Visualizing Uncertainty in Time Series Forecasts: The Impact of Uncertainty Visualization on Users' Confidence, Algorithmic Advice Utilization, and Forecasting Performance

Abstract

Time series forecasts are always associated with uncertainty. However, experimental studies on the impact of uncertainty communication provide inconclusive results on the effect of providing this uncertainty to end users. In this study, we examine the impact of uncertainty visualizations on advice utilization in the context of time series forecasts with line charts. Based on a literature review, we identified probabilistic framing versus frequency framing as a theoretical foundation for studying the topic. We then used the Judge Advisor System (JAS) as a framework to create an experimental design with two treatments (95% prediction interval [PI] and ensemble plots), one control group (point plot), and various mediating variables (e.g., confidence, graph literacy, and domain knowledge). The results of an online experiment ( $N=239$) indicate a U-shaped relation between uncertainty visualization and forecasting performance. Additionally, we examine how confidence, advice utilization, and other factors mediate the effect of uncertainty visualizations. This paper highlights the benefits of PI plots for researchers and practitioners engaged in the development of effective uncertainty visualizations for decision-making processes.