Co-trading networks for modeling dynamic interdependency structures and estimating high-dimensional covariances in US equity markets

Yutong Lu, Gesine Reinert, Mihai CucuringuarXiv 2023 · DOI · arXiv

Abstract

The time proximity of trades across stocks reveals interesting topological structures of the equity market in the United States. In this article, we investigate how such concurrent cross-stock trading behaviors, which we denote as co-trading, shape the market structures and affect stock price co-movements. By leveraging a co-trading-based pairwise similarity measure, we propose a novel method to construct dynamic networks of stocks. Our empirical studies employ high-frequency limit order book data from 2017-01-03 to 2019-12-09. By applying spectral clustering on co-trading networks, we uncover economically meaningful clusters of stocks. Beyond the static Global Industry Classification Standard (GICS) sectors, our data-driven clusters capture the time evolution of the dependency among stocks. Furthermore, we demonstrate statistically significant positive relations between low-latency co-trading and return covariance. With the aid of co-trading networks, we develop a robust estimator for high-dimensional covariance matrix, which yields superior economic value on portfolio allocation. The mean-variance portfolios based on our covariance estimates achieve both lower volatility and higher Sharpe ratios than standard benchmarks.

Introduction

In modern finance, understanding how stocks move together (covariance) is essential for risk management and portfolio optimization. However, traditional methods often rely on static industry classifications, like the Global Industry Classification Standard (GICS), which assume relationships between companies are fixed and based solely on their business sector.

Furthermore, estimating covariance for hundreds of stocks using limited time-series data leads to "noisy" or singular matrices that are mathematically difficult to work with. This research addresses two main challenges:

  1. Capturing Latent Structure: Static sectors fail to account for the high-frequency reality of the market, where algorithmic and high-frequency traders (HFTs) often trade "baskets" of stocks across different sectors simultaneously.
  2. Robust Estimation: Creating a high-dimensional covariance matrix that is both invertible and reflective of these dynamic, data-driven relationships.

Methodology

The authors propose a multi-step pipeline to bridge the gap between micro-level trading behavior and macro-level price movements.

Co-trading Score and Network Construction

The core of the approach is the "co-trading score," a similarity measure based on trade co-occurrence. If two stocks trade within a small time window (e.g., 500 microseconds), it is counted as a co-trading event.

Illustration of co-trading events within a time window $\delta$.

This score is normalized to create a weighted adjacency matrix, forming a co-trading network where nodes are stocks and edges represent the strength of their co-trading relationship. The authors use eigenvector centrality to identify influential stocks and Maximum Spanning Trees (MST) to visualize the market's backbone.

The co-trading network visualized as a Maximum Spanning Tree (MST), with nodes colored by GICS sectors.

Spectral Clustering and Covariance Estimation

To find communities of stocks that behave similarly, spectral clustering is applied to the co-trading matrix. Unlike GICS, these clusters are updated daily, capturing shifting market regimes. This information is then used to improve standard factor models (like PCA) by imposing a "block-diagonal" structure on the idiosyncratic (stock-specific) part of the covariance matrix. This allows for correlations between stocks within the same data-driven cluster while zeroing out noise between unrelated clusters.

Data

The study utilizes high-precision data from the US equity markets:

  • LOBSTER Database: High-frequency limit order book data for 457 NASDAQ-traded stocks that were constituents of the S&P 500, spanning from January 3, 2017, to December 9, 2019.
  • CRSP & Compustat: Daily stock prices and official GICS sector labels used as a baseline.

Results and Insights

Network Dynamics and Regime Shifts

Eigenvector centrality reveals that the market structure is heavily influenced by a core group of tech giants (e.g., Microsoft, Apple) and financial institutions (e.g., JPMorgan, Citigroup), acting as hubs in the co-trading network. Analysis of cluster similarity over time (using the Adjusted Rand Index) identified three distinct temporal regimes, showing that cross-sector co-trading increased from 2017 to 2019. This suggests the market is becoming more interconnected and less reliant on traditional industry boundaries.

Statistical and Economic Performance

The co-trading networks showed a statistically significant positive relationship with realized covariance on 98.51% of the days studied. In backtests, the Global Minimum Variance (GMV) portfolio built using co-trading clusters achieved an annualized Sharpe ratio of 1.40, significantly outperforming the GICS-based baseline (0.97) and the broader market (0.83). The proposed estimator consistently produced portfolios with lower annualized volatility (around 8%) compared to standard benchmarks.

Conclusion

This paper demonstrates that trade arrivals measured in microseconds have significant explanatory power over price correlations measured in minutes or days. While the current model assumes symmetric co-trading, future work could explore asymmetric lead-lag relationships and incorporate trade direction (buys vs. sells) to further refine covariance estimates. The use of network time series also presents potential for forecasting future market volatility and structural shifts using deep learning.

Summary updated

Abstract

The time proximity of trades across stocks reveals interesting topological structures of the equity market in the United States. In this article, we investigate how such concurrent cross-stock trading behaviors, which we denote as co-trading, shape the market structures and affect stock price co-movements. By leveraging a co-trading-based pairwise similarity measure, we propose a novel method to construct dynamic networks of stocks. Our empirical studies employ high-frequency limit order book data from 2017-01-03 to 2019-12-09. By applying spectral clustering on co-trading networks, we uncover economically meaningful clusters of stocks. Beyond the static Global Industry Classification Standard (GICS) sectors, our data-driven clusters capture the time evolution of the dependency among stocks. Furthermore, we demonstrate statistically significant positive relations between low-latency co-trading and return covariance. With the aid of co-trading networks, we develop a robust estimator for high-dimensional covariance matrix, which yields superior economic value on portfolio allocation. The mean-variance portfolios based on our covariance estimates achieve both lower volatility and higher Sharpe ratios than standard benchmarks.
BibTeX
@misc{lu2023co,
  author = {Yutong Lu and Gesine Reinert and Mihai Cucuringu},
  title = {{Co-trading networks for modeling dynamic interdependency structures and estimating high-dimensional covariances in US equity markets}},
  year = {2023},
  eprint = {2302.09382},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2302.09382},
}
Figure