How to use ChatGPT or Claude to analyze a crypto chart honestly
How to ask ChatGPT or Claude about a crypto chart so it counts past cases instead of guessing: what to give it, prompts that work and checks to run.
Ask the AI to count past cases, not to predict, and give it real price bars instead of a screenshot. A screenshot has no history to count. With real bars it can tell you, for example, that Binance Bitcoin was higher 24 hours later in 51.88% of 62,335 hourly starting points, so a chart reading that just says "up" is right about half the time before it says anything clever.
This guide shows how to ask an AI app questions that give you counts and past cases instead of confident stories. These are past cases, not a forecast.
For base rates on more coins and horizons, see Crypto base rates.
Why is a screenshot a weak input for AI?
A screenshot is a picture. It shows shapes, but it often leaves out what you need to judge them:
- Which exchange and which contract (spot or perp, a futures contract that never expires).
- The bar size, such as 1 hour or 1 day.
- The exact dates, including any gaps in the data.
- Fees, and what happened on all the other, similar-looking charts.
We did not test how well AI apps read screenshots. The problem is simpler: a picture cannot tell the AI how often that shape was followed by a rise. If you ask for a story, you get a story.
What does the AI need instead?
It needs real data and a question with a number in the answer. Many AI apps can use MCP, a standard way for AI apps to call outside tools. With a data tool connected, the AI can fetch price bars and do the counting. Our guide to connecting Claude or ChatGPT to crypto data shows the settings, and ChatGPT vs Claude for crypto research compares the two apps.
Even then, ask the AI to say which numbers came from the tool and which it worked out itself.
Which questions get useful answers?
Questions that name a market and ask for counts are useful. Open questions about the future invite a made-up forecast.
| A question that invites a guess | A question that asks for the past |
|---|---|
| "Will Bitcoin go up tomorrow?" | "On Binance BTCUSDT hourly bars from [start] to [end], how often was the close higher 24 hours later? Show the count." |
| "What does this pattern mean?" | "Find past charts like this one for this coin on this exchange. Show what followed each one, including the ones that went the other way." |
| "Is this a good time to get in?" | "What were the median, 10th and 90th percentile moves over the next 24 hours, before and after a 0.05% fee per side?" |
A percentile is a rank cut-off. The 10th percentile is the move that only 1 case in 10 fell below. The 90th is the move that only 1 in 10 rose above.
What does a base rate look like?
A base rate is how often something happened without using any pattern at all. Every chart reading has to beat it. Here is Binance BTCUSDT, using every hourly close from September 2019 to October 2026 as a starting point.
| Time ahead | Cases | Price rose | Median move | 10th percentile | 90th percentile | Positive after fees |
|---|---|---|---|---|---|---|
| 12 hours | 62,347 | 51.30% | +0.03% | -2.09% | +2.28% | 47.32% |
| 24 hours | 62,335 | 51.88% | +0.07% | -3.07% | +3.49% | 49.30% |
| 48 hours | 62,311 | 52.23% | +0.15% | -4.33% | +5.06% | 50.75% |
The fee column assumes 0.05% to enter and 0.05% to exit.
Three lessons sit in this table:
- The middle is tiny, the spread is large. A median of +0.07% over a day sits next to a 90th percentile of +3.49%.
- Fees change the story. Price rose in 51.88% of 24-hour cases, but only 49.30% ended positive after fees.
- The coin and dates matter. Over 24 hours, price rose in 52.03% of Binance ETHUSDT cases but 49.66% of SOLUSDT cases.
Dates matter too. For Binance Bitcoin, the 24-hour up share was 41.39% in 2019 (a partial year, September to December), 58.03% in 2020 and 46.88% in 2022. The cases overlap, so 62,335 is not 62,335 separate tests.
What happened when we asked past charts to call the direction?
We tested one clear idea. At 3,600 chosen times, across 12 markets, we took the last 48 hourly closes and found the 10 closest earlier charts of the same coin on the same exchange. Then we took the median of what followed those 10 cases. If it was above zero, the rule said "up". The times run from January 2020 to September 2026, with 300 per market.
The pooled results:
| Time after the test | Look-alike direction correct | Always guessing up | Average rule return after fees |
|---|---|---|---|
| 12 hours | 50.44% | 51.67% | -0.05% |
| 24 hours | 50.22% | 50.92% | -0.07% |
| 48 hours | 50.25% | 51.39% | -0.08% |
The rule was right about as often as not, and lost to "always up" at every horizon. At 24 hours it beat "always up" in only 6 of the 12 markets. For Binance Bitcoin alone, it scored 50.67% against 49.33% for always guessing up, and its average trade after fees was -0.30%. Accuracy ranged from 43.67% for BNBUSDT to 54.67% for AVAXUSDT.
