← All tools · Guides

How to choose AI stock-picking tools: four kinds, credibility, and what is legal

AI is a research assistant. It is not a machine that earns money for you.

Categories: Four kinds: an AI assistant inside a market-data app, a general language model you ask directly, AI factors on a quant platform, and an open-source research project you deploy yourself. The first two suit news and a write-up of one company. The third suits factor research. The fourth suits someone who can install software and wants the process on their own machine. All four are research assistants. You still decide the trade.

How to choose: If you only want news and a write-up of one company, use the AI assistant in a market app you already open, or ask a general model, and check every number against the company filing. If you want factor research or machine-learning stock selection, use a quant platform such as BigQuant, or run Qlib yourself. If you can install software and you want a research report whose figures you can trace, use an open-source project such as Vibe-Research or the TradingAgents family, and read the licence first. If the research is finished and you want a program to send orders, that is the QMT and PTrade choice, and you paper-trade first. This page is not the tool-order guide and not the site FAQ.

Who it is for: Someone who wants AI to sort information, look at a company from more than one side, and do research, while the buy or sell decision stays theirs. That includes a person reading news and company notes, a person doing factor research who writes a little Python, and a person who can install software and wants the whole process in their own hands. Each of those jobs maps to one of the four kinds above.

Who it is not for: Someone looking for a tool that promises a return, advertises a win rate, charges for stock tips, or pulls them into a group to copy trades. Treat that as a scam first. Charging other people for investment advice without a securities investment consulting licence is not allowed under the rule this page cites. This page does not recommend any stock. It is not the tool-order guide and not the site FAQ.

The same article in Chinese is how to choose AI stock-picking tools (Chinese). The tool order of data, then a backtest, then paper trading, then an agent is a different article, and so is the site FAQ. Neither is a language version of this page. Broker terminals and hosted research sites are also a different article: how to choose a quant trading platform.

This site has no commercial relationship with the platforms, apps, or projects named below, and it does not offer download or account-opening links. The facts were checked on 21 Oct 2026 against the CSRC text, the Securities Association of China site, vendor homepages, and the project notes. This page does not recommend any stock.

01 · Four kinds of AI stock-picking tools

02 · What AI can do, and what it cannot

• It can: read filings, table several years of one company, list reasons for and against, write a screen from your conditions, and try factors in batches.
• It cannot: promise that a forecast is right. Prices move for too many reasons. A pattern an AI finds in history often fails on the next period.
• It makes mistakes: a language model invents figures and events that look real. When it is offline, it does not know the latest quote or filing. Check every figure against the original filing.

03 · The AI assistant inside a market app

This is the easiest start: you ask inside an app you already use. Eastmoney's homepage says the Eastmoney app is connected to the Miaoxiang large model and can be used free of charge, and that there is also a Miaoxiang research assistant and a Miaoxiang app. Many brokers and quote apps have connected a language model too. Before you rely on one, check three things: where its data comes from, how recent that data is, and what the disclaimer on the page says.

This kind of assistant fits news, one company, and an industry. Its answer can still be wrong. If it prints "buy" or "sell", treat that as one opinion, not as an instruction.

04 · Asking a general model such as DeepSeek

A general model is a useful research assistant when the question is the right shape. Ask it to sort and analyse. Do not ask it which stock to buy. Useful questions look like this:

• "Table this company's annual revenue, net profit, operating cash flow, and debt-to-asset ratio for the last five years, and mark the line that changed the most."
• "List three main reasons to be bullish on this industry over the next year, and three main reasons to be bearish. Say what each reason rests on."
• "I want companies whose revenue grew for three years and whose debt ratio is below the industry average. Write that as a screen."

Watch the date of the data. A model that is not online only knows what it was trained on. DeepSeek's API docs list web search and page fetch as separate capabilities. On the website, check whether online search is turned on. Connecting a language-model API and a market-data source to your own research loop is market data and an LLM API. The same article in Chinese is connecting a data source and an LLM API. That note is not this article.

05 · AI stock selection on a quant platform

For factor research and machine-learning stock selection, use an online quant platform. BigQuant's homepage says it offers a library of 2,000+ base factors and that AI can mine more derived factors. Registration is free, and there is an enterprise edition. A result on this kind of platform is a backtest or a paper-trading result. It is not a live return. How the platforms differ is how to choose a quant trading platform. Running the machine-learning selection yourself can start with Qlib.

