Multiviz: a smart visualization system for real-time financial trading operations

Multivariate visualization in financial contexts

Today’s financial sector is one of the most data intensive domains, characterized by the real-time generation of large volumes of multidimensional, interconnected, and volatile data. Analysts and operators are routinely required to monitor variables such as credit default swap (CDS) spreads, market depth, bond yields, sector performance indicators, order book activity, volatility surfaces, macroeconomic announcements, and geopolitical risk signals. These tasks usually occur in parallel and are often performed under time pressure (Keim, 2002). Furthermore, the heavy cognitive workload remains the same throughout the workday (Sarlin & Peltonen, 2013). The big challenge is not only in accessing the data but also in turning it into actionable insights during rapidly changing scenarios such as sudden market changes and policy shifts (Sackett et al., 2006). Recent approaches also move beyond isolated indicators toward integrated multi signal architectures for anomaly detection. For example, Shu, Wang & Liang (2024) proposed a multi-signal integration framework that synthesizes market microstructure metrics, sentiment measures, and cross asset signals, achieving earlier detection of financial anomalies across diverse regimes. This line of work highlights the growing emphasis on integrated signal processing, which parallels Multiviz’s goal of combining heterogeneous indicators into a unified perceptual display.

Traditional financial information systems have relied heavily on tabular displays and chart-based dashboards, sometimes seasoned with colors and additional markers. Table based monitoring lists or Bloomberg Terminals™ are widely used in the financial sector due to their familiarity and flexibility to users. However, they still impose considerable cognitive demands on users, especially in high dimensional settings where relevant variables are not spatially grouped or perceptually pre-attentive (Guo et al., 2022). Users are required to perform manual filtering, sorting, and cross-comparison across multiple views, leading to high levels of visual search effort, memory load, and error susceptibility.

The limitations of such approaches have been well documented especially in high dimensional settings (Marghescu, 2007). Cleveland & McGill (1984) demonstrated that common graphical encodings vary significantly in their perceptual accuracy, with position and length-based encodings outperforming area and color under most circumstances. Although bar charts and line plots work well for isolated comparisons, they are less effective in conveying hierarchical relationships, attribute clusters, or anomaly patterns in large data sets. In the finance domain, where datasets usually exhibit hierarchical and multiscale properties such as “sector → sub-sector → firm”, the need for more expressive and compact visualization methods becomes apparent.

In response to these difficulties, financial analytics platforms have started to incorporate more sophisticated visualizations. Heatmaps and dashboard tiles are frequently used to encode variable magnitudes through color gradients. Some platforms offer simple tree navigation or filterable panels to support hierarchical analysis. However, these systems rarely optimize for perceptual salience or cognitive efficiency, especially when addressing multi-dimensional data such as environment, social and governance (ESG)-based investment factors (Khoo et al., 2025; Huang, Liang & Nguyen, 2009). As noted in Vliegen, van Wijk & van der Linden (2006), many real-time monitoring dashboards are not suitable to support exploratory sense-making under stress but rather for routine surveillance of static key performance indicators (KPIs). Such tools often struggle with scalability issues. As the number of variables or assets increases, traditional visualizations become cluttered and difficult to work with, and interactions such as sorting or filtering become time consuming and cognitively disruptive.

Several academic and commercial efforts have explored techniques to enhance visual performance in finance. For example, Kong, Heer & Agrawala (2010) investigated multi-criteria ranking tools for financial instruments by combining sliders and scatterplots for portfolio optimization. Medeiros et al. (2023) and Limberger et al. (2023) proposed event-driven visual interfaces that adapt their focus based on detected anomalies. Yet these systems generally stop short of integrating hierarchical layouts with real-time feedback mechanisms or perceptual guidance layers.

One underaddressed challenge is the alignment between data structure and user task. For example, when identifying underperforming firms within a sector, it is not sufficient to sort a column in a table. The user must simultaneously evaluate firm level indicators, their sector wide context, and how the anomaly deviates from temporal baselines. All of these benefit from visualization techniques that support visual grouping, comparison, and pattern detection at multiple levels of abstraction. Existing tools that rely on linear representations or disconnected panels fail to provide this integrative visual context.

A recent systematic review further highlights the broadening application of visualization in finance. Du et al. (2025) surveyed visualization approaches aimed at improving financial literacy and decision making, emphasizing how visual design principles can support not only expert analysis but also public understanding of financial risk and investment. This underscores the expanding role of visualization in both professional and educational financial contexts, complementing operational systems such as Multiviz with evidence from the domain of financial literacy.

