Posted by buddhagamer on 10/3/2018 4:30:00 PM (view original):
A team winning 55% of the time is almost not significantly better in HD and most coaches who did win 55% of the time wouldn't say its out of line to role switch and be the 45% winner. Its the Team A should win 80% of the time and in terms of HD programming they are getting the shaft in the first half due to the variances of sim engine of only having some 40 possessions that don't fall within expectations.

It’s already been stated that the number of upsets was roughly in line with real life. I’d like to see how often those legitimate “80%” teams were losing. Was it more than 20%? If not, *the problem was user expectation*, not a faulty sim.

1. Not buying it. Unless you’re telling us that right now the simulation consists of one single outcome (final score), and the rest is just window dressing, re-configuring the parameters that determine possession outcomes isn’t any more complex (at least it wouldn’t have to be) than adjusting the “player roles” from the user settings page.

Reconfiguring the outcomes of possessions only makes a meaningful difference in the outcomes of games if you do it in an asymmetric way (IE you make possessions on which the team was most likely to score more favorable while simultaneously making the possessions on which the team was least likely to score less favorable). Basically actually make the good teams, players, possessions more successful relative to the lesser teams, players, possessions. I strongly feel this would be a massive net negative for the game because it would shrink the number of viable strategies. Everybody would move towards offenses focused around elite scorers because those guys are getting bumped relative to average scorers. So triangle would become a dominant offense. Almost no way to avoid that in this scenario. And teams built for defense and turnovers (IE fastbreak/press) would likely become less viable. Shrinking strategy is not a good way to fix a game.

The reality is that, at least based on the explanations we heard, this correction factor is designed only to kick in during significant outlier results. And it may not even be that big. Plenty of people get up huge in blackjack and then wind up broke. Every dealer who's been at the tables for any amount of time at all can tell you plenty of stories about these people. It's part of the reason it's a psychologically challenging job for some people - it's hard to watch people make an amount of money that could in some cases be significant in their lives and then blow it all in half an hour. But they do it all the time. All. The. Time.

Nobody changes the deck. Nobody changes the odds. Most of those players aren't changing the way they play. You just go through a cold streak at some point. It's bound to happen. If you play at an outlier level for part of a game of HD you would expect significant regression to the mean in any case. The extent to which this is accelerated is difficult to ascertain. We know there is some. Frankly, if you create a histogram of player or team shooting performance for every game at the end of the season the distribution is not significantly narrower than you might expect. Plenty of good shooters go 2/11 once or twice, 10/13 once or twice, and mostly hang out around .500. It's hard to say exactly how broad the distributions should be since we don't know the extent of natural variation depending on defending. But like I said, nothing looks particularly amiss.

I've long wondered about the algorithms that run this game and obviously unless we can peek behind the curtain we'll never truly know what the inputs are. And that's fine. The main debate in this thread regarding probability and competitiveness seems to be either:

- WhatIf should tweak their formula so that if a close game is the intended outcome, the PBP should reflect this from the tip and be generated, possession by possession, because the inputs of the ratings and gameplan are evenly matched. This would be the only reason a close game was intended in the first place. Outliers aside, if two relatively even teams are going to play a game that should be decided by < 5 points, one team shouldn't be up 23 at halftime only to inevitably (some may argue predictably) see the trailing team come charging back and make it a close game. Do these types of games happen? Sure. But I believe the argument shoe is putting forth is that this "NBA Jam" programming carries too much weight and creates unrealistic outcomes due to it's lazy design. These are fair points. If you're up 20 at halftime on a "good team", if shouldn't become predictable that the game will narrow. It should be possible - but not predictable. Additionally, I believe an improvement in this department would likely lead to less 20 point halftime leads in these instances that then need to be "corrected" to ensure the most probable outcome (aka a close game).

vs.

- This is the best solution WhatIf has to a difficult problem. Because of the small sample sizes contained in 1 game, 40 minutes, 60 shots, etc., random variance has the potential to create heavily skewed results if not held in check by an outside force. People would get frustrated if "better" teams won every single time because of their mathematical advantages OR if small sample size variance allowed bad teams to regularly upset good teams because of the sheer randomness of an isolated sample of dice roles. The only way to hold this in check is to implement some type of NBA Jam coding that narrows gaps in games where the most probable outcome is a close result. This is also fair and was apparently more of a problem "back in the day" with coaches being upset all the time about losing to bad teams.

Does this seem like an accurate representation of the arguments here shoe & debater?

With regards to the box score in question, any team going from 16 points in the first half to 61 in the second must have had one helluva halftime speech. And it is definitely frustrating to think that if you play a smart lineup and come up with a strong gameplan that rightly grants you a solid early lead, that that will then be diminished because of a stacked deck, not the counter-punch of your opponent. I was up 20 points in the 2nd half of an NC game a while back before my opponent went on a 22-1 run over about 8 minutes. I wound up still winning on a GW 3-pointer but to see my team completely fall apart like that was a little odd considering how strong we'd been though 30 mins. My attitude towards that was simply that the game should have been close, and it was. Took a roundabout way to get there but that is simply a flaw in the sim. Box score was fine, PbP was weird. If I'd lost, though, I probably would have been a bit more annoyed! Also - it matters when the NBA Jam switch gets flipped on and off. Giving an NFL team the ball at their own 25, with 1:52 to go and 1 timeout down 6 is still a bit of an uphill climb. But if you give them that scenario when 7 mins earlier they were down 34-14, suddenly the 40-ish% chance they have to win the game is BS if you're the team with the lead because the losing team's odds of winning should be at about 5%. The fact that you still have a slight advantage is no consolation.

Perhaps the middle-ground here is that shoe is right - ideally, close games should be caused because the inputs of two good teams will cause that. Not because WhatIf needs to use the old school Madden technique of boosting losing teams to keep games close. But debater has a point as well insofar as I have no idea how WhatIf could actually precisely hone their algorithms to create this type of honest competition without having a fail safe in place to ensure, at the very least, that the box scores remain accurate. Even if the PBPs occasionally get a little off track. I'm not sure there will ever be a consensus here, mainly because there are too many variables to determine reality. Are you up by 20 in the first half because of small sample size variance or a great gameplan? Overall I think WhatIf does a pretty good job of allowing "better" teams to take care of business but also allowing close underdogs enough of a punchers chance to keep things fun and competitive. But again to shoe's point, I definitely agree that this is "a thing", and something to be aware of.

Posted by shoe3 on 10/3/2018 4:38:00 PM (view original):

Posted by buddhagamer on 10/3/2018 4:24:00 PM (view original):
Programmer here for like 40 years, and I can tell you that making it work like shoe wants would by totally complex, take massive amounts of code/CPU cycles to just SIM a single game (nevermind 10 worlds with 500ish games going). I wouldn't want to code review those algorithms either.

And just to be clear, shoe are you advocating adjust shooting %s on a per possession basis (assuming the previous ones are outliers) instead of just doing it at halftime (and only giving it a single adjustment)? How is that different?

All this because you think your team should of won that game because yes your team got lucky for the first half, got an outlier result (holding your opponent to 16 points), and they wouldn't get a similar outlier result (61 points of which only 46 were non-intentional FTs) plus you previously lost to the same opponent earlier in non-con.

1. Not buying it. Unless you’re telling us that right now the simulation consists of one single outcome (final score), and the rest is just window dressing, re-configuring the parameters that determine possession outcomes isn’t any more complex (at least it wouldn’t have to be) than adjusting the “player roles” from the user settings page.

2. No. I’m not talking about shooting percentage. I’m talking about adjusting the parameters that determine the outcome of a possession. My goal is not to produce a simulation with x number of upsets. My goal would be to produce a reasonable and fair process of simulating college basketball games.

