AI Chart Analysis: What It Can and Cannot See

A chart is a compressed record of decisions. Every bar contains a range, a close, a volume and a relationship to the bars around it. Reading it well is not a matter of finding a magic pattern; it is a matter of observing structure, character and context in the same order every time, then writing conditional scenarios instead of predictions.

AI can help with that work. It can describe what is visible, organize the read, compare timeframes and draft scenarios. It cannot see what is not in the image, verify its own levels, or know what the market will do next.

This guide is the analysis pillar of the AI Trading Desk. It covers what AI can and cannot see, the screenshot-to-brief workflow, the four-pass chart review, evidence grading and the prompts that keep the output checkable. It is educational material, not financial advice.

A model can describe a chart. It cannot be responsible for the trade. The value is in the structure it imposes on your reading, not in the confidence of its answer.

What Chart Analysis Actually Is
What AI Can See
What AI Cannot See
The Screenshot-To-Brief Workflow
The Four-Pass Chart Review
Levels And The Line In The Sand
Multi-Timeframe Context
Evidence Grading
Prompt Patterns That Work
Common Failure Modes
Frequently Asked Questions
Related Guides
Sources

What Chart Analysis Actually Is

Most traders describe analysis as a prediction problem: “Where is price going?” That framing hides the actual work.

Chart analysis has four layers, and only the last one is a decision.

Layer	The question	Example
Observation	What is visible and measurable?	“Price is below the 60-minute EMA stack and made a lower high at 29,635.”
Interpretation	What is the best current explanation?	“The trend is down, but the last break failed and the market is two-sided.”
Scenario	What would confirm or invalidate that explanation?	“Above 29,635 on a 5-minute close, the bearish case is invalid; below 29,505, continuation is live.”
Decision	Is the opportunity worth its defined risk?	“No trade until one boundary is accepted.”

AI is useful in the first three layers. It is not qualified for the fourth, because the fourth requires accountability, risk tolerance and context that are not in the chart.

The distinction matters because most bad analysis collapses the layers. A trader sees a lower high (observation), concludes the market is weak (interpretation), assumes it will fall (prediction), and enters without a scenario for being wrong (no decision). AI can be trained by prompt to keep the layers separate. That alone improves the quality of the read.

What AI Can See

When you give a model a chart screenshot, it can work with what is visible.

Structure. Higher highs and higher lows, lower highs and lower lows, ranges, failed breaks, swing points and the position of price within the visible range.
Levels. Horizontal lines, prior highs and lows, round numbers, visible moving averages, VWAP lines, session boundaries and annotated zones.
Indicators. The values printed on the chart — RSI, stochastic, moving averages, volume, MACD — and their direction or divergence, if the panel is visible.
Candles and character. Bar size, wicks, close location, overlap, gaps and follow-through.
Annotations. Any boxes, lines or labels drawn on the chart, provided they are legible.
Time and instrument. The chart title, timeframe, contract month and timestamp, if it is in the image.

It can then organize those observations into a summary, compare them with a second screenshot, and draft scenarios in a consistent format. That is genuinely useful: it reduces the time spent formatting notes and makes the read more comparable across days.

But every one of those outputs is a description of an image, not a measurement of the market. If the image is stale, cropped, low-resolution or missing a panel, the description inherits the defect.

What AI Cannot See

The limits are more important than the capabilities.

It Cannot See Data That Is Not In The Image

A chart screenshot is a snapshot. It does not contain the order book, the true bid/ask delta, the full volume profile, the economic calendar, the news, the correlation with other markets or what happened after the screenshot. If you ask “what is the market going to do?”, the model has no more information than the image. It will still answer, because language models are designed to produce plausible text. That is the risk.

It Cannot Verify Its Own Levels

A model can read a line that is drawn on the chart, but it cannot confirm that the line is at the right price, that it is the correct session boundary, or that it has not misread a label. A level that appears in the output but not on the chart is a fabrication until proven otherwise.

It Cannot Perform Reliable Arithmetic

Models can produce arithmetic that looks right and is wrong. Position size, R multiples, P&L and risk must be calculated independently. Do not let a model restate a number you will rely on without checking it.

