ECO529 Notes

Lectures on Macro, Money, and Finance
A Heterogeneous-Agent Continuous-Time Approach1
Markus K. Brunnermeier2 Sebastian Merkel3

Princeton University Princeton University
Yuliy Sannikov4
Stanford University
August 31, 2021

– Preliminary and Incomplete –
Latest Version: [Click Here]
1 These lecture notes would not have seen the light without the great input of Ziang Li and
Daojing Zhai. We are grateful to Jonathan Payne for numerous discussions, Max Alston, Fer-
nando Mendo and Luca Neri for great help with the project.
2 Email: markus@princeton.edu.
3 Email: smerkel@princeton.edu.
4 Email: sannikov@gmail.com.
CONTENTS
Contents
List of Symbols 5
1 Introduction 6
1.1 The Three Watershed Moments in Macroeconomics . . . . . . . . . . . . 6
1.2 Fire sales, Pecuniary Externalities, Strategic Complementarities and Am-
plification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Continuous-time Modeling with Itô Processes . . . . . . . . . . . . . . . 6
Part I Real Models with Financial Frictions
2 A Simple Heterogeneous Agents Model 8

2.1 Model Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Solution Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 A Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4 Discussion: Time-varying Risk-free Rate . . . . . . . . . . . . . . . . . . . 25
2.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3 Endogenous Risk Dynamics 29

3.1 Model Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 Solution Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.3 Stationary Distribution and Fan Charts . . . . . . . . . . . . . . . . . . . 53
1 / 176
CONTENTS
3.4 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4 A Model with Jumps 67

4.1 Jump Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2 Model Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.3 Solution Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Part II Monetary Models with Aggregate and Idiosyncratic Risk
5 A One-Sector Monetary Model with Idiosyncratic Risk 82

5.1 Models of Money as a Store of Value . . . . . . . . . . . . . . . . . . . . . 83
5.2 Model Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.3 Solution Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.4 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.5 Connections to Monetary Theory . . . . . . . . . . . . . . . . . . . . . . . 95
5.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6 Cash versus Cashless Economy and the I Theory of Money 106

6.1 Cash versus Cashless Economy . . . . . . . . . . . . . . . . . . . . . . . . 107
6.2 The I Theory of Money . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.3 Policy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
7 Welfare and Optimal Monetary Policy 136

7.1 Welfare Function with Log Utility . . . . . . . . . . . . . . . . . . . . . . . 136
Part III International Models
8 A Symmetric International Model with Runs / Sudden Stops 139
2 / 176
CONTENTS
8.1 Model Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

8.2 Contrasting Different Degrees of Market Completeness . . . . . . . . . . 144
8.3 Welfare Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
8.4 Sudden Stops and Runs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
3 / 176
CONTENTS
4 / 176
CONTENTS
List of Symbols
i Sector type i (experts, households...)

ĩ Agents of sector i with idiosyncratic risk, ĩ ∈ [0, 1]
ct Consumption
yt Output goods
ρ Discount rate
a Productivity
A (·, ·) Aggregate productivity
κ Capital share
χ Equity share
η Wealth share
ψ Share of capital wealth
ϑ Share of wealth that agents hold in money (fraction of nominal wealth)
ι Investment rate
ω Investment opportunities
ϕ Adjustment costs factor in investment function
Φ(·) Investment function
P Money price of goods
qK Price of capital
qM Rescaled price of money, q M = PM·K
M positive net supply of money
ζ Consumption-wealth ratio
ξ Stochastic discount factor
V Value function
µt Drift
σt Volatility
ςt Price of risk
dZt Brownian motions
5 / 176
CHAPTER 1. INTRODUCTION
Chapter 1
Introduction
1.1 The Three Watershed Moments in Macroeconomics
1.2 Fire sales, Pecuniary Externalities, Strategic Comple-

mentarities and Amplification
1.3 Continuous-time Modeling with Itô Processes
6 / 176
Part I
Real Models with Financial Frictions
7 / 176
CHAPTER 2. A SIMPLE HETEROGENEOUS AGENTS MODEL
Chapter 2
A Simple Heterogeneous Agents

Macro-Finance Model in Continuous
Time
To start with, we study the heterogeneous agents model of Basak and Cuoco (1998),
a simple yet classic model. The main objective of this lecture is to illustrate some main
building blocks of a large body of macro-finance models that employs continuous time
methods.
2.1 Model Setup
Environment. Time is, of course, continuous. There is no labor in this model and
hence capital is the only factor used in production. The economy consists of two types
of agents – experts and households. We denotes the two types by i = {e, h}.1 There is a
continuum (with mass one) of both types, ĩ ∈ [0, 1].2 Experts have access to production
technologies while households only consume.
1 In general, i can denote different types/sectors, or different subgroups within the same sector.
2 Individual-specific analysis will only be needed in environments with idiosyncratic risk.
8 / 176
Expert. Experts have a CRS production function yet = aket . Denote experts’ consump-
tion and investment rate by cet and ιet .3 An expert’s capital stock evolves according to
dket
= (Φ(ιet ) − δ)dt + σdZt ,
ket
where investment function Φ satisfies Φ(0) = 0, Φ′ (0) = 1, Φ′ (·) > 0, and Φ′′ (·) < 0.
Φ(·) reflects adjustment costs in capital stock and its concavity captures technological
illiquidity, i.e., the adjustment cost due to converting output to new capital and vice
versa (Bernanke, Gertler and Gilchrist, 1999). δ is the depreciation rate. σdZt is the risk
faced by investors.4 Note that there is no idiosyncratic risk in this model. Experts have
a CRRA utility function and they maximize
"Z #
∞ ( c e )1− γ
E0 e−ρt t dt .
0 1−γ
Household. Denote household’s consumption rate by cth . They maximize

"Z #
∞ ( c h )1− γ
E0 e−ρt t dt .
0 1−γ
Financial Friction. Risk sharing is not allowed. Experts can only issue risk-free debts.
The balance sheets of the two sectors are as following:
A Expert L A Household L
Net worth
Capital Debt Loans
𝑞" 𝐾" − 𝑁"
𝑞" 𝐾"
𝑁"
Figure 2.1: Balance sheets of experts and households

3 Again, we will suppress the superscript ĩ throughout this chapter.
4 Unlike most macroeconomic models, we impose risk on the capital accumulation process, instead
of shocks to productivity.
9 / 176 1
The net worth of an expert equals the value of her capital stock minus her debt. The
net worth of a household equals the loans he provides.
2.2 Solution Method
Before starting with the solution method, it is useful to introduce three Itô’s formula,
which will be used extensively throughout this class.
The following results hold for geometric Itô processes Xt , Yt where
dXt dYt
= µtX dt + σtX dZt , = µYt dt + σtY dZt .
Xt Yt

′ 1 ′′
Itô’s lemma: d f ( Xt ) = f ( Xt )(µtX Xt ) + f ( Xt )(σt Xt ) dt + f ′ ( Xt )(σtX Xt )dZt .
X 2
2
d( Xt Yt )
Itô’s product rule: = (µtX + µYt + σtX σtY )dt + (σtX + σtY )dZt .
Xt Yt
d( Xt /Yt ) h i
Itô’s quotient rule: X Y Y Y X
= µt − µt + σt (σt − σt ) dt + (σtX − σtY )dZt .
Xt /Yt
2.2.0 Postulate aggregates, price processes and obtain return processes
In general, the aggregate capital stock is obtained by

Z i
Kt ≡ ke,t ĩ dĩ.
0
With only aggregate risk, all experts (households) are identical (i.e., ke,t ĩ = ket , ∀ĩ), so
total capital stock, expert net worth and household net worth can be simply obtained
by Kt = ket , Nte = net , Nth = nth .
Denote the price of capital by qt . The total wealth of the economy is qt Kt . The wealth
10 / 176
share of experts is ηt = Nte /( Nte + Nth ) = Nte /qt Kt . We then postulate that qt follows
dqt q q
= µt dt + σt dZt .
qt
Importantly, volatility of price loads on the same Brownian motion as capital stock.
Given the price process and the consumption-investment decision of experts, we can
calculate the return rate to capital, rtK (ιet ).5 Using Itô’s product rule,
Dividend Yield Capital Gain

z }| { z }| {
e
a − ιt d(qt ket )
drtK (ιet ) = dt +
qt qt ket
" #
a − ιet q q q
= + Φ(ιet ) − δ + µt + σσt dt + (σ + σt )dZt . (2.1)
qt
We then postulate that the stochastic discount factor (SDF, e.g., ξ ti = e−ρt u′ (cit )) follows
dξ ti
= −rt dt − ςit dZt , i ∈ {e, h}, (2.2)
ξ ti
where rt is the risk-free rate.6
2.2.1 For given SDF processes, derive individual equilibrium condi-

tions
Having studied the returns to the two assets avaliable to the experts (capital and
debt), we can formally state expert’s problem as
"Z #
∞ ( c e )1− γ
max E0 e−ρt t dt
{ι t ,θ e t ,cet }∞
t =0 0 1−γ
dnet cet
s.t. =− dt + (1 − θte )drtK (ιet ) + θte rt dt (2.3)
net net
5 For
superscripts, we use lowercase letters for different types and capital letters for different assets.
6 Recall that in discrete time, the consumption Euler equation implies E SDF t + 1
t (1 + rt ) = 1. The
" SDFt
Z t #
SDFt
continuous-time analogue is Es exp rτ dτ = 1.
SDFs s
11 / 176
n0 given,
where θte < 0 is expert’s (short) position on risk-free asset (i.e., loans from the house-
holds).
First, the choice of investment rate is a static and time-separable problem. An expert
chooses ιet to maximize her return
" #
a − ιet q q q
drtK (ιet ) = + Φ(ιet ) − δ + µt + σσt dt + (σ + σt )dZt .
qt
The first-order condition yields the classic Tobin’s q equation:
1
= Φ′ (ιet ).
qt
Since all experts are identical, aggregate capital stock Kt also follows
dKt
= (Φ(ιet ) − δ)dt + σdZt .
Kt
For simplicity, we will use a special functional form Φ(ι) = 1

ϕ log(ϕι + 1), which implies
ϕιet = qt − 1.
Next, we solve for the asset market equilibrium. For pedagogical purpose, we solve
the problem using three different methods – the HJB approach, the stochastic maximum
principle, and the martingale approach. Starting from next chapter, we primarily use
the martingale approach.
The HJB approach. HJB approach is the most common method to optimal stochas-
tic control problems for macroeconomists. Once writing all agents’ decision problems
recursively, one needs to specify explicitly the state space on which value and policy
functions are defined. Then start with deterministic functions of state variables, one
gets PDEs to solve numerically for equilibrium.
Let’s start with expert’s problem (2.3), where that experts’ net worth evolves accord-
12 / 176
ing to
dnet cet
= − dt + (1 − θte )drtK (ι t ) + θte rt dt
net net
 ! 
c e a − ιt e
q q q
= − te + (1 − θte ) + Φ(ι t ) − δ + µt + σσt + θte rt  dt + (1 − θte )(σ + σt ) dZt ,
nt qt | {z }
: = σ ne
| {z } t
e
:=µnt
that is, expert makes decision under a set of variable controls cet , θte , ιet , and a set of ex-
q q
ternal variables rt , qt , µt , σt . Suppose we know that net worth share ηte is sufficient to
characterize external variables, and note that net worth net depends on variable controls
and external variables.7 Then two state variables (net , ηte ) and a value function V e (net ; ηte )
can fully characterize the system.8 The HJB equation writes

 ( c e )1− γ
ρV e (net ; ηte ) = max t
+
e e e ct ,θt ,ι t  1−γ

e 2
1 1 2
ηe ηe ηe
e
e

net µnt ∂n Vte + ηte µt ∂η Vte + net σtn ∂nn Vte + ηte σt ∂ηη Vte + net σtn ηte σt ∂nη V e .
| 2 {z 2 }
=Et [dV e (net ;ηte )]/dt
ηe ηe
In equilibrium, rt , qt , µt , σt are functions of the aggregate state ηte , and we will show
that the dynamics of ηte is given by (2.17). The first-order conditions w.r.t. cet , θte , and ιet
are
(cet )−γ = ∂n Vte ,

 
a − ιet q q ∂nn Vte e
∂nη Vte e q
+ Φ(ιet ) − δ + µt + σσt − rt = − net σtn + ηte σt (σ + σt ),
η
qt ∂n Vte ∂n Vte
| {z } | {z }
Myopic Demand Hedging Demand
1
Φ′ (ιet ) = (2.4)
qt
Note the last condition is exactly the Tobin’s q condition.
7 We will verify if in section 2.2.4.

8 The value function under CRRA utility is investigated in section 2.2.3.
13 / 176
For simplicity we focus on log-utility form. Let’s first make a guess of value function
that separates individual state net from aggregate state ηt , i.e. V e (net , ηte ) = α log(net ) +
f (ηte ), and verify it later.9 Hence, ∂n Vte = α/net , ∂nn Vte = −α/(net )2 , ∂enη Vte = 0, i.e.,
hedging demand is always zero under log utility. The first-order conditions are then
1 e
cet = n (2.5)
α t
a − ιet q q e q
+ Φ(ιet ) − δ + µt + σσt − rt = σtn (σ + σt ). (2.6)
qt
Substituting all of conditions into HJB and rearranging it, we have the following differ-
ential equation of f (η e )10
 !2 
r K ι e ( η e ); η e − r ( η e )

1
(ρα − 1) log ne = −ρ f (η e ) − log α − 1 + α r (η e ) + 
2 σ + σq (η e )
e 1 ′′ e η e e e 2
+ f ′ (η e )µη (η e )η e + f (η ) σ (η )η . (2.7)
2
Note right-hand side does not depend on net so it must be that ρα = 1. In conclusion,
the solution to expert problems is given by
• Optimal investment choice (2.4).

• Optimal consumption choice (2.5).
• Optimal portfolio choice (2.6).
• Characterization of value function (2.7)
Similarly we can solve the much simpler household’s problem:
• Optimal consumption choice ch = ρnh .

• Characterization of value function.11
1 e 1 e 2
ρg(η e ) = log ρ − 1 + r (η e ) + g′ (η e )µη (η e )η e + g′′ (η e ) ση (η e )η e (2.8)
ρ 2
9 Similarly, we can guess the value function for CRRA utility is in the form V e ( ne , η e ) = ( n e )1− γ f ( η e ).
f (η e )
will enter in ce /ne
(unless EIS=1) and in hedging demand (unless RA=1). To verify our guess, one
only needs to check the (2.7,2.8) have well-behave solutions for f , g.
e
10 Where r K = a−ι t + Φ ( ιe ) − δ + µq + σσ q .
t qt t t t
11 Like expert’s problem, we guess V h ( n h , η e ) = β log( n h ) + g ( η e ), and find β = 1/ρ
t t t t
14 / 176
Finally we combine two sectors’ problems with market clearing conditions and solve
the model. We will see the price of capital qt is actually a constant in section (2.2.4).
Pontryagin’s stochastic maximum principle. Pontryagin’s maximum principle is a

method to solve optimal control problems, which is complementary to dynamic pro-
gramming.
Stochastic maximum principle. Consider a control problem
dXt = µ( Xt , At )dt + σ( Xt , At )dZt ,
where At are the controls and Xt are the states. The stochastic maximum principle
is formulated for finite-horizon problems with the objective function of the form
"Z #
T
E0 g (t, Xt , At ) dt + G ( XT ) ,
0
where the payoff flow g (t, Xt , At ) depends on t and hence can accommodate dis-
counting.
To solve the optimization problem, one can work with the special adjoint pro-
cess pt , which is the dynamic Lagrange multiplier on the state variable Xt . We label
pt and its volatility qt as costates of the system, and then optimize the Hamiltonian
h i
H = g(t, X, A) + ⟨ p, µ( X, A)⟩ + tr q T σ( X, A) . (2.9)
Under necessary convexity conditionsa , the stochastic maximum principle says that pt
must satisfy the BSDE

dpt = − HX t, Xt , At , pt , qt dt + qt dZt (2.10)
with terminal condition p T = G ′ ( XT ).

a See the convexity conditions in Yuliy’s “Overview of Stochastic Calculus."
We label the two costates (ξ te , −ςet ξ te ) for the expert’s prblem12 . Conventionally, the
second costate is the (arithmetic) volatility of the first, but to be consistent with our
12 it will turn out that they coincide with the SDF and the price of risk.
15 / 176
specification of the SDF (2.2), we let the second costate be minus the arithmetic volatility
of ξ te . Now the Hamiltonian (2.9) writes
e 1− γ
−ρt ( ct )t
e
e

e
H =e + ξ te net µnt − ςet ξ te net σtn
1−γ
 
e 1− γ
!
−ρt ( ct )t a − ιet q q
=e + ξ te −cet + net (1 − θte ) + Φ(ιet ) − δ + µt + σσt + net θte rt 
1−γ qt
q
− ξ te ςet net (1 − θte )(σ + σt ).
The first-order condition w.r.t. cet , θte , ιet are
e−ρt (cet )−γ = ξ te (2.11)

a − ιet q q q
+ Φ(ιet ) − δ + µt + σσt − rt = ςet (σ + σt ). (2.12)
qt
1
Φ(ιet ) = (2.13)
qt
In addition, now the dynamics of costate ξ te follows BSDE (2.10),
∂H e
dξ te = − dt − ςet ξ te dZt ,
∂ne
then drift of ξ te is given by

 ! 
ξe ∂H e a − ιet q q q
µt ξ te = − = −ξ te (1 − θte ) + Φ(ιet ) − δ + µt + σσt + θte rt − ςet (1 − θte )(σ + σt )
∂ne qt
= −rt ξ te .
Hence,
dξ te
= −rt dt − ςet dZt ,
ξ te
confirming that the costates ξ te and ςet are indeed the SDF and the price of risk.
16 / 176
Also, it can be shown that for log utility, the costate ξ te and its volatility ςet ξ te follows13
∂Vte 1 e
ξ te = e = e, ςet = σtn ,
∂nt ρnt
so the FOCs (2.5)-(2.6) and (2.11)-(2.12) are equivalent.
The martingale approach. Now we introduce the martingale approach, a powerful

tool for many macro-finance models.
Martingale approach in discrete time.

Consider a standard dynamic portfolio choice problem in discrete time:
∞
" #
1
max
{θτ ,cτ }∞
Et ∑ (1 + ρ ) τ − t u ( c τ )
τ =t τ =t
s.t. θt pt = θt−1 ( pt + dt ) − ct ∀t,
where {θt , pt , dt } are the vectors of holdings, prices and dividends of different as-
sets. WLOG, we focus an environment with one asset. The FOC w.r.t. θt is
ξ t p t = Et ξ t +1 ( p t +1 + d t +1 ) ,

u′ (c )
where ξ t = (1+1ρ)t u′ (c t ) is the (multi-period) SDF. Consider a self-financing trad-
0
ing strategy A where one reinvests dividend dt in every period. The price of the
strategy ptA satisfies
h i
ξ t ptA = Et ξ t+1 ptA+1 ,
i.e., the process ξ t ptA is a martingale.
Martingale approach in continuous time.
13 We can guess and verify it like the treatment of value function. More intuitively, we will see later
that costate pt is necessarily the derivatives of the value function with respect to the state variable n.
17 / 176
Consider a similar portfolio choice problem in continuous time:
∞
Z
−ρt
max E0 e u(ct )dt
{ιet ,θet ,ct }∞
t =0 0
dnt ct
= − dt + ∑ θt drt + labor income/endowment/taxes
j j
s.t.
nt nt j
n0 given.
j
Here nt is the net worth of the agent. rt denotes the return of asset j. Let xtA be
the value of a self-financing trading strategy A where one reinvests all dividends.
Again, define the SDF as ξ ti = e−ρt u′ (cit ). Then it must be that ξ t xtA follows a mar-
tingale. (For proof, see Brunnermeier and Sannikov (2016, pg. 19) or separate notes
prepared by Sebastian.) Let
dxtA
A
= µtA dt + σtA dZt .
xt
As in (2.2), assume that the SDF follows
dξ t
= −rt dt − ς t dZt .
ξt
Using Itô’s product rule,
d(ξ t xtA ) h
A A
i
= −rt + µt − ς t σt dt + volatility terms.
ξ t xtA
Since ξ t xtA follows a martingale, its drift equals zero, i.e.,
µtA = rt + ς t σtA .
Example 1. For risk-free asset, σtA = 0. Hence, rtF = rt .
Example 2. For any two assets A, B, we have µtA − µtB = ς t (σtA − σtB ).
Now we apply example 2 above to our context. Letting asset A be capital and asset
18 / 176
B be the risk-free debt, the following equation holds for experts14
K K
µrt − rt = ς t σtr .
Plugging in the return equation (2.1), we have
a − ιet q q q
+ Φ(ιet ) − δ + µt + σσt − rt = ςet (σ + σt ). (2.14)
qt
e
As we will see later in section 2.2.3, with log utility cet /net = ρ, ςet = σtn , so (2.14) is also
equivalent to (2.6) and (2.12). Thus we have shown that all three methods deliver the
same result.
Comparison of Different Methods. When working with the HJB approach, we for-
mulate agents’ decision problems recursively, explicitly specifying the state space, and
then get PDEs for value functions. Typically, we need good guesses of functional forms.
In the end, we solve for the equilibrium (PDEs) using iterative method. The HJB char-
acterizes value function on full individual state space (not just along the equilibrium
path).
In the implementation of Stochastic Maximum Principle, we keep the stochastic pro-

cess formulation (as opposed to the recursive formulation) in the derivations, charac-
terize the system in terms of forward and backward stochastic differential equation (in-
stead of differential equations), and only transform them to PDEs if we need numerical
solutions.15 A potential advantage of this approach is that we do not solve for equi-
librium conditions in the whole state pace, and instead only focus on the equilibrium
path.
Stochastic Maximum Principle is connected with HJB method: 1) the costate pt is nec-
essarily the derivatives of the value function with respect to the state variable n, i.e.
pt = ∂n Vt . Note that this happens only along the optimal (equilibrium) path. Essen-
tially, the Stochastic Maximum Principle allows us to ignore what happens in the rest
of state spaces. 2) pt must satisfy the BSDE that comes from the Hamiltonian, and the
14 We use µ X and σ X to denote the drift and volatility of variable X.

15 A recursive structure is still required to solve numerically using PDEs.
19 / 176
controls At must maximize the sum of payoff flow and expected drift of the value func-
tion, as in the HJB approach. As a complementary to the HJB approach, the Stochastic
Maximum Principle is useful particularly in non-stationary settings, in which the state
space may be high dimensional or difficult to specify — even in those settings it can
provide conditions that are necessary and sometimes sufficient for optimization.
Martingale approach works best in endowment economy, where the consumption

steam and stochastic discount factor are known.
For a production economy, we have to combine the martingale approach with the
value function method, which provides information about the SDF process. In the end,
we still characterize the equilibrium systems by PDEs and solve them using a iterative
method.
2.2.2 Evolution of state variable ηt
In this model, agents start with some initial endowments of capital. Over time, they
allocate their wealth between the assets available to them by solving their respective
utility maximization problems, subject to budget constraints and taking prices as given.
Given prices, markets for capital and and consumption goods have to clear. We define
an equilibrium as a map from shocks to prices satisfying all constraints.
Definition 2.1. Given any initial allocation of capital among the agents, an equilibrium is a
map from histories { Zs , s ∈ [0, t]} to price qt , investment rate ιit , price of risk ςit and experts’
portfolio choice θte , such that
(1) all agents choose portfolios and consumption rates to maximize utility,
(2) all markets, for capital and consumption goods, clear.
From now on, we use the wealth share of experts ηt as the state variable.16 Recall
16 In chapter 3, we combine martingale approach with value function, and characterize the equilibrium
with PDEs. Then we will see the (de-scaled) function only depends on wealth share ηt .
20 / 176
that
Nte
ηt = ∈ [0, 1].
qt Kt
The total wealth of experts Nt follows
dNte dnet (2.3) cet e

h
K
i
= == = − dt + r t dt + ( 1 − θ t ) dr t − r t dt
Nte net net
" # 
(2.1) c e  a − ιte 
q q q
=== − te dt + rt dt + (1 − θte ) + Φ(ιet ) − δ + µt + σσt − rt dt + (σ + σt )dZt
nt  qt 
(2.14) cet e
n
e q q
o
==== − e dt + rt dt + (1 − θt ) ς t (σ + σt )dt + (σ + σt )dZt .
nt
Also,
d( q t K t ) h
q q
i
q
= µt + Φ(ιet ) − δ + σσt dt + (σ + σt )dZt
qt Kt
" #
a − ιet q q
= rt − + ςet (σ + σt ) dt + (σ + σt )dZt .
qt
Apply Itô’s quotient rule to ηt = Nte /qt Kt :

" #
dηt cet a − ιet q

q

q
= − e+ − θte (σ + σt ) ςet − (σ + σt ) dt − θte (σ + σt )dZt . (2.15)
ηt nt qt
2.2.3 Value functions
A property of CRRA utility is that if the agent’s wealth changes by a factor of x,

then his optimal consumption at all future states changes by the same factor. Hence,
the agent’s value function takes a power form
u(ωti nit )
V (nit ) = ,
ρ
21 / 176
where ωt reflects current investment opportunity or “net worth multiplier”. The opti-
mal consumption condition (see Sebastian’s notes) yields
cit
V ′ (nit ) = u′ (cit ) ⇐⇒ i
= ρ1/γ (ωti )1−1/γ .
nt
In this chapter, we study the special case of log utility u(c) = log c, which is a CRRA
utility function with γ = 1. Hence, cit /nit = ρ for any ωti process. Let
dcit ci ci dnit ni ni
= µ t dt + σt dZt , = µ t dt + σt dZt .
cit nit
i i
Because cit /nit is constant, σtc = σtn .17 By Itô’s lemma,
d(cit )−1
i 2
ci c ci
= − µ t + σt dt − σt dZt .
(cit )−1
Because ξ ti = e−ρt u′ (cit ) = e−ρt (cit )−1 , Itô’s product rule implies that
dξ ti d(cit )−1
i 2
ci i
i
= −ρdt + i −1 = −ρ − µt + σtc dt − σtc dZt .
ξt (ct )
Recall the postulated SDF process (2.2), we can see the price of risk is
i i q
ςit = σtc = σtn = (1 − θte )(σ + σt ). (2.16)
under the log-utility. Plugging in (2.15), the evolution of state variable ηt , we have
" #
dηt a − ιet q q
= − ρ + (θte )2 (σ + σt )2 dt − θte (σ + σt )dZt . (2.17)
ηt qt
17 To see this, apply Itô’s quotient rule to (cit /nit ).
22 / 176
2.2.4 (Value function iteration) and goods market clearing
Finally, we can close the model using market clearing conditions. Consumption
good market clearing yields
Ct = ( a − ιet )Kt =⇒ ρqt Kt = ( a − ιet (qt ))Kt =⇒ ρqt = a − ιet (qt ).
q q
qt is constant! So µt = σt = 0, ∀t. Our special functional form implies that
1 + ϕa a−ρ
q= , ιe = . (2.18)
1 + ϕρ 1 + ϕρ
Capital market clearing yields
qt Kt 1
1 − θte = e = . (2.19)
Nt ηt
Using portfolio choice and market clearing, we obtain the risk-free interest rate
(2.14) a − ιe q q q
rt ==== + Φ(ιe ) − δ + µt +σ σt −ςet (σ + σt )
q |{z} |{z} |{z}
=0 =0 =0
(2.16) a − ιe
==== + Φ(ιe ) − δ − (1 − θte )σ2
q
(2.19) a − ιe σ2
==== + Φ(ιe ) − δ −
q η
! t
(2.18) 1 1 + ϕa σ2
==== ρ + log −δ− .
ϕ 1 + ϕρ ηt
Plugging (2.19) in the process of state ηt (2.17), we end up with
dηt (1 − ηt )2 2 1 − ηt
= 2
σ dt + σdZt ,
ηt ηt ηt
a simple one-dimensional stochastic differential equation (SDE).
23 / 176
2.3 A Numerical Example
We parametrize the model with a = 0.11, ρ = 5%, σ = 0.1 and ϕ = 10.
Figure 2.2: Numerical results of the simple model, a = .11, ρ = 5%, σ = .1 and ϕ = 10.
Several key observations about equilibrium characteristics are worth pointing out.
First, because qt is constant, as illustrated in the top left panel, there is no endoge-
nous risk, no amplification and no volatility effects.
Second, as experts are levered, their wealth share ηt fluctuates with macro shocks.
A negative shock erodes ηt , and experts require a higher risk premium to hold risky
assets. Experts must be convinced to keep holding risky assets by the increasing Sharpe
24 / 176
ratio
σ ρ + Φ(ιe ) − δ − rt
= ,
ηt σ
which goes to ∞ as ηt goes to 0. Strangely, this is achieved due to the risk-free rt =
ρ + Φ(ιe ) − δ − σ2 /ηt going to −∞, rather than due to a depressed price of the risky
asset, as illustrated in the top right panel of Figure 2.2.
η
Moreover, as the drift of wealth share µt is alway positive, in the long run ηt con-
verges to one and experts will overwhelm the whole economy. This feature is typical
of models in which one group of agents has an advantage over another group – in this
case only experts can invest in the risky asset. It is possible to prevent the expert sec-
tor from becoming too large through some additional assumptions. For example, in
Kiyotaki and Moore (1997), experts are less patient than households and have higher
discount rates, so a greater consumption rate prevents the expert sector from becoming
too large. And in Bernanke, Gertler and Gilchrist (1999), some experts are randomly hit
by a shock that turns them into households.
2.4 Discussion: Time-varying Risk-free Rate
The variation of the risk-premia is driven in this model to 100 percent by the time
variation of the risk free rate. This requires a highly volatile (real) risk-free rate. For
very low net worth shares of the experts the real interest rate has to be significantly
negative. Caballero and Simsek (2019) enforce a zero lower bound on the interest rate
in a setting with sticky prices – in order to study a New Keyensian demand recession
crisis with underutilization of physical capital.
25 / 176
2.5 Exercises
2.5.1 Capital (Quality) and Technology Shocks
Consider the simple model from chapter 2. There, we assumed that expert capital
follows
dket e
Φ

= ( ι t ) − δ dt + σdZt (2.20)
ket
and produces an output flow18
yet = āket . (2.21)
In this problem set you are asked to consider a different specification with consumption-
specific technology shocks instead of capital shocks. Specifically, suppose instead of
equations (2.20) and (2.21) that capital evolves according to
!
dket

ā e
= Φ ι − δ dt (2.22)
ket at t
and produces an output flow

yt = at k t (2.23)
where at is now a stochastic process given by19
dat
= ψ log ā − log at dt + σdZt . (2.24)
at
For ψ = 0 this specification implies a geometric Brownian motion for productivity, for
ψ > 0, at mean-reverts to the level ā in the long run. The additional term ā/at in the Φ
function implies that only consumption production is impacted by changes of at .
(a) Show that without productivity mean reversion (ψ = 0), the model with capi-
18 I use here ā instead of a from the lecture to distinguish this more clearly from the process a defined
below.
19 One gets to this equation by imposing that log a follows an Ornstein-Uhlenbeck process, the
t
continuous-time equivalent of a discrete-time AR(1) process, and correcting by a deterministic time drift,
such that the long-run mean of At is not growing/shrinking over time. The equivalent in discrete time
is often taken as a productivity process in standard macro models.
26 / 176
tal shocks (evolution (2.20) and output (2.21)) and the model with consumption-
specific technology shocks (capital evolution (2.22), output (2.23) and productiv-
ity process (2.24)) are isomorphic in the sense that they imply the same dynamics
for output, consumption, net worth, the expert wealth share η e and the risk-free
rate.
(b) How are the two models related, if ψ > 0?
(c) Explain economically, why the two shock types are not equivalent for neutral
technology shocks (i.e. if the investment technology is Φ ιet ) and an inconstant

ψ function (1-2 sentences are sufficient).
2.5.2 The Basak-Cuoco Model with Heterogeneous Discount Rates
Consider the model from chapter 2 (now again with capital shocks), but unlike there
assume that households are more patient than experts, i.e. they have a discount rate
ρh < ρe . This is the simplest way to generate both a nondegenerate stationary distribu-
tion and some endogenous capital price dynamics.
e e
(a) Derive closed-form expressions for ιe , q, σq , µη and ση as a function of η and
model parameters.20 You do not actually have to follow the order of steps in the
lecture. In this simple model it pays off to start with goods market clearing.
(b) Replicate the Figure 2.2 (δ is not stated there, so choose just some parameter that
generates similar numbers for the risk-free rate), then add to each plot the corre-
sponding line for the model with ρh < ρe using ρh = 2% (and all other parameters
as before).
(c) Assume ϕ > 0. Show that in this model asset price movements mitigate exoge-
nous risk. Explain economically, why this happens.
(d) Argue that the model must have a nondegenerate stationary distribution (just
give some intuition, not a fully spelled-out formal proof). Compute the station-
20 As in the lecture, assume the specific functional form Φ(ι) = 1
ϕ log(1 + ϕι) for Φ(·).
27 / 176
BIBLIOGRAPHY
ary density of η e by numerically solving the ODE stated on page 16 of Yuliy’s

stochastic calculus notes using the same parameters as in part 2. What is the sta-
tionary density of q?
Bibliography
Basak, Suleyman and Domenico Cuoco, “An equilibrium model with restricted stock
market participation,” The Review of Financial Studies, 1998, 11 (2), 309–341.
Bernanke, Ben, Mark Gertler, and Simon Gilchrist, “The Financial Accelerator in a
Quantitative Business Cycle Framework,” in “Handbook of Macroeconomics,” Vol. 1,
Elsevier, 1999, pp. 1341–1393.
Brunnermeier, Markus K. and Yuliy Sannikov, “Macro, money, and finance: A

continuous-time approach,” in “Handbook of Macroeconomics,” Vol. 2, Elsevier,
2016, pp. 1497–1545.
Caballero, Ricardo J and Alp Simsek, “A Risk-centric Model of Demand Recessions

and Speculation,” Working Paper, 2019.
Kiyotaki, Nobuhiro and John Moore, “Credit Cycles,” Journal of Political Economy,
1997, 105 (2), 211–248.
28 / 176
CHAPTER 3. ENDOGENOUS RISK DYNAMICS
Chapter 3
A Macro-model with Endogenous Risk

