Accurate time series forecasting is vital in areas such as finance, weather prediction, and energy management. Traditional forecasting methods often struggle to effectively model the intricate patterns and nonlinearities present in real-world data. This study proposes the feedback-matching neural network (FMNN), a deep learning model that evolves from the feedback-matching algorithm (FMA). By embedding the core concepts of FMA into a neural network structure, the FMNN can recognize and match historical patterns in time series data, leading to more accurate predictions. Extensive experiments reveal that the FMNN outperforms several conventional statistical models and modern neural networks in terms of forecasting accuracy, as evaluated by the weighted absolute percentage error (WAPE). The FMNN enhances prediction accuracy by offering a sophisticated method for identifying and leveraging repeating trends within the data.