Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Trading Graph Neural Network

Xian Wu

TL;DR

Trading Graph Neural Network (TGNN) addresses price formation in trading networks by jointly modeling asset features, dealer features, and relationship features within a structurally grounded graph-based estimator, yielding a unique equilibrium $\boldsymbol{v}^*$ found via a contraction mapping $T$. It blends the Simulated Method of Moments philosophy with neural graph learning to directly minimize prediction error and produce interpretable parameter estimates with bootstrap confidence intervals. Across dense, sparse, and core-periphery networks, TGNN outperforms traditional OLS with centrality controls and accurately recovers latent primitives such as bargaining power and holding costs, enabling counterfactual analyses in OTC and decentralized markets. By micro-founding the message-passing mechanism with economic structure, TGNN provides economists with a tool for price formation analysis, regulatory insight, and systemic-risk assessment in heterogeneous trading networks.

Abstract

This paper proposes a new algorithm -- Trading Graph Neural Network (TGNN) that can structurally estimate the impact of asset features, dealer features and relationship features on asset prices in trading networks. It combines the strength of the traditional simulated method of moments (SMM) and recent machine learning techniques -- Graph Neural Network (GNN). It outperforms existing reduced-form methods with network centrality measures in prediction accuracy. The method can be used on networks with any structure, allowing for heterogeneity among both traders and assets.

Trading Graph Neural Network

TL;DR

Trading Graph Neural Network (TGNN) addresses price formation in trading networks by jointly modeling asset features, dealer features, and relationship features within a structurally grounded graph-based estimator, yielding a unique equilibrium found via a contraction mapping . It blends the Simulated Method of Moments philosophy with neural graph learning to directly minimize prediction error and produce interpretable parameter estimates with bootstrap confidence intervals. Across dense, sparse, and core-periphery networks, TGNN outperforms traditional OLS with centrality controls and accurately recovers latent primitives such as bargaining power and holding costs, enabling counterfactual analyses in OTC and decentralized markets. By micro-founding the message-passing mechanism with economic structure, TGNN provides economists with a tool for price formation analysis, regulatory insight, and systemic-risk assessment in heterogeneous trading networks.

Abstract

This paper proposes a new algorithm -- Trading Graph Neural Network (TGNN) that can structurally estimate the impact of asset features, dealer features and relationship features on asset prices in trading networks. It combines the strength of the traditional simulated method of moments (SMM) and recent machine learning techniques -- Graph Neural Network (GNN). It outperforms existing reduced-form methods with network centrality measures in prediction accuracy. The method can be used on networks with any structure, allowing for heterogeneity among both traders and assets.

Paper Structure

This paper contains 22 sections, 1 theorem, 16 equations, 12 figures, 8 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Literature
  3. Model
  4. Remark
  5. Parameterization
  6. Estimation
  7. Estimate Customer Values
  8. Estimation with the Trading Network
  9. Bootstrap Confidence Intervals
  10. Test Cases
  11. Network Structure and Synthetic Data Generation
  12. Network Dimensions:
  13. Node Generation:
  14. Edge Structure:
  15. Latent Variables:
  16. ...and 7 more sections

Key Result

Theorem 1

There exists unique fixed point of equilibrium values $\{v_{ikt}^*\}_{i,k,t}$ given $\{c_{ikt},u_{ikt}\}_{i,k,t}$ and $\{\pi_{ijkt}\}_{i,j,k,t}$. And for each asset $k$ and date $t$, we can find the equilibrium value $\boldsymbol{v}_{kt}^*=(v_{1kt}^*, \dots, v_{nkt}^*) \in \mathbb{R}^n$ with the fol

Figures (12)

  • Figure 1: The Structure of Dense Random Networks
  • Figure 2: Estimated Parameters in Dense Random Networks
  • Figure 3: Comparison of Predicted vs. Actual Latent Variables in Dense Random Networks
  • Figure 4: Prediction Comparison: OLS with Centrality Interactions vs. TGNN in Dense Random Networks
  • Figure 5: The Structure of Sparse Random Networks
  • ...and 7 more figures

Theorems & Definitions (2)

  • Theorem 1: Equilibrium Existence and Uniqueness
  • proof : Proof of Theorem \ref{['thm:eqlm']}