<p>The Channel Quality Indicator (CQI) is an essential metric in 5G networks, supporting adaptive infrastructure optimization to ensure a high level of Quality of Service (QoS). Recent studies have explored enhancing CQI estimation using machine learning techniques. A key factor in training an accurate model is selecting an appropriate loss function. Two widely used loss functions are Mean Squared Error (MSE) and Mean Absolute Error (MAE). Generally, MSE emphasizes outliers more heavily, whereas MAE prioritizes the majority of the data points. In this work, we highlight the advantages of the Huber loss function for CQI prediction, as it effectively integrates the strengths of both MSE and MAE. The Huber loss achieves this by smoothly transitioning between the two, governed by a user-defined hyperparameter called delta. However, manually selecting the optimal delta to balance the sensitivity to minor errors (as in MAE) and robustness to outliers (as in MSE) is a challenging task. To overcome this limitation, we introduce a novel loss function called Residual-based Adaptive Huber Loss (RAHL). RAHL incorporates a learnable residual into the delta parameter, allowing the model to adapt dynamically to the error distribution in the data. This approach enhances robustness against outliers while preserving precision for inlier data. Additionally, we leverage various features to improve CQI prediction accuracy. To evaluate RAHL’s effectiveness, we evaluate its performance across multiple architectures, including Long Short-Term Memory (LSTM), parallel 1D Convolutional Neural Networks (CNN1D), and the Informer model. Experimental results confirm that RAHL consistently enhances performance across these models, demonstrating its adaptability and reliability. These findings establish RAHL as a promising solution for improving CQI prediction in 5G networks.</p>