The authors present a neural-network-based approach to time series modeling (TSM) in which a time series is classified into one of the autoregressive moving-average (ARMA) models. The main feature of this approach lies in extraction of regularities from the extended sample autocorrelation function (ESACF) which is derived from a given time series being considered. The role of the neural network is to recognize the ESACF patterns whose interpretation is essential for a successful TSM. The backpropagation learning algorithm is used to learn the ESACF patterns within the framework of a multilayered neural network. Through extensive computer experiments with real time series, the neural-network-based TSM proved promising due to its robust pattern-recognition ability in two aspects: it not only avoids statistical difficulties, but also provides more user-friendly decision-making aids for forecasting purposes.