UniVQ: Vector quantization is effective for unified time series forecasting.

Abstract

Time series forecasting is crucial for applications such as climate, energy, and transportation. Existing models, whether task-specific or pretrained, face limitations due to constrained sample size of datasets and generalization. Models trained on single datasets struggle with real-world scenarios where training data represents only a subset of actual distributions. We introduce UniVQ, a framework using unified vector quantization for latent space forecasting. UniVQ employs a uniform discrete vector database covering intrinsic representations across multiple domains. An Encoder-Decoder architecture converts time series into discrete latent space, where a forecaster predicts future discrete representations. We use Finite State Quantization to unify distributions and capture temporal variations, mapping time series into uniform discrete space while preserving patterns. UniVQ utilizes a unified prediction head for one-stage multi-horizon forecasting. Evaluation results show that UniVQ surpasses state-of-the-art models in both zero-shot and in-domain experiments across six datasets. Code will be publicly available.