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.
Introduction
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 (PCC) to see if two stocks move in tandem. However, traditional approaches suffer from three major flaws: dependency on prior human-defined patterns (like linear correlation), "stop word" bias where market noise overwhelms signals, and a lack of adaptability to evolving market patterns.
The goal of this research is to infer "co-investment activities"—when the same group of investors buys or sells a pair of stocks—which is usually unobservable data. To test this, the study utilizes the S&P 500 dataset covering 470 companies over six years (2010–2016), analyzing over 110,000 possible pairwise combinations.
Methods: The DeepCNL Framework
The authors propose DeepCNL, a framework that treats the market index as a form of "distant supervision" to back-infer relationships between individual stocks.
The architecture consists of several key components:
- Convolutional Layer: Instead of using a fixed formula, a CNN learns "co-investment patterns" directly from pairwise stock data (price and volume), producing "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: Once the model is trained to predict the market index, the authors extract the inner weights of the LSTM. These weights represent the "intensity" of the co-investment relationship between specific stock pairs.
- Graph Generation: By applying "Occam’s Razor," weak connections are filtered out to create a sparse, meaningful network of the most influential stock pairings.
Key Results and Financial Validity
DeepCNL demonstrates superior ability in identifying market structures that align with real-world financial principles. A key finding is that stocks with higher market capitalization had a higher "degree" (more connections) in the network, which is consistent with the S&P 500 being a capitalization-weighted index. Specifically, the top 10 "hub" stocks identified by DeepCNL had an average market cap of $223.1 Billion, significantly higher than the $67.0 Billion found by standard Pearson correlation.
Furthermore, the network structure found by DeepCNL covered approximately 60% of the top-performing stocks mentioned in annual financial news reports, outperforming traditional baselines. While traditional correlation networks appeared noisy and erratic year-over-year, DeepCNL identified stable, evolving clusters (e.g., strong links between Apple and Bank of America) that match known mutual fund holdings.
Conclusion
DeepCNL offers a novel way to mine the "black box" of neural networks by treating internal parameters as primary outputs. While computationally expensive due to the $O(N^2)$ nature of pairwise enumeration, it provides a more accurate and evolving map of market structure. Future work may explore using continuous return values instead of binary rise/fall targets and further interpreting the specific patterns learned by the CNN kernels.
Summary updated