Motivation Model Equilibrium Efficiency BoundedRecall Conclusions

Attention, Coordination, and Bounded Recall

Alessandro Pavan

Northwestern University

Chicago FED, February 2016 Useful modelization for: - production or network externalities - incomplete markets - business cycles - large Cournot-Bertrand games - elections ...

Most of the literature: exogenous information structure

Many phenomena of interest: (info. acquisition) is central

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Motivation

Many socioeconomic environments - large group of agents - actions under dispersed information Most of the literature: exogenous information structure

Many phenomena of interest: attention (info. acquisition) is central

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Motivation

Many socioeconomic environments - large group of agents - actions under dispersed information

Useful modelization for: - production or network externalities - incomplete markets - business cycles - large Cournot-Bertrand games - elections ... Many phenomena of interest: attention (info. acquisition) is central

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Motivation

Many socioeconomic environments - large group of agents - actions under dispersed information

Useful modelization for: - production or network externalities - incomplete markets - business cycles - large Cournot-Bertrand games - elections ...

Most of the literature: exogenous information structure Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Motivation

Many socioeconomic environments - large group of agents - actions under dispersed information

Useful modelization for: - production or network externalities - incomplete markets - business cycles - large Cournot-Bertrand games - elections ...

Most of the literature: exogenous information structure

Many phenomena of interest: attention (info. acquisition) is central Equilibrium and effi cient allocation of attention

- perfect recall

- bounded recall

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions This paper

Flexible (yet rich) framework

- complementarity or substitutability in actions

- rich of payoff interdependencies Motivation Model Equilibrium Efficiency BoundedRecall Conclusions This paper

Flexible (yet rich) framework

- complementarity or substitutability in actions

- rich set of payoff interdependencies

Equilibrium and effi cient allocation of attention

- perfect recall

- bounded recall Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Questions

What payoff interdependencies create ineffi ciency in eq. allocation of attention?

How does ineffi ciency in attention relate to ineffi ciency in use of information?

What is the effect of bounded recall?

What policies can alleviate such ineffi ciencies? (related work) Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Related literature (incomplete)

Effi cient use of information and social value of information Radner (1977), Vives (JET 1984, 2013) Morris and Shin (AER 2002) Angeletos and Pavan (AER, 2004, Ecma 2007, Jeea, 2009) ... Information acquisition/(in)attention in coordination settings Vives and Van Zandt (2007) Hellwig and Veldkamp (Restud, 2009) Amir and Lazzati (2014) Ma´ckowiak and Wiederholt (AER, 2009, 2012) Myatt and Wallace (Restud 2012) → Szkup and Trevino (2013), Yang (2013) Colombo, Femminis and Pavan (Restud 2014) → Tirole (2014), Denti (2016) ... Benabou Tirole (JPE 2004) Wilson (2004), Kocer (2010) ... Analogy-based equilibrium Jehiel (JET 2005) 2 Equilibrium allocation of attention

3 Effi cient allocation of attention

4 Bounded recall

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Plan

1 Model (perfect recall) 3 Effi cient allocation of attention

4 Bounded recall

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Plan

1 Model (perfect recall)

2 Equilibrium allocation of attention 4 Bounded recall

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Plan

1 Model (perfect recall)

2 Equilibrium allocation of attention

3 Effi cient allocation of attention Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Plan

1 Model (perfect recall)

2 Equilibrium allocation of attention

3 Effi cient allocation of attention

4 Bounded recall Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Model Actions and gross payoffs

ui ki , k j j=i, θ { } 6  Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Model Actions and gross payoffs

Continuum of agents with payoffs:

2 u k , K , θ, σ k   where: k R —individual action ∈ K = k0dΨ(k0) —aggregate action σ 2 =R (k K)2dΨ(k ) — dispersion k 0 − 0 θ RR—underlying uncertainty ("fundamentals") ∈

Assumptions: u( ) quadratic in (k,K,θ), linear in σ 2 · k u( ) s.t. equilibrium and first-best unique and bounded · Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Examples

Investment spillovers (Angeletos and Pavan AER 2004)

ui = Rki c(ki) − 1 2 R = (1 a)θ + aK and c(ki) = k − 2 i Beauty contest (Morris and Shin AER 2002)

