Pablo_Stafforini

• Pablo_Stafforini 19 Mar 2019 14:41 UTC

  12 points

  on: Un­der­stand­ing in­for­ma­tion cascades

  Two re­cent ar­ti­cles that re­view the ex­ist­ing eco­nomic liter­a­ture on in­for­ma­tion cas­cades:

  • Sushil Bikhchan­dani, David Hir­sh­leifer and Ivo Welch, In­for­ma­tion cas­cades, The new Pal­grave dic­tio­nary of eco­nomics (Macmil­lan, 2018), pp. 6492-6500.

  • Ok­sana Do­herty, In­for­ma­tional cas­cades in fi­nan­cial mar­kets: re­view and syn­the­sis, Re­view of be­hav­ioral fi­nance, vol. 10, no. 1 (2018), pp. 53-69.

  An ear­lier re­view:

  • Maria Grazia Ro­mano, In­for­ma­tional cas­cades in fi­nan­cial eco­nomics: a re­view, Gior­nale degli Economisti e An­nali di Econo­mia, vol. 68, no. 1 (2009), pp. 81-109.

• Pablo_Stafforini 19 Mar 2019 14:17 UTC

  14 points

  on: Un­der­stand­ing in­for­ma­tion cascades

  In­for­ma­tion Cas­cades in Multi-Agent Models by Arthur De Vany & Cassey Lee has a sec­tion with a use­ful sum­mary of the rele­vant eco­nomic liter­a­ture up to 1999. (For more re­cent overviews, see my other com­ment.) I copy it be­low, with links to the works cited (with the ex­cep­tion of Chen (1978) and Lee (1999), both un­pub­lished doc­toral dis­ser­ta­tions, and De Vany and Walls (1999b), an un­pub­lished work­ing pa­per):

  A sem­i­nal pa­per by Bikhchan­dani et al (1992) ex­plains the con­for­mity and frag­ility of mass be­hav­ior in terms of in­for­ma­tional cas­cades. In a closely re­lated pa­per Ban­er­jee (1992) mod­els op­ti­miz­ing agents who en­gage in herd be­hav­ior which re­sults in an in­effi­cient equil­ibrium. An­der­son and Holt (1997) are able to in­duce in­for­ma­tion cas­cades in a lab­o­ra­tory set­ting by im­ple­ment­ing a ver­sion of Bikhchan­dani et al (1992) model.

  The sec­ond strand of liter­a­ture ex­am­ines the re­la­tion­ship be­tween in­for­ma­tion cas­cades and large fluc­tu­a­tions. Lee (1998) shows how failures in in­for­ma­tion ag­gre­ga­tion in a se­cu­rity mar­ket un­der se­quen­tial trad­ing re­sult in mar­ket volatility. Lee ad­vances the no­tion of “in­for­ma­tional avalanches” which oc­curs when hid­den in­for­ma­tion (e.g. qual­ity) is re­vealed dur­ing an in­for­ma­tional cas­cade thus re­vers­ing the di­rec­tion of in­for­ma­tion cas­cades.

  The third strand ex­plores the link be­tween in­for­ma­tion cas­cades and heavy tailed dis­tri­bu­tions. Cont and Bouchaud (1998) put for­ward a model with ran­dom groups of imi­ta­tors that gives rise to stock price vari­a­tions that are heavy-tailed dis­tributed. De Vany and Walls (1996) use a Bose-Ein­stein al­lo­ca­tion model to model the box office rev­enue dis­tri­bu­tion in the mo­tion pic­ture in­dus­try. The au­thors de­scribe how sup­ply adapts dy­nam­i­cally to an evolv­ing de­mand that is driven by an in­for­ma­tion cas­cade (via word-of-mouth) and show that the dis­tri­bu­tion con­verges to a Pareto-Lévy dis­tri­bu­tion. The abil­ity of the Bose-Ein­stein al­lo­ca­tion model to gen­er­ate the Pareto size dis­tri­bu­tion of rank and rev­enue has been proven by Hill (1974) and Chen (1978). De Vany and Walls (1996) pre­sent em­piri­cal ev­i­dence that the size dis­tri­bu­tion of box office rev­enues is Pareto. Sub­se­quent work by Walls (1997), De Vany and Walls (1999a), and Lee (1999) has ver­ified this find­ing for other mar­kets, pe­ri­ods and larger data sets. De Vany and Walls (1999a) show that the tail weight pa­ram­e­ter of the Pareto-Levy dis­tri­bu­tion im­plies that the sec­ond mo­ment may not be finite. Lastly, De Vany and Walls (1999b) have shown that mo­tion pic­ture in­for­ma­tion cas­cades be­gin as ac­tion-based, non­in­for­ma­tive cas­cades, but un­dergo a tran­si­tion to an in­for­ma­tive cas­cade af­ter enough peo­ple have seen it to ex­change “word of mouth” in­for­ma­tion. At the point of tran­si­tion from an un­in­formed to an in­formed cas­cade, there is loss of cor­re­la­tion and an on­set of tur­bu­lence, fol­lowed by a re­cov­ery of week to week cor­re­la­tion among high qual­ity movies.

• Pablo_Stafforini 19 Mar 2019 13:29 UTC

  4 points

  in reply to: Jan_Kulveit's comment on: Un­der­stand­ing in­for­ma­tion cascades

  In­for­ma­tion cas­cades does not seem to be the best choice of key­words.

  I wouldn’t say that ‘in­for­ma­tion cas­cades’ isn’t the best choice of key­words. What’s hap­pen­ing here is that the same phe­nomenon is stud­ied by differ­ent dis­ci­plines in rel­a­tive iso­la­tion from each other. As a con­se­quence, the phe­nomenon is dis­cussed un­der differ­ent names, de­pend­ing on the dis­ci­pline study­ing it. ‘In­for­ma­tion cas­cades’ (or, as it is some­times spel­led, ‘in­for­ma­tional cas­cades’) is the name used in eco­nomics, while net­work sci­ence seems to use a va­ri­ety of re­lated ex­pres­sions, such as the one you men­tion.

• Pablo_Stafforini 15 Mar 2019 19:35 UTC

  4 points

  in reply to: habryka's comment on: Un­der­stand­ing in­for­ma­tion cascades

  Thanks, Oli!

• Pablo_Stafforini 15 Mar 2019 16:28 UTC

  3 points

  in reply to: Pablo_Stafforini's comment on: Un­der­stand­ing in­for­ma­tion cascades

  [meta] Not sure why the link to the overview isn’t work­ing. Here’s how the com­ment looks be­fore I sub­mit it:

  https://​imgur.com/​MF5Z2X4

  (The same prob­lem is af­fect­ing this com­ment.)

  In any case, the URL is:

  https://​aiim­pacts.org/​ev­i­dence-on-good-fore­cast­ing-prac­tices-from-the-good-judg­ment-pro­ject-an-ac­com­pa­ny­ing-blog-post/​

• Pablo_Stafforini 15 Mar 2019 16:20 UTC

  11 points

  in reply to: Davidmanheim's comment on: Un­der­stand­ing in­for­ma­tion cascades

  Thanks for this.

  Re ex­trem­iz­ing, the re­cent (ex­cel­lent) AI Im­pacts overview of good fore­cast­ing prac­tices notes that “more re­cent data sug­gests that the suc­cesses of the ex­trem­iz­ing al­gorithm dur­ing the fore­cast­ing tour­na­ment were a fluke.”

• Pablo_Stafforini 1 Mar 2019 17:52 UTC

  13 points

  on: Un­con­scious Economies

  Here’s an in­sight I had about how in­cen­tives work in prac­tice, that I’ve not seen ex­plained in an econ text­book/​course.

  There are at least three ways in which in­cen­tives af­fect be­havi­our: 1) via con­sciously mo­ti­vat­ing agents, 2) via un­con­sciously re­in­forc­ing cer­tain be­havi­our, and 3) via se­lec­tion effects. I think per­haps 2) and prob­a­bly 3) are more im­por­tant, but much less talked about.