Most of the time the ten cases disagreed with each other. In 62.72% of the 24-hour tests, only 4 to 6 of the 10 matches went up. A single "the past says up" verdict hides that split. We look at splits more closely in when past crypto cases disagree.
For the full test, read do chart look-alikes predict crypto.
How should you ask an AI to read a chart?
Use this short routine:
- Name the market. Exchange, symbol, contract type and bar size.
- Give the dates. A start and an end, in UTC (the shared world clock).
- Ask for the base rate first. Count, share that rose, median and the 10th and 90th percentiles.
- Then ask for past cases. Ask for the closest earlier charts of the same coin, and what followed each, up and down.
- Ask for fees and the spread. A result that vanishes after fees is weak.
- Ask what the data cannot show. A good answer lists limits.
Be wary when an AI gives a price target or a chance with no count behind it. Ask it to show the cases.
What does this not show?
We did not test AI screenshot reading. The look-alike test is one rule on one setup: 48 hourly closes, 10 matches and a median. Other settings or other rules may do differently. Both tests use surviving large coins, ignore funding and trading slippage, and describe the past only.
How we measured this
For the base rates, we used hourly closing prices for Binance BTCUSDT, ETHUSDT and SOLUSDT perps. From every hourly close we measured the percent change to the close 12, 24 and 48 hours later, using only stretches with no missing hours. A move of exactly zero counted as not up. The fee version assumed 0.05% on entry and on exit.
For the look-alike test, we picked 300 evenly spaced times per market and found the 10 closest non-overlapping earlier charts from the same market. Each match had to end, and its outcome had to finish, before the test time. We scored each time as a separate long-or-short trade after fees. The tests are not fully independent.
Try this in Comparand
With Comparand connected to your AI app, ask: "Use shape_analogs to find the 10 closest past charts to Binance BTCUSDT's last 48 hourly bars. For each match, list its date and its later moves at 12, 24 and 48 hours in basis points. Count how many went up and how many went down. Do not average them."
- shape_analogs matches the last window of completed bars only against the same coin's own past on the same exchange and bar size. Matches never overlap your chart and come one per episode.
- Each match shows its own path and its own later moves. A basis point is one hundredth of one percent.
- It gives no average or hit rate. Your AI does the counting. It costs $0.01 per search, paid from credits.
To add context from Hyperliquid, ask: "Show BTC's per-minute labels on Hyperliquid between my chosen start and end times, and say what period they cover."
- hl_labels returns per-minute labels that describe the past for one coin. Each answer lists the data behind it and what the call cost.
- It is $0.005 per call on pay-as-you-go, and part of the data tools on a plan under fair-use limits.
Read the docs to connect, or try 3 coins free. Those are past cases, not a forecast.
The short version
- Ask an AI for counts and past cases, not for a forecast.
- Binance Bitcoin rose in 51.88% of 24-hour windows, but only 49.30% were positive after fees.
- Chart look-alikes picked the 24-hour direction 50.22% of the time, below the 50.92% of always guessing up.
- In 62.72% of tests, only 4 to 6 of the 10 matches went up, so they disagreed.
- Name the market, dates and bar size, and ask what the data cannot show.
Questions people ask
Can ChatGPT analyze a crypto chart screenshot?
It can describe one, but a picture leaves out the exchange, bar size, exact dates, fees and missing data. We did not test how accurate screenshot reading is. A better input is real price data pulled by a tool, plus a question that asks for counts.
What should I ask an AI instead of 'will Bitcoin go up'?
Ask how often price was higher a set time later, on a named exchange and date range, with the number of cases, the median move and the 10th and 90th percentile moves. For Binance BTCUSDT, price was higher 24 hours later in 51.88% of 62,335 cases.
How often does Bitcoin go up in 24 hours?
On hourly Binance BTCUSDT perpetual prices from September 2019 to October 2026, price was higher 24 hours later in 51.88% of 62,335 cases. After an assumed 0.05% fee on each side, 49.30% were positive. This is the past, not a forecast.
What is a base rate in crypto?
A base rate is how often something happened with no pattern used at all. For Binance BTCUSDT hourly prices from September 2019 to October 2026, price was higher 24 hours later in 51.88% of cases. Any chart reading has to beat that number to add anything.
Do similar past charts predict the next move?
Not in our test. Across 3,600 tests on 12 markets, a rule built from the 10 closest past charts picked the 24-hour direction right 50.22% of the time. Always guessing up scored 50.92% on the same dates.