A screenshot of a candlestick chart feels like a natural input for artificial intelligence. Human traders look at charts, identify structure, and describe momentum, support, resistance, or a possible breakout. Modern vision-language models can inspect the same image and produce a fluent explanation within seconds. Yet a market chart is only a rendering of underlying data, and the rendering process discards information while adding visual conventions. This article explains when chart images are useful, when structured OHLCV data is superior, and how to design a fair experiment instead of mistaking a persuasive visual narrative for a forecast.
What a chart image actually contains
A candlestick encodes open, high, low, and close for a selected interval. Color usually indicates direction, a lower panel may display volume, and optional overlays show transformations such as moving averages or Bollinger Bands. Once rendered, those values become pixels. Exact prices may be recoverable from axis labels, but precision depends on resolution, compression, scaling, crop, theme, and typography. The image also hides information that was never drawn: individual trades, bid and ask quotes, corporate actions, missing bars, timezone conventions, and metadata about data quality.
The same underlying series can create radically different pictures. Change the vertical scale from linear to logarithmic, resize the window, alter candle width, or choose a different lookback and the perceived pattern changes. A vision model may respond to those presentation choices even though the economic history is identical. This is the first reason to avoid treating screenshots as neutral data. They are a view created by a charting configuration.
Why vision models still have value
Images are not useless. They are appropriate when the phenomenon of interest is genuinely visual or when the system must communicate with a human workflow built around charts. A vision model can describe annotated research charts, detect whether an analyst has drawn a channel, compare layouts from different platforms, or retrieve visually similar historical configurations. It may also provide a convenient interface for discretionary traders who want a second opinion on a chart already open on screen.
Vision-language models are particularly helpful for chart understanding rather than pure price prediction. They can identify visible elements, translate a chart into a structured hypothesis, point out conflicting indicators, and ask which timeframe or scale is in use. Used this way, the model is an interface and reasoning assistant. It is not expected to infer hidden data or manufacture precision that the screenshot does not contain.
The advantage of structured OHLCV
Structured data preserves exact numerical values. It allows returns, ranges, gaps, volatility estimators, volume changes, and corporate-action adjustments to be computed consistently. A model can receive long histories without converting them into enormous images. Researchers can normalize inputs, flag missing observations, add timestamps, and reproduce every transformation. The same pipeline can run across thousands of instruments without worrying about font size or pixel artifacts.
OHLCV is also easier to audit. If a suspicious prediction occurs, the exact input rows can be retrieved and compared with the vendor source. With images, a team must preserve the rendering code, viewport, theme, scale, overlays, and compression settings to reproduce the input. That is possible, but it creates additional failure modes without necessarily adding information.
Information density and context length
A 1920-pixel-wide image may display a few hundred candles clearly. A numerical model can often process a much longer sequence and can summarize multiple frequencies directly. Conversely, an image can compact several panels—price, volume, momentum, and annotations—into one view that is easy for a multimodal model to discuss. The difference is not simply more or less information. It is information optimized for computation versus information optimized for human perception.
For forecasting, context must match the horizon. A five-minute prediction may benefit from order-book and trade features that a chart cannot show. A monthly allocation decision may depend on macroeconomic variables, valuations, and cross-asset relationships. A screenshot of one ticker silently encourages a univariate technical-analysis frame even when the real decision is multivariate.
A fair chart-versus-data experiment
To compare modalities, begin with one immutable market dataset. At each timestamp, create two inputs from precisely the same history. The image branch receives a deterministic chart generated with fixed dimensions, scale, colors, and overlays. The numerical branch receives the corresponding OHLCV rows and declared transformations. Neither branch may access later data. The target, horizon, asset universe, and evaluation dates must be identical.
The prediction format should also be identical. Asking the vision model for a paragraph and the numerical model for a return estimate produces an unfair comparison. Both should output, for example, probabilities of negative, neutral, and positive return buckets; or a median return plus lower and upper quantiles. A strict parser should reject invalid outputs rather than allowing a human to reinterpret them after the fact.
| Criterion | Chart image | Structured OHLCV |
|---|---|---|
| Numerical precision | Limited by rendering | Exact stored values |
| Human annotations | Naturally visible | Must be encoded separately |
| Long context | Visually compressed | Efficient sequence representation |
| Reproducibility | Requires rendering controls | Straightforward data versioning |
| Cross-asset inputs | Can become cluttered | Naturally multivariate |
| Explainable interface | Strong for humans | Needs visualization layer |
Beware of visual leakage
Chart datasets can leak the answer in surprising ways. A platform watermark or timestamp may identify the period. Automatically scaled axes can reveal the magnitude of unseen context. Indicators may be centered or calculated with future values. Images collected from published analyses may include arrows, labels, filenames, or surrounding text that states what happened next. Even the folder structure used during training can correlate with the target.
Preventing leakage requires generating charts from raw, timestamped data inside the evaluation pipeline. Crop every image consistently, strip identifying text where appropriate, compute indicators causally, and separate train and test periods chronologically. Duplicate and near-duplicate images should be detected using perceptual hashes. A model that memorizes a familiar historical chart is not demonstrating general market reasoning.
From chart to chart: a better product design
The most useful “chart-to-chart” system does not need to hallucinate future candles. It can accept a chart or ticker context, recover the observable structure, request the underlying data, run numerical forecasting models, and return scenarios overlaid on a clean chart. The output should show a central path only if appropriate, probability bands, forecast origin, horizon, and assumptions. Alternative regimes can be displayed as separate scenarios rather than one falsely precise continuation.
This architecture separates responsibilities. The vision layer understands the user’s view and annotations. The data layer retrieves exact prices and metadata. A forecasting model estimates distributions. A risk layer checks scale, missing data, and plausibility. Finally, a visualization layer communicates the result. The user still gets the intuitive experience of giving the system a chart and receiving a chart, but the analytical core is not limited to pixels.
What should be measured
Evaluation should include accuracy, calibration, stability, and usefulness. Directional hit rate alone can be misleading when one class dominates. Measure log loss or Brier score for directional probabilities, quantile loss for intervals, and coverage for uncertainty bands. Test sensitivity by rendering the same data with different harmless themes and dimensions. If predictions change materially when candle colors are swapped, the image model has learned a fragile visual shortcut.
A human-centered test is also valuable. Give analysts either the original chart, the model explanation, or the model’s scenario chart and measure whether decisions become more consistent—not merely more confident. A fluent model can amplify conviction without improving outcomes. Record confidence before and after assistance, along with the eventual result and the time required.
Recommended division of labor
Use structured time series for the forecasting core whenever raw data is available. Use chart images for interaction, annotation, retrieval, quality inspection, and explanation. Combine both when the user’s markup or a visually encoded feature is itself part of the question. This division captures the strengths of each modality and reduces the temptation to claim that a model can see hidden market intent in colored rectangles.
For implementation, general multimodal models can support the interface, while dedicated time-series foundation models such as Chronos, TimesFM, or Moirai handle numerical forecasting. The exact combination should be selected through a controlled benchmark, not by comparing attractive demonstrations.
Conclusion
A chart screenshot is a powerful communication object but an imperfect scientific record. Vision models can understand its visible grammar and make chart-based workflows more accessible. Structured OHLCV remains the stronger default for precision, scale, auditability, and causal evaluation. The best system joins the two: pixels express what the human sees, data preserves what the market recorded, and probabilistic forecasts communicate what remains uncertain. That design is more useful—and more honest—than drawing a confident set of future candles from an isolated screenshot.