Time Series Prediction with Neural Networks: a Review

One dimensional time series prediction is a major problem nowadays. These series can describe physical phenomenon, traffic flow, economic transactions, etc. Anomaly detection, congestion control and bandwidth allocation require predictions with minimal error. In this paper a critical overview is provided of time series prediction techniques based on neural networks and their applications. The accent is on papers published last two years about discrete processes including both short term, long term and multi-step ahead forecasts. Many different approaches have been applied such as long-short term echo state network, deep hybrid neural network, intelligent models and so on. New methods for processing, analysis and segmentation of information are discussed. There are a few milestones like algorithms for training, prediction strategy and criteria for early stopping. The performance of different neural network architectures is reviewed also. Many real-world problems have been modelled with related time series and multiple seasonal patterns.