Weather forecasts have become more accurate
ourworldindata.org
ourworldindata.org
Rather than letting some aggregator simplify the weather for you, you can just look at the raw data yourself: https://weather.cod.edu/forecast/
For big events, the media briefings by the National Weather Service are good resources. But they often stop the briefings early; a few weeks ago we had a high probability of a large amount of snowfall. The updates stopped at like 9AM, the snow was forecast to start around 1PM. Watching the short term models showed that the probability for snow was decreasing (NYC was just below the snow/rain line), and indeed we got pretty much no snow. (It snowed, but it didn't accumulate and the change to rain happened early.) To be fair, the briefing from the weather service said that the changeover time between snow and rain was very uncertain and that it would be the difference between a little rain and major snow event. But my point is, you can always go get yourself some more data; the closer you get to the event, the more accurate the forecast is.
(I don't know if any of you watch Skip Talbot, but he was looking at helicity swaths on the HRRR a few hours out, found a big one, and where HRRR predicted the strong rotation in the storm is pretty much exactly the path of a major tornado. HRRR is never going to be perfect, but it is right a lot.)
>”These observations are then fed into numerical prediction models to forecast the weather.”
In other words, the forecasts come from models, not necessarily real-time station readings. Those readings are inputs into the model, and the models may not get updated fast enough to reflect current conditions.
Windy.com lets you compare different models for a specific location, it also includes the size of the area per model: https://www.windy.com/?49.339,5.054,5
GFS is area is 22km, ECMWF 9km, ICON-D2 2.2Km, Arome 1.3Km, and UKV is 2Km. Even in a 1.3x1.3Km area it may not rain everywhere at the same time.
And then there's also the time element, so it's 1.3Kmx1.3Kmx1Hrs (or 3Hrs). So lot's of variation possible.
"Hey Siri, is it going to rain?"
"It doesn't look like it's going to rain today."
"It's raining right now."
"It isn't raining right now."
So unless you are sitting next to the the weather station that Siri is getting data from, I would not expect it to know 100% of the time.
I don't, but as a result, I expect it not to guess.
I live in Toronto, Canada, which stretches about 40km east-west, and 20km north-south:
If the west-end (Sherway) gets hit with rain, but the east-end is dry, did it rain "in" Toronto when folks in Scarborough didn't experience it? Was the forecast wrong?
If it snows in North York but is dry at Billy Bishop, was the precipitation forecast "wrong" for one particular group of people?
Meanwhile the Google Weather app constantly insists I live in Frankfurt while I'm in Warsaw.
On a serious note I'm dealing with this as well - I live in a middle-sized city in the centre of the country, not in Kraków!
The technical limitation would be on the weather data size: what is the granularity/resolution of the radar data on where rain is actually falling?
* https://en.wikipedia.org/wiki/Canadian_weather_radar_network
Further: what is the geography of the area, and how does that effect things as well? Toronto specifically has (a) all sort of heat island effects, (b) certain areas are effected by the lake and how weather systems cross it at certain angles, and (c) has enough of an elevation change going north of the lake (e.g., Niagara Escarpment) that there are a few ˚C change in temperature that makes the differences between snow and rain.
The serious answer is that the way you'd try to figure this out is by combining weather radar, satellite imagery, and a nearby surface observation to try to estimate the current conditions. But there can be a latency of up to a few minutes from these sources, and they could disagree with one another. You have to use them to bootstrap your near-term or nowcast product, but enforcing consistency with recent real-time and the nowcast is quite hard.
It's a surprisingly nuanced technical challenge. Most of the time, it works out just fine (e.g. if there is no weather). But people are awfully good at remembering when these sorts of analyses end up being wrong!
If you're interested in providing on-the-ground condition reports, install mPING: https://mping.nssl.noaa.gov/
I keep this app on my homescreen and try to report when very light rain starts, since it's not always obvious from the reflectivity data. Ultimately the user reports get fed into things like improving the model, and more data is always good.
You'd think that if their users are accepting that level of communications with the mothership that they could ship some AI model to hear rainfall in the wild, and thus improve their live weather data.
