Pattern Matching for Multivariate Time Series Forecasting

ABSTRACT This article presents a new approach to multivariate time series forecasting. While most existing techniques in the literature focus on forecasting a single time series, forecasting multiple time series is a common goal in many applications. To deal with this, we introduce a new method, Weighted Nearest Neighbors for multivariate time series (WNN multi ). This method forecasts the future of a given series by identifying similar patterns not only in its own past but also in the past of related time series. Once the k nearest neighbors are identified, forecasts are made by averaging their future values. We evaluate the proposed approach on several real‐world datasets and compare its performance against state‐of‐the‐art forecasting techniques. The results demonstrate that our method achieves comparable or significantly improved performance, showcasing its effectiveness.