It Cannot Know The Context You Did Not Provide

The same chart means different things on a Fed day, a holiday, a contract roll, or a day after a large gap. If the calendar, the session and the instrument are not in the prompt, the model will assume a generic context. That assumption may be invisible in the answer.

It Cannot Carry Accountability

No model can be the named owner of a trade. It cannot be disciplined for you, cannot feel the loss, and cannot decide that the right action is no trade. The accountability remains with the human.

It Cannot Replace Screen Time

The model can describe a pattern, but it cannot build the pattern recognition that comes from watching hundreds of sessions. Use it to structure your reading, not to skip the practice.

This is the same boundary EOG applies to AI in proposals and reports: the model transforms and organizes; the human verifies and owns. See What AI Can (and Cannot) Do in Trading Analysis for the detailed task-by-task version.

The Screenshot-To-Brief Workflow

The goal is not to ask AI “what do you see?” and accept the answer. The goal is to produce a brief that a trader can check against the chart in a few minutes.

Step 1: Capture A Complete, Timestamped Chart Set

For each session, capture the same views:

Higher timeframe (for example, 60-minute) with the trend, moving averages, VWAP and session levels.
Lower timeframe (for example, 5-minute) with the same overlays plus volume.
Any panel that summarizes character, volume or structure.
The economic calendar for the day.
The account and order view, if you are doing an execution review.

Include the instrument, contract month, timeframe and time zone in the capture. A chart without a timestamp is not evidence; it is an illustration.

Step 2: Extract The Observations Before Asking For Interpretation

Prompt the model to list only what is visible: swing structure, range position, key levels, indicator values, candle character. Do not ask for a bias yet. The separation forces the model to show its work, and it makes the errors easier to spot.

Step 3: Reconcile The Observations With The Chart

Open the chart beside the output and check every level and every indicator value. Delete anything you cannot see. If the model has invented a level, stop and correct the prompt; do not build on it.

Step 4: Ask For The Three Scenarios

Once the observations are clean, ask for three branches: bullish, bearish and no-trade. Each branch needs a trigger, an invalidation and a first target. Ask the model to state what evidence would change its mind. The no-trade branch should be a real option, not an afterthought.

Step 5: Write The Final Read Yourself

The trader writes the bias, chooses the primary scenario, sets the levels and owns the plan. The model’s output is an input to that work, not a substitute for it. The final read should fit on one page and be traceable to the screenshots.

This workflow is slower than asking for a prediction. It is also the only version that produces a result you can check, review and improve.

The Four-Pass Chart Review

The four-pass review is a mechanical way to apply the workflow. It is adapted from a chart-review overlay used in a live MNQ process, but the logic is instrument-agnostic.

Pass 1 — Clean Price And Structure

Ignore the indicators. What is the swing structure? Are highs and lows rising, falling or overlapping? Where is price in the visible range — upper, middle or lower third? Has the market broken a range and accepted the break, or broken it and failed?

The most useful distinction in this pass is acceptance versus rejection. A break that holds and builds a base is different from a break that pierces a level and closes back inside. The first supports continuation; the second supports a fade. A model can describe which one it sees, but you need to confirm it on the chart.

Pass 2 — Character And Volume

Now look at the bars. Is the spread typical or wide? Is volume typical or unusual? Where does each bar close relative to its range? Is there overlap, or clean separation? Are tails frequent? Is there follow-through after a break?

This pass decides the environment. Overlap and frequent tails mean two-sided trade: demand cleaner triggers and smaller size. Clean separation and follow-through mean trend: pullback entries work better than breakouts. A model can summarize the character, but it cannot hear the market; use the summary as a second opinion, not a substitute for your own read.

Pass 3 — Context

Zoom out. Is the higher timeframe above or below its mean? Is it inside a range or outside it? Is the trend accelerating or flattening? Does the lower-timeframe signal agree or conflict?

The context pass is the binding constraint. A lower-timeframe long inside a higher-timeframe downtrend is a counter-trend trade; it can work, but it needs a smaller size and a nearer target. A model can compare the two timeframes, but it cannot decide which one should govern the trade. That is a risk decision.

Pass 4 — Conditional Scenarios

Write three branches:

Bullish: if price accepts above X and holds, the bullish case is live. Trigger, entry, stop, target, invalidation.
Bearish: if price breaks below Y and fails to reclaim, the bearish case is live. Trigger, entry, stop, target, invalidation.
Wait: if neither boundary is accepted, do nothing. This is the correct answer more often than most traders admit.

