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

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.

Introduction and Motivation

Financial forecasting traditionally relies on numerical time-series (TS) signals, but textual information such as news remains largely underutilized in this field While these signals capture certain market dynamics, textual information such as financial news remains largely underutilized despite its significant role in reflecting market-relevant events and information Recent advances in natural language processing and representation learning have made it increasingly feasible to incorporate textual information into financial prediction models However, progress in this direction remains limited due to the lack of standardized benchmarks that systematically integrate financial TS data with aligned textual signals Many existing datasets focus primarily on sentiment analysis or provide loosely associated news streams, making it difficult to rigorously evaluate how effectively models leverage textual information for financial forecasting tasks To address this limitation, the paper constructs a dataset that aligns financial TS data with curated financial text by leveraging FinTexTS (Lee et al., 2026) This generation process yields aligned and context-aware financial news signals, enabling a more rigorous evaluation of newsgrounded forecasting models

XForecast Benchmark and Dataset Construction

XForecast is introduced as a benchmark and competition for news-grounded financial forecasting where the task is to predict stock price movements four weeks ahead using both historical price data and paired news signals The challenge covers 100 large-cap stocks from 2019–2023, with training on 2019–2022 and evaluation on 2023 To construct this dataset, the authors leverage FinTexTS (Lee et al., 2026), a multimodal dataset generation method that produces multi-level paired financial texts capturing macroeconomic, sector-level, related-company, and target-company contexts This structured pairing framework allows models to incorporate broader financial context beyond firmspecific news and enables more meaningful alignment between textual signals and financial TS dynamics

Modeling Approach

The core task of the XForecast challenge is formally defined as predicting the future closing price as Cˆi,t = f (Ci,τ≤t−4 weeks, Ti,τ≤t−4 weeks), where f(·) represents the forecasting model that integrates both numerical TS data and paired textual information To incorporate textual information while ensuring practical feasibility in the Kaggle competition environment, direct processing of raw text using large language models during training or inference is not allowed due to limited computational resources Text representations are restricted to embeddings generated by six allowed models: SBERT (Reimers & Gurevych, 2019), LGAI-Embedding-Preview (Choi et al., 2025), Linq-Embed-Mistral (Choi et al., 2024), llama-embed-nemotron-8b (Babakhin et al., 2025), Qwen3-Embedding-8B (Zhang et al., 2025), and gemini-embedding-001 (Google, 2024)

Evaluation Metric

The performance of models is evaluated using a weighted hit rate metric, which measures the accuracy of predicting the direction of stock price movements The directional correctness indicator is defined as Hi,t = sign(Cˆi,t − Ci,t−4 weeks) = sign(Ci,t − Ci,t−4 weeks)i Unlike a standard hit rate that assigns equal importance to all instances, the weighted average is computed as weighted hit rate = 1/P i,t wi,t X i,t wi,tHi,t The weights wi,t are designed to emphasize cases where news information is expected to play a stronger role in explaining subsequent price movements To compute the weight for each prediction instance, the paired news content is decomposed into atomic statements and GPT assigns each statement a positive or negative tag according to its expected impact on the target company The instance weight wi,t is determined from the proportion of atomic statements whose tagged impact direction agrees with the realized price direction

Competition Setup and Results

The XForecast challenge focuses on large-cap U.S. equities, specifically selecting the top 100 companies by market capitalization that were continuously tradable during 2019 to 2023 Data from 2019–2022 is used for training, while 2023 is reserved exclusively for evaluation Competition rules strictly prohibit any form of look-ahead bias, requiring models to use only information available up to t − 4 weeks when predicting the closing price at time t Participants must utilize the provided text embeddings and submit fully reproducible code implemented as Kaggle notebooks The competition attracted 341 submissions, with the first-place solution submitted by finx hanyang from Hanyang University Representative solution strategies among high-ranking teams included ensembles, cross-sectional information pooling to estimate common market regimes, and transforming text into compact auxiliary signals before fusion with price momentum and volatility features

Conclusion

XForecast introduces a benchmark and competition for news-grounded financial forecasting that integrates historical stock price data with curated textual signals By combining structured financial texts generated using FinTexTS with historical stock price data, the dataset enables systematic evaluation of how financial news can be leveraged for stock price forecasting The challenge aims to encourage research on multimodal financial forecasting and advance methods that more effectively integrate numerical market data with textual information The competition demonstrates substantial interest in news-grounded financial forecasting through 341 submissions

References

Araci, D. Finbert: Financial sentiment analysis with pretrained language models. In ACL Financial NLP Workshop, 2019. URL https://www.kaggle.com/competitions/xforecast-challenge-kdd (Note: The reference list provided in the paper contains an entry for Araci et al., 2019, but the anchor points to a different context based on the text flow).

