[This post was inspired by this one]

You may have noticed that I award the occasional bounty - I like to give one when I see a particularly good answer or a question I think should get an answer. (What do I need 125K reputation for? It's not doing me any good there.)

However, I only see a fraction of the posts on site, and while I have broad interests, I don't pay close attention to every kind of question. So it's a sure thing that there are many good answers I haven't seen, or that I saw once but have since forgotten.

I'd like to see suggestions of good posts that deserve a bounty (one per answer*) -- answers that more people should see, or even questions that it would be good for the site to get a good answer to. Feel free to upvote any you agree with (whether you have nominated some or not); upvotes will certainly influence the judges but are not the only factor (encouraging good answers from newer users is another, for example)

* you can nominate more than one in separate answers, but for this exercise you can't nominate your own questions or answers (though I may do a self-nomimation one at a later time). You also can't nominate my answers (well, okay, you actually can -- I can't stop you -- but I couldn't bounty myself even if I wished to).

In a few days from now, if there have been more than zero nominations, I will place the first bounty (it takes at least a day to before you can award one, though - longer if you let them stay in the queue to get exposure).

A few days after placing that, I'll look here for another to start - up to the limit of available free slots for (maximum of three for any one user). As slots free up I'll add more bounties.

As a bounty is awarded, I'll mark it here (in the answer). I may telegraph my intent on some depending on how things go.

A typical bounty I'd seek to offer for a good answer will be 100 points, but I may offer more or less as circumstances suggest (you can also suggest a higher bounty when you nominate, or in comments under a nomination). I will generally try to avoid the highest amounts (500 is possible under the bounty system but I don't want to outshine other people's bounties too much; all bounties deserve good attention).

Here are some things you might consider if you're having trouble coming up with one (but you can nominate for other reasons):

• Has an answer been especially helpful to you or given some special insight?

• Is there an answer where you thought "why didn't the textbooks explain it like this?"

• Is there an answer you wish you could show everyone?

• Is there a great question you really wish had an answer?

That's the sort of thing that needs promotion, and would be a good post to suggest.

I hope to spend at least 1000 points (i.e. perhaps about 10 bounties give or take) this way -- to award that number of bounties would take several weeks (since deserving posts should have some time in the bounty queue spotlight so people can see them) -- however, if it seems to be working well, I'd run it longer. [NB at this stage the total points awarded to nominated posts exceeds that target, which counts as a success.]

[If anyone else feels particularly that they want to offer bounties on posts that come up, feel free -- but try to keep in mind that we shouldn't flood the queue, so I think we should keep it to no more than say five bounties running at one time across all nominations here. If you want to claim one in advance of setting the bounty - because you want to wait a few days - just say so on the post. If you do award a bounty, mark the post you bountied as was done in the linked post.]

• Hi @Tim -- that's fine even if it doesn't motivate people, since that's not my prime motivation for the bounties on answers (which is mostly better visibility); however, it may motivate early participators a bit more -- I know bounties made a difference to my perception of reward-per-effort when I otherwise found points hard to come by. It could be the case for pulling in answers on questions that lack answers (I already put plenty of bounties on questions regularly -- I know already that it's not always effective). ... ctd Feb 2, 2016 at 9:34
• @Tim ctd ... Anyway, worst case is I spend a thousand or so points; I've already spent about 8 times that on bounties, it's not like I don't know what I'm getting into. Feb 2, 2016 at 9:36
• Even if it just makes people feel happier about being on site that would be a good thing, I think. Feb 2, 2016 at 10:14
• @Tim: would you say that active users on other SE sites are relatively more motivated by points? I can say for myself that they do motivate me - I tend to answer a lot more questions when a recent answer of mine seems to have been appreciated than if I put effort into one that was largely ignored (as measured by upvotes). Feb 2, 2016 at 13:12
• @Tim Whether you're right or not, the exercise is the same here; if you don't wish to nominate any posts, you're free not to. It won't cost you a thing. If you don't think it will do any good, but for some reason want to suggest a post anyway, go ahead; it still won't cost you anything. Feb 2, 2016 at 13:52
• @Glen_b forget it, it was just an off-topic comment, in general I like your idea :)
– Tim Mod
Feb 2, 2016 at 14:06
• I wish I could apply the featured tag to this post. I think it has huge benefits for the community and could do with more attention.
– Sycorax Mod
Feb 3, 2016 at 16:20
• @user777: flagged for mod attention, requested tagging this as featured. Let's see what the mods think. Feb 7, 2016 at 2:56
• One might consider an indicating metric like question upvotes to answers ratio as an initial way to select a good question/good answer rich sample. Feb 7, 2016 at 13:52
• I reviewed many Q&A that I find good and interesting and I must say that it seems that the common power-law rule is true for them: vast majority of the good answers are made by small minority of answeres from the CV's top-50.
– Tim Mod
Feb 18, 2016 at 8:14
• It looks like 1000 points of bounty were started by the end of March (a little under 9 weeks from the initial post), of which 850 was successfully awarded to a post. I'm very pleased with the way that worked out. Apr 3, 2016 at 1:44
• Glen: bounty this: stats.stackexchange.com/a/410543/805 May 29, 2019 at 2:27