The fourth pass is where AI is most useful as a drafting tool, because it can generate the structure quickly and check for missing invalidation. It is also where the trader must be most careful: a scenario that sounds coherent can still be untradeable. The plan is not the prose; it is the level, the trigger and the risk.

See The Four-Pass Chart Review for the full checklist.

Levels And The Line In The Sand

A chart has hundreds of possible lines. A plan needs a handful.

Use a fixed vocabulary so the read is comparable across days:

Prior day high / low / close.
Overnight high / low.
Prior session high / low.
Initial balance high / low (for intraday sessions).
VWAP (anchored and session).
Moving averages (the ones you actually use, not all of them).
Swing highs and lows from the visible range.
Volume shelves (high-volume nodes, value area boundaries).
Round numbers that the market has respected.

For each level, record the source and whether it is a zone or a line. A prior session low from one indicator may differ from the same label on another timeframe; print both and treat the gap as a zone rather than choosing one and hiding the other.

Then choose one line in the sand: the level that defines the day. Above it, one scenario is live; below it, another. The line should be a place where the market has made a decision, not a level chosen because it is halfway between two others.

The line in the sand is the most useful output of the analysis, because it converts a bias into a testable condition. “I am bearish” cannot be reviewed. “Below 29,505 on a 5-minute close, I look for continuation; above it, I stand aside” can.

Multi-Timeframe Context

A single timeframe is a partial view. The higher timeframe tells you where you are; the lower timeframe tells you when. AI can help compare them, but the comparison needs a rule.

A simple rule set:

Higher timeframe trend + lower timeframe pullback = the highest-quality continuation setup.
Higher timeframe range + lower timeframe breakout = lower quality; the breakout is more likely to fail at the range edge.
Higher timeframe trend + lower timeframe counter-trend signal = a counter-trend trade; half size, nearer target, and a clear invalidation.
Both timeframes in a range = no trend trade; wait for a boundary or stand aside.

The model can summarize the two timeframes and flag the conflict. It cannot decide whether the conflict is acceptable for your risk tolerance. That is your call.

Evidence Grading

Not every statement in a chart read deserves the same weight. Grade the evidence before you act on it.

Grade	What it is	Example	How to use it
A — Measured	A number read directly from the chart or data	“5-minute close at 29,505.75”	Can be used in a plan
B — Observed	A visible pattern or relationship	“Lower highs since the open”	Can support a scenario
C — Interpreted	An explanation of the observation	“Sellers are in control”	Useful, but not a fact
D — Predicted	A forecast	“Price will bounce here”	Treat as a hypothesis, not a plan
X — Unverified	A level or claim you cannot find on the chart	“Support at 29,480” (not visible)	Discard

AI output often blends these grades seamlessly. “Price is in a downtrend and will likely bounce at support” mixes B, C and D in one sentence. Your job is to separate them. A plan built only on A and B is robust; a plan built on C and D is a story.

Prompt Patterns That Work

The prompt is not the system. But a good prompt makes the system easier to follow. These patterns are designed to produce checkable output.

Pattern 1: Observations Only

“You are a chart-reading assistant. Do not predict. List only what is visible in this image: instrument, timeframe, timestamp, swing structure, position in the visible range, the five nearest levels with their labels, and the indicator values printed on the chart. If a value is not visible, write ‘not visible’. Do not infer.”

Pattern 2: Three Scenarios With Invalidation

“Using only the observations above, write three scenarios: bullish, bearish and no-trade. For each, give the trigger, the invalidation level, the first target and the evidence that would change your mind. Do not assign a probability unless I ask. Do not invent a level that is not in the observations.”

Pattern 3: Challenge The Read

“Here is my current read and plan. Argue the opposite case. Identify the three assumptions the plan depends on and the evidence that would falsify each one. Do not rewrite the plan; list the weaknesses.”

Pattern 4: Reconcile Two Timeframes

“Compare the higher-timeframe and lower-timeframe reads. Where do they agree, where do they conflict, and which one should govern a trade taken on the lower timeframe? State the conflict explicitly; do not average it away.”

