It turns out F_A saw what I could not. As soon as I got rid of the Sports Madness and the (old) CBGR numbers, most of the outliers disappeared.

I test the average of the four services to find if the sum of the differences of each of the numbers from the mean is greater than 20. There's a "AVDEV" function for that. If it is greater than 20, I use the GEOMETRIC MEAN. If not, I use the MEAN.

I may change this in the future, but for now
Pros
• It is fairly easy to implement
• It seems to be more fair
• No data is thrown away
• It is consistent without any judgement on my part
Cons
• No judgement on my part, purely mathematical
• Assumes each service is equally competent and honest

I mark the averages that have outliers, i.e. the ones which use the GEOMETRIC MEAN, so I can study if some services are consistently better than others. (I intend to test this in other ways also.) Perhaps i will find it better to eliminate high or low scores in the future depending own how the players with "outlier" averages perform.

P.S. It improves the "averages" of Mikayla Vaughn, Abby Prohaska, Alasia Hayes, and Amirah Abdur-Rahim.

P.P.S. I believe the large number of outliers among the ND recruits is accidental. I harbor no conspiracy theories, but will be looking at the performance of those recruits to see if Muffet can pick them better than (some) rating services.

could be a determining factor for eliminating outliers. Of course those picks can be subjective as well - they are determined by coaches/committees who have their own preferences/agenda. But generally, they are made up of Top 20 players or close to it.

For example, anyone who watched Sonia Citron play in International tourneys, knows you can throw out that 33 rating.

And of course there is the Benjamin Disraeli/Mark Twain quote:
"There are three kinds of lies: lies, damned lies, and statistics."

by a few things, one of those is playing on the national team. But, if the girl's rating is already below 12, there is no benefit (assuming the top twelve belong on the team). In Citron's case, she currently has a 19.5 rating, so I give her another point benefit, 18.5. That may be not enough. On the other hand I don't want to penalize girls who don't have the parental time and money to work towards the national team.

But you may be right to give the national team members more benefit. Thanks.

Not by a famous person, but I also like "73.6% of statistics are made up".

solution is to drop the highest and the lowest scores and average the rest. That doesn't require you to make any decisions as to what constitutes an outlier, but utilizes slightly more information than the median. But if you have only 3 or 4 data points, the median and the mean (after removing the first and last values) are identical because there are exactly 4 values total. If there were 5 ratings, the number would NOT be the same.

Another candidate would be to take the geometric mean (Nth root of the product of N numbers), a nice way of finding central tendency because it tends to mitigate outliers.

An aside on the ratings: I would just use the 4 main ratings because most players aren't rated on Sports Madness and Dan Olsen's CGBR is redundant if you are already using Hoopgurlz. Dan Olsen's pay site (CGBR) updates more frequently than Hoopgurlz, but it's the same guy's rating. So CGBR would be preferable, but it's hard to get it unless you are a subscriber. Because CGBR ratings are highly dependent on WHEN the ratings were acquired, it seems less valuable than the Hoopgurlz rating. For example, Shelby Calhoun was highly ranked earlier in her development, but sank in the ratings over time. Dan Olsen didn't find her worth a top 100 rating in the final rankings of the 2020 class, but your listed CGBR rating was 15th! Maybe you got that rating from some time in the past when she was more highly regarded. I checked the VTech site and they only mention her high school accolades, but never mention her high CGBR rating. Maybe it's an old rating? (Don't really know.)

So here are the ratings of the women you listed, only counting ratings from HG, BS, PN and ASGR. The numbers are MEAN, MEDIAN, GEOMETRIC MEAN:

Edwards 84 87 49
Abdur-Rahim 113, 117, 108
Hayes 62, 43, 52
Calhoun 76, 72, 72
Todd-Williams 58, 45, 51
Thompson 103, 101, 101
Levings 75, 73, 70
Hutcherson 54, 46, 40

Sonia Citron (2021 of course) has one outlier among her 4 scores (14, 15, 16, 33) but but her mean (19.5) is not too different from her median (15.5) or geometric mean (18.2).

So which one is your favorite?

First of all, your advice on only using the standard four services makes total sense. I'm sure you are right about Dan Olsen's CGBR. I just stumble across those once in a while. I never understood why they would be different than Hoopgurlz. Now I know—they must be old ratings. Sports Madness is also a "catch as catch can". I thought more information is better information, but, as you say, that is not true.

Also, thanks for providing the numbers for the sample data. The GEOMETRIC MEAN is intriguing and often (Edwards, Abdur-Rahim, Hutcherson) gave what I would consider the best "average". Theoretically, however, these should be related numbers, not requiring the GEOMETRIC MEAN. Of course, theoretically, there shouldn't be such outliers. Dillon does give good reasons for them, however. Also, there is a GEOMETRIC MEAN in Numbers (which I use, also in Excel, of course).

(Note: I learned some new stuff, namely the GEOMETRIC MEAN ("A child is half-way between a cell and the Earth")

So, I will drop the other ratings, which will help, especially with Shelby Calhoun. I still must decide whether I think outliers are the result of negligence and/or corruption and drop some data or whether, as I typically think, people and organizations are competent and forthright. Perhaps I will use the MEAN (or MEDIAN) until I come to a outlier > 20 from the MEAN and then use the GEOMETRIC MEAN. That keeps my faith in mankind.