The first clustering algorithm you will learn about is k-means. That is nice and good, but unfortunately people will sometimes think that k-means is the One Tool to solve all their clustering problems and neglect finding out about drawbacks and alternatives to k-means.

In such cases, I find both Anony-Mousse's and David Robinson's answers to the question How to understand the drawbacks of K-means invaluable. Both answers are crafted with love, with lots of wonderful graphics, and I find myself recommending this question and both answers over and over. Both definitely qualify as "answers I wish I could show everyone."

[First of two bounties begun 9 Feb 2016. 2nd bounty started 15 Feb. Now awarded.]

• Yes, though this thread (the Q and both answers) has staggering amount of views and upvotes already. Feb 3, 2016 at 10:26

We have a question on Anscombe-like datasets with the same box and whiskers plot (mean/std/median/MAD/min/max), which currently has five upvotes and no answers.

I like Anscombe's quartet a lot, since it wonderfully illustrates how summary univariate or multivariate statistics lose (you can also say "compress") information. Frank Harrell's example of very different data leading to identical dynamite plots is also very instructive.

Therefore, I'd really love to see an answer to the question linked above for pedagogical reasons, and I think encouraging this through a bounty or similar would be worthwhile.

[Started 27 March 2016. No answer was received in the bounty period, so it went unawarded.]

• I spent several hours over a couple of days working on an answer to that before life intervened and I forgot about it. I think it's a soluble problem. Feb 3, 2016 at 9:58
• I think so, too, but I also suffer from this condition called "life". I'd love to see one of your nice answers on this question. Feb 3, 2016 at 10:08
• In any case, in the meantime it could definitely stand to have a bounty. Feb 3, 2016 at 10:39
• I've been pondering about how to get an answer on that question. If one doesn't arise from the bounty, rather than wait until I have a full answer, I might later post some thoughts on a partial solution as a CW-answer in the hopes that others will improve it. Mar 27, 2016 at 1:42
• Stephan Kolassa, you might appreciate my modest contribution to Anscombe's quartet stats.stackexchange.com/questions/89547/… Jan 5, 2019 at 17:45

Antoni Parellada's answer in this question! Why is ANOVA equivalent to linear regression?

With nice graphical demonstration and math!

[Awarded 21 May 2017]

• I did not bounty this right away as I had just given Tony a bounty. I've only just remembered to come back to it. May 16, 2017 at 6:13

Getting Information out of Blackbox Models - RandomForest / XGBoost

Although I closed the thread as a duplicate (and I continue to think that was correct), the answer is really good and has not gotten the votes it deserves, IMO.

Update:
(Actually, it isn't clear that it's possible to add a bounty to a closed thread. That makes @amoeba's argument for merging stronger.)

[Bounty started 11 Feb 2016. Now awarded.]

• Should this thread maybe be merged into the other one so that the answer gets moved there? I always find it unfortunate and unhelpful when there is a really good answer in a thread closed as a duplicate. Feb 9, 2016 at 10:19
• It has been merged, thanks @whuber. stats.stackexchange.com/questions/172761. Feb 10, 2016 at 0:28
• @Glen_b, the answer I meant was Soren Havelund Welling's. Your comment to the bounty names RUser4512's. I don't know if you want to / can change that. Feb 10, 2016 at 21:30
• @gung Thanks for pointing it out. That appears to have been a hiccup with my browser as I scrolled down to see who the author was -- there was a short delay as I scrolled and it seems to have jumped from somewhere in the middle of the answer you linked to the middle of the next one as I scrolled (the similarity of some of the plots helped me to think nothing was amiss). I can't change the message. I'll roll back the edit to your post so that I also place a bounty on the one you wanted; Soren's answer will get a 200 point bounty as a result. Feb 10, 2016 at 21:55
• @Glen_b: Just to see if I understand it correctly: so after you wrote an erroneous bounty description message you considered yourself morally obliged to award it RUser4512? You could still have awarded it to Soren, despite the message. Feb 12, 2016 at 0:02
• @amoeba I could have, but as you say, I felt a moral obligation to do as I had said. It's decent post though, I'd happily have given it 50 anyway. If it had not been a good post I'd have awarded it where it should have gone and posted an apology in comments under the other answer. Feb 12, 2016 at 0:06
• @Glen_b, by the "Now awarded", are you indicating I should delete? Feb 17, 2016 at 23:56
• @gung No. I'd actually prefer you didn't delete. Feb 18, 2016 at 0:07