2 ui = (1 r) (ki θ) r (L(ki) L¯) − − · − − · −

2 2 2 2 L(ki) k0 ki dΨ(k0) = (ki K) +σ and L¯ = L(k)dΨ(k) = 2σ ≡ − − k Z Z  Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Examples

Monetary economies (Woodford 2005, Colombo, Femminis and Pavan, 2014, Llosa and Venkateswaran, 2015)

u(θ,Ci,Ni) V (Ci) Ni ≡ − v v 1 v 1 −v − Ci = chi dh Z[0,1]  α Yi = θ Ni

phchidh piYi T [0,1] ≤ − Z

Cournot and Bertrand games (Vives JET 1984)

1 2 ui = (a θK) ki k − · − 2 i N = 1,234,576 sources of information:

1 yl = θ + εl with εl N(0,η− ) l = 1,...,N ∼ l

i i N Agent i’s "impressions" x = (xl)l=1 with

1 i i i i − x = yl + ξ with ξ N 0, z tl l = 1,...,N l l l ∼ l ·     where

ηl : accuracy tl : transparency/clarity i zl : attention

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Model Information and attention

Common prior: 1 θ N(0,π− ) ∼ θ i i N Agent i’s "impressions" x = (xl)l=1 with

1 i i i i − x = yl + ξ with ξ N 0, z tl l = 1,...,N l l l ∼ l ·     where

ηl : accuracy tl : transparency/clarity i zl : attention

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Model Information and attention

Common prior: 1 θ N(0,π− ) ∼ θ N = 1,234,576 sources of information:

1 yl = θ + εl with εl N(0,η− ) l = 1,...,N ∼ l Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Model Information and attention

Common prior: 1 θ N(0,π− ) ∼ θ N = 1,234,576 sources of information:

1 yl = θ + εl with εl N(0,η− ) l = 1,...,N ∼ l

i i N Agent i’s "impressions" x = (xl)l=1 with

1 i i i i − x = yl + ξ with ξ N 0, z tl l = 1,...,N l l l ∼ l ·     where

ηl : accuracy tl : transparency/clarity i zl : attention Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Model Attention cost and net payoffs

i i i N Attention cost: C(z ) where z = (zl)l=1

C0 zi > 0, all zi = 0 · n 6  i limzn ∞ Cn0 (z ) = ∞ · → convex (results extend to concave, e.g., entropy reduction) ·

i i E.g. C(z ) = c ∑l zl  i i E.g. C(z ) = ∑l g(zl)

...but also C(zi) = µ(zi;y) (entropy reduction)

Net payoff 2 i u ki,K,σ ,θ C(z ) k −   Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Model Timing

agents allocate attention zi

update their beliefs based on xi

commit their actions ki

payoffs realized Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Plan

1 Model (perfect recall)

2 Equilibrium allocation of attention

3 Effi cient allocation of attention

4 Bounded Recall Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Equilibrium use of information (Angeletos and Pavan, Ecma 2007)

Optimality: j j k j = E[ κ + α(K κ) x ; z ] − | where

κ = κ0 + κ1θ (complete-info. equilibrium action)

α ukK equilibrium degree of coordination ukk ≡ | | −→ Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Equilibrium allocation of attention

Theorem There exists a unique symmetric equilibrium. In this eq., the attention zˆ that each agent assigns to the various sources of information is s.t., for any source n = 1,...,N that receives strictly positive attention,

ukk zˆn = κ1γn | | s 2Cn0 (zˆ)tn where (1 α)πn − 1 αρn γn − is "influence" of the source ≡ N (1 α)πs πθ + ∑s=1 1−αρ − s and where

ηszˆsts πs πs = is endogenous precision and ρs = is endogenous "publicity" zˆsts + ηs ηs Given equilibrium allocation of attention zˆ, equilibrium actions are given by

i N i i N k = κ0 + κ1 ∑ γnxn all i [0,1], almost all x R . n=1 ∈ ∈   Agent’seq. continuation payoff (fixing k( ;z ˆ)): · i ukk i i Ui(z ;z ˆ) = E[u(K,K,σ k,θ)] + Var[ki K z ,zˆ,k( ;z ˆ)] C(z ) 2 − | · − Private value of attention

ukk ∂Var[k K z,k( ;z)] | | − | · − 2 · ∂zn private aversion to dispersion reduction in dispersion (fixing eq. strategy· k( ;z)) · Result generalizes Colombo, Femminis, Pavan (Restud 2014)

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Private value of attention

Envelope reasoning: hold k( ;z ˆ) fixed · Private value of attention

ukk ∂Var[k K z,k( ;z)] | | − | · − 2 · ∂zn private aversion to dispersion reduction in dispersion (fixing eq. strategy· k( ;z)) · Result generalizes Colombo, Femminis, Pavan (Restud 2014)

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Private value of attention

Envelope reasoning: hold k( ;z ˆ) fixed ·

Agent’seq. continuation payoff (fixing k( ;z ˆ)): · i ukk i i Ui(z ;z ˆ) = E[u(K,K,σ k,θ)] + Var[ki K z ,zˆ,k( ;z ˆ)] C(z ) 2 − | · − Result generalizes Colombo, Femminis, Pavan (Restud 2014)

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Private value of attention

Envelope reasoning: hold k( ;z ˆ) fixed ·

Agent’seq. continuation payoff (fixing k( ;z ˆ)): · i ukk i i Ui(z ;z ˆ) = E[u(K,K,σ k,θ)] + Var[ki K z ,zˆ,k( ;z ˆ)] C(z ) 2 − | · − Private value of attention

ukk ∂Var[k K z,k( ;z)] | | − | · − 2 · ∂zn private aversion to dispersion reduction in dispersion (fixing eq. strategy· k( ;z)) · Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Private value of attention

Envelope reasoning: hold k( ;z ˆ) fixed ·

Agent’seq. continuation payoff (fixing k( ;z ˆ)): · i ukk i i Ui(z ;z ˆ) = E[u(K,K,σ k,θ)] + Var[ki K z ,zˆ,k( ;z ˆ)] C(z ) 2 − | · − Private value of attention

ukk ∂Var[k K z,k( ;z)] | | − | · − 2 · ∂zn private aversion to dispersion reduction in dispersion (fixing eq. strategy· k( ;z)) · Result generalizes Colombo, Femminis, Pavan (Restud 2014) Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Plan

1 Model (perfect recall)

2 Equilibrium allocation of attention

3 Effi cient allocation of attention

4 Bounded Recall Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Effi ciency

Welfare : ex-ante utility of representative agent

Definition

Effi cient allocation consists of attention z∗ along with action rule k∗( ;z∗) that jointly maximize · 2 E[u(k,K,σ ,θ) z] C(z) k | − Team problem

Planner’s problem: control incentives but cannot transfer information Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Effi cient use of information (Angeletos and Pavan, Ecma 2007)

Given attention z, effi ciency in actions requires that k ( ;z) solves ∗ ·

k∗(x;z) = E[κ∗ + α∗(K κ∗) x ; z ] x, − | ∀ where κ∗ = κ∗ + κ∗θ FB 0 1 −→

uσσ 2ukK uKK aversion to volatility α∗ − − = 1 ≡ ukk + uσσ − aversion to dispersion

socially optimal degree of coordination Recall that eq.

ukk zˆn = κ1γn | | s 2Cn0 (zˆ)tn

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Effi cient allocation of attention

Theorem

Effi ciency in attention requires that, for any n for which zn∗ > 0,

ukk + uσσ zn∗ = κ1∗γn∗ | | s 2Cn0 (z∗)tn where

(1 α )πn − ∗ 1 αρn γn∗ − is effi cient "influence" of the source ≡ N (1 α∗)πs πθ + ∑s=1 1−α ρ − ∗ s

ηszs∗ts πs∗ πs = is endogenous precision and ρs = is endogenous publicity z∗s ts + ηs ηs Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Effi cient allocation of attention

Theorem

Effi ciency in attention requires that, for any n for which zn∗ > 0,

ukk + uσσ zn∗ = κ1∗γn∗ | | s 2Cn0 (z∗)tn where

(1 α )πn − ∗ 1 αρn γn∗ − is effi cient "influence" of the source ≡ N (1 α∗)πs πθ + ∑s=1 1−α ρ − ∗ s

ηszs∗ts πs∗ πs = is endogenous precision and ρs = is endogenous publicity z∗s ts + ηs ηs

Recall that eq.

