SHARE: A Distributed Learning Framework for Multivariate Time-Series Forecasting: A Distributed Learning Framework for Multivariate Time-Series Forecasting

In this work, we propose a distributed time series forecasting framework for scenarios where multiple computing nodes collaboratively make predictions using only locally available data streams. Our proposed framework utilizes a hierarchical architecture, consisting of multiple local models and a global model, and provides an efficient training algorithm to achieve high prediction accuracy. Key features of the framework are: (i) Simple collaboration, where a carefully designed global model enables the system to leverage the correlations among different local streams; (ii) Flexibility, where each node can choose its local model from a wide variety of prediction models (e.g., RNNs or Transformers) suitable to its compute resources; (iii) Privacy and communication efficiency, where messages communicated between the nodes are low-dimensional embeddings. The proposed approach's effectiveness is demonstrated theoretically and numerically using a number of time series forecasting tasks, with the (surprising) observation that the proposed distributed models can match or even outperform centralized ones.