Forecasting stock market liquidity with machine learning : an empirical evaluation in the German market

The study benchmarks four machine-learning algorithms- Random Forest, XGBoost, CatBoost and Long Short-Term Memory (LSTM) networks-for forecasting stock market liquidity in Germany's DAX equity market. Using data from January 2006 to May 2025, a Liquidity Score is constructed as a turnover-to-volatility ratio, designed to penalize wide intraday price swings while rewarding active trading behavior. This metric captures key microstructural aspects of liquidity and serves as the dependent variable throughout the analysis. It is paired with 41 independent variables that capture volatility, price ranges, return dynamics, technical indicators and cross-asset linkages. Empirical testing shows that the two gradient-boosting ensembles consistently outperform both Random Forest and the LSTM model, tracking sudden liquidity swings more accurately and delivering the tightest forecast errors. The evidence highlights (i) the practical superiority of tree-based boosting for high-frequency liquidity prediction, (ii) the value of rich, carefully engineered feature sets in modelling non-linear market micro-structure effects and (iii) the limitations of standard LSTM architectures when financial sequences are short and noisy. The findings offer actionable insights for traders, treasurers and regulators seeking real-time early-warning indicators of liquidity stress in European blue-chip equities.