Comparison of LLM-based forecasting models for intermittent time series

Determine the comparative forecasting performance of large language model–based time series forecasting methods on intermittent, non-negative, zero-inflated series by conducting a systematic experimental comparison against established global neural forecasting approaches for probabilistic prediction of intermittent demand.

Background

The paper conducts the first systematic comparison of local and global probabilistic models for intermittent demand time series, evaluating feed-forward networks, DeepAR, transformer-based architectures, and D-Linear, each coupled with distribution heads suitable for intermittent data (negative binomial, hurdle-shifted negative binomial, and Tweedie). Across five large real-world datasets, D-Linear consistently provides strong accuracy with low computational cost, while transformer-based models are less effective and more expensive.

Although the study thoroughly benchmarks widely used neural architectures, it explicitly does not evaluate LLM-based forecasting methods. The authors highlight this gap and defer it as future research, indicating the need to determine how LLM-based approaches compare to established global neural architectures in the context of probabilistic forecasting for intermittent demand.