Why are rain weather predictions (apparently) so unreliable still in the year 2021?

Question

A year or so ago, I finally implemented a little mechanism which, every five minutes, fetches weather data from SMHI (Swedish official organization, having kept track of temperatures and stuff since the mid-1700s) for a GPS coordinate quite close to my house. (Less than 1 km, for privacy.)

Since I really love it when it rains, it then extracts all predictions of rain and puts them into my calendar which I constantly see on my screen, so that I can at a glance see exactly when rain is predicted for the current days and 13 more days into the future. (Obviously, each time, it deletes the old data first.)

But this is where it gets weird. It might say right now that it's going to rain sometime tomorrow, but yesterday, or hours before, it was five days until the next rain. And the very same day, it might say that it's going to rain in 10 minutes, but then it increases to 20 minutes, then an hour left, then 3 hours left, then no rain at all for 3 days. Then no rain for 7 days. Then rain tomorrow. It seems to vary so wildly as to be completely unreliable and useless.

It feels like I might as well just remove this mechanism as I could just as well just make my own random guesses and probably be as accurate.

But why is this? I thought that weather predictions were extremely reliable at this point, with global, advanced models processing enormous amounts of information with powerful computers around the clock and whatnot. Yet they seem to have no clue? Is predicting rain especially difficult, perhaps? Are just SMHI specifically very incompetent?

Answer 1

You are expecting too much. Weather forecasting has become quite reliable.

There are exceptions. One exception is predicting what will happen more than seven days into the future. At ten days into the future, forecasts remain pretty much a random guess. The 1960s discovery that weather is chaotic (Predictability: Does the flap of a butterfly's wings in Brazil set off a tornado in Texas?) remains a challenge, and acts as a limit to long-term weather forecast accuracy.

Another exception is predicting what will happen in the next few hours. The various high precision weather models used by weather agencies are only run a few times a day. In between, lower precision heuristic algorithms are used to fill in the blanks. Those high precision models run on some of the world's largest supercomputers, and it can take hours to complete a run. For example, the European Centre for Medium-Range Weather Forecasts runs its high resolution Integrated Forecast System twice a day. The ECMWF is not going to make a special-purpose run because you want a forecast for the next ten minutes.

Predicting when and where a little popup thundershower will occur remains highly challenging. For example, the ECMWF has a 9 km horizontal resolution. A popup thundershower might affect less than a square kilometer. Predicting when and where a squall line or a front will hit is a bit easier, and is much more important. The heuristics used for short term forecasts focus on the important things -- things that can cause damage. The little popup thundershowers that make your day or spoil someone's picnic does not qualify as such.

Answer 2

Weather prediction has been (https://www.ecmwf.int/en/newsletter/166/news/forecast-performance-2020) and will be getting better, but it can only get so good. Even if we build the perfect weather model and feed it with the perfect initial and boundary conditions (practically impossible at the moment), the forecasts will still be overshadowed by chaos and/or stochasticity at some spatial and temporal scales. Moist convection and precipitation are particularly prone to this, due to the involvement of a zoo of processes (inter)acting at scales from nanometres to kilometres and from seconds to days. Running the exact same model two times with slightly different initial conditions may produce two equally-plausible forecasts that disagree on whether it will rain or not in your neighborhood in 12 hours from now. The big forecast centres nowadays reflect on this inherent uncertainty by producing many forecasts at a time. In this regard, you are better off looking at probabilistic forecasts (e.g., 70% chance of rain) rather than deterministic (e.g., 25mm of rain) or categorical (e.g., rain vs no rain) ones. When you then assess the model performance, you should do that in a probabilistic sense as well (when the forecast says 70% chance of rain, does it actually rain 70% of the times?).

Unfortunately, even probabilistic forecasts fail at times and produce a range of outcomes that is far from what was actually observed. Associated with that, one should keep in mind that not all regions in the globe and weather regimes are equally predictable. Rainfall associated with a (synoptic-scale) cyclone and its organised frontal systems is easier to predict than rainfall associated with (small-scale) isolated thunderstorms. Wind over a large plain is easier to predict than wind over a mountainous region.

I am not sure about the SMHI model, but I suppose it is a competent high-resolution regional model focused on Scandinavia. If they do not issue probabilistic forecasts, you can look at global models like ECMWF, GFS, ICON. They are not tuned to Sweden, but this is not necessarily bad; an ensemble of forecasts (from a single or several models) will surely give you a better overview of the possible outcomes.

Regarding the short-term predictions of the next 0-6 hours, we are in the nowcasting realm (https://public.wmo.int/en/resources/bulletin/nowcasting-guidelines-–-summary). For this range, you should primarily look at real-time observations from radars, satellites, lightning networks, surface stations, etc; especially if there are enough of those near you. Some weather apps/services make a synthesis of these observations and extrapolate into the near future so that people know what actually comes their way.

Answer 3

Fun fact about rain percentages. When it says there is a 70% chance of rain, it does not always mean there is a 70% chance that you personally will get rained on. Weather stations are divided into sections. Depending on what app you use some predictions can be interpreted differently. But most of the time the 70% chance of rain means that there is nearly a 100% chance of rain in 70% of the area covered by the station that you are looking at. If there is less that a 100% of rain in the 70% area of coverage, a more reliable app will calculate the new average. I don’t know if there is a way to see the coverage area of any given station, but I can say that watching the radar predictions and coming up with your own prediction based on your exact location is much more reliable, albeit more time consuming.