Synthesis of Statistical Knowledge from Time-Dependent Data

This paper presents a general approach to analyzing multivariate time-dependent system processes with discrete-valued (both nominal and ordinal) and/or continuous-valued outcomes. The approach is based on an event-covering approach which selects (or covers) a subspace from the outcome space of an n-tuple of variables for estimation purposes. From the covered subspace, statistically interdependent events are selected as statistical knowledge for forecasting unknown events. The event-covering method presented here is based on the use of restricted variables with only a subset of the outcomes considered. Extension to event-covering method based on the selection of joint outcomes is also discussed. We have tested this method using climatic data and simulated data which model situations in real life. The experiments successfully demonstrate both its flexibility and efficacy.