November 2017 C.E.T. forecasts

[Summer Sun]

7.4 to the 29th

0.8 above the 61 to 90 average

0.2 above the 81 to 10 average

___________________________________

Current high this month 10.9 to the 1st

[NorthYorksWeather]

5.7c here to the 29th, 0.8c below the 1981-2010 average.

With just a high of 2c forecast today, likely to finish at 5.5c so not as cold as last November.

 

[BornFromTheVoid]

If the maximum today is 4.4C or higher, we finish on 7.3C before corrections, anything else and we finish 7.2C. So at this stage, I'd say a finish of 6.8C to 7.3C after corrections.

[Summer Sun]

November confirmed as 6.8

https://www.metoffice.gov.uk/hadobs/hadcet/mly_cet_mean_sort.txt

[BornFromTheVoid]

Big downward correction this month, 0.41C, from 7.23 to 6.82C

[NorthYorksWeather]

Finished on 5.6c here, 0.9c below the 1981-2010 average.

November was the coldest month relative to the mean so far this year, and the 2nd month after January to finish below average.

 

[The PIT]

Sunny Sheffield finished on 6.3C -0.3C below normal. Rainfall 44.4mm 55.9% of average.

[JeffC]

11 hours ago, Summer Sun said:

November confirmed as 6.8

https://www.metoffice.gov.uk/hadobs/hadcet/mly_cet_mean_sort.txt

Ye Gods I got it right!! Cue Blackadderish type hyperbole....

[J10]

November 2017 Results

Monthly
4 players got the 6.8c Correct simshady, JeffC, DiagonalRedLine, syed2878

Seasonal
All Change at the top with the  winner being Polar Gael, simshady and stewfox.

Overall
In the overall a slight change in the Top 3.

3rd davehsug (from 2nd)
2nd Weather26 (from 3rd)
1st I remember Atlantic 252 (stays in 1st)

Files
PDF Nov 2017 CET New.pdf
Excel Nov 2017 CET New.xlsx

PDF Summary Nov 2017 CET Summary.pdf

Edited December 2, 2017 by J10

[J10]

November 2017

Seasonal Autumn 2017 Final

Overall 2017 Final

 

Edited December 2, 2017 by J10

[J10]

2016/17 Stats

Edited December 3, 2017 by J10

[simshady]

Well done everyone!

[Roger J Smith]

Thanks, J10 for all those summaries and scoring the contest diligently. I am glad to be able to help out with the tables of entries and opening threads, otherwise I would have no idea when it was the 19th of any given month. 

Now, here's the final report on our three hidden talents, 1981-2010 normal, 1987-2016 normal (played Dec as 1986-2015) and Consensus, which is basically like average only it doesn't matter if LG goes totally haywire and says 150, consensus is the median and not the mean. And I'm not being mean to LG, I know he doesn't really believe his above-record-high forecasts will verify every time, just Dec 2015.

(when he was too low, can you believe it?)

Well anyway, these three are top ten material, as I will demonstrate in the table below ... I have shortened the "Forecaster" portion to get the last month (Nov 2017) to fit the page ... view at 100% to avoid line wrap problems ... the phrase Fx is short for "Forecaster" and the word Normal is dropped from their identities.

Explanation ... For each entity, the first row gives their errors for each month, a negative number means they were too cold and a positive number too warm. After the 12 entries, the mean is the average of absolute errors, so ignoring the anomaly sign. Then the final entry is how that ranks in the table of actual contest entrants. Following that is a row of "points" for that same entity, and these are derived from the monthly scoring tables as if these entities had entered before any contestants who predicted values of equal error. Those might have started out with different values on the other side of normal (example, the first January error of +0.4 or -0.4 would get the highest number of points so this would determine points for 1981-2010 in that case.) As there are often 5-10 tied in these middling values near normal, you can assume that if I scored them against the last entrants of equal value, their ranks would drop 5 or so positions, so they are partly good because they are early (like always first to enter early). However, any way you cut it, humans find it hard to outdo these three ... and their friend "0.5 above recent normal" is even sharper, finishing just about 2nd or 3rd.

Fx ________ Dec _ Jan _ Feb _ Mar _ Apr _ May _ Jun _ Jul _ Aug _ Sep _ Oct _ Nov __ mean/T __ position

1981-2010 _  -1.4 _ +0.4_ -1.7_ -2.1_ -0.4 _ -1.5 _ -1.5 _ -0.1 _ +0.8 _ +0.5 _ -1.7 _ +0.3 __ 1.03 __t-5th

__ points ___ 79.6_ 69.9_ 56,1_ 59.0_ 86.8 _49.7__7.5 _ 93.7 _ 61.8 _ 86.3 _ 61.4 _ 76.0 _ 787.8 __ 3rd

 

1987-2016 _  -1.0 _ +0.6_ -1.3_ -2.0__ -0.1 _-1.4_ -1.5 _ -0.1 _ +0.9 _ +0.6 _ -1.5 _ +0.4 __ 0.95 __ 4th

__ points ___ 89.8_ 58.0_ 67.4_ 61.8_  98.4_ 53.0__7.5 _93.7 _ 56.6 _ 82.9 _ 69.8 _ 68.9 807.8 __ 2nd

 

Consensus _-2.1 _ -0.1_ -1.8_ -2.2 _ +0.9 _-1.3 _ -0.5 _+0.4 _ +0.9 _ +1.0 _ -1.8 _ +0.2 __ 1.10 _ t-8th

__ points ___50.0 _95.7 _52.4 _56.2 _57.1 _56.4 _72.4_ 70.1 _ 56.6 _ 60.7 _ 56.4 _ 87.3 __ 771.3 __ 5th

____________________________________

____________________________________

These ranks are determined independently, obviously if Normal 1987-2016 and Normal 1981-2010 were both playing, Normal 1981-2010 would finish one rank lower than shown.

Our consensus was slightly lower ranking than either of these recent normals but it should be noted that only one or two players in the field did better than consensus without also doing better than either Normal. It should also be noted that I have not done the complete scoring regime including accuracy points for these three. So they are only being compared in two key statistics against the contest field.

What's the absolute best "robot" forecaster to emulate? 

Well, taking 1981-2010 as a base (average error was 1.03), these are the average errors for various departures from those 1981-2010 normal values:

0.2 above 1981-2010 ____ +0.98

0.3 above 1981-2010 ____ +0.93

0.4 above 1981-2010 ____ +0.95

0.5 above 1981-2010 ____ +0.95

0.6 above 1981-2010 ____ +0.95

0.7 above 1981-2010 ____ +0.95

0.8 above 1981-2010 ____ +0.95

... so it appears that any strategy around taking the more recent and warmer 1987-2016 normals to adding almost a full degree to the 1981-2010 normals will yield roughly the same outcome but adding 0.3 to 1981-2010 normals is your best option (if you want to let a robot do your work). 

I suppose an even better outcome would happen if you selected the most likely looking option from 0.5 below 1981-2010 to 1.5 above, based on the first half of the month as available from GFS 16-day output. Assuming that would be non-random, you should be able to drop your average error to about 0.7 using that technique. 

 

 

[Polar Gael]

On 12/3/2017 at 12:50, J10 said:

2016/17 Stats

Thank you, J10 and Roger and to all who make these monthly competitions happen. Always good fun (and interesting, too). 

Well done to everyone else, too. 

Am now off to crown myself 'King of Autumn' or suchlike. ;-)>