Cart: A Future-Oriented Text Impacts Alignment Framework for Time Series Forecasting

Time series forecasting is crucial in many real-world applications. Unlike unimodal time series forecasting models relying solely on historical data, recent multimodal methods align event texts with historical time series to enhance performance. However, they fail to clearly model the impact of external events on future time series. So we propose a framework which explicitly establishes an alignment between the impact of external event texts and the future time series, called Cart. As a future-oriented text impacts alignment framework, Cart first leverages large language models to extract indicative texts and then converts these texts into impact curves. Then, it uses temporally-constrained optimal transport and contrastive learning to align these impacts with the future time series, clearly showing how event texts impact future time series. Moreover, we also enhance the unimodal time series forecasting model using Cart in a plug-and-play fashion. The framework achieves significant improvements for unimodal time series forecasting models and achieves state-of-the-art results across multiple public datasets.