Financial candlestick charts are a staple of trading analysis, but teaching AI to read them is not a simple image-classification problem. According to a new arXiv paper, the task requires locating visual elements, aligning them across trading dates and panels, and then combining those observations with numerical and financial reasoning.

The paper's key critique is that answer accuracy alone does not reveal whether a model actually performed these steps. A correct final answer could come from a shortcut or guess, masking failures in the underlying reasoning process.

The title points to a code-interactive approach, though the abstract excerpt does not describe the method in detail. Still, the implication is clear: evaluation for such visual reasoning tasks should look beyond the final output, perhaps by inspecting intermediate reasoning or executable steps.