Atmospheric Rivers - The Long And Short of It

Atmospheric Rivers - The Long And Short of It

(Originally posted January 10, 2017 on Blogger)

On January 3rd I posted a link and brief write-up about a recent peer-reviewed publication out of the Massachusetts Institute of Technology (MIT) on my Google+ profile. The paper describes how atmospheric river (AR) events over California could become more common over the next several decades. I felt the timing of the publication was almost eloquent as early model runs on the GFS indicated an AR event was in the works. I also mentioned that long-range forecasts are dicey and referenced Lorenz' Chaos Theory. I briefly wrote about Lorenz' theory in a blog back in August; that long-range forecasts suffer from algorithmic error amplification that exacerbates the further out from the forecast target date one gets.

Some odd runs on the GFS showed Armageddon-like precipitation (ppt) potential for California as far back as December 25th, but nothing was consistent. I thought these odd-ball forecasts that popped out of different GFS runs at random would provide me with a perfect opportunity to demonstrate to whoever reads my boring blogs just how unreliable long-range forecasts can be. In fact, I began collecting multi-level charts for the continental United States (CONUS) as well as daily weather stats for several major cities across the United States as well. These latter two data hoards will be used later this spring when I plan to share two separate studies about Gary Lezak's "LRC Hypothesis", and AccuWeather's 90-day forecasts. Stay tuned for those, as I think you'll find them entertaining. A thorough undermining of psuedoscience will be performed before your very eyes.

This blog is about the forecasts leading up to the recent AR event over the eastern Pacific into California. I'm already hearing folks exclaim how well the GFS did in forecasting this event, while others are upset because the GFS underestimated it. Let's simply look at exactly what the GFS did, and didn't do with regard to forecasting this impressive event, then look at some actual rainfall totals, and conclude with what I hope will be a decent explanation of why so many people continue to take long-range forecasts so seriously when they're most often wrong, and why we continue to see long-range forecasts in print, online, in apps, and on that dinosaur, televisionasaurus.

This will be tricky, as an AR event indeed happened... but I don't want to cherry pick in order to make a point about long-range forecasts. I think the situation I have observed between the last week of December and now may actually be better in explaining why long-range forecasts continue to dupe the masses, than if there was no AR event at all and those random Armageddon-runs from late December were just anomalies. So let's get started...

Sixteen days out from the AR event and I somehow was lucky enough to save forecasts for the right days and times; I don't know how I managed to do that... but I have them, and have made animated GIFs with them, so let's take a look... The following image is a collection of GFS forecast runs for 06Z January 8, 2017. I began saving runs from December 25th to January 7th and put them together in chronological order to demonstrate what consecutive runs look like greater than 5 days out, and what they look like within 5 days of the target forecast date:

FS SLP and Precipitation forecasts for 06Z January 8th. This is a compilation I put together of runs from December 25th to January 7th in chronological order. "Long Range" is greater than 5 days out from the AR event.

Notice how runs of the GFS could not get a handle on the AR event beyond 5 days. The images vary widely in depicting SLP and precipitation. It isn't until about 6 days out (the tail end of the long-range stuff) that we begin to see a consistent pattern showing an atmospheric river. The short-range runs had a better handle on things and consistently showed the AR. Now, I want to point out the very first image that shows up after the frame that reads "Long Range"... see image to the right >>.

THAT is what I first saw pop up on the GFS... THAT was the "Armageddon" forecast, the AR event of the century... A forecast that popped out of the GFS a whopping 336 hours out... and never seen again. In fact, the next two runs on the GFS had California bone dry. This is the inherent chaos in algorithms my friends. The GFS had no handle on this event whatsoever, but this monster run was enough to wake up sleepy weather nerds from Eureka to San Diego, and the online weather world discussions and arguments began! "Ahh these QPFs are junk!", "This is going to be the biggest storm in over 100 years", "This stuff will miss us completely"... and on and on and on the posts in nerdy forums go... I admit to reading them... kind of my version of awful reality TV, but I know better than to participate.

Let's take a gander at some more GFS runs, this time runs that forecast what was to come at 12Z on January 8th...

GFS SLP and Precipitation forecasts for  12Z  January 8th. This is a compilation of runs from December 25th to January 7th in chronological order. "Long Range" is greater than 5 days out from the AR event.

GFS SLP and Precipitation forecasts for 12Z January 8th. This is a compilation of runs from December 25th to January 7th in chronological order. "Long Range" is greater than 5 days out from the AR event.

