XForecast: The KDD 2026 Finance Day Challenge on News-Grounded Financial Forecasting

summary

The gist

The gist: XForecast introduces a benchmark and competition for news-grounded financial forecasting that integrates historical stock price data with curated textual signals to predict stock price

In short

XForecast creates a benchmark and competition for predicting stock price movements four weeks ahead by integrating historical stock data with curated financial news signals. It addresses the lack of standardized datasets by using FinTexTS to align financial time series with context-aware text, pushing multimodal forecasting methods.

Key concepts

News-grounded Financial Forecasting: This is a prediction task where models use both numerical historical stock prices and relevant textual information (like news articles) to forecast future stock movements. The goal is to see if incorporating real-time or historical news helps predict price changes more accurately than using only numbers.
FinTexTS: This is a method used to create a high-quality dataset by pairing financial time series data with structured, multi-level financial text. It captures different contexts, such as macroeconomic trends and company-specific news, ensuring the text signals are highly relevant to the stock prices.
Weighted Hit Rate: This is a specific evaluation metric used to judge model performance. Instead of treating all correct predictions equally, this metric assigns higher importance (weights) to predictions where news information was expected to be most influential in explaining the price change.

Terminology used across episodes

Reinforcement Learning (RL) — explained in 25 episodes
Sample Complexity — explained in 20 episodes
Retrieval-Augmented Generation (RAG) — explained in 18 episodes
Federated Learning — explained in 17 episodes
Quantum Fisher Information (QFI) — explained in 16 episodes
Flow Matching — explained in 15 episodes

This episode discusses

The paper

XForecast: The KDD 2026 Finance Day Challenge on News-Grounded Financial Forecasting · Read on arXiv

Jaehoon Lee, Jun Seo, Seunghan Lee, Tae Yoon Lim, Dongwan Kang, Hwanil Choi, Minjae Kim, Sungdong Yoo, Junhyeok Kang, Sangjun Han, Soonyoung Lee, Yongjae Lee

Ulsan National Institute of Science and Technology (UNIST)

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "XForecast: The KDD 2026 Finance Day Challenge on News-Grounded Financial Forecasting".

Jane: The gist: XForecast introduces a benchmark and competition for news-grounded financial forecasting that integrates historical stock price data with curated textual signals to predict stock price movements four weeks ahead.

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we're wrapping up our talk on XForecast, focusing on how this benchmark helps us understand news-grounded financial forecasting with the XForecast: The KDD two thousand twenty-six Finance Day Challenge on News-Grounded Financial Forecasting paper. We've seen how the challenge sets a standard for integrating historical stock price data with curated text signals to predict future movements.

Jane: Specifically, they introduce XForecast as a benchmark and competition where the task is to predict stock price movements four weeks ahead using both historical price data and paired news signals <ref:2610.16131#pg1>. It covers one hundred large-cap stocks from two thousand nineteen to two thousand twenty-three with training happening on two thousand nineteen to two thousand twenty-two and evaluation done on two thousand twenty-three.

Lu: The core idea is the construction of a curated dataset using FinTexTS, which is a multimodal dataset generation method that produces multi-level paired financial texts capturing macroeconomic, sector-level, related-company, and target-company contexts <ref:2610.16131#pg2>.

Meng: The authors also propose a weighted hit rate metric that places greater emphasis on instances where news information is expected to have stronger influence on price changes <ref:2610.16131#pg2>. That means the metric isn't just about getting the final direction right, but weighting those predictions based on how much news is actually expected to matter.

Tom: So we're not just looking for accuracy; we're focusing on the instances where news truly drives those price movements, which is a big shift in how we look at this problem.

Jane: That means future models won't just be looking at past prices; they'll be actively reading and interpreting financial news as part of their decision-making process to predict what happens next <ref:2610.16131#pg3>.

Lu: The big picture here is pushing multimodal AI to a point where the model has to understand both time series and language simultaneously to get an answer, which is what they are trying to encourage research on <ref:2610.16131#pg2>.

Tom: And even though three hundred forty-one submissions came in for the challenge, that competition itself is what really tests if these integrated systems can perform under real pressure when you’re trying to predict those stock price movements four weeks out <ref:2610.16131#pg3>.

Jane: This whole effort is all wrapped up in the paper titled "XForecast: The KDD two thousand twenty-six Finance Day Challenge on News-Grounded Financial Forecasting" <ref:2610.16131#pg2>.

Conclusion: Jane: This research moves us beyond just using numbers on a chart by including context from financial news, which is what this entire effort is about <ref:2610.16131#pg3>. It’s about building models that actually understand market events as they happen.

Lu: The authors show how you can combine historical stock price data with these curated textual signals to predict what happens next four weeks out, which is a structured way to test if financial news actually matters in forecasting <ref:2610.16131#pg2>.

Meng: They’re proving that you don't need just the numbers; you need the context of the news to get a real edge in predicting those movements <ref:2610.16131#pg2>.

Tom: It’s about moving from simple time-series prediction toward something that actually understands market events as they happen, which is a big step forward for AI models <ref:2610.16131#pg3>.

Jane: And this is what we mean when we talk about news-grounded forecasting—it’s not just noise; it's signal <ref:2610.16131#pg3>.

Lu: This work gives us a structured way to test if financial news actually matters in forecasting beyond just looking at the numbers on a chart by itself <ref:2610.16131#pg2>.

Tom: Right. And this structured testing is exactly what makes this paper important for the broader field of AI and finance today <ref:2610.16131#pg3>.

More episodes

← Home