Abstract: This paper considers a hybrid fuzzy neural networks (HFNN) for time-series prediction based on error distribution analysis. Firstly, a new hybrid FNN (HFNN) structure is established, where the last two layers are replaced by a combination of a full connection layer and nonlinear activation function. Thus, more parameters can be updated in training process to guarantee the prediction accuracy. Secondly, a novel attention loss function is proposed to make a sample with a certain error distribution, leading to gains in training process. Rule analysis with probability density function, indicates that the proposed method can provide a more uniform and stable prediction output. The prediction errors of HFNN converge to a compact set. Finally, two benchmark problems are applied to demonstrate the hybrid model performance on a time series prediction. The comparisons with other prediction models verify the efficiency and accuracy of the proposed HFNN model.