Paper: arXiv 2409.05192

Authors: Tejas Ramdas, Martin T. Wells

Abstract

In this study, we leverage powerful non-linear machine learning methods to identify the characteristics of trades that contain valuable information. First, we demonstrate the effectiveness of our optimized neural network predictor in accurately predicting future market movements. Then, we utilize the information from this successful neural network predictor to pinpoint the individual trades within each data point (trading window) that had the most impact on the optimized neural network’s prediction of future price movements. This approach helps us uncover important insights about the heterogeneity in information content provided by trades of different sizes, venues, trading contexts, and over time.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced non-linear machine learning (neural networks, AutoML) and interpretable methods, requiring sophisticated mathematical modeling. It demonstrates strong empirical rigor through the use of large-scale, high-frequency trade data, a multi-stage prediction and optimization pipeline, and quantifiable results on real market data.

Research Flowchart

  flowchart TD
  A["Research Goal: Identify trades most influential in predicting future price movements"] --> B["Data Inputs: Historical Market Data<br>Individual Trades, Context, Venue"]
  B --> C["Step 1: Build Predictor<br>Train Neural Network to predict future price movements"]
  C --> D["Step 2: Isolate Impact<br>Use Network to identify specific trades with highest impact on prediction"]
  D --> E["Step 3: Analyze Characteristics<br>Examine size, venue, and context of influential trades"]
  E --> F["Key Findings: Heterogeneous Information Content<br>Trade specifics (size/venue) significantly impact price prediction"]

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 an interpretable machine‑learning framework that identifies trade characteristics most influencing an optimized neural‑network predictor of future price movements.

Source sentence
In this study, we present a framework using interpretable methods in machine learning to identify characteristics of trades that most influence our optimized neural network’s prediction of future price movements.

Assumes 2

  • The approach does not require or assume a model of the trading process. model
    Source sentence
    Our approach to identifying trades that most influence an optimized predictor’s prediction of future market movements does not
  • The choice of a lower window‑size threshold (289 trades) may retain a degree of arbitrariness. data
    Source sentence
    The selection of a lower threshold may still retain a degree of arbitrariness, and exploring different dimensions of window sizes may yield insightful results.

Data

  • NYSE Daily Trades and Quotes database · NYSE order-book commercial
    Source sentence
    we extract trades and quotes information from the NYSE Daily Trades and Quotes database.

Recognised sources: Cboe

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • Lower window sizes provide superior predictive power, which plateaus as window size increases. predictive-accuracy
    Source sentence
    We observe that lower window sizes provide superior predictive power, which plateaus with increasing window size.
  • Trades of size 101‑1000 shares have average influence‑score around 46.1, while larger trades have lower scores. statistical
    Source sentence
    Trades between sizes 101-1000 show similar influence-score averages of 46.12 and 46.13 respectively.
  • A longer time since the last trade is associated with higher influence scores for predicting positive market movements (coefficient 0.10, p < 0.001). statistical
    Source sentence
    In the model for predicted positive movements, the coefficient is 0.10 (p < 0.001), indicating that a longer time since the last trade may be associated with a higher influence score.

Checked by others

Contradicted by

  • Predictive AI with External Knowledge Infusion for Stocks (2025)
    On the question: How does the size of the sliding window affect the predictive performance of trade‑based signals?
    Compare the findings

    This paper: Lower window sizes provide superior predictive power, which plateaus as window size increases.

    That paper: Moderate sliding‑window sizes generally yield stronger forecasting performance, as shown by the effect of window size W on Top‑5 return ratio across different holding periods.

    Finding A reports lower window sizes give superior predictive power that plateaus as size grows, while Finding B reports moderate window sizes generally yield stronger forecasting performance.

