Enhancing Uncertainty Communication in Time Series Predictions: Insights and Recommendations

TL;DR

This paper addresses how uncertainty in time-series forecasts is communicated to diverse users and evaluates how visualization choice and user characteristics shape uncertainty estimation. It employs two user studies (n=94 and n=31) comparing five uncertainty-visualization variants (Confidence Band, Overlapping Bands, decreasing color-saturation band, decreasing-area Circular Glyphs, Colored Markers) and analyzes task performance with nonparametric Friedman/Nemenyi tests and a normalized absolute-error metric. Key contributions include guidelines for standardized uncertainty terminology, meeting diverse informational needs (statistical and model information), mitigating numeracy effects, reducing clutter and optimizing aesthetics, and enabling interactive features to improve comprehension. The findings inform dashboard design for epidemiological forecasts, enabling more reliable decision-making by a broad range of stakeholders and highlighting directions for future research on interactive uncertainty tools and terminology standardization.

Abstract

As the world increasingly relies on mathematical models for forecasts in different areas, effective communication of uncertainty in time series predictions is important for informed decision making. This study explores how users estimate probabilistic uncertainty in time series predictions under different variants of line charts depicting uncertainty. It examines the role of individual characteristics and the influence of user-reported metrics on uncertainty estimations. By addressing these aspects, this paper aims to enhance the understanding of uncertainty visualization and for improving communication in time series forecast visualizations and the design of prediction data dashboards.As the world increasingly relies on mathematical models for forecasts in different areas, effective communication of uncertainty in time series predictions is important for informed decision making. This study explores how users estimate probabilistic uncertainty in time series predictions under different variants of line charts depicting uncertainty. It examines the role of individual characteristics and the influence of user-reported metrics on uncertainty estimations. By addressing these aspects, this paper aims to enhance the understanding of uncertainty visualization and for improving communication in time series forecast visualizations and the design of prediction data dashboards.

Figures (7)

• Figure 1: For 20 independent samples from the population, nineteen of their 95% confidence intervals (marked by black squares) contain the true mean, while one (marked by a hollow circle) does not tantan_ci.
• Figure 2: A simplified version of the knowledge generation model Sacha2016_US
• Figure 3: A snapshot of the Epidemiological Scenarios for Infectious Diseases (ESID) application, a visual analytics application developed to show the results of epidemiological simulations. On the left hand side is a map of the German counties with a customizable heat legend. At the top section is a purple scenario card representing one simulation scenario. Left of the scenario cards is a list of infection states or aggregations of infection states, like infected, hospitalized, and dead. The bottom portion contains a line chart comparing the scenarios over a predefined timeline. betz2023esid.
• Figure 4: Uncertainty represented as A. Confidence Band, B. Overlapping Bands, C. Blur, D. Circular Glyphs and E. Colored Markers
• Figure 5: Boxplots showing the performance of visualization techniques. A: The five visualization techniques from the initial study ($p < 0.001$). The performance under Colored Marker is significantly worse than all the other techniques. B: The three visualization techniques from the second study ($p \leq 0.05$). Colored Marker is not considered for the second study due to its poor performance in the initial study. Confidence Band is also excluded as it did not show a significant difference compared to other techniques and does not qualitatively depict the nature of a credible interval. In the second study, Circular Glyphs shows a significant difference with Blur ($p = 0.058$) but no other significant differences were observed. Circular Glyphs demonstrates higher median performance and smaller variability.