Page 1 of 1

2048: "Curve-Fitting"

Posted: Wed Sep 19, 2018 6:21 pm UTC
by LucasBrown
Image
Title text: Cauchy-Lorentz: "Something alarmingly mathematical is happening, and you should probably pause to Google my name and check what field I originally worked in."

The text accompanying the piecewise plot is missing an ending quotation mark.

Re: 2048: "Curve-Fitting"

Posted: Wed Sep 19, 2018 7:24 pm UTC
by Soupspoon
What Randall doesn't know is that actually those dots are all points on a Radar Plot. Something to do with responses to Rorschach tests, I think.

Re: 2048: "Curve-Fitting"

Posted: Wed Sep 19, 2018 8:00 pm UTC
by Old Bruce
Soupspoon wrote:What Randall doesn't know is that actually those dots are all points on a Radar Plot. Something to do with responses to Rorschach tests, I think.

... probable responses to Rorschach tests...

Re: 2048: "Curve-Fitting"

Posted: Wed Sep 19, 2018 8:49 pm UTC
by Sableagle
2048.png

Re: 2048: "Curve-Fitting"

Posted: Wed Sep 19, 2018 8:51 pm UTC
by SAI Peregrinus
Reminds me a bit of the "Datasaurus Dozen" graphs.

https://www.autodeskresearch.com/publications/samestats

Re: 2048: "Curve-Fitting"

Posted: Wed Sep 19, 2018 9:44 pm UTC
by kkt
Is there a trick to the hidden text punchline? Lorentz was in theoretical physics, especially electromagnetism. Cauchy was a civil engineer working on a canal for Napoleon. Neither seem especially disreputable.

Re: 2048: "Curve-Fitting"

Posted: Wed Sep 19, 2018 10:30 pm UTC
by gmalivuk
It's not a joke about the dudes, it's about their probability distribution, which is in many ways "pathological" as a distribution, though I don't know how often it's used to fit curves.

Re: 2048: "Curve-Fitting"

Posted: Wed Sep 19, 2018 11:03 pm UTC
by Mjb
Ah, the eternal source of statistics jokes.

https://algotrading101.com/blog/1543426 ... timization
http://ucanalytics.com/blogs/machine-le ... ed-part-1/
https://xkcd.com/1347/
https://xkcd.com/1725/

An image search brings up plenty more opportunities for suspicious fitting - B-spline, devil's curve (if you can present your data in polar) - and the occasional guide on how to dress for one's body shape.

Re: 2048: "Curve-Fitting"

Posted: Wed Sep 19, 2018 11:09 pm UTC
by Flumble
I'm loving Sableagle's interpretation of the datas. :lol:

The "house of cards" diagram reminds me of when I first learned about polynomial interpolation. Despite some occasional overshooting they result in nice smooth motion ...until you try and see what prediction power it has and *BLAM* everything explodes.

Re: 2048: "Curve-Fitting"

Posted: Thu Sep 20, 2018 6:27 am UTC
by Graham Finch
But what message do the polynominal people send?

Re: 2048: "Curve-Fitting"

Posted: Thu Sep 20, 2018 1:35 pm UTC
by wumpus
Graham Finch wrote:But what message do the polynominal people send?


I thought "House of Cards" was what you got with a polynomial curve fitting (or at least a sufficiently large polynomial for such a good fit).

Re: 2048: "Curve-Fitting"

Posted: Thu Sep 20, 2018 9:36 pm UTC
by CorruptUser
This is why I just use a regression where appropriate, multivariate if there's sufficient data, and call it a day. Anything more is the mathematical equivalent of sophistry.

Re: 2048: "Curve-Fitting"

Posted: Thu Sep 20, 2018 10:52 pm UTC
by Archgeek
Mjb wrote:Ah, the eternal source of statistics jokes.

https://algotrading101.com/blog/1543426 ... timization
http://ucanalytics.com/blogs/machine-le ... ed-part-1/
https://xkcd.com/1347/
https://xkcd.com/1725/

An image search brings up plenty more opportunities for suspicious fitting - B-spline, devil's curve (if you can present your data in polar) - and the occasional guide on how to dress for one's body shape.

I love how the Teacher's T distribution kinda looks like Barad-dûr.

Re: 2048: "Curve-Fitting"

Posted: Fri Sep 21, 2018 1:12 am UTC
by ucim
The key to curve fitting is to ask the question "how necessary is the complication of this particular equation when I remove some data. That shows how robust the curve fit is. This can be taken to the limit of removing all the data; the resulting curve should be perfect.

