Comments for Win-Vector Blog http://www.win-vector.com/blog The Applied Theorist's Point of View Sun, 28 Feb 2010 17:07:12 +0000 http://wordpress.org/?v=2.9.2 hourly 1 Comment on Something I don’t get about business and bailouts by jmount http://www.win-vector.com/blog/2008/10/something-i-dont-get-about-business-and-bailouts/comment-page-1/#comment-1929 jmount Sun, 28 Feb 2010 17:07:12 +0000 http://www.win-vector.com/blog/?p=23#comment-1929 Update- one of the big businesses added value to its "to be paid on time you must off a discount scheme": they went to net-90 day payment. For you consultants out there Alan Weiss had a preemptive tactic for dealing with this. He found out that many billable departments had a policy that required them to always payment terms if a vendor discount was offered- so he offers a discount if they pay in full before the work is performed. Update- one of the big businesses added value to its “to be paid on time you must off a discount scheme”: they went to net-90 day payment. For you consultants out there Alan Weiss had a preemptive tactic for dealing with this. He found out that many billable departments had a policy that required them to always payment terms if a vendor discount was offered- so he offers a discount if they pay in full before the work is performed.

]]> Comment on Sorting Used in Anger by jmount http://www.win-vector.com/blog/2008/04/sorting-in-anger/comment-page-1/#comment-1883 jmount Wed, 17 Feb 2010 01:12:20 +0000 http://www.win-vector.com/blog/?p=13#comment-1883 This sort of problem with sorting seems like it has always been with us. For example: Bentley quoting Knuth: "in section 6.2.1 of his 'Sorting and Searching,' Knuth points out while the first binary search was published in 1946, the first published binary search without bugs did not appear until 1962" (Programming Pearls 2nd edition, "Writing Correct Programs", section 4.1, p. 34). See also: Nearly All Binary Searches and Mergesorts are Broken (2006) http://googleresearch.blogspot.com/2006/06/extra-extra-read-all-about-it-nearly.html This sort of problem with sorting seems like it has always been with us. For example: Bentley quoting Knuth: “in section 6.2.1 of his ‘Sorting and Searching,’ Knuth points out while the first binary search was published in 1946, the first published binary search without bugs did not appear until 1962″ (Programming Pearls 2nd edition, “Writing Correct Programs”, section 4.1, p. 34).

See also: Nearly All Binary Searches and Mergesorts are Broken (2006) http://googleresearch.blogspot.com/2006/06/extra-extra-read-all-about-it-nearly.html

]]> Comment on Google AdSense Channels IDs and the Cramer Rao Inequality by jmount http://www.win-vector.com/blog/2009/10/google-adsense-channels-ids-and-the-cramer-rao-inequality/comment-page-1/#comment-1784 jmount Tue, 19 Jan 2010 01:35:51 +0000 http://www.win-vector.com/blog/?p=898#comment-1784 <a href="#comment-1780" rel="nofollow">@John </a> A very interesting point. I would caution that the likelihoods you would want to calculate (say in the section 3.2.1 example) in a Bayesian procedure would essentially be functions of the summary statistics that are not very tightly related to the parameters you are trying to estimate (due to the unfortunate censoring process of only being allowed a constant number nearly co-linear measurements). So I think you run into similar problems. As the number of (hidden) data items gets large I would expect the Bayesian estimate to get near the linear algebra estimate (which is itself having trouble). I emphasize that the unfortunate sum-ups were not part of the estimation procedure, but part of the externally imposed problem structure. @John
A very interesting point. I would caution that the likelihoods you would want to calculate (say in the section 3.2.1 example) in a Bayesian procedure would essentially be functions of the summary statistics that are not very tightly related to the parameters you are trying to estimate (due to the unfortunate censoring process of only being allowed a constant number nearly co-linear measurements). So I think you run into similar problems. As the number of (hidden) data items gets large I would expect the Bayesian estimate to get near the linear algebra estimate (which is itself having trouble). I emphasize that the unfortunate sum-ups were not part of the estimation procedure, but part of the externally imposed problem structure.

