative application often rely on huge amounts of historical data and large, complicated models. These models range from autoregressive and diffusion-based models through to architecturally simpler models such as the temporal-attention bilinear layer. Agent-based approaches to modelling limit order book dynamics can also recreate trading activity through mechanistic models of trader behaviours. In this work, we demonstrate how a popular agent-based framework for simulating intraday trading activity, the Chiarella model, can be combined with one of the most performant deep learning models for forecasting multi-variate time series, the TABL model. This forecasting model is coupled to a simulation of a matching engine with a novel method for simulating deleted order flow. Our simulator gives us the ability to test the generative abilities of the forecasting model using stylised facts. Our results show that this methodology generates realistic price dynamics however, when analysing deeper, parts of the markets microstructure are not accurately recreated, highlighting the necessity for including more sophisticated agent behaviors into the modeling framework to help account for tail events.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper combines advanced mathematical concepts from deep learning (TABL models) and agent-based modeling (Chiarella model) with stylized facts and simulation outputs to validate its synthetic LOB generation.

Research Flowchart

  flowchart TD
  A["Research Goal: Improve synthetic LOB generation using hybrid AI methods"] --> B{"Key Methodology: TABL-ABM Framework"}
  B --> C["Chiarella Model<br>(Agent-Based Simulation)"]
  B --> D["TABL Model<br>(Time Series Forecasting)"]
  C --> E["Matching Engine &<br>Deleted Order Flow"]
  D --> E
  E --> F{"Stylized Facts Analysis<br>vs. Real Market Data"}
  F --> G["Outcomes: Realistic Price Dynamics"]
  F --> H["Outcome: Inaccurate Microstructure & Tail Events"]

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 hybrid TABL-ABM model that combines the Chiarella agent‑based framework with a Temporal Attention Bilinear Layer forecasting model and integrates it with a simulated matching engine for synthetic LOB generation.

Source sentence
In this work, we demonstrate how a popular agent-based framework for simulating intraday trading activity, the Chiarella model, can be combined with one of the most performant deep learning models for forecasting multi-variate time series, the TABL model.

Assumes 6

  • The matching engine operates in event time rather than real clock time, so it does not model real‑world temporal intervals between events. model
    Source sentence
    First, it operates in event time rather than clock time, meaning it does not model real-world temporal intervals be-tween events.
  • The probability of an order being deleted given its depth is approximated by P(Depth = x) = Nx p / Nt, a simplifying assumption for relative activity at each depth. statistical
    Source sentence
    The assumption P(Depth = x) = Nx p Nt is made, as this serves as a measure of relative activity at each depth.
  • Tracking individual orders for precise deletion probabilities is computationally heavy, so the simplifying assumption above is used. model
    Source sentence
    This is very computationally heavy and involves in depth LOB data, hence we use this simplifying assumption.
  • Traders act according to the dynamics of the extended Chiarella model (fundamentalists, chartists, noise traders). market
    Source sentence
    By leveraging this structure, we use the Chiarella model to statistically infer the direction of the next order under the assumption that traders act according to these dynamics.
Show 2 more
  • The data used are level‑10 order‑book snapshots for Apple (AAPL) from the LOBSTER sample dataset. data
    Source sentence
    For this project we used the level 10 orderbook data from Apple (AAPL).
  • Order flow direction is inferred from the sign of demand, with positive indicating buy and negative indicating sell. model
    Source sentence
    In this framework, the sign of the demand reflects the likely direction of the next order: a positive demand suggests a buy order, while a negative demand implies a sell.

Data

  • LOBSTER sample limit order book data · LOBSTER order-book public
    5 symbols (MSFT, AAPL, GOOG, INTC, AMZN)
    Source sentence
    6.1 Data and Preprocessing In this work, we use the LOBSTER sample limit order book data [23].