  Jon Elster dis­t­in­guishes these three differ­ent ways in Ex­plain­ing So­cial Be­hav­ior. He first draws a dis­tinc­tion be­tween 1-2 (“re­in­force­ment”) on the one hand, and 3 (“se­lec­tion”), on the other. He then draws a fur­ther dis­tinc­tion be­tween 1 (“con­scious ra­tio­nal choice”) and 2 (“un­in­ten­tional choice”). Here are the rele­vant ex­cerpts from ch. 11 (em­pha­sis in the origi­nal; I have added num­bers in square brack­ets to make the cor­re­spon­dence be­tween your dis­tinc­tions and his more con­spicu­ous):

  In this chap­ter, I dis­cuss ex­pla­na­tions of ac­tions in terms of their ob­jec­tive con­se­quences… There are two main ways in which this can hap­pen: by re­in­force­ment [1-2] and by se­lec­tion [3]… If the con­se­quences of given be­hav­ior are pleas­ant or re­ward­ing, we tend to en­gage in it more of­ten; if they are un­pleas­ant or pun­ish­ing it will oc­cur less of­ten. The un­der­ly­ing mechanism could sim­ply be con­scious ra­tio­nal choice [1], if we no­tice the pleas­ant or un­pleas­ant con­se­quences and de­cide to act in the fu­ture so as to re­peat or avoid re­peat­ing the ex­pe­rience. Often, how­ever, the re­in­force­ment can hap­pen with­out in­ten­tional choice [2].

• Pablo_Stafforini 23 Feb 2019 11:41 UTC

  3 points

  in reply to: Raemon's comment on: Min­i­mize Use of Stan­dard In­ter­net Food Delivery

  In game the­ory, the costs and benefits in terms of which defec­tion is defined oc­cur in a well-defined con­text of strate­gic in­ter­ac­tion. My ob­jec­tion was to the use of defec­tion in a way that im­plied that the situ­a­tion de­scribed in the post had a par­tic­u­lar game-the­o­retic struc­ture, when in fact no clear ac­count was given of what that struc­ture was sup­posed to be.

• Pablo_Stafforini 22 Feb 2019 11:48 UTC

  2 points

  on: What’s your fa­vorite LessWrong post?

  Feed the Spinoff Heuris­tic!

• Pablo_Stafforini 22 Feb 2019 11:45 UTC

  2 points

  on: What’s your fa­vorite LessWrong post?

  Para­psy­chol­ogy: the con­trol group for science

• Pablo_Stafforini 11 Feb 2019 21:48 UTC

  3 points

  in reply to: gjm's comment on: Min­i­mize Use of Stan­dard In­ter­net Food Delivery

  The view you ar­tic­u­late is perfectly in­tel­ligible. I’m just not sure it cor­re­sponds to the view ex­pressed in the OP. Why in­voke no­tions like defec­tion, if all you want to say is that you should not im­pose a great cost on oth­ers when you can do so at a small cost to your­self?

• Pablo_Stafforini 11 Feb 2019 14:25 UTC

  10 points

  in reply to: habryka's comment on: Min­i­mize Use of Stan­dard In­ter­net Food Delivery

  I think the point of the OP is not to en­courage peo­ple to go to phys­i­cal restau­rants, but to dis­cour­age the use of on­line de­liv­ery ser­vices rel­a­tive to other ways of plac­ing or­ders . As they write (bold­faced added):

  If you like the restau­rant and want those work­ing there to earn a liv­ing, and the place to con­tinue to ex­ist, do not or­der via on­line ser­vices like Seam­lessWeb, GrubHub, De­liv­ery.com or Caviar, if there is an­other way to con­tact the restau­rant.

  I do find the post con­fus­ing in cer­tain ways; for ex­am­ple, the fol­low­ing quote ex­presses a view which I find hard to un­der­stand, let alone agree with:

  If you would cost your lo­cal place $5 to save the cost of a fif­teen sec­ond phone call, make no mis­take. You are defect­ing. You are play­ing zero-sum games with those who should be your al­lies. You are bad, and you should feel bad.

• Pablo_Stafforini 7 Feb 2019 1:30 UTC

  13 points

  on: Disen­tan­gling ar­gu­ments for the im­por­tance of AI safety

  See Ben Garfinkel’s talk at EA Global Lon­don 2018 (which I’ll link when it’s available on­line).

  Ben’s talk is now on­line.

  (Loved the post, BTW.)

  EDIT: A tran­script is now also available.

• Pablo_Stafforini 9 Jan 2019 4:34 UTC

  12 points

  on: Changes to my workflow

  Any fur­ther changes since the post was pub­lished?

• Pablo_Stafforini 18 Oct 2018 16:52 UTC

  7 points

  on: Book re­view: 23 things they don’t tell you about capitalism

  Yet that last chap­ter also show­cases one of the book’s main failings. Thing 21′s ti­tle is about the benefits of big gov­ern­ment, but its con­tent is only about the welfare state. I’m happy to grant that so­cial safety nets can be benefi­cial for job mo­bil­ity, while still strongly be­liev­ing that in­creased reg­u­la­tion and state-sec­tor em­ploy­ment have the ex­act op­po­site effect.

