The figures are already there, the answer is not

History, and enough of it. Two years is the practical minimum once there is seasonality involved, because with one year the model has seen exactly one summer and cannot tell whether that is the pattern or the weather that year.

And then the trap: only the columns you genuinely have at the moment of forecasting. This sounds obvious and it almost always goes wrong. You want to predict whether a customer will leave, and the export contains a column “reason for leaving”. It is filled in for a leaving customer and empty for a staying one, so the model becomes brilliant, and it is worthless: at the moment you want to use it that column is still empty.

So I go through the columns with you one by one with the same question each time: did you already know this at the moment you needed the forecast? Everything with a no goes out, however much it helps.

Not with a random split of the rows, because then the model trains on next week and tests on last week. It is measured backwards through time: learning on everything up to a date, predicting what happened after, and shifting that forward a number of times. That measures what the model would have done on a Monday morning.

And it is set against the dumbest reasonable rule. For busyness that is usually “this week is like the same week last year”. If the model does not beat that, you do not need a model but that rule, and that is an outcome I would rather send you home with than a dashboard. Surprisingly often that simple rule is tougher than people expect.

The result is a range rather than a number. “Between 180 and 240 visitors” is usable; “211 visitors” is false certainty that you take the wrong decisions on.

Information from the future. The trap above, and it is so often the cause of a result that is too good that a model which works well immediately makes me suspicious before it makes me happy. “What do we actually know at this point?” is the most important question of the whole project.

A model that predicts the average. If you ask how many customers will leave and three percent leave per year, then “nobody leaves” is right ninety-seven percent of the time. That is the same pitfall as with image recognition in a different coat, and the answer is the same: do not look at one number but at what you catch and what you miss.

And a break in the history. A move, a price change, a year in which the world was different: the model takes that along as a pattern. Sometimes the answer is to leave that period out, sometimes to include it separately. That is a conversation, not a setting.

Not for decisions about people. Creditworthiness, pre-selecting job applicants, assessments about health: those uses fall partly into the high-risk category of the AI Act, with obligations a one-person business cannot carry and consequences you do not want to explain when it goes wrong.

And not when the question is really “why”. A forecasting model says what is likely to happen, not what causes it. If you want to know whether your action caused something, that is a different kind of research, and a model is the wrong tool.