I hear they performed well during the summer. Well that makes sense seeing as that's the time of the year when the jet is somewhat less vigorous. These East Vs West standoffs are much more influenced by even slight changes in jet energy. 

Will we get a mid winter Easterly? Perhaps. Though I think we all know it isn't going to be a clean 80s style countdown. Times change, as do expectations, so I suggest we *try* to enjoy the twists and turns over the coming days 👍

I think that's all very fair. 

We're not seeing a 2018/2010 style countdown in the ens. In fact,  both the ECM and GFS ens have never really suggested an Easterly to be any more than a 1 in 3 shot, though this is infinitely better than most winters since the end of 2017/18!

The Easterly is still there as an option and I personally do think we'll eventually get one, just not the screamer shown in some of the more extreme solutions, and probably not until sometime early in February. 

I think a bigger problem than climate change is precision of reanalysis charts. If AI is using data from 30 years ago, then it is relying on initialization data that is much less precise than it is today. Think of the difference between the 0Z T+6g forecast, and the 6Z T+0h forecast in any model run; they are so minute you are playing spot the difference. Do we believe reanalysis charts from 30 years ago are that level precise particularly when concerning the uncertain Arctic?

I worry that AI is basically the 'october pattern index' on steroids when it comes down to it. I'm not getting the sledge out until the real models are also showing these crazy charts with more consistency. I will say, though, that the ensembles (and many OPs) do leave us very much open for interest which is why I still remain somewhat optimistic.

I'm a big fan of AIFS in particular it's stats suggest it is the best performing model in the world. But it isn't perfect and the easterly is still 9 days away so plenty of time for it to go wrong. 

Yes but the best by what measure? It probably means its the best at matching some upper level height pattern over the entire globe; which is nice but that doesn't mean its the best in forecasting what is going to happen in our little patch given this very specific and rare scenario.

Me too - you can't beat them (in terms of accuracy) for large scale weather patterns. The devil is in the detail however and whilst they're great for picking up broad changes it's the detail that's a challenge.

The stats are for the NH yes. But I followed the model closely in summer and it was the best at predicting the heatwaves and settled weather the UK had. It isn't as good in winter and it didn't handle the New Year cold spell particularly well but certainly no worse than the other models. I'm not saying it's got this easterly right and it may well drop it soon but it shouldn't be dismissed either. 

That's probably fair, but all models struggle with detail. 

If what the posters on here are saying is correct (I use their knowledge on weather prediction because mine is more limited) surely the AI element is slightly different tool to use that can contribute to the overall picture (predictions) at a particular time.

I think it comes down to, what approach works better in this context. The empirical (AI) method where we ignore the Physics and just look at what happened in similar scenarios, or a pure physics approach modelling what we would actually expect the atmosphere to do slaved to some Navier stokes approximation.

I think the answer will depend on the context, I don't have one for this scenario but I can think of pros and cons to each. In terms of the MO reading in general; the cold easterly is clearly (imo) still on the table but we have probably running out of time now for that Svalbard bubble to be in a more useful place than most of the 6Zs have it.

The influence relationship combines all available met data - pressure, temperature, humidity, dew points, wind speeds & directions through all atmosphere levels and stuff like SSTs. It then does a massive matrix mash-up of all the data items set against all the data items for each 6-hour run in the history set; develops the current value-weight-factors and makes a prediction. It will then re-do this for every run from scratch again, so if it maintains a forecast position over time, it is a useful prediction. 

But it is only a prediction, so it can be wrong, just like any other prediction model that is based either directly on pre-defined physics equations or physics equations with human oversight. It really is a very interesting period of model watching. Keep it up everyone.

I actually wonder if it is a factor of latitude. The big difference between the AI and legacy models towards the end of next week, is how they predict pressure over NW Russia/Eastern Scandi and the incursion of the Arctic low (or not).

Are AI models fed with sufficient data from Arctic regions for example? Do they pick up on shortwaves to the same degree as legacy models?

Will be very interesting to see how this all plays out. 

The Meteociel wind roses on the ensemble chart do suggest an easterly or northeasterly influence later in the run:

https://modeles16.meteociel.fr/modeles/gensp/runs/2026011606/graphe8_1000___-0.12574_51.5085_Londres.gif 

P5 and in particular P28 give a delayed pulse of really cold stuff, close to -20 incoming....

There's a simple answer: check the verification stats 

Ah okay, that's very interesting. Do you know how it works from there? Do they then compare the initiation data to reanalysis charts and let it run from that? 

Are there any papers you know of explaining exactly how they operate, as I'd love to learn more if so 👍

EDIT: Apologies for all the questions, it's just a fascinating subject. 

One other question Brian - How are verification stats quantified? Are there any articles you know of that explain this? 

To clarify I was referring to the training data rather than the current initialization data. Essentially all the reanalysis data that taught the model how to pattern match; my point was that imprecisions in the reanalysis data could lead to poor pattern matching in certain contexts in particular.

Another fascinating set from the GEFS. We even have one crazed coldie going off on one from the 22nd! (I've christened this little fella Moomin 😘)

The mean again meanders around the -5 line, though there's now less obvious clustering around a frigid cold solution. 

Issue with AI models is that they are basically super sophisticated pattern matching systems that try to discern the coming weather based on previous parameters (This is how all AI works, first you must feed it data to learn before you ask the questions).

The issue is that previous wether patterns are based on previous climate and with the climate changing and warming, this means that in affect the AI models are predicting what would have happened previously from the current patterns and data, as we don't know how the climate is affecting current patterns (but it is), then the AI model is in affect already out of date.

They will for the AI weather models I suspect read all the initial data that is fed to the numeric models and look into the database for a set of data that closely matches this input and will then look for the outcome of said initial data in previous model runs. Hopefully they also read the failed model runs into the AI database as well to show the engine what happens when it goes wrong as well as the actual outcome. The engine will then determine the likely outcome from these.

Makes sense. I suppose through time, the accuracy of their output, should improve steadily.

If it (proper cold from the east) is to happen the timescale to get the cold in has always been around the 25th to 27th at the earliest. Some op runs have tried to bring it in sooner but they have been isolated.

GFS 12Z has the svalbard bubble further NW; it lands over central scandanavia at T+120h. That's quite a bit better than recent runs, but the ICON has gone the other way so who knows!