Title: ODTQA-FoRe: An Open-Domain Tabular Question Answering Dataset for Future Data Forecasting and Reasoning

Abstract:

While the swift evolution of Large Language Models (LLMs) has markedly improved the field of tabular question answering, most existing systems lack the capability to execute forward-looking numerical predictions. To bridge this critical gap, we introduce a new task designated as Open-Domain Tabular Question Answering for Future Data Forecasting and Reasoning (ODTQA-FoRe). We present the inaugural dataset designed to encompass both time-series forecasting and reasoning based on forecasts, utilizing real estate data. This task introduces significant hurdles, including the retrieval of accurate historical information, the mitigation of LLM forecasting constraints, and the standardization of answers across a wide variety of queries.

To tackle these difficulties, we developed TimeFore, a framework driven by an LLM agent that divides the problem into three distinct, collaborating roles: a Retriever that autonomously writes SQL queries to extract data, a Forecaster that leverages external time-series models to enhance precision, and an Analyzer that integrates these outputs to formulate a final answer that is both accurate and consistent. Our extensive experimental results confirm the efficacy of the TimeFore approach.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC