modelradar 0.1.3

Aspect-based Forecasting Accuracy

Model Radar 🎯

A framework for aspect-based evaluation of time series forecasting models based on Nixtla's ecosystem.

Overview

Model Radar introduces a novel aspect-based forecasting evaluation approach that goes beyond traditional aggregate metrics. Our framework enables:

• Fine-grained performance analysis across different forecasting aspects
• Better understanding of model behavior in varying conditions
• More informed model selection based on specific use case requirements

🚀 Getting Started

Check the notebooks folder for usage examples and tutorials.

Installation

You can install modelradar using pip:

pip install modelradar

[Optional] Installation from source

To install modelradar from source, clone the repository and run the following command:

git clone https://github.com/vcerqueira/modelradar
pip install -e modelradar

Prerequisites

Required dependencies:

utilsforecast==0.2.11
numpy==1.26.4
plotnine==0.14.5

Example outputs

• Spider chart with overall view on several dimensions:

• Parallel coordinates chart with overall view on several dimensions:

• Barplot chart controlling for a given variable (in this case, anomaly status):

• Grouped bar plot showing win/draw/loss ratios wrt different models:

📑 References

Cerqueira, V., Roque, L., & Soares, C. (2024). "Forecasting with Deep Learning: Beyond Average of Average of Average Performance." arXiv preprint arXiv:2406.16590

Check DS24 folder to reproduce the experiments published on this paper. The main repository and package contains an updated framework.

⚠️ WARNING

modelradar is in the early stages of development. The codebase may undergo significant changes. If you encounter any issues, please report them in GitHub Issues

Project Funded by

Agenda “Center for Responsible AI”, nr. C645008882-00000055, investment project nr. 62, financed by the Recovery and Resilience Plan (PRR) and by European Union - NextGeneration EU.