Deep Co-Investment Network Learning for Financial Assets

Yue Wang, Chenwei Zhang, Shen Wang, Philip S. Yu, Lu Bai, Lixin Cui

TL;DR

DeepCNL (Deep Co-investment Network Learning) is a novel framework that automatically discovers hidden "co-investment" relationships between stocks by analyzing how their joint movements influence the overall market index. By using deep learning to move beyond simple linear correlations, it creates a more accurate and evolving map of market structure that aligns closely with real-world financial principles and news reports.

What Problem Does This Paper Address?

Understanding the "market structure"—how different stocks relate to one another—is vital for economists and investors. Traditionally, this is done using statistical methods like Pearson correlation to see if two stocks move in tandem. However, the authors identify three major flaws in these traditional approaches:

  1. Prior Pattern Dependency: They rely on human-defined patterns (like linear correlation), potentially missing complex, non-linear relationships.
  2. Stop Word Bias: Frequent but uninformative market noise can overwhelm truly interesting signals, much like "stop words" (the, a, is) in text analysis.
  3. Evolving Patterns: Markets change over time, but fixed statistical formulas do not adapt to new dynamics.

The goal is to infer "co-investment activities"—when the same group of investors buys or sells a pair of stocks—which is usually unobservable data.

Methods

The authors propose DeepCNL, a framework that treats the market index as a form of "distant supervision" to back-infer relationships between individual stocks.

  • Convolutional Layer: Instead of using a fixed formula, a CNN learns "co-investment patterns" (kernels) directly from pairwise stock data (price and volume). This produces "co-investment evidence."
  • Recurrent Neural Network (LSTM) Layer: This layer takes the evidence from all possible stock pairs and attempts to predict the rise or fall of the market index.
  • Weight Extraction: This is the "aha" moment of the paper. Once the model is trained to predict the market index, the authors look at the inner weights of the LSTM. They argue that the weights assigned to specific stock pairs represent the "intensity" of their co-investment relationship.
  • Graph Generation: By applying "Occam’s Razor," they filter out weak connections to create a sparse, meaningful network of the most influential stock pairings.

Datasets

The study uses the S&P 500 dataset (sourced from Kaggle/NYSE):

  • Scope: 470 companies over 6 years (2010–2016).
  • Features: Five daily features per stock (Open, High, Low, Close, and Volume).
  • Target: The "SPY" ticker (an ETF tracking the S&P 500) serves as the ground truth for the market index's rise/fall trends.
  • Scale: With 470 stocks, there are over 110,000 possible dyadic (pairwise) combinations to analyze.

Key Insights and Ideas

  • Distant Supervision via the Index: You don't need private transaction records to find co-investors. If the joint movement of Stock A and Stock B consistently predicts the movement of the entire market index, there is a high probability they are being traded by the same large institutional players.
  • Mining the "Black Box": Rather than just using the neural network for prediction, the researchers treat the trained model's internal parameters as the primary output. The weights are the discovery.
  • Automatic Feature Engineering: By using CNN kernels, the model discovers its own "indicators" for what constitutes a meaningful co-movement, allowing it to adapt as market regimes change.

Key Results

  • Financial Validity: The model discovered that stocks with higher market capitalization had a higher "degree" (more connections) in the network. This aligns with the S&P 500 being a capitalization-weighted index.
  • Superior Hub Identification: The top 10 "hub" stocks identified by DeepCNL had an average market cap of $223.1 Billion**, compared to only **$67.0 Billion for those identified by standard Pearson correlation.
  • Real-World Correlation: The network structure found by DeepCNL covered 60% of the top-performing stocks mentioned in annual financial news reports, significantly outperforming traditional baselines.
  • Stable Evolution: While traditional correlation networks appeared noisy and erratic year-over-year, DeepCNL identified stable, evolving clusters (e.g., strong, persistent links between Apple and Bank of America) that match known mutual fund holdings.

Limitations and Future Work

  • Computational Cost: The method requires enumerating all pairwise combinations ($|V|^2$). For a market with thousands of stocks, this becomes an $O(N^2)$ problem that is computationally expensive, even with GPU acceleration.
  • Binary Supervision: The model currently uses a binary (rise/fall) target for the market index. The authors suggest that using continuous return values might provide more granular insights.
  • Interpretability of Kernels: While the weights tell us which stocks are related, the specific "patterns" learned by the CNN kernels remain somewhat of a black box and could benefit from further financial interpretation.

Summary updated

Abstract

Most recent works model the market structure of the stock market as a correlation network of the stocks. They apply pre-defined patterns to extract correlation information from the time series of stocks. Without considering the influences of the evolving market structure to the market index trends, these methods hardly obtain the market structure models which are compatible with the market principles. Advancements in deep learning have shown their incredible modeling capacity on various finance-related tasks. However, the learned inner parameters, which capture the essence of the finance time series, are not further exploited about their representation in the financial fields. In this work, we model the financial market structure as a deep co-investment network and propose a Deep Co-investment Network Learning (DeepCNL) method. DeepCNL automatically learns deep co-investment patterns between any pairwise stocks, where the rise-fall trends of the market index are used for distance supervision. The learned inner parameters of the trained DeepCNL, which encodes the temporal dynamics of deep co-investment patterns, are used to build the co-investment network between the stocks as the investment structure of the corresponding market. We verify the effectiveness of DeepCNL on the real-world stock data and compare it with the existing methods on several financial tasks. The experimental results show that DeepCNL not only has the ability to better reflect the stock market structure that is consistent with widely-acknowledged financial principles but also is more capable to approximate the investment activities which lead to the stock performance reported in the real news or research reports than other alternatives.
BibTeX
@inproceedings{wang2018deep,
  author = {Yue Wang and Chenwei Zhang and Shen Wang and Philip S. Yu and Lu Bai and Lixin Cui},
  title = {{Deep Co-Investment Network Learning for Financial Assets}},
  booktitle = {ICBK},
  year = {2018},
  doi = {10.1109/icbk.2018.00014},
}

Loading related content...

Figure