Halloween, measured
Did Halloween really sell, or is it noise in October?
A few good days at the end of October are easy to see. Whether they were Halloween, and whether November paid for them, is not. Measure what Halloween added, with its interval, and keep Black Friday out of it.
Invitation only while in beta. Tell us who you are and we get in touch.
Weekly sales, with and without Halloween
late September to November, sample retailerAn illustration on one of our sample markets. On your file the line without Halloween is estimated from your own years, and every number comes with its confidence interval.
Four ways to misread Halloween
Some reports find nothing in October, others blame Halloween for November. Both read the wrong weeks.
The weeks before
Costumes, sweets and decorations are bought in the two weeks before 31 October, not on the day. A Halloween dated only on the 31st misses most of its own effect.
The week after
Some of the October spending would have come in early November anyway. The first week after Halloween gives a part of it back.
Black Friday’s waiting
From mid November shoppers wait for the discounts. That dip is Black Friday’s, and a month-on-month report charges it to Halloween.
Small effects, wide bands
A lift of a few percent over three weeks sits close to the noise. The interval says whether your file can see it at all, and “no measurable effect” is an answer a budget can use.
From the booked lift to the real one
Start from what the Halloween weeks sold over the line, take back what the week after lost, and leave Black Friday’s weeks to Black Friday. On the sample retailer, of the booked lift was new. Small, and real.
The grey weeks of November are not in the sum: they are measured on the Black Friday page, with that event’s own dates.
Where the booked lift went
sample retailer, 2026The path in tea
Three analyses, on the same file, each answering one part of the question.
-
Seasonality & Event Impact
Declare Halloween with the days before it, and Black Friday on its own dates. Each gets its effect with its 95% interval; windows that overlap are flagged, because two events on the same days cannot be told apart. How it works
-
Promo Elasticity
If Halloween came with a discount, the lift per point of discount comes out net of the weeks after. It needs at least three promotions in the file, and says so. How it works
-
Contribution Decomposition
If a campaign ran, a Halloween column enters the model next to media, price and the season, so the October weeks are split between them instead of all going to the event. How it works
What your file needs
Two Halloweens at least
Two years of rows, daily if you have them: an effect of a few days blurs in a weekly file.
date day
The number to explain
Sales, orders, units. One column, the same on every row.
sales orders
What else moved
Discount depth, media spend, price. Optional, and each one is a part of the answer.
discount search_spend
Halloween is yours to date, with the days before it, and the Data Scout tells you before the run if the file is too thin for the question.
Halloween incrementality, in a paragraph
The incremental effect of Halloween is the difference between what the weeks around 31 October sold and what they would have sold without it. For most categories it is small, and the honest measure comes with a band: an effect whose interval crosses zero is a finding, not a failure.
The window matters more than the day. Declared with the two weeks before it and the week after, Halloween takes what is its own; declared next to Black Friday, it stops being blamed for November.
How the event model reads a calendar, and where it misleads: Seasonality and event impact, explained.
Does Halloween affect sales outside sweets and costumes?
Often it does not, and that is a result worth having. Measured with its interval, a Halloween effect that crosses zero says the event is noise for your category, and the October budget can go elsewhere.
Why is November lower after Halloween?
Partly because of Halloween, which pulled some purchases into October, and mostly because shoppers start waiting for Black Friday. The two belong to different events: a model that dates both separates them, a month-on-month comparison charges Halloween with Black Friday’s waiting.
How many weeks does Halloween last?
For sales, about three: costumes, sweets and decorations are bought in the two weeks before 31 October, not on the day. Declaring only the day misses most of the effect; declaring a window, with the days before it, catches it.
How much data do I need?
Two years at least, so two Halloweens, and daily data helps: an effect of a few percent over a few weeks sits close to the noise, and the interval says whether the file can see it.
What does it cost?
During the beta, nothing: every run is measured and no balance moves. After the beta, every plan includes a monthly token allowance: 150 on Free, 3,000 on Pro, 6,000 on Plus, 20,000 on Business. What is left of it expires when the month renews; tokens you buy never expire. The plan never limits which questions you can ask.
Where it goes next
Give October the budget it earns, measured.
Drop the file, measure the event, keep the number that holds. Free while in beta, by invitation.
Join the betaLearning
A report says what happened. An analysis says why.
Most of what lands on a marketing desk is reporting. Useful, and a different job. The blog tells one case a day where the difference decides the answer.
Sales were up 11% in April, TV ran six weeks. What happened, by channel, side by side.
4.2 of those points came from TV, 5.1 from Easter moving. What caused it, and what the next euro does.
When a budget optimiser says move 60% into one channel
Rolling cross-elasticity: checking whether a rivalry is stable
