Having done previews in the past and enjoying reading previews of others there are two issues with most previews in my mind:

1. You don't see the quality gaps between teams. Many previews present only the final rank prediction. That is fine and i am already happy about anything but it only tells you a part of the story. Often predictors will have a couple of teams that look just about equally strong with close to no difference in between.
Whether they put a certain team at e.g. 3. or 7. can look like a rather big difference but the creator might see them as basically similar strong likely seperated by 50 or 100 points while the 2. and 3. team might be seperated by a huge gap of 200 points in the mind of the creator. Those gaps are very interesting to see imo but only few prediction format can show them. Jandal shows them a bit by grouping teams into different categories and Marco shows them with a difference in "some artificial Team-OVR" but I'm looking for a better visable gap indicator.

2. Many predictions expect teams to follow a certain team building guideline to get a good ranking predictions. Like: "You need one super good main leader, bonus points if it's a climber. You also need another at least sub-top leader in every terrain. All of them need at least one good lieutenant. etc.". I am a firm believer that quality is quality. You don't have to follow traditional thinking to make a good team. You can be succesful by simply having only many sub-top riders, you can focus on one or two terrains only and be succesful (see Gazelle, Strava, Evonik or 2014 Euskaltel). But the most important point is the quality you have in your team.

I've tried to implement these two thoughts into a formula. I feel like OVR is already a very good indicator for the quality of a rider and better than every improvised formula i could come up with so i came up with a formula that directly translates the overall of a rider into expected points. This formula comes straight from the 2016 final rankings of the division.

Adding the best 20 riders (=min number needed) of each team together leads to the pure-OVR based prediction. Every rider over that limit would reduce the RDs used of the better rider so I decided to ignore them to have a fair basis to treat the different teams. (it can be a nice help in some situation to have an additional specialized helper but in most cases it doesnt matter much whether #20 or #21 is the filler in the race)

This is the pure-OVR based prediction:

	Team	Manager	PT
1	Festina - Dexia	SotD	7299
2	Moser - Sygic	Roman	6920
3	Team Puma - SAP	cio93	6477
4	Pendleton's	wackojackohighcliffe	6432
5	Vesuvio - Accumalux	SportingNonsense	6379
6	Aker - MOT	ember	6363
7	Aegon - Lavazza	roturn	6170
8	Wiesenhof - Andritz	Heine	6114
9	Gazelle	DubbelDekker	5965
10	Hugo Boss	dev4ever	5942
11	RBC Pro Cycling	jt1109	5763
12	eBuddy	alexkr00	5585
13	Tinkoff Sport Academy	OlegTinkov	5565
14	Jayco - Cobra9	tsmoha	5541
15	Porto - Prio	Gustavovskiy	5500
16	Spotify - Haglöfs	viking90	5421
17	Evonik - ELKO	knockout	5419
18	Iberia - Team Degenkolb	beagle	5238
19	Strava	Shonak	5000
20	Project: Africa	aidanvn13	4972
21	Nemiroff - ABBYY	fjhoekie	4402
22	Ayubowan!	jph27	4203

I then went through the list of riders and adjusted expected points of many riders based on how i rate their stats and how well they performed in 2016 (measured in points). That were especially downgrades for one dimensional climbers without acceleration and 76-78 sprinters and upgrades for versatile riders like Bille, low AVG TTers like Stannard and other underrated riders like Chavanne. This step was fully suspective and since i basically looked over every rider the influence of the original formula was widely reduced but is only used as a baseline for my grading of better riders and as a grade for lesser rider. The result is a list of expected points for each rider.

I think of gained points of a rider in a season as a statistical random variable that (if we played a season often enough) would be distributed randomly around the expectated points with a certain variance. For this preview I assumed a Gaussian distribution eventhough that's obviously not really the ideal case but it is simple and good enough to work for a model like this. The variance depends on the highest main stat so that a TTist or a climber will have a higher variance than a puncheur.

I then simulated 500 seasons based on these numbers to see how likely each team is to finish in a certain spot of the rankings. The results are seen in the matrix posted below. How to read the chart:
Place = The predicted finishing position as seen in all other predictions. This should be used to compare my method with other predictions for a fair comparison.
AVG = Average finishing position of 500 simulations
1 = Observed probability to finish 1st
...

Please note: I haven't adjusted anything for the new calendar because i don't know yet how much the expected points should change. I also didnt looked at any announced race planning information to change the numbers. These numbers should simply be seen as an idea for expected finish, best case and worst case instead of any real realistic numbers.

knockout wrote: