Nowcasting, weather and seasonal forecasting: which horizon fits which decision?

A date that will not move

Every February the world buys roses, and a good share of them are grown on the slopes of the Ecuadorian highlands. For a stem to arrive fresh on the 14th, the bulk of the cutting happens in mid-January; and for a stem to be ready on 15 January, its cycle was scheduled some twelve to sixteen weeks earlier. Said plainly: the biggest decision of the year on a flower farm is taken in October, when Valentine's Day has not yet come up in a single conversation. (One clarification before we go on, because in English the word forecast quietly does two jobs. Everything here is about the weather: rain, temperature, light. The other forecast, the one that projects how many stems the market will take, is a different exercise and we set it aside from the start.)

That countdown (sixteen weeks, then days, then hours) answers a question we are asked often, and answers it better than any table would: which forecast horizon serves which decision? Let us follow the count.

October: the season sets the calendar

In October the farm's question covers a whole season: are the coming months going to run warmer or cooler than usual? This is not almanac curiosity. The speed of the rose cycle depends on temperature: a warm season shortens it, a cool one stretches it, and the production calendar is built by counting backwards from the week of the cut. Whoever schedules on the assumption of a normal season and then gets warm months arrives with the crop running early; and roses, once they are ready, do not wait.

Notice which forecast cannot answer this: the app on the phone, with its ten days ahead. The October question is a question about months, and no forecast of days reaches it, however good it is; consulting one here means letting the decision make itself, badly. What does speak in months is the seasonal statement, which says something different and says it in a different language: not "it will rain on 3 November", but "the coming months are more likely than not to be warmer than usual". Less detail, yes; exactly the detail the October decision needs.

January: the week gives the orders

January arrives and the question changes size. No longer the season: the week. Which days to cut, with how many people, when to fill the cold room, how to order the blocks so a stem waits as little as possible between the shears and the cold. Here the useful forecast is the weather forecast, the one covering the next few days: with it you pick windows, move a crew, pull a batch forward or push it back. The margins are short and well known: a cut rose lives two or three weeks if the cold chain is respected, so every cutting day chosen well or badly travels intact all the way to the buyer's vase. The week of the cut does not tolerate October's vagueness; it needs days with names on them.

And here the mirror image of the October mistake appears: demanding Tuesday from the seasonal statement. "They said warm season and Tuesday came up cold" does not describe a failure; it describes somebody asking a claim about months to speak about a day. It never spoke about Tuesday, and it could not have.

The night: a question of hours

One leg of the countdown is left, the shortest. The greenhouse on a highland night: if the minimum comes in lower than expected, the screen goes down in time and with somebody there to lower it; if the small hours turn mild, that night shift is saved. It is a decision of hours, and only a forecast of hours serves it: nowcasting, which tells you what is about to happen over your greenhouse tonight, not this week. (Whether that night shift is worth paying for is a question of threshold, and we already did that arithmetic in the previous guide; what this stage adds is the clock.)

The same decision, passed from hand to hand

Seen whole, the countdown teaches something the definitions on their own do not: the three horizons are not alternatives among which you pick one. They are the same decision moving down a relay. October prepared the calendar, January scheduled the week, the night lowered the screen; and each leg lived in a different horizon for a simple, almost physical reason: every action takes a different amount of time to carry out. Rescheduling a production cycle takes months; organising a week of cutting, days; lowering a screen, hours.

And as in any relay, each runner inherits whatever the one before handed over. October does not choose the day of the cut, but it decides which days will be available to choose from; January cannot repair a calendar that October scheduled wrong, however fine its forecast; and the night only saves what is still savable in hours. No horizon replaces another: each one works inside the margin the previous one left it.

Your operation sets the horizon

In the previous guide we said that the threshold is not set by the meteorologist, and we promised to come back to how long each action takes to prepare. Here is the rule, and it also fits on one line: the forecast horizon you need is not a matter of taste and is not assigned to you by the meteorologist; it is set by the lead time of each action in your operation. Measure how long passes between the order and the fact (months, days or hours) and you will know which forecast serves that action and which one arrives too late for it, however good it is. That inventory of actions with their times is your company's map of horizons; and, like the threshold, you already have it, even if you have never written it down.