I'll probably do a little more testing. Thanks again.

1. High school rankings are a valid discussion point to use for incoming frosh and perhaps sophomores. At some point they cease to be a good measure of the player's ability.

2. The ability of players, if you had a perfect ranking system, would resemble the normal distribution curve. The implication is that there may be a greater difference between players ranked 3 and 10 than players ranked 30 and 50.

3. Ideally I'd like to assign a certain number of points (say 5 -25) to each ranking position. For instance the top 3 rankings would be worth 24 points (25 points for a Griner or Stewart "generational player"). The next grouping might only be 22 points (to reward getting the truly elite recruits). This mitigates the unranked outliers a bit, although it becomes more work.

4. Your team ranking is the points for a mix of say 2 tall players and 3 guards/wings

1. I've incorporated an "experience" factor based on minutes of playing time. In addition, the first 1000 minutes are worth more than the next 1000, and the next.

I have started using Martin Manley Efficiency rating for how the athletes perform in college. (PER is just not worth it.) So far, I've kept performance separate from rating, but who knows?

2 and 3. I think I got this idea from you. I produced a complicated formula using various logs to value the better rated athletes from the others. I take the top 150 from each class and apply this formula. The result is that the difference between No. 1 and No.2 is 2.2, (i.e. If your avg. rating is "1", you end up ".77". If your avg rating is 2, you end up 2.96) between 2 and 3 is 1.9 (a little less then between 1 and 2.) The difference between No. 149 and 150 is .85. A lot of fluctuation between ratings occurs in the first 5 or 6 players and by 20 or 30, there is less than 1 between the players. I toyed with assigning points, like you say, for the first 5, then the next 5 or 10 or so, but, as you say, it gets complicated.

4. That's an idea I haven't really dealt with. It sounds good though.

especially how to group players.

The concept of #4 was based on a shortcoming the post-Wooden UCLA team had in that they seemed to recruited the same player (wing/shooting forward) and lacked a "good mix". I think we ended up with that sort of problem kast year, especially when Mik was hurt. It's the same concept that Georgia had with its QBs - great to have a top 3 recruit 3 years in a row, but only one can play at a time.

Then you do not have to disregard any of the rankings and get a better mix of opinions.

Edit: Need a valid definition of outliers.

I've been thinking of a deviation from the mean (or median) of greater than 20 invokes some modification. Perhaps it would invoke a formula for weighting the values based on how far they vary from the mean or median. You've opened up new areas of enquiry. Thanks.

although a lot of work. I suppose you could set up a spread sheet with formulas to do it.

Plus I've automated much of the input. Among other things, I'm a programmer. I'm all for doing the least amount of work.

- Can you include both a median and mean? Might satisfy folks who are looking for certain aspects of their numbers.

- I think it was Tim K. who wanted some more specificity of "outlier" --I get that -- and suggested "weighting" outliers.

As you know, I've always thought that some services have drastically different ratings of some prospects for reason varying from: 1) they have small rating staffs and have not seen the player; 2) a player might be a frequent attendee of camps they're at/sponsor, etc.; and/or 3) They just think the player is the cat's meow or a not-so-shiny penny.

Can't tell what is an outlier per se: you kinda know it when you see it.
Five to 10 points doesn't catch my eye, but jumps of 20 points or more tend to do that. Even worse, non-inclusion, so I usually tend to drop the low-rated ones. However, your adding points to certain candidates -- if feasible -- seems right and tight.

- Lastly, Most of these services stop rating players once they're in school so they don't matter quite as much, save for going back to see how accurate a picture they gave.

TopDrawer Soccer continues to rate players throughout their entire career and All-White Kit may do one of the most thorough sketch of draft-eligible players I've ever read.

Could we go by various ratings based on game statistics? Sure, but would love to get some Chris Henderson "All White Kit' comments and insights there.

Including both a median and a mean for 1,000 or so high schoolers over 5 years takes about 20 seconds to implement. Is it really satisfactory though? I think what people really want is less stats but more meaningful ones.

I have the same problem with Efficiency and something called PCR which is Efficiency that better reflects guard contributions of assists, blocks and steals. A complaint about Efficiency is that it favors centers and forwards. For example, Ruthy Hebard had a slightly better Efficiency rating than Sabrina Ionescu last year, but Sabrina's PCR was higher. I could use both, but I think it would be more confusing than it's worth (and less consistent). Everyone knows the value of both those players.

As you wisely say in you last paragraph. OK, I'm going to paraphrase here. What are we looking for on the Bench, the mood of a stat class or the character of a sports bar—a high class sports bar? The banter and repartee is what I'm looking for (along with a few games), but a few understandable, meaningful stats can get the conversation going.

It is naturally more resistant to outliers. If there were more rating agencies I would 100% prefer the median to the mean, but with so few points for each player you might have to proceed with caution. Still addresses your issue though.

I have some testing to do. Thanks.