  I agree with this. Ja­son Bren­nan draws a use­ful dis­tinc­tion be­tween the “so­cial in­surance state”, which seeks to provide var­i­ous goods and ser­vices, and the “ad­minis­tra­tive state”, which seeks to reg­u­late the econ­omy. Since these two func­tions of the state are clearly very differ­ent, it makes lit­tle sense to frame the dis­cus­sion as one where one should be ei­ther for or against “big gov­ern­ment”.

• Pablo_Stafforini 30 May 2017 20:21 UTC

  0 points

  on: The Best Text­books on Every Subject

  Sub­ject: His­tory of Economics

  Recom­men­da­tion: Eco­nomics Evolv­ing, by Ag­nar Sandmo

  Rea­son: A su­perbly clear overview of the his­tory of eco­nomics, from Adam Smith un­til the 1970s. Each chap­ter pro­vides a guide to fur­ther read­ing. I found this book much bet­ter than the al­ter­na­tives in the genre that I con­sulted, in­clud­ing Lionel Rob­bins’ opinionated A His­tory of Eco­nomic Thought and Joseph Schum­peter’s chaotic His­tory of Eco­nomic Anal­y­sis.

  As a com­pan­ion, I recom­mend Keynes’ Es­says in Biog­ra­phy, a col­lec­tion of won­der­fully writ­ten (and as­ton­ish­ingly well-re­searched) es­says on some of the great English economists, in­clud­ing Malthus, Jevons, Edge­worth and Mar­shall.

• Pablo_Stafforini 3 Mar 2017 6:52 UTC

  0 points

  in reply to: gjm's comment on: How does MIRI Know it Has a Medium Prob­a­bil­ity of Suc­cess?

  Thanks for the feed­back. I agree that a com­ment worded in the man­ner you sug­gest would have com­mu­ni­cated my point more effec­tively.

• Pablo_Stafforini 26 Feb 2017 21:09 UTC

  0 points

  in reply to: gjm's comment on: How does MIRI Know it Has a Medium Prob­a­bil­ity of Suc­cess?

  My point is that these early pro­nounce­ments are (limited) ev­i­dence that we should treat Eliezer’s pre­dic­tions with more cau­tion that we would oth­er­wise.

• Pablo_Stafforini 26 Feb 2017 20:46 UTC

  0 points

  in reply to: gjm's comment on: How does MIRI Know it Has a Medium Prob­a­bil­ity of Suc­cess?

  Yes, I am aware that this is what Eliezer has said, and I wasn’t im­ply­ing that those early state­ments re­flect Eliezer’s cur­rent think­ing. There is a clear differ­ence be­tween “Eliezer be­lieved this in the past, so he must be­lieve it at pre­sent” and “Eliezer made some wrong pre­dic­tions in the past, so we must treat his cur­rent pre­dic­tions with cau­tion”. Eliezer is en­ti­tled to ask his read­ers not to as­sume that his past be­liefs re­flect those of his pre­sent self, but he is not en­ti­tled to ask them not to hold him re­spon­si­ble for hav­ing once said stuff that some may think was ill-judged. (If Eliezer had com­mit­ted se­ri­ous crimes at the age of 18, it would be ab­surd for him to now claim that we should re­gard that per­son as a differ­ent in­di­vi­d­ual who also hap­pens to be called ‘Eliezer Yud­kowsky’. Epistemic re­spon­si­bil­ity seems analo­gous to moral re­spon­si­bil­ity in this re­spect.)

• Pablo_Stafforini 25 Feb 2017 22:17 UTC

  0 points

  in reply to: CarlShulman's comment on: How does MIRI Know it Has a Medium Prob­a­bil­ity of Suc­cess?

  Eliezer has not pub­lished a de­tailed ex­pla­na­tion of his es­ti­mates, al­though he has pub­lished many of his ar­gu­ments for his es­ti­mates.

  Eliezer wrote this in 1999:

  My cur­rent es­ti­mate, as of right now, is that hu­man­ity has no more than a 30% chance of mak­ing it, prob­a­bly less. The most re­al­is­tic es­ti­mate for a seed AI tran­scen­dence is 2020; nanowar, be­fore 2015.