Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bellwether Trades: Characteristics of Trades influential in Predicting Future Price Movements in Markets

Tejas Ramdas, Martin T. Wells

TL;DR

This paper addresses which trades carry information that predicts future price movements in financial markets. It develops an AutoML–optimized CNN predictor that operates on grid-structured trading windows and then derives per-trade influence scores via a first-order Taylor expansion of the predictor, $f(\cdot;\theta)$. A regression model then relates these influence scores to trade attributes (size, venue, time, year) and their interactions, revealing a long-tail distribution of predictive power and systematic differences across assets and venues. The framework is adaptable to varying window sizes and asset classes and provides a robust template for downstream inference in finance without relying on explicit microstructure models.

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.

Bellwether Trades: Characteristics of Trades influential in Predicting Future Price Movements in Markets

TL;DR

This paper addresses which trades carry information that predicts future price movements in financial markets. It develops an AutoML–optimized CNN predictor that operates on grid-structured trading windows and then derives per-trade influence scores via a first-order Taylor expansion of the predictor, . A regression model then relates these influence scores to trade attributes (size, venue, time, year) and their interactions, revealing a long-tail distribution of predictive power and systematic differences across assets and venues. The framework is adaptable to varying window sizes and asset classes and provides a robust template for downstream inference in finance without relying on explicit microstructure models.

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.
Paper Structure (12 sections, 14 equations, 11 figures, 14 tables)

This paper contains 12 sections, 14 equations, 11 figures, 14 tables.

Table of Contents

  1. Introduction
  2. Data
  3. Initial Overview
  4. Raw data and Inputs for Prediction
  5. Input Variables (Prediction)
  6. Outcome Variable (Prediction)
  7. The Neural Network Architecture
  8. Hyper-parameter Optimization
  9. Inference Problem: Identifying Influential Trades
  10. Extracting Attributes of Influential Trades
  11. Results and Discussion
  12. Discussion and Future Directions

Figures (11)

  • Figure 1: Each window has $n^2$ trades, arranged in a grid. Each of the 35 inputs of each trade are then $n \times n$ matrices arranged in a vector of dimensions $n \times n \times 35$
  • Figure 2: Lower window sizes offer the best predictive power, and this predictability plateaus with increasing window size.
  • Figure 3: An illustration of the convolutional, pooling and dense layers of our optimal neural network architecture.
  • Figure 4: Prediction accuracies generated from applying optimized predictor to two separate valiation sets, by year
  • Figure 5: Visualizing conditional distributions of influence scores by trade attribute category
  • ...and 6 more figures