Causal interventions in bond multi-dealer-to-client platforms.

Area of Science:

• Quantitative Finance
• Computational Economics
• Machine Learning in Finance

Background:

• Financial markets are increasingly digital, shifting trading from voice to electronic channels.
• Multi-Dealer-to-Client (MD2C) platforms facilitate simultaneous quote requests (RfQs) for financial instruments.
• The opacity of competitor pricing on MD2C platforms necessitates robust analysis for dealer profitability.

Purpose of the Study:

• To introduce a novel framework for analyzing the RfQ process on MD2C platforms.
• To explore inferential questions crucial for dealer profitability, including optimal pricing and revenue estimation.
• To compare generative and discriminative modeling approaches for RfQ analysis.

Main Methods:

• Utilized probabilistic graphical models and causal inference for a general analytical framework.
• Developed a generative model based on Fermanian, Guéant, & Pu (2017).
• Employed machine learning techniques for discriminative models, including LightGBM.

Main Results:

• Generative models achieved predictive accuracy comparable to leading discriminative algorithms (ROC-AUC: 0.742 vs. 0.743).
• The proposed framework successfully incorporates critical business constraints, such as spread monotonicity.
• Demonstrated the utility of the framework for pricing, revenue estimation, and client identification.

Conclusions:

• A novel framework using probabilistic graphical models and causal inference effectively analyzes MD2C RfQ processes.
• Generative models provide a viable alternative to discriminative machine learning, balancing predictive power with business rule adherence.
• The findings offer valuable insights for dealers navigating competitive electronic trading environments.