Can AI Read a Trading Chart? We Tested 3 Models on 200 Charts

Paste a chart into an AI and ask where support is. You'll get an answer straight away, with a price, and it'll sound sure of itself.

But is that price on the chart? We wanted to know, so we tested it. We drew 200 charts from real candle data, which meant we knew every true price on every one. Then we asked three of Google's Gemini models to read them. That's 600 answers, each one checked against the candles.

How we tested it

We took real candles for 67 markets: 151 stock charts, 21 ETF charts, 14 crypto charts and 14 forex charts. They're spread across four timeframes: 1-hour, 4-hour, daily and weekly. Each chart has 120 candles.

We drew each one the way most traders see a chart. Dark background, green and red candles, a price scale on the right and dates along the bottom. The ticker and timeframe sit in the top corner, and the last price has a coloured tag on the scale. It looks like a TradingView screenshot.

Then we gave each screenshot to three models: Gemini 3.5 Flash-Lite, Gemini 3.8 Flash, Gemini 3.1 Pro. Pro is the biggest and the most expensive. We asked each one what a typical AI chart tool asks: the ticker, the timeframe, the last price, the chart's highest and lowest price, and the nearest support and resistance.

To grade the prices fairly across a $1 coin and a $500 stock, we measured every error in candles. One candle means the chart's average candle size over its last 14 bars. Traders call it the ATR (average true range). Missing by half a candle is a lot when your stop sits just past a level.

What AI reads perfectly: the label

Every model read every ticker and every timeframe right. Not one miss in 600. It read "EURUSD · 1H" as EURUSD on the hourly, and "BTCUSD · 1W" as bitcoin on the weekly.

The last price was nearly as good, at 99%. But that's because it's printed in a tag on the price scale. The model read the number, the same way it read the ticker.

That's the pattern. If the answer is written on the chart as text, AI reads it. If it has to work out where a candle sits on the price scale, it's estimating.

Where it starts guessing: the prices

Nothing tells you a chart's high in words. You have to find the tallest wick and read across to the price scale. The models got the high to within a tenth of a candle 31% of the time, and the low 24% of the time.

Most misses were small. But about 1 in 5 highs and lows was off by more than half a candle, and 4% by more than a whole one. If you put a stop "just below the low", that's the gap between a stop that sits behind the level and one that sits in the middle of the price action.

Support and resistance: barely better than random

This is the part that matters most, because it's what people ask AI for. A support level is only useful if price really turned there. So we checked every level the models named against the candles: did it sit within half a candle of a real turning point?

A turning point here is a swing low or swing high: the lowest low, or highest high, within three candles either side. Then we checked how often a random price would land that close to one. Charts have lots of small turns, so a random price does better than you'd think.

The models' levels hit 60% to 63% of the time. A random price hit 58%. The best model, Gemini 3.1 Pro, beat a random price by four points.

So the levels look right. They're near the candles, on the right side of price, at sensible round-ish numbers. But they aren't drawn from where price actually turned. They're drawn from roughly the right part of the picture.

One chart, three models, three different answers

Here's one chart from the test: Oracle (ORCL) on the 4-hour. Its high is easy to find by eye. It's a long wick on the second candle, up to $220.50.

No model got it. Gemini 3.1 Pro said $213.90, which is that candle's open, the top of its body. It missed the wick completely. Gemini 3.8 Flash said $225.50, above any candle on the chart. Gemini 3.5 Flash-Lite said $218.00.

Say you wanted a stop just above that high. Using Pro's answer, you'd put it at about $213.90, which is $6.60 below a price Oracle had already traded at.

The most expensive model wasn't the most accurate

You'd expect the biggest model to read best. It didn't. Gemini 3.1 Pro read the high off by more than half a candle on 28% of charts. Gemini 3.8 Flash, a cheaper model, did it on 6%.

Pro was a little better at support and resistance, but only by a couple of points. A bigger model doesn't fix the problem, because the problem isn't intelligence. It's that the price isn't in the picture as a number.

