ML-Powered Corporate Bond Liquidity Scoring: A Multi-Measure XGBoost Framework on FINRA TRACE Data with SHAP Interpretability

Introduction: Corporate bond illiquidity remains a persistent challenge in fixed income markets, with significant implications for pricing accuracy, transaction cost analysis, and portfolio risk management. Traditional single-measure proxies — such as the Amihud illiquidity ratio or Roll spread — capture only partial dimensions of bond market liquidity, leaving systematic mispricing risk unaddressed. This paper introduces a machine learning framework that integrates multiple classical and market-microstructure liquidity measures into a unified, interpretable scoring system for US corporate bonds.

Methods: Using FINRA TRACE transaction data, we engineer 40+ features spanning four liquidity dimensions: price impact (Amihud illiquidity ratio, Kyle's lambda), trading costs (Roll spread, effective bid-ask spread), market depth (inter-trade duration, trade size distribution), and flow toxicity (volume-weighted price impact). An XGBoost classifier is trained on investment-grade and high-yield bond universes, with Optuna-driven hyperparameter optimisation and time-series-aware cross-validation to prevent look-ahead bias. Model interpretability is addressed through SHAP (SHapley Additive exPlanations) analysis, identifying the marginal contribution of each feature to individual liquidity scores.

Results: The multi-measure XGBoost model achieves superior classification accuracy compared to single-measure baselines (Amihud, Roll), with SHAP analysis revealing that inter-trade duration and Kyle's lambda are the dominant predictors of bond illiquidity episodes. The model identifies liquidity deterioration signals 1–3 days ahead of spread widening in high-yield segments.

Conclusions: This framework provides a scalable, ML-powered approach to corporate bond liquidity scoring with direct applications in RFQ pricing desks, portfolio construction, and regulatory liquidity stress testing. Code and data pipeline are available on GitHub.