Dynamics: Amplification, Fire sales,
and Speculation
In last chapter, we studied a very simple model with closed form solutions to illus-
trate the basic structure of countinuous-time macro-finance models.
In this chapter, we present a more complex model where apart from risk-free debt,
experts can also issue outside equity. In terms of economic insights, we enrich the
model to obtain the following properties:
• The risk as well as the price of risk is endogenous and hence time-varying de-
pending on the wealth distribution across the heterogeneous agents in the econ-
omy.
• Equilibrium dynamics contains two regimes – a normal regime around the steady
state and a crisis regime. The economy should be relatively stable near the steady
state, where experts are adequately capitalized and able to absorb most shocks.
However, an unexpected large shock or a series of negative shocks can signifi-
cantly damage the experts and bring the economy to the crisis regime. In a crisis,
experts are undercapitalized and financially constrained. As a result, market liq-
uidity can suddenly dry up and shocks do affect demand for and prices of assets.
29 / 176
This generates endogenous risk and volatility through feedback effects of fire-sales
and financial constraints can become occasionally binding.
• Volatility is high in the crisis regime, which might push experts’ net worth to-
wards zero. In this case, the economy needs a long time to recover. Ex ante, the
system will spend a large amount of time away from the steady state and the
stationary distribution can be bimodal.
• Assets are more correlated during crises due to endogenous risk.
• Endogenous risk-taking gives rise to a volatility paradox, meaning that the econ-
omy does not become more stable when fundamental risk σ is lower. This is be-
cause when risk is lower, experts take on greater leverage, making the economy
more prone to crises.
• Financial innovations (e.g., securitization) that improve risk-sharing among ex-

perts might de-stablize the economy in equilibrium. The logic is similar. Be-
ing able to diversify (idiosyncratic) risk emboldens the experts, leading to higher
leverage and amplifying systemic risk.
In terms of modeling, we highlight the essential techniques for solving large-scale

macro-finance models in continuous time:
• We introduce an occasionally binding constraint in this setting. The “skin in the

game constraint” is not binding in the normal regime, shile it binds in the crisis
regime in which fire sales occur and volatility spikes.
• We rely on the “Fisher separation theorem” in order to solve the model from the
viewpoint of a “price-taking” social planner.
• We introduce a change from a consumption numeraire to a total wealth numeraire,

which simplifies many algebraic steps.
• We explicitly show the value function iteration method in this setting, which is
needed as soon as we relax the assumption of log utility.
30 / 176
This chapter builds on Brunnermeier and Sannikov (2016), which expands on Brunner-
meier and Sannikov (2014).
3.1 Model Setup
Environment. Like before, there is no labor and the economy is populated by experts
and households, i ∈ {e, h}. However, now households can also produce consumption
goods but with an inferior technology. The agents can issue both equity and debt, but
subject to certain financial firctions.
Experts. Experts have a CRS technology yet = ae ket . Denote their consumption and
investment rate by cet , ιet . Experts’ capital stock evolves according to
dket
= (Φ(ιet ) − δ)dt + σdZt .
ket
Still, we have only aggregate risk in the environment. Experts have a CRRA utility
function and they each maximize
"Z #
∞ (cet )1−γ
−ρe t
E0 e dt .
0 1−γ
Households. Households also have a CRS technology yth = ah k ht with ah ≤ ae . House-

holds’ capital accumulation process is
dk ht
= (Φ(ιht ) − δ)dt + σdZt .
k ht
We let households to hold capital to capture fire-sales. Households are more patient
than the experts, i.e., ρh ≤ ρe . As we have discussed in section 2.3, this is a modeling
trick to ensure that the experts do not hold all the capital in the long run. The house-
holds maximize
"Z #
∞ (cth )1−γ
−ρh t
E0 e dt .
0 1−γ
31 / 176
Financial Friction. The financial friction in this chapter is due to incomplete markets
(see, e.g., Dumas and Luciano, 2017). Although experts are allowed to issue equity,
they must hold at least α fraction of their risk. The balance sheets of the two sectors are
as following:
A Expert L A Household L
Debt Loans
Capital Net worth
!"# $"% &" Equity $"% &" − '"#
Outside
'"# Capital
equity
!"( $"% &"
*"# ≥ ,!"#
Eco 529: Brunnermeier & Sannikov
Figure 3.1: Balance sheets of experts and households
The skin-in-the-game constraint can be expressed as χet ≥ ακte , where χet is the frac-
tion of risk held by experts and κte is the fraction of capital held by experts. We will
discuss carefully this relationship in later sections.
3.2 Solution Method
Again, with only aggregate risk, all experts (households) are identical, so total cap-
ital stock and net worth in each sector are obtained by Kti = kit , Nti = nit , i = {e, h}.
Denote the price of capital by qt . The total wealth of the economy is qt ∑i Kti = qt Kt =
Nt = ∑i Nti . Define the capital shares as
Kti
κti = ′,
∑i′ Kti
32 / 176
and the net worth share as
Nti Nti
ηti = ′ = .
∑i′ Nti qt Kt
We then postulate that qt follows
dqt q q
qt
Given the price process and the consumption-investment decision of the expert, we can
calculate the return rate to capital for both sectors, rti,K (ιit ). Same as (2.1), we have
" #
ai − ιit q q q
drti,K (ιit ) = + Φ(ιit ) − δ + µt + σσt dt + (σ + σt )dZt . (3.1)
qt
We then postulate that SDF (ξ ti = e−ρt u′ (cit )) follows
dξ ti
= −rt dt − ςit dZt , (3.2)
ξ ti
where rt is the risk-free rate.

tions
With an additional asset available, sector i’s problem becomes

"Z #
∞ (cit )1−γ
− ρi t
max E0 e dt
{ιit ,θit ,cit }∞
t =0
0 1−γ
dnit cit
s.t. i
= − i
dt + θti,K drti,K (ιit ) + θti,OE drOE i,D
t + θt rt dt (3.3)
nt nt
n0i given,
where rOE
t is the return to outside equity. Note that the outside equity has the same risk
(volatility) as capital but possibly different expected returns (drifts) due to the skin-in-
33 / 176
the-game constraint. The experts’ allocation satisfies
θte,K ≥ 0, −(1 − α)θte,K ≤ θte,OE ≤ 0, θte,D ≤ 0, and θte,K + θte,OE + θte,D = 1. (3.4)
The households’ allocation satisfies
θth,K ≥ 0, θth,OE ≥ 0, θth,D ≥ 0, and θth,K + θth,OE + θth,D = 1. (3.5)
Optimal investment ι. The choice of investment rate is still a static and time-separable
problem. An agent chooses ιet to maximize her return rtK (ιit ). The first-order condition
yields the Tobin’s q equation
1
= Φ′ (ιit ).
qt
With the special functional form Φ(ι) = 1

ϕ log(ϕι + 1), ϕιit = qt − 1.
Asset and risk allocation. We solve the portfolio choice problem via the “price-taking
planner’s problem”, which is widely applicable to environments with multiple assets.
Intuitively, the price-taking planner’s Theorem means that a social planner that takes
prices as given chooses a real asset (capital) allocation κt and risk allocation χt that
coincides with the choices implied by all individuals’ portfolio choices. The planner’s
problem is often of the form
max E Capital Return − (weighted ave. price of risk) × (incremental capital risk),

{κ,χ}
s.t. Financial Friction(s).
Let’s see how this seemingly magical result works in our environment. The price-
taking planner’s problem is
(Weighted) Risk Premium

 z  }|  {
 h i 
∑
K
max Et drtK (κt ) /dt −  ςit χit  σr
× |{z} s.t. χet ≥ ακte .
{κt ,χt } | i ={e,h}
 | {z }
{z } Risk Financial Friction
Expected Return | {z }
(Weighted) Price of Risk
34 / 176
Here rtK (κt ) is the overall return to capital: (ι t = ιet = ιht = ϕ1 (qt − 1))
" #
∑i κti ai − ιt
∑ κti drti,K (ιit ) =
q q q
drtK (κt ) = + Φ(ι t ) − δ + µt + σσt dt + (σ + σt )dZt .
i
qt
Hence, the planner’s problem can be written as

 ! 
 
∑ ςit χit
q
max − (σ + σt ) s.t. χet ≥ ακte . (3.6)
{κt ,χt }  qt i

Theorem 3.1. The equilibrium allocation of physical capital, κte , as well as the allocation of
risk, χet , that arises from agents’ portfolio decisions can be more directly obtained by solving the
“price-taking social planner problem” (3.6).
Proof. The proof takes three steps
1. By Fisher’s Separation Theorem1 , each individual’s portfolio maximization is

equivalent to the following maximization problem of a “firm”. In our model,
individuals in sector i solve2
h i h i i,K
max θti,K Et drti,K (ιit ) /dt + θti,OE Et drOE
t /dt + θti,D rt − ςit (θti,K + θti,OE )σr
{θti,K , θti,OE , θti,D }

(3.4) if i = e
s.t. .
(3.5) if i = h
2. Aggregate {ηt }-weighted sum of the two sectors’ problems:3

h i h i
max
{θit }i={e,h}
∑ tt t t t
η i i,K
θ E dr i,K i
( ι ) /dt + ∑ t t t t /dt + ∑ ηti θti,D rt
η i i,OE
θ E dr OE
i i i
−∑
K
ςit η i (θti,K + θti,OE )σtr , s.t. (3.4) and (3.5).
i
1 We postpone the proof of this result till chapter 7.

2 Recallthat outside equity and capital have the same risk (volatility).
3 Note that σri,K = σr K as the two sectors face the same aggregate risk.
35 / 176
3. Market clearing conditions are
Capital: ηti θti,K = κti , ηti (θti,K + θti,OE ) = χit ,

Outside Equity: ∑ ηti θti,OE = 0,
i
Debt: ∑ ηti θti,D = 0.
i
Note that (3.4) togerther with the capital market clearing condition implies
χet = ηte [θte,K + θte,OE ] ≥ ηte [θte,K − (1 − α)θte,K ] = ακte .
Therefore, the aggregated problem can be simplified to

!
h i
∑ κti Et ∑ ςit χit
K
max drti,K (ιit ) /dt − σtr , s.t. χet ≥ ακte ,
{κt ,χt } i i
which is equivalent to the planner’s problem (3.6).
Although we proved the theorem for this specific model, the three-step argument is
generally valid for more complicated models. Now we can solve the planner’s problem
(3.6) to obtain the risk/capital allocations. The KKT conditions are
n o
χt : min ςet − ςht , χet − ακte = 0, (3.7)
( )
ae − ah q
κt : min − α(ςet − ςht )(σ + σt ), 1 − κte = 0. (3.8)
qt
The FOCs can be visualized as following:
36 / 176
!"# < 1 !"# = 1
.# − .( 4 .# − .( 4
= +(0"# − 0"( ) 2 + 2" > +(0"# − 0"( ) 2 + 2"
$" $"
,"# = +!"# ,"# > +!"#

Figure 3.2: First-order conditions of the planner’s problem
The definition of equilibrium is similar to the previous chapter – a map from histo-
ries of shocks to equilibrium prices.
map from histories { Zs , s ∈ [0, t]} to price qt , investment rate ιit , price of risk ςit and capital/risk
allocation κti , χit , such that
(2) all markets, for capital, equity and consumption goods, clear.
Drift of ηti . One might solve for the process of the state variable ηti = N i /(qt Kti ) by
brute-force, i.e., combining process (3.3) and the market clearing conditions by Itô’s
quotient rule. However, this method involves a formidable amount of algebra and
is thus error-prone. As an alternative, we introduce a new method – “change of nu-
ηi
meraire” – to compute the drift of wealth shares, µt .
Note that as a ratio of two equilibrium quantities, the wealth share ηti (and its
volatility) remain unchanged under a different numeraire, so the change-of-numeraire
ηi
method does not provide any new information on σt .
Change of Numeraire.
C). Let xtA
Consider two different numeraires – call them dollars ($) and euros (A
37 / 176
be the value of a self-financing strategy in $. Denote the exchange rate by Yt :
dYt
= µYt dt + σtY dZt .
Yt
Then, xtA /Yt is the value of the self-financing strategy in A C. Applying the martin-
gale approach, ξ t$ xtA and ξ tAC ( xtA /Y ) are both martingales. Recall that for any two
assets A, B, the martingale approach implies
µtA − µtB = ς$t (σtA − σtB ),

µtA/Y − µtB/Y = ςACt (σtA/Y − σtB/Y ).
By Itô’s quotient rule,
µtA/Y − µtB/Y = (µtA − µtB ) − σtY (σtA − σtB ),

σtA/Y − σtB/Y = (σtA − σtY ) − (σtB − σtY ) = σtA − σtB .
Hence,
(µtA − µtB ) − σtY (σtA − σtB ) = ςACt (σtA − σtB )

⇐⇒ (ς$t − σtY )(σtA − σtB ) = ςACt (σtA − σtB )
⇐⇒ ςACt = ς$t − σtY . (3.9)
We change the numeraire from consumption goods to the total wealth in the econ-
omy Nt = ∑i Nti . Consider two assets:
• Asset A: sector i’s portfolio return in terms of total wealth, that is Nti /Nt = ηti .
The return to this asset is
!
dηti + (Cti /Nt )dt ηi Cti ηi
= µt + dt + σt dZt .
ηti Nti
• Asset B: a benchmark asset that everyone can hold (e.g., risk-free asset or money
in terms of total wealth). In this chapter, asset B is the risk-free loan from the
households to the experts, which has return rt dt.
38 / 176
The martingale asset pricing formula implies
ηi Cti ηi
µt + − rt = ( ςi − σ N ) σt . (3.10)
Nti | t {z t }
price of risk
under Nt numeraire
Aggregate {ηt }-weighted sum of the two sectors

′ ′
′ ηi Ct ′ ′ ηi
∑′ ηti µt +
Nt
− rt = ∑′ ηti (ςit − σtN )σt , (3.11)
i i
| {z }
=0
′
where the first item equals zero because it is the drift of ∑i′ η i = 1. Subtracting (3.11)
from (3.10), the drift of ηti is
!
ηi ηi i′ i′ ηi
′
Cti Ct
µt = (ςit − σtN )σt − ∑ ηt (ς t − σtN )σt − −
i′ Nti Nt
!
ηi i′ i′ ηi
′
Cti Ct
= (ςit − σ − σt )σt − ∑ ηt (ς t − σ − σt )σt −
q q
− ,
i′ Nti Nt
q
where the second equality holds because Nt = qt Kt , and hence σtN = σ + σt .
Volatility of ηti . Recall that the wealth share ηti is numeraire invariant, so we still use
Itô’s quotient rule to solve for its volatility. Since ηti = Nti /Nt ,
′
" #
ηi i′
′ χit i
i′ χt χit − ηti
= σt − σtN = σt − ∑ ηt σt
Ni Ni Ni
∑
q q
σt = − η t i′ (σ + σt ) = (σ + σt ),
i′ ηti i′ ηt i
ηt
where the third equality follows from (3.3) and market clearing conditions, as
i i q χit q
σtN = σtn = (θti,K + θti,OE )(σ + σt ) = i
(σ + σt ).
ηt
Amplification. Applying Itô’s lemma to q(ηte ),
q q′ (ηte ) e η e q′ (ηte ) χet − ηte q

σt = ( η σ ) = (σ + σt ).
q(ηte ) t t q/ηte ηte
39 / 176
The total volatility is
q σ
σ + σt = > σ.
q′ (ηte ) χet − ηte
1−
q/ηte ηte
The amplification effect arises due to fire-sales of capital from the experts to the house-
holds. A negative shock increases market illiquity, leading to expert losses on their
capital stock. Since the experts are levered, when hit by a negative shock, they are
forced to sell their capital stock to the households, causing a further price drop, and so
on – a loss spiral.
Martingale approach works well in endowment economy, where the consumption

steam is exactly the endowment stream, hence the marginal utility is given exogenously
by stochastic discount factor. While for our production economy case, we mix the mar-
tingale approach with value function method.
By the same arguments presented in Section 2.2.3, CRRA utility implies that cit /nit -
ratio is invariant in nit , that is4 .