Is there an intuitive interpretation of $A^TA$ for a data matrix $A$? should be definitively awarded as it provides short but clear and pretty detailed answer for the very basic and very important question.

[Bounty started 7 April 2016, and has since been awarded]

• Yes, that's an excellent suggestion. Apr 7, 2016 at 9:49
• agreed that answer saved my a while back Sep 5, 2018 at 16:31

Our soon-to-be-infamous question At each step of a limiting infinite process, put 10 balls in a urn and remove one at random. How many balls are left? is spawning its share of answers I wish I could show everyone (along with the, ahem, other kind). In particular:

Yes, neither one needs bounty rep. But both answers definitely pass the first three bullet points test in this meta question.

(bounty awarded to amoeba)

• whuber has asked me not to give him bounties, sorry (and hasn't yet indicated a change of heart). I agree that this answer (like many of his answers) deserves one. I will have to try to content myself with a mere upvote there. I have placed a bounty for amoeba's answer Nov 27, 2017 at 21:59
• Stephan and @Glen_b thanks. I am planning to share one half of Glen's bounty with whuber :-) Nov 28, 2017 at 8:12
• I wonder if whuber's answer is the highest upvoted deleted answer on our site (Cc to @Glen_b). Nov 29, 2017 at 7:56
• I've been working on an answer that addresses "physicality" fallacies which have been appearing throughout the question's answers. No bounty-begging here, but I am hoping some people see it and enjoy. It has been fun to write! stats.stackexchange.com/a/316129/11646
– Paul
Dec 1, 2017 at 16:15
• I'd like to share part of @Glen_b's generous bounty with some other answer(s), so Stephan and Glen_b if you have some suggestions, let me know. Dec 2, 2017 at 9:19
• Also, ekvall's answer IMHO definitely needs to have more net votes than Neil's one so check your votes :) Dec 2, 2017 at 9:20

I think that lacerbi's answer deserves some recognition because its bibliography is incredibly good. The articles cited here are incredibly helpful to me and I had not previously encountered them.

[A bounty has been started on this answer, 5 Feb 2016. Now awarded.]

I do not know if it is a memorable answer by the overall standards of CV, but the accpeted one here is the one I found most useful of all questions I have asked:

Example for a prior, that unlike Jeffreys, leads to a posterior that is not invariant

[Bounty started 9 Feb 2016. Now awarded.]

[Bounty started 27 March 2016. Now awarded]

• I agree that this is a worthy answer, but I've already awarded David a bounty (to another already-highly-upvoted) answer in this round, so I may leave a bit of a gap before awarding another to a similarly upvoted post. Feb 17, 2016 at 23:34
• @Glen_b ok, this policy makes sens :)
– Tim Mod
Feb 18, 2016 at 8:10
• Even if it's only a week or so perhaps -- I only just awarded the other in the last day or so. Feb 18, 2016 at 8:13

Glen,

Thank you very much for your largess!

I constantly go back to some posts that have helped me understand certain concepts. A good number of them are yours...

Right off the top of my head:

1. @gung's post on the linearity of polynomial regression with absolutely self-explanatory plots is fabulous.

2. @amoeba got a prior 50-pt bounty for one of the all-time classics on understanding PCA, a post worthy of higher recognition.

3. Recently I found @ttnphns' post on the geometric interpretation of PCA loadings very useful.

To be clear, I am abstaining from recommending any of the many unbelievable posts by @whuber because of his condition as moderator.

[Bounty started on #1. Awarded 30th May 2017]

• +1 for other suggestions, but I definitely don't need another bounty on that PCA answer: it has well enough upvotes & visibility :) May 8, 2017 at 9:32
• The Q calls for one nomination per post. I'll take your first. Even if amoeba had not demurred I would not have awarded one there - while it's well deserving of recognition, I think it has that already, the idea is to spread the recognition around to good posts that don't have the recognition they should -- already-bountied posts will usually be out of the running May 28, 2017 at 23:43

This answer comparing OLS algorithms is a well-written introduction into the topic.

