Where do the numbers come from?
Were the levels computed from the bars by ordinary code, or written by the model? If the tool cannot say, assume the model wrote them.
Purffle / Purffle Chartwright / Guides
Can AI read a stock chart? How to tell if it is guessing.
Ask a chatbot where support is and you will get an answer — precise, confident, and sometimes a price the instrument never traded. Here is how to check any AI chart tool before you trust a number from it.
Why language models are bad with prices
A language model writes the most plausible next words. Asked about a chart it cannot see, the most plausible answer is a round, confident price — and nothing in the model checks it against the bars. Compare two answers to the same question:
| A chatbot, asked the same question | An answer with computed figures |
|---|---|
| "Support sits around 1,360 with strong resistance near 1,400. A stop below 1,350 gives you a good risk-reward here." | "Support at 1,359.95 — the swing low of 12 Aug, retested four times since, held each time. ATR(14) is 7.03, so a stop there is 2.50 ATR below price." |
The first is fluent and unattached to anything. The second can be traced: which swing, how many touches, which volatility figure.
What a published test found
One question can be settled without predicting anything: is the price it just told you a price that
exists? The Chartwright benchmark held three small local models fixed — granite4.1:3b,
granite4.2:3b and qwen3.5:4b — and changed only whether the arithmetic was done first.
The headline
34 of the 72 prices given without chart data sat outside everything the instrument traded, against 0 of 38 once the computed analysis was supplied.
What it does not show matters as much: it is a small run — three instruments (NSE:RELIANCE 15m, BINANCE:BTCUSDT 1h,
AAPL 1d), two questions each — measuring one property. It says nothing about whether any tool makes money, and it is
not a comparison against frontier models: GPT, Claude and Gemini were not tested. The full tables are
on the Chartwright page, and the harness ships with the app as
scripts/benchmark.py.
Five questions to ask any AI chart tool
Can you trace a figure back?
Pick any price it quotes and ask which bars produced it. A trustworthy tool shows its working.
Does it ever decline?
A tool that always has an answer is guessing some of the time. "Not enough evidence" is a feature.
Has it published a test?
Look for a method you could re-run — not a testimonial or an accuracy percentage with no definition.
Does it say what it is not?
Watch for trade calls, promised win rates and order buttons. A tool that describes charts should not also be selling you trades.
How Chartwright answers those questions
Purffle Chartwright is a desktop chart analyst built around one rule: the model speaks last, and only about numbers it was handed.
- The engine computes first — structure, level clusters, trend, volatility and position arithmetic in about 0.2 seconds, with no language model involved.
- The model never sees raw bars. It can only call tools, and every tool returns values the engine computed.
- Every answer keeps its receipts: the tools that ran, what each returned and how long it took, shown on demand.
- No trade calls, no broker connection. There is no order path in the application.
Vela
The limits, stated plainly
Grounding fixes the numbers, not the judgement: a model can still draw a poor conclusion from figures that are all correct. A local model is slow on a laptop CPU — minutes per answer on four cores without a GPU. Free market data can be delayed, gapped or adjusted. None of this forecasts price. The full list is in What Chartwright is not.
Risk warning. Trading carries substantial risk of loss. This guide and Chartwright are educational technical-analysis material, not investment advice or a recommendation to buy or sell anything.