AFAC 2026 Financial AI Finals in Shanghai Draw 5,027 Teams, Release Massive Open Dataset

The competition, which is regarded as an annual benchmark for financial AI innovation, was co-launched by more than 30 institutions, including the Shanghai Municipal Science and Technology Commission, the China Computer Federation, Peking University, Ant Group and the Global Finance & Technology Network. On-site contestants included students from top universities such as Tsinghua, Peking, Fudan and MIT, as well as core engineers from major technology firms.

This year's contest set four challenge tracks designed to simulate real-world financial pressures: identifying trading behavior and fund flows from market data, restoring complex financial documents, automating experiments with sparse feedback, and compressing dynamic context for long-text financial questions. According to the report, the questions required contestants to build systems that can replicate results, explain their reasoning and generalize to changing conditions, rather than merely chasing high scores on a fixed benchmark.

Experts involved in judging pointed to the gap between benchmark performance and real deployment. Wang Chao, a professor at the University of Science and Technology of China, noted that the market is dynamic while contestants only have static historical data, so their models must be robust and interpretable. Zhang Junping, a professor at Fudan University, said complex tables in financial documents remain a major challenge for large models, especially when templates change. Zhang Yu, a professor at Harbin Institute of Technology, said the real differentiator is the system's ability to control its context window, not just its maximum length. Xu Wanqing, head of Ant Group's Wealth AI Lab, explained that in the AI era, every extra user consumes tokens that correspond to real costs, making cost management an inherent part of system capability.

The newly released AFAC Million-Scale Financial Intelligence Dataset was built from 14 high-value competition tasks between 2023 and 2025, containing 130,000 evaluation samples, more than 2,000 professional financial documents and 130 award-winning solutions, covering data items on the scale of tens of billions. The dataset is intended to give researchers a common foundation to build upon, creating a data flywheel for the financial AI community.

Since the contest's first edition in 2023, it has attracted more than 20,000 teams and 60,000 participants in total, covering over 600 universities and 700 enterprises. This year, the organizers expanded their network to 33 co-organizers, adding the Global Finance & Technology Network as a partner. The final featured a combined conference, exhibition and competition format, with a prize pool of one million yuan, computing support, direct job opportunities and a six-month acceleration program for early-stage teams.

At the finals, SGInnovate Director Jae Annie Tay, who was attending AFAC for the first time, said smaller teams especially need a larger external ecosystem that connects them with talent, capital and industrial resources. The event also showcased previous achievements and opened frontier products such as Alibaba's AI glasses for hands-on experience.

The evolution of the contest's tracks over four years reflects a broader shift in financial AI from demonstrating model capabilities to building useful, cost-aware systems. This year's tasks focused on agent behavior, sparse feedback and token economics, themes that have dominated recent AI industry discussions.