Interpretable, non-mechanistic forecasting using empirical dynamic modeling and interactive visualization.

Area of Science:

• Epidemiology
• Data Science
• Computational Biology

Background:

• Forecasting methods often lack interpretability, hindering user integration of domain knowledge.
• Mechanistic models offer interpretability but require detailed dynamical system knowledge.
• Non-mechanistic methods can be opaque, limiting trust and applicability.

Purpose of the Study:

• Introduce EpiForecast, an interpretable, non-mechanistic forecasting tool.
• Enhance user understanding of forecast generation through interactive visualization.
• Provide a data-focused forecasting technique based on empirical dynamic modeling.

Main Methods:

• Developed EpiForecast as an in-browser web application for FAIR data principles and privacy.
• Implemented a four-plot interactive dashboard for visualizing forecast generation.
• Utilized kernel density estimation for distributional forecasts, visualized with color gradients.

Main Results:

• EpiForecast provides interpretable, non-mechanistic forecasts.
• Interactive visualizations aid in understanding forecast generation processes.
• Distributional forecasts offer intuitive visual summaries of future estimations.

Conclusions:

• EpiForecast enhances the applicability of forecasts by enabling user interpretation.
• The tool promotes transparency in forecasting through interactive data visualization.
• An accessible, in-browser application ensures widespread usability and adherence to FAIR principles.