The thing is, even if you did have a super reliable in situ "rain detector", how do you combine it with the existing datasets like weather radar, which is a gridded product? This is actually a really, really difficult sensor fusion problem when you then super-impose product requirements like the general location real-time detection map and the inputs necessary for whatever internal nowcasting system they use.
Actual ground observation weather stations are fairly rare outside of places like airports and major news stations.
“Is it raining here right now?” Is a harder question than you’re giving it credit for. Radar can show rainfall at as low as a couple thousand feet altitude, but if conditions are right/wrong (depending on how you look at it) it never reaches the ground.
But rain or other objective information? I suppose it works, maybe a bit like “limit IP address tracking” — cloudflare or other edge provider could mediate so Apple gets the data, knows it comes from an iPhone (to prevent bad data attacks), but Apple can’t tell what phone sent which data.
(the privacy concern being documentation of when you were inside/outside/etc).
But this sort of technique only works for medium range forecasts. Short-range precipitation nowcasts are almost always a single, deterministic run of a model that extrapolates from patterns in recent radar imagery. They aren't bias corrected at all, so you can't use observations in the same way to improve them.
And the ship has been towed beyond the environment.
There is nothing out there, all there is is sea, and birds, and fish. And 20,000 tons of crude oil. And a fire.
Less sarcasticaly speaking I think there is always weather. Maybe what you mean is “no significant change in the weather” neither in time, nor in space.
I've also noticed that Met.ie will typically predict more rain, and they're usually right. (e.g., last weekend was basically rain/drizzle/wind the whole time, met.ie nailed it, apple weather said that there would be an hour on Sat and all Sunday morning would be wet.
Of course, predicting rain in Ireland is not difficult.
[0]: https://windy.com
But things like "rain in X mins" is a feature multiple providers & apps have (including Apple once they bought Dark Sky), it's not specifically what Dark Sky was nor is it exclusive to them/Apple. (And actually, Dark Sky was probably the best weather app all round, yet Apple despite buying them and using some of their tech still produce one of the worst weather apps in my experience.)
It's subscription based though.
Is that some idiom or some meme reference? Unfamiliar.
Google changes it to "road to Oz" when I search.
I had read the original book, which is different, as a kid, but not seen the film.
Wow, tornados. That is something I would never want to be stuck in. I have experienced a few cyclones and earthquakes, including a major one.
Three day forecast in the 80s and probably early 90s are about where, crap, 15 days out is, actually the 15day is probably better.
Modern forecasting long term identifies the front movements, really well long term, even if they might be off a day and 5degrees.
The old forecasts would be completely off.
Source 50something
It's June in Texas, so we'll just say the 10-day is going to be sunny, hot, no rain. It's California in June, so we'll just say warm with June gloom burning off in the afternoon; warmer to hot inland. I didn't use any science data, and my forecast will probably have just as good of a chance as one that did.
Nothing like the old Aggie weather station consisting of a chain suspending a rock. If rock is wet, it's raining. If rock is moving, it's windy.
HRRR is OK, but usually <10 hours for better numbers.
You should relllllly look at the soundings, though.
So, I don't care if tomorrow there's a 50% chance of rain. I care that at a precise time of the day , say 9am, has a 10% precipitation and at noon is 90% , because i commute at 9am, not at noon. Wind is also an important factor if its raining. Temp as well. I need all this info presented as a mosaic.
For this purpose, I find the NOAA forecast local by hour is unrivaled. https://www.weather.gov/okx/ . Enter ZIP and then in enter local forecast by hour.
I have this URL bookmarked in my browser. I haven't looked back since. Example:
https://forecast.weather.gov/MapClick.php?lat=33.797&lon=-11...
I'd love to know if there is an android app that gives this level of detail, preferably, without spying into my microphone...
which is interesting, as i'm noticing the "within 15 minutes" level of notice on rain starting/stopping to have been close enough. the daily forecast last week said no rain even though the conditions really looked like it could at any moment. my iDevices pinged with rain starting soon even though the same apps forecast still did not suggest rain. it started raining with in "good enough" range of the app's notifications.
the update to the native weather app have all been very good over the past 2 OS updates. maybe they have integrated whatever company they purchased for good, but for my local area on the globe, it has been pretty good. i haven't traveled in a good while, so maybe my market is in the sweet spot of getting a lot of attention??? BigD in case you're wondering
This website allows you to select which weather model you want to use: https://www.pivotalweather.com/model.php?fh=loop&dpdt=&mc=&r...
> In The Weather Machine, Andrew Blum takes readers on a fascinating journey through an everyday miracle. In a quest to understand how the forecast works, he visits old weather stations and watches new satellites blast off. He follows the dogged efforts of scientists to create a supercomputer model of the atmosphere and traces the surprising history of the algorithms that power their work. He discovers that we have quietly entered a golden age of meteorology—our tools allow us to predict weather more accurately than ever, and yet we haven’t learned to trust them, nor can we guarantee the fragile international alliances that allow our modern weather machine to exist.
* https://www.andrewblum.net/the-weather-machine-2
* https://www.goodreads.com/en/book/show/42079139
For the very early history of meteorology, see perhaps The Invention of Clouds about Luke Howard:
* https://www.goodreads.com/book/show/1148768.The_Invention_of...
To get what people judge to be a ‘good forecast’, the chance of rain has to be adjusted to be wildly too high - so that’s what consumer-focused forecasters do.
I've come to think of that as "it is going to rain 50% of the time. I don't know if that's what really is meant by "50% chance of rain," but it seems to fit.
And overall I tend to believe that the forecast is astonishingly accurate. This is in the Midwest (Chicago market) where weather has to cross large portions of the country or Canada before it gets to us. I suppose there are areas on the coast where weather is more volatile and harder to predict.
So it of course won't rain for exactly 50% of the time on a given day, but over the long run, it will.
"The probability of precipitation (POP), is defined as the likelihood of occurrence (expressed as a percent) of a measurable amount of liquid precipitation (or the water equivalent of frozen precipitation) during a specified period of time at any given point in the forecast area. Measurable precipitation is equal to or greater than 0.01 inches. Unless specified otherwise, the time period is normally 12 hours. NWS forecasts use such categorical terms as occasional, intermittent, or periods of to describe a precipitation event that has a high probability of occurrence (80%+), but is expected to be of an "on and off" nature."
The percentage chance of rain includes whether or not it might rain in your specific dot of a given forecast area, which might be a suburb or entire city, as spelt out in your quote "at any given point in the forecast area".
The first time I drove over the Nullarbor in Australia, which has an entirely flat and straight 100+ KM section of road, I got to see rain far in the distance and experience driving into it having been able to clearly see it's edge from far out. That was an experience I had never had in the costal city I live in (Perth). That also led into similar realisations as the above.
It sounds so simple in theory but was not obvious to me for a long time :)
In an urbanized area most "is it going to rain?" questions are short-term, e.g. is now or 30 minutes later a good time to bike home?
Perhaps this wouldn't be as useful in other areas. The Netherlands gets very spotty rain. So even if you've got a 100% chance today it's probably 1-2 hours spread throughout the day, and sometimes very heavy rain followed by a dry spell.
The only time I've seen it to be incorrect is if a moving rain cloud just barely misses you due to changes in wind patterns.
Never really thought about it, but I've opened the "Rain radar" more frequently than any weather app including the native one during the last couple of years, too.
Though I don't know anyone else who does this that doesn't cycle.
Of course, there are countless ways for the for the forecast to be wrong, and only a couple ways for the forecast to be right!
Less so in FL where drought begins 4 minutes after the last rainfall. The 13th month of summer can have us begging for days w/o the migraine-making cancer ball.
I am not affiliated, but I recommend checking out https://www.forecastadvisor.com/ to see what forecasts are best for your city. I totally changed weather providers and it seems much better now.
'The Secret World of Weather: How to Read Signs in Every Cloud, Breeze, Hill, Street, Plant, Animal, and Dewdrop' by Gooley is a fun read for anyone interested in figuring out weather without a forecast (or to supplement).
The local apps pull data from the Japan Meteorological Agency, so does Apple Weather, and so does Carrot Weather since a recent update (though those 2 still give me different results). Outside of Japan, when I travel, I have no idea so I just leave the Carrot Weather source on Apple Weather, because that at least pulls data from local weather services if available (https://developer.apple.com/weatherkit/data-source-attributi...)
