We forecast Amazon's October sale before it happened

The usual way to check a tool like ours is to take a question whose answer is already known, run the tool on it and report how close it came. Even when the tool was kept blind, the answer was sitting somewhere while it ran. So this time we did it the other way round: we wrote the forecast down before the event, locked it, and waited for the real numbers to come out.

average miss on the six shopper percentages

1.5 pts

average miss on the two dollar figures

4.1%

average miss for the same AI asked the questions directly

16 pts

What we forecast

Amazon held Prime Big Deal Days on 6 and 7 October 2026. The shopper-research firm Numerator tracks every one of these sales: it asks real shoppers what they did and publishes the answers. Did they know about the sale before buying? Did they compare prices at other stores? Did they place more than one order, buy holiday gifts, feel the deals were good? How much did their household spend?

On 1 to 2 October 2026, Planetarium forecast the answers. Numerator published its final figures on 8 October 2026. We scored the forecast against them on 22 October 2026. Nothing was re-run or changed in between.

How the forecast was locked

A forecast is only worth something if nobody can quietly change it later. So before the sale, the forecast was saved, given a fingerprint (a short code that changes if a single number in the file changes) and committed to our benchmark records with its date. The written plan for the test said which forecast would be judged, and how, before any real figure existed.

The forecast was given Numerator's published results from earlier Amazon sales, only what was public before this one, and a correction for spending that was worked out on the 2023 to 2025 sales. The model inside was Gemini 3.8 Flash.

What happened

Here is every figure Numerator published, against what we said beforehand.

Every percentage landed within three points of the real one. Household spend came in 4.3% low: we said $109.15, shoppers spent $114.01.

Against other ways to guess

A close number means little until you know how close an easy guess would have come. So we scored three other ways to answer the same questions.

• Ask the same AI directly. The question alone, with no audience behind it. It missed by 16 points on the percentages, about 10 times further off. It said 21.4% of households would place two or more orders; the real answer was 54%.
• Copy June's Prime Day. 4.8 points off on the percentages and 14% on spending.
• Copy last October's sale. 1.5 points off on the percentages, as close as the forecast, and 7.5% on spending, where the forecast was closer at 4.1%.

That last one matters, and we would rather say it plainly than have you find it. Shoppers at this October sale behaved a lot like shoppers at last October's. A forecast that simply repeated last year would have done as well on the percentages. Where the forecast pulled ahead was money.

Two other forecasts we locked

We locked two other versions before the same sale, to learn from. An earlier version, from before the Amazon tuning, missed by 9.1 points and 41.6%. A general version, with no Amazon-specific tuning at all, was close on the percentages (2.7 points) but put household spend at $190.17 against $114.01, 57% off on the dollar figures. The plan named the tuned forecast as the one to judge before any result existed, and that is the one reported above.

What this does not prove

• It is one sale and eight figures. A good result on one event is a reason to keep testing, not a record.
• It did not beat every simple method. Copying last year matched it on the percentages.
• Two questions could not be scored. We also forecast how many shoppers were more careful with money because of the economy, and how many used AI shopping tools. Numerator did not publish those after the sale.
• It is shopper behaviour, not an ad campaign. It says nothing yet about how an ad or a landing page will perform.

Why we test this way

A forecast made before the event cannot be fitted to the answer, because there is no answer yet. That is the test we want Planetarium judged on. We will keep locking forecasts before events and publishing them when the real numbers arrive, including the ones that miss.