Why AI reads the label and guesses the price

To an AI model, a screenshot is a grid of pixels. Text like "ORCL · 4H" is something it's very good at reading. A candle's high isn't text. It's a coloured line that stops somewhere between two gridlines.

To turn that into a price, the model has to estimate how far up the scale the line ends. A person would do the same thing, and we'd be off too. That's why traders read prices from the data, not from the picture.

How our in-house trading algorithm does it

TradeWize's chart analyzer is built around this exact split. AI does the one job it got right 600 times out of 600. Everything else comes from the real data.

1. 1AI reads the label. When you upload a screenshot, AI reads only the ticker and the timeframe off it. Or you can type the ticker yourself.
2. 2The algorithm pulls the real candles. It fetches the price history for that market and timeframe, the same data the chart was drawn from.
3. 3The algorithm reads the chart. It finds the support and resistance from where price actually turned, along with the pattern, the trend and the breakouts. Then AI explains those finished numbers in plain English.

On the Oracle chart, that means the high reads $220.50, because it comes from the candle, not from a guess about where a wick ends. Every level you see has a price behind it that really happened.

Our algorithm on the same test

We ran our algorithm on the same 200 charts and graded its levels by the same rule. Then we graded the AI models and a random price again, on exactly the charts where our algorithm named a level.

Within half a candle of a real turn, our levels hit 76%. The AI models hit 64%, and a random price 61%.

The gap gets much bigger when you ask for precision. Within a quarter of a candle, which is where a stop actually lives, ours hit 70%. AI hit 45%, barely above a random price at 41%.

It also knows when to say nothing. Our algorithm named a support on 136 of the 200 charts and a resistance on 145. Where there wasn't a clean level, it didn't make one up. The AI models named one on every chart.

To be fair to the AI: our levels come from where price actually turned, so you'd expect them to do well here. That's the whole point. A level should come from the candles.

Our read of the same Oracle chart

Remember the Oracle chart, where three AI models gave three different highs? Here's what our algorithm reads on the same 120 candles.

It doesn't stop at the high. Its headline: "ORCL broke down from a rising wedge 2 bars ago on heavier than normal volume. Above $148.50 the break has failed." That's the pattern, the break and the price that decides it, $148.50, all read off the candles. Every line on that chart is a price that really traded.

What a Trade Setup looks like

A read tells you what a chart is doing. A Trade Setup turns it into a plan you can act on. Here's a real one our algorithm built for one of the test charts: silver (SLV) on the daily.

Silver broke above $56.75 and came back to test it. The plan waits for the next test: buy between $56.27 and $57.23, stop at $55.30, first target $61.04. That's 3 times what the trade risks. If a day closes back below $56.75 first, the plan is off.

Not every chart gets one. Of the 200 charts in the test, 102 held a plan our algorithm would stand behind. It turned the Oracle chart down: the only stop that made the numbers work sat inside an ordinary day's range, and a stop like that gets hit by noise.

How to check any AI chart tool

• Give it a chart you know. Ask for the high and the low, then compare them with the real ones. A few cents off is fine. Half a candle off isn't.
• Ask where its levels come from. If it can't point to the candles where price turned, it's reading the picture.
• Check a level against the chart. A real support level should line up with lows that held. If you can't find them, don't trust the level.
• Look for the price tag trick. Most tools get the last price right because it's printed on the scale. That tells you nothing about the rest.
• Watch for confident round numbers. A level at exactly $200.00 on a chart that turned at $198.40 is a guess with a nice shape.

What this test can't tell you

• Our screenshots were clean and sharp, 1280 by 720 pixels. Real ones are often cropped, compressed or taken on a phone, so AI's errors there are probably bigger, not smaller.
• We tested Google's Gemini models, three of them. We didn't test ChatGPT or Claude. Any model reading prices off the picture faces the same estimate, but we haven't measured them.
• "A real turning point" is our definition: a swing low or high within three candles either side, within half a candle. With a stricter half of that, AI's levels hit 42% and a random price 39%. Still barely better.
• Each chart was asked once, with each model's default settings. A different prompt or a second try could give different answers.
• The candle data is real, from Yahoo Finance. The test ran on 2026-10-20, and models change often.