Again, long-range (greater than 5 days out) the GFS is all over the place with what is happening. Monster single-cell storms, scattered moderate stuff, nothing at all, NorCal Armageddon... the GFS didn't know what the heck was going to happen! One day it's raining cats and dogs, and the next it's perfect beach weather! But then notice how things start to become consistent towards the tail-end of the long range stuff... you begin to see hints of an AR, and then nearly perfect consistency in the short range outputs where the AR is obvious. That my friends is long-range chaos math vs. short range statistically-reliable math. Irregardless of whether there was an AR event or not, long-range forecasts are notoriously unreliable and--if you read Lorenz' work--unavoidably so.

And let's have a look at the GFS forecasts for January 8th at 18Z...

GFS SLP and Precipitation forecasts for  18Z  January 8th. This is a compilation of runs from December 25th to January 7th in chronological order. "Long Range" is greater than 5 days out from the AR event.

GFS SLP and Precipitation forecasts for 18Z January 8th. This is a compilation of runs from December 25th to January 7th in chronological order. "Long Range" is greater than 5 days out from the AR event.

At this point, it should be no surprise as to what we're seeing... inconsistent long-range outputs that only begin to see some semblance of consistency towards the tail-end of the long-range, becoming realatively consistent in the short range.

I don't want to be redundant, but this stuff took a lot of time and patience to collect, and I'm going to post it dammit! Here's the GFS forecasts for January 9th at 00Z....

GFS SLP and Precipitation forecasts for  00Z  January 9th. This is a compilation of runs from December 25th to January 7th in chronological order. "Long Range" is greater than 5 days out from the AR event.

GFS SLP and Precipitation forecasts for 00Z January 9th. This is a compilation of runs from December 25th to January 7th in chronological order. "Long Range" is greater than 5 days out from the AR event.

And for Jan. 9th at 06Z...

GFS SLP and Precipitation forecasts for  06Z  January 9th.This is a compilation of runs from December 25th to January 7th in chronological order."Long Range" is greater than 5 days out from the AR event.

GFS SLP and Precipitation forecasts for 06Z January 9th.This is a compilation of runs from December 25th to January 7th in chronological order."Long Range" is greater than 5 days out from the AR event.

 Annnnnnnnd 12Z..

GFS SLP and Precipitation forecasts for  12Z  January 9th. This is a compilation of runs from December 25th to January 7th in chronological order. "Long Range" is greater than 5 days out from the AR event.

GFS SLP and Precipitation forecasts for 12Z January 9th. This is a compilation of runs from December 25th to January 7th in chronological order. "Long Range" is greater than 5 days out from the AR event.

Last one, I promise....a look at forecasts focused towards the end of the event (according to the GFS), here's Jan. 9th at 18Z...

GFS SLP and Precipitation forecasts for  18Z  January 9th. This is a compilation of runs from December 25th to January 7th in chronological order. "Long Range" is greater than 5 days out from the AR event.

GFS SLP and Precipitation forecasts for 18Z January 9th. This is a compilation of runs from December 25th to January 7th in chronological order. "Long Range" is greater than 5 days out from the AR event.

I think it's evident how unreliable long-range forecasts can be, and how much better forecasts become within a week of a target date. Now let's look at total precipitation forecasts. Since it was December 25th--16 days out--I didn't want to chance it by collecting total precip forecasts during the AR event, or worse... before it! ...so made sure to collect them for a date that would be well after (if an AR event occurred). So I semi-arbitrarily chose to look at how much rain areas would get by January 11th.

Total precipitation forecasts are put out by the GFS (and other models) with the intent on predicting how much precipitation will fall between "now" and the target date.. the target date in this case being January 11th. Each new run of the GFS puts out a new total precip forecast.. and as you'll see, the same chaotic nature is apparent in the long-range forecasts, becoming more consistent in the short-range...

GFS forecasts for total precipitation in inches on January 11th for the Los Angeles, CA area. These are taken from runs starting December 27th through January 7th. Note the chaotic nature of long-range forecasts (left side), some varying by over half a foot from previous runs, and how relatively more consistent the short-range forecasts become (right side).

GFS forecasts for total precipitation in inches on January 11th for the Los Angeles, CA area. These are taken from runs starting December 27th through January 7th. Note the chaotic nature of long-range forecasts (left side), some varying by over half a foot from previous runs, and how relatively more consistent the short-range forecasts become (right side).

