Recursive Multi-step Time-Series Forecasting for Residual-Feedback Artificial Neural Networks: A Survey

Residual-feedback artificial neural networks are a type of artificial neural network (ANNs) that have shown better forecasting performance on some time series. One of the challenges of residual-feedback ANNs is by utilizing the previous time step’s observed value, they are only capable of predicting one step ahead in advance. Therefore, it would not be possible to apply them directly in a recursive multi-step forecast strategy. To shed light on this challenge, a systematic literature review was conducted in this paper to find answers to the following three research questions: What are the main motivations behind introducing residual feedback to ANNs? How good are the existing residual-feedback ANNs compared to other forecasting methods in terms of forecasting performance? And what are the existing solutions for recursive multi-step time series forecasting using residual-feedback ANNs? An analysis of 19 studies was conducted to answer these questions. Furthermore, several potential solutions that can be further practically explored are suggested in an attempt to overcome this challenge.