Empirical Evaluation of Various Deep Neural Network Architectures for Time Series Forecasting

Time series forecasting is regarded amongst the top 10 challenges in data mining. Lately, deep learning based models have garnered a lot of attention from researchers in time series forecasting. However, which deep neural network architecture is most appropriate in time series forecasting domain has not been researched extensively.In this research performance of 4 deep neural network architectures MLP (Multilayer Perceptron), Traditional RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Units) were evaluated on two synthetic and two real-world time series exhibiting strong chaos, trend, and seasonality. Mackey Glass and Lorenz chaotic time series were simulated in this study to test our DNN models against chaos, while Apple Stock and Melbourne Minimum temperature were two real-world datasets showing increasing trend and seasonality. Experiments demonstrate that GRU based deep learning models outperform all other DNN models in forecasting both real world and synthetic time serie