Given this landscape, it becomes clear that an effective financial visualization solution must:

• (1)

  support real-time updates and responsiveness,

• (2)

  scale to high dimensional data while preserving interpretability,

• (3)

  allow rapid anomaly detection and comparison, and

• (4)

  align visual structure with domain specific decision tasks.

The design of Multiviz addresses these needs through employing a hierarchical enhanced treemap based structure which is coupled with visual agents guiding attention. It takes the cognitive burden away from manual data scanning toward perception driven insight. This transition is becoming increasingly necessary as financial operations move toward automation assisted human oversight.

Treemaps and hierarchical visual encodings

Treemaps are space filling visualization techniques designed to represent hierarchical structures using nested rectangles, where each rectangle encodes a data item and its size and color represent one or more quantitative attributes. First introduced by Shneiderman (1992), treemaps have proven useful in domains with dense, multilevel data requiring efficient screen utilization, such as file systems, system monitoring, and stock market summaries.

Initial treemap layouts, such as the “slice and dice” method, produced long, thin rectangles that hindered interpretability and comparison. This limitation prompted the development of squarified treemaps, which improved readability by ensuring more balanced aspect ratios (Bruls, Huizing & van Wijk, 2000). Other innovations followed, including strip, pivot, and ordered treemaps, each balancing competing design goals: maintaining item order, improving rectangle shape, or stabilizing layouts during updates (Elmqvist & Fekete, 2010; Van Wijk & Van de Wetering, 1999; Scheibel et al., 2020). Building on these foundational refinements, recent work has increasingly explored how treemap variants can address the challenges of temporal dynamics, anomaly detection, and cognitive optimization in high frequency environments.

Recent advances have also investigated hierarchical visualization of temporal data streams. For example, Meng et al. (2025) introduced ChronoDeck, a visual analytics framework designed for hierarchical time series analysis, which enables scalable exploration of multilevel temporal structures while preserving hierarchical context. This work reflects a broader trend of enhancing hierarchical visual encodings with temporal awareness, reinforcing the relevance of treemap based approaches for financial data characterized by both hierarchy and rapid change.

Despite these advances, the adoption of treemaps in operational financial systems has historically remained limited. Early implementations, such as SmartMoney’s Map of the Market, demonstrated the appeal of using treemaps to visualize sector wise stock performance with color coded returns (Wattenberg, 1999). However, these systems lacked interactivity, had delayed updates, and were largely disconnected from users’ decision workflows. Other attempts to use treemaps for portfolio composition or credit scoring faced similar issues, namely, insufficient real-time responsiveness, lack of perceptual guidance, and poor integration with user tasks.

Recent research has emphasized the value of dynamic treemap adaptations that move beyond static layouts. Animated heat treemaps were introduced to visualize evolving data access hotspots in distributed microservices architectures, demonstrating how treemaps can be extended to capture temporal dynamics and activity patterns (De Rycke et al., 2025). This illustrates the continuing relevance of treemap variants for monitoring time evolving, high frequency environments, reinforcing their suitability for financial contexts where anomalies and trends unfold rapidly.

One critical limitation of traditional treemaps is their semantic neutrality. They represent all items visually, but without embedded intelligence to indicate relevance or urgency. In high stakes domains such as financial trading, users require more than compact structure. They need interfaces that guide attention, prioritize anomalies, and help them parse and assimilate complexity. This has led to the emergence of hybrid systems that combine treemaps with task aware agent models or enhanced visual analytics mechanisms.

For example, treemap variants are proposed that adapt based on semantic constraints or dynamic reordering strategies (Vernier, Burch & Weiskopf, 2020). Another study integrated financial indicators into interactive layouts with user-adjustable filters (Carvalho, de Oliveira & Campos, 2016). However, these systems usually remain limited to static or semi dynamic contexts, lacking robust real time architecture and fine grained attentional cues (Balzer, Deussen & Lewerentz, 2005; Limberger et al., 2023). Additional recent work has compared treemaps with alternative hierarchical visualizations in terms of efficiency and usability. A recent work evaluated circle packing, treemaps, and network graphs for ontology relationship tasks, using both task accuracy and usability measures (Maukar, 2025). The results showed that treemaps offered the fastest completion times and highest usability scores, while circle packing achieved higher accuracy but suffered from lower usability. These findings reinforce the value of treemaps for contexts requiring efficiency and scalability, while also highlighting the importance of perceptual and interaction design trade-offs.

In parallel, treemap literacy was investigated through a constructivist classroom intervention, comparing a Treemap only interface with a coordinated Node-Link + Treemap design (Firat et al., 2025). Their findings indicated that students who directly engaged with treemaps developed stronger interpretive skills than those using the combined view, suggesting that treemaps themselves can provide sufficient scaffolding for comprehension when designed carefully. This reinforces the argument that treemaps, when perceptually optimized, can serve as both effective analytical tools and pedagogical instruments, aligning with Multiviz’s emphasis on self explanatory, cognitively efficient treemap designs.

Taken together, this body of research shows a trajectory from foundational treemap designs toward increasingly sophisticated adaptations that integrate temporal dynamics, semantic guidance, anomaly detection, and usability considerations. Multiviz extends this trajectory by adopting a squarified treemap layout optimized for real-time performance, embedding multi variable glyphs, and integrating a rule based agent layer that enhances perceptual salience. In doing so, it combines visual scalability, responsiveness, and cognitive optimization in ways not addressed by prior systems, particularly in the context of high frequency financial trading.

In addition to the basic treemap rendering, Multiviz incorporates a smart agent layer that enhances treemap regions based on context sensitive triggers. Developed agents monitor live data streams. When detecting relevant events such as anomalies, threshold breaches, or volatility spikes, they modify the visual appearance of the corresponding rectangles using perceptual cues (e.g., brightness modulation, edge emphasis). This changes the treemap from a passive structure into an interactive attention guiding system supporting rapid orientation and decision making.

By addressing the perceptual challenges of traditional treemaps alongside the operational demands of real time financial environments, Multiviz contributes a hybrid visualization model combining layout stability, semantic awareness, and user-centered responsiveness.

Cognitive optimization and attention-guiding interfaces

Effective visual interfaces must do more than simply present data. They must align with the user’s perceptual and cognitive capabilities. In high-stakes domains such as financial trading, where decision speed and accuracy are critical, most users are often overwhelmed by the volume and volatility of information. Traditional tabular or chart based layouts provide flexibility and the ability to drill down further, but they require high visual effort and manual synthesis. This leads to suboptimal decisions due to pressure where time is critical.

The field of cognitive optimization in visualization focuses on minimizing this mental burden by designing displays that support preattentive processing, reduce visual search time, and lower working memory load (Card, Mackinlay & Shneiderman, 1999). As early as the 1980s, Julesz (1981) demonstrated that certain visual features, such as color, orientation, and spatial proximity, are processed by the human visual system with minimal effort, enabling rapid detection of outliers and groupings. More recently, Franconeri et al. (2021) emphasized the importance of leveraging these perceptual channels in complex multivariate displays.

Other contemporary efforts have also introduced novel visualization metaphors designed specifically for anomaly signaling and early warning. Wu et al. (2025) proposed the Bubble-Wall Plot, a dynamic analytical processing visualization tool that serves as a visual warning system for anomaly detection tasks in complex data environments. This work underscores the importance of perceptually salient, easily interpretable cues in financial early warning systems, aligning closely with Multiviz’s use of rule based agents and visual emphasis to draw user attention to emergent anomalies.

Interfaces following such principles try to make important patterns “pop out” automatically to the user, eliminating the need for exhaustive scanning. Tufte (2001) indicates it is possible to transform dense numerical data into intuitive visual structures and improve decision clarity.

For example, using consistent spatial layouts, color to indicate magnitude or directionality, and size to represent importance enables users to perform visual inference. Inferences such as drawing conclusions directly from perceptual patterns instead of through manual calculation or cross referencing become possible. This approach is particularly valuable in financial contexts where users should track dozens of critical variables across instruments, sectors, and time windows simultaneously.

Conventional table or spreadsheet based tools are limited in this regard. They rely on serial data inspection, demand short term memorization, and impose frequent context switches. In contrast, cognitively optimized interfaces structure information spatially and semantically, allowing users to visually prioritize, compare, and filter without shifting mental context.

This brings the concept of attention guiding interfaces to mind described as a class of systems that use visual cues and interaction design to direct the user’s focus toward the most relevant information. These cues can be passive (e.g., color saturation, border emphasis) or active (e.g., animation, dynamic reordering). While such systems do not completely replace user control, they reduce the burden of decision critical navigation. It is anticipated that the current status of artificial intelligence tools may further enhance attention guiding capabilities, which is left as a future study.

This principle is related to the broader idea of mixed-initiative interfaces, originally proposed in the human–computer interaction literature, but also has important differences (Horvitz, 1999). In full fledged mixed initiative systems, the machine models user intent, intervenes adaptively, and engages in a collaborative workflow. Examples include intelligent assistants, dialogue systems, or predictive recommendations.

