LG AI Research: The Rise and Classification of Time Series Foundation Models

Time series data—such as stock prices, sensor data, and weather—has a time axis, and AI-based forecasting is crucial for responding to future uncertainty. Recently, research on Time Series Foundation Models, which can solve various time series problems without being limited to a specific task, has been actively pursued.

Time series foundation models are broadly classified into two methodologies. The first is the approach of utilizing an LLM (Large Language Model), and the second is the approach of directly building a separate large-scale model exclusively for time series data.

The LLM-based methodology is further divided into three types depending on how the model is utilized.

• Without Adaptation: This focuses on prompt design without any separate training. The example PrompCast utilizes text input/output, but performance degradation can occur due to numerical Tokenization issues. To address this, the LLM-Time approach, which represents numbers as individual tokens, has also been proposed.
• Adapt LLM: Based on the premise that the LLM is not familiar with the time series domain, this approach adds an Adaptation Layer to connect the time series data and the LLM, or fine-tunes part of the model.
• Adapt to LLM: (Body text omitted)