Why a published forecast cannot answer your question

The subscription

After the 36,000 pieces went to a competitor, Marrash did the sensible thing and bought data. A market service at USD 4,800 a year: national apparel trade figures by month, plus an outlook for the category. Two years of it came to USD 9,600.

This lesson is what Marrash found when it checked whether the money had worked. The conclusion is not that the service was bad. It is that a published number and a factory's decision are different shapes, and no amount of accuracy in the first will make it fit the second.

Adding numbers up cancels the thing you need to see

Start with the arithmetic. It settles the question before any argument about data quality.

Three factories supply the same market. Last season each shipped 100. This season they ship 130, 85 and 88.

The total goes from 300 to 303. It moved 1.0%. The three factories moved +30.0%, −15.0% and −12.0%.

Not one of them moved anything like 1.0%. Anyone reading the total would report a flat market. A total is a sum, and a sum cancels out movements that go in opposite directions. So a total is at its quietest about exactly the thing one supplier is trying to see: its own share moving.

This is not a criticism of the data. It is how addition works. A number that describes many firms cannot describe one of them, and the better it describes the many, the less it says about you.

Five steps down, and what they do to the answer

The published figure covers a country, all categories, all price points, all seasons, all buyers. Marrash's decision covers one buyer, one category, one price point, one season, one delivery window. Getting from the first to the second takes five steps, and Marrash has to guess at every one.

Marrash was honest about its guesses and put a band on each one. It reckoned it could get each of the five shares right to within 20%. That is a confident merchandiser, not a gloomy one.

Stack five of those guesses and the top of the band is 2.488 times the published figure. The bottom is 0.328 times it. The top of the range is 7.6 times the bottom.

The number you started with was accurate. The number you ended with spans a factor of nearly eight. And the decision it was meant to help is worth eight line-weeks of work.

The geography is not your geography

Marrash competes for Ovenden's autumn jersey against four factories. One of them is in Jordan. Three are not.

So the national figures contain one of the four firms that could actually take this order, mixed in with several thousand firms that could not. Whatever the figures say, they are an average dominated by companies that were never in the room.

A data service splits the world by country because that is how customs authorities split it. That is not how a buyer splits it. A buyer splits it by who can hit the date at the price. That group has four members, and it does not match any border.

The clock

The service publishes a month's figures 68 days after that month closes. Marrash makes its capacity decision for weeks 30 to 33 on 19 March.

The figures describing August, the month Marrash is deciding about, come out on 7 November. That is 233 days after the decision was due, or 33.3 weeks.

Lesson 1 measured a 127-day relay and called it the shape of a supply chain. This is a 233-day relay, and it sits after the decision instead of before it. A number cannot help a decision it arrives eight months behind, however good it is.

And everyone else has it too

The last problem is the one nobody mentions in the sales meeting.

The subscription is a product. The three other factories quoting for Ovenden's autumn jersey can buy the same product for the same USD 4,800, on the same day, and read the same chart. Whatever it tells Marrash, it tells them.

A signal anyone can buy cannot be an edge, because an edge is something your competitor does not have. What it can be is insurance against surprise. Being the last of four to hear something really is worse than being the second. That is a real benefit, and it is worth deciding whether it is worth USD 4,800 a year. That is a different question from whether the data is good.

The audit

Marrash asked itself one question about the two years: which decisions did we make differently because of this subscription?

None. It found one conversation with Hallgarten that went better because the owner knew a shipping figure. It found zero decisions changed.

Over the same two years it found three decisions changed by the behaviour log from lesson 3, which cost nothing but a column and a habit.

Cost per decision changed: for the log, zero. For the subscription there is no ratio at all, because you cannot divide by zero. That is the honest answer, not a debating point. Marrash kept the subscription and moved it out of the planning budget into the marketing budget, where knowing things for conversations belongs.