Multiviz adopts a lightweight form of this principle. Rather than learning or predicting user behavior, it employs rule based visual agents that monitor live financial data streams. These agents are configured with domain specific thresholds and conditions (e.g., a firm’s CDS spread increasing while its interest coverage ratio drops below a critical value). When such patterns are detected, the agent layer triggers non intrusive visual cues, such as brightness highlighting, border thickening, or gentle animation, within the treemap interface. This allows the system to direct user attention toward emergent anomalies without interrupting the user’s flow or removing agency.

This mechanism transforms the interface into a form of cognitive scaffold, a visualization adapting in real time to guide sense-making under complexity. The goal is not to completely replace user decision making and automate. The actual goal is to amplify perception and reduce unnecessary visual work. In this way, Multiviz integrates the cognitive optimization principles of visual structure, stability, and preattentive encoding with a task aware, agent enhanced display system.

Multiviz embeds these capabilities directly into the layout engine instead of relying on alert panels or external notifications. It ensures that the most critical data becomes visually dominant within the user’s natural field of view, reducing the need for serial scanning or manual filtering. This design choice is critical in trading environments, where reaction time and clarity are critically important and where attention is a scarce cognitive resource.

Eye-tracking for visual interface evaluation

Eye-tracking has been in use as a valuable tool in the evaluation of visual interfaces. It offers insights into how users interact with complex data displays. By recording gaze patterns, fixation points, and saccadic movements, researchers can infer how attention is distributed across the interface and which elements users engage with most. While eye-tracking is often used in cognitive psychology to estimate workload or stress, its application in interface evaluation is generally focused on assessing visual efficiency, layout effectiveness, and usability.

In the context of financial systems, eye-tracking has received limited but growing attention. Studies such as Carvalho, de Oliveira & Campos (2016) and Alsayani et al. (2025) have shown that gaze data may help reveal whether users are overloaded by visual complexity, misdirected by suboptimal layouts, or aided by perceptual grouping and focus mechanisms. These works generally emphasize that effective visual interfaces naturally direct user attention to the most critical information, minimizing the time and effort required for manual searching.

In the present study, eye-tracking was employed in both a measurement and observational capacity. The goal was to compare gaze behaviors between users interacting with the Multiviz interface and those using a conventional table based layout. Gaze heatmaps, scan paths, and fixation clusters were used to identify high level differences in visual search strategies and attention allocation. Typical layouts showing heatmaps from the experiment depicting treemap and tabular style screens used are given in Fig. 1.

It is important to emphasize that eye-tracking always provides supplementary evidence. These results complement more traditional performance metrics such as task completion time and error rate by offering an additional layer of insight into how users experience and navigate the interface. More granular gaze analytics for each element (e.g., fixation duration distributions, transition matrices, or attention entropy) could generally further quantify the impact of interface design on cognitive effort.

The development of Multiviz was evaluated through controlled eye-tracking experiments, comparing it to a baseline Excel interface using realistic financial analysis tasks. Eye-tracking in this study served as a diagnostic tool, reinforcing that Multiviz’s structured layout and visual guidance mechanisms align with natural attentional strategies. This supports the broader system goal of making complex financial information more accessible, interpretable, and actionable.

Overview

The Multiviz system was designed to meet the real-time cognitive and operational demands of financial analysts by providing a perceptually optimized, intelligent, and responsive data visualization interface. The architecture employs a layered design that integrates data acquisition, visual structuring, and cognitive support mechanisms, enabling seamless monitoring of multidimensional financial information. Multiviz is designed to operate within the Bloomberg Terminal™ framework for data acquisition and deployment.

The system comprises three primary layers:

• (1)

  Visual data layer: Responsible for data mapping, layout generation, and rendering of treemaps and visual encodings.

• (2)

  Smart agent layer: Monitors incoming data streams in real time, applies rule based logic to detect significant patterns or anomalies, and triggers visual emphasis operations.

• (3)

  Interaction and navigation layer: Manages user interface controls, filters, and navigational interactions, allowing users to manipulate focus, apply constraints, or explore data hierarchies interactively.

These components work together to transform raw financial data into an interactive visual workspace where structure, priority, and change are encoded both visually and semantically. The overall architecture of Multiviz is depicted in Fig. 2, illustrating the relationship between data ingestion, smart agent evaluation, and visual rendering.