[Bounty awarded Jun 20]

• Ah, so that's why I got that bounty; thanks! Jun 20, 2016 at 13:42
• You're welcome! cheers
– Sycorax Mod
Jun 20, 2016 at 13:44

How to represent kWh usage by year against average temperature is a wonderful illustration of how to build scientific knowledge into a model. (In this case it's science that many of us know but wouldn't think to use nevertheless.)

• [Scortchi placed a bounty on the post mentioned here] Jun 21, 2016 at 8:54

https://stats.stackexchange.com/users/84394/bill-woessner

One could probably pick any of this user's answer for a bounty (as I have done already for his most-upvoted answer). The one comparing eigen- and Cholesky-factorization is particularly good.

[Bounty placed and awarded.]

• Thanks. Any specific suggestions would be welcome. Aug 3, 2016 at 21:26
• @Glen_b (1) This one takes a deep dive into some FORTRAN code. (2) This one is some useful linear algebra.
– Sycorax Mod
Aug 3, 2016 at 21:52
• Placed a bounty on the second, since the first already had one. Aug 9, 2016 at 18:48

Matthew wrote a good answer here about the computational steps involved in linear regression using linear algebra.

Least Squares Regression Step-By-Step Linear Algebra Computation

[Bounty placed and awarded]

I think that amoeba's answer on "Building an autoencoder in Tensorflow to surpass PCA" is extremely elegant and insightful; it provides an excellent and coherent example on how an auto-encoder compares to PCA.

To quote another user's comment: "Fantastical! Stupendousness!" - I agree with that assessment.

• Thank you. (I will not vote up or down for obvious reasons.) To be honest, at the moment that answer has too much code and not enough explanations to be really considered insightful. I was not planning on working on it any further but if somebody decides to put a bounty on that thread, I will try to add some additional comments. (It's a bit annoying though that OP never left any comments let alone marked my answer as accepted.) Nov 6, 2017 at 23:03
• @amoeba: I think it is great for someone who looks for solid examples. As a side-note: I found it also extremely portable; my Python is rusty and I recoded in R within minutes and it works fine. While vastly unappreciated in the expense of "code-fu" and adherence to style, clarity of presentation is important and your post is excellent at that. Nov 6, 2017 at 23:17
• Oh, I did not even realize that Keras/Tensorflow are available in R. Nov 7, 2017 at 8:10
• bounty started. Nov 16, 2017 at 23:43
• Bounty now awarded Nov 23, 2017 at 0:19
• @Glen_b: Thanks. I added an R version of amoeba's example too. I waited till the bounty got awarded to avoid any confusion. Nov 23, 2017 at 1:55

Alecos Papadopoulos wrote this really nice answer to the question "Testing for autocorrelation: Ljung-Box versus Breusch-Godfrey". It's one of the most solid posts I have seen so far on CV, providing a well-researched canonical answer.

• Yes. that's a good one. Bounty started. Nov 16, 2017 at 23:46
• Bounty now awarded. Nov 23, 2017 at 0:19

Wolfies made an especially valuable contribution to this thread, in which s/he traced back a particular distribution to its originating author. Doing so provided direct refutation to the Wikipedia article.

Does this distribution have a name? $f(x)\propto\exp(-|x-\mu|^p/\beta)$

[bounty started 14 March 2016. Now awarded.]

• Since Wikipedia, like CV, is made up of popular contribution, Wolfies loses my vote for pointing out the error, but not editing (or even linking to!) the Wikipedia article. Jan 5, 2019 at 18:10

Brian Borchers provided very nice and detailed answer comparing gradient descent to other optimization methods in machine learning describing when and why it is used.

[Bounty awarded 23 May 2016]

What is the difference between a loss function and decision function?

MånsT answers nicely about the difference between loss function and decision function. The question is trivial, but the answer provides condense but clear introduction decision-theoretic thinking in statistics.

[Bounty awarded 19 Jun 2016]

Rob Hyndman's answer to his own question How should I transform non-negative data including zeros? at https://stats.stackexchange.com/a/1630/60613 brings a nice transformation function that I had never heard being mentioned. There are other great answers in that thread, but this was the one that most helped me (I was actually searching for another transformation for positive/negative data, and the arcsinh really suited my application as it resembles the log1p transformation).

