Fifteen cereal products priced one at a time on 169,677 store-product weeks. Nine cleared to rise, two held, and four refused outright.
A grocery chain can raise prices, or it can keep volume. Nobody in the business could say which products would survive a rise, because nobody had measured what a price move actually costs in volume. I measured it product by product and store by store, then turned each answer into the one number a finance director can act on: how much volume that product's own margin can afford to lose.
Three questions, in the order a finance director asks them. Which prices can move. By how much, before the volume loss costs more than the rise earns. And how do we find out cheaply that we were wrong, instead of expensively?
Prices are never set at random. They are raised into strong demand and discounted into weak, which is exactly the thing that makes a naive elasticity worthless. So the first question is not what the number is. It is where the number comes from.
A hundred products each carrying a recommended price reads as more work and less evidence. Every product here lands in one of three levels, and the level decides what may be done with it.
What qualifies. A precise estimate, sitting clear of the margin line, on enough price movement to trust it.
What happens. Act on it. Raise in steps, or hold where the price is already at its optimum.
What qualifies. The direction is supported but the size is not: a margin too thin to carry the loss, or an estimate inherited from the category rather than measured on the product.
What happens. Move 5% and re-measure before anything larger.
What qualifies. The estimate reads between −0.41 and −0.59, which no cereal sustains. The number is not usable, and a precise wrong number is worse than a missing one.
What happens. No price is recommended, and none of their money is in the headline.
An elasticity is not a constant. It moves with competition and with the health of the category. So the category was fitted twice, on the first 24 months and on the last 24, and the difference was priced rather than described.
No price goes chain-wide on a coefficient. The first move is small, reversible, and designed to fail loudly if it is wrong.
Test stores read against controls of near-identical revenue. One pair differs by 0.1%.
Same promotions, same displays, same shelf. A display week moves volume up to 73% on its own, and one slipping in would corrupt the read.
Any product breaching its affordable-loss threshold two weeks running reverts to its old price. That product only. No meeting.
One product stays at its old price everywhere. If it moves, something other than our prices is moving the market, and the read is flagged.
The same shape on a client file. Two days deciding whether the data can honestly carry the study, before anything more is billed.
Stated here rather than left to be found.
This public file carries none. Optimal prices and dollar figures re-price the day real costs arrive. The price responses themselves do not move.
The data sees this chain only. If a whole category rises sharply, some shoppers change store, which is part of why the moves are capped and staged.
Cereal is a known-value aisle. Whether a flagship carries a +30% optimum is a commercial decision this analysis informs but does not make.
Four products were described as unmeasured in the text and their money was still inside the headline, because the code computed a figure for every row and the judgement lived in a footnote. Honesty written beside a number is not honesty. It has to be enforced where the number is produced, or it will not survive the next rollup.
The optimum, the first rung, and what sits between are three different numbers. Quote one of them alone and you will be held to it in year two. So all three appear together, every time, and the first one asked for is always the smallest.
Not the fit statistics, not the model choice, not the forecast accuracy. Where the price variation comes from, and whether it is caused by the demand it is supposed to explain. Everything else is downstream of that one question, and it deserves its own page.
An elasticity moves between years. A price anchored to what the business has actually charged does not. Capping the recommendation inside the observed range is what makes a moving coefficient survive contact with a real price list.
Start with the question, not a proposal. If your export cannot support the study you had in mind, that is worth knowing in twenty minutes.