The Multiviz architecture consists of three layers that operate together in real time. Financial data is ingested through the Bloomberg application programming interface (API) and passed to the visual data layer for treemap generation and attribute encoding. The smart agent layer continuously evaluates incoming values against predefined rules and triggers visual emphasis cues when relevant conditions are met. The interaction and navigation layer provides filtering, drill-down, and control mechanisms, enabling users to explore and interpret high-density financial data efficiently.

Visual data layer

The visual data layer is responsible for converting complex, real-time financial data into perceptually coherent visual structures. Serving as the core of Multiviz, this layer handles data driven layout generation, attribute encoding, and screen rendering. It is designed to support high frequency data updates and visual scalability, accommodating dozens of financial indicators across hundreds of entities without sacrificing readability in the terminal. It includes a list view generator, treemap generator, dynamic filters, and other standard user interface rendering components.

Visual encoding of financial attributes

The visual glyphs in a typical Multiviz interface are designed to encode up to nine distinct variables simultaneously, enabling dense information display without overloading the user. Accommodated within the treemap structure, each treemap tile functions as a composite perceptual unit, layering multiple encodings such as geometric form, color, label, icon, and numeric overlay to support quick, parallel interpretation, as shown in Fig. 3.

Variables include:

• v₁: rectangle size (e.g., market cap)

• v₂: circle size (e.g., risk exposure)

• v₃: rectangle color (e.g., share price change)

• v₄: circle color (e.g., volatility class)

• v₅: circle area texture (e.g., rating stability)

• v₆: text label (e.g., firm name or rating)

• v₇: icon (ordinal, e.g., trend arrow or alert symbol)

• v₈: binary flag (e.g., warning triangle indicating rule violation)

• v₉: overlaid numeric value (e.g., yield or return delta)

The layered encoding strategy allows users to interpret multiple dimensions without additional navigation or clicking. It performs this by embedding semantic cues directly into the visual representation, reducing the need for external legends or modal popups, and allowing for faster cognitive parsing of the interface. Such multi channel glyph design follows established visualization principles, such as those described by Bertin (1983) and Cleveland & McGill (1984), emphasizing parallel, non conflicting visual mappings (Friendly, 2006).

Each treemap node encodes multiple attributes using a combination of visual channels, such as:

• Rectangle size: Market capitalization or total outstanding debt

• Rectangle color: Direction and magnitude of share price or CDS change

• Circle color (overlaid): Volatility or risk-class value

• Circle border: Spread behavior or external score (e.g., rating outlook)

• Brightness and texture (triggered by agents): Used to emphasize or de-emphasize nodes dynamically

• Visual glyph: Supplementary shape, arrow, and text as options

The encodings were selected based on perceptual research emphasizing separable channels to avoid cognitive interference and ensure rapid attribute differentiation (Bertin, 1983; Cleveland & McGill, 1984).

This design aligns with established principles of visual perception and encoding separability, ensuring that users can rapidly interpret the state of each entity. By embedding both semantic value and interaction cues directly into the shape, external lookup requirements are minimized, supporting fast, context-rich comparisons across the treemap.

Furthermore, the display adapts to zoom levels and viewport size, automatically rescaling visual elements to preserve interpretability as an additional feature.

Update mechanism and rendering performance

Financial data is streamed into the system via a data ingestion module that connects to structured data sources (e.g., SQL-based financial feeds or APIs). In this study, the Bloomberg API was used to retrieve data within the Terminal framework. Each record is parsed and pushed into the rendering pipeline with a maximum update interval of 200–300 ms, depending on system load. This provides real time perception to the user. To minimize frame drops or interface lag, the system uses incremental update queues and batched rendering for grouped changes. The information architecture of this mechanism is shown in Fig. 4.

The rendering engine is built using a custom graphics library, allowing Multiviz to maintain consistent update rates even under high variable update loads which is a crucial requirement for real-time trading environments.

Smart agent layer

The smart agent layer in Multiviz acts as a domain aware, rule driven component that actively monitors the financial data stream and modifies the visual interface in real time to support attention guidance and anomaly detection. Unlike traditional alert systems that rely on external notifications or intrusive pop-ups, this layer enables the interface itself to react dynamically to context changes and financial market signals.

Visual emphasis without interruption

A key design principle of the smart agent layer is non intrusiveness. Rather than interrupting the user or masking the data with overlays, agents enhance existing visual properties to steer user attention naturally. This attention steering technique is based on literature in visual cognition and human–computer interaction (HCI). Subtle but perceptually salient changes, such as motion, contrast, or brightness shifts, are more effective and less cognitively disruptive than modal alerts (Franconeri et al., 2021; Julesz, 1981). Multiviz incorporates visual emphasis and additional channels on enhanced treemap design as shown in Fig. 5.

Visual cues used by agents include:

• Pulse highlighting: temporal glow around a tile’s border

• Brightness emphasis: increased luminance of the target rectangle

• Motion cues: oscillation or bounce effect to signal volatility

• Color shift: saturation change to encode anomaly severity

These cues are layered carefully to avoid over salience, especially in dashboards where multiple agents may be triggered simultaneously. A priority queuing mechanism ensures that the most critical alerts are visually dominant, while others are subdued or stacked using layered overlays.

Interaction and Navigation layer

The interaction and navigation layer provides user facing functionality for exploration, filtering, and manipulation of the data visualizations. As the visual data layer presents a perceptually optimized snapshot of the current state of financial instruments and the smart agent layer selectively guides user attention, the interaction layer ensures that users remain in control, dynamically tailoring the interface to match their informational needs and decision contexts. Dynamic features reduce visual clutter and support real-time decision making (Biedert, Schwarz & Roth-Berghofer, 2008).

Machine learning based parameter filtering

The number of traditional variables received from the Bloomberg data server exceeds 30. Although all are considered important and are thus offered through terminal services, traders usually focus on certain parameters frequently and monitor those closely, while other parameters are examined infrequently. To determine which parameters are most significant for continuous display, a simple machine learning algorithm was used and results were analyzed.

To reduce the number of parameters displayed in real time to about 5 or 6, a user study was conducted using the traditional system. A machine learning based selection process was performed to identify a usable data filtering technique and determine the five most critical features. Initially, data was collected over two days, yielding 1105 samples with 34 features for each stock, along with whether each feature was actually examined (“viewed/not viewed”) by professional traders. This constituted a classification problem with two classes: viewed (+) and not viewed (–). Given the class imbalance in this data (most features are not actively used), precision, recall, and ROC curves were used as evaluation metrics rather than accuracy alone. Several classification algorithms, including decision tree based methods (J48, jRip, PART), instance based (IBk), and ensemble methods (Bagging, RandomForest), were evaluated and produced adequate results. The decision tree and rule-based classifiers (J48, jRip, PART) performed well in terms of balancing accuracy and interpretability, and also provided short test times. As a result, these algorithms were identified as suitable for selecting the most relevant parameters for display, based on their reliable classification of viewed parameters and overall system responsiveness in near real-time scenarios.

User study with eye-tracking

To evaluate the effectiveness of the Multiviz interface in supporting rapid visual analysis of multidimensional financial data, a controlled laboratory user study was conducted. The study compared the performance and gaze behavior of participants using two different interface designs: the Multiviz system with hierarchical visual layout and smart agents, and a baseline tabular interface resembling traditional grid based displays. The primary goal was to assess whether Multiviz could improve task performance and reduce visual search effort in complex analytical scenarios.

Task completion time and accuracy

Participants completed tasks more quickly using Multiviz than with the baseline tabular interface. The average completion time with Multiviz was 113.11 ± 13.81 s compared to 241.47 ± 42.68 s for the tabular interface, t(11) = –10.63, p < 0.000001, Cohen’s d = –3.07, representing a 53% reduction. The largest time savings were observed in tasks requiring multi-variable filtering and anomaly detection, where spatial grouping and agent-based cues enabled users to rapidly locate relevant data regions.

Accuracy was also significantly higher with Multiviz (9.17 ± 0.27%) compared to the tabular interface (7.75 ± 0.26%), t(11) = 11.76, p < 0.000001, d = 3.39. Errors in the baseline condition were mostly due to missed variable combinations or misinterpreted rows, confusion, while errors in Multiviz were less frequent and typically related to over selection during early exploration. A comparison of mean fixation times, saccade durations, task completion times, and accuracy across both interfaces is provided in Table 1, which shows the overall advantage of the Multiviz treemap design.

Participants made fewer mouse interactions when using Multiviz. The number of clicks decreased from 52.6 ± 5.4 with the tabular interface to 34.9 ± 3.4 with Multiviz, t(11) = –10.63, p < 0.000001, d = –3.07. This suggests that users required less trial and error and navigation overhead, supporting the hypothesis that visual encodings and perceptual layout reduce the need for additional manipulation and search. This aligned with the post-experiment interviews.

Gaze behavior

Eye-tracking heatmaps and scan paths revealed distinct behavioral differences between conditions, as well as differences in fixation and saccade measurements.

