Does a chart image add information beyond its numbers?

A chart image can improve a trading agent only if it helps interpret a fixed information set or contributes additional relevant information. Those mechanisms need different tests. We propose a matched-input study that reconstructs an image and numeric snapshot from the same saved observations, with results still unmeasured.

The continuous record companion shows why a rendered surface deserves experimental treatment: historical candidate visibility influenced selection. A chart introduces another surface through axis ranges, color, cropping and annotations. Studying it as an information channel requires controlling those rendering decisions as carefully as the underlying prices.

Build the pair from one object

An illustrative case contains twelve hourly closes, volume for each hour and one declared moving-average formula. The numeric arm receives those values with units and observation times. The image arm receives a chart generated from precisely those values, using a fixed size and visible axes. A combined arm receives both. A fourth arm receives the numeric snapshot plus an unrelated chart with the same layout, clearly identified as a representation control.

For a small arithmetic check, take closes of 100, 102 and 101. Their simple average is 101. A displayed line at 102 would add false evidence through rendering even though the numeric packet is correct. Before inference, compare plotted coordinates with the source table and verify that the chart contains no later candle. A screenshot of a live interface could include a forming bar, a future annotation or a separate data feed and would fail this matching contract.

Ask every arm the same forecast question and the same action question. A representation benefit would appear as better forecasts or fewer factual reading errors at equal information. A richer-data benefit would require a separate experiment where the extra information is explicitly enumerated. The first study should keep model version, inference budget and action policy fixed.

Read the disagreements

Record mistaken value extraction, unsupported pattern claims, forecast scores and policy-valid actions. If the image arm says a price rose from 100 to 102 when it actually closed at 101, the failure is traceable before a trade outcome arrives. If forecasts agree while actions differ, inspect how the model interpreted risk constraints rather than assigning the change to visual predictive skill.

The controls paper companion supplies the mandate-to-action trace discipline for that inspection. Our proposed extension adds a rendering manifest and a matched numeric reconstruction. Evaluation should include several axis scales and accessible color choices as registered perturbations, with the canonical image frozen before the main comparison. A gain confined to one attractive chart style would justify a narrow representation claim. A durable gain across those checks would support further testing of multimodal input, still within the saved cases and execution assumptions used.