What a great recommendation, it’s sadly US only. I have previously used an app called Climendo which claimed to digest over 15k forecasts and use the most accurate one in my city.
It doesn't get a ton of press, but as this article highlights, progress has been steady and significant.
This article asserts that improving forecasts in low-income countries is underrated--does anyone know of studies that predict the impact better forecasts would have? Helping the poor with tech seems like the kind of project that many philanthropists could get excited about, and hopefully more effective than gravity lights and the like.
Modern technology is amazing.
Any replacement for it on iOS? Maybe I am crazy but Apple's weather alerts just don't seem like the same sauce.
Before Apple bought them, my Android phone was its own party trick at the bar. I'd be able to tell people down to the minute when it would start and stop raining. It was amazing for bar hopping on bad weather days.
Switching between providers on Carrot, Apple Weather often doesn't predict any amount of rain for the entire week, meanwhile I'm soaked in water in a thunderstorm, and NOAA and others predicted rain the entire week (which it did).
Apple Weather will tell me it won't rain today or all week.
Meanwhile NOAA will tell me I'm currently in a thunderstorm and that it will rain all week - And it was right.
Carrot is nice because you can switch between several providers.
The future of weather forecasting is likely to rely heavily on AI models. The article discusses Pangu Weather and HN comments mention GraphCast as examples. Interestingly, on the first of March, the European weather forecast center ECMWF released their new AI weather model AIFS as open data. This model is not only more accurate than their existing numerical model, but also requires significantly less computing power to run. They've published comparisons showing AIFS outperforms other models in terms of forecast precision: https://www.ecmwf.int/en/about/media-centre/aifs-blog/2024/f...
I've seen similar things in our area (Minnesota) where you drive through a snowstorm, but the radar shows nothing in theare.
I can't see how any weather predictor could be correct in that situation.
The latter is somewhat common because the models (AFAIK) use probabilistic estimates, where different initial conditions generate potentially distant outcomes. The number of “rainy outcomes” defines the probability of rain, and doesn’t necessarily get updated with real conditions.
Fwiw, I agree with your bemusement and scorn - it's not good enough! (I say this as someone who has had roles where I issued these 'always stale' forecasts)
We just had to make a prediction for the next day's weather. Then compare our prediction to the forecasted prediction. It didn't matter how accurate we were for the grade. Just that we systematically performed this exercise.
It really made one appreciate the quality of forecasts, and that no, the "weather man" is not always wrong at all. A lot of huffing and puffing is from people who lack any rigor in their observations. And if you're trying to contest the accuracy of weather forecasts, or any form of forecasting really, then you really should provide some hard evidence.
I think it was in that episode where one said that every 10 years we improve the forecast by 1 day.
It was recorded in 2019, so AI wasn't really that much of a topic as it is today, considering that Google published an AI weather model in November of last year [1].
[0] https://omegataupodcast.net/326-weather-forecasting-at-the-e...
[1] https://deepmind.google/discover/blog/graphcast-ai-model-for...
I plan my motorcycling based on rain, and the number of times I've gotten caught in rain when it wasn't supposed to rain at all that day is non-zero just this year.
That statement is pretty much only used as a thought-terminating cliche that means "you're not allowed to have an opinion".
Pointing this out is not an attempt to silence you.
1. High frequency 5G has thrown off rain forecasts in urban areas. Average prediction accuracy has still improved because rural/suburban areas don't have high frequency 5G.
2. The weather app now shows rain forecasts in time blocks as small as 15 minutes, even though predictions this granular are still inaccurate. This has inflated our expectations for forecast accuracy.
Further, it's only the upper range of high frequency spectrum that's being used (not sure who owns it) so it's not even every carrier that could interfere.
Finally, the most powerful radars are transmitting in the kilowatts range of output. It's hard for me to imagine that the microwatt output of cellphones are often the cause of radar interference.
The net is the improved raw accuracy of the weather forecast is offset by the difficulty of reversing the clickbait layer slathered on top.
