experiment-stats

name	experiment-stats
description	This skill should be used when the user asks 'is this A/B test significant', 'did the variant actually win', 'compute lift and confidence interval', 'score a marketing experiment', 'run the promote-gate', or wants a statistically defensible promote/hold/reject call on an A/B or multivariate marketing experiment. Computes lift, a bootstrap confidence interval, and a significance test (Mann-Whitney U or two-proportion z) using pure Python standard library, without scipy or numpy. Every p-value is computed from the input data, never estimated. Use before promoting any experiment variant to production.
version	0.1.0

Experiment Stats

Compute real inferential statistics for A/B and multivariate marketing experiments, then apply a dual promote-gate to decide whether a variant actually won. The p-value and confidence interval behind that call come from bootstrap resampling and a hand-implemented Mann-Whitney rank-sum test, run directly against the outcome data you provide.

What it does

Given a control group and one or more variant groups of outcome data, scripts/ab_stats.py:

1. Computes lift: (variant_mean - control_mean) / control_mean.
2. Computes a bootstrap confidence interval on that lift (seeded resampling, default 10,000 iterations).
3. Runs a significance test:
   • metric: "conversion" (binary 0/1 outcomes) -> two-proportion z-test.
   • metric: "value" (continuous outcomes) -> Mann-Whitney U test, with the rank-sum computed by hand (tie correction + continuity correction).
4. Applies the dual promote-gate and returns PROMOTE, HOLD, or REJECT with the numbers behind the call.

When to use

Invoke this any time a promote/hold/reject call needs to rest on a computed p-value and a real lift interval. Typical callers:

• The closed-loop-marketing discipline, at the point where it decides whether to promote a winning variant to full rollout.
• Any A/B or multivariate test report on marketing copy, pricing, onboarding flows, or ad creative that ends in a promote/hold/reject call.
• Any experiment dashboard that flags a result "significant" without also checking whether the lift clears the bar for shipping it.
• Any place a lift number is about to be quoted without the confidence interval or p-value behind it.

How to run it

echo '{
  "control": [0, 1, 0, 0, 1, 0, 0, 0, 1, 0],
  "variants": {"B": [1, 1, 0, 1, 0, 1, 1, 0, 1, 0]},
  "metric": "conversion",
  "alpha": 0.05,
  "min_lift": 0.15
}' | python3 scripts/ab_stats.py

Or from a file: python3 scripts/ab_stats.py experiment.json.

Input JSON fields:

Field	Required	Default	Notes
control	yes	--	array of outcomes (0/1 for conversion, numbers for value)
variants	yes	--	object of {name: [outcomes]}, one or more
metric	no	"conversion"	"conversion" or "value"
alpha	no	0.05	significance threshold
min_lift	no	0.15	minimum relative lift to promote (15%)
bootstrap_iterations	no	10000	resampling count
seed	no	1234	RNG seed, for reproducible bootstrap output
confidence	no	0.95	bootstrap CI width

Output is JSON, one result block per variant, each with lift, bootstrap_ci, test (which test ran), z, p_value, decision, and a human-readable reason.

How the promote-gate decision works

A variant is promoted only when both conditions hold:

• p_value < alpha (statistically significant), and
• lift >= min_lift (the effect is large enough to matter).

p_value	lift	Decision
>= alpha	any	HOLD -- not enough evidence yet
< alpha	>= min_lift	PROMOTE -- ship it
< alpha	<= -min_lift	REJECT -- variant is significantly worse
< alpha	between -min_lift and min_lift	HOLD -- significant, but the effect is too small to act on

That last row is what the min_lift floor exists for: at large enough sample sizes, a trivial lift can clear p < 0.05 on significance alone, and the floor stops that from reading as a win. See scripts/test_ab_stats.py case (d) for a worked example.

Validation and tests

Run python3 scripts/test_ab_stats.py from this directory. It runs each fixture below, prints PASS or FAIL for every assertion, and exits 0 only if all of them pass.

• a hand-verifiable Mann-Whitney U case (no ties, fully separated samples)
• a clear-winner case that must PROMOTE
• a no-difference case (identical distributions) that must HOLD
• a large-sample, small-lift case that is statistically significant but must still HOLD -- proving the dual gate does what it's for

Caveats

• The Mann-Whitney normal approximation is asymptotic; on very small samples (roughly n < 10 per group) it will disagree somewhat with the exact permutation p-value. It gets closer as sample size grows, and the tie/continuity corrections keep it well-behaved in the meantime.
• ab_stats.py fails with a non-zero exit and a message on stderr for bad input (missing fields, wrong types, empty arrays, non-binary values under metric: "conversion") -- it never silently substitutes a default or drops bad rows.

Attribution

The idea of pairing a bootstrap lift interval with a promote-gate for marketing experiments is drawn from the growth-engine skill in ericosiu/ai-marketing-skills (MIT licensed). This implementation is an independent clean-room reimplementation in pure Python: the statistical routines (rank-sum Mann-Whitney with tie/continuity correction, two-proportion z-test, the bootstrap resampler, and the dual-gate logic) were written from scratch for this skill and share no code with the original repository.