This repository incorporates several analytics tools used during processing of CPG (consumer packaged goods) datasets at Suburbia as well as notebooks demonstrating their usage at Suburbia. They are separated into four different categories:
Set of functions for classification of stable merchants based on four criteria:
- lifespan in days,
- relative activity (percentage of days with sales activity),
- volatility of sales quantities,
- longest gap (longest period without activity).
Stability of a merchant is determined by comparing these criteria with a set of thresholds.
Thresholds can be chosen using an associated Stability settings Shiny app.
Tools for detection of outliers in a set of specified metrics, based on Prophet.
We fit a time series model for each metric we are interested in, using data up until a reference date. All values collected after the reference date will be then compared with predicted values. If observed value is outside of predicted interval or is missing, the algorithm labels it as such.
For each time series, we provide an indication of correctness, a chart of actual values together with a predicted interval, and a summary of problematic dates.
Next time the outlier detection is run, user can choose to exclude certain problematic dates to prevent the model from training on this data.
An algorithm for tagging new line items as "unbranded" given a list of items that have already been tagged as "unbranded".
Tools for simulation of data for demonstration of data processing at Suburbia,
available through example Jupyter notebook notebooks/Data-Science at Suburbia.ipynb.
Install:
$ pip install .For an IPython notebook and example files, go to notebooks/.
To start Stability settings app, run:
$ R -f R/stability_settings/app.R