Pattern 5: Check The Plan For Omissions

“Review this one-page plan. Flag any scenario without an invalidation, any stop that sits inside the last swing, any target nearer than the stop, any risk number that exceeds the daily limit, and any rule that is not testable. Return a list of required changes, not a rewrite.”

The prompts share three properties: they constrain the output, they require traceability, and they leave the decision with the human. That is the opposite of “tell me what to trade.”

Common Failure Modes

The stale screenshot. A chart captured hours ago is not the current market. The model will describe it confidently. Always check the timestamp and the last bar.
The invented level. A model produces a precise price that is not on the chart. It looks authoritative. Discard it.
The indicator soup. Ten indicators produce ten conflicting signals. Use a small, fixed set; the model cannot resolve what you have not defined.
The narrative trap. A compelling story (“the market is distributing”) is not evidence. Ask for the observation behind it.
The confirmation loop. If you show the model a chart and say “I think this is a short,” it will find reasons to agree. Ask it to argue the other side.
The arithmetic error. A model calculates position size or R multiple incorrectly. Recalculate independently.
The false precision. A scenario probability of 63% is not more useful than “more likely than not.” Do not let the model dress up uncertainty.
The missing no-trade branch. A plan with only bullish and bearish scenarios is incomplete. The no-trade branch is often the correct one.

Frequently Asked Questions

Can AI Read A Chart Screenshot Accurately?

It can describe visible structure, levels and indicator values with reasonable accuracy if the image is clear and complete. It cannot verify its own output, see data outside the image, or perform reliable arithmetic. Treat every level and number as unverified until you check it.

Which Chart Timeframes Should I Give The Model?

At least one higher timeframe and one lower timeframe, plus a panel that shows character or volume if you use one. The comparison is more useful than a single image. Always include the instrument and timestamp.

Can AI Identify Support And Resistance?

It can identify lines that are drawn or visibly respected in the image. It cannot know which levels are institutionally significant, how they were formed, or whether they will hold. Use it to organize the levels you have already marked, not to replace the marking.

Can AI Tell Me The Trend?

It can describe the swing structure and the position of price relative to moving averages. Whether that constitutes a tradeable trend depends on your timeframe, your rules and the context. The model’s label is a description, not a signal.

Can AI Analyze Volume Profile Or Order Flow?

Only from the data you provide. A screenshot of a volume profile can be described; true order flow requires trade-level bid/ask data that a screenshot does not contain. Do not let a model infer delta from candles and present it as fact.

Is It Safe To Upload My Charts?

Charts can contain account information, order lines and position markers. Crop or redact account numbers, balances and order quantities before uploading. Use an approved business-tier tool and follow your data rules. If in doubt, redraw the chart.

How Do I Stop The Model From Inventing Levels?

Constrain the prompt: “Use only the levels visible in the image; if a level is not visible, write ‘not visible’.” Then check the output against the chart. No prompt eliminates the risk; verification does.

What Is The Best AI Model For Chart Analysis?

The model matters less than the workflow. A model with strong image reading and instruction-following is useful; a model with a confident narrative is dangerous. Test candidates on the same chart and see which one follows the “observations only” constraint. Do not choose on brand.

Sources

US Commodity Futures Trading Commission, Customer Advisory: AI Won’t Turn Trading Bots into Money Machines (25 January 2024). AI cannot predict the future; guaranteed returns are a fraud red flag.
CME Group, Micro E-mini Nasdaq-100 (MNQ) contract specifications, used for the worked-example vocabulary. Verify current specifications with the exchange.
NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), for the principle that AI output requires human review in high-stakes decisions.
Barber, B. M., and Odean, T. (2000), “Trading Is Hazardous to Your Wealth,” Journal of Finance, 55(2), 773–806. Active trading carries a performance penalty after costs.
Chague, F., De-Losso, R., and Giovannetti, B. (2020), “Day Trading for a Living?” SSRN 3423101. Most persistent day traders lost money; no evidence of learning by trading.

No statistic in this article is invented. Verify figures against the primary source before republication.

Trading involves substantial risk of loss. This article is educational information, not financial, investment, legal or tax advice, and it is not a personal recommendation or solicitation. AI output can be inaccurate or incomplete; every level and figure must be verified on the chart and in the broker platform. Past performance and backtests do not guarantee future results.