ukk zˆn = κ1γn | | s 2Cn0 (zˆ)tn Welfare under effi cient use of information (for given attention z)

w∗(z) E[u(κ∗,κ∗,0,θ)] L ∗(z) C(z), ≡ − − where u(κ∗,κ∗,0,θ) is welfare under FB allocation and

ukk + 2ukK + uKK L ∗(πx,πz) | |Var[K κ∗ k∗( ;z),z] ≡ 2 − | · ukk + uσσ + | |Var[k K k∗( ;z),z] 2 − | · Holding k ( ;z), Var[K κ k ( ;z),z] independent of z ∗ · − ∗ | ∗ · Social value of attention

ukk + uσσ ∂Var[k K z,k ( ;z)] | | − | ∗ · − 2 · ∂zn social aversion to dispersion reduction in dispersion (fixing eff. strategy· k ( ;z) ) ∗ ·

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Effi cient allocation of attention

Envelope reasoning Holding k ( ;z), Var[K κ k ( ;z),z] independent of z ∗ · − ∗ | ∗ · Social value of attention

ukk + uσσ ∂Var[k K z,k ( ;z)] | | − | ∗ · − 2 · ∂zn social aversion to dispersion reduction in dispersion (fixing eff. strategy· k ( ;z) ) ∗ ·

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Effi cient allocation of attention

Envelope reasoning Welfare under effi cient use of information (for given attention z)

w∗(z) E[u(κ∗,κ∗,0,θ)] L ∗(z) C(z), ≡ − − where u(κ∗,κ∗,0,θ) is welfare under FB allocation and

ukk + 2ukK + uKK L ∗(πx,πz) | |Var[K κ∗ k∗( ;z),z] ≡ 2 − | · ukk + uσσ + | |Var[k K k∗( ;z),z] 2 − | · Social value of attention

ukk + uσσ ∂Var[k K z,k ( ;z)] | | − | ∗ · − 2 · ∂zn social aversion to dispersion reduction in dispersion (fixing eff. strategy· k ( ;z) ) ∗ ·

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Effi cient allocation of attention

Envelope reasoning Welfare under effi cient use of information (for given attention z)

w∗(z) E[u(κ∗,κ∗,0,θ)] L ∗(z) C(z), ≡ − − where u(κ∗,κ∗,0,θ) is welfare under FB allocation and

ukk + 2ukK + uKK L ∗(πx,πz) | |Var[K κ∗ k∗( ;z),z] ≡ 2 − | · ukk + uσσ + | |Var[k K k∗( ;z),z] 2 − | · Holding k ( ;z), Var[K κ k ( ;z),z] independent of z ∗ · − ∗ | ∗ · Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Effi cient allocation of attention

Envelope reasoning Welfare under effi cient use of information (for given attention z)

w∗(z) E[u(κ∗,κ∗,0,θ)] L ∗(z) C(z), ≡ − − where u(κ∗,κ∗,0,θ) is welfare under FB allocation and

ukk + 2ukK + uKK L ∗(πx,πz) | |Var[K κ∗ k∗( ;z),z] ≡ 2 − | · ukk + uσσ + | |Var[k K k∗( ;z),z] 2 − | · Holding k ( ;z), Var[K κ k ( ;z),z] independent of z ∗ · − ∗ | ∗ · Social value of attention

ukk + uσσ ∂Var[k K z,k ( ;z)] | | − | ∗ · − 2 · ∂zn social aversion to dispersion reduction in dispersion (fixing eff. strategy· k ( ;z) ) ∗ · Social value of attention

ukk + uσσ ∂Var[k K z,k ( ;z)] | | − | ∗ · − 2 · ∂zn social aversion to dispersion reduction in dispersion (fixing eff. strategy· k ( ;z) ) ∗ ·

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Equilibrium vs effi cient allocation of attention

Private value of attention

ukk ∂Var[k K z,k( ;z)] | | − | · − 2 · ∂zn private aversion to dispersion reduction in dispersion (fixing eq. strategy· k( ;z)) · Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Equilibrium vs effi cient allocation of attention

Private value of attention

ukk ∂Var[k K z,k( ;z)] | | − | · − 2 · ∂zn private aversion to dispersion reduction in dispersion (fixing eq. strategy· k( ;z)) ·

