AI Chart Analysis: How AI Reads Trading Charts, and How Far to Trust It
AI chart analysis means giving a chart, as an image or as price data, to an AI that describes trends, levels and patterns. It is quick at labelling what is on the chart, but it cannot see the future in it, and image readings can misplace exact prices. Test any AI chart reader on your own charts before you act on it.
How AI chart analysis works
There are two ways to hand a chart to a model. One is an image: a screenshot with candles, lines, and labels. The other is the prices themselves, as numbers. Image reading is what people mean when they drop a picture into a chat. Number reading is what you want when the exact price matters. The image is fast to label. The numbers are what an order would actually use.
Three vendor docs, each read November 22, 2026, describe sending an image to the model. They are the OpenAI images and vision guide, the Anthropic vision guide, and the Gemini image understanding guide.
A label can be useful and still be the wrong price. A model can call a region "support" because the picture looks flat there, while the wick you care about is a few percent away. If you trade the label without reading the axis, you traded a caption. Captions are cheap. Fills are not.
What a chart reader cannot see
It cannot see the future in the chart. No AI can predict markets. Past candles are a record, not a script. The CFTC customer advisory SI Won't Turn Trading Bots into Money Machines says SI technology can't predict the future or sudden market changes. A neat pattern name does not amend that sentence. Can AI predict the stock market is the longer version of why a tidy hit rate still may not pay.
It also cannot see your fees, your size, or whether the token is a trap, unless you paste those facts. A breakout caption on a chart with fake activity is a caption about a lie. AI trading signals is the page for turning a picture into a rule you could test. A signal you cannot write down is a mood. Moods do not belong in the order ticket.
Worked example: how often a coin flip looks skilled
Assumed numbers for the math, not a quote and not a price call. Example, not a recommendation. Suppose a reader calls direction on ten separate charts and gets 6 of 10 correct. That score, counting 6 of 10 correct or better, happens by a coin flip about 37.7% of the time. It is not rare. It is a normal week for a coin. Ten charts is a story you can screenshot. It is not a career.
Stretch the same fair coin to 100 calls. 60 of 100, or better, happens by chance about 2.8% of the time. That is rarer. It is still not a trading result, because a call is not a fill. Round-trip cost is about 1.5% to 3.6% at small sizes: about 1.5% when 0.75% applies each way, and about 3.6% when a $50 ticket pays the $0.90 minimum twice. A coin that is slightly right, and a fee that is always due, is a leak. The risk reward calculator is where you put the gross win and the gross loss before you admire the hit rate.
Image error stacks on top. If the model misreads the axis, the "correct" call was about the wrong level, and the order you send is not the order the caption described. Test the reader on charts where you already know the prices. If it cannot repeat the high and the low you can see, do not let it name a pattern. Swaps and agent market trades pay 0.75% with a $0.90 minimum per swap. Limit, take-profit, stop-loss and trailing orders pay a flat 0.75% with no minimum. Onchain fees apply too. A 0.75% swap fee (minimum $0.90 per swap), plus onchain fees. Fees on your agent's trades earn up to 10% cashback in USDC. It is not a discount on the swap fee.
Numbers you can put next to the picture
Through the MCP server, your agent can pull any window up to 48 hours, or a 1 hour, 4 hour and 24 hour summary of whale buying and selling. Moby's MCP server includes a whale trade window tool and a smart money pulse tool. Those tools are not the image model. They are a way for your agent to pull whale buying and selling, either as a window up to 48 hours or as a 1 hour, 4 hour, and 24 hour summary, and to use a smart money pulse beside that. The chart caption does not contain that flow. If you want it, ask for it as data, then read it. Smart money tracker is the background page. A caption that says "whales are buying" without the window is a decoration.
Keep the two outputs in separate lines of your note. Line one: what the image said, including any price it claimed. Line two: the prices you read off the axis yourself. Line three: the whale window, if you pulled one. If line one and line two disagree, throw away line one. Do not average them. An average of a misread and a real price is still a misread.
How to test a chart reader on your own charts
Pick ten charts from your own history, including ones that did nothing. Hide the later candles if you can. Ask for a label: trend, level, and what would invalidate it. Then reveal the later candles and score the label yourself. Do not let the same model score its own homework. A self-score is a compliment, not a test.