u ω i (ηt , Kt )nit 1 (ωti nit )1−γ cit
V i (nit ; ηt , Kt ) = = , = (ρi )1/γ (ωti )1−1/γ . (3.12)
ρi ρi 1 − γ nti
Let’s first take a look at a special case. For constant investment opportunities ωti =
i i i i
ω, cit /nit is constant, and hence µct = µnt , σtc = σtn . Furthermore, by Itô’s lemma,
" #
d(cit )−γ i γ (1 + γ ) i 2
i
= −γµct + σtc dt − γσtc dZt .
(cit )−γ 2
4 The value function for individuals i donates V i (nit ; ηt , Kt ), where (ηt , Kt ) are state variables. For
n-sector problem, ηt is a n − 1 vector, while in this chapter ηt is a scalar ηte . For simplicity, later we use
the notations Vti = V i (nit ; ηt , Kt ), ωti = ω i (ηt , Kt ), vit = vi (ηt ).
40 / 176
i i
Because ξ ti = e−ρ t u′ (cit ) = e−ρ t (cit )−γ , Itô’s product rule implies that
" #
dξ ti d ( c i )−γ i γ ( 1 + γ ) i 2 i
i
= −ρi dt + i t −γ = −ρi − γµct + σtc dt − γσtc dZt .
ξt (ct ) 2
i i
Recall the SDF process (3.2), and we now see that ςit = γσtc = γσtn and
i γ (1 + γ ) ci 2
rt = ρi + γµct − σt . (3.13)
2
Consider a self-financing strategy that reinvests consisting of an agent’s net worth with
i
consumption reinvested. By construction, the value of this strategy pnt follows
i
dpnt dnit cit
= + dt.
pnt
i
nit nit
i
The martingale approach tells us that ξ ti pnt follows a martingale and the following asset
pricing equation holds
pn
i
pn
i
i cit i ni (ςit )2
µt − rt = ςit σt ⇐⇒ µnt + − r t = ς σ
t t = .
nit γ
The net worth then follows

" #
dnit n i n i ( ς i )2
t c i
t ςit
= µ t dt + σt dZt = r t + − dt + dZt .
nit γ nit γ
i i i i
Since µct = µnt , σtc = σtn , (3.13) implies
" # " #
( ς i )2 c i γ ( 1 + γ ) ( ς i )2 c i γ − 1 ( ς i )2
rt = ρi + γ rt − t − ti − t
2
=⇒ ti = ρi + r t − ρi + t .
γ nt 2 γ nt γ 2γ
(3.14)
This will turn out to be a useful relationship for the next chapter, but now we focus on
more general investment opportunity processes.
i i
For arbitrary opportunity processes ωti , we still have ςit = γσtc and that ξ ti pnt follows
41 / 176
a martingale. By Itô’s product rule,
i
d(ξ ti pnt ) d(ξ ti nit ) cit
= + dt.
ξ ti pnt
i
ξ ti nit nit
Rewrite (3.12) as
1 i 1− γ i − γ i i 1
(cit )−γ = i
( ωt ) ( nt ) ⇐⇒ eρ t e−ρ t (cit )−γ nit = i (ωti )1−γ (nit )1−γ .
ρ | {z } ρ
ξi | {z
t
}
(1−γ)Vti
Hence,
i
! i
dVti d(eρ t ξ ti nit ) ci d(ξ ti pnt )
= = ρi − ti dt + . (3.15)
Vti i
eρ t ξ ti nit nt ξ i pn
i
| t{zt }
Martingale
Unfortunately, we can not use Itô’s formula on Vti to get the drift of dVti /Vti , as nit (ηt )
and ωti (ηt ) are not differentiable when qt (ηt ) has a kink5 . Instead, we can de-scale the
value function with regard to Kt and define the "de-scaled value function" vit :
1− γ
1 (ωti nit )1−γ wit nit /Kt 1− γ
Kt
Vti = = , (3.16)
ρi 1−γ ρi 1−γ
| {z }
vit :=
By such a de-scaling, we separate two state variables ηti and Kt 6 , hence can work on
them independently.
By Itô’s product rule,

h i
1− γ
dVti d vit Kt
i 1

vi
= 1− γ
= µvt + (1 − γ)(Φ(ι t ) − δ) − γ(1 − γ)σ + (1 − γ)σσt dt + [· · · ]dZt .
2
Vti vit Kt 2
5 Aswe will see in section 3.2.6, this is indeed the case.

6 We will see later vi only depends on state variable ηti , and is twice differentiable in ηti . Besides, state
variable Kt is easy to handle due to scale invariance.
42 / 176
Recall (3.15), the drift of Vti equals
i 1 i ci
µvt + (1 − γ)(Φ(ι t ) − δ) − γ(1 − γ)σ2 + (1 − γ)σσtv = ρi − ti .
2 nt
This gives us the following backward stochastic differential equation (BSDE)

" #
dvit cit 1 vi i
i
= ρ − i − (1 − γ)(Φ(ι t ) − δ) + γ(1 − γ)σ − (1 − γ)σσt dt + σtv dZt .
i 2
vt nt 2
(3.17)
This is a BSDE that we can solve using standard numerical methods (Tourin, 2011). But
before that, we need to study the evolution of the state ηti in order to pin down terms
like cit /nit and ιit .
Price of risk ςit . The value function (3.12) implies

u ω i (ηt , Kt )nit
V i (nit ; ηt , Kt ) =
ρi
1− γ !1− γ
∂V i (nit ; ηt , Kt ) ω i (η t , Kt ) (ωti nit /Kt )1−γ Kt
i −γ
⇒ = (n ) = (nit )−γ
∂nit ρi ρi nit
| {z }
vit :=
Applying6 the envelop condition ∂Vt

∂nt = u ′ ( c t ),
!1− γ
∂Vti Kt ∂u(cit )
i
= vit (nit )−γ = (cit )−γ = .
∂nt nit ∂cit
In equilibrium Nti = nit and Cti = cit , plugging in Nti = ηti qt Kt , the condition ends up
becoming
!1/γ
Cti ηti qt
= . (3.18)
Kt vit
43 / 176
Applying Itô’s quotient rule and comparing the volatility terms, we have7

i 1 i ηi q
σtc −σ = −σtv + σt + σt .
γ
The prices of risk are then
i i ηi q
ςit = γσtc = −σtv + σt + σt + γσ. (3.19)
Consumption propensity Cti /Nti . Note that we can express Cti /Nti in terms of ηt , qt
and vit . Plug in Kt = Nti /ηti qt and rewrite (3.18) as
Cti cit (ηti qt )1/γ−1

= = . (3.20)
Nti nit (vit )1/γ
On the aggregate level,

!1/γ
i ηti qt
Ct i Ct 1
Nt
= ∑ t N i = qt
η ∑ vit
. (3.21)
i t i
3.2.4 Value function iteration and goods market clearing
To apply the finite difference method, we postulate that vit = vi (ηte , t) (Note in two-
sector model we only use η e as state variable). By Itô’s formula, it follows
ηe ηe ηe
dvit ∂t vit + (ηte µt )∂η vit + 12 (ηte σt )2 ∂ηη vit (ηte σt )∂η vit
= dt + dZt .
vit vit vit
Comparing with the BSDE (3.17), we get the growth equation

h
ηe
i 1 e η e 2
∂t vit + ηte µt ∂η vit + ηt σt ∂ηη vit
2
 " #
i
 cit 1 η e ∂η vt

= ρi − − (1 − γ) (Φ(ιit ) − δ) − γσ2 + σ ηte σt vit , (3.22)
 nit 2 vti 
i
7 Note that σtK = σ because Kt = ∑i κ i kit and σtk = σ, ∀i.
44 / 176
where
!
ηi q ηi i′ i′ q ηi
′
Cti Ct
µt = (ςit − σ − σt )σt −∑ ηt (ς t − σ − σt )σt − i
− , (3.23)
i′ Nt Nt
ηi χit − ηti q
σt = i
(σ + σt ). (3.24)
ηt
In order to solve this PDE, we need to know all the terms in red. Luckily, we already
have all the building blocks from the previous sections:
• Tobin’s q gives us the investment rate ιit :
1
ιit = ( q t − 1). (3.25)
ϕ
• The price of risk ςit is given by (3.19):
ηe
vi ηi q vi (η e σ )∂η vit
ςit = −σt + σt + σt + γσ where σt = t t i . (3.26)
vt
q
• The amplification equation yields σt :
q q′ (ηte ) χet − ηte q

σt = e e (σ + σt ). (3.27)
q/ηt ηt
• The first-order conditions to the planner’s problems gives χit , κti . It can be shown
that the FOCs (3.7)-(3.8) are equivalent to the following8
 ! 
 ae − ah ∂η ve ∂η vth 1 q

min −α − et + + (χet − ηte )(σ + σt )2 , 1 − κte = 0,
 qt vt vth (1 − ηte )ηte 
(3.28)
χet = max{ακte , ηte }. (3.29)
• In (3.20) and (3.21), we have obtained the consumption ratios cit /nit , Ct /Nt from
8 Proof will be added soon. For now, please see relevant parts of the lecture slides.
45 / 176
the optimal consumption condition:

!1/γ
Cti cit (ηti qt )1/γ−1 Ct i
i Ct 1 ηti qt
= i = , = ∑ ηt i = ∑ . (3.30)
Nti nt (vit )1/γ Nt i Nt qt i vit
• Finally, goods market clearing jointly constraints qt and κti :

!1/γ
Cti ηti qt
∑ t
κ i i
a − ι t = ∑ Kt = ∑ vit
. (3.31)
i i i
In the following algorithm, we first guess two functions ve (ηte , T ), vh (ηte , T ) and use
them as terminal conditions. We then solve the PDE (3.22) backwards on a discretized time
grid. In each step (time t), we solve for time-t equilibrium quantities as functions of ηt
using conditions (3.25)-(3.31).
More specifically, the algorithm goes as following
1. Start by guessing two functions ve (η e , T ), vh (η e , T ) over a grid of η e
2. Loop over t = { T, T − ∆t, · · · , 0} until changes in ve -functions are small. In

each step, do the following:
(a) Compute ∂η vit by first-order differences

(b) Start at q(0) (the autarky economy) and solve to the right. Use different
procedures for two η regions
i. If κte < 1, solve ODE for q(ηte ) using conditions (3.31), (3.28) and
(3.27). Specifically, we have to first plug (3.25) into (3.31) and (3.29)
into (3.28). Then solve the three equilibrium conditions to the right
using Newton’s method
q
ii. If κte = 1, (3.28) is no longer informative about σt . Instead, we
solve (3.25) and (3.31) for q(ηte ), again, using Newton’s method
ηe ηe i i
(c) compute µt (ηte ), σt (ηte ), µvt (ηte ), σtv (ηte ) using equations (3.23) and (3.24)
(d) make time-step – back in time – and update the vit (t, ·) functions to
vit (t − ∆t, ·)
46 / 176
𝑣 " (𝜂 " , 𝑡) 𝑣 " (𝜂 " , 𝑇)
𝑡
𝑇
𝑇−Δ
𝑇 − 2Δ
1 𝜂"
Figure 3.3: Visualization of the time step

4
Newton’s Method.
The classic one-dimensional Newton’s Method finds the root of a real-valued
function by successively computing the intercept of the tangent line approximation
of this function. Mathematically, it iteratively computes
zn+1 = zn − [ f ′ (zn )]−1 f (zn ).
Figure 3.4: One-dimensional Newton’s method
47 / 176
For a multi-dimensional system F (z) = 0, the Newton’s method proceeds as
zn+1 = zn − Jn−1 F (zn ), (3.32)
where Jn is the Jacoban matrix, i.e., Ji,j = ∂ f i (z)/∂z j .
In the algorithm above, we use Newton’s method to approximate the solution to

our ODE. Specifically, we let F (·) be the the three equilibrium conditions and z =
{q, κ e , (σ + σq )}. Starting from the autarky solution z0 , we then iteratively compute
to the right zn by (3.32) on the η e grid. Note that for each grid point, we essentially only
conduct the first step in the Newton’s method. It can be shown that the error is of order
O((z − z∗ )2 ), where z∗ is the true solution. Since our η e grid will be very dense and all
variables are continuous in η e , the error should be negligible in practice.
If using only the first step is insufficient, one can easily switch to a multi-step New-
ton’s method. For example, in the main program of Section 3.2.5, one only needs to add
an additional loop for codes between line 43 and 68. However, it is important to keep
in mind that Newton’s method does not guarantee global convergence. In practice, a
multi-step Newton’s method might diverge, and one should check the gain from each
additional step to deter possible divergence.
3.2.5 Implementation in MATLAB
Following is main code executing the algorithm.

1 %% Housekeeping
2 close all ; clear ; clc
3
4 %% Parameters
5 rho_e = 0.06; rho_h = 0.05; delta = 0.05; a_e = 0.11; a_h = 0.03;
6 sigma = 0.10; alpha = 0.50; gamma = 2; phi = 10;
7
8 %% Grid
9 % uneven grid on [0 , 1] with 1000 points , with more points near 0 and 1
10 N = 1000; zz = linspace (0.001 ,0.999 , N ) ';
11 Eta = 3* zz .^2 - 2* zz .^3; dEta = Eta (2: N) - Eta (1: N -1) ;
12
13 %% Terminal conditions for value functions V_e and V_h
14 V_e = a_e ^( - gamma )* Eta .^(1 - gamma ); V_h = a_e ^( - gamma ) *(1 - Eta ) .^(1 - gamma );
15
48 / 176
16 Q = ones (N ,1) ; Qp = zeros (N ,1) ; SSQ = zeros (N ,1) ; Kappa_e = zeros (N ,1) ;
17 Chi_e = zeros (N ,1) ; MU = zeros (N ,1) ; MV_e = zeros (N ,1) ; MV_h = zeros (N ,1) ;
18 G = zeros (N ,1) ; S = zeros (N ,1) ; SV_e = zeros (N ,1) ; SV_h = zeros (N ,1) ;
19
20 for t = 1:1000
21 %% Initialization
22 GS1 = ( Eta ./ V_e ) .^(1/ gamma ); GS2 = ((1 - Eta ) ./ V_h ) .^(1/ gamma ); GS = GS1 + GS2 ;
23 Vp_e = ( V_e (2: N) - V_e (1: N -1) ) ./ dEta ; Vp_h = ( V_h (2: N) - V_h (1: N -1) ) ./ dEta ;
24 Vpl_e = Vp_e ./ V_e (2: N); Vpl_h = Vp_h ./ V_h (2: N);
25 VVlp = Vpl_h - Vpl_e + 1./( Eta (2: N) .*(1 - Eta (2: N) ));
26
27 %% Initial condition at Eta (1)
28 kappa_e = 0; ssq = sigma ;
29 qL = 0; qR = a_h * phi + 1;
30 for k = 1:30
31 q = ( qL + qR ) /2; iota = (q - 1) / phi ; A0 = a_h - iota ;
32 if log (q)/ gamma + log ( GS (1) ) > log ( A0 )
33 qR = q;
34 else
35 qL = q;
36 end
37 end
38 Q (1) = q; q_old = q;
39
40 %% Newton method
41 % find q , kappa and ssq = sigma + sigma ^ q from value functions
42 for n = 2: N
43 % errors given guess
44 ER = [ log (q )/ gamma + log ( GS (n)) - log ( a_e * kappa_e + a_h *(1 - kappa_e ) - (q -1) / phi
);
45 ssq *( q - ( q - q_old ) *( alpha * kappa_e - Eta (n) )/ dEta (n -1) ) - sigma *q;
46 a_e - a_h - q* alpha *( alpha * kappa_e - Eta (n ))* ssq ^2* VVlp (n -1) ];
47 % matrix of derivatives of errors ( could shorten it since q_old = q)
48 QN = zeros (3 ,3) ;
49
50 QN (1 ,:) = [1/( q* gamma ) + 1/(( a_e - a_h )* kappa_e + a_h - (q - 1) / phi )/ phi , ...
51 -( a_e - a_h ) /(( a_e - a_h )* kappa_e + a_h - (q - 1) / phi ) , 0];
52
53 QN (2 ,:) = [ ssq *(1 - ( alpha * kappa_e - Eta (n))/ dEta (n -1) ) - sigma , ...
54 - ssq *( q - q_old )* alpha / dEta (n -1) , q - (q - q_old ) *( alpha * kappa_e - Eta (n ))/ dEta (n
-1) ];
55
56 QN (3 ,:) = [ - alpha *( alpha * kappa_e - Eta ( n))* ssq ^2* VVlp (n -1) , ...
57 -q * alpha ^2* ssq ^2* VVlp (n -1) , -2* q* alpha *( alpha * kappa_e - Eta (n))* ssq * VVlp (n -1)
];
58
59 EN = [q; kappa_e ; ssq ] - QN \ ER ;
60
61 % if the boundary of the crisis regime has been reached . kappa_e = 1 from now
on
62 if EN (2) > 1
63 break ;
49 / 176
64 end
65
66 % new guesses
67 q = EN (1) ; kappa_e = EN (2) ; ssq = EN (3) ;
68
69 % save results
70 Q(n ) = EN (1) ; Kappa_e ( n) = EN (2) ; SSQ (n) = EN (3) ;
71 Qp (n) = (Q( n) - q_old )/ dEta (n -1) ; q_old = EN (1) ;
72
73 end
74
75 %% Newton method for kappa_e = 1, for remaining eta
76 n1 = n;
77 for n = n1 :N
78 ER = log (q) / gamma + log ( GS (n)) - log ( a_e - (q - 1) / phi );
79 QN = 1/( q* gamma ) + 1/( a_e - (q - 1) / phi )/ phi ;
80 EN = q - ER / QN ;
81
82 q = EN ; Q(n ) = EN ; Kappa_e (n ) = 1; Qp (n) = ( EN - q)/ dEta (n -1) ;
83 SSQ (n) = 1./(1 - ( max ( alpha , Eta (n)) - Eta (n) )* Qp (n )/Q(n ))* sigma ;
84 end
85
86 %% Computing the PDE
87 Chi_e = max ( alpha * Kappa_e , Eta );
88 S = ( Chi_e - Eta ) .* SSQ ; S(N ) = 0;
89 Iota = (Q - 1) / phi ;
90 Phi = log ( Q)/ phi ;
91 A = a_e * Kappa_e + a_h *(1 - Kappa_e ) - Iota ;
92
93 CN_e = GS1 .* Q .^(1/ gamma - 1) ./ Eta ;
94 CN_h = GS2 .* Q .^(1/ gamma - 1) ./(1 - Eta );
95
96 SV_e (2: N ) = Vpl_e .* S (2: N);
97 SV_h (2: N ) = Vpl_h .* S (2: N);
98
99 VarSig_e = - SV_e + S ./ Eta + SSQ + ( gamma - 1) * sigma ;
100 VarSig_h = - SV_h - S ./(1 - Eta ) + SSQ + ( gamma - 1) * sigma ;
101
102 MU (2: N -1) = (( a_e - Iota (2: N -1) ) ./ Q (2: N -1) - CN_e (2: N -1) ) .* Eta (2: N -1) + ...
103 S (2: N -1) .*( VarSig_e (2: N -1) - SSQ (2: N -1) ) + ...
104 Eta (2: N -1) .* SSQ (2: N -1) .*( VarSig_e (2: N -1) - VarSig_h (2: N -1) ) *(1 - alpha );
105
106 MV_e = rho_e - CN_e - (1 - gamma ) *( Phi - delta - gamma * sigma ^2/2 + SV_e * sigma );
107 MV_h = rho_h - CN_h - (1 - gamma ) *( Phi - delta - gamma * sigma ^2/2 + SV_h * sigma );
108
109 %% Updating V_e and V_h
110 % lambda0 is dt * rho if dt is small , can be at most 1 (1 = policy iteration )
111 % it is more agressive to set lambda0 closer to 1 , but code may not converge
112
113 lambda0 = 0.8;
114 V_e = payoff_policy_growth ( Eta , MV_e , MU , S , G , V_e , lambda0 );
115 V_h = payoff_policy_growth ( Eta , MV_h , MU , S , G , V_h , lambda0 );
50 / 176
116
117 end
The time step is summarized in the function payoff_policy_growth.m:

