Why You Can't Just Act on AI Trading Advice: Four Examples Checked Against Binance's Own Pages
Read it, don't act on it. The trading advice a language model gives you is a reorganization of the enormous quantity of text it has read, not a judgment made about your account. This isn't a question of how well one product has been tuned; it falls straight out of three structural limits that models of this kind have: the knowledge has a cut-off date, it doesn't know your positions, and it will describe things that don't exist in convincing detail. In a plain chat with no tools attached, not one of the three can be worked around.
First, draw a line: it explains well, it decides badly
In one sentence: anything of the "what does this mean" kind you can hand to it; anything of the "what should I do right now" kind is not something you can lean on.
Explaining is work like turning an English API error into plain language, making "mark price", "maintenance margin ratio" and "slippage" clear and then giving a counter-example, or restating a strategy you have only half thought through. Questions like that have settled answers, written down long ago across countless documents and textbooks; the model has seen them many times over and restates them fairly reliably.
Deciding is work like "can I add to this position now", "is this line about to break out", "how much have I actually lost on this trade". It answers those three as fluently as it answers the others — and that is exactly the problem. Fluency has nothing to do with correctness. It does not change its tone of voice because it doesn't know.
The steadier way to use it is to let it ask the questions and leave the call to you: have it list ten things worth confirming, then confirm them yourself one at a time.
Limit one: the knowledge stops at some point in time, but it won't say "I don't know"
Every model's training data has a cut-off, and it doesn't know what happened afterwards. That in itself isn't a defect; any source of knowledge has a shelf life. The problem is that it doesn't fall silent because of it — it takes information from before the cut-off and answers questions about after the cut-off in exactly the same confident tone.
The facts that change in trading are exactly of that kind: where a feature is reached, fee rates, regional availability, trading pairs. The fourth example below is a live case. The platform fee on Binance stock trades is at a promotional price that ends on 31 December 2026, after which the correct answer to the same question becomes a different number. In the old "snapshot of the world" the model is holding, the number stays where it was on the day of training.
Cut-offs differ from model to model, and vendors update them, so don't go memorizing a particular date — anything you memorize goes out of date soon enough. What you want to build instead is an unconditional habit: whenever it is telling you "here is what the rule is now", treat that as information that may already have expired.
Web search relieves part of this, but relief isn't a cure. What it retrieves may be a third-party article that is equally out of date, and once that has been paraphrased you can't even see where a given sentence came from any more. To confirm something properly you still have to go back to the Help Center or the fee page of the platform you are actually using.
Limit two: it doesn't know what you are holding
"You could consider adding a little here" is two opposite pieces of advice depending on whether it is said to someone holding nothing or to someone already fully committed with leverage on. The same sentence: acted on by one person it opens a position, acted on by the other it pushes them towards liquidation.
The model has none of your account context. It doesn't know your cost basis, how much margin you have left, how long you can stand to be down on this money, or where next month's rent is coming from. What it gives you is a general formulation aimed at the "average reader", and in trading there is no such thing as an average reader.
Even if you type your position data in, what it receives is a static snapshot. The market has moved on by the time you finish typing, and it won't know. Never mind the dimension of what you can absorb — "a 30% drawdown is within the normal range" may be statistically unobjectionable, but to somebody who put their living costs in, that sentence means nothing, and it has no way of telling which of the two you are.
· If anybody else asking the same question would get the same answer, then it isn't advice for you.
· An answer genuinely aimed at you has to have your position, your cost basis and the loss you can absorb in it as variables. The model has none of those variables, so it cannot produce an answer of that kind.
· You, however, do have them. So the correct division of labour is: let it lay out the options and what each one costs, and you put your own numbers into it.
Limit three: the more specific the answer, the more likely it was invented
The dangerous error isn't it saying "I don't know", it is it inventing on the spot — and inventing in great detail. It takes three shapes:
| What gets invented | What it looks like | Why you believe it |
|---|---|---|
| Menu paths | "Open the app → Trade → More → Strategy Center" | The hierarchy looks real and so does the wording; it is only the final entry point that doesn't exist |
| Parameter and field names | An API field that looks perfectly standard | The naming style is identical to the real documentation, and nobody questions a field that "just looks right" |
| Numbers | A fee, a cap or a threshold given to several decimal places | Precision is itself a hint that "this was looked up"; if it were vague, it wouldn't dare say it |
Keep in mind that precise and correct are different things. A number given to three decimal places looks as if it was looked up; to the model, it is just the slot where a number belonged, filled with something that looks like a number.
It isn't lying to you. It is continuing along the probabilities of language: a sentence has reached the point where a field name belongs, so it generates a string that looks like a field name. Which makes the standard a blunt one — anything that could be falsified by looking at a specific page (steps, field names, numbers) should not be used directly.
Four examples: which sentence on the official page holds the answer
Each of the four questions below has a definite answer, and each answer is written on one of Binance's own pages.
Example 1: what happens to a spot grid when the price falls below the lower limit
The answer is in the Help Center page "What Is Spot Grid Trading and How Does It Work?" It says: "The system will no longer execute orders when the market price falls below the Lower Price." The page's own example adds that above the upper price or below the lower price the strategy is suspended, and that it resumes when price returns to the range.
The drift to watch for is "suspended" turning into "automatically stopped out" or "the grid ends". Those are two other things the page covers elsewhere. The stop-loss is a separate setting, and the grid only stops working when the market price reaches it. Whether coins are sold when a grid stops depends on the Sell All Base Coins on Stop switch; with it off, the base coins stay in your Spot Account. So a grid with no stop-loss that falls through its lower limit still holds the coins it already bought, and they keep falling with the price.
To check a reply, find the Lower Price, Stop Loss and Sell All Base Coins on Stop passages on that page and see whether the reply has folded three things into one. The full rules are in Binance Grid Trading: Spot Grid Rules, Settings, a Worked Example and Why Grids Lose.
Example 2: how often funding is settled on a perpetual contract
The answer is in "Introduction to Binance Futures Funding Rates". The page first says funding on perpetual contracts is settled every 8 hours, at 00:00, 08:00 and 16:00 (UTC). Further down there is an exception: since 2 May 2025, when a USDⓈ-M perpetual's funding rate hits its cap or floor at a settlement, settlement moves to once an hour, whether the contract was on an 8-hour or a 4-hour interval before. Once on the hourly schedule, it goes back to every 4 hours from the 17th cycle, after the rate's absolute value has stayed at or below 0.025% for 16 consecutive cycles.
The drift here is answering only the first half, "a fixed eight hours". That sentence can be found on the page, but it is not the whole rule, and whoever is answering does not know which contract you hold or which interval applies to it right now. The page names the place to check: the real-time funding rate and funding rate history pages show the latest funding interval and the cap or floor.
Two more rules on the same page are useful for checking a reply. The funding amount is the nominal value of the position times the funding rate, so one settlement on a 10,000 USDT nominal position at 0.01% is 10,000 × 0.01% = 1 USDT. And if you close the position before the funding time, you neither pay nor receive funding for that period.
Example 3: after a 429, can you keep sending with a different API key
The answer is in the limits section of Binance's Spot REST API documentation: "The limits on the API are based on the IPs, not the API keys." The same page says that on a 429 you are expected to back off, that failing to do so leads to an automated IP ban returned as HTTP 418, that bans run from 2 minutes to 3 days for repeat offenders, and that a Retry-After header comes with a 418 or 429 response giving the number of seconds to wait.
There are two ways a reply drifts. One is offering generic fixes such as switching keys or adding concurrency; by the sentence above, that solves nothing on the Binance spot API, and continuing to send turns a 429 into a 418. The other is quoting a specific per-minute limit. For how much weight has been used at this moment, the docs point to the X-MBX-USED-WEIGHT header carried on every response.
To check, have your program print the Retry-After value and the used-weight header from the 429 response, and wait for as long as the response says.
