griffinqiu / trials

Tiny Bayesian A/B testing library

trials

Tiny Bayesian A/B testing library

Installation

pip install git+git://github.com/bogdan-kulynych/trials.git@master

pip might not install all the system packages needed for scipy. To install them on Debian:

sudo apt-get install libatlas-dev libatlas-base-dev liblapack-dev gfortran

Run the tests:

nosetests trials/tests

Usage

Import package

from trials import Trials

Start a split test with Bernoulli (binary) observations

test = Trials(['A', 'B', 'C'])

Observe successes and failures

test.update({
    'A': (50, 10), # 50 successes, 10 failures, total 60
    'B': (75, 15), # 75 successes, 15 failures, total 90
    'C': (20, 15)  # 20 successes, 15 failures, total 35
})

Evaluate some statistics

dominances = test.evaluate('dominance', control='A')         # Dominance probabilities P(X > A)
lifts = test.evaluate('expected lift', control='A')          # Expected lifts E[(X-A)/A]
intervals = test.evaluate('lift CI', control='A', level=95)  # Lifts' 95%-credible intervals

Available statistics for Bernoulli observation variations: expected posterior, posterior CI, expected lift, lift CI, empirical lift, dominance, z-test dominance.

Print or visualize results

for variation in ['B', 'C']:
    print('Variation {name}:'.format(name=variation))
    print('* E[lift] = {value:.2%}'.format(value=lifts[variation]))
    print('* P({lower:.2%} < lift < {upper:.2%}) = 95%' \
        .format(lower=intervals[variation][0], upper=intervals[variation][2]))
    print('* P({name} > {control}) = {value:.2%}' \
        .format(name=variation, control='A', value=dominances[variation]))

Examine the output:

Variation B:
* E[lift] = 0.22%                       # expected lift
* P(-13.47% < lift < 17.31%) = 95%      # lift CI
* P(B > A) = 49.27%                     # dominance
Variation C:
* E[lift] = -31.22%
* P(-51.33% < lift < -9.21%) = 95%
* P(C > A) = 0.25%

Interpreting and analyzing results

As per the output above there's 50% chance that variation B is better than A (dominance). Most likely it is better by about 0.2% (expected lift), but there's 95% chance that real lift is anywhere betwen -13% to 17% (lift CI). You need more data to know if B is better or worse for sure.

There's 100% - 0.25% = 99.75% chance that variation C is worse than A. Most likely it is worse by about 31%, and there's 95% chance that real lift falls betwen -51% to -9%. The data was sufficient to tell that this variation is almost certainly inferior to both A and B. However, if this 99.75% chance still doesn't convince you, you need more data.

Theory

Explanation of mathematics behind and usage guide are coming soon as a blog post.

Meanwhile, see the notebook for comparison of Bayesian lift (blue) and empirical lift (green) errors in a theoretical benchmark with equal sample sizes. Bayesian approach is a little better at predicting the lift, but no miracles here. Bayesian p-values and frequentist (z-test) p-values yield almost identical results.