In the tabular interface, gaze patterns were characterized by horizontal sweeps across rows, repeated fixations on column headers, and frequent re-checking of variable combinations. The mean fixation duration was higher in the tabular interface (441.0 ms vs. 414.0 ms with, t(11) = –3.45, p < 0.005, d = –1), suggesting potential difficulties in cognitive processing. In contrast, users of Multiviz exhibited more clustered fixations in relevant treemap regions. They also made rapid returns to agent highlighted nodes and demonstrated fewer long saccades or visual backtracking, ultimately reducing mean saccade times. Due to significantly longer task completion times, mean number of fixations were also higher in Tabular interface (602.2 ± 108.8 vs. 259.3 ± 43.1 fixations, p < 0.0000017).

Figures 6 and 7 showed a sample heatmap comparison from a task involving multivariable filtering (DAX market data). In the baseline tabular interface, attention was widely dispersed, whereas in Multiviz, fixations were concentrated in visually prioritized zones.

These patterns suggested that Multiviz supported more focused visual attention, likely due to the combined effect of semantic grouping, perceptual encoding, and agent-based cueing.

The observed patterns indicated that users interacting with Multiviz exhibited more concentrated fixations, with shorter scan paths and fewer returns to previously visited regions. In contrast, users working with the tabular interface spent more time scanning horizontally across rows and vertically down columns, often revisiting cells to verify contextual information. These differences suggest that the spatial structuring and perceptual cues in Multiviz helped users form a coherent visual strategy more quickly. For example, in tasks requiring the identification of firms with specific financial criteria (e.g., high CDS but low interest coverage), gaze behavior in Multiviz tended to center around the visual clusters emphasized by agent driven highlighting. In the list view, the same task required participants to scan and locate relevant rows, mentally combine values across columns, and jump frequently between sections.

Participant feedback

Qualitative feedback gathered through short post-task interviews indicated a clear preference for Multiviz. Participants described the interface as “Visually clean and smart,” “Helpful in pointing me to the problem area,” and “Much faster to work with once I understood what the colors and highlights meant.” Some participants also noted brief initial confusion when interpreting the layered glyph encodings before becoming familiar with the visual mappings, indicating a short learning period associated with the multivariate design.

In addition, several participants mentioned that the agent-driven highlights increased their trust in the visualization, particularly for complex filtering and anomaly detection tasks.

Visual structuring enables faster cognition

The observed reduction in task time and interaction count suggests that preattentive visual structuring, enabled by treemaps, size and color encodings, and layout stability, allowed users to “recognize” rather than “search.” This finding is consistent with the theory of visual cognition advanced by Ware (2019) and supported by empirical work from Cleveland and McGill, emphasizing that position, color, and size are rapidly processed channels for visual comparison.

In particular, the use of consistent spatial regions for sector and firm groupings enabled subjects to build spatial memory, reducing re-orientation effort and providing fast access to familiar locations across tasks. This aligns with the principle of layout stability as a cognitive anchor, as explored in Bederson, Shneiderman & Wattenberg (2002) work on dynamic interfaces.

Agent-driven cues amplify situational awareness

The smart agent layer contributed to non-disruptive attention guidance, particularly in tasks requiring synthesis of multiple financial indicators. Unlike alarm-based systems that interrupt user flow, Multiviz agents operated inline with the visual representation, using subtle cues to direct gaze rather than distract. This model draws from the concept of ambient display, where low-cost perceptual signals can enable situational awareness without demanding conscious switching (Pousman, Stasko & Mateas, 2007).

The effectiveness of these agent cues also reflects principles from feature integration theory by Treisman & Gelade (1980), wherein attention is automatically drawn to feature conjunctions such as color plus motion or brightness plus border. By embedding rule based intelligence into the perceptual field, the system allowed users to maintain peripheral awareness of non focal anomalies without the cognitive cost of constant scanning.

Gaze-based validation adds depth

Although limited in data size, the gaze analysis provided valuable insight into the visual search efficiency enabled by Multiviz. The tighter fixation clusters and reduced scan path complexity suggested that the interface allowed for more targeted attention allocation, consistent with literature on perceptual grouping and guided visual search by Wolfe (2007). This evidence confirms that the design reduced extraneous cognitive load, even without relying on subjective workload ratings such as NASA-TLX (Hart & Staveland, 1988).