Recognised sources: LOBSTER

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • The hybrid TABL‑ABM framework generates realistic price dynamics that capture volatility and mean‑reverting behaviour, as shown by mid‑price paths matching stylised facts. performance
    Source sentence
    These paths illustrate the model’s ability to capture key features of the market, such as the volatility and mean-reverting behaviour.
  • When analysing deeper microstructure, parts of the market are not accurately recreated, indicating the need for more sophisticated agent behaviours. negative
    Source sentence
    Our results show that this methodology generates realistic price dynamics however, when analysing deeper, parts of the markets microstructure are not accurately recreated, highlighting the necessity for including more sophisticated agent behaviors into the modeling framework to help account for tail events.
  • The simulated market impact experiment shows an initial sharp price movement followed by gradual reversion, reflecting market resilience. performance
    Source sentence
    This figure highlights how the simulated market initially responds with a sharp price movement. This is followed by a gradual reversion toward the original price trajectory, indicating a degree of market resilience or recovery once the imbalance subsides.
  • Bid/ask autocorrelations decrease as a power law and match historical levels, supporting alignment with the efficient market hypothesis. statistical
    Source sentence
    This is observed in the results from the model and the results are well matched to the historical levels.
  • The model exhibits short‑term volatility memory but lacks persistence at longer lags, showing weaker volatility clustering than historical data. statistical
    Source sentence
    This shows that our model does exhibit short term volatility memory but lacks persistence.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 22 related papers). Coverage is partial, so this is neither replication nor refutation.

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 (32)

  1. Financial Time Series Forecasting using CNN and Transformer — → 2304.04912
  2. DeepLOB: Deep Convolutional Neural Networks for Limit Order Books (2019) — link
  3. Modality-aware Transformer for Financial Time series Forecasting (2024) — link
  4. Prediction of Stock Price Time Series using Transformers (2023) — link
  5. Temporal Attention-Augmented Bilinear Network for Financial Time-Series Data Analysis (2019) — link
  6. Data Normalization for Bilinear Structures in High-Frequency Financial Time-series (2020) — link
  7. Multi-head Temporal Attention-Augmented Bilinear Network for Financial time series prediction (2022) — link
  8. Metadata Matters for Time Series: Informative Forecasting with Transformers (2024) — link
and 24 more
  1. mWDN-Transformer: Utilizing Cyclical Patterns for Short-Term Forecasting of Limit Order Book in China Markets (2024) — link
  2. Controllable Financial Market Generation with Diffusion Guided Meta Agent — → 2408.12991
  3. 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
  4. MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model — → 2409.07486
  5. Deep learning models for price forecasting of financial time series: A review of recent advancements: 2020–2022 — → 2305.04811
  6. Text2TimeSeries: Enhancing Financial Forecasting through Time Series Prediction Updates with Event-Driven Insights from Large Language Models (2024) — link
  7. Predicting stock price trends based on financial news articles and using a novel twin support vector machine with fuzzy hyperplane (2021) — link
  8. Priority Rules! (1998)
  9. Neural Network Time Series Forecasting of Financial Markets (1994)
  10. Long Short-Term Memory Neural Network for Financial Time Series (2022) — link
  11. Selected indicators of financial stability (2007)
  12. A dynamic limit order market with fast and slow traders (2014) — link
  13. A technical note on bilinear layers for interpretability (2023) — link
  14. LOB-Based Deep Learning Models for Stock Price Trend Prediction: A Benchmark Study — → 2308.01915
  15. Limit order books (2013) — link
  16. Decoupled Weight Decay Regularization (2019) — link
  17. Get Real: Realism Metrics for Robust Limit Order Book Market Simulations (2019) — link
  18. The impact of heterogeneous trading rules on the limit order book and order flows (2009) — link
  19. Understanding intra-day price formation process by agent-based financial market simulation: calibrating the extended chiarella model (2022) — link
  20. Forecasting High Frequency Order Flow Imbalance — → 2408.03594
  21. LOBSTER : Limit Order Book Reconstruction System, Technical Documentation (2011)
  22. The Price Impact of Order Book Events (2013) — link
  23. Co-existence of Trend and Value in Financial Markets: Estimating an Extended Chiarella Model (2018) — link
  24. Deeper Hedging: A New Agent-based Model for Effective Deep Hedging — → 2310.18755

8 of 32 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.