AI chart analysis explained

Two very different things get called "AI chart analysis"

The label covers two architectures that behave nothing alike.

Model-reads-the-chart. A screenshot or a raw price series is handed to a language or vision model, which is asked what it sees. Everything in the output — levels, percentages, volume figures — is generated by the model. This is fast to build and produces confident, fluent text. It is also where fabricated numbers come from: a model asked for a support level will supply one whether or not the data contains it.

Engine-computes, model-writes. A deterministic program calculates the values first — the profile, the delta, the levels, the session statistics — and the model is given those numbers and asked only to phrase them. The model never invents a figure because it is never asked to produce one.

The second is meaningfully harder to build and is the only version where the numbers in the paragraph can be traced back to the chart.

What an AI can actually read from a chart

With the underlying data available, the readable material is descriptive and largely mechanical:

• Where volume concentrated — the point of control, the value area, the thin bands price moved through quickly.
• Which side was aggressive — delta per candle and cumulatively across the session.
• Whether aggression was met — heavy one-sided volume that did not move price.
• Session structure — opening range, prior-session levels, where the current price sits relative to both.
• How today compares to this instrument's own recent sessions — the only comparison that means anything, since absolute values differ wildly between instruments.

All of that is description. It is a statement about what has already traded.

What it cannot tell you

These limits are structural, not a matter of a better model:

• It cannot tell you what price will do next. Nothing in the data contains the future. Any system presenting a forecast as an output is presenting a guess in the grammar of a fact.
• It cannot identify participants. Exchange tick data is anonymous; size and repetition can suggest a large order being worked, but no feed names anyone.
• It cannot see orders that were never placed, or resting size away from the touch.
• It cannot replace a decision. A description of a session is an input to a trader's process, not a substitute for it.

Under Indian regulation this distinction is not stylistic. A platform that analyses and describes market data is a tool; making specific recommendations to buy or sell securities is a regulated activity requiring SEBI registration as an investment adviser or research analyst. Factual analysis and personalised advice are different products with different legal footing.

How to tell whether an AI analysis is grounded

Four checks separate a grounded analysis from generated text, and all of them can be run in a few seconds:

1. Can you find every number on the chart? Each figure cited should correspond to something visible in a toolbar, an indicator or a profile. A number with no on-screen counterpart came from the model.
2. Does it change when the data changes? Reload on a later candle. Grounded output moves with the data; generated text tends to restate the same shape regardless.
3. Does it decline to answer sometimes? A system reading real data will occasionally have nothing notable to report. Output that always finds something significant is producing narrative, not analysis.
4. Does it speak in comparisons or absolutes? "Delta is in the top decile of this stock's own sessions" is a measurement. "Delta is very high" is an impression.

How VolumeLens builds it

VolumeLens uses the second architecture. A deterministic analysis engine reads the order flow — footprint, delta, volume profile, big trades — and computes the session's own distributions. Detectors work in percentiles of that instrument's own session rather than fixed thresholds, so a reading means the same thing on an index future and on a mid-cap.

Only after the engine has produced its numbers is a language model involved, and only to write them up. Output passes a validator that checks every figure against the computed values, with a deterministic template as the fallback when validation fails. The analysis is also generated once per symbol, timeframe and candle and shared by everyone viewing it — so two traders looking at the same chart see the same reading, which would not be true if the text were generated per request.

Analysis can be delivered as text or spoken aloud, in English, Hindi, Hinglish and four other Indian languages. Prices, levels and indicator names always stay in English regardless of the language selected.

Frequently asked questions

Keep learning

• What is order flow trading?
• Volume profile explained
• Price action trading, and how order flow differs
• All learn articles