How we know which weather models perform best

I have answered this before but each year I learn more. So here is an update.

The best way to judge a weather model isn’t by a single forecast — it’s by tracking its performance over time.

We look at how often a model gets:

• Timing right (when rain or storms occur)

• Placement right (where it actually happens)

• Scale right (isolated storms vs widespread rain)

In the very short term — like this afternoon or tonight — high-resolution models that focus on current observations usually perform best. For broader patterns over several days, global models tend to do better because they capture the big picture.

Why models can never be perfect

Weather models can’t be perfect because, in theory, they would need to map every molecule of air, moisture, heat, and motion in the atmosphere — which is impossible. What we can do is forecast within a certain radius and confidence range, not an exact street-by-street outcome.

That’s why forecasts talk about risk, likelihood, and ranges, not guarantees.

Why rain falls unevenly

Just like the experiment we did with the grandkids today, clouds don’t behave neatly. They form in irregular shapes, store moisture unevenly, and release rain through weak points.

That’s why:

• One suburb gets drenched

• The next one over barely gets wet

• Even in a heavy downpour, rain doesn’t fall evenly

• This isn’t forecast failure — it’s how clouds actually work.

The danger of pretending long-range certainty

What we shouldn’t do is pretend we can forecast 8, 16, or more days out with precision, then only point to the times we were right and quietly ignore when we were wrong.

As a teacher, I tell students this:

> Mistakes are how you learn.

If you only accept when you’re right, you kid yourself — and mislead those who trust you.

Owning uncertainty builds credibility. Ignoring it destroys it.

What timeframes really mean

Here’s what the science supports:

0–24 hours:

• Difficult, especially for storms. Small changes make big differences.

1–4 days:

• Reasonably accurate for large synoptic systems because most models agree on the pattern.

5–11 days:

• Increasingly uncertain. Outcomes can change run to run as models “chase” evolving conditions.

12–16 days:

• Based more on broad atmospheric signals. Useful for trend watching, not specifics.

30 days:

• A blend of signals and historical patterns. Not a forecast — more a climate-style outlook.

3–7 months:

• Primarily driven by large-scale climate signals and historical behaviour. These show leanings, not outcomes.

The bottom line

Weather forecasting isn’t about being right all the time.

It’s about being honest about uncertainty, learning from misses, and improving over time.

And sometimes, the best explanation comes from a glass of water, some shaving foam — and a curious kid asking why.