Papers

Bayesian Learn­ing in Unsta­ble Set­tings: Exper­i­men­tal Evi­dence

We study learn­ing in a ban­dit task in which the out­come prob­a­bil­i­ties of six arms switch (“jump”) over time. In the exper­i­ment, opti­mal Bayesian learn­ing tracks the jumps by dis­cov­er­ing the prob­a­bil­ity of a jump or direct jump detec­tion. Although more com­plex than the nat­ural alter­na­tive of learn­ing, through adap­tive expec­ta­tions, when com­bined with a par­tially myopic deci­sion rule, Bayesian learn­ing bet­ter matches the behav­ior observed in the lab. This result sug­gests com­plex­ity does not always ham­per ratio­nal learn­ing. We hypoth­e­size that the lat­ter can emerge if sev­eral con­di­tions are met: (i) imple­ment­ing Bayesian learn­ing, rather than bound­edly ratio­nal alter­na­tives, markedly improves eco­nomic per­for­mance (as in our task and, arguably, in many finan­cial con­texts), and sub­jects are pro­vided with sig­nif­i­cant mon­e­tary incen­tives; (ii) the task at hand is eco­log­i­cally rel­e­vant; (iii) agents are “nudged” into pay­ing atten­tion to the right things

Swiss Finance Insti­tute Research Paper No. 10 – 28 — Most recent ver­sion: Feb­ru­ary 2012

View on SSRN

Risk, Unex­pected Uncer­tainty, and Esti­ma­tion Uncer­tainty: Bayesian Learn­ing in Unsta­ble Set­tings

The abil­ity of humans to learn like Bayesians implies that they per­ceive, at a min­i­mum, three lev­els of uncer­tainty: risk, which reflects imper­fect fore­sight even after every­thing is learned; (para­me­ter) esti­ma­tion uncer­tainty, i.e., uncer­tainty about out­come prob­a­bil­i­ties; and unex­pected uncer­tainty, or sud­den changes in the prob­a­bil­i­ties. We describe how these lev­els of uncer­tainty evolve in a six-arm rest­less ban­dit task in which human choices reli­ably reflect Bayesian updat­ing, and how their evo­lu­tion changes the learn­ing rate. We then zoom in on esti­ma­tion uncer­tainty. The abil­ity to sense esti­ma­tion uncer­tainty (also known as ambi­gu­ity) is a virtue because, besides allow­ing one to apply Bayesian learn­ing, it may guide more effec­tive explo­ration; but aver­sion to esti­ma­tion uncer­tainty may be mal­adap­tive. Here, we show that par­tic­i­pant choices reflected aver­sion to esti­ma­tion uncer­tainty. We dis­cuss how past imag­ing stud­ies fore­shad­owed the abil­ity of humans to dis­tin­guish the dif­fer­ent notions of uncer­tainty. Also, we doc­u­ment that the abil­ity of par­tic­i­pants to do such dis­tinc­tion relies on suf­fi­cient rev­e­la­tion of the pay­off gen­er­at­ing model. When we induced struc­tural uncer­tainty, par­tic­i­pants did not gain aware­ness of the jumps in our rest­less ban­dit task, and fell back to model-free rein­force­ment learn­ing.

With P. Bossaerts — PLoS CB, 2011

View online ver­sion Down­load

Esti­ma­tion Uncer­tainty, Unex­pected Uncer­tainty and Action
Uncer­tainty as Mod­u­la­tors of Deci­sion Mak­ing: How Adap­tive?

Using sim­u­lated data, we study the per­for­mances of model-free and model-based approaches to learn­ing in a rest­less ban­dit prob­lem.

With P. Bossaerts

Positions

Assis­tant Pro­fes­sor of Finance — Australian School of Busi­ness, Syd­ney
Vis­it­ing Asso­ciate in Eco­nom­ics — California Insti­tute of Tech­nol­ogy, Pasadena, CA

Student Opportunities

Are you a UNSW stu­dent look­ing for oppor­tu­ni­ties to explore neu­ro­fi­nance? See what’s avail­able and pos­si­ble here

Contacts

Elise Payzan-Le Nestour
Room 338 - Level 3
Australian School of Business, UNSW Sydney
+61 (2) 9385 4273 | +1 626 407 3330
elise@elisepayzan.com