Extract Transform and Load (ETL) processes are vital for accurate banking liquidity reporting, ensuring data is extracted, transformed, and loaded from various sources into centralized systems. These processes support regulatory compliance by enabling timely and consistent liquidity risk assessments. However, ETL faces challenges such as handling complex financial data, maintaining data quality, and adapting to evolving regulatory requirements. In this manuscript, a machine learning-driven optimization framework for enhancing ETL processes in banking liquidity reporting (MLDO-ETL-BLR-CKAN) is proposed. Firstly, input data is gathered from cash liquidity forecasting dataset. Then the input data is pre-processed utilizing Robust Adaptive Error State Kalman Filter (RAESKF) which is used to cleaning the data. Then the pre-processed data are provided to Red-billed blue magpie optimizer (RBMO) for feature selection. RBMO is employed to select the relevant features. Then the selected features are given to Convolutional Kolmogorov-Arnold Network (CKAN) method is used to predict future liquidity positions, such as cash reserves, shortfalls, and funding. The proposed method implemented in python, demonstrates substantial improvements in Root Mean Squared Error (RMSE), Mean Squared Error (MAE), R-squared. The proposed MLDO-ETL-BLR-CKAN model achieves the best results with the lowest MSE of 0.015, highest R-squared of 0.78, and lowest RMSE of 0.022, outperforming IPAJ-ANN, LR-SR-CNN, and OCBM-BC-DNN in predictive accuracy and consistency, compare with existing methods such as From liquidity risk to systemic risk: A use of knowledge graph (LR-SR-CNN), Neural network-based liquidity risk prediction in Indian private banks (NNLR-ANN), and Operationalization of the construct “Business model of a Bank”: clustering analyses with deep neural networks (OCBM-BC-DNN).