In this thesis, a new approach using pattern recognition techniques is suggested for time series modeling (TSM) where a time series is classified into one of autoregreeive moving- average (ARMA) models. The underlying concept of the approach is pattern matching which can be described in such a manner that unknown pattern may be appropriately assigned to a class which best matches the pattern. In this study, a pattern is derived from a time series under consideration by using extended sample autocorrelation function and the time series is classified into one of ARMA models whose prototypical pattern best matches the time series pattern. For TSM, two approaches- inductive learning approach and decision tree classifier approach- are investigated and integrated into a unified framework. For inductive learning approach, a learning procedure is used to recognize various types of patterns by using linear discriminant whose goal is to discriminate one pattern from the others. The recognition results are stored in pattern base. When a time series pattern is to be identified, the pattern is compared with the stored patterns in pattern base. If the degree of similarity is proved to be highest for a stored pattern, then the time series pattern is assigned to a model relating to the stored pattern. For decision tree classifier approach, a decision tree classifier is developed to divide decision procedures involved in TSM into a set of simple and local decisions at each node of the tree. Fuzzy set theory is used to deal with the imprecision inherent in TSM. A tree search algorithm suitable for the characteristics of TSM is suggested based on fuzzy decision values. Knowledge-based approach is applied to provide intelligence for the algorithm. These two approaches are integrated into a unified TSM-solving framework. A prototype system is designed and implemented. Experimental results with several examples show that a pattern recognition-based approach can yield a promising sol...