Non-performing loans (NPLs) hurt financial institutions by raising risks, decreasing cash flow, and lowering capital, making it critical for lenders to evaluate credit risk appropriately. With the increasing complexity of credit risk assessment, machine learning algorithms have become essential for the early detection and mitigation of non-performing loans (NPLs), allowing financial institutions to make better decisions to lower credit risk by properly forecasting NPLs. Compared with traditional statistical models, machine learning algorithms are better at predicting default probabilities and identifying patterns. In order to predict non-performing loans (NPLs), this study examines the efficacy of seven machine learning algorithms: Random Forest, Decision Tree, Lasso Regression, Support Vector Machine (SVM), Bidirectional Long Short-Term Memory (BiLSTM), Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting (XGBoost). The analysis is conducted using a dataset from the DSE-listed commercial banks of Bangladesh, covering the period from 2013 to 2023. Various performance matrices, such as the mean absolute error (MAE), mean square error (MSE), root mean squared error (RMSE), and mean absolute percentage error (MAPE), are used to train and assess the accuracy of the models. The empirical results show that while BiLSTM shows promise in capturing temporal relationships in loan performance, ensemble learning models—specifically, XGBoost and LightGBM—display stronger predictive competencies when compared to conventional tree-based classifiers. By providing insightful information for researchers, banking institutions, and legislators on how to improve risk assessment frameworks, this comparative analysis adds to the expanding reservoir of work on machine learning applications in financial risk management.