IEEE Logo Columbia Logo The Fin AI Logo

Introduction

Financial reinforcement learning FinRL [1,2] is an interdisciplinary field that applies reinforcement learning to financial tasks, such as portfolio management, algorithmic trading, and option pricing. The recent breakthroughs of large language models (LLMs) is driving open finance, which provides affordable and scalable solutions for customers to make intelligent decisions and created personalized financial search and robo-advisors.

The FinRL Contest 2025 explores and evaluates the capability of machine learning methods in finance, with the following features:

  1. FinRL-DeepSeek. DeepSeek’s groundbreaking open models have shown strong capabilities on par with popular close models. FinRL-DeepSeek [3] integrates financial reinforcement learning with large language models to enhance stock trading strategies. It adds risk assessment and trading recommendation signals generated by DeepSeek models from financial news. In this contest, we encourage participants to explore and innovate with FinRL-DeepSeek.
  2. FinRL-AlphaSeek. Factors play a critical role in driving trading decisions, enabling traders to design efficient, data-driven strategies. Two stages can be set up for training robust trading agents – factor engineering and ensemble learning. In this contest, we encourage participants to independently perform factor engineering and selection and utilize ensemble learning to create their trading strategies.
  3. Open FinLLM Leaderboard. The Open FinLLM Leaderboard [4] serves as an open platform to evaluate LLMs’ performance on various financial tasks. It encourages the community to explore the models’ capability and openness. In addition, complex financial regulations and industry standards are critical to the financial services, but the leaderboard haven’t included such benchmark datasets. We expanded and upgraded the regulations datasets developed during the Regulations Challenge at COLING 2025 [5]. In FinRL Contest 2025, we selected three domains for digital regulatory reporting: the Common Domain Model (CDM), the Model Openness Framework (MOF), and eXtensible Business Reporting Language (XBRL). We aim to contribute these datasets to the leaderboard so that the participants can explore both existing tasks and the new regulatory reporting task. We also encourage participants to utilize reinforcement learning to develop their LLMs.

We design four tasks to promote open finance: (1) FinRL-DeepSeek for Stock Trading, (2) FinRL-AlphaSeek for Crypto Trading, (3) Open FinLLM Leaderboard – Models with Reinforcement Fine-Tuning (ReFT), and (4) Open FinLLM Leaderboard – Digital Regulatory Reporting (DRR). These challenges allow contestants to participate in various financial tasks and contribute to open finance using state-of-the-art technologies. We welcome students, researchers, and engineers who are passionate about finance and machine learning to partake in the contest.

Tasks

Each team can choose to participate in one or more tasks. The prizes will be awarded for each task.

In addition, we will provide tutorials for participants to learn FinRL and the Open FinLLM Leaderboard.

Task I: FinRL-DeepSeek for Stock Trading

This task is about developing automated stock trading agents trained on stock prices and financial news data by combining reinforcement learning and large language models (LLMs). Participants can build upon the FinRL-DeepSeek project (e.g., with new prompts, new ways to inject LLM-processed news signals into the RL agent, new RL algorithms like GRPO) or explore more computationally intensive directions, such as adapting variants of the DeepSeek R1 training method to this stock trading task

Datasets

The Financial News and Stock Price Integration Dataset (FNSPID) [6] comprises stock prices and 15 million time-aligned financial news records for Nasdaq companies, covering the period from 1999 to 2023. The processed training dataset based on the FNSPID will be provided. Participants are also encouraged to utilize publicly available data, such as Twitter, or develop scraping/API AI agents for this purpose. Some teams can choose to focus their submission on improving the dataset, others on improving trading agents.

Task II: FinRL-AlphaSeek for Crypto Trading

This task aims to develop robust and effective trading agents for cryptocurrencies through factor mining and ensemble learning. In this task, participants are expected to explore useful factors and ensemble methods for crypto trading. Participants are free to apply various techniques to the factor engineering process, design component models, and use innovative methods to increase the diversity of component models in the ensemble. They also need to specify the state space, action space and reward function in the environment. The final model should be able to interact with the provided trading environment.