Babakhin, Y., Osmulski, R., Ak, R., Moreira, G., Xu, M., Schifferer, B., Liu, B., and Oldridge (Note: The reference list provided in the paper contains an entry for Babakhin et al. 2025). URL https://arxiv.org/abs/2511.07025

Bai, S., Kolter, J. Z., and Koltun, V. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling (Note: The reference list provided in the paper contains an entry for Bai et al., 2018). arXiv preprint arXiv:1803.01271, 2018.

Bollerslev, T. Generalized autoregressive conditional heteroskedasticity (Note: The reference list provided in the paper contains an entry for Bollerslev, 1986). Journal of Econometrics, 31(3):307–327, 1986.

Box, G. E. P., Jenkins, G. M., Reinsel, G. C., and Ljung, G. M. Time Series Analysis: Forecasting and Control (Note: The reference list provided in the paper contains an entry for Box et al., 2015). Wiley, 2015.

Choi, C., Kim, J., Lee, S., Kwon, J., Gu, S., Kim, Y., Cho, M., and yong Sohn Linq-embed-mistral technical report (Note: The reference list provided in the paper contains an entry for Choi et al. 2024). URL https://arxiv.org/abs/2412.03223.

Choi, J., Kim, H., Jang, H., Jun, C., Bae, K., Choi, H., Choi, S. J., Lee, H., and Yun (Note: The reference list provided in the paper contains an entry for Choi et al. 2025). URL https://arxiv.org/abs/2506.07438.

Ding, X., Zhang, Y., Liu, T., and Duan Deep learning for event-driven stock prediction (Note: The reference list provided in the paper contains an entry for Ding et al., 2015). In International Joint Conference on Artificial Intelligence, 2015.

Giantsidi, S. and Tarantola, C. Deep learning for financial forecasting: A review of recent trends (Note: The reference list provided in the paper contains an entry for Giantsidi & Tarantola, 2025). International Review of Economics & Finance, 104:104719, 2025.

Lee, J., Park, S., Lim, T. Y., Lee, S., Seo, J., Kang, D., Choi H., Kim M., Yoo S. Fintexts: Financial text-paired time-series dataset via semantic-based and multi-level pairing (Note: The reference list provided in the paper contains an entry for Lee et al. 2026).

1. textbfWeighting Metric Refinement for News Influence Stability. The current method assigns weights based on atomic statement agreement with realized price direction; this could be improved by incorporating temporal decay or uncertainty measures into the weighting scheme to prevent single, noisy news events from disproportionately influencing the final score, addressing the implicit assumption that instances for which a larger share of the news statements correctly indicates the subsequent movement receive greater weight.

2. textbfContextual Feature Integration via FinTexTS. The model can be improved by explicitly leveraging the multi-level contexts produced by FinTexTS; instead of simply feeding precomputed embeddings, models should be designed to utilize semantic-based and multi-level pairings between financial texts and stock market data, capturing macroeconomic context, sector-level information, related-company signals, and company-specific events to generate richer input features.

3. textbfEnsemble Strategy Optimization for Multimodal Fusion. To enhance robustness, the system can be improved by moving beyond simple ensembles; participants should explore strategies that combined heterogeneous model families or separated the prediction into market-direction and stock-specific components, specifically testing how different architectures (e.g., diffusion models versus temporal neural networks) interact with the fused textual and TS representations.

4. textbfModel Architecture Selection for Temporal Alignment. Since careful temporal alignment is highlighted as important, the system can be improved by prioritizing architectures like TFT or TimesNet that are explicitly designed for multi-horizon forecasting, ensuring that the integration of historical price dynamics and paired financial news signals respects the required look-back window without introducing look-ahead bias.

5. textbfControlled Text Representation Utilization. Given the restriction to precomputed embeddings from a specific set of models, the system can be improved by developing an embedding novelty feature or a pooled news representations approach, which are explicitly mentioned as strategies explored by high-ranking teams, allowing the model to learn how to weigh different embedding sources effectively.

Financial forecasting has traditionally relied on numerical time-series (TS) signals, while textual information such as news remains largely underutilized. We introduce XForecast, a benchmark and competition for news-grounded financial forecasting, where the task is to predict stock price movements four weeks ahead using both historical price data and paired news signals. XForecast combines stock price data with curated financial text generated using FinTexTS, a multimodal dataset generation method. It covers 100 large-cap stocks from 2019--2023, with training on 2019--2022 and evaluation on 2023. To better evaluate models that leverage textual signals, we adopt a weighted hit rate metric that places greater emphasis on instances where news information is expected to have stronger influence. The XForecast Challenge, hosted on Kaggle as part of KDD 2026 Finance Day, aims to benchmark methods for integrating financial TS and textual information.