Example 4: what a US-stock trade on Binance costs
Right now this question has two answers, each with an official source. The Help Center's stock trading page gives the standard price: a platform fee of 0.35 USD per order at 350 USD and below, and a 0.10% spread above that. The stock fee page in October 2026 showed the promotional price: a platform fee of 0.17 USD per order at 340 USD and below, and a 0.05% spread above, until 2026-12-31 23:59 (UTC).
Nothing has to be invented for a reply to be wrong here. The issue is timing: a reply gives whichever number sits on the page and date it drew from. Both numbers can be found on Binance's domain, and only the fee page reflects what is charged today. When the promotion ends, today's correct 0.17 USD becomes the stale answer.
To check, treat the fee page as the only source for any rate that involves a promotion, and read the end date. The other fee lines and worked examples are in Binance US-Stock Fees.
Two lists: ask and don't ask
Translate the three limits above into everyday use and you get the two lists below: one for what you can hand it, one for what you shouldn't.
Fine to hand over, as long as you look over the result:
- Explaining a concept. Have it make a term clear, then follow up with "give me an example where this concept doesn't apply". A working counter-example means it has caught the edges of the concept rather than reciting a definition.
- Translating error messages. Paste an error response in and ask what the line is saying and what usually causes it. It does this quickly and well, because the meaning of an error code is relatively stable and doesn't change when an interface is redesigned.
- Generating a checklist. Describe the operation you are about to perform and ask for "ten things to confirm before starting". You won't have thought of everything on its list, and anything it missed you will spot at a glance — the two complement each other neatly.
- Restating your thinking to find the holes. Talk it through your strategy logic and have it restate it, setting out each implicit assumption. Where it restates things wrongly, either you hadn't thought that part through or it has misread you; both are worth stopping to check. This is the use of it I find most underrated.
- Reading code and sketching a skeleton. Have it explain what somebody else's order-placing logic is doing and which edge cases have been left out. If it is really going to run, you still have to read it line by line yourself.
Not to hand over to it:
- Giving levels and calling direction. In a plain chat it usually has no real-time data; hook it up to web or market-data tools and the numbers it pulls still need checking for timestamp and price definition. And even with live data, it has no reason at all to know better than you where the next candle goes.
- Working out your actual profit and loss. It doesn't know your fee tier, your real fill price, or whether the order filled partially. The figure it produces looks tidy and doesn't reconcile with your account.
- Confirming how a feature is reached right now. See limit one. The right destination for questions like that is the platform's own Help Center.
- Judging the rules, the fees or regional availability as they stand today. Same as above — and being wrong about these costs real money.
Two small checking habits
- Paths and numbers go back to the official page for a check, without exception. If it gave you steps, open the Help Center of the platform you are going to use and walk through them as described; if it gave you a fee or a cap, turn to the fee page and compare. Note that it has to be the one you are actually using — entry points and parameters for features with the same name often differ between platforms, and taking one platform's description over to another leaves you tapping at nothing.
- Have it separate out the parts it isn't sure about. Add a line to your question: "list separately any parts you are unsure about or are inferring." This doesn't guarantee honesty, but whatever it puts on that list is exactly what you need to check first.
There is a more fundamental layer worth a sentence of its own: even if every rule it tells you is right, the price it is looking at may not be the same number as the one you are looking at. Where the data comes from, which definition of price is being used, and how long ago it was taken — those three differences turn "the market at one moment" into several versions of itself. That is covered in The Market an AI Tool Sees Isn't the One You See.
And if you intend to take a step further and wire the model's output into a program that runs by itself, there is a whole runtime layer to get through: disconnects, rate limits, parameters quietly going stale, none of which anybody will warn you about. See Trading Bot Not Placing Orders or Throwing Errors?
Tools that feed in web, market or account data can cover part of the first two limits, the expired knowledge and the missing account context, and whatever they bring in still has to be checked for time and definition. The third, describing paths, fields and numbers that don't exist in convincing detail, does not go away when a tool is attached.