1 function F = payoff_policy_growth (X , R , MU , S , G , V , lambda )
2 % X , R , MU , S , and G are column vectors of the same length
3 % X = [X (1) , X (2) ... X(N )]' is the state space ( an increasing grid )
4 % R is the discount rate minus growth
5 % MU is a drift vector of length N , with MU (1) >= 0 and MU (N) <= 0
6 % S is a volatility vector of lenght N , with S (1) = S(N) = 0
7 % G is a payoff flow of length N
8 % We are solving the value function equation R(X)* F(t ,X) = G(X) + MU (X)* F_x (t ,X ) + S(X)
^2* F_xx (t ,X) /2 + F_t (t ,X)
9 % To solve the equation over a small time interval dt with continuation value function
V , set lambda = dt /(1 + dt ) ( or simply lambda = dt )
10 % To get the stationary solution of this equation , set lambda = 1: then V has no effect
on the solution
11 % Given value F(t + dt , X ) = V , the value of F(t , X ) is found through an implicit
scheme , by solving the matrix equation (R - MU *D - S ^2* DD /2 + I/ dt )*F(t) = G + F(t+
dt )/dt , where D denotes the first derivative operator ( in the upwind direction ) and
DD denotes the second derivative operator
12 % Multiplying both sides by lambda = dt /(1 + dt ) , we obtain the actual equation solved ,
(( R - MU *D - S ^2* DD /2) * lambda + I *(1 - lambda ))*F (t) = G* lambda + F (t+ dt ) *(1 -
lambda ) where the matrix multiplying F(t) is the matrix which is inverted
13
14 if or ( MU (1) < 0 , MU ( end ) > 0) ,
15 display ( ' error : not true that MU (1) >= 0 and MU (N) <= 0 ');
16 end
17 if or ( S (1) ~= 0 , S( end ) ~= 0) ,
18 display ( ' error : not true that S (1) and S( N) both zero ') ;
19 end
20
21 N = length (X );
22 dX = X (2: N) - X (1: N -1) ;
23
24 S0 = zeros (N ,1) ; S0 (2: N -1) = S (2: N -1) .^2./( dX (1: N -2) + dX (2: N -1) );
25 DU = zeros (N ,1) ; DU (2: N) = - ( max ( MU (1: N -1) ,0) + S0 (1: N -1) ) ./ dX * lambda ;
26 DD = zeros (N -1 ,1) ; DD = - ( max (- MU (2: N) ,0) + S0 (2: N)) ./ dX * lambda ;
27
28 D0 = (1 - lambda )* ones (N ,1) + lambda *R;
29 D0 (1: N -1) = D0 (1: N -1) - DU (2: N) ; D0 (2: N) = D0 (2: N) - DD ;
30 A = spdiags ( D0 ,0 ,N ,N) + spdiags ( DU ,1 ,N ,N) + spdiags ( DD (1: N -1) , -1,N ,N );
31 F = A \( G* lambda + V *(1 - lambda ));
To ensure numerical stability, it is important to:
• Use implicit method in step 2.(b)
51 / 176
• Use an upwind scheme when taking derivatives in step (a):


 vi (t, η i (n + 1)) − vi (t, η i (n)) ηi
if µt ηti > 0


η i ( n + 1) − η i ( n )



∂η vit (t, η i (n)) = (3.33)
 vi (t, η i (n)) − vi (t, η i (n − 1)) ηi
if µt ηti < 0



η i ( n ) − η i ( n − 1)


A good initial guess is usually crucial to the success of a numerical procedure. Here
are some common ways of choosing initial guesses for vi :
• Take an arbitrary constant, e.g. a vector of ones. It is the easiest way, but doesn’t
always work / may take a long time to converge.
• Take a specific constant, namely the value at the boundary steady state (η e = 0 or
η e = 1) where only one type exists (if that is a valid equilibrium) – this is typically
also very easy.
• Assume there are no financial contracts and compute for each η e the autarky value
of the agent types, when the initial wealth distribution is described by η e
• Along the same lines, assume complete markets and compute first-best utility as
a function of η e (this certainly bounds utility from above)
• If the log utility model is simple to solve, solve it first. Use the consumption path
of agents in that model, but compute the implied CRRA utility.
• If you have solved the model for different parameters that are “close”, use that
solution as an initial guess.
3.2.6 Numerical Results
In this section, we demonstrate the solutions generated by the code in Section 3.2.5
and discuss their implications. The baseline parameters are as follows.
ρe ρh ae ah δ σ α γ ϕ
0.06 0.05 0.11 0.03 0.05 0.10 0.50 2 10
52 / 176
Figure 3.5 illustrates the equilibrium with baseline parameter values. Note that
q(ηte ) indeed has a kink, which marks the boundry between the cirsis region near η e = 0
and the normal region near η e = 1. In the crisis region, κ e < 1, and households hold
some capital, while in the normal region, experts hold all capital in the economy.
Figure 3.5: Equilibrium for the baseline set of parameters (η should read η e )
3.3 Stationary Distribution and Fan Charts
3.3.1 Stationary distribution
Recall that in the simple model in Chapter 2, experts hold all the capital in the long
run. In this chapter, we assumed differente discount rates for experts and households
to avoid a degenerate stationary distribution. Figure 3.6 shows the drift and volatil-
ity of η e . There exists an η ∗ where the drift of η e becomes zero. η ∗ can be viewed as
the “steady state” of this model. In the absence of shocks, the system will converge to
and stay at the steady state. In response to small shocks, drifts of η e can still push the
economy back to the steady state. Moving away from the crisis regime, risk premia de-
cline, which boosts experts’ consumption and lowers the drift of η e . At η ∗ , risk premia
decline sufficiently so experts’ income are exactly offset by their higher consumption
propensity (recall ρe > ρh ), and hence their wealth share stays constant.
53 / 176
The region where η e ≥ α = 0.5 reflects perfect risk sharing between experts and
households, where the volatility of η e is zero. Since of drift is negative, the system will
never stay in this region. If we start there, the system deterministically moves to η e = α.
Figure 3.6: Drift and Volatility of η e
The stationary distribution of the system can be derived using Kolmogorov Forward
Equation. Consider a n-dimensional Itô diffusion X with law of motion
dX = µ( X )dt + σ( X )dZ, (3.34)
where µ : Rn → Rn , σ : Rn → Rn×m and Z is a m-dimensional Brownian motion. The

stationary KFE for f X is
n 1 n n ∂2

∂
0=−∑ µi ( x ) f X ( x ) + ∑ ∑ σ( x )σ( x ) T
f X (x) .
i =1
∂xi 2 i=1 j=1 ∂xi ∂x j ij
This is a linear equation for the function f X , namely T f X = 0 with the differential
operator T defined by
n 1 n n ∂2

∂
T f := − ∑ µi f + ∑ ∑ σσ T
f . (3.35)
i =1
54 / 176
One can show by integration by parts that

 
n n n
1 ∂2
Z Z
∂
f ( x )  ∑ µi ( x ) g( x ) + ∑ ∑ σσ T

T f ( x ) g( x )dx = (x) g( x ) dx,
i =1
∂xi 2 i =1 j =1 ij ∂xi ∂x j
| {z }
=:(Sg)( x )
so T is the adjoint of the operator S. We then discretize the differential operator S by a

matrix A. In finite dimensions, forming adjoints means taking transposes, so that A T
is a finite-dimensional approximation of T. We interpret A as the transition matrix of
a continuous-time Markov chain. Then its basis y (i.e., A T y = 0) is a multiple of the
invariant distribution of this Markov chain, so one can divide y by its (unweighted)
sum to obtain invariant probabilities. Then we can form the cumulative sum to get the
CDF and approximate the density by taking finite differences.
In next section we provide MATLAB program KFE.m to solve stationary and time-
dependent Kolmogorov forward equation. Figure 3.7 plots the stationary distribution.
Note that any monotone transformation of η e is also a valid state variable, including
the CDF of η e .
15
10
f (η e )
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ηe
Figure 3.7: Stationary distribution
55 / 176
3.3.2 Fan Charts
In standard macroeconomic models with a fixed steady state, it is commonplace to

start at the deterministic steady state and shock the state variable with a one standard
deviation (negative) shock (bad unanticipated realization). Subsequently one can ob-
serve how the system/economy converges back to the steady state, i.e., by plotting the
the impulse response function.
The mathematical tool for studying transition paths in a continuous-time environ-

ment is given by the time-dependent Kolmogorov forward equation. For a process X
with the evolution (3.34), the KFE is
n 1 n n ∂2

∂ ∂
f X ( x, t) = − ∑ µi ( x ) f X ( x, t) + ∑ ∑ σ( x )σ( x ) T
f X ( x, t)
∂t i =1

or, shorter, ∂ f X ( x, t) /∂t = T f X (·, t) ( x ) with T as defined in (3.35). To solve such a
linear parabolic PDE, we can borrow the A T approximation matrix from the stationary
case and iterate the right hand side of KFE forward.9
The following MATLAB function outlines this solving method for both stationary
and time-dependent Kolmogorov forward equation:
1 function [ pdf_stat , varargout ] = KFE (X ,MU ,S ,T , F0 )
2 % KFE Solve one dimensional stationary and time - dependent KFE by finite
3 % difference method The process being studied is dX = MU (X) dt + S(X) dZ_t .
4
5 % REQUIRED INPUT :
6 % X: [X (1) , X (2) ... X(N)] ' is the state space ( can be uneven grid )
7 % MU : a drift vector of length N , with MU (1) >= 0 and MU ( N) <= 0
8 % S: a volatility vector of lenght N , with S (1) = S(N ) = 0
9 % OPTIONAL INPUT FOR TIME - DEPENDENT KFE :
10 % T: time grid , M *1
11 % F0 : initial distribution vector , N *1
12
13 % INPLEMENT
14 % 1. For stationary distribution , [ pdf_stat ] = KFE (X ,MU ,S );
15 % 2. For distribution diffusion , [ pdf_stat , cdf ] = KFE (X ,MU ,S ,T , F0 );
16
17 % NOTE : 1. Fokker Planck operator ( KFE ) is the adjoint operator of Feynman Kac
18 % operator ( KBE ). We first build Feynman Kac operator and then transpose it .
19 % 2. We use upwind scheme and implicit scheme for monotonicity and stability .
20
9 In the one-dimensional case can also be solved with MATLAB built-in solver pdepe.
56 / 176
21 N = length (X );
22 dX = X (2: N) - X (1: N -1) ;
23 %% 1. Build Fokker - Planck operator
24 % approximate drift terms with an upwind scheme
25 % upper diagonal
26 AU = max ( MU (1: N -1) ,0) ./ dX ;
27 % lower diagonal
28 AD = - min ( MU (2: N) ,0) ./ dX ;
29 % main diagonal
30 A0 = zeros (N ,1) ; A0 (1: N -1) = A0 (1: N -1) - AU ; A0 (2: N) = A0 (2: N) - AD ;
31 % matrix A
32 A = sparse (1: N ,1: N ,A0 ,N ,N) + sparse (1: N -1 ,2: N ,AU ,N ,N) + sparse (2: N ,1: N -1 , AD ,N ,N );
33
34 % approximate volatility terms
35 % sigma ^2/( x_ {n +1} - x_ {n -1})
36 S0 = zeros (N ,1) ; S0 (2: N -1) = S (2: N -1) .^2./( dX (1: N -2) + dX (2: N -1) );
37 % upper diagonal
38 BU = S0 (1: N -1) ./ dX ;
39 % lower diagonal
40 BD = S0 (2: N ) ./ dX ;
41 % main diagonal
42 B0 = zeros (N ,1) ; B0 (1: N -1) = B0 (1: N -1) - BU ; B0 (2: N) = B0 (2: N) - BD ;
43 % matrix B
44 B = sparse (1: N ,1: N ,B0 ,N ,N) + sparse (1: N -1 ,2: N ,BU ,N ,N) + sparse (2: N ,1: N -1 , BD ,N ,N );
45
46 % Fokker Planck operator
47 FP = (A+B) ';
48
49 %% 2. Find stationary distribution .
50 % MATLAB doesn 't have build - in kernel solver for sparse matrix , for higher
51 % efficiency one can use online package like spnull , etc .
52 F_stat = null ( full ( FP )) ;
53 cdf_stat = cumsum ( F_stat (: ,1) ./ sum ( F_stat (: ,1) ) );
54 pdf_stat = [0;( cdf_stat (2: end ) - cdf_stat (1: end -1) ) ./ dX ];
55
56 %% 3. Solve the time - dependent KFE .
57 if nargin == 5
58 F = F0 ;
59 DT = [0 T (2: end ) -T (1: end -1) ];
60 pdf_diffusion = zeros ( length (T) , length (F)) ;
61
62 for i = 1: length (T)
63 F = ( speye (N ,N) - DT ( i)* FP ) \F;
64 pdf_diffusion (i ,:) = F;
65 end
66 varargout {1} = pdf_diffusion ;
67 end
68
69 end
To visualize the transition, we can use fan charts originally introduced by the Bank
57 / 176
of England. The first type of fan chart plots the evolution of (the distribution of) the
state variable after a shock – the “distributional impulse response”. The idea is sim-
ilar to the impulse response function in the DSGE literature, but instead of imposing
a one-time shock, it studies the dynamics of the whole system. Figure 3.8 plots the
convergence back to the stationary distribution after a shock to an economy originally
at the median of the stationary distribution. To simulate a negative shock, we set the
original Brownian shock at its 1% quantile (dZt = −2.32dt) for a period of ∆t = 1.
The dashed lines indicate different quantiles of the distribution and the solid line is the
median response. The color is monotone in density.
0.5
Shock
0.45
0.4
0.35
0.3
-10 0 25 50 75 100
Figure 3.8: Distributional impulse response at stochastic steady state
As an alternative, we can start the transition from the stationary distribution. See
the following video for a more direct display of the transition. (Adobe Acrobat needed.)
58 / 176
59 / 176
More interestingly, the second type of fan chart plots the difference between distribu-
tions with and without the shock. As we find in Figure 3.9, the difference converges to
zero in the long run.
0.01
Shock
-0.01
-0.02
-10 0 25 50 75 100
Figure 3.9: Distributional impulse response (difference to unshocked path), σ = 0.15
3.4 Discussions
In this section we study how specific parameters (ah , σ, α and γ) affact the equilib-
rium.
Figure 3.10 shows the effect of σ on the equilibrium. The steady state η ∗ drops as
σ declines, while risk premia fall in the normal region, until η ∗ reaches the boundary
q
of the crisis region (the kink in q).10 A volatility paradox emerges: endogenous risk σt
does not necessarily fall as σ declines. Note that as σ → 0, the boundary of the crisis
region does not converge zero – there is always some positive endogenous risk.
10 This happens for σ = .01 in Figure 3.10.
60 / 176
Figure 3.10: Equilibrium for σ= .1 (blue), .05 (red) and .01 (black)
The effect of relaxed financial friction is similar. As shown in Figure 3.11, endoge-
q
nous risk σt increases as α falls.
Figure 3.11: Equilibrium for α= .5 (blue), .2 (red) and .1 (black)
The household productivity ah has a major impact on stability of the system. ah re-
flects how households value capital, when they are forced to hold it. Figure 3.12 shows
the equilibrium dynamics for different values of ah . Endogenous risk significantly in-
61 / 176
creases as ah declines while the behavior in the normal regime and η ∗ are insensitive to
ah .11 It is surprising that although expert leverage responds to fundamental risk σ, it
does not respond to endogenous tail risk.
Figure 3.12: Equilibrium for ah = .03 (blue), -.03 (red) and -.09 (black)
3.5 Exercises
3.5.1 He and Krishnamurthy (2013)
The goal of this problem is to characterize equilibria in the model of He-Krishnamurthy

