tw-experimentation 0.1.1.13

Wise AB platform

TW Experimentation: A library for automated A/B testing and causal inference

TW Experimentation is a library to design experiments, check data, run statistical tests, causal inference and make decisions

Table of Contents

• TW Experimentation: A library for automated A/B testing and causal inference
  • What can this do for you?
    • 1. Designing experiments
    • 2. Evaluating results
  • Installation
  • Quick Start
    • Notebooks
    • Streamlit web app
  • For Developers
    • Testing

What can this do for you?

The experimentation library can help you with:

• Sample Size Calculator link
• Integrity checks + Evaluation link
• Evaluation (Bayesian A/B testing) link

1. Designing experiments

By using TW Experimentation you can design your experiments, choose sample size, evaluate the experiment and calculate features and metrics.

2. Evaluating results

You can use different statistical tests and causal inference techniques.

Installation

You can install the package via the dependency manager poetry after cloning/git pull/download it as a zip from this repository.

To do so, clone the repository by running

git clone [e-mail]:transferwise/tw-experimentation.git

from terminal. To set up poetry, run

make set-up-poetry-mac

for mac (or linux) and

make set-up-poetry-windows

for windows. Then, run

make run-streamlit-poetry

from the root of the package folder.

Alernative: TW Experimentation requires the following libraries to work which you can find in the .yml file. To install requirements please make sure you have installed the package manager Anaconda and then run the following commands in the terminal:

conda env create -n <my_env> -f envs/environment.yml
conda activate <my_env>

If you are using Windows, please do these additional steps:

1. pick a jaxlib-0.3.7 wheel from here https://whls.blob.core.windows.net/unstable/index.html and install it manually (pip install <wheel_url>)
2. Install jax==0.3.7

Quick Start

Make sure you have followed the installation instructions.

Notebooks

You can use the jupyter notebooks 1_pre_experiment.ipynb or 2_integrity_checks + evaluation.ipynb for experiments design and evaluation. The tw experimentation package can be used for different things, for example for analyzing results:

df = pd.read_csv('experiment.csv')

ed = ExperimentDataset(
    data=df,
    variant="T",
    targets=['conversion', 'revenue'],
    date='trigger_dates',
    pre_experiment_cols=None,
    n_variants=2,
)
ed.preprocess_dataset(remove_outliers=True)

This code will generate the data model for experiment analysis

Streamlit web app

Open terminal and navigate to the repository. Then navigate to the folder ./tw_experimentation/streamlit.

Now run the command streamlit run Main.py and the app should open in your browser.

Tip on navigation: ls - show files in current directory pwd - print current directory address cd - change directory, e.g. cd ./tw_experimentation/streamlit

For Developers

Testing

We use PyTest for testing. If you want to contribute code, make sure that the tests in tests/ run without errors.