F. Yan et al. Visual analysis approach for mutual fund selection, Frontiers in Computer Science, 19(5):#195604, 2025. DOI.
F. Yan et al. FundSelector: A visual analysis system for mutual fund selection, Visual Informatics, 9(4):#100258, 2025. DOI.
This Case Study Report was authored in 2026 by Philip Beaucamp, University of Oxford, UK
This case study is based on the progression from Yan et al. [1] to Yan et al. [2] as a post-hoc assessment of a visual analytics design influenced by mixed-initiative machine learning. The earlier research presents the core workflow: mutual fund data are used to compute seven technical indicators covering profitability, risk resistance, and price performance, followed by training a pairwise preference-based classifier on investors' annotations of misordered funds. The latter research extends this workflow to FundSelector, a visual analytics system that combines the ranking model with six coordinated views to specify investment preferences, market context, fund rankings, indicator overviews, manager views, and adjustable temporal fund comparisons. Thus, the latter work resolves multiple symptoms partially addressed by the former system, including explainability, contextual understanding, and granular comparison of mutual funds.
The process under the lens of the ontological framework [3] is outlined in Fig 1. The baseline workflow that investors encounter is characterised by high interaction costs, as investors navigate a sizable candidate space to compare the intricacies of different mutual funds, where the trade-offs are complex. Both papers [1] and [2] highlight the difficulty of mutual fund selection due to the large number of funds and the heterogeneity among investor preferences, whereas the latter research further puts forth that existing platforms provide many different attributes or basic return-based rankings without enabling investors to translate their investment strategy into practical selection criteria. This can be abstracted into the first symptom (Int-High-Ct), as the cost of investigating, comparing, recalling, and deciding is simply too high. A credible cause is insufficient statistical and algorithmic compression: the platforms offer too many alternatives or poor filtering and summarisation tools, reflected in the causes (Alg-Low-AC, Stat-Low-AC). A response to this is the first remedy highlighted in the earlier system: increasing algorithmic compression (Alg-High-AC) utilising a mixed-initiative ranking model. By converting the selection issue into pairwise comparisons and injecting human input via learned preference weights, the classifier narrows the initial candidate space to a more manageable ranked list of funds. Following multiple rounds of refinement, the quantitative MAP and NDCG assessments demonstrate that the learned model can be useful for supporting the preference selection process, thereby substantially reducing cognitive costs.
Fig. 1. The Workflow for Optimising Workflow (WF4OWF) of the case of VA + ML for mutual fund selection.
However, the first remedy creates a side effect that receives limited attention in both papers but becomes apparent through abstract reasoning: once ranking is delegated to a machine learning model, investors may often blindly trust the ranking or exhibit suspicion without the means for further investigation. This results in a second symptom (Alg-High-PD), materialising as imperfect model performance and uncertainty about recommendation quality. The cause of this symptom may stem from insufficient training data and user feedback compared to the size of the total candidate space, as well as temporal changes in the information space (Stat-High-PD). The performance of mutual funds is strongly dependent on changing market conditions, sector rotations, and changes in fund managers, so the information space itself shifts significantly over an extended period. While the weights of a trained model may be reasonable for one period, they may become less reliable over time. This explains why even a sensible machine learning algorithm remedy may still lead to significant residual distortion in deployment.
Subsequently, the 2025 FundSelector workflow can be viewed as a second major remedy for the residual distortion and explainability issues (Alg-High-PD, Int-High-Ct). The core contribution is not only additional visualisation but also access to original, more contextual information via the coordinated views, thereby reducing the excessive compression of the algorithm (Alg-High-AC). The investment preference view presents and validates the current preference weights, the Market Overview provides temporal stock, bond and sector market context, the Fund List provides context to ranking outcomes, the Fund Indicator view allows for multidimensional comparison, the Fund Manager view provides further investment style insights, and the Fund Comparison view provides temporal mutual fund view details via elastic trend charts. Furthermore, the rank-informed bipartite contribution bar chart provides insights into the positive and negative effects of the ranking process, thereby increasing the model's transparency. The latter 2025 work addresses Alg-High-AC with Vis-Low-AC, as more reasoning is inspectable by the investor, less information remains concealed within the classifier. The new remedies may cause further side effects, including high interaction and learning costs, particularly for ordinary inexperienced investors (Int-High-Ct), in the form of interface complexity, navigation difficulty, and higher response times, which could be addressed in turn by progressive, personalised information disclosure (Vis-High-AC, Int-High-AC). Furthermore, this confirms the framework’s central claim that workflow optimisation is iterative.
However, compared to previous iterations, the improved system is more efficient, specifically because it does not replace complex human judgment with machine learning alone, but instead supports decision-making through the strengths of algorithmic preference learning and human-driven analysis utilising visualisation. At a higher level of abstraction, this case study illustrates an issue between machine learning development and deployment workflows. During the development process, model accuracy is improved through relabeling, retraining, validation, and testing, whereas in deployment, the model has to operate on evolving data, relying on the investor’s soft knowledge to detect outdated weights and discrepancies, i.e., to monitor the changing information space.
[1] F. Yan et al. Visual analysis approach for mutual fund selection, Frontiers in Computer Science, 19(5):#195604, 2025. DOI.
[2] F. Yan et al. FundSelector: A visual analysis system for mutual fund selection, Visual Informatics, 9(4):#100258, 2025. DOI.
[3] 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.