New Framework Uses Vision Models for Time Series Forecasting

Researchers have introduced ICI-Time, an innovative approach that reimagines time series forecasting by treating it as a visual inpainting problem. This method capitalizes on the robust generalization capabilities of large vision models (LVMs), sidestepping the need for bespoke temporal architectures or extensive, domain-specific training typically associated with time series analysis. The core of ICI-Time involves converting time series data into structured visual formats, specifically area charts. These visual representations are then processed using visual in-context learning, allowing pre-trained vision transformers to complete patterns within a grid-structured prompt. This novel paradigm establishes a consistent and invertible mapping between numerical time series data and its visual counterpart, enabling LVMs to perform forecasting without fine-tuning or architectural modifications. Extensive evaluations across diverse fields such as epidemiology, meteorology, and power systems demonstrate that ICI-Time achieves competitive performance against established deep learning baselines. The framework also shows promising adaptability in scenarios with limited data, marking a significant step in bridging the temporal and visual domains for forecasting tasks.