3. Call sour grapes if you want. You’re incorrect, but it’s cool, I understand. I’ve talked about it before, this is the most recent, and most egregious example. And as I told mully, the whole point of bringing it up here is that once folks understand what is actually happening, and can spot it, they will feel “royally screwed”. Because forced or speeding up regression is fundamentally unfair. It is re-loading the deck. It is turning a coin into an unfair coin once it shows heads 3 times in a row.

What part aren't you buying? The fact that WIS technically has approximately 10 minutes per world to generate 450 played games, generate PbP for those games and publish web pages and you want a complex simulation which keeps track of previous possessions, regenerate shooting % based on adjusting parameters, all in real-time.

I'm taking a guess here but what I expect is happening is the run the algorithm once at the start of the game, randomize each possession based on that value, adjust at half-time (including the supposed adjustment along with the coaches' adjustments) and generate the 2nd half.

PbP is all window dressing likely as the result of the possession is either made shot, missed shot, turnover, foul, etc. and then a decision tree will indicate how and whom did it.

Buddha, the inputs for possession outcomes are actually recalculated far more often; in fact, I think it's done each possession. It used to be done every time a substitution happened but I think around 2012 they ramped it up to every possession because of fatigue/proportion of shots taken by one player changing on a per-possession basis.

If that is actually how it’s done - one sim, lumping in all the factors together, and spitting out possessions as “window dressing” at the end - I feel like that’s a much bigger deal, and a much bigger shock to more people than anything discussed here. We may need a new thread, because I would have a whole new rant.

OK, granted, *if* that is how the sim is actually run, it would be a bit more complicated to adjust to the way I’m talking about.

Now that I think about it, dahs is likely correct since players are substituted in and out but its likely not tracking prior results (i.e. if your 15 PER C has made 2 x 3Pt shots so far, I doubt they are going to adjust his chances of making the next one downwards on the next possession but leaving it identical to the previous possession assuming all other factors are identical).

I guess I"m wondering how is adjust the shooting % on a per-possession basis any superior than adjusting it only at half-time? At least in RL, there is the possibility of other half-time factors (HC giving them a pep-talk, underdog becomes over-confident, veterans on the better team not wanting to get embarrassed).

Regarding shoe's 55/45 question, two points:

First, you really can't define the parameters of the model by the outcome, it has to come in the other order. In other words, you can't say the team "should" win 55% of the time, how much do they win? If the model says they should win 55% of the time they win about 55% of the time. If they don't then they aren't, by definition, a 55% win team. So this doesn't really mean anything.

But I understand what you're asking. You're asking "how much would a team that would have won 45% of the time against a certain opponent without forced regression win with forced regression." The answer to that question is also 45%. Keep in mind that this effect only kicks in during significant outlier events, IE in the tails of the distribution. A 55/45 matchup is pretty even. What you have to remember is that the 45% team is just as likely to experience a negative outlier half and experience positive regression, and the 55% team just as likely to overperform and experience negative regression, as vice versa. When the distributions are largely overlapped these effects basically cancel one another and the overall distribution of game outcomes is pretty similar. All you'd really observe over a large number of simulations would be that the largest margins of victory would tend to be smaller than they would have been without stat corrections. If you plot what this would look like in terms of points per possession, it looks something like the top half of this figure:



What bringing outliers back towards the center does is shrink the tails (to the dotted lines). When the distributions are largely overlapped - and in fact, that top distribution is a bigger split than 55/45, a 55/45 matchup would be even more closely overlapped - it doesn't affect very much in terms of long-term winning percentage. But when the teams are more uneven, as in the bottom panel, you see the probabilities are significantly reduced. It may turn 80% into 90%. It may turn 90% into 97%. But at most it turns 55 into like 55.5%.

Posted by buddhagamer on 10/3/2018 6:37:00 PM (view original):
Now that I think about it, dahs is likely correct since players are substituted in and out but its likely not tracking prior results (i.e. if your 15 PER C has made 2 x 3Pt shots so far, I doubt they are going to adjust his chances of making the next one downwards on the next possession but leaving it identical to the previous possession assuming all other factors are identical).

