A forecast we built, and do not show you

Prism Microclimate shows a prediction only if it beats the simple guess on days it has never seen. One of our own soil-moisture forecasts failed that test, so growers never see it. This note explains the rule, and why we think every field forecast should follow it.

The simple guess is harder to beat than it looks

For many field conditions, the simplest possible forecast is surprisingly good: tomorrow will be like today. Soil moisture usually changes slowly. A frost-prone low spot was frost-prone last week too. A forecast that cannot do better than "no change" is not adding anything, however sophisticated it is.

So before any prediction reaches a grower, we ask one question: over days the model was never trained on, does it make smaller errors than simply assuming tomorrow is like today?

How the test works

We build a prediction from older data, then score it on the most recent stretch of days, which it has never seen. We also score the simple guess on those same days. The prediction is shown only if its errors are smaller.

Testing on unseen, later days matters. A model can look excellent on the data it learned from and still fail on tomorrow. Testing in time order is the closest thing to how it will actually be used.

What happened to our own forecasts

We built soil-moisture outlooks for more than one depth.

• At one depth, the outlook passed. Its errors were about 8% smaller than the simple guess. That is a modest improvement, and we describe it as modest.
• At the shallower depth, it did not pass. Its errors were no better than assuming no change. So it is not shown to growers at all.

We also tried predicting how much colder frost-prone parts of a field get on calm nights, using the shape of the land. It did worse than simply assuming no difference. We recorded that result and do not use the model.

A failed forecast is still a result. Hiding it would make the ones that pass harder to trust.

Why this matters on a farm

A number on a screen looks equally confident whether it is good or not. That is the problem. A grower deciding when to plant, spray or run equipment over a field cannot tell a tested prediction from a hopeful one by looking at it.

The test does not make a forecast perfect. It makes a narrower promise: anything shown has earned its place against the plainest alternative, on days it never saw. This is not agronomic advice; for decisions on your own fields, check with your agronomist.

If you are weighing up any field forecast, from anyone, three questions help:

1. What does it beat, and by how much?
2. Was it tested on days it had never seen, in time order?
3. What did the provider build that did not work, and did they stop showing it?

Where Prism fits

Prism Microclimate is Prism's own product for North Dakota: field-level weather, soil moisture, frost, spray and field-work windows, and growing degree days for any of the state's 961,439 field boundaries. It runs on the same engine as everything we make, which answers only from stored evidence; see why AI answers need sources. For background on the field conditions themselves, read microclimate: why the nearest weather station is not your field and frost risk explained.