Evaluating System 1 vs. 2 Reasoning Approaches for Zero-Shot Time Series Forecasting: A Benchmark and Insights

Haoxin Liu, Zhiyuan Zhao, Shiduo Li, B. Aditya Prakash · 2025

Read Evaluating System 1 vs. 2 Reasoning Approaches for Zero-Shot Time Series Forecasting: A Benchmark and Insights by Haoxin Liu, Zhiyuan Zhao, Shiduo Li, B. Aditya Prakash (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-01895.txt.

Description

Haoxin Liu, Zhiyuan Zhao, Shiduo Li, and Aditya Prakash ask whether the reasoning tricks that sharpen language models also sharpen forecasters. Time series forecasting with no task-specific training, zero-shot forecasting, is a hard test bed. Their benchmark, ReC4TS, applies popular reasoning strategies, from chain-of-thought at run time to post-training methods, across data sets from eight domains.

Three lessons stand out. Self-consistency, sampling several answers and voting, was the best test-time strategy. A group-relative training method best built reasoning in during post-training. And forecasts that mix text with numbers gained more from reasoning than pure number series did. The team also releases a data set of reasoning traces for forecasting and a simple scaling rule for self-consistency, offered as starting blocks for future work.

Catalog record

CREATORHaoxin Liu, Zhiyuan Zhao, Shiduo Li, B. Aditya Prakash
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-01895

Browse the full library · About the Tower of Babel