(1) is irrelevant for weather forecasting.
A few years ago I was able to stop my friend's outdoor wedding (on the terrace as opposed to the hall, the venue had both ready) from getting rained out by reading the radar and catching a small pocket storm that had formed and coming right towards us. Sure enough it down poured, but everyone was inside for the ceremony. Reading just the weather report, there wasn't even rain forecasted.
"I remember reading in The Signal and The Noise* that people _think_ that forecasts are bad if it rains, but the chance of rain was reported as below 50%. Getting rain when the forecast told you there would probably not be rain is annoying; getting a sunny day when the forecast predicted likely rain is a pleasant surprise. To get what people judge to be a ‘good forecast’, the chance of rain has to be adjusted to be wildly too high - so that’s what consumer-focused forecasters do."
Unfortunately, the most important part of any forecast IMO is intensity. I don't care if we're going to get snow flurries all day, but if we're going to get a foot of snow, I would like to know -- and not just when the winter storm warning goes into effect!
Similarly, I don't care if we're going to get scattered showers all day. But if we're going to get a downpour in the afternoon, I'd like to know so I can avoid getting caught in a flash flood on a trail or on the road.
Same thing applies with temperature: if it's going to be cold all day, good to know. But if a rainstorm is going to remain active during a deep freeze and create a layer of ice on every exposed surface, I need to be prepared for walking, biking, or driving.
Fortunately there's a somewhat local weather station near me that provides an RSS feed of longform weather forecasts. But I notice that more and more people wind up surprised by slightly-abnormal weather events as they rely more and more on smartphone weather apps. Weather apps that utterly lack the nuance that a paragraph of text can provide.
We objectively know that weather forecasts are more accurate than ever. We subjectively know that they are bad/gotten worse, because last Thursday I brought my umbrella to work for nothing.
https://weather.gov/ still works (type your zip code into the box on the top left), thankfully.
I’m worried weather.gov is only one election cycle away from being decommissioned.
Lobbyists from commercial weather sites nearly got the US weather service killed under Trump.
If/when they finally kill NOAA, global shipping will probably collapse, which is why killing it was blocked last time.
But, although you get good spatial coverage, the drawback is 'the map is not the territory' - the model's representarion of reality doesn't perfectly mesh with the weather on the ground.
I am not sure if this has to do with radar capability but all the old time hams seem to corroborate this.
I use both, a bit, but I don't know much about weather science, so can't evaluate, except by comparing it with the real, for which both do seem to be at least somewhat accurate.
More warming == more energy in the system.
More energy in the system -> more volatile weather.
More volatile weather -> harder to predict weather.
But don't forget! That doesn't mean nice weather is counter-evidence of climate change. Nice weather is /also/ evidence of climate change, because it's merely the lull before the weird weather.
Got it yet?
Also, it's worth noting that GraphCast's outputs are a tiny subset of what we would traditionally forecast as weather parameters. You can't out-perform a competitor on a task that you aren't solving!
Where are your data to back this claim up? And over what time horizon?
I haven’t found an app or tool outside of building a grafana dash that beats it.
I paid far less attention to weather forecasts 30 years ago than I do now, but I have numerous anecdotal examples of how weather forecasting models and information provided by publicly available weather services have trended towards uselessness.
There is no publicly accessible weather information service that can accurately forecast weather at my house. One of the first purchases I made when I moved in to the house was an Ambient Weather Station resulting from pure curiosity that has evolved into an interest in keeping a historical record of "actual weather". Daily hi/low temperatures generally have positive correlation with forecasted temperatures, but the spread between forecasted temperatures and actual temperatures is generally ten degrees less than forecasted.
Long term qualitative temperature trends ("above average for the winter" and similar) are positively correlated.
But ...
- Forecasted storm intensities are wildly inaccurate. Forecasted high-intensity rain storms end up being all-day drizzle events or on and off rain showers, and visa versa. A forecast of “a passing afternoon shower” ends up being an all-day wash-out.