I guess I"m wondering how is adjust the shooting % on a per-possession basis any superior than adjusting it only at half-time? At least in RL, there is the possibility of other half-time factors (HC giving them a pep-talk, underdog becomes over-confident, veterans on the better team not wanting to get embarrassed).

In 2011 seble announced that they were evaluating double teams on a possession-by-possession basis. So presumably other things are also recalculated on that basis.

This has gone in a bunch of different directions, but to clarify:

1. I’m not talking about how it should recalculate every possession. I’m saying if the outcomes feel too random (too many upsets) to the developers, tweaks should be made at the level of the original parameters determining the outcome of possessions, and importantly that they should *never* be readjusted or recalculated in the course of a game, except when variables themselves (like setting, defender, etc) change. I’m not saying change it as it goes along, I’m saying don’t change it.

2. I’m not asking about how often a 45% winning team should win, and in effect you’re making my argument for me. You don’t engineer a result, you construct a reasonable sim based on reasonable and fair parameters, and then you let it run. If it feels like there are too many upsets, see above. It’s a problem with how the parameters are set up. My preference, if the results are somewhat in line with real life frequency of upsets, is that you let the simulation run, and don’t worry about the people who can’t accept an occasional bad sim. We have coaching tools to mitigate the risk of a bad sim. The game *should not be forcing or speeding regression* because that is a fundamental perversion of probability and statistics, and building it into the game means we are not dealing with a simulation, we’re dealing with a god mode.

The game *should not be forcing or speeding regression* because that is a fundamental perversion of probability and statistics, and building it into the game means we are not dealing with a simulation, we’re dealing with a god mode.

This really isn't true since in effect this implementation preserves a predictable distribution of outcomes, it just narrows it. But you're free to have your own take on it.

One way of thinking about this is that rather than flipping a fair coin it's more as if the sim is randomly picking from a set of papers marked with 1s and 0s. In that case if you pick a significantly non-equilibrium number of 1s, you're going to naturally experience a correction in which subsequent draws have a greater probability of coming up 0 than the initial distribution. The more off-center your existing results are, the bigger the correction. The more you've drawn, the bigger the potential correction. This is still a random statistical process. The mathematics driving it are not identical but share a lot in common with the way I presume this regression correction would have been programmed.

Posted by MWalpole on 10/1/2018 1:34:00 PM (view original):
I know it isn’t the main subject of this thread, but I don’t think anyone can argue that a D1 prospect signing with a D2 over a D1 is ridiculous. A high school basketball player from the state of Michigan, who is aspiring to play D1 ball, is going to sign with Eastern/Western/Central Michigan over Ferris State 100 times out of 100 if they showed even slight interest and offered a scholarship. They won’t care that Ferris State is an “A+” D2 team (they just won the D2 title this year, which is why I used them as an example).

Or a D3 team beating A+ UConn for a recruit. So realistic you guys!

https://www.whatifsports.com/hd/TeamProfile/Schedule.aspx?tid=14347. I tend to lean to the side of shoe. THEN explain to me how you beat a team by 29 then lose to them by 5? Does HC have an advantage at DII? I was told no! This isn't in game adjustment, but game to game adjustment. This is a SIM so no gameplanning or adjustment done like would have been if human coached. Makes no sense at all!!

You're actually arguing the opposite point from shoe. His point is that things are not variable enough. I thought about pointing out how often season series are split, but it didn't actually feel important to the theoretical aspects of the discussion.

Also, HCA exists at all levels. It just isn't very big.

His point is in in-game programming, like how he held team to 16 points in first half and then gave up 61 in second half. How the engine weighted the coin flip to even out the game. My point is my game should never have happened like that and season series against SIMS should not be that different if there is no adjustment done. If HCA isn't that big then no one can explain the difference in my game with AK FB besides. So yes my game didn't have in game fluctuation, but the outcome was so different then the first game so my game I scored 30 points less then game 1, so where was my adjustment, but not at halftime at the beginning of the game. Makes no sense