The quantitative analysis further confirmed that fixation duration was significantly shorter in Multiviz (414.0 ± 13.2 ms) than in the table (441.0 ± 22.4 ms), t(11) = –3.45, p = 0.005, d = –1.00. Scanpath length was also greatly reduced (43,477 ± 7,554 px vs 89,210 ± 19,989 px), t(11) = –8.23, p < 0.00001, d = –2.38. Saccade duration showed a nonsignificant trend (38.1 ± 2.2 ms vs 36.2 ± 3.1 ms), p = 0.067. These metrics corroborate the performance results by demonstrating that Multiviz supported more efficient and focused visual strategies.

These qualitative insights serve as grounded justifications for interface design decisions, bridging the gap between engineering choices, such as the use of border glow, and user experience.

Domain specificity and generalizability

Although Multiviz was designed and evaluated specifically for financial market monitoring, its underlying architectural principles, hierarchical treemaps, multi variable glyphs, and inline rule-based agents, may have relevance in other time-sensitive, data-intensive domains.

However, such applicability remains prospective. The present study does not provide direct empirical validation outside the financial context, and claims of cross domain transferability should be treated with caution. Future research should investigate Multiviz inspired designs in settings such as cybersecurity, healthcare, or industrial operations, with appropriate adaptation to domain-specific data structures and decision tasks. Only through such targeted evaluations can the broader utility of the architecture be established with confidence.

Limitations

This study has several constraints that qualify the interpretation of the results. First, the sample size was relatively small (N = 12) and composed of domain familiar participants (finance professionals and advanced students) recruited through existing networks, which may bias performance toward expert strategies and limit generalizability to novice users. Second, participants received only a brief familiarization period (≈2–3 min per interface). Such short onboarding may depress early performance for the composite glyph encodings. Longer training could attenuate early errors and alter relative differences. Third, the evaluation was conducted in a controlled laboratory setting with fixed and difficult hardware (desktop display and SMI eye tracker at ≥120 Hz) and scripted tasks. These conditions do not capture typical production constraints (interruptions, multi-display setups, concurrent applications) on a trading floor. The evaluation therefore does not simulate the time pressure, concurrent task demands, or environmental complexity of an operational trading floor. Instead, it examines perceptual and interaction performance under controlled laboratory conditions. Fourth, although a within-subjects counterbalanced design was used, residual learning or fatigue effects cannot be fully ruled out. Fifth, our gaze analysis relied on aggregate measures (mean fixation duration, saccade duration, scanpath length) and did not include subjective workload scores (e.g., NASA-TLX) or trial level models. We therefore treat eye-tracking results as convergent evidence rather than definitive workload estimates. Finally, the present version evaluates a rule based, non adaptive agent; results may differ with adaptive or predictive agents that personalize thresholds or learn user intent over time.

Design implications for high-density interfaces

The findings suggest several actionable design guidelines for real-time, multivariate decision support systems:

• Preserve layout stability to support spatial memory and reduce re-orientation time.

• Use perceptual cues such as border pulse or brightness to guide attention without intruding on the data.

• Design visual elements as multi attribute glyphs, leveraging separable channels like size, color, shape, and text.

• Avoid global sorting and re-layouts during updates; prefer inline visual emphasis.

• Embed decision logic visually, highlighting what matters directly within the interface rather than in external alert panels.

These principles may offer a blueprint for designing cognitively efficient interfaces in domains where human operators must process complex data under constraints of time and uncertainty.

Future work

Several directions emerge from this study. The present evaluation focused on a rule based, non adaptive agent layer. Future extensions may include adaptive or user trainable agents that learn from user feedback or interaction histories, as well as generative approaches that propose new visual encodings and anomaly cues in real time. Such developments would align with recent advances in agent-assisted and AI driven visualization design, and could improve personalization, trust, and decision accuracy.

Second, although our study focused exclusively on financial data, similar architectures could be tested in other high stakes domains such as cybersecurity, healthcare monitoring, and industrial process control. Cross domain validation would help establish the scalability and transferability of the Multiviz design.

Third, the participant pool was limited to twelve domain familiar users. Future studies should include larger and more diverse samples spanning novice and expert populations, different financial roles, and potentially users from other data intensive domains. Multi user or collaborative settings could further test how attention guiding agents perform in team based decision contexts.

Finally, our study examined short term task performance in a controlled laboratory setting. Longitudinal evaluations, in which participants adopt Multiviz over extended periods of use, are needed to capture learning effects, trust calibration, fatigue, and integration into everyday work practices. These future investigations will help establish the external validity, scalability, and sustainability of the Multiviz approach.

Multiviz contributes not only a novel interface but also a replicable design pattern for cognitively optimized, task aligned visualization in expert settings. As data environments grow increasingly complex, such systems will be critical in ensuring that human operators remain informed and empowered.