A new paper on arXiv, titled "Lotus: Learning to Understand Time-series by Learning to Forecast," challenges how time-series language models are trained. The authors observe that current models align numerical signals with language using supervision like labels, captions, or question-answer pairs. While this encourages linguistic representations, the abstract argues that it does not explicitly require those representations to be predictive of future values.
The paper's central proposal is to shift the learning objective toward forecasting. By learning to forecast, the model may develop a deeper understanding of the underlying temporal structure, rather than merely mapping numbers to text. The abstract suggests that this forecasting-based approach is what distinguishes Lotus from existing methods.
As the only source, the abstract provides limited detail on implementation or results. Still, the core idea is clear: understanding time-series may be better achieved by predicting what comes next, not just by describing what has already happened.