Nobody stopped caring. They stopped buying.
Halloween candy sales ran down 6.7% in the month to October 17 and down 11.7% in mid-September. The usual reading is that the shopper is tapped out and the category is softening. We measured the demand side, and that reading is wrong in a way that changes the 2027 plan. Five Halloween search terms against their own six-year averages, four comment corpora with every theme registered before the pull, and a costume control that decides whether the price anxiety is about this category or about the economy.
The occasion is at a six-year high. What you buy for it is not. Candy is the thing people protect, and candy still lost 6.7%.
September searches for “when is halloween” ran 35.7% above their own 2019–2024 September average and “halloween candy” 15.5% above. Over the same window decorations fell 58.6% and costumes 36.4%. Across 409 comments about trick-or-treating, with five terms registered in advance, “fewer” and “skipping” each appeared zero times.
The decision this is for
A brand manager at a confectionery business writing the 2027 plan, with the director above them approving it. The question in the room is whether to take price again, hold, or rebuild the entry point through pack.
The registered question was narrower and decides the rest: did the occasion shrink, or did the basket shrink inside an intact occasion? A wallet problem is price-pack architecture and recoverable volume. A ritual problem is a three-year brief that no pack size fixes. A shrinking occasion and a shrinking basket look identical in a sales number and different here.
The occasion, measured against itself
Seasonal terms cannot be read on rolling twelve-month retention: a term that peaks every October contains exactly one peak in any twelve-month window and reports near-total retention whether the category is healthy or dying. The comparison that carries information is the same month, year over year.
| Term | What it measures | 2019–2024 mean | Sept 2026 | Change |
|---|---|---|---|---|
| when is halloween | do people know it is coming | 271,167 | 368,000 | +35.7% |
| halloween candy | the category | 95,250 | 110,000 | +15.5% |
| trick or treating | the ritual itself | 82,750 | 74,000 | -10.6% |
| halloween costumes | a thing you buy | 2,878,333 | 1,830,000 | -36.4% |
| halloween decorations | a thing you buy | 727,500 | 301,000 | -58.6% |
The further a category sits from the ritual itself, the harder it has been cut. Decorations first, costumes next, candy last.
2025 is excluded from every baseline. September 2025 for costumes printed 7,480,000 against a 2019–2024 range of 2.24m to 3.35m, with identical Augusts either side — the base year is an anomaly. The raw year-over-year fall is 75.5% and we are not using it. 2025 is dropped for all five terms rather than only where it would have helped.
Nobody is talking about not going
| Term | Comments | Share |
|---|---|---|
| kids | 16 | 3.91% |
| tradition | 2 | 0.49% |
| neighborhood | 2 | 0.49% |
| fewer | 0 | 0 of 409 |
| skipping | 0 | 0 of 409 |
A household that was going to skip does not buy a small bag; it buys nothing. Only small bags grew this season. Large and extra-large fell hardest. That is a price-pack signal, not a demand signal.
The price anxiety is about this category
Five price words were registered before the pull and counted against both corpora. The costume row is the control, and nothing else on this page is readable without it.
| Term | Comments | Share |
|---|---|---|
| price | 31 | 5.84% |
| expensive | 26 | 4.90% |
| afford | 23 | 4.33% |
| inflation | 10 | 1.88% |
| shrinkflation | 3 | 0.56% |
| Term | Comments | Share |
|---|---|---|
| afford | 2 | 0.65% |
| expensive | 0 | 0 of 307 |
| price | 0 | 0 of 307 |
| inflation | 0 | 0 of 307 |
| shrinkflation | 0 | 0 of 307 |
17.51% against 0.65%. A factor of twenty-seven.
Same holiday, same week, same five words. This is not a general affordability mood being read onto candy. Candy and gum prices rose 44% over five years against 22.5% for all food eaten at home, and the shopper has noticed about this aisle and not the one beside it.
Where the value-seeking went
| Term | 3-year mean | Sept 2026 | Change |
|---|---|---|---|
| candy in bulk | 5,133 | 9,900 | +92.9% |
| candy clearance | 907 | 1,300 | +43.4% |
| candy deals | 1,307 | 1,600 | +22.4% |
| cheap halloween candy | 2,567 | 2,900 | +13.0% |
| halloween candy costco | 7,700 | 8,100 | +5.2% |
| halloween candy target | 3,600 | 3,600 | +0.0% |
| halloween candy sale | 5,600 | 5,400 | -3.6% |
| halloween candy walmart | 8,700 | 8,100 | -6.9% |
| bulk candy | 40,233 | 33,100 | -17.7% |
| halloween candy coupons | 210 | 170 | -19.0% |
Generic value-seeking rose; retailer-named and coupon terms did not. That split was observed after the data arrived and is recorded as an observation. It does not revive the prediction it failed.
