Interpretable Uncertainty Diagnostics
Building the tools to make probabilistic forecasts trustworthy, auditable, and actionable.
The Problem
Most forecasting models produce uncertainty estimates, but few practitioners know how to interpret them. Coverage gaps, overconfidence, and poor calibration go undetected — leading to decisions based on false confidence.
The Innovation
The k-diagram framework introduces a polar coordinate system for visualizing multiple dimensions of forecast quality simultaneously — coverage, calibration, severity, and reliability — in a single, interpretable chart.
A Concrete Application: Subsidence Forecasting
How k-diagram diagnostics improved real-world model evaluation
Generate Forecasts
Run XTFT or PINN-based model to produce probabilistic subsidence predictions with confidence intervals.
Apply k-Diagram
Feed predictions into the k-diagram toolkit to produce polar diagnostic plots across all coverage levels.
Interpret & Improve
Identify coverage gaps, overconfident intervals, or systematic biases and retrain accordingly.
What Can k-Diagram Answer?
The k-diagram toolkit turns abstract uncertainty metrics into visual diagnostics.
"Is my model well-calibrated?"
Visualize whether prediction intervals actually contain the true value at the stated confidence level.
"Where does coverage fail?"
Identify specific coverage levels or time horizons where the model systematically under- or over-predicts.
"How severe are the errors?"
Quantify the magnitude of coverage gaps and their potential impact on downstream decisions.
Research Outcomes
- Developed the k-diagram open-source Python toolkit for multi-dimensional forecast diagnostics.
- Demonstrated application on urban subsidence forecasting with XTFT and PINN-based models.
- Submitted to Zenodo with full documentation and reproducible notebooks.
- Framework applicable to any probabilistic forecasting domain — climate, hydrology, finance.
Tools & Resources
Related publications
- k-diagram: Rethinking Forecasting Uncertainty via Polar-Based VisualizationKouadio, K. L. · Journal of Open Source Software (JOSS) · 2025
- k-diagram: Technical Report — Derivations and DetailsKouadio, K. L. · Zenodo (Technical Report) · 2025