Forecasting Seasonal Time Series with Neural Networks: A Sensitivity Analysis of Architecture Parameters

Neural networks are widely applied in time series forecasting. However, no consensus exists on their capability of forecasting seasonal time series. As seasonal patterns frequently occur in empirical time series, it is imperative to establish their efficacy in forecasting seasonality. This paper seeks to evaluate the usefulness of multilayer perceptrons in forecasting time series with different forms of seasonal and trend components. Using eight synthetic time series, we systematically evaluate the impact of different combinations of hidden nodes, input nodes and activation functions on the distribution of the forecasting errors. We aim to a) establish the sensitivity of different architectural choices for neural networks in forecasting and b) analyze the relative accuracy of one or multiple neural network architectures as forecasting methods for seasonal time series. Results are presented in order to guide future selection of network parameters. We find that neural networks show sensitivity to selected architecture decisions but generally provide a robust and competitive forecasting performance on seasonal data.