[Paper Review] Stock markets reconstruction via entropy maximization driven by fitness and density

Tiziano Squartini, Guido Caldarelli+1 more|arXiv (Cornell University)|Jun 1, 2016

Complex Systems and Time Series Analysis10 references3 citations

TL;DR

This paper proposes a novel entropy maximization method to reconstruct stock market networks from aggregate data, using firm capitalizations and link density to infer connection probabilities and position weights via a density-corrected gravity model. The approach produces more accurate, less dense networks that better predict systemic risk than standard methods.

ABSTRACT

The spreading of financial distress in capital markets and the resulting systemic risk strongly depend on the detailed structure of financial interconnections. Yet, while financial institutions have to disclose their aggregated balance sheet data, the information on single positions is often unavailable due to privacy issues. The resulting challenge is that of using the aggregate information to statistically reconstruct financial networks and correctly predict their higher-order properties. However, standard approaches generate unrealistically dense networks, which severely underestimate systemic risk. Moreover, reconstruction techniques are generally cast for networks of bilateral exposures between financial institutions (such as the interbank market), whereas, the network of their investment portfolios (i.e., the stock market) has received much less attention. Here we develop an improved reconstruction method, based on statistical mechanics concepts and tailored for bipartite market networks. Technically, our approach consists in the preliminary estimation of connection probabilities by maximum-entropy inference driven by entities capitalizations and link density, followed by a density-corrected gravity model to assign position weights. Our method is successfully tested on NASDAQ, NYSE and AMEX filing data, by correctly reproducing the network topology and providing reliable estimates of systemic risk over the market.

Motivation & Objective

• To address the challenge of reconstructing financial networks when individual position data is unavailable due to privacy constraints.
• To overcome the limitations of standard reconstruction methods that produce unrealistically dense networks, underestimating systemic risk.
• To develop a method specifically tailored for bipartite market networks of investment portfolios, not just interbank exposures.
• To improve the accuracy of systemic risk estimation by preserving higher-order network properties from aggregate balance sheet data.
• To validate the method on real stock exchange data (NASDAQ, NYSE, AMEX) for reliable network topology and risk prediction.

Proposed method

• The method begins with maximum-entropy inference to estimate connection probabilities based on firm capitalizations and overall link density.
• It incorporates fitness (capitalization) and density constraints to ensure realistic network structure under given aggregate information.
• A density-corrected gravity model is then applied to assign weights to inferred positions, improving realism of link weights.
• The approach is specifically adapted for bipartite market networks, where institutions hold equity positions in other firms.
• The reconstruction process is validated by comparing topological properties and systemic risk measures against real market data.
• Statistical mechanics principles are used to ensure the most unbiased network reconstruction consistent with known aggregate constraints.

Experimental results

Research questions

• RQ1 Can we reconstruct accurate stock market networks from only aggregate balance sheet data, without access to individual position details?
• RQ2 How can we improve network reconstruction to avoid unrealistically dense networks that underestimate systemic risk?
• RQ3 What role do firm capitalization and overall link density play in generating realistic network topologies?
• RQ4 Can a gravity model corrected for density enhance the accuracy of position weight estimation in reconstructed networks?
• RQ5 Does the proposed method preserve higher-order network properties and yield reliable systemic risk estimates?

Key findings

• The proposed method successfully reproduces the topological structure of real stock markets such as NASDAQ, NYSE, and AMEX using only aggregate data.
• The reconstructed networks are significantly less dense than those produced by standard methods, leading to more accurate systemic risk assessments.
• Incorporating fitness (capitalization) and density constraints via maximum-entropy inference improves the realism of connection probabilities.
• The density-corrected gravity model enhances the accuracy of position weight assignment, contributing to better network fidelity.
• Systemic risk estimates derived from the reconstructed networks closely match those from real market data, validating the method's predictive power.
• The method outperforms standard reconstruction techniques by preserving higher-order network properties critical for risk analysis.