cardtale 0.1.0

Data, Model, and Algorithm Cards for Time Series

cardtale

cardtale is a Python package for generating automated model and data cards for time series, streamlining the documentation process for machine learning models and datasets.

Key Features

• Automated generation of PDF reports with comprehensive time series analysis
• Built-in statistical analysis and visualization of temporal patterns
• Support for univariate time series data (will support other structures in future iterations)

Each time series is studied from multiple dimensions, including:

• Data Overview: Fundamental characteristics and statistical properties analysis
• Trend Analysis: Long-term growth patterns and level stabilization assessment
• Seasonality Detection: Multiple seasonality levels with strength metrics
• Variance Analysis: Heteroskedasticity testing and variance stabilization methods
• Change Point Detection: Identification of structural changes and regime shifts

Basic Example

Here's a basic example of cardtale.

from datasetsforecast.m3 import M3
from cardtale.cards.builder import CardsBuilder

df, *_ = M3.load('./assets', group='Monthly')

freq = 'ME'
uid = 'M1080'

series_df = df.query(f'unique_id=="{uid}"').reset_index(drop=True)

tcard = CardsBuilder(series_df, freq)
tcard.build_cards()
tcard.get_pdf(path='examples/cards/example.pdf')

Screenshots

⚠️ WARNING

cardtale is in the early stages of development. It is designed to cover datasets containing single univariate time series, focusing on forecasting tasks.

cardtale has been developed for monthly time series. So, the output for other frequencies may not be as reliable. Especially time series with complex seasonality.

If you encounter any issues, please report them in GitHub Issues

Installation

Prerequisites

Required dependencies:

arch==7.1.0
pandas==2.2.3
lightgbm==4.5.0
neuralforecast==1.7.5
mlforecast==0.13.4
statsforecast==1.7.8
datasetsforecast==0.0.8
numerize==0.12
plotnine==0.13.6
statsmodels==0.14.4
jinja2==3.1.4
ruptures==1.1.9
weasyprint==62.3

You can install cardtale using pip:

pip install cardtale

[Optional] Installation from source

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

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

License

Cardtale is released under the MIT License. See the LICENSE file for more details.

Mission

In the machine learning lifecycle, proper documentation of models and datasets is crucial for transparency, reproducibility, and responsible AI practices. However, creating comprehensive model and data cards can be time-consuming. The Python package cardtale aims to partially automate this process, making it more efficient and consistent.

The goal of cardtale is to generate a set of analyses, visualizations, and interpretations based on input model metrics and dataset characteristics. While it doesn't replace the expertise of data scientists or domain experts, Cardtale speeds up the creation of model and data cards, guiding analysts towards key insights and areas that may require further exploration.