Forecasting time series with artificial neural networks


Having accurate time series forecasts helps to be prepared for upcoming events. As many real world time series have nonlinear and irregular behavior, traditional approaches may be lacking performance. A suitable alternative method is artificial neural network models, that can achieve high accuracy in various difficult tasks. The objective of given thesis is to give theoretical and practical guidelines for applying neural networks in time series forecasting with packages h2o and neuralnet for statistical programming language R, and library Keras for programming language Python. An empirical study was conducted on five different datasets to compare multilayer perceptron model performance with long short-term memory model, and iterative, direct and multi-neural network modeling strategies with each other. The performance of neural network models were compared with liner baseline models to expose whether the results have any practical gain. When comparing the network structures, the results indicate the superiority of long short-term memory models. Furthermore, long short-term memory models offered improvement over linear baseline model almost in case of all datasets. Based on these results, neural networks proved to have great performance for time series forecasting, and should be considered as an alternative to linear models.