Dataset

We provide second-level Limit Order Book (LOB) data for Bitcoin. Participants are permitted to use additional external datasets.

Task III: Open FinLLM Leaderboard – Models with Reinforcement Fine-Tuning (ReFT)

This task aims to encourage the community to learn, maintain, and update the Open FinLLM Leaderboard – add new models (rows) to the leaderboard. In this task, participants are expected to submit their models and compete for high rankings in the leaderboard. Participants are free to train or fine-tune their models, which will be evaluated across all tasks in the leaderboard. We encourage participants to explore reinforcement fine-tuning (ReFT) techniques to enhance LLMs’ capabilities in financial tasks.

Dataset

The public benchmark datasets are described at the leaderboard. Participants can collect the data themselves to train or fine-tune their models.

Task IV: Open FinLLM Leaderboard - Digital Regulatory Reporting (DRR)

This task aims to challenge the community to explore the strengths and limitations of LLMs in digital regulatory reporting and contribute the new task to the Open FinLLM Leaderboard – add new datasets (columns) to the leaderboard. Participants are expected to train or fine-tune their LLMs to perform tasks in the three domains: the CDM, the MOF, and XBRL:

  • CDM is a machine-oriented model for managing the lifecycle of financial products and transactions.
  • XBRL is a standard for electronic communication of business and financial data but often has a high error rate in the filing process.
  • MOF evaluates and classifies the completeness and openness of machine learning models. We also encourage participants to explore reinforcement learning fine-tuning solutions.

Dataset

We developed and upgraded the question datasets for the three domains. We will provide the data sources so that participants can collect data themselves. The full question datasets will be released during the evaluation period.

[1] X.-Y. Liu, Z. Xia, H. Yang, J. Gao, D. Zha, M. Zhu, Christina D. Wang*, Zhaoran Wang, and Jian Guo. Dynamic datasets and market environments for financial reinforcement learning. Machine Learning Journal, Springer Nature, 2023.

[2] X.-Y. Liu, Z. Xia, J. Rui, J. Gao, H. Yang, M. Zhu, C. Wang, Z. Wang, J. Guo. FinRL-Meta: Market environments and benchmarks for data-driven financial reinforcement learning. NeurIPS, Special Track on Datasets and Benchmarks, 2022.

[3] Mostapha Benhenda. 2025. FinRL-DeepSeek: LLM-Infused Risk-Sensitive Reinforcement Learning for Trading Agents. arXiv preprint arXiv:2502.07393 (2025).

[4] Shengyuan Colin Lin, Felix Tian, Keyi Wang, Xingjian Zhao, Jimin Huang, Qian-qian Xie, Luca Borella, Christina Dan Wang Matt White, Kairong Xiao, Xiao-Yang Liu Yanglet, and Li Deng. 2024. Open FinLLM Leaderboard: Towards Financial AI Readiness. International Workshop on Multimodal Financial Foundation Models (MFFMs) at 5th ACM International Conference on AI in Finance (MFFM), 2024.

[5] Keyi Wang, Jaisal Patel, Charlie Shen, Daniel Kim, Andy Zhu, Alex Lin, Luca Borella, Cailean Osborne, Matt White, Steve Yang, Kairong Xiao, and Xiao-Yang Liu Yanglet. 2025. A Report on Financial Regulations Challenge at COLING 2025. FinNLP-FNP-LLMFinLegal-2025 Shared Task: Regulations Challenge. Proceedings of the Joint Workshop of the 9th Financial Technology and Natural Language Processing (FinNLP), the 6th Financial Narrative Processing (FNP), and the 1st Workshop on Large Language Models for Finance and Legal (LLMFinLegal).

[6] Zihan Dong, Xinyu Fan, and Zhiyuan Peng. 2024. FNSPID: A Comprehensive Financial News Dataset in Time Series. arXiv preprint arXiv:2402.06698 (2024).

Contact

Contestants can communicate any questions on Discord.

Contact email: finrlcontest@gmail.com