Abstract
Large language models (LLMs) have recently been utilized to construct forecasting models for time series data. Both classical statistical models and newer LLM-based models offer distinct advantages and disadvantages. Significant effort is often invested in selecting and customizing suitable models. Current blending techniques, often employed in ensemble learning, involve training an additional layer during the training phase. This additional layer combines the prediction results of individual forecasting models to generate the final outcome. These techniques necessitate an extra training process, preventing direct use at inference time, and lack full customization for time series data forecasting. This disclosure describes techniques comprising the following components: a set of existing forecasting models, another set of forecasting models designed to predict the forecasting accuracy of the first set, and a blending layer. Additionally, the described techniques include an inference framework utilizing these components, enabling inference-time use cases, or alternatively, a training framework utilizing the same components.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Yim, K. S., "Blending Forecasting Models via Meta Forecasting", Technical Disclosure Commons, (October 22, 2024)
https://www.tdcommons.org/dpubs_series/7466