Slashdot Mirror
Modern Weather Forecasts Are Stunningly Accurate (theatlantic.com)
An anonymous reader quotes a report from The Atlantic: Meteorologists have never gotten a shiny magazine cover or a brooding Aaron Sorkin film, and the weather-research hub of Norman, Oklahoma, is rarely mentioned in the same breath as Palo Alto. But over the past few decades, scientists have gotten significantly -- even staggeringly -- better at predicting the weather. How much better? "A modern five-day forecast is as accurate as a one-day forecast was in 1980," says a new paper, published last week in the journal Science. "Useful forecasts now reach nine to 10 days into the future." "Modern 72-hour predictions of hurricane tracks are more accurate than 24-hour forecasts were 40 years ago," the authors write. The federal government now predicts storm surge, stream level, and the likelihood of drought. It has also gotten better at talking about its forecasts: As I wrote in 2017, the National Weather Service has dropped professional jargon in favor of clear, direct, and everyday language. "Everybody's improving, and they're improving a lot," says Richard Alley, an author of the paper and a geoscientist at Penn State.
Understanding months-long events like El Niño, for instance, has allowed meteorologists to go beyond the seven-day forecast. Alley, the Penn State professor, says that he is awed by the new models. Well-studied features of Earth's climate -- like the temperate Gulf Stream in the Atlantic Ocean -- emerge in computer models, even though developers have written code that only mimics basic physics. We are now surrounded by the products of these miraculous models. In 2009, a back-of-the-envelope study estimated that U.S. adults check the weather forecast about 300 billion times per year. Perhaps in all that checking we have forgotten how strange the forecast is, how almost supernatural it is that people can describe the weather before it happens. More than 1,000 years ago, the Spanish archbishop Agobard of Lyon argued that no witch could control the weather because only God could understand it. "Man does not know the paths of the clouds, nor their perfect knowledges," he wrote. He cited the Book of Job for authority, which asks: "Dost thou know when God caused the light of his cloud to shine? Dost thou know the balancings of the clouds ?"
Understanding months-long events like El Niño, for instance, has allowed meteorologists to go beyond the seven-day forecast. Alley, the Penn State professor, says that he is awed by the new models. Well-studied features of Earth's climate -- like the temperate Gulf Stream in the Atlantic Ocean -- emerge in computer models, even though developers have written code that only mimics basic physics. We are now surrounded by the products of these miraculous models. In 2009, a back-of-the-envelope study estimated that U.S. adults check the weather forecast about 300 billion times per year. Perhaps in all that checking we have forgotten how strange the forecast is, how almost supernatural it is that people can describe the weather before it happens. More than 1,000 years ago, the Spanish archbishop Agobard of Lyon argued that no witch could control the weather because only God could understand it. "Man does not know the paths of the clouds, nor their perfect knowledges," he wrote. He cited the Book of Job for authority, which asks: "Dost thou know when God caused the light of his cloud to shine? Dost thou know the balancings of the clouds ?"
I do not doubt that significant progress has been made in modeling weather and predicting some aspects of future weather. But so far this only changed weather forecasts from being "mostly random, not better than just predicting that tomorrow's weather will be just the same as today's weather" into the current "we can state some trend that is reasonably likely to be correct for the next few days". I can still read weather forecasts from yesterday evening in the news that say "0% probability of precipitation today for the city I live in", while I see rain falling outside the window.
"Stunningly accurate" would be a whole different thing, like the forecast being able to tell me "rain will start to fall at my location from 10:34h to 11:27h tomorrow" - and that they very much still cannot.
Here in Norway we have learned to depend upon https://yr.no/ which provides both short-term (2+ days) and long-term forecasts:
When the short-term forecast states that it will be 0.5 to 0.8 mm rain (or snow equivalent) between 10:00 and 11:00 tomorrow, and that it will clear up starting at 13:00, this is very likely to be correct. If it isn't exactly right it is usually because the changes happen a little bit before or after the maximum likelihood prediction.
The presentation of the weather data is so good that many people in our neighboring countries have started to use YR instead of their local weather service.
Terje
"almost all programming can be viewed as an exercise in caching"
I wish every weather service published a graph that showed the progress of the correlation between their 1 to 7 day forecasts and what actually happened, somewhat like the graph for hurricane tracks in the referenced article. Published confidence levels would also help to know how locked-in a prediction was.
My experience has been that forecasts a day or two ahead are amazingly accurate, but that you can't rely on forecasts a week out for scheduling an important event.
A big part of this is that forecasts and current conditions have vastly smaller location granularity than in the past. In 1980, for a given state, you'd be lucky to obtain specific forecasts for maybe 10-15 large cities in the entire state (less for smaller states). I'm sure most of you have seen weather where it rained at your house, but just a few blocks down the street they didn't get rain at all. When your forecast granularity is representative of hundreds of square miles, then of course you can never be very accurate for that entire area.
Now the forecast is latitude and longitude based, and the precision is vastly finer. That alone increases the accuracy tremendously. Weather forecasts now are also down to "minutely" (as in hourly or daily) time spans. Again, same thing. When your forecast broke the entire day into "night" and "day" periods, you can never be very accurate. Most weather apps now forecast what will happen in the next hour down to the minute ("Light rain will begin in around 12 minutes"). It's easy to be accurate when you can forecast such a small time into the future.
There are many reasons weather is more accurate now, everything from the lead time (if your forecast has to be in to the newspaper before 5 AM so it can meet the press deadline, then you're accuracy will be reduced compared to a forecast calculated the moment it is asked for), to the technology that allows people to ask for and view data when they want it for a very specific area.
Better known as 318230.
We have a lot more data and we have a lot better understanding of the correlation betweeen past and present conditions around the world and future contitions at any given spot than we did 40 years ago.
Yes, there has been a huge improvement.
No, I am not surprised.
Knowledge is how to play a game, intelligence is how to win, wisdom is knowing what game to play.
I see lots of people claiming they know better than the study but it doesnâ(TM)t seem to be based on any rigorous research.
The issue with complex modelling has traditionally been in computational power vs usefulness. With weather modelling, you can generate a very accurate prediction a week ahead. Problem with it was in 1980, you needed so much computational power, you would get this "forecast" finally calculated a few years after it was relevant. Computational power to do it in time frame that was useful was simply not there.
Today it is there, so meteorologists can get calculations for days ahead done in time frame which is useful, i.e. before the events they're trying to predict occur, rather than long after.
Models improved too, but most of that improvement has actually been "add even more detail to do something with all the increasing computational power".
In 2016, in the mid-Atlantic region, we had a mild winter. There was one snow event only - but that event dumped about 33-36 inches in the Baltimore/DC metro region. That event was informally called Snowzilla, and it was predicted to the tee, 8 days out. That was probably the most amazing forecast achievement I've seen.
Of course, the meteorologists still screw it up, sometimes fantastically. But other times, they knock it out of the park.
"Cold front will arrive next Tuesday morning and lows will be in the teens and highs in the 30s by Thursday afternoon."
That's funny that you put it that way, because that's how they USED (~ 1980s) to predict the temperature: in the lower 70s, mid-20s, etc. Nowadays, all the TV stations here in Chicago give impossibly exact numbers: 32 for the high, 14 for the low, etc.
I spoke to a guy at the UK met office about this in the 90s and he explained how they were basically compute limited. They run a number of sims with randomized perturbations at the start and see which outcomes are the most common across perturbations. They were using all their Crays full whack and that's what determined and limited the accuracy of the results. 30 years later, compute power is somewhat cheaper and my desktop is faster than one of those Crays.
I should use this sig to advertise my book ISBN-13 : 978-1501515132.