Can LLM Coding Agents Reason About Time Series?

This study investigates the ability of large language models (LLMs) to analyze time series data through three approaches: raw numerical data input, coding agents that query data using Python, and a hybrid method. Results from two time series benchmarks indicate that coding agents can outperform raw data models by up to 10%, though they still answer 22-34% of questions incorrectly. The findings highlight that while coding agents can select suitable statistical tests, they struggle with nuanced reasoning, suggesting a need for improved model strategies in automated decision-making systems.