FinGenius: an A-share AI financial agent that argues its way to a report

FinGenius is an AI financial analysis app aimed at China's A-share market. The idea is to hand a research team's routine - several specialists looking at the same question from different angles, then arguing it out - to a set of agents. The materials describe a multi-agent game-theoretic architecture in which 16 specialist agents (sentiment, hot money, risk control and others) divide the work and produce a combined financial analysis report.

The difference from a single-turn chatbot is the pipeline. A Research environment runs the collaborative analysis, a Battle environment runs a structured debate, and the full result is passed from one to the other. The materials also mention applying game theory to decisions and a memory algorithm that records a user's investing habits to personalise the analysis. What it produces is analysis material, not investment advice.

What it does

• Multi-agent debate system. Sentiment, hot money and risk-control agents each examine a different side of the same target, then their findings are merged into one report. The materials describe this as simulating how market participants interact, rather than having one model answer in a single pass.
• Two-environment architecture. The Research environment handles collaborative analysis and the Battle environment handles structured debate, mirroring the path from research to decision.
• Cumulative debate. The Battle environment carries all earlier expert discussion forward as context, so each round builds on the previous ones instead of producing unconnected statements.
• Deep data integration and reporting. Several data sources are combined into a structured, multimodal report that pulls out key information and risk points.
• Real-time processing and risk alerts. The materials claim millisecond-level handling of large volumes of financial data and real-time monitoring for market anomalies, used to flag risks early.
• MCP tool integration. Model-to-model collaboration runs over the Model Context Protocol, which the materials describe as supporting a plug-and-play ecosystem of financial tools.

How to use it

1. Open fingenius.cn, or install the mobile app (the materials say it is listed on the Honor, Xiaomi and Vivo app stores), then register and sign in.
2. Set your analysis parameters and alert conditions. The materials call this customisation; which settings actually exist is best checked inside the product.
3. Submit the target or question and let one full run happen: the agents analyse together first, then the debate stage begins.
4. When reading the report, look at three things: which data the conclusion rests on, which risk points were flagged, and which disagreements never resolved. The last is often the most useful.
5. To self-host, take the code from the HuaYaoAI/FinGenius repository on GitHub and install it with conda or uv as the repository guide describes.
6. After deploying, edit the configuration file to add your API key, then use command-line flags to control output format, saving results and text-to-speech. Agent steps and debate rounds are adjustable too.

Where it fits

• Investment research input: analysis material for individual stocks and portfolios that shortens information gathering and first-pass reasoning.
• Risk management: anomaly monitoring and risk flags as a routine step, instead of an after-the-fact investigation.
• Market trend tracking: an ongoing record of sector and market movement that can be compared over time.
• Research desk work: data integration and a first draft for analysts, with a person deciding which conclusions hold.
• Quantitative research: starting points for signals and strategy hypotheses, still validated by backtesting.

Limits worth knowing

• The materials give no pricing or plan information, so whether it is free and how usage is metered has to be checked against the current official description.
• Claims such as first of its kind and precise come from the materials and are the vendor's wording; this article does not treat them as independently verified.
• Millisecond processing and 16 agents are the materials' figures; actual latency, agent count and division of labour can change between versions.
• The materials say the app is listed on the Honor, Xiaomi and Vivo stores and do not say whether other platforms are supported.
• The report is analysis material, not investment advice. Markets are uncertain; verify the conclusions yourself, since gains and losses stay with the user.
• Multi-agent debate surfaces disagreement but does not resolve it. The disagreement is often the part most worth reading.
• The materials do not disclose the data-source list, update frequency, historical depth or regulatory qualifications; confirm with the vendor before wiring it into a formal process.

Frequently asked questions

Is FinGenius for individuals or institutions?

Both, with different uses. Individuals lean on it for analysis of stocks and portfolios and for watching risk points; institutions use it for data integration, first-draft reports and risk alerts. The materials do not separate personal and institutional feature sets.

How is it different from a normal AI chat?

A normal chat answers in one turn. FinGenius runs a pipeline: several agents analyse separately, then enter a debate stage and question each other, then the result is merged. The materials attribute that difference to the multi-agent architecture and the two-environment design.

Can the report be used as investment advice?

Not advisable. It is analysis material that puts information and reasoning on the table, but market judgement and the profit or loss stay with the user. The materials make no return promises either.

Do I have to self-host it?

No. The website and the mobile app can be used directly. Self-hosting the GitHub code makes sense when you want to change the pipeline, swap models or connect your own data sources, and it requires an API key and a Python environment.

Does raising the debate rounds make it better?

Not necessarily. More rounds accumulate more context, but can also amplify a disagreement and lengthen the run. The materials only say rounds are configurable and give no recommended value, so test against your own task.