]]> Comment on Google AdSense Channels IDs and the Cramer Rao Inequality by John http://www.win-vector.com/blog/2009/10/google-adsense-channels-ids-and-the-cramer-rao-inequality/comment-page-1/#comment-1780 John Mon, 18 Jan 2010 23:01:52 +0000 http://www.win-vector.com/blog/?p=898#comment-1780 This is exactly why you shouldn't be using unbiased estimation. Bayesian estimation avoids all these problems and is consistent with what your actually trying to do: maximize profit. This is exactly why you shouldn’t be using unbiased estimation. Bayesian estimation avoids all these problems and is consistent with what your actually trying to do: maximize profit.

]]> Comment on Relative returns: a banker versus trader paradox by Steve Cole http://www.win-vector.com/blog/2010/01/relative-returns-a-banker-versus-trader-paradox/comment-page-1/#comment-1753 Steve Cole Fri, 08 Jan 2010 00:52:59 +0000 http://www.win-vector.com/blog/?p=1296#comment-1753 So... *Is* there a reasonable language for *me*, the "I'm just barely paying attention" investor to use to rationally describe gains and losses? Somehow, saying "I had a 0.0458 dB gain" doesn't really flow off the tongue (and dB isn't exactly right, either, but we need some kind of unit). Or, maybe it does. If we use the EE definition of dB, and multiply your logarithms by 20, then the numbers actually enter the realm of useful for human thinking. That 10% loss is a drop of .916 dB. Hm. Now if I could just get my broker to give me those kind of numbers... So… *Is* there a reasonable language for *me*, the “I’m just barely paying attention” investor to use to rationally describe gains and losses? Somehow, saying “I had a 0.0458 dB gain” doesn’t really flow off the tongue (and dB isn’t exactly right, either, but we need some kind of unit).

Or, maybe it does. If we use the EE definition of dB, and multiply your logarithms by 20, then the numbers actually enter the realm of useful for human thinking. That 10% loss is a drop of .916 dB. Hm. Now if I could just get my broker to give me those kind of numbers…

]]> Comment on Statistics to English Translation, Part 2a: ’Significant’ Doesn’t Always Mean ’Important’ by Nina Zumel http://www.win-vector.com/blog/2009/12/statistics-to-english-translation-part-2a-%e2%80%99significant%e2%80%99-doesn%e2%80%99t-always-mean-%e2%80%99important%e2%80%99/comment-page-1/#comment-1671 Nina Zumel Thu, 17 Dec 2009 17:40:18 +0000 http://www.win-vector.com/blog/?p=1186#comment-1671 <a href="#comment-1670" rel="nofollow">@Yizhar Toren </a> You are absolutely correct. Thanks for the nice comments, and the additional reminder. @Yizhar Toren
You are absolutely correct. Thanks for the nice comments, and the additional reminder.

]]> Comment on Statistics to English Translation, Part 2a: ’Significant’ Doesn’t Always Mean ’Important’ by Yizhar Toren http://www.win-vector.com/blog/2009/12/statistics-to-english-translation-part-2a-%e2%80%99significant%e2%80%99-doesn%e2%80%99t-always-mean-%e2%80%99important%e2%80%99/comment-page-1/#comment-1670 Yizhar Toren Thu, 17 Dec 2009 11:37:27 +0000 http://www.win-vector.com/blog/?p=1186#comment-1670 Great writing man. Very crisp and clear. Just wanted to add an important reminder: ** Never forget that p-value is a wonderful way of answering the wrong question. ** As you've said we would like to answer the question: "what is the probability that groups a and b have the same distribution?" P-value answers the question: "how likely is it to get the difference we measure between the groups, when we assume that both groups have the same distribution?" The two answers are of course related (and a small p-value implies small probability of the two groups having the same distribution) but not necessarily in a simple fashion. For that you have to go Bayesian, but that's a completely different story... :) Great writing man. Very crisp and clear. Just wanted to add an important reminder:
** Never forget that p-value is a wonderful way of answering the wrong question. **

As you’ve said we would like to answer the question: “what is the probability that groups a and b have the same distribution?”
P-value answers the question: “how likely is it to get the difference we measure between the groups, when we assume that both groups have the same distribution?”

