Abstract
An interval time series (ITS) assigns to each period an interval covering the values taken by the variable. Each interval has four characteristic attributes, since it can be defined in terms of lower and upper boundaries, center and radius. The analysis and forecasting of ITS is a very young research area, dating back less than 15 years, and still presents a wide array of open issues. One main issue with time series in a big data context consists of deciding if to handle it as classic time series (CTS) or to proceed with some kind of aggregation in order to get a time series of symbolic data like ITS. Using the k-Nearest Neighbours (kNN) method, in this paper both approaches are applied to forecast exchange rates. Based on usual distances for interval-valued data such as Haussdorff, Ichino-Yaguchi and so on; the reduction in mean distance error using ITS instead of CTS suggests that the ITS approach could be a better way to forecast exchange rates using large data or data streaming. Some interesting conclusions about monthly and daily aggregation horizons are obtained and further research issues are proposed.