Jose

Re: 2048: "Curve-Fitting"

Posted: Fri Sep 21, 2018 11:18 am UTC
by cellocgw
Sableagle wrote:2048.png
..mmmpphhh guess you'll have to look back to see the image I tried to quote.

It could be (and often is, these days) a lot worse: just imagine what the results would be if you tried to apply machine-learning to distinguish among those groupings.

Also, as the local representative of The World Convocation of Carls, I resent the implication :P

Re: 2048: "Curve-Fitting"

Posted: Fri Sep 21, 2018 11:10 pm UTC
by zjxs
LucasBrown wrote:Comic strip and alt-text.


I have done all of these. But not to the same data.

Re: 2048: "Curve-Fitting"

Posted: Tue Sep 25, 2018 3:07 pm UTC
by MarkGyver
Congrats on the big round-number milestone! Xkcd now needs 0xC bits to represent its comic numbers.

Re: 2048: "Curve-Fitting"

Posted: Thu Sep 27, 2018 12:46 am UTC
by phlip
kkt wrote:
Is there a trick to the hidden text punchline? Lorentz was in theoretical physics, especially electromagnetism. Cauchy was a civil engineer working on a canal for Napoleon. Neither seem especially disreputable.

It's a bit of an ambiguous pronoun, but I think it's supposed to mean that you'd want to Google the credentials of whoever is using the Cauchy-Lorentz distribution in a statistical analysis, rather than Cauchy and Lorentz themselves. Because if you're trying to use that distribution in an actual applied setting, you're probably either doing something very clever, or very wrong.

My favourite one of these was a couple years back, where someone plotted tournament attendance figures for some game, slapped on a 6th-degree-polynomial regression line (probably because that's the highest Excel will go) and saw it started to go down slightly at the end (because, of course, the ends are the most stable part of a high-degree polynomial fit) and then took to Twitter/reddit to chicken little about how the game had peaked and attendance was falling and the game was being mismanaged and etc etc and soapbox about all their favourite gripes with how the game was being run...

Re: 2048: "Curve-Fitting"

Posted: Fri Sep 28, 2018 12:56 am UTC
by teal4two
Post disapproved:

"2048: "Curve-Fitting""

Reason: The reported message has the only purpose to advertise for a website or another product. If you’re a bot, posting the last rejection is a giveaway. If you’re not a bot.... you kinda look like a bot right now..

Thu Sep 27, 2018 1:37 pm America/Los_Angeles
-----------------

????
I posted a link to a similar plot from the journal Science.
If this was handled by a human moderator, it should have been clear from the link that it was unlikely to be an ad. A human might actually have looked at the link and been able to make a rational decision.
So presumably this was addressed by AI, with only the A being appropriate.

PS I am not even that sure of what a bot is, much less spend my time studying their behavior.

Re: 2048: "Curve-Fitting"

Posted: Fri Sep 28, 2018 1:44 am UTC
by SecondTalon
Read

The goddamn

Rules

The Goddamn Rules wrote:3. You cannot post any links before you have reached five posts.

Please do not post any links or perform any form of self-promotion until you have a post count of at least five. We have our reasons. Not all spammers are bots, and it can be hard to tell the difference between somebody pointing out somebody else's cool webpage, somebody advertising out their cool webpage, or someone posting malicious links. Because of this, if you have any links in your first post, the entire post will be deleted. Also, pushing your post count up to five with useless posts and then posting your link and vanishing will go over poorly.

“But my link was relevant to the conversation and I think by post was blocked without a moderator even reading it!”

Oh. Gosh. Whatever shall we do. A moderator didn’t read an entire post before blocking it. Gosh. If only there were rules explaining this that should have been read first. Gee golly. What a shame.

That said, there is a list of acceptable and unacceptable words or fragments that first time posters can and cannot use. www will be immediately flagged as spam, as will .com or http. This means that if you fail to hit space between a period and the word commence, it will be blocked as spam as it hits the ".com" flag. You will get a message stating "This message has been flagged as spam and has been blocked" if you attempt to do so. This will go away once you have a few posts under your belt. We do apologize for this inconvenience, but your momentary inconvenience is outweighed by our convenience of not having to block 500+ spam messages a day.

And no, the list of blocked words will not be made public.