- Precipitation forecasts are wildly inaccurate, without correlation. Actual precipitation can be far less than forecasted or far more than forecasted, even when compared to short term forecasts--to include same day and intrahour forecasts. Just this past weekend we had accumulating whiteout snow squalls on an off all day long on Sunday, yet there was never any mention of any possibility of snow by any local meteorologists or by any weather forecasting service I routinely check.
Dark Sky was the best app I ever used for weather forecasting. Its short and long term forecasts were more than sufficient for planning purposes, but where the app to this day has had no equal was in its intrahour local forecasts and precipitation forecasts. If Dark Sky alerted me that there was going to be tornado in my area within the next 15 minutes, I saw a funnel cloud 15 minutes later. If Dark Sky alerted me that it was going to stop snowing in 15 minutes, the snow stopped 15 minutes later. Sadly, Apple lobotomized the service when they claimed to have integrated Dark Sky functionality in to Apple Weather. Even though I fairly regularly report weather accuracy issues to Apple via the Weather app, the reporting and forecasting provided by Apple Weather has never improved.
- Seasonal precipitation forecasts are wildly inaccurate without correlation. Modeling (from NOAA, local meteorologists, etc.) suggested we were to have "above average snowfall" this winter, with the official average winter snowfall being 48 inches. We have received 20 inches so far this winter. Either winter will go out with a bang in the next few weeks (which would be nice, IMO), or modeling will have predicted more than 140% of the actual snowfall. This is an altogether unfair comparison, but why not: if the executives of a publicly traded company forecasted 140% more revenue to shareholders than the company they preside over realized, they would all be immediately fired, sued, jailed, etc.
If society collectively will not tolerate 140% inaccuracy in financial matters (stock price manipulation, value destruction, and so forth), should we be content with weather forecasting and modeling that is just as inaccurate? After all, weather is treated as (only) a financial matter by insurance companies. On an individual level, viewing weather's impact through financial optics still makes sense--from lost days of work and lost wages, to insurance premiums, to food prices, to transportation costs, to taxes, to paying for the ability to get your money back for a concert ticket you bought months ago if the weather is too bad.
Climate change is certainly wreaking havoc on weather modeling, but it has been doing so for a significant period of time and the models do not appear (to me) to be getting better at adequately accounting for the effects of climate change. If current weather forecasting models cannot be adapted to accurately account for the effects of climate change, it may be time to either fundamentally change the way weather modeling and forecasting is done, or not do it at all. Taking out my broad brush and bucket of paint: are there any companies relying on AI to develop a more accurate weather forecasting service?
And if anyone has a weather service to recommend that will not “Night at the Roxbury” me with ads and that has accurate 3-day-or-less weather forecasts, I am all ears. Please post them here.
> are there any companies relying on AI to develop a more accurate weather forecasting service?
Sure there are. But AI isn't a silver bullet, and existing weather forecasting technologies are _really freaking good_. For all of the hullabaloo over AI-NWP systems like Google's GraphCast and Huawei's PanguWeather, these state-of-the-art systems are about _on par_ with the best-in-class existing numerical weather models; they offer incremental improvements in tuned forecast accuracy, but these improvements are statistical descriptions of a very, very large number of forecasts - end users really wouldn't see any practical difference in forecast quality if they relied on these forecasts. But to my point above - even AI-NWP outputs would be filtered through statistical post-processing to boost their accuracy/utility.
There are a lot of companies that _claim_ they use AI at different parts of the weather value chain to improve forecasts. A lot of them stretch the truth as to what extent they really use AI or ML. The simple reality is that the weather community has used ML since the 1970's to improve weather forecasts.
Consumer weather prediction isn't about being right. It's about pleasing the customer by appearing to be helpful. Which often means exaggerating the chances of abnormal weather, so if it happens you can be a hero.
Real prediction is boring.
E.g. Darwin in Australia's tropics - persistence forecasting (as you describe above, just predicting the weather the day before) does very well on a metric like 'mean absolute error'. But has no practical skill at forecasting a severe tropical cyclone (aka hurricane/typhoon)! Many are willing to accept some level of false positives and a higher mean absolute error, because the cost of a surprise cyclone is so devestating.
I imagine using the previous day would have a much lower skill score in more variable climes.
I also couldn't find a link to this, and if you have one I'm interested in reading more.