← Back to home

AI Stock Trading — Where Does the Data Actually Come From? From Zero Yuan to 39,000 a Year, I Tested Every Route

Long-Form Video · EP0070 July 27, 2026 07:05
What this episode covers

Quite a few people are using AI to trade stocks and do investment analysis, and the more technical ones are doing quant work. But for this kind of job, how smart the model is comes second — the information source is the thing your life depends on. Whatever price AI is holding, it won't tell you what time it's from, and it won't raise an error either. I tested a free MCP on Moutai and got the previous day's close. On the same Hong Kong interface, two letters in the symbol were the difference between a live price and one 22 minutes old.

This episode works through every route for feeding trading data to AI. How the three-layer architecture (channel / method / query) divides the work, why the three free sources can cross-verify each other, what each rung of the price ladder from zero yuan to 39,000 a year actually buys you, and three rules that make data trustworthy. The strategy is calculated by humans and AI only speeds up execution — this episode is technical only and is not investment advice. The complete comparison table is linked at the end for download.

Companion downloads · Feed them to your Claude Code

Two premises first

Strategy is something humans work out. What AI does is speed up execution — helping you find data faster, watch the tape, run the process. It is not going to conjure up a strategy that makes money for you. So this episode is not investment advice. No tickers, no timing calls. Technical only: how AI actually gets data, and what it takes for that data to be trustworthy.

Information sources split into two halves, trading data and sentiment. Sentiment runs deeper and I won’t get into it here. Trading data only.

Three-layer architecture: pick the wrong layer and it’s all wasted

There are several completely different routes for feeding stock market data to AI. A lot of people assume getting data just means finding an interface. It really isn’t like that — there are three layers here.

The channel layer: a command-line interface or a direct API, giving you tick-level to second-level quotes, and taking up almost none of AI’s attention. The public quote interfaces from Tencent, Sina, and East Money all live at this layer.

The method layer: all the various Skills packages — technical analysis, Chan Theory, the Greeks. A skills package doesn’t generate data itself. How fresh it is depends entirely on which channel it’s wired to. Wire it to a second-level channel and it can compute second-level things; wire it to next-day data and what it computes is next-day.

The query layer: the various MCPs. Financials, shareholders, and macro data are its strong suit, but in testing most free MCPs give you 15-minute-delayed or even next-day data. What it’s good at is breadth, not real time.

The free route: one command, and it can even cross-verify itself

Second-level A-share data, zero cost, one command to connect, the barrier is absurdly low. In testing, three sources hit the same stock (Moutai) simultaneously, and all three prices came back not a cent apart — free stuff that can cross-verify itself, which was the biggest surprise of this whole investigation.

The Hong Kong trap is the one worth remembering. Same domain: add two letters in front of the symbol and you get the real-time price (tested intraday against the clock on the wall, same second). Leave those two letters off and you get a price from 22 minutes ago, and it raises no error. Ask a free MCP about Moutai and it gives you the previous day’s close, no error raised, no sense that anything is wrong — this “no error but stale data” pattern is more dangerous than an outright error.

There’s also a hidden cost. Tushare’s official MCP has over 200 tools and Longbridge’s has over 140, all of them resident in AI’s context. The more you install, the easier it is for AI to pick the wrong tool, and the faster tokens burn — the “tool tax” I covered last episode.

The price ladder: what each rung buys

  • Zero yuan: second-level A-shares plus cross-verification across three free sources, enough for watching prices intraday
  • 19 to 29 dollars a month: buys history and breadth (for backtesting), still on 15-minute-delayed data
  • 99 dollars a month: the minimum for true real-time across all US exchanges
  • 199 dollars a month: tick-level data plus 20 years of history plus options
  • Tens of thousands of yuan a year: a terminal like Wind (39,800 a year for a single seat, per media reporting from a few years back) — what you’re buying isn’t just data, it’s licensing, stability, and a compliant basis you can cite externally

Brokers are another route. Longbridge shipped an official command-line tool, explicitly for AI use — tick-level, with order book and broker queue, able to place orders directly, and exchange-licensed. Futu takes the same approach through a local gateway. The institutional route (Anthropic’s financial edition wired into FactSet, S&P Global, LSEG, Morningstar) doesn’t publish pricing. Just know it exists.

Three rules: having data doesn’t mean you can trust it

One, every number must come back carrying its own timestamp. AI is not allowed to just say “latest price” — you want the hour, minute, and second.

Two, hit important numbers from two or three sources at once. If they agree, proceed. If they don’t, stop.

Three, plenty of interfaces don’t raise an error when you exceed your quota — they hand you a null or a stale value, which is more dangerous than an error. Use the timestamp to verify freshness yourself.

On May 6, 2010, US stocks fell nearly 1,000 points in nine minutes, and there’s a detail in the regulatory report: NYSE quotes reaching the national market system were delayed by as much as five minutes, yet the time stamped on those quotes read “current.” (The report’s conclusion was that this delay wasn’t the primary cause of that day’s crash, but it confirmed the phenomenon really did occur.) If even an official system can have timestamps that don’t mean what they say, how can you not verify data you wire up yourself.

FinanceBench (a 2023 benchmark) found that with the same model, feeding it an entire raw financial report versus having it retrieve on its own produced results that differed by a factor of four. The model didn’t change. The only thing that changed was how it got its information.

This methodology isn’t just about stocks. Exchange rates, oil prices, competitor pricing, inventory — for any scenario where you feed real-time data to AI, the thinking is the same. Map out the layers of your data sources, pick the route that fits the use case, then set the rules for how the data gets used.

The commercial red line

Free public interfaces have no exchange licensing and no service commitment, and one redesign at the source site breaks them — only suitable for your own research and prototypes. To put something in a product or publish externally, go through a broker or a commercial data source, and settle the license first.

Don't ask AI what the price is right now. Ask it what time your number is from, and where it came from.