We propose a novel approach based on large language causal models to perform the task of time-series forecasting, and we use the proposed approach to effectively forecast the concentration of polluting substances in a water treatment plant; we address both short- and mid-term forecasting. As opposed to the classical state-of-the-art approaches for time-series forecasting, that handle numerical and categorical features following a standard deep learning approach, we transform the input features into a textual form and we then feed them to a standard causal model pre-trained on natural language tasks. Our empirical results provide evidence that large language models are more effective than state-of-the-art forecasting systems, and that they can be practically used in time-series forecasting tasks. We also show promising results on zero-shot learning. The results of this study open up to a wide range of works aimed at predicting future temporal values by leveraging natural language paradigms and models.
