A k-Nearest Neighbours Approach for Interval-Valued Time Series Prediction
Interval-valued time series arise in applications where observations are affected by uncertainty or intrinsic variability. In such settings, classical forecasting methods are often applied to the interval bounds separately, which may ignore relevant structural information.
This work proposes kNN-ITSP, an adaptation of the k-nearest neighbours time series prediction framework designed to operate on interval-valued observations. The method extends the standard kNN-TSP scheme by defining similarity between interval-valued windows and by incorporating window-wise centering transformations to account for local level effects, while preserving interval coherence by construction.
An experimental study on long-term daily temperature data shows that the proposed approach achieves predictive accuracy comparable to kNN-TSP, suggesting that similarity-based forecasting can be naturally extended to interval-valued time series without collapsing interval information.
Palabras clave: Interval-valued time series k-nearest neighbours forecasting interval similarity