AI chart analysis: what a screenshot can and cannot establish

First separate visible, inferred and unavailable

A useful chart explanation should distinguish text visibly printed on the image, an interpretation of the drawn series, and information the image does not contain. Asking for those categories makes the answer easier to review than asking for a confident market opinion.

For example, a visible timeframe label can support a statement about that label. A cropped header cannot establish the interval merely because the candles look familiar. A volume panel can show a relative shape, but a missing or unreadable scale cannot support an exact volume number.

Our existing chart pattern tool page and support and resistance tool page cover those individual tool topics. This guide addresses the evidence needed to review an image-based explanation. It supplies no current levels, entry prices, targets or predictions.

“An entity not appearing in the list should not be assumed to be authorised by the RBI.”

A chart-reading app does not establish platform authorisation

Searches for AI chart analysis also surface platform-specific apps and unofficial downloads. If a chart app leads to a trading platform, check that platform's status separately. An AI feature, a familiar-looking chart or a free download does not supply authorisation evidence.

RBI keeps an Alert List of entities that are not authorised to deal in forex or run forex trading platforms, and the list is not exhaustive. Use RBI's authorised-person and authorised-ETP registers for the authorisation question. Do not rely on an assistant to infer status from a logo or from an absent search result. For a stock broker, the matching check is SEBI's list of registered intermediaries: search the name and registration number yourself rather than trusting a number shown inside the app or in a screenshot. This is a plain-English summary, not legal advice.

Twelve cases to audit before trusting chart descriptions

Proposed audit, prepared 14 October 2026. No screenshots or model responses were collected for this grid; the last column states the expected review action.

Case	Controlled image change	Expected review action
1 · Interval	Keep the same image but remove the timeframe label.	Mark interval unknown; do not guess it from candle count.
2 · Identifier	Crop the ticker and exchange prefix.	Ask for the missing identifier; avoid naming a security.
3 · Price labels	Make the price axis too small to read.	Withhold exact values rather than manufacture precision.
4 · Volume	Remove the volume panel entirely.	State that volume confirmation cannot be assessed.
5 · Volume units	Keep bars but crop the unit and numerical axis.	Separate relative bar height from absolute quantity.
6 · Time boundary	Hide whether the last candle is complete.	Flag completion status as unknown.
7 · Scale	Compare regular and logarithmic presentations.	Read the declared scale before interpreting geometry.
8 · Line identity	Draw visually similar solid and dashed indicators.	Check the legend; acknowledge ambiguous identification.
9 · Adjustment	Compare declared adjusted and unadjusted histories.	Keep adjustment settings with the interpretation.
10 · Timezone	Crop the timezone and capture timestamp.	Do not infer the exchange session from image placement.
11 · Annotation	Add a clearly labelled author-drawn line.	Distinguish the annotation from observed market data.
12 · Missing history	Crop earlier observations while keeping the recent shape.	Limit the conclusion to the visible window.

Author-designed controls informed by vision limitations and chart-setting documentation. They are neither measured failure rates nor trading recommendations.

Sources: [1], [2]

Read the grid as a test design

The cases isolate a missing input or a visual ambiguity. They are not twelve examples of models failing. We have no response archive, model settings or independently checked screenshot dataset for a vendor comparison, so assigning pass rates would be unsupported.

To run the audit, create a baseline capture with readable labels and preserve its underlying observations. Make one controlled alteration at a time. Retain both images, use the same question, and save the complete response before supplying any hints. Otherwise a better answer may reflect a different prompt rather than a more reliable reading.

Define the answer key in advance. For a hidden timeframe the expected answer can be unknown; the test should reward recognition of missing evidence. For a readable label, compare the transcription with the source capture. Keep ambiguous cases in their own category instead of silently treating them as correct or excluding them from the denominator.

Use made-up diagnostic plots, or market data more than 30 days old, for this educational audit. Do not use a successful description as a reason to trade the depicted instrument. The test concerns faithfulness to an image, not future returns.

Preserve the chart settings with the image

TradingView documents regular, percent, indexed-to-100 and logarithmic displays, alongside dividend adjustment and back-adjustment of continuous futures when the contract changes. A picture without those settings can conceal why two captures of an apparently similar series have different shapes.

Treat the visible chart as one view of a declared dataset. Preserve the exact instrument, interval, timezone, adjustment choice and last completed observation in a note outside the image as well. If any item is unknown, say so. An assistant cannot recover a missing convention merely by describing the lines more confidently.

OpenAI's vision documentation warns that small text and differing graph colours or styles can be difficult, that descriptions can be wrong, and that original file names or metadata are not processed. Supply essential context in the prompt or visible image rather than rely on a meaningful filename. These are provider-documented limitations; they do not give us a measured error rate for any trading app.

A screenshot worth reviewing

• A readable header and both axes

  Keep the identifier and interval visible. If enlarging labels would hide part of the series, provide a second capture and explain that it belongs to the same source window.

• A declared last observation

  Record whether the final bar is complete, and preserve the capture time. A screenshot date is not necessarily the date of the observations drawn in it.

• A plain legend

  Name each indicator and annotation. Explain whose line or label it is, so a user drawing does not become independent confirmation.

• A separate value sheet where precision matters

  Use exported observations or another licensed source for exact numerical checks. Keep the image for visual context and mark differences between the sources explicitly.

Ask for observations you can disprove

A bounded prompt is: transcribe the readable labels, describe only the visible series, identify missing context, and separate observations from uncertain interpretations. Ask the assistant to leave unreadable values blank. Avoid requesting an entry, target or probability of a profitable trade.

Review each statement against the original capture. If the model names an indicator that is absent, records a timeframe whose label was cropped, or reports volume when no panel exists, flag that statement directly. Do not let a plausible paragraph compensate for an invented input.

If you save an explanation beside your own research, retain what you accepted, what you corrected and the source used for the correction. Later outcomes cannot retroactively make a fabricated observation accurate. For options-specific limitations, the AI and options guide examines payoff and valuation assumptions.

Keep the screenshot, the exact prompt and the reply together, so a later reviewer can see what the model was actually shown. Image interpretation remains a review aid with visible evidence limits. The twelve-case plan above still needs a dated response archive before anyone can describe it as a completed accuracy test.

Sources

Every figure on this page comes from one of these. Each was checked on the date shown.

1. [1]OpenAI: images and vision limitations · checked 14 Oct 2026
2. [2]TradingView: configuring chart scales, values and adjustments · checked 14 Oct 2026
3. [3]RBI: Alert List, updated 19 November 2025 · checked 14 Oct 2026
4. [4]SEBI: PR No.48/2026, 17 August 2026 · checked 14 Oct 2026