The search shelf is not your shelf
Five retailer domains were registered before the pull as the ones expected to hold page one for halloween candy: walmart.com, target.com, costco.com, amazon.com, instacart.com. They hold none of it.
| # | Domain |
|---|---|
| 1 | vermontcountrystore.com |
| 2 | allcitycandy.com |
| 3 | cvs.com |
| 4 | twix.com |
| 5 | dylanscandybar.com |
| 6 | tasteofhome.com |
| 7 | foodnetwork.com |
| 8 | apartmenttherapy.com |
| 9 | nuts.com |
CVS is a retailer and we failed to register it — part of that miss is ours. The finding survives the error: the big boxes hold nothing, page one belongs to specialty sellers and recipe publishers, and twix.com is a brand site ranking where the retailers are not. Meanwhile, in 531 comments, the most-named place to buy is Walmart at 4.90%, then dollar stores at 3.58% and Costco at 3.01%. “dollar” matches both a retailer and a price and was registered knowing it overcounts.
The other side of the same question
Measured separately, on a separate instrument, and published October 25: the price increases did not reach margin either.
| Filer or group | Detail | Change |
|---|---|---|
| 15 filers outside the category | the control | +1.99 |
| 42 food and household filers | the category median | −0.92 |
| Mondelez | FY2019 40.0% → 28.4% | −11.59 |
| Hershey | FY2019 45.4% → 33.5% | −11.81 |
Cocoa rose 109% between 2022 and 2024 while soybean oil fell 27%. Costs did not fall after the peak; they split. So the increases went to covering inputs rather than to margin, and they still cost the category its volume.
And the category already ran the other experiment this season. Non-chocolate Halloween candy averaged $6.08 a pound in the nine weeks to October 17, about 3% below last year — and sales did not rise. One caveat, stated plainly: 3% is a small cut against a 44% five-year build, and nothing here locates the threshold at which a larger cut would move units. A plan that reads this as proof elasticity is dead is over-reading one data point.
What we registered before the data was pulled
Five predictions, each labeled with what it was worth, all written before anything was retrieved. Two came out wrong and one held on a band it could not have failed.
| # | Prediction | Band | Basis | Result | What happened |
|---|---|---|---|---|---|
| H1 | September search within 10% of last year | likely | belief | Held, and the band was defective | Flat — 110,000 against 110,000. But Google bands these figures and the rungs either side of 110,000 are 90,500 and 135,000, so the instrument’s own resolution there is about ±10%. A band no wider than the noise cannot fail informatively. |
| H2 | Every price-seeking term at or above its 3-year September mean | likely | belief | Missed | Four of ten below. The falsifier was written strictly and fired. Generic value-seeking rose; retailer-named and coupon terms did not — an observation made after the data arrived, recorded as one, and not scored. |
| H3 | Candy price vocabulary clears 5% of comments | roughly even | belief | Held | 17.51% across 531 comments. |
| H4 | Costume price talk runs under half the candy rate | likely | measured-narrow | Held | 0.65% against 17.51% — a factor of 27. The only prior in this pass grounded in an earlier measurement, and the most decisive result in it. |
| H5 | Registered retailers hold half of page one | likely | belief | Missed | Zero of nine. Part of the miss is ours: CVS is a retailer and was not registered. |
The full record of every prediction we have registered, including these, is at the prediction ledger.
What this study cannot tell you
- Search intent is not purchase. NielsenIQ and Circana measure units sold; this pass does not and cannot check their figures.
- Theme frequency is not sentiment. “price” appearing in 5.84% of comments says the subject came up, not which way anyone felt about it.
- October is absent. The engine excludes the incomplete current month, so the season’s own month cannot be measured until November. A re-pull costs $0.011.
- Google bands every search figure here. Directional and multi-band moves are safe; a single-band move is not resolvable. Every change in the September table crosses at least one band.
- Comment corpora are drawn per row and the denominators differ. Each is printed beside its rate, and no share is compared across rows that do not share one. On a pass five days earlier the same query returned 973, 973, 653 and 101 comments across four rows.
- A YouTube comment corpus over-represents people who chose to comment, and on a shopping subject it skews toward deal and haul channels — which biases toward the price finding, not away from it. Stated before the pull.
- The margin evidence is a different study on a different instrument. Joining it to these demand figures is an argument, not a measurement.
- Costumes are a durable purchase and candy is a consumable, so the control is not perfect. A factor of 27 survives that objection where a factor of 2 would not have.
How this page changes
A standing study, not an article. Corrections are published in place and dated, never applied quietly.
| Measurement | Cadence |
|---|---|
| September search, all five terms | Annually, each September |
| October search — the season's own month | As soon as October closes, $0.011 |
| Comment corpora and the costume control | With each re-pull, denominators printed |
| Page one for “halloween candy” | With each re-pull, retailer list corrected to include CVS |
| Date | What changed |
|---|---|
| 2026-10-26 | First publication. Two registered predictions came out wrong and are published above. A third held on a band no wider than the instrument's own resolution and is reported as a defective prediction rather than a clean hit. The raw year-over-year fall in costume search is 75.5% and is not used anywhere on this page, because the base year is an outlier against its own history. |
October is the month this question actually turns on, and it is not measurable yet.
Sound Decisions carries the re-pull when it lands, and the registered question is written before the data is.
Subscribe to Sound Decisions