Are you ready for Winter?
Winter is fast approaching and it is a time of year that often sees Weather playing a very critical role in the energy market. The general consensus in the market is to follow the ECMWF medium range forecast as the safest bet, after all on average it performs the best out of all the models.
However behind the skill score statistics often lies a very different story, let's flash back to last November.
Below: EC ENS Forecast from 17th Nov 2017 (Light Blue shows ENS members - Red ENS mean - Dark Blue EC Op). Black lines to the right of the image is the Forecast from 24th Nov 2017 - forecast flip of 2-7C colder
The forecast issued on the 17th Nov depicted a very warm outlook with a warming trend - however that hid the truth - in out-turn it turned much colder in fact 6-7C cooler as a daily mean across an entire country around day 8-9 in the forecast. Not only was the day 8-9 forecast wrong but the entire outlook shifted from long-lasting warmth to long-lasting cold.
But it didn't end there.....
Below: EC ENS Forecast from 24th Nov 2017 same as black lines above but this time - Light Blue shows ENS members - Red ENS mean - Dark Blue EC Op. Black lines to the right of the image is the Forecast from 30th Nov 2017 - forecast flip of 5C warmer
The cold forever outlook wasn't right either - it ended up flipping 5C warmer! Once more in the 8-9 day window onwards - making it look like a brief cold snap - back to long-lasting warmth in the medium range for what was December.
However this wasn't the truth either and it actually continued to flip like this for a good few weeks afterwards.
Why do the models do this?
It's easier to analyse this if we break it down into two components:
- Models predict the correct pattern but have the wrong variables (i.e. cold air sits for too long in a westerly pattern)
- Models predict the wrong pattern completely
For the first component you can protect yourself against this.
After these model flips occurred, I felt it was necessary to develop a sophisticated tool that compared what the model variables were forecasted to be, against what normally out-turns in that pattern. The tool is called MetSet and you can find more about it here: https://www.met-set.com/
Luckily I developed it quick enough to make very good use of it in February and there was a very clear edge using this tool verses what was being said in the market at the time.
Let's re-wind to the end of Jan/start of Feb 2018 - everyone remembers the big SSW (Sudden Stratospheric Warming) event and the strong cold that was correctly predicted for late Feb into March - but was is not well remembered was how poorly models performed before this event.
Late January heading into February, the general consensus in the Market was for warm Westerly however a stubborn high pressure flared up in the forecast at the very end of Jan for early Feb and we had a big SSW event in the forecast - models had gone cold forever so the outlook/market consensus went very cold for the whole of Feb - because of this.
Below: Forecast Days left to Right - EC ENS sorted into pattern probability per day - colour coded to show whether the model thought it was warm or cold from https://www.met-set.com/.
Trouble was that the cold in early Feb was from a temporarily high pressure that moved over from Russia - this was not SSW induced cold in the outlook here - instead it only looked cold because the models could not see past the cold front-half. If you look at what the pattern normally brought you'd see a much more balanced view .
Below: Forecast Days left to Right - EC ENS sorted into pattern probability per day - colour coded to show whether the pattern normally out-turns as warm or cold from https://www.met-set.com/.
This enabled me to correctly predict a more seasonal spell for mid-Feb (and colder view later due to SSW still) at a time when the market was dead certain of a cold spell forever from the start of Feb.
Then when the SSW event was changing the Pattern for the back-end of Feb - the tool came in useful again - the models only wanted 3degC belows - but the pattern that was heavily favoured, was on average a good 3degC colder than this - and that's not adding the extra factors of being SSW induced and a colder than normal air source over Russia - meant I was able to correctly predict the extremely of the cold as well which the models did badly on even 5 days out - although did very well with the general pattern.
At least this combats model biases when the pattern is right but the model is blinded - certainly worth a look: https://www.met-set.com
For the second component - when models are completely wrong - looking at models will not help you in this scenario - this is where you need a Meteorologist with medium-long range expertise to help you.
If you are a Met - the great news is that you can use the MetSet tool as a platform to add teleconnections biases onto it to enable a fully integrated model/teleconnection cluster analysis.
The traditional approach to teleconnections is to take the model mean – see what the teleconnection bias is (+NAO/MJO Phase 4-5/solid PV = warm/windy) and then add/subtract your bias to the model mean. Trouble is looking at teleconnection this way does nothing to tell you about risk of busts etc.. With this tool you can take the current model forecast, then add in what normally out-turns in an MJO Phase 4-5 (there is a downloadable excel file with dates to enable you to look up all kinds of stats – in this case looking up when MJO is in Phase 4-5 – and match that to what was pattern outturn) - this gives you a probability of being in each pattern based on MJO - which you can combine with your model probabilities. This enables you not only to have a main scenario view but also gives insight into possible alternative scenarios to consider.
You can even apply different MOS calculations when in different pattern as model biases for "city locations" or wind forecasts will have different errors depending on the flow.
If you are not a Met but need sensible advice on weather forecasts, don't leave yourself unprotected, I would seriously recommend that you send me an email: firstname.lastname@example.org