06 · Open-source AI research projects

If you can install software, you can deploy an open-source project: several AI roles read the data, argue, and write a research report. What they share is that they output a report and a rating, and they do not touch your live account. You connect your own language-model API and pay per call. Star counts and the latest update stay on each tool page.

• Vibe-Research: a personal research desk for China A-shares, Hong Kong, and US stocks. It writes a report in stages (company profile, financials, earnings expectations, valuation, risk). Figures in the report carry an evidence id that can be traced to the source. The portfolio feature is a ledger. It cannot send an order. MIT licence. The directory stopped listing this project on 25 Oct 2026, so this page does not link a tool URL. The description follows the 21 Oct 2026 check.
• TradingAgents-astock: a TradingAgents fork adapted to China A-shares. Several analyst roles debate, then it writes a rating and the reasons. The project note says one analysis calls the language model 30 to 50 times, so price the run before you start. Apache-2.0.
• TradingAgents-CN: a Chinese multi-agent research system, positioned as research help, with a web interface. The licence is mixed: the core engine is open source, and the application-layer source is visible but redistribution and commercial use are forbidden.
• easy-stock: a local-first China A-share research desk. Its note says it will not send orders or change a live portfolio. The licence is non-commercial.

If you want to ship a product or a service on one of these projects, read its LICENSE first. More projects, and which ones rose fastest over seven days, are in open-source AI quant projects. The same list in Chinese is open-source AI quant projects. A further kind of project can send orders on a paper account, but it often asks for the app's login credentials, such as a cookie. That hands the account over. Be careful even when the account is only a simulator.

07 · Is it credible: what the CSRC rule says

The CSRC Interim Provisions on the Securities Investment Advisory Business were issued in 2010, revised in October 2020, and are the text in force. They say that investment advice, or a similar function, delivered through software or a terminal must describe the function objectively, disclose built-in defects and the risk of use, and state the data source. A function that selects stocks or times a buy or a sell must state its method and its limits. The same provisions forbid promising or guaranteeing an investment return, in any form. This page does not cite article numbers. The current text controls.

Five questions follow from that, for any AI stock-picking product:

1. Where does the data come from, and how recent is it? If it cannot say, the conclusion cannot be checked.
2. What is the method? A language-model write-up, a factor score, or something else, and is that written down.
3. Are the limits and the risks written down? Including the kind of market where it fails.
4. Does it promise a return or advertise a win rate? If it does, drop it.
5. Can the track record be checked? A backtest alone is a weak claim. Tuning a rule on history until the curve looks good is overfitting, and the live result is often different. Run it on a paper account first. The same article in Chinese is the Chinese walkthrough.

08 · Is it legal

• You analyse with AI and you decide: legal.
• You charge other people for investment advice: under the provisions above, securities investment advisory business is carried out by securities firms and securities investment consulting firms. The person providing the service needs a securities investment consulting qualification and a registration with the Securities Association of China. A group that tips stocks, or a paid signal service, without that qualification is not legal.
• A program sends the orders: that is programmatic trading. Before it starts, it has to be reported through the broker to the exchange. See section 07 of how to choose between QMT and PTrade. The same guide in Chinese is QMT and PTrade (Chinese).

09 · Common scams, and how to check

The CSRC column on illegal securities and futures risk warnings includes a December 2025 item on AI financial traps, and other warnings about illegal stock-recommendation schemes. Stop when you see any of these:

• A promised return, capital protection, or a screenshot of a high win rate.
• A group where a "teacher" uses AI to pick stocks and lead orders, free at first, then paid.
• An app you cannot place, or money sent to a personal account or an offshore platform.
• An "AI quant robot" that earns automatically once you pay.

How to check: a licensed securities investment consulting firm can be found in the Securities Association of China's disclosure of securities investment consulting firms. A practitioner's qualification can be checked in the association's disclosure of securities practitioners. For a suspicious case, call the CSRC hotline 12386.

10 · How to choose, by situation

• News and one company: the AI assistant in the market app you already use, or a general model. Check the figures.
• Factor research or machine-learning selection: a quant platform such as BigQuant, or Qlib on your own machine.
• A research report whose figures you can trace, and you can install software: Vibe-Research or the TradingAgents family. Read the licence first.
• The research is done and you want a program to send orders: see how to choose between QMT and PTrade, and paper-trade first. The same guide in Chinese is QMT and PTrade (Chinese).

This is not investment advice, and it does not recommend any stock. An AI write-up can be wrong, and a backtest is not a future return. Features and fees follow each product's own note.