Time-VLM: Exploring Multimodal Vision-Language Models for Augmented Time Series Forecasting

Recent progress in time series forecasting has increasingly incorporated textual and visual modalities to enhance predictive performance. While textual data offer rich contextual insights, they often overlook detailed temporal dynamics. In contrast, visual representations can capture complex temporal variations but lack semantic understanding, reducing the synergy between modalities. To overcome these limitations, we introduce Time-VLM, a multimodal forecasting framework that harnesses pre-trained Vision-Language Models (VLMs) to unify temporal, visual, and textual information. The framework integrates three core modules: (1) a Retrieval-Augmented Learner that enriches temporal representations via memory-based retrieval; (2) a Vision-Augmented Learner that transforms time series into informative visual features; and (3) a Text-Augmented Learner that generates contextual textual semantics. These modules collaboratively interact with frozen VLMs to produce multimodal embeddings, which are fused with temporal encodings for final prediction. Extensive experiments show that Time-VLM delivers state-of-the-art results, particularly in few-shot and zero-shot forecasting tasks, highlighting its potential as a new paradigm for multimodal time series modeling.