Social value of attention

ukk + uσσ ∂Var[k K z,k ( ;z)] | | − | ∗ · − 2 · ∂zn social aversion to dispersion reduction in dispersion (fixing eff. strategy· k ( ;z) ) ∗ · Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Effi cient allocation of attention

Effi ciency in attention requires

- effi ciency in use of information: k( ;z) = k ( ;z) · ∗ ·

- private = social aversion to dispersion uσσ = 0 ⇔ Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Plan

1 Model (perfect recall)

2 Equilibrium allocation of attention

3 Effi cient allocation of attention

4 Bounded Recall Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Bounded Recall

Idea: posteriors correct, but agents cannot recall influence of individual sources Given attention z j, posterior beliefs about θ continues to be Normal with mean

j N i πn ηszsts x¯ = ∑n=1 δ nxn, with δ n N and πs ≡ πθ + ∑s=1 πs ≡ zsts + ηs N and precision πθ + ∑s=1 πs However, agent is unable to decompose x¯j into various impressions x j (x j ,...,x j ). ≡ 1 N Equivalently, unable to decompose his posteriors into

θ˜ xi | n Measurability constraint on k(x j;zz) Distinction relevant only in strategic setting Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Bounded Recall

For simplicity: πθ = 0 Theorem In unique symmetric equilibrium, given allocation z#, actions given by

i i k = κ0 + κ1x¯ For any source that receives strictly positive attention in eq.,

# # # # ukk ∂Var k K;z ,k( ;z ) ukk ∂Var K κ;z ,k( ;z ) C (z#) = | | − · | | (1 α) − · n0 2 z 2 z −  ∂ n  − −  ∂ n  Novel effect:

# # ukk ∂Var K κ;z ,k( ;z ) | | (1 α) − · − 2 − ∂zn private aversion to volatility of own’saverage action reduction in volatility (fixing eq. strategy k#( ;z)) · · Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Bounded vs Perfect Recall

Theorem Let zˆ be eq. allocation of attention with perfect recall. There exist publicity thresholds ρ0,ρ00 [0,1] s.t., starting from zˆ, any agent with bounded recall is better off by ∈

(a) locally increasing attention to sources for which ρ [ρ ,ρ ]; n ∈ 0 00 (b) locally decreasing attention to sources for which ρ / [ρ ,ρ ]. n ∈ 0 00 Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Bounded vs Perfect Recall

Reallocation of attention towards sources of average (endogenous) publicity zsts ρn = zsts + ηs

Sources of low publicity: useful to forecast θ

Sources of high publicity: useful to forecast K

Sources of intermediate transparency: good compromises Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Bounded vs Perfect Recall

Previous result about best responses extends to equilibrium N Suppose C(z) = c ∑s=1 zs Theorem  Let zˆ be eq. attention with perfect recall and z# eq. attention with bounded # recall. There exist thresholds t0,t00 R++ s.t. zn > zˆn only if tn [t0,t00]. ∈ # # ∈ Furthermore for any n for which tn [t ,t ], z < zˆn only if z = 0. ∈ 0 00 n n Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Effi ciency under Bounded Recall

...see paper! Effi ciency in allocation of attention requires (a) absence of externalities from action-dispersion (b) effi ciency in use of information

Bounded recall: reallocation of attention towards sources with intermediate transparency

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Conclusions

Attention in large economies with - complementarity / substitutability in actions - rich set of payoff interdependencies - rich information structure Bounded recall: reallocation of attention towards sources with intermediate transparency

Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Conclusions

Attention in large economies with - complementarity / substitutability in actions - rich set of payoff interdependencies - rich information structure

Effi ciency in allocation of attention requires (a) absence of externalities from action-dispersion (b) effi ciency in use of information Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Conclusions

Attention in large economies with - complementarity / substitutability in actions - rich set of payoff interdependencies - rich information structure

Effi ciency in allocation of attention requires (a) absence of externalities from action-dispersion (b) effi ciency in use of information

Bounded recall: reallocation of attention towards sources with intermediate transparency Motivation Model Equilibrium Efficiency BoundedRecall Conclusions Conclusions

Future work - endogenous sources / social (e.g., capital mkts information aggregation) →

- "optimal" recall strategy

- dynamics (optimal stopping)

- fully flexible info. structures (attention-based correlated eq.) Motivation Model Equilibrium Efficiency BoundedRecall Conclusions

Thank You!