Paper: arXiv 2504.07923

Authors: Xian Wu

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.

Complexity vs Empirical Score

  • Math Complexity: 8.0/10
  • Empirical Rigor: 5.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper combines advanced mathematical concepts like fixed point theory, contraction mappings, and structural estimation (high math complexity), while providing simulation-based validation and discussing real-world applications (moderate empirical rigor).

Research Flowchart

  flowchart TD
  A["Research Goal"] --> B["Data Sources"]
  B --> C["Model Architecture"]
  C --> D["SMM Optimization"]
  D --> E["Prediction & Validation"]
  E --> F["Key Findings"]

  subgraph A ["Research Goal"]
      A1["Estimate impact of asset,<br>dealer, & relationship features<br>on asset prices in trading networks"]
  end

  subgraph B ["Data/Inputs"]
      B1["Trading Network Data<br>(Assets, Dealers, Relationships)"]
  end

  subgraph C ["Methodology: TGNN"]
      C1["Combine Simulated Method of Moments<br>(SMM) + Graph Neural Network (GNN)"]
  end

  subgraph D ["Computational Process"]
      D1["Structural Estimation<br>via SMM Optimization"]
  end

  subgraph E ["Validation"]
      E1["Prediction Accuracy Test<br>vs. Network Centrality Methods"]
  end

  subgraph F ["Outcomes"]
      F1["✓ Superior prediction accuracy<br>✓ Works on any network structure<br>✓ Handles trader/asset heterogeneity"]
  end

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a new algorithm, Trading Graph Neural Network (TGNN), that structurally estimates the impact of asset, dealer, and relationship features on asset prices in trading networks.

Source sentence
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.

Assumes 6

  • Asset features X, dealer features Y, and relationship features E are assumed to be 1‑dimensional for simplicity. model
    Source sentence
    For simplicity, we assume that the asset features X, dealer features Y and trading relationship E are 1-dimensional.
  • Asset features Xkt are drawn independently from a standard normal distribution across assets and time. statistical
    Source sentence
    Each asset k at time t has feature Xkt ∼N(0, 1) are drawn from standard normal distributions.
  • Dealer features Yit are drawn independently from a standard normal distribution across dealers and time. statistical
    Source sentence
    Each dealer i at time t has feature Yit ∼N(0, 1) are drawn from a standard 10 normal distribution.
  • Relationship features Eij are independent standard normal draws. statistical
    Source sentence
    Each edge (i, j) is assigned relationship features drawn from independent standard normal distribution Eij ∼N(0, 1).
Show 2 more
  • Edges are generated as Erdos–Rényi random graphs with each potential edge included with probability 70%, creating an undirected sparse network. model
    Source sentence
    We generate Erdos–R´enyi (ER) random graphs where each potential edge is included with probability 70%, resulting in an undirected sparse network where each dealer can potentially trade with multiple other dealers.
  • The network structure may change over time, but this is not required for applications. other
    Source sentence
    We allow the network to change over time to highlight the flexibility of the algorithm, but this step is not necessary in applications.

Data

  • Synthetic trading network data synthetic unknown
    10 dealers, 2 assets, 5 days
    Source sentence
    4.1 Network Structure and Synthetic Data Generation Our trading network model represents an over-the-counter market with dealers trading assets over multiple days. In this test case, the network is constructed as follows: Network Dimensions: We create a synthetic trading environment with 10 dealers (Ni = 10) and 2 different assets (Nk = 2) observed over 6 trading days (Nt = 5).

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • TGNN outperforms existing reduced‑form methods with network centrality measures in prediction accuracy. predictive-accuracy
    Source sentence
    It outperforms existing reduced-form methods with network centrality measures in prediction accuracy.
  • The estimates of asset, dealer, and relationship parameters obtained with TGNN are close to the true parameters in test cases. statistical
    Source sentence
    The estimates of the parameters on the asset features, dealer features and relationship features with TGNN are close to the true parameters.
  • TGNN can recover bargaining power, dealer holding cost, dealer values, and potential transaction prices with high accuracy. statistical
    Source sentence
    It can recover the bargaining power, dealer holding cost, dealer values and potential transaction prices with high accuracy.
  • TGNN outperforms the traditional reduced‑form approach with centrality measures in explaining observed prices across dense, sparse random, and core‑periphery networks. performance
    Source sentence
    It outperforms the traditional reduced-form approach with centrality measures in explaining the observed prices in dense, sparse random networks and core-periphery networks.
  • TGNN achieves the highest R² and the lowest MAE and MSE compared to OLS regression with centrality measures and interaction terms. performance
    Source sentence
    We can see that TGNN still has the highest R2, lowest MAE and MSE.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 21 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (19)

  1. Simulation and the asymptotics of optimization estimators (1989)
  2. A method of simulated moments for estimation of discrete response models without numerical integration (1989)
  3. Core-periphery trading networks (2016)
  4. Dealer networks (2019)
  5. The value of trading relations in turbulent times (2017)
  6. Bid-ask spreads, trading networks, and the pricing of securitizations (2017)
  7. Graph Neural Networks: Theory for Estimation with Application on Network Heterogeneity (2024) — link
  8. Attention-based Graph Neural Networks in Firm CDS Prediction (2024)
and 11 more
  1. Graph Neural Networks for Causal Inference Under Network Confounding (2022) — link
  2. Decentralized exchange (2017)
  3. Trading and information diffusion in over-the-counter markets (2018)
  4. Public Information in Over-the-Counter Markets (2025)
  5. Comparing search and intermediation frictions across markets (2022)
  6. Inventory, market making, and liquidity in OTC markets (2024)
  7. Sequential search for corporate bonds (2023)
  8. OTC intermediaries (2023)
  9. Should Corporate Bond Trading Be Centralized? (2017)
  10. Linear regression with centrality measures (2022) — link
  11. Bootstrap methods: another look at the jackknife (1992)

0 of 19 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.