Estimating Market Liquidity from Daily Data : Marrying Microstructure Models and Machine Learning

Estimating market liquidity based on low-frequency (daily) data is important for both empirical research and investment practice. We apply machine learning to estimate market liquidity by combining human-engineered liquidity proxies based on microstructure models and widely available low-frequency data. Nonlinear machine learning models can effectively learn a good estimate of the high-frequency spread from the raw low-frequency data alone, achieving performances far better than classical human-engineered liquidity proxies. In addition, combining human-engineered proxies and the raw low-frequency data further improves performances, with the best model achieving cross-sectional correlations of over 0.95 and time-series correlations of over 0.70 between model estimates and the ground-truth average spread. We show that the machine learning models can easily be interpreted using feature importance scores and partial dependence plots, with the latter showing nonlinear relationships that are otherwise not captured by prior approaches. Our results suggest that significant improvements in liquidity estimation can be achieved by combining human knowledge and machine