[Bounty awarded]

• Speaking for myself, I found it a fairly obvious transformation to consider, but he's right that people are unaware of it (for reasons that aren't really clear to me -- do people not do hyperbolic trig in school any more, or do they just never look at the function?). On the other hand, I think some reasons why it's not used more widely among people who are aware of it aren't so hard to guess at. In any case, I agree it's a very good answer, and indeed, I had considered putting a bounty on it before. I have done so now, thanks. Jul 21, 2016 at 18:22
• @Glen_b Thanks! I totally studied hyperbolic functions during secondary school, but when I started studying machine learning during my undergrad I didn't really associate both. I of course knew the tanh was useful for data transformation, but I was completely oblivious to how the asinh could be applied similarly to a log transformation. Jul 27, 2016 at 11:16
• $\sinh$ is odd, like $x$ near 0 and goes to $\frac12 e^x$ for large $x$; its inverse is therefore going to be odd, look like $\log(x)-\log(\frac12)$ for large $x$ and like $x$ for $x$ near $0$. This led me to try it as a transformation for data that's both positive and negative and heavy-tailed in both directions as an undergrad, way back when. It sometimes works fairly well but it's not easy to explain to people. May 28, 2017 at 23:40

I think that kjetil b halvorsen answer on "Why does logistic regression become unstable when classes are well-separated?" is very well-thought and addresses a common misconception/issue regarding logistic regression applications.

[Bounty awarded 30th May 2017]

Benjamin Christoffersen's answer to the question on: Residual standard error difference between optim and glm is deceptively simple, under-appreciated and requiring a solid appreciation on how standard errors actually are computed.

• deceptively rather than deceiving, I suggest. Dec 12, 2018 at 1:57
• Ooops! Auto-correct! Dec 12, 2018 at 20:01

I haven't seen any suggestions for a question so far (only answer), but I would like to break the ice with understanding logistic regression p-values for random predictors

[minor edit]

The problem it addresses is very interesting. The post is not great (too much code and too many questions for a single post, I asked the OP to fix both) but the question behind the post is in my opinion intriguing. Also, the user seems to be fairly new but he's already asking an interesting question, so I think it would deserve a little encouragement.

• -1 because I think this suggestion is off-topic here as Glen_b explicitly asked about answers. Dec 4, 2017 at 11:11
• @amoeba no; in fact the post explicitly allows both questions and answers, and one question has already been bountied -- see this one. Of course, the bounty can only be awarded to an answer to the bountied question -- if none get posted, the bounty is gone. Dec 4, 2017 at 11:14
• @amoeba hey, that's unfair! Glen_b specifically wrote "I'd like to see suggestions of good posts that deserve a bounty (one per answer*) -- answers that more people should see, or even questions that it would be good for the site to get a good answer to. " Dec 4, 2017 at 11:17
• @Glen_b ah, ok, like I said! I didn't see the Anscombe question. So it looks like the ice had already been broken :-) Dec 4, 2017 at 11:19
• I don't think amoeba was unfair; it's easy to miss. (I did the first read-through and I wrote the thing.) Dec 4, 2017 at 11:22
• @Glen_b Thanks for the clarification. DeltaIV, could you please make any edit to this post so that I could remove the downvote? In my defense, bountying unanswered questions is not really "spreading the wealth" as the title of this thread says; it's a different bounty reason (from the official list of bounty reasons) altogether. That's why I assumed that this thread refers to only one of those reasons, namely rewarding existing answers. Dec 4, 2017 at 11:40
• @amoeba done. No problem, and I know downvotes here don't affect reputation, I just didn't want to appear as someone who hadn't read the question :-) Dec 4, 2017 at 13:06
• Downvote removed. However, this question is less than 48 hours old which means it's not possible to put a bounty on it. This is for the good reason - many questions on CV get answered after a day or two. Usually it makes sense to curate the question, edit it a couple of times, and try to improve, before putting a bounty. IMHO it does not make sense to post an answer here suggesting for Glen to put a bounty there before all these normal routes of attracting attention to the question are exhausted. Dec 4, 2017 at 13:08
• @amoeba yes, yes, I know...I do admit that "the post is not great", and I did ask the OP to improve it. In case s/he improves it, then it might become worthy of a bounty, but I wasn't suggesting to put a bounty right now. Also, I respect the fact that it's an "absolute first" post - to me it seems a very interesting question, for someone who has never used Stack Exchange before. Anyway, I didn't wan't to create an argument - I can delete this answer of mine, ok? I don't think it's that bad, but if you feel strongly against it, I will. Dec 4, 2017 at 13:22