Intelligent Aggregation and Time Series Smoothing

Abstract

Predicting future values of a variable from past observations is fundamental task in many modern domains. This process, often referred to as times series smoothing, involves an aggregation of the past observations to predict the future values. Our objective here is to use recent advances in computational intelligence to suggest new and better approaches for performing the necessary aggregations. We first look at some special features associated with the types of aggregations needed in times series smoothing. We show how these requirements impact on our choice of weights in the aggregations. We then note the connection between the method of aggregation used in times series smoothing and that used in the intelligent type aggregation method known as the Ordered Weighted Averaging (OWA) operator. We then take advantage of this connection to allow us to simultaneously view the problem from a times series smoothing perspective and OWA aggregation operations perspective. Using this multiple view we draw upon the large body of work on families of OWA operators to suggest families for the aggregation of time series data. A particularly notable result of this linkage is the introduction of the use of linear decaying weights for time series data smoothing.