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 (59)

  1. Continuous Auctions and Insider Trading (1985)
  2. Bid, ask and transaction prices in a specialist market with heterogeneously informed traders (1985) — link
  3. Stealth trading and volatility: Which trades move prices? (1993) — link
  4. Stealth-trading: Which traders’ trades move stock prices? (2001) — link
  5. Which Trades Move Stock Prices in the Internet Age? (2005) — link
  6. Measuring the Information Content of Stock Trades (1991) — link
  7. One Day in the Life of a Very Common Stock (1997)
  8. The market impact of a limit order (2012) — link
and 51 more
  1. Permanent trading impacts and bond yields (2012) — link
  2. Impacts of trades in an error-correction model of quote prices (2004) — link
  3. Measuring price impact and information content of trades in a time-varying setting — → 2212.12687
  4. Microstructure in the Machine Age [The risk of machine learning] (2021) — link
  5. Sparse Signals in the Cross-Section of Returns (2019) — link
  6. Predicting Stock Market Returns with Machine Learning (2018) — link
  7. Deep neural networks, gradient-boosted trees, random forests: Statistical arbitrage on the S&P 500 (2017) — link
  8. Empirical Asset Pricing via Machine Learning (2020) — link
  9. Advances in Financial Machine Learning (2018)
  10. Financial Machine Learning (2023) — link
  11. One Security, Many Markets: Determining the Contributions to Price Discovery (1995) — link
  12. What’s Not There: Odd Lots and Market Data (2014) — link
  13. An Introduction to Convolutional Neural Networks (2015) — link
  14. Deep Learning (2016)
  15. Tree-Structured Parzen Estimator: Understanding Its Algorithm Components and Their Roles for Better Empirical Performance (2023) — link
  16. Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification (2015)
  17. Why gradient clipping accelerates training: A theoretical justification for adaptivity (2019) — link
  18. Backpropagation Applied to Handwritten Zip Code Recognition (1989) — link
  19. Optuna: A Next-generation Hyperparameter Optimization Framework (2019)
  20. Deep inside convolutional networks: Visualising image classification models and saliency maps (2014)
  21. Dropout: A Simple Way to Prevent Neural Networks from Overfitting (2014) — link
  22. Price Impact Asymmetry of Block Trades: An Institutional Trading Explanation (2001) — link
  23. Visualizing and Understanding Convolutional Networks (2014)
  24. Efficient Processing of Deep Neural Networks: A Tutorial and Survey (2017)
  25. A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects (2020) — link
  26. Algorithms for Hyper-Parameter Optimization (2011) — link
  27. Random Forests (2001) — link
  28. A Unified Approach to Interpreting Model Predictions (2017) — link
  29. “Why should i trust you?” Explaining the predictions of any classifier (2016)
  30. Learning deep features for discriminative localization (2016)
  31. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization (2019) — link
  32. Axiomatic attribution for deep networks (2017)
  33. SmoothGrad: removing noise by adding noise; 2017. 10.48550 (2017)
  34. Full-gradient representation for neural network visualization (2019)
  35. Machine Learning Methods Economists Should Know About (2019)
  36. Big Data: New Tricks for Econometrics (2014)
  37. The Elements of Statistical Learning: Data Mining, Inference, and Prediction (2009)
  38. Deep Learning (2016)
  39. Pattern Recognition and Machine Learning (2006)
  40. Microstructure in the Machine Learning Age (2019)
  41. Sparse Signals in the Cross-Section of Returns (2019)
  42. Predicting Stock Market Returns (2018)
  43. Deep neural networks, gradient-boosted trees, random forests: Statistical arbitrage on the S&P 500 (2017)
  44. Empirical Asset Pricing via Machine Learning (2020)
  45. Advances in Financial Machine Learning (2018)
  46. Deep Learning for Finance: Deep Portfolios (2017)
  47. Universal Features of Price Formation in Financial Markets: Perspectives from Deep Learning (2019)
  48. Deep Learning in Portfolio Allocation (2020)
  49. Deep Learning for Asset Pricing (2019)
  50. Deep Learning in Asset Pricing (2020)
  51. Financial Networks and Asset Returns (2023) — link

1 of 59 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.