Abstract
Computer-implemented time-series forecasting systems, such as those executing on a server or distributed computing platform, may have difficulty anticipating demand volatility caused by real-world events due to a potential reliance on historical quantitative data. This disclosure relates to systems and methods for adjusting short-term forecasting by processing unstructured text data, for instance from news articles, to identify and quantify events that could influence future demand. A data processing pipeline can use a large language model to extract event descriptions, generate numerical embeddings for those events, and aggregate them into structured features. These features can then be integrated as dynamic covariates into a time-series forecasting model. This approach may allow the model to be more responsive to forward-looking indicators not yet present in historical data. An explainable interface can also be provided to offer traceability from a forecast modification back to the source text that influenced it, potentially improving user comprehension of the model's output.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Kuligin, Leonid and Zettl, Thomas, "Time-Series Forecasting Using Dynamic Covariates from Text-Based Event Extraction", Technical Disclosure Commons, (December 02, 2025)
https://www.tdcommons.org/dpubs_series/8972