ABSTRACT
A multiple set of discrete time series can be visualized as a set of tracks running along together. If the data points are connected by a wire diagram then we have a surface. This surface is fuzzy (influenced by error) and dynamic, in that interrelationships (transient and permanent) are pro pogated through time. Our focus is on procedures which can help detect fundmental patterns and time-varying relationships on such a surface. This paper proposes a new approach called FEVA, feature vector analysis, where the key idea is to replace each scalar data point with a vector of information which represents what the point Sees” around itself. The main processing unit consists of a clustering module to detect patterns in the enhanced (feature vector) data set, and a neural net simulator to estimate flexible functional relationships. The cluster analysis results are presented graphically using pattern–grids and the neural net simulator builds a decision surface. These two visual objects can be used to perform both forecasting and “what-if” analyses. A small but real illustrative example involving stock market information is used to demonstrate the FEVA approach.