Write the misses down in prices, not in adjectives. "It was early" is an adjective. "It marked 2.10 and the wick was 1.94" is a price. After ten charts, compare the hit count with the coin-flip line above before you compare it with your hopes. Then put fees in. A 6 of 10 that would have paid 3.6% each round trip on a small ticket is not a green week. Backtesting with AI and the backtest demo are for when the label becomes a rule. Until then it is a caption.
Prompts that keep the test honest live on AI trading prompts. Use review a trade after a fill, and ask it to compare the caption with the fill, not to defend the caption. AI trading bot is the page if you are tempted to wire the caption straight to an order. Do not wire it until the ten-chart test is boring and the prices match.
How to use a label if you still want one
Use it as a checklist, not as a trigger. The checklist can say: where is the level, what invalidates it, what is the size, what is the fee. You answer those from numbers. The model may draft the list of questions. Turn on Ask before every trade and your AI asks you first. Moby does not hold or approve trades. You set your agent's per-trade and daily spending limits and market-cap range; Moby blocks red-rated tokens. Limits are only as tight as you set them. Sells and exits skip the market-cap and safety checks, and tokens with no safety rating are not blocked. A red-rated token does not become tradable because the candles look tidy. A token with no safety rating is not blocked. The chart will not tell you that. The guardrail settings will, and only if you set them.
One agent per user, with its own wallet. Agents trade Solana tokens only. The agent cannot withdraw, transfer or send funds. AI Trading and copy trading currently need an access pass in the app. Download Moby if you want the app. Bring screenshots if you like. Bring the axis too. The order ticket needs a price, a size, and a slippage you chose. It does not need a pattern name. If you cannot point to the price without the caption, you are not ready to send anything.
What to save from a chart session
Save the image, the model's words, and your own reading of the high and the low. A week later the image is the only way to see whether the caption was even about that picture. Memory will upgrade the caption. The file will not. If you traded, add the fill price and the fee in dollars. The distance between the caption's level and the fill is the number that tells you whether image reading is close enough for your size.
If that distance is larger than the move you hoped to catch, stop using images for entries. Use them to notice that a chart exists, then switch to prices. Nothing in the vendor docs cited above says an image reading is an order. They describe sending an image to the model. The rest of the distance is yours to measure.
Frequently asked questions
Can AI read a trading chart?
Yes, as a labeler. You can send an image or send prices. Three vendor docs read on November 22, 2026 describe sending an image to the model. The label can name a trend or a level. It can also misplace the price. Read the axis yourself before you treat the caption as a level.
Can an AI chart reader see the future?
No. Past candles are a record, not a script, and no AI can predict markets. A pattern name does not say what the next candle will be. Test labels on your own charts, including charts that went nowhere. A tool that only sounds right on winners has not been tested.
How often does a coin flip look like a good reader?
6 of 10 correct, or better, happens by a coin flip about 37.7% of the time. 60 of 100, or better, happens by chance about 2.8% of the time. A short streak is easy to screenshot. It is not a trading result until fees are in, and small round trips cost about 1.5% to 3.6%.
Should I trust an image for an exact price?
Not by itself. Image readings can misplace exact prices even when the trend label sounds right. If the stated high or low does not match the axis you see, ignore the level. Use prices for the order. Keep the image as a note, not as the ticket.
Can an agent see whale flow next to a chart?
Through the MCP server, your agent can pull any window up to 48 hours, or a 1 hour, 4 hour and 24 hour summary of whale buying and selling. Moby's MCP server includes a whale trade window tool and a smart money pulse tool. That is separate data. A caption that mentions whales without the window is decoration.
How do I test an AI chart reader?
Use ten of your own charts, hide later candles if you can, and score the labels yourself. Do not let the model grade itself. Compare a 6 of 10 with the coin-flip rate before you call it skill. Then subtract a round trip of about 1.5% to 3.6% at small sizes.
What should I do with a chart label before I trade?
Turn it into questions: level, invalidation, size, and fee. Answer them from numbers you read. Turn on Ask before every trade and your AI asks you first. Moby does not hold or approve trades. A tidy pattern is not a reason to skip the cap you set.