Choose the comparison that answers the question
A prior week, forecast, matched store and randomized holdout are different sources of evidence. A previous period can be operationally convenient while differing in seasonality, stock availability, distribution or competitor activity. A forecast can account for some differences, but its training information and error still matter.
If an experiment is feasible, specify the assignment unit and analyze at that unit. NIST's experimental-design guidance describes randomization and blocking as design tools; matching a product or store after an event does not, by itself, recreate random assignment.
- Was the product available in both windows?
- Did selling locations, assortment or customer eligibility change?
- Did the weekday mix, holiday calendar or distribution change?
- Was another offer running on the same product or substitutes?
- Was the comparison chosen before seeing whether it made the campaign look good?
Normalize duration without overclaiming
MarginTrial can multiply baseline units by promotion days / baseline days when both window lengths are supplied. This aligns exposure time arithmetically. It assumes that the source window's average daily sales are a useful benchmark for the campaign window.
A seven-day comparison drawn from a 28-day month does not inherit the month's weekday mix, seasonality or stock position automatically. If those differences are material, prepare an adjusted expected volume outside the calculator and document how it was obtained.
Worked example: a window error reverses the interpretation
A source period contains 2,800 units over 28 days. A seven-day promotion sells 1,000 units at €18, versus a €20 regular price and €8 unit cost, with €200 fixed spending. The duration-aligned baseline is 700 units.
| Comparison | Baseline contribution | Promotion contribution | Change |
|---|---|---|---|
| Correct duration: 700 baseline units | €8,400 | €9,800 | +€1,400 |
| Unaligned: all 2,800 baseline units | €33,600 | €9,800 | −€23,800 |
Record the evidence in the worksheet
The aligned-baseline template contains the numeric example. Keep the comparison source, exclusions and known differences in the report note or an accompanying evidence record. The CSV day fields describe lengths, not calendar dates or a causal-control design.
| Worksheet item | CSV field | Example |
|---|---|---|
| Source baseline quantity | baseline_units | 2800 |
| Source window length | baseline_days | 28 |
| Campaign window length | period_days | 7 |
| Gross campaign quantity | promo_units | 1000 |
Stress-test the decision
Calculate the result with a lower and higher plausible no-promotion volume. If the sign changes within that range, report the threshold and the evidence needed to narrow it. A hand-chosen range is a sensitivity analysis; it is not a statistical confidence interval.
Do not silently reduce baseline units to account for returns if the economic benchmark also needs different costs or refund treatment. The workbench does not separately model baseline refunds, so a material baseline return process needs a fuller comparison.
Common baseline questions
Can I use last year's matching week? Only if the relevant differences are addressed. The same week number is not proof of comparable availability, customers or holiday timing.
What if both day fields are blank? The workbench assumes that the supplied baseline is already aligned to the promotion window. Supply both lengths together when normalization is needed.
Put it to work
Source context
Guide text and examples were prepared independently for MarginTrial. Sources provide the stated methodological or product context. They do not validate the hypothetical commercial outcomes.