The two answers are of course related (and a small p-value implies small probability of the two groups having the same distribution) but not necessarily in a simple fashion. For that you have to go Bayesian, but that’s a completely different story… :)

]]> Comment on CRU graph yet again (with R) by Kevin Gaab http://www.win-vector.com/blog/2009/12/cru-graph-yet-again-with-r/comment-page-1/#comment-1663 Kevin Gaab Tue, 15 Dec 2009 01:45:35 +0000 http://www.win-vector.com/blog/?p=1195#comment-1663 A related blog post you might be interested: http://junkcharts.typepad.com/junk_charts/2009/12/the-real-climategate.html There was an interesting article in Science a while back that discussed why the three sigma variance in climate models will probably always be large. I'll have to find it. The argument, if I correctly understood it, was essentially that the propagation of uncertainties in all the component variables (and there are a lot of variables in these models, and they interact in complex, nonlinear ways) is such that the predicted uncertainties in temperature will ALWAYS be of the same magnitude as the mean change in temperature we're trying to model. It wasn't very hopeful that climate model accuracy will ever improve. A related blog post you might be interested:

http://junkcharts.typepad.com/junk_charts/2009/12/the-real-climategate.html

There was an interesting article in Science a while back that discussed why the three sigma variance in climate models will probably always be large. I’ll have to find it. The argument, if I correctly understood it, was essentially that the propagation of uncertainties in all the component variables (and there are a lot of variables in these models, and they interact in complex, nonlinear ways) is such that the predicted uncertainties in temperature will ALWAYS be of the same magnitude as the mean change in temperature we’re trying to model. It wasn’t very hopeful that climate model accuracy will ever improve.

]]> Comment on CRU graph yet again (with R) by Nina Zumel http://www.win-vector.com/blog/2009/12/cru-graph-yet-again-with-r/comment-page-1/#comment-1657 Nina Zumel Mon, 14 Dec 2009 08:02:21 +0000 http://www.win-vector.com/blog/?p=1195#comment-1657 <a href="#comment-1656" rel="nofollow">@jmount </a> : It isn't discussed explicitly in the article above (or in its follow-up article either), but the significance of a regression coefficient is taken against the null hypothesis that the coefficient is zero. In other words, PC5 and PC10 are the only coefficients that are really bounded away from zero in this analysis (well, PC4, PC7, and PC8, too, if you feel generous). That means that most coefficients -- especially the ones that actually matter, the first 3 -- have uncertainty that is of the same magnitude as their estimated value. It's just not a very good model. @jmount : It isn’t discussed explicitly in the article above (or in its follow-up article either), but the significance of a regression coefficient is taken against the null hypothesis that the coefficient is zero. In other words, PC5 and PC10 are the only coefficients that are really bounded away from zero in this analysis (well, PC4, PC7, and PC8, too, if you feel generous).

That means that most coefficients — especially the ones that actually matter, the first 3 — have uncertainty that is of the same magnitude as their estimated value. It’s just not a very good model.

]]> Comment on CRU graph yet again (with R) by jmount http://www.win-vector.com/blog/2009/12/cru-graph-yet-again-with-r/comment-page-1/#comment-1656 jmount Sun, 13 Dec 2009 19:56:56 +0000 http://www.win-vector.com/blog/?p=1195#comment-1656 By the way- the tiny principle components elements tricking the modeling system into thinking they are significant (when they are in fact unimportant noise, just not independent noise as expected by significance calculations) is great example of some of the issues discussed in: http://www.win-vector.com/blog/2009/12/statistics-to-english-translation-part-2a-’significant’-doesn’t-always-mean-’important’/ By the way- the tiny principle components elements tricking the modeling system into thinking they are significant (when they are in fact unimportant noise, just not independent noise as expected by significance calculations) is great example of some of the issues discussed in: http://www.win-vector.com/blog/2009/12/statistics-to-english-translation-part-2a-’significant’-doesn’t-always-mean-’important’/

]]>