GFS forecasts for total precipitation in inches on January 11th for the Las Cruces, CA area. These are taken from runs starting December 27th through January 7th. Again, the same pattern of wildly-varying forecasts in the long-range giving way to relatively more consistent forecasts in the short range runs. (Note: I mislabeled this chart as Point Conception...my bad!)

GFS forecasts for total precipitation in inches on January 11th for the Las Cruces, CA area. These are taken from runs starting December 27th through January 7th. Again, the same pattern of wildly-varying forecasts in the long-range giving way to relatively more consistent forecasts in the short range runs. (Note: I mislabeled this chart as Point Conception...my bad!)

Both Los Angeles and Las Cruces received rain between December 27th and the AR event; 0.85" and 0.34" respectively, but the key point here is to look at how less chaotic forecasts became within a week of the target forecast date as compared to the erratic outputs > 5 days out. As it turns out, from December 27 through "now" (05Z Jan. 9th), Los Angeles received 1.69" and Las Cruces received 1.73"... not too far off from what the GFS predicted in the short-range forecasts.

You can see how Quantitative Precipitation Forecasts (QPFs) like this can either get praise or ridicule, depending on the temporal perspective one takes when arguing for or against them. And this holds for other long-range forecast parameters as well. We're funny creatures you and I... we often see patterns where there are none, and remember things at the expense of forgetting others (such as how much more often long-range predictions are wrong rather than right, or that a long-range prediction that turned out to be sort of right was followed by runs that were completely wrong).

Context often becomes relative, and if we take all this together and mix it in a social pot of who is 'right' and who is 'wrong', we end up having all the ingredients needed to perpetuate the need for and faith in long-range forecasts. Couple this with a host of known cognitive biases we all suffer from to varying degrees; such as the Availability Heuristic, or Focalism, Berkson's Paradox, or the fallacy of Illusory Correlation to name a few... there are many dozens more cognitive biases.

Of course, cognitive biases aren't the only set of mental hurdles we sentient beings must overcome; there are memory errors and biases as well. Had I not saved every run of the GFS from that Armageddon run back in December, I might have remembered the GFS performing incredibly well in the long-range.

This sort of memory error is known as the Von Restorff effect, wherein that one run stuck out like a beacon in the night, and I would have remembered it over all the subsequent runs. There are plenty of other memory delinquencies; we can suffer from source confusion, misinformation effect (thank the media for this one), hindsight bias and others.

Long-range meteorological forecasts are shots in the dark... they don't have a handle on things, and at best serve to get us interested in something potentially big; as was the case for me when I spotted that random Armageddon run on the GFS, but not in a statistically-reliable sense. Climatologically, there are long-range predictions that can be made as certain synoptic patterns can be expected... but is this an axiom? Absolutely not.

Look at last year's El Niño, or this year's La Niña which seems unable to overcome the atmospheric influence of cold and relatively warmer waters interacting with each other out in the Pacific. Earth's atmosphere is a dynamic system that reacts to many billions of different small and large-scale influences. Like it or not, we live in a mathematically chaotic world, and in a chaotic world like this, only the stochastic Monte Carlo simulation is king... for about a week.

Well, since we've gotten this far, let's take a look at what actually happened out there... the following is a high-resolution animation of water vapor taken by GOES (you can read more about GOES in my blog from October here).

1745 Zulu January 07 to 1715 Zulu January 09, 2016.

1745 Zulu January 07 to 1715 Zulu January 09, 2016.

There are plenty of write ups and YouTube videos out there regarding this event, so I'll spare you from reading any more of my mumbo jumbo! Mumbo aside, I hope you all enjoyed this... This blog will come back to life this spring when I tackle the LRC hypothesis, AccuWeather, and start forecasting tornadoes again!

Oh, some pet peeves for the news media...

Mud doesn't slide... mud flows!  grrrrrr....

EDIT: I purposely did not use other models (NAM for instance) in this blog, mainly because it was too much of my time to additionally collect that data, but also because it is a short range model... but... I did collect all of the spaghetti plots from the GEFS from the last week of December through January 7 for your entertainment... :)   As one would expect, the ensembles are all over the map in the long range, becoming more coherent in the short range...

Nothing shows long-range forecasts lack of a handle on things like a spaghetti plot.

Nothing shows long-range forecasts lack of a handle on things like a spaghetti plot.

GOES-R Series

GOES-R Series

Aeolus and His Bag of Hot Air

Aeolus and His Bag of Hot Air