Weatherkind: A New Kind of Weather Intelligence

The app says sunny and 74, in a cheerful yellow circle. What arrives is hail the size of peas, delivered with real conviction, while you stand there in a t-shirt reconsidering your choices. Anyone who has planned a day around a mountain forecast knows the feeling. The app was not exactly lying. Somewhere within a few miles of you it probably was sunny and 74. It simply wasn’t where you were standing, and where you were standing was the only place that mattered.

That gap, between the weather a forecast describes and the weather that actually falls on your head, is why Weatherkind exists.

The map is not the mountain

To understand why your weather app does this to you, it helps to know a slightly depressing fact about how forecasts are made. The world’s great weather models divide the atmosphere into a grid of boxes and solve the physics inside each one. This is an extraordinary achievement, one of the genuine triumphs of twentieth-century science, and we don’t want to be glib about it. Predicting the behavior of a chaotic fluid three days out is closer to sorcery than most people appreciate.

But the boxes are big. Depending on the model, one box might be nine or thirteen or twenty-five kilometers on a side, and inside that box the model has to settle on a single elevation, a single surface, a single answer. In Kansas this is a defensible simplification. In Colorado, one box can contain a river valley at 7,900 feet, a ski town at 8,700, a lake, a south-facing slope of scrub oak, a north-facing wall of spruce that holds snow into June, and a ridge at 12,400 feet where the wind has never once stopped blowing. The model averages all of that into a number. An app then draws the number in a rounded rectangle, puts a little sun on it, and out you go in a t-shirt.

Meteorologists have known this forever. There is a well-worn term for it, representativeness error, which is a wonderfully bloodless name for a phenomenon whose real-world expression is somebody on a mountainside being hit repeatedly by ice.

The frustrating part is that this is a solvable problem, or at least a substantially reducible one. The model does not know your terrain, but we do. Elevation is mapped. Slope and aspect are mapped. The way temperature falls off with height, the way cold air pools in valleys overnight and refuses to leave, the way a ridge line wrings the moisture out of an approaching storm and hands the next valley over a dry, disappointing afternoon: none of these are mysteries. They are, in the most literal sense, the lay of the land.

What we actually do

Weatherkind starts with the same world-class model guidance everyone else does. We are not claiming to have secretly out-physicked the European Centre for Medium-Range Weather Forecasts in a garage in Colorado. What we do is refuse to stop there.

We take that guidance and correct it down to the ground you are standing on: your elevation rather than the grid box’s notional average, your slope, your aspect, your local quirks. In mountain towns we don’t even pretend a single number will do. We show the valley floor, the mid-mountain and the summit as separate tiers, because they are separate weather, and anyone who has driven up a canyon in a snowstorm knows it in their bones. We blend in observations from real instruments, including, if you have one, your own weather station in your own yard, because a thermometer forty feet from your front door has an opinion worth hearing.

Then we do the thing that we think genuinely sets us apart, and it is almost embarrassingly simple.

We keep score, and we show you the score

Every weather company on earth says it is the most accurate. The claim is usually supported by a study the company commissioned, conducted over a period the company selected, against competitors the company chose, with the results summarized in a press release. It is not that these studies are fraudulent. It is that they are marketing, and everyone involved knows it.

So we archive our forecasts. All of them, for every location in our verification network. Then, once the future arrives and stops being a prediction and starts being a measurement, we go back, compare what we said with what happened, and publish the result. Anyone can look at it right now. It is on the site: a rolling thirty-day record, hourly or daily, and you can open any row to see the mean error, the bias, and exactly how many comparisons went into it. If we ran two degrees warm in your area last week, that is in there. If we had a bad Thursday, that is in there too.

It is worth saying plainly why we do this, because “radical transparency” is the sort of phrase that has been drained of meaning by overuse. Part of it is simply that it is the right thing to do. But we also do it because it is the only mechanism we know of that reliably makes a forecast better. A number you publish is a number you have to live with. It creates a scoreboard, and a scoreboard creates the kind of pressure that fixes real problems: the persistent cold bias in a particular valley, the overnight low that is always too high in October, the model that is brilliant in flat country and hopeless in a canyon. You cannot fix what you refuse to measure, and the industry’s long habit of not measuring in public has, we suspect, cost everyone years of progress.

There is a second reason, and it is a practical one. A published record is a promise that is expensive to fake. Anyone can say they are accurate. Publishing the score whether it flatters you or not is a costly signal, and costly signals are the only ones worth much.

Weather is a decision, not a number

The other thing we believe, and it shapes nearly everything we build, is that nobody has ever actually wanted to know the weather. What people want is to know what to do.

Take the early flight or the late one. Send crews out at four in the morning to salt the lot, at considerable expense, or stay in bed. Pour the concrete or don’t. Bring the jacket or risk it. Every one of these is a decision made under uncertainty, and a forecast is useful only insofar as it improves the decision. Which means the honest thing to hand you is not a confident-looking icon but the shape of the uncertainty: how likely, how much, how soon, and how sure anyone can be.

That is why the app shows minute-by-minute precipitation when rain is closing in rather than a single percentage for the whole afternoon, and why the radar renders at sub-kilometer resolution instead of the smeared porridge most consumer apps serve up. It is why Mission Control lets an operations team set real thresholds on real sites and be told when reality crosses them, rather than staring at a dashboard hoping to notice. It is why there is an API, and why we designed it so that software agents can query it as comfortably as a developer can. Increasingly, the thing making the four-in-the-morning decision is a program, and it deserves good data too.

It is also why we called the company Weatherkind, in the way you would say humankind. The word takes in the whole of something. All of the weather, not one number at a time: the drizzle, the downpour, the ordinary Tuesday, the ridge that makes its own afternoon. And it carries the second meaning we care about just as much, which is how all of that reaches you. Kindly. Told straight, with the uncertainty left in, the record open for inspection, and enough context to decide what to do next.

We are building this in the mountains of Colorado, which is either an act of extraordinary hubris or the only sensible place to do it, and quite possibly both. If you can get the forecast right in a state where the atmosphere is repeatedly and violently interrupted by geology, the rest of the world starts to look tractable. And if we get it wrong here, we will have nowhere to hide, because we will have published that too.

Thanks for reading. This is the first of what we intend to be many field notes, on forecasting, on verification, and on the odd and wonderful things we find in the data. Bring a jacket.