(“Intermediary Asset Pricing") and to use the iterative method to compute equilibria.
There are two agent types: experts and households. Households have log utility, while
experts have CRRA utility with relative risk aversion γ, and both with discount rate ρ.
New households are born continuously, and the newborn receive labor income at rate
lKt .
11 As
noted in Brunnermeier and Sannikov (2016), for log utility, it can be analytically proved that the
dynamics in the normal regime is independent of ah .
62 / 176
Aggregate capital follows the law of motion
dKt
= gdt + σdZt .
Kt
Capital produces dividend of aKt . The price of capital per unit is denoted by qt and
follows
dqt q q
qt
Only experts can hold capital, and they can finance capital by borrowing through risk-
free debt and by issuing equity to households, but they must retain fraction of at least
χ = 1/(1 + m) of risk.
(a) Write down the expression for drtK for the return on capital.
Experts make optimal consumption and portfolio decisions: they choose how
much to borrow and how much outside equity to issue (up to fraction 1 − χ)
to buy capital. Denote the required risk premium of experts by ςet and recall that
e e
ςet = −γσtC , where σtC is the volatility of aggregate consumption of experts. De-
note the value function of a representative expert by
1− γ
Kt
vet .
1−γ
(b) Write down the law of motion of aggregate net worth of experts Nte as a function
of the risk-free rate rt , the experts’ equity share χet , the experts’ net worth share
q
ηte , the price of capital qt , the volatility capital return σ + σt , the experts’ risk pre-
mium ςet and process vet . To write the law of motion of Nte , you need to express the
experts’ consumption rate Cte /Nte as a function of ηte , qt and vet .
(c) He and Krishnamurthy assume that inside and outside equity of experts earn
the same returns. Thus, the experts’ equity held by households earns the risk
premium of ςet , even though households’ required risk premium is higher. Under
this assumption, write down the law of motion of world wealth qt Kt , as a function
q
of the risk-free rate rt , the price of capital qt , the volatility capital return σ + σt ,
the experts’ risk premium ςet and output parameter a.
63 / 176
(d) From your answers to parts (b) and (c), derive the law of motion of the experts’
wealth share ηt = Nte /(qt Kt ).
(e) Write down the market-clearing condition for output. Hint: Recall that total
world output is ( a + l )Kt , including dividend and labor income of newborn house-
holds.
Next, you should determine the size of the “constrained region" where χet = χ
and the size of the unconstrained region where χet > χ. To do that, you should use
the following assumptions of He and Krishnamurthy. Assume that fraction λ of
households (i.e. the net worth share of these households is (1 − ηte )λ) are “debt"
households who can only hold the risk-free asset. Fraction 1 − λ are “equity"
households who can hold outside equity of experts and the risk-free asset. He and
Krishnamurthy furthermore assume that equity households cannot use leverage,
i.e. the risk of their net worth can be at most equal to the risk of experts’ net worth
(who hold their own inside equity). Assume (you can verify this later), that it is
this constraint that determines the amount of equity that experts can issue.
(f) Derive the value of χet as a function of ηte implied by the constraint that equity
households cannot use leverage.
The goal of the next questions is to formulate a procedure to compute equilibria

using Matlab, using the value function iteration in section 3.2.5. You should use
Matlab function payoff_policy_growth.m to perform the “time step" of the iterative
procedure.
(g) Formulate a procedure for the static step. That is, suppose you are given function
ve (t, η e ) for all η e at time t. Find the price of capital q(t, η e ) for all η e at time t.
Then, given this function, derive the law of motion of η e . Provide an expression
e
for µvt .
(h) Formulate a procedure for the time step. That is, for the function
F = payoff_policy_growth(X, R, MU, S, G, V, lambda0),
what values of X, R, MU, S, G, V, lambda0 should you use?
64 / 176
BIBLIOGRAPHY
(i) Program the iterative procedure using the terminal condition ve ( T, η ) = a−γ (η e )1−γ .
Use N = 1000. Compute an example for the parameters of He and Krishna-
murthy, ρ = 0.04, g = 0.02, m = 4, a = 1, l = 1.84, σ = 0.09, γ = 2 and λ = 0.6.
(See Table 2 of HK - for these parameters you should be able to get convergence
by setting lambda0 for payoff_policy_growth aggressively to 0.9).
Plot, as a function of η e , the price of capital q, the risk-free rate rt , the drift and
ηe ηe
volatility of ηt η (i.e. σt η e and µt η e ), the fraction of equity χet held by experts and
the experts’ consumption rate Cte /Nte .
(j) Replicate Figure 2 from He and Krishnamurthy, where the vertical axis displays
q
the risk premium for capital, i.e. ςet (σ + σt ).
3.5.2 Brunnermeier and Sannikov (2014)
In Brunnermeier and Sannikov (2014), experts and less productive households are
risk neutral. However, while consumption of households can go negative, for experts
it has to stay non-negative. Hence, experts become extremely risk averse when con-
sumption approaches zero. As a result, the stationary distribution is bimodal.
Try to replicate the paper with tools studied in this chapter.
Bibliography
Brunnermeier, Markus K. and Yuliy Sannikov, “A Macroeconomic Model with a Fi-

nancial Sector,” American Economic Review, 2014, 104 (2), 379–421.
and , “Macro, money, and finance: A continuous-time approach,” in “Handbook

of Macroeconomics,” Vol. 2, Elsevier, 2016, pp. 1497–1545.
Dumas, Bernard and Elisa Luciano, The Economics of Continuous-Time Finance, MIT
Press, 2017.
He, Zhiguo and Arvind Krishnamurthy, “Intermediary asset pricing,” American Eco-
nomic Review, 2013, 103 (2), 732–70.
65 / 176
BIBLIOGRAPHY
Tourin, Agnes, “An Introduction to Finite Diffference Methods for PDEs in Finance,” in
Nizar Touzi, ed., Optimal Stochastic Target problems, and Backward SDE, Fields Institute
Monographs, Springer, 2011, pp. 201–212.
66 / 176
CHAPTER 4. A MODEL WITH JUMPS
Chapter 4
A Model with Jumps
In Chapter 3, we studied a benchmark model with financial frictions and endoge-

nous risk dynamics. Despite having highly non-linear dynamics, the model is driven
by Brownian shocks and the paths of all variables are continuous, implicitly ruling out
discrete jumps in prices and state variables.
Nevertheless, there is a long tradition of modeling both unanticipated and antici-

pated jumps in macro-finance models. Some notable examples include models of bank
runs (Diamond and Dybvig, 1983), liquidity spirals (Brunnermeier and Pedersen, 2009),
sudden stops (Calvo, 1998; Mendoza, 2010), currency attacks (Obstfeld, 1996; Morris
and Shin, 1998), twin crises (Kaminsky and Reinhart, 1999), and the loss of safe asset
status.
In this chapter, we illustrate how to incorporate self-fulfilling jumps into the model
outlined in Chapter 3, based on a simpified version of Mendo (2020).
4.1 Jump Processes
Previously, we focused on Itô processes in the form
dXt = µtX Xt dt + σtX Xt dZt ,
67 / 176
where the Brownian “shocks” dZt are i.i.d. and small, such that time paths are contin-
uous. In this chapter, we allow for discontinuities by considering a more general class
of processes with i.i.d. increments: Levy processes.
The Levy-Itô decomposition states that any Levy process Lt can be additively de-
composed into three independent components: a linear time drift, a scaled Brownian
motion and a Levy jump process, that is
dLt = adt + bdZt + dJt ,
where a, b are constants, dZt is a Brownian motion and dJt is a Levy jump process.
Processes driven by Levy-noise therefore look formally like Itô processes with an addi-
tional jump component
dXt = µtX Xt dt + σtX Xt dZt + jtX Xt− dJt .
There is a fairly rich class of Levy jump processes dJt . Here we restrict our attention
to Poisson processes.1 Consider a Poisson process with intensity (or arrival rate) λ >
0: Jt takes only values in N0 = {0, 1, 2, . . . } and increments Jt+∆t − Jt are Poisson-
distributed with parameter λ∆t. THe following are some important properties of this
Poisson process:
• Jt is weakly increasing. That is, Jt is either locally constant or has a jump of size 1.
• Conditional on Jt = n, the random time to the next jump, τ := inf{s ≥ 0 | Jt+s >
n}, is exponentially distributed with parameter λ (i.e. the expected time to the
next jump is constant and given by 1/λ).
• the stochastic integral with respect to a Poisson process simply sums the values
of the integrand at the jump times:
Z T JT
0
at dJt = ∑ aτn
n =1
1 Note:this is not as restrictive as it may seem: general Levy jump processes can be written as integral
with respect to Poisson random measures, a generalization of sums of integrals with respect to Poisson
processes.
68 / 176
where τn is the time at which Jt jumps from n − 1 to n.
To capture time-varying macroeconomic dynamics, we will allow for a slightly more

general process, sometimes called a Cox process, where λ does not need to be a con-
stant, but can be time- and state-dependent (λt ).
It is important to note that, while a Brownian motion dZt is a martingale, a jump

process Jt is not — it is expected to drift up. To get a martingale, we have to “compen-
Rt
sate” the jump process by its intensity. In other words, Jt − 0 λs ds is a martingale.2
The following Itô formulae holds for jump diffusions driven by Brownian and Pois-
son noise.
Consider geometric jump diffusions Xt , Yt
dXt dYt
= µtX dt + σtX dZt + jtX dJt , = µYt dt + σtY dZt + jtY dJt ,
Xt − Yt−
where dZt is a standard Brownian motion and dJt is a Poisson process. The follow-
ing results hold.
Itô’s lemma:

′ 1 ′′
d f ( Xt ) = f ( Xt )(µtX Xt ) + X 2
f ( Xt )(σt Xt ) dt+
2
f ′ ( Xt )(σtX Xt )dZt + ( f ((1 + jtX ) Xt− ) − f ( Xt− ))dJt
Itô’s power rule:
γ
d( Xt )
= (γµtX + γ(γ − 1)(σtX )2 )dt + γσtX dZt + ((1 + jtX )γ − 1)dJt .
Xt −
Itô’s product rule:
d( Xt Yt )
= (µtX + µYt + σtX σtY )dt + (σtX + σtY )dZt + ( jtX + jtY + jtX jtY )dJt .
Xt Yt
2 More
Rt
generally, if Xt = 0 as dJs (and a is “predictable”, i.e. at uses information only up to right
Rt
before time t, but does not contain information about potential jumps at time t), then Xt − 0 as λs ds is a
martingale.
69 / 176
Itô’s quotient rule:
d( Xt /Yt ) h i j X − jtY
= µt − µt + σt (σt − σt ) dt + (σtX − σtY )dZt + t
X Y Y Y X
dJt .
Xt /Yt 1 + jtY
Notice that these equations are the same as our earlier rules for geometric Itô pro-
cesses, but with new terms for the jump process. The new terms in the power,
product and quotient rules can be expressed more simply as:
(Xγ )
1 + jt = (1 + jtX )γ
1 + jtXY = (1 + jtX )(1 + jtY )
1 + jtX
1 + jtX/Y =
1 + jtY
4.2 Model Setup
The model setup is exactly as described in Section 3.1. The innovation comes from
our postulated price process, which will now include a jump term. That is, the model
has the same primitives as before, but we now allow for self-fulfilling / sunspot jumps.
4.3 Solution Method
We introduce jumps by postulating that qt follows
dqt q q q
= µt dt + σt dZt + jt dJt .
qt
As before, we can calculate the return rate to capital for both sectors, rti,K (ιit ), using
Itô’s product rule to calculate the capital gains rate. The process is similar to (3.1), but
70 / 176
with a new jump term

" #
ai − ιit q q q q
drti,K (ιit ) = + Φ(ιit ) − δ + µt + σσt dt + (σ + σt )dZt + jt dJt . (3.1’)
qt
Similarly, the return on defaultable debt is
drtD = rti dt + jtD,i dJt .
We then postulate that the SDF (ξ ti = e−ρt u′ (cit )) follows
dξ ti
i
= −rtF,i dt − ςit dZt − νti (dJt − λt dt), (3.2’)
ξt
where rtF,i is the risk-free rate, ςit is the price of Brownian risk, νti is the price of jump risk
and λt dt is the Sunspot arrival rate (which, as discussed above, ensures that the jump
process Jt is a msartingale). Note that in contrast to Chapter 3, the risk-free rate now
depends on i and may vary across agents in the model. This is because the introduction
of jumps means that the debt traded in the model is no longer risk-free.

tions
Optimal investment ι. As before, the choice of investment rate is a static and time-
separable problem. An agent chooses ιit to maximize her return rti,K (ιit ). The first-order
condition yields the Tobin’s q equation
1
= Φ′ (ιit ).
qt

ϕ log(ϕι + 1), ϕιit = qt − 1.
Goods market clearing. The goods market clearing condition is the same as in Chap-
ter 3.
71 / 176
Asset and risk allocation using the martingale approach. To derive the optimal port-
folio choice, we can again use the martingale approach, with a slight modification to
incorporate the newly added jumps.
Martingale approach with jumps.

Consider a portfolio choice problem in continuous time:
∞
Z
−ρt
max E0 e u(ct )dt
{ιet ,θet ,ct }∞
t =0 0
dnt ct
= − dt + ∑ θt drt + labor income/endowment/taxes
j j
s.t.
nt nt j
n0 given.
j
nt is the net worth of the agent and rt denotes the return of asset j. Let xtA be
the value of a self-financing trading strategy A where one reinvests all dividends.
Define the SDF as ξ ti = e−ρt u′ (cit ). As before, ξ t xtA follows a martingale. Let
dxtA
A
= µtA dt + σtA dZt + jtA dJt .
xt
Assume that the SDF follows
dξ ti
i
= −rtF,i dt − ςit dZt − νti (dJt − λt dt)
ξt
Using Itô’s product rule,
d(ξ ti xtA )
i A
=(−rtF,i + µtA − ςit σtA + νti λt )dt + (σ A − ςit )dZt +
ξ t xt
( jtA − νti − νti jtA )dJt
=(−rtF,i + µtA − ςit σtA + λt jtA − λt νti jtA )dt + (σ A − ςit )dZt +
( jtA − νti − νti jtA )(dJt − λt dt).
Where (σ A − ςit )dZt + ( jtA − νti − νti jtA )(dJt − λt dt) is a martingale, given the inclu-
sion of the compensating λt dt term. Then, since ξ t xtA follows a martingale, its drift
72 / 176
equals zero, giving us that
µtA + λt jtA = rtF,i + ςit σtA + λt νti jtA .
Where rtF,i is the (shadow) risk-free rate, ςit is the price of Brownian risk, ς t σtA is the
required Brownian risk premium, λt νti is the price of Poisson upside risk if jtA > 0.
For any two self-financing strategies A, B, the martingale approach implies
µtA − µtB + λt ( jtA − jtB ) = ςit (σtA − σtB ) + λt νti ( jtA − jtB ).
Using the martingale approach on expert capital for each i, we get
ae − ι t q q q q
+ Φ(ι t ) − δ + µt + σσt + λt jt = rtF,i + ςit (σ + σq ) + λνti jt
qt
Summing across types, we can express this in terms of outside equity ownership
" #
ae − ι t χ e χ e
q q q
+ Φ(ι t ) − δ + µt + σσt − et rtF,e + (1 − et )rtF,h + λt jt
qt κt κt
" # " #
χet e χet h χ e χ e
q
= ς + (1 − e )ς t (σ + σq ) + et νte + (1 − et )νth λνti jt
κte t κt κt κt
Similarly, we can derive the household portfolio choice condition by taking the dif-
ference between the drift of household capital and defaultable debt
ah − ιt q q q q
+ Φ(ι t ) − δ + µt + σσt − rt + λt ( jt − jtD ) ≤ ςht (σ + σq ) + νth λt ( jt − jtD ),
qt
This condition holds with equality if κ e < 1 (i.e., if households hold a non-zero amount
of capital).
MA: slides have rtF,i rather than rt . Markus, when we discussed this on Zoom you
asked whether it made sense to derive household capital and expert capital using the
same method rather than these two slightly different approaches.
Asset and risk allocation using the price-taking social planners problem. As in
Chapter 3, we can also solve for the equilibrium risk and asset allocation using the
73 / 176
social planner’s problem. In this environment, we can generalize the price-taking

planner’s theorem to include the choice of jump risk. The price-taking social plan-
ner chooses an allocation of physical capital, κte , Brownian risk, χet and jump risk, ζ te to
solve
 ! ! 
 
∑ ςit χit ∑ t t t t
K q
max − σtr − λ ν i i
ζ j (3.6’)
{κt ,χt ,ζ t }  qt i i
s.t. χet ≥ ακte .
Theorem 4.1. The equilibrium allocation of physical capital, κte , as well as the allocation of
risk, χet , that arises from agents’ portfolio decisions can be more directly obtained by solving the
“price-taking social planner problem” (3.6’).
Proof. The proof takes three steps
1. By Fisher’s Separation Theorem3 , each individual’s portfolio maximization is

equivalent to the following maximization problem of a “firm”. In our model,
individuals in sector i solve4
h i h i
max θti,K Et drti,K (ιit ) /dt + θti,OE Et drOE
t /dt + θti,D Et [drt ]/dt
{θti,K , θti,OE , θti,D }
i,K
h D
i
q
− ςit (θti,K + θti,OE )σtr − λt νti (θti,K + θti,OE ) jt + θ i,D jtr

(3.4) if i = e
s.t. .
(3.5) if i = h
2. Aggregate {ηt }-weighted sum of the two sectors’ problems:5

h i h i
max
{θit }i={e,h}
∑ tt t t t
η i i,K
θ E dr i,K i
( ι ) /dt + ∑ t t t t /dt+
η i i,OE
θ E dr OE
i i
3 We postpone the proof of this result till chapter 7.

4 Recallthat outside equity and capital have the same risk (volatility).
5 Note that σri,K = σr K as the two sectors face the same aggregate risk.
t t
74 / 176
∑ ηti θti,D Et [drt ]/dt − ∑ ςit η i (θti,K + θti,OE )σtr

K
−
i i
h i
∑ t
λ t ν i i
η ( θ i,K
t + θ i,OE q
t ) jt + θ i,D r D
jt ,
i
s.t. (3.4) and (3.5).
3. Market clearing conditions are
Capital: ηti θti,K = κti , ηti (θti,K + θti,OE ) = χit = ζ ti ,

Outside Equity: ∑ ηti θti,OE = 0,
i
Debt: ∑ ηti θti,D = 0.
i
Note that (3.4) together with the capital market clearing condition implies
χet = ηte [θte,K + θte,OE ] ≥ ηte [θte,K − (1 − α)θte,K ] = ακte .
Therefore, the aggregated problem can be simplified to

! !
h i
∑ ∑ ςit χit rK
∑ λt νti ζ ti
q
max κti Et drti,K (ιit ) /dt − σt − jt ,
{κt ,χt ,ζ t } i i i
s.t. χet ≥ ακte ,
which is equivalent to the planner’s problem (3.6’).
q
Since σtr,K = σt + σt and ζ ti = χit , we can set up the Lagrangian as
κte ae + (1 − κte ) ah − ι t q q
L= + Φ(ι t ) − δ + µt + σt σt − (χet ςet + (1 − χet )ςht )(σ + σq )
qt
q
− λt (χet νte + (1 − χet )νth ) jt + m1 (χet − ακte ) + m2 (1 − κte )
75 / 176
The KKT conditions yield two cases: χet = ακte or χet > ακte . For the first case
κte ae + (1 − κte ) ah
q q

= α (ςet − ςht )(σt + σt ) + λt (νte − νth ) jt =⇒ κte = 1,
qt
κte ae + (1 − κte ) ah
q q

> α (ςet − ςht )(σt + σt ) + λt (νte − νth ) jt =⇒ κte < 1.
qt
Similarly, for χet > ακte
ae − ah
> 0 =⇒ κte = 1,
qt
q q
(ςht − ςet )(σ + σt ) + λt (νth − νte ) jt = 0 =⇒ ςht = ςet and νth = νte .
These conditions are almost identical to 3.6, with the added result that the price
of jump risk is equalized across households and experts when the skin in the game
constraint does not bind.
MA: I’ve left out the slide on the invariance of capital demand because I wasn’t sure
how to derive the equation.
MA: In the slides, the change in numeraire approach with jumps is introduced in
this section. I’ve moved it to the next section because it seems to fit there better.
Drift of ηt . We calculate the drift of ηt by changing to the total weath Nt numeraire.

The change of numeraire approach is similar to the case without jumps, with an addi-
tional equation for νt
Change of numeraire with jumps.

Consider two different numeraires – call them dollars ($) and euros (AC). Let xtA
be the value of a self-financing strategy in $. Denote the exchange rate by Yt :
dYt
= µYt dt + σtY dZt + jtY dJt .
Yt
76 / 176
Then, xtA /Yt is the value of the self-financing strategy in A C. Applying the martin-
gale approach, ξ t$ xtA and ξ tAC ( xtA /Y ) are both martingales. For any two assets A, B,
the martingale approach implies
µtA − µtB + λt ( jtA − jtB ) = ς$t (σtA − σtB ) + νt$ λt ( jtA − jtB ),
µtA/Y − µtB/Y + λt ( jtA/Y − jtB/Y ) = ςACt (σtA/Y − σtB/Y ) + νtAC λt ( jtA/Y − jtB/Y ).
By Itô’s quotient rule,
µtA/Y − µtB/Y = (µtA − µtB ) − σtY (σtA − σtB ),

σtA/Y − σtB/Y = (σtA − σtY ) − (σtB − σtY ) = σtA − σtB ,
jtA − jtB
jtA/Y − jtB/Y =
1 + jtY
Hence,
jtA − jtB
(µtA − µtB ) − σtY (σtA − σtB ) = ςACt (σtA − σtB ) + (νtAC λt − λt )( )
1 + jtY
jtA − jtB
(ς$t − σtY )(σtA − σtB ) + (νt$ λt − λt )( jtA − jtB ) = ςACt (σtA − σtB ) + (νtAC λt − λt )( )
1 + jtY
Comparing terms, we get
ςACt = ς$t − σtY and νtAC = νt$ + νt$ jtY − jtY
slides have νtAC = νt$ − νt$ jtY − jtY
As before, we change the numeraire from consumption goods to total wealth. Con-
sider two assets:
• Asset A: sector i’s portfolio return in terms of total wealth, that is Nti /Nt = ηti .
The return to this asset is
!
dηti + (Cti /Nt )dt Cti ηi ηi ηi ηi
= + µt + λt jt dt + σt dZt + jt dJt .
ηti Nti
• Asset B: a benchmark asset that everyone can hold expressed in terms of the total
77 / 176
wealth numeraire. The return to this asset is
rtbm dt + σtbm dZt + jtbm dJt .
The martingale asset pricing formula gives
ηi Cti bm ηi bm i ηi bm i η
i
bm
µt + − r t + λ t ( jt − jt ) = ς̂ (
t t σ − σt ) + ν̂t t − jt ),
( j
Nti
where ς̂it and ν̂ti are the prices of Brownian and jump risk in the total wealth numeraire,
respectively.
Summing across types

′ ′ ′ ′
′ ηi Ct ′ ηi ′ ′ ηi ′ ′ ηi
∑
′
ηti µt +
Nt
− rtbm + ∑ ηti λt jt −λt jtbm = ∑ ηti ς̂it (σt − σtbm ) + ∑ ηti ν̂ti ( jt − jtbm )
i ∈I i′ ∈ I i′ ∈ I i′ ∈ I
| {z } | {z }
=0 =0
Subtracting from the first equation
ηi ηi Ct Ci ′ ′ ηi
′
ηi ′ ′ ηi
′
− ti + ς̂it (ση − σbm ) − ∑ ηti ς̂it (σt − σtbm ) + ν̂ti ( jt − jtbm ) − ∑ ηti ν̂ti ( jt − jtbm )
i
µt + λt jt =
Nt Nt i′ ∈ I i′ ∈ I
i
MA: slides have −ς̂it (ση − σbm )
For experts
ηe ηe Ct Cte e h
µt + λt jt = − e + (1 − ηte )ς̂et (ση − σbm ) − (1 − ηte )ς̂ht (ση − σbm )+
Nt Nt
ηe ηh
(1 − ηte )ν̂te ( jt − jtbm ) − (1 − ηte )ν̂th ( jt − jtbm )
e
MA: slides have −(1 − ηte )ς̂et (ση − σbm ))
jtD − jtN
In this context, the benchmark asset is risky debt, so σtbm = −σtN and jtbm = 1+ jtN
.
The price of Brownian risk in the total wealth numeraire is ς̂it = ςit − σtN and the price
of jump risk is ν̂ti = νti + νti jtN − jtN .
78 / 176
BIBLIOGRAPHY
Volatility of ηti We can calculate the volatility of ηti using Itô’s quotient rule. Since
ηti = Nti /Nt we have
ηi ′ i′ i−
∑
i i i
σt = σtN − σtN = σtN − ηti σtN = (1 − ηti )(σ N − σtN )
′
i ∈I
Jumps in ηti Similarly,
i i ′ i′ i i−
ηi jtN − jtN jtN − ∑i′ ηti jtN (1 − ηti )( jtN − jtN )
jt = = i′
= i−
1 + jtN ′
1 + ∑i′ ηti jtN
i
1 + ηti jtN + (1 − ηti ) jtN
The value function can be derived following the same process as Section 3.2.3.
Bibliography
Brunnermeier, Markus K. and Lasse Heje Pedersen, “Market liquidity and funding
liquidity,” The Review of Financial Studies, 2009, 22 (6), 2201–2238.
Calvo, Guillermo A., “Capital flows and capital-market crises: the simple economics
of sudden stops,” Journal of Applied Economics, 1998, 1 (1), 35–54.
Diamond, Douglas W. and Philip H. Dybvig, “Bank runs, deposit insurance, and liq-
uidity,” Journal of Political Economy, 1983, 91 (3), 401–419.
Kaminsky, Graciela L and Carmen M. Reinhart, “The twin crises: the causes of bank-
ing and balance-of-payments problems,” American Economic Review, 1999, 89 (3), 473–
500.
Mendo, Fernando, “Risky low-volatility environments and the stability paradox,”

Working Paper, 2020.
Mendoza, Enrique G., “Sudden stops, financial crises, and leverage,” American Eco-
nomic Review, 2010, 100 (5), 1941–66.
79 / 176
BIBLIOGRAPHY
Morris, Stephen and Hyun Song Shin, “Unique equilibrium in a model of self-
fulfilling currency attacks,” American Economic Review, 1998, pp. 587–597.
Obstfeld, Maurice, “Models of currency crises with self-fulfilling features,” European

Economic Review, 1996, 40 (3-5), 1037–1047.
80 / 176
Part II
Monetary Models with Aggregate and

Idiosyncratic Risk
81 / 176
CHAPTER 5. A ONE-SECTOR MONETARY MODEL WITH IDIOSYNCRATIC RISK
Chapter 5
A One-Sector Monetary Model with

Idiosyncratic Risk
The vast literature on monetary economics has identified three roles of money, namely,
medium of exchange, store of value, and unit of account. In this chapter, we study
money as a store of value against idiosyncratic risk in a simple one-sector economy.
We investigate the monetary equilibrium and its implications for welfare and policy.
Lastly, we discuss the relation of our model to the fiscal theory of price level and other
roles of money.
In an economy with idiosyncratic risk and incomplete markets, money serves as

a safe asset since it provides absolute nominal insurance. Following Brunnermeier,
Merkel and Sannikov (2020b), a safe asset can be characterized by two distinguishing
features:
• The “good-friend analogy”: A safe asset is like a good friend in that it is there
when one needs it, in times of market stress.
• The “safe-asset tautology”: A safe asset is safe because others perceive it to be

safe. In times of financial distress, agents coordinate on investment in safe assets.
In the simple closed economy presented in this chapter, one can always resort to
money as a store of value in times of high idiosyncratic risk, and importantly, coordi-
nated attacks on money will be ruled out. To focus on the importance of safe assets, we
82 / 176
will define money broadly in this chapter, which includes a wide array of government-
issued bonds and privately issued inside money, as opposed to the narrow definition
based on transaction cost.
The model of this chapter is based on Brunnermeier and Sannikov (2016a, Section
3) and Brunnermeier and Sannikov (2016b).
5.1 Models of Money as a Store of Value
The tradition of modeling money as a store of value dates back to the seminal work
of Samuelson (1958), who illustrates how money as a bubble can facilitate transfers
between generations in an OLG model. Diamond (1965) follows this tradition and in-
troduces capital in Samuelson’s OLG model. Bewley (1980) and Aiyagari (1994) study
money as a store of value in dynamic macroeconomic models with uninsurable idiosyn-
cratic risk. While Bewley and Aiyagari focus on endowment risk, Angeletos (2007)
studies money under the presence of investment risk.
In the models of Diamond (1965) and Aiyagari (1994), market outcome is dynami-
cally inefficient, meaning that the equilibrium interest rate is too low and agents over-
accumulate capital. As we will see in Chapter 7, our model predicts opposite results
where the interest rate is too high and the investment rate is inefficiently low.
5.2 Model Setup
Environment. Unlike before, there is only one sector in the economy, which consists
of a continuum of agents indexed by ĩ ∈ [0, 1]. They each maximize
∞
Z
−ρt
E0 e log(cĩt )dt .
0
Technology. Each household operates a firm, with a CRS production function
yĩt = akĩt .
83 / 176
Households face idiosyncratic risks in their capital accumulation processes. The capital
stock of agent ĩ evolves according to
dkĩt
= (Φ(ιit ) − δ)dt + σ̃t dZ
etĩ .
kĩt
where the idiosyncratic risk follows1
dσ̃t = µ(σ̃t )dt + ν(σ̃t )dZtσ̃ ,
subject to an aggregate risk dZtσ̃ .
Financial Friction. Markets are incomplete in the sense that agents cannot share the
e ĩ .
idiosyncratic risk dZ t
Money. Denote the supply of “narrow money” by Mt , which can serve as a medium
of exchange and facilitate transactions. Nevertheless, we shall ignore the medium of
exchange role for now2 and focus on money as a safe asset. Hence, we adopt the notion
of “broad money” which includes government bonds. For simplicity, we assume all
bonds are zero-coupon and have an interest rate of zero, so they have the same func-
tionality as narrow money. In this chapter (except for section 5.5.3), we use the words
“money” and “bonds” interchangeably. Denote the total supply of broad money by Bt .
That is,
Bt = Mt + government bonds.
Agents hold bonds for consumption smoothing as the value of bonds is only affected
by the aggregate risk. The quantity of bonds is controlled exogenously by a central
bank. We assume that bond supply follows
dBt
= µtB dt + σtB dZtσ̃ ,
Bt
that is, the Monetary Policy is implemented with two instruments µtB , σtB .
1 For
√
example, the idiosyncratic risk σ̃t might follow a CIR process, dσ̃t = α(σ̄ − σ̃t )dt + ν σ̃t dZtσ̃ .
2 It will be incorporated in Section 5.5.3.
84 / 176
We also assume that seignorage is transferred to agents proportionally to their cap-

ital holdings. The following figure shows balance sheets of the agents
A L
A L
A L
A L
Net worth
Money
𝑛$̃
𝑘"$̃
Figure 5.1: Household balance sheets
5.3 Solution Method

R1
The aggregate capital stock is Kt = 0 kĩt dĩ. As we will see in section 5.3.1, invest-
ment rate is equalized across agents. Hence,
dKt
= (Φ(ι t ) − δ)dt.
Kt
where ι t is the investment rate chosen by all agents.
Prices will be denoted in the following way.
• Let the price of the final consumption good be the numeraire. Let Pt denote the
price level (units of money/bond needed to buy one unit of consumption good).
• Let qtK donate the real price of physical capital, so the aggregate capital stock is
worth qtK Kt units of consumption goods.
• Observe that the real price of bonds is the number of consumption goods that
it takes to purchase one unit of bond, i.e., 1/Pt . It will turn out that the price
85 / 176
of bonds will be trend stationary after scaling by Kt /Bt so, for convenience, we
define the scaled price of money as qtB = Bt /( Pt Kt ), or
Bt
qtB Kt = = real value of all bonds.
Pt
Bt
That is, all bonds in the economy is worth Pt := qtB Kt units of consumption goods.
The value of physical capital is thus qtK Kt and the value of nominal bonds (outside
money) is qtB Kt . Then, the fraction of nominal wealth is denoted by ϑt := qtB /(qtK + qtB ).
We postulate that
dqtK qK qK dqtB qB qB
K
= µt dt + σt dZtσ̃ , B
= µt dt + σt dZtσ̃ ,
qt qt
and that SDF (ξ tĩ = e−ρt u′ (cĩt )) follows
dξ tĩ
= −rt dt − ςĩt dZtσ̃ − ς̃ĩt dZ
etĩ , (5.1)
ξ tĩ
where rt is the (individual) risk-free rate3 .
Given the price process and the consumption-investment decision of the agents, we
can calculate the return rate to bonds rtB
d(1/Pt ) d(qtB Kt /Bt )

drtB = = (5.2)
1/Pt qtB Kt /Bt
( )
qB qB qB
= Φ(ι t ) − δ + µt − µtB + (σt − σtB )σtB dt + (σt − σtB )dZtσ̃ .
principle, the drift of individual SDF should be rtĩ . Since the agents can trade bonds freely in a
3 In
competitive market, rtĩ = rt .
86 / 176
Transfers and the return rate to capital. The cumulative seigniorage Tt follows4
( )
1 1 qB
dTt = dBt + d ,B = qtB Kt µtB + (σt − σtB )σtB dt + σtB dZtσ̃ . (5.3)
Pt P t
Since seigniorage is transferred to money holders proportionally to their capital, the

return rate to capital rtĩ,K (ιit ) is
a − ιĩt d(qtK kĩt ) dTt /Kt

drtĩ,K (ιĩt ) = K dt + +
qt qtK kĩt qtK
( )
a − ιĩt q K q B
q B
= K
+ Φ(ιĩt ) − δ + µt + Kt µtB + (σt − σtB )σtB dt
qt qt
!
K q B
q
+ σt + Kt σtB dZtσ̃ + σ̃t dZ̃tĩ . (5.4)
qt
5.3.1 For given SDF processes, derive equilibrium conditions
Agent ĩ’s utility maximization problem is
∞
Z
−ρt
max E0 e log(cĩt )dt
{ιĩt ,θtĩ ,cĩt }∞
t =0
0
dnĩt cĩt
s.t. =− dt + 1 − θtĩ drtĩ,K (ιĩt ) + θtĩ drtB (5.5)
nĩt nĩt
n0ĩ given,
where θtĩ is the fraction of agent ĩ’s wealth in bonds.
4 The value of one unit of money pbt := qtB Kt /Bt = 1/Pt . Consider a discrete-time economy where
government issues money ( Bt+∆t − Bt ) during time [t, t + ∆t] and pays it out at time t + ∆t as transfers.
Total seigniorage follows
Tt+∆t − Tt = pbt+∆t ( Bt+∆t − Bt ) = pbt ( Bt+∆t − Bt ) + ( pbt+∆t − pbt )( Bt+∆t − Bt ).
Taking ∆t → 0, we obtain the continuous-time limit (5.3).
87 / 176
Optimal investment rate. The choice of ιĩt is given by the Tobin’s q equation
1
= Φ′ (ιĩt ).
qtK

ϕ log(ϕι + 1),
ϕιĩt = qtK − 1. (5.6)
Optimal portfolio choice. Recall the martingale approach studied in Chapter 2. For
capital and money, the asset pricing equation is respectively
a − ιĩt qtB

qK qB
h i
Et drtĩ,K (ιĩt )/dt = ĩ B B B
+ Φ(ι t ) − δ + µt + K µt + (σt − σt )σt
qtK qt
!
qK qB
= rt + ςĩt σt + Kt σtB + ς̃ĩt σ̃,
qt

qB qB
h i
B B B B
Et drt /dt = Φ(ι t ) − δ + µt − µt + (σt − σt )σt
qB
= rt + ςĩt (σt − σtB ).
Subtracting the second equation from the first,

!
a − ιĩt σB

qK qB 1 qB qK qB
+ µt − µt + B
µ + (σt − σt )σt = ςĩt
B B
σt − σt + t + ς̃ĩt σ̃.
qt 1 − ϑt t 1 − ϑt
(5.7)
As in previous chapters, under log-utility the prices of risk equal the agent’s risk load-
ings on the respective risk5
!
ĩ qB qK qtB M qB qK
ςĩt = σtn = θtĩ (σt − σtM ) + (1 − θtĩ ) σt + σt = θtĩ σt + (1 − θtĩ )σt
qtK (5.8)
nĩ
ς̃ĩt = σ̃t = (1 − θtĩ )σ̃t .
5 We useσ A and σ̃ A to denote variable A’s risk loading on dZt and dZ

et , respectively. To derive a clear
formula for risk premium ς t , we should consider capital market clearing condition ϑt = θtĩ .
ĩ
88 / 176
Recall the fraction of nominal wealth is denoted by ϑt := qtB /(qtK + qtB ). By Itô’s lemma,
qB qK σtϑ
σt − σt = ,
1 − ϑt
(5.9)
qB qK 1 ϑ q
B
µt − µt = µ + σt (σt − σt ) .
ϑ ϑ
1 − ϑt t
Plugging (5.8) and (5.9) into the asset pricing equation (5.7), we have the optimal port-
folio choice
h i
( 1 − ϑt ) 2 a − ιĩt (1 − ϑt ) µtB − (σtB )2 − µϑt + (σtϑ )2
1 − θtĩ = + (5.10)
σtϑ (σtB − σtϑ ) + (1 − ϑt )2 σ̃t2 qtK σtϑ (σtB − σtϑ ) + (1 − ϑt )2 σ̃t2
Since ιĩt = ι t = (qt − 1)/ϕ is constant, all agents hold the same portfolio in equilibrium.
Optimal consumption. Again, cĩt /nĩt = ρ under log utiliy.
Not applicable to this one-sector model.
Not applicable to this one-sector model.
5.3.4 (Value function iteration and) goods market clearing
Since the optimal portfolio choice is the same for all agents, market clearing re-
quires6
1 − θ t = 1 − ϑt . (5.11)
6 Note that we dropped the individual superscript ĩ.
89 / 176
Also, goods market clearing requires
Ct = ρNt = ρ(qtK + qtB )Kt = ( a − ι t )Kt .
Hence,
qtK + qtB a − ιt a − ιt ρ
ρ K
= K =⇒ K
= . (5.12)
qt qt qt 1 − ϑt
Combining (5.11), (5.12) with the asset pricing equation (5.7) gives us
q
ρ + µtB − (σtB )2 − µϑt + σtϑ σtB
1 − ϑt = . (5.13)
σ̃t
Substituting (5.6) in (5.12), we can solve for qK and ι:
1 + ϕa ( 1 − ϑt ) a − ρ
qtK = (1 − ϑt ) , ιt = . (5.14)
1 − ϑt + ϕρ 1 − ϑt + ϕρ
Finally, q B is obtained by
ϑt 1 + ϕa
qtB = qtK = ϑt . (5.15)
1 − ϑt 1 − ϑt + ϕρ
Stationary Equilibria with constant idiosyncratic risk. Let’s consider a Stationary

Equilibria with a constant idiosyncratic risk σ̃. In a steady state, qK , q B , ϑ are constant,
qK qB qK qB
so µt = µt = σt = σt = 0. Now the fraction of nominal wealth ϑt writes
p
ρ + µ B − ( σ B )2
1−ϑ = . (5.16)
σ̃
Proposition 5.1. For Φ(ι) = log(ϕι + 1)/ϕ, there exists for σ̃ >
p
ρ + µ B − (σ B )2 a sta-
tionary equilibrium, in which qK , q B , ι are given by (5.14) - (5.16). For all σ̃ there is a moneyless
1+ϕa a−ρ
equilibrium, in which q B = 0, qK = 1+ϕρ , and ι = 1+ϕρ .
90 / 176
5.4 Discussions
5.4.1 Endogenously Sticky Price of Risk
We can rewrite equation (5.13) as the following money evaluation equation
µϑt = ρ − (1 − ϑ )2 σ̃t2 + µtB − (σtB )2 + σtϑ σtB . (5.17)
In a steady state without monetary policy where µϑt = σtϑ = µtB = σtB = 0, we have
ĩ √
ς̃ t = σtc = (1 − ϑt )σ̃t = ρ.
√
The price of risk is constant at ρ. Recall equation (3.13) shows that the equilibrium
risk free rate is
2
ĩ ĩ
rt = ρ + µct − σtc .
ĩ
Consequently, the steady-state risk-free rate rt = µct = Φ(ι t ) − δ := g equals the growth
rate of the economy.
As idiosyncratic risk changes, agents’ portfolio decision adjusts to keep the con-
sumption risk and the price of risk constant. Why is this the case? If the price of risk
√
is above ρ, capital stops being attractive and agents demand more money, thereby
√
driving up ϑt and down ς̃ t = (1 − ϑt )σ̃t . Similarly, if the price of risk was below ρ,
money is unattractive and agents demand less money, thereby driving ϑt down and ς̃ t
up. In the long run, the price of risk is sticky (constant), the quantity of risk adjusts to
changes in the level of risk.
5.4.2 Alternative Ways to Derive the Money Evaluation Equation
Above, we obtained the money evaluation equation by applying the martingale

asset pricing condition to capital and bonds. Now we offer two alternative derivations.
First, the equation can be obtained from the asset pricing condition associated with
91 / 176
net worth and money. Remember in Section 3.2.3, we used a self-financing strategy that
reinvests consisting of an agent’s net worth with consumption reinvested. The return
on this strategy is
dnĩt + cĩt dt
drtn,ĩ = = ϑt dRtB + (1 − ϑt )dRtK,ĩ + ρdt
nĩt
The martingale asset pricing condition is then

h i K,ĩ B
(1 − ϑt )Et drtK − drtB /dt + ρ = ςĩt (1 − ϑt )(σtR − σtR ) + ς̃ĩt (1 − ϑt )σ̃t
which also leads to (5.17).
Second, we can apply the method above in the Nt -numeraire (recall Section 3.2.2).
In the Nt -numeraire, the return on the net worth self-financing strategy is
d(nĩt /Nt ) + (cĩt /Nt )dt

drtn,ĩ = = (1 − ϑt )σ̃t dZ̃tĩ + ρdt.
nĩt /Nt
The last equality is due to the fact that in equilibrium, nĩt and Nt have the same drift
and exposure to aggregate risk. In the Nt -numeraire, the return on money is
d(qtB Kt /Bt Nt ) d(ϑt /Bt ) h i

drtB = = = ( µ ϑ
t − µ B
t ) + σt
B B
( σt − σt
ϑ
) dt + (σtϑ − σtB )dZt .
qtB Kt /Bt Nt ϑt /Bt
The martingale asset pricing condition is

h i
ρ − (µϑt − µtB ) + σtB (σtB − σtϑ ) = ςĩt (σtB − σtϑ ) + ς̃ĩt (1 − ϑt )σ̃t .
Equation (3.9) implies that in the Nt -numeraire, ςĩt = 0 and ς̃ĩt = (1 − θtĩ )σ̃t 7 , so this is
another (and arguably easier) way to also arrive at (5.17).
7 As a general principle, ςĩt and ς̃ĩt can always be derived as volatility of the SDF. Since we know ςĩt
and ς̃ĩt in consumption numeraire, it is straightforward to calculate their values in the Nt -numeraire.
92 / 176
5.4.3 Bubbles, Retrading, and Transversality
Government bonds in this model are purely bubbles in the sense that they are not
backed by any primary surpluses or transaction services. The bonds are valued because
they provide (idiosyncratic-)risk-free store of value. Bonds allow for self-insurance
through retrading in response to idiosyncratic shocks. For example, when agent ĩ is
hit by a negative capital shock dZtĩ < 0, she can sell bonds and buy capital to mitigate
the decline in networth (dZtĩ < 0 =⇒ nĩt ↓ =⇒ θtĩ nĩt ↓), and vice versa if dZtĩ > 0.
Such retrading reduces the volatility of nĩt and allows the agent to better insure against
idiosyncratic risk.
Typical macroeconomic models invoke transversality conditions (TVC) to eliminiate

bubbles in government bonds. In this model, the individual TVC for bonds is
" #
h i 1
lim Et ξ ĩT θ Tĩ nĩT = 0 ⇐⇒ lim Et e−ρt θ Tĩ nĩT = lim e−ρt Et [ϑT ] = 0,
T →∞ T →∞ ĩ
ρn T T →∞
which holds trivially. However, more importantly, the (steady-state) valuation of all
government bonds is
RT Z RT
Et e − t rs ds
θt nĩT dĩ = Et e − t rs ds B
qT KT = e − r ( T − t ) q B K T = e ( g − r ) T q B K0
where r, g are the steady-state risk-free rate and economic growth rate, respectively.
We see that this expression does not converge to zero if r ≤ g, which is satisfied in the
steady state as we have shown in section 5.4.1.
What is the difference between individual TVC and social valuation of bonds? The
answer lies in the discount rate used in these calculations. The effective discount rate in
the individual TVC is the discount rate for stochastic bond portfolio θtĩ nĩt , which equals
the risk-free rate plus the risk premium of government bonds, with the latter being
positive due to retrading opportunities. By contrast, he discount rate for aggregate
bond stock has to equal the risk-free rate in the long-run, which is smaller than the
individual discount rate.
93 / 176
5.4.4 Comparison Static w.r.t. σ̃
As σ̃ increases, the store-of-value role of money becomes more important. As a

result, the value of money relative to capital rises (∂(q B /qK )/∂σ̃ > 0). Another inter-
pretation is that when σ̃t is large, there is more retrading of bonds, resulting in a larger
bubble component in the value of bonds. Furthermore, the total value of capital and
money qK + q B = (1 + ϕa)/( σ̃1 ρ + µ B − (σ B )2 + ϕρ) is also increasing in σ̃. Hence, an
p
increase in idiosyncratic risk results in a higher total net worth of the economy. Through
a decline in the value of capital qK , a higher σ̃ boosts consumption while lowers the in-
vestment rate ι and the growth rate g := Φ(ι) − δ. A higher µ M has opposite effects as
inflation makes money less desirable as a store of value.
Figure 5.2: q B , qK as functions of σ̃. Parameter values: a = 0.1, ρ = 0.05, ϕ = 10, µ B =

0.05, σ B = 0.05.
5.4.5 Impossibility of “as if” Representative Agent Economies
Note that it is impossible to construct an “as if” representative agent economy with
the same aggregate output and investment streams and same asset prices that mimics
the equilibrium outcome of our heterogeneous agents economy. In any representative
94 / 176
agent economy, absence of individual-level idiosyncratic risk, capital returns strictly

dominate money and hence money could never have some positive value.
5.5 Connections to Monetary Theory
5.5.1 Interpreting the Money Evaluation Equation
To understand the money evaluation equation (5.17), we can rewrite it in an inte-

geral form.
µϑt = ρ − (1 − ϑt )2 σ̃t2 + µtB − (σtB )2 + σtϑ σtB ,
It follows from Itô’s lemma that

d(e−ρt ϑt ) = e−ρt −ρ + µϑt ϑt dt + e−ρt σtϑ ϑt dZtσ̃
Z T Z T
−ρT −ρt −ρs
⇒ e ϑ T − e ϑt = e −ρ + µs ϑs ds +
ϑ
e−ρs σsϑ ϑs dZsσ̃
t t
After taking expectations, rearranging and taking the limit as T → ∞, we get
∞
Z
−ρ(s−t)
ϑt = E t e ρ − µϑs ϑs ds
t
∞
Z
−ρ(s−t)
= Et e (1 − ϑs )2 σ̃s2 − µsB − σsB (σsϑ − σsB ) ϑs ds
t
This equation illustrates the forces that determine ϑt . First, an increase in the future
path of idiosyncratic risk will lead to an increase in ϑt . This is because higher idiosyn-
cratic risk leads to greater precautionary saving, which increases the portfolio share in
bonds. Second, ϑt is larger when future µtB is lower. This is because the government
is printing less money and so the distortion that comes from redistributing seignorage
revenue is lower. Third, ϑt is affected by the uncertainty in monetary policy σtB , while
the effect depends on the difference between σsB and σsϑ .
95 / 176
5.5.2 Relation to The Fiscal Theory of Price Level
Our analysis is related to the fiscal theory of price level (FTPL). It also emphasizes
the store-of-value role of money and hence money is defined in a broad sense which
includes cash and government bonds. While in our theory, money has value due to a
bubble component, in the FTPL the value is derived purely from the ability to pay taxes
with money.
In a setting in which money supply growth has fiscal implications, our analysis
resembles the FTPL with difference that an additional bubble component is added to
the standard FTPL continuous time equation8 :
∞
Z
Bt ξs ξ T BT
qtB Kt = = Et ( Ts − Gs ) