Inferring Event Descriptions from Time Series with Language Models

Mingtian Tan, Mike A. Merrill, Zack Gottesman, Tim Althoff, David Evans, Tom Hartvigsen

Abstract

Time series data measure how environments change over time and drive decision-making in critical domains like finance and healthcare. When analyzing time series, we often seek to understand the underlying events occurring in the measured environment. For example, one might ask: What caused a sharp drop in the stock price? Events are often described with natural language, so we conduct the first study of whether Large Language Models (LLMs) can infer natural language events from time series. We curate a new benchmark featuring win probabilities collected from 4,200 basketball and American football games, featuring 1.7M timesteps with real value data and corresponding natural language events. Building on the recent wave of using LLMs on time series, we evaluate 16 LLMs and find that they demonstrate promising abilities to infer events from time series data. The open-weights DeepSeek-R1 32B model outperforms proprietary models like GPT-4o. Despite this impressive initial performance, we also find clear avenues to improve recent models, as we identify failures when altering the provided context, event sequence lengths, and evaluation strategy.

Introduction

While significant research has focused on using Large Language Models (LLMs) to forecast future numerical values in time series, the inverse problem—inferring the events that drive those numerical changes—remains largely unexplored. Understanding the "why" behind data fluctuations in domains like finance or healthcare is critical for decision-making. This paper introduces GAMETime, a benchmark designed to evaluate if LLMs can "reverse-engineer" natural language event descriptions from numerical time series data.

Illustration of the GAMETime task: inferring NBA play-by-play events from win probability time series.

Methodology and Datasets

The core of this study is the GAMETime dataset, which leverages real-world sports data. It includes 4,200 NBA and NFL games, totaling 1.7 million data points. The dataset pairs ESPN’s "win probability" time series with corresponding play-by-play descriptions. To test generalizability, the researchers also evaluated models on four additional domains: cryptocurrency prices, international trade, public health (influenza rates), and energy consumption.

The task is framed as a multiple-choice challenge: given a sequence of timestamps and values, the model must select the correct sequence of events from four options. To optimize performance, the researchers explored:

  • Chain-of-Thought (CoT) Prompting: Encouraging logical deduction before answering.
  • Knowledge Distillation: Fine-tuning smaller models (e.g., Qwen2.5 1.5B) on reasoning traces from larger models like DeepSeek-R1 32B.
  • Self-Improvement (GRPO): Using Group Relative Policy Optimization to reward correct answers and formatting.

Key Results and Insights

The evaluation of 16 LLMs revealed that reasoning-focused models are significantly better at this task than standard models. OpenAI’s o1 achieved the highest accuracy on NBA data (83%), while standard models like GPT-4o performed more modestly. Notably, the open-weights DeepSeek-R1 32B outperformed several proprietary models.

Performance of various LLMs on the NBA event inference task, comparing results with and without Chain-of-Thought (CoT) prompting.

A major finding was the "Underdog Win": a base Qwen2.5 1.5B model, which initially performed near random (5.5% accuracy), reached 75.5% accuracy after distillation and GRPO training. This suggests that even small models can be "supercharged" to expert levels for specific time-series reasoning tasks.

The researchers also identified a "Goldilocks" reasoning length, where performance peaks at approximately 1,400 tokens. "Rushed reasoning" leads to errors, while "overthinking" causes models to get stuck in self-reflective loops that degrade accuracy. Furthermore, the models' reliance on event order confirms they are performing logical deduction rather than simple pattern matching.

Conclusion

This work demonstrates that LLMs possess a promising, yet nascent, ability to map numerical trends to natural language narratives. While the current benchmark uses multiple-choice questions, future research should aim for open-ended event generation. Such capabilities could revolutionize automated reporting in fields like medical chart summarization and financial analysis, where explaining the "story behind the numbers" is essential.

Summary updated

Abstract

Time series data measure how environments change over time and drive decision-making in critical domains like finance and healthcare. When analyzing time series, we often seek to understand the underlying events occurring in the measured environment. For example, one might ask: What caused a sharp drop in the stock price? Events are often described with natural language, so we conduct the first study of whether Large Language Models (LLMs) can infer natural language events from time series. We curate a new benchmark featuring win probabilities collected from 4,200 basketball and American football games, featuring 1.7M timesteps with real value data and corresponding natural language events. Building on the recent wave of using LLMs on time series, we evaluate 16 LLMs and find that they demonstrate promising abilities to infer events from time series data. The open-weights DeepSeek-R1 32B model outperforms proprietary models like GPT-4o. Despite this impressive initial performance, we also find clear avenues to improve recent models, as we identify failures when altering the provided context, event sequence lengths, and evaluation strategy. (All resources needed to reproduce our work are available: https://github.com/BennyTMT/GAMETime)
BibTeX
@inproceedings{tan2026inferring,
  author = {Mingtian Tan and Mike A. Merrill and Zack Gottesman and Tim Althoff and David Evans and Tom Hartvigsen},
  title = {{Inferring Event Descriptions from Time Series with Language Models}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.157},
}
Figure