J. Hao, M. Yang, Q. Shi, Y. Jiang, G. Zhang, and W. Zeng. FinFlier: Automating graphical overlays for financial visualizations with knowledge-grounding large language model, IEEE Transactions on Visualization and Computer Graphics, 31(9):6353–6369, 2025. DOI.
This Case Study Report was authored in 2026 by Philip Beaucamp, University of Oxford, UK
This case study is a post-hoc analysis of Hao et al.’s FinFlier [1], a visual analytics system that produces graphical overlays to visualise financial narratives. It addresses a workflow issue in financial analysis: textual data and charts are commonly examined side by side, increasing the cognitive burden for readers to compare values, trends, turning points, or other patterns related to textual and graphical data. FinFlier tackles this issue by combining a knowledge-grounded LLM for text-data binding with a graphical overlay module that maps the retrieved narrative to appropriate visual annotations. The method is based on a survey of 1,752 layered financial and academic charts, identifying common narrative structures, overlay methods, and correspondence patterns [1]. Hao et al. [1] utilise output constraints, chain-of-thought reasoning, and dynamic few-shot prompting to detect subjects, trend patterns, and numerical values and connect them to tabular data [1], resulting in automatic prompt generation and graphical overlays.
Applying Chen and Ebert’s method [2], the symptoms, causes, remedies, and multi-order effect analysis can be subdivided into branches, as outlined in Fig. 1. In the baseline workflow, users need to inspect financial narrative statements in one view and analyse a chart in the other, while mentally establishing a connection between them, which results in high interaction cost (Int-High-Ct) due to split attention and slow comprehension. The most likely causes are the disconnection between the text and the visualisation, as well as the insufficient structure for encoding key narrative-data patterns (Int-Low-AC, Vis-High-Ct). As a first remedy, graphical overlays are directly embedded into the chart, including highlights, labels, descriptions, and trend lines, thereby externalising reasoning and encoding narrative meaning as part of the visualisation, which naturally maps to Vis-High-AC and Vis-Low-PD. However, this leads to side effects that the original paper only partially addresses, such as a denser, more complex representation. While overlays substantially decrease the cognitive cost of identifying key patterns, they risk clutter, occlusion, and competition among elements within the visualisation (Vis-High-PD, Vis-High-Ct). These could be addressed by introducing progressive, priority-based overlay disclosure (Int-High-AC, Vis-Low-PD), but this may result in Int-High-Ct due to additional actions required to access optional annotations. Another multi-order effect is that users over-trust the overlay narrative as an authoritative interpretation of the data (Int-High-PD), even though it is only one among multiple plausible alternatives. A possible remedy could be multi-interpretation support with uncertainty and alternatives in the form of a toggle between possible annotations (Int-Low-PD, Alg-Low-AC), but this may result in Int-High-Ct or Vis-High-Ct, as multiple alternatives make the interface and analysis more complex.
Fig.1. The Workflow for Optimising Workflow (WF4OWF) for the case of Prompts4LLMs
The second symptom concerns the difficulty of producing useful graphical overlays for complex financial narratives (Alg-High-PD). Initially, there is no correspondence between financial narrative types and graphical overlay patterns, so an automated system would have to address the mapping issue arising from underspecified inputs (Stat-High-PD). The remedy is a development workflow based on a survey of 1,752 real-world charts, resulting in a formalisation of common narrative structures, overlay types, and correspondence patterns (Stat-Low-PD). This remedy leads to dependence on coverage and the survey body's bias, as well as the need for rigorous human controls, since underrepresented financial narratives might be misrepresented (Alg-High-PD, Int-High-Ct). However, it represents a clear improvement to the baseline workflow.
The third symptom is the machine learning bottleneck: Alg-High-PD due to errors in identifying relational patterns between text and data, including hallucination, unit mismatches, or incorrect subject detection. Likely causes include Stat-High-PD and Stat-High-AC, as financial language is specialised, context-dependent and can exhibit sparse labelled training data. FinFlier addresses this issue by using knowledge-grounded prompting that constrains the LLM output to verifiable structures and human controls (Int-Low-PD, Alg-Low-PD), demonstrating improved text-data binding accuracy compared to baseline LLMs [1]. However, the adverse side-effect is Int-High-Ct at the later stages of the process during required verification and manual correction. FinFlier does not simply represent automation, but a reconfiguration of earlier workflows, thereby reallocating cognitive costs from manual authoring to control and correction functions.
[1] J. Hao, M. Yang, Q. Shi, Y. Jiang, G. Zhang, and W. Zeng. FinFlier: Automating graphical overlays for financial visualizations with knowledge-grounding large language model, IEEE Transactions on Visualization and Computer Graphics, 31(9):6353–6369, 2025. DOI.
[2] M. Chen and D. S. Ebert. An ontological framework for supporting the design and evaluation of visual analytics systems. Computer Graphics Forum, 38(3):131-144, 2019. DOI.