The Effect of Liquidity on Deep Learning Errors

Abstract

This study investigates the impact of market liquidity and macroeconomic variables on the forecasting performance of deep learning models in financial markets. The primary objective is to forecast price movements for ten stocks listed on the BIST 30 index using a single-layer Long Short-Term Memory (LSTM) model and identify market conditions associated with prediction failures. A two-stage empirical framework is employed using daily data covering 2006–2025. First, LSTM hyperparameters are optimized through Bayesian optimization and walk-forward validation. Second, daily Absolute Percentage Error (APE) is modeled using OLS-HAC and EGARCH-X regressions to examine the effects of stock-level liquidity measures and global macroeconomic variables. Results show that the LSTM model generates well-calibrated forecasts across all ten equities, with MAPE values ranging from 3.29% to 8.40%. Trading volume and illiquidity, measured by the Amihud ratio, are positively associated with prediction errors for most stocks. At least one liquidity proxy is statistically significant for nine of the ten equities, with consistent coefficient signs where significance is observed. In contrast, macroeconomic and global indicators exhibit limited and stock-specific effects. These findings highlight the importance of incorporating liquidity conditions into deep learning-based financial forecasting models to improve forecast reliability.