Liquidity’s Many Faces, Deep Learning’s Many Minds: Hybrid and Baseline Deep Networks for Multidimensional Liquidity Forecasting in the Moroccan Stock Market

This paper investigates multidimensional liquidity forecasting in the Moroccan Stock Market—an order-driven, less dynamic frontier venue—using deep-learning architectures (LSTM, CNN, LSTM–CNN, CNN–LSTM, and an attention-augmented hybrid). With fixed look-back windows (5 and 21 days) and a consistent out-of-sample protocol (scalers fit on Train; one-step forecasts), we evaluate MAE, MSE, RMSE, MASE, and R², and compare models via Diebold–Mariano tests based on MSE and QLIKE. The attention hybrid model delivers the best overall accuracy, and the 21-day window provides the most robust gains. These results support better trading decisions, earlier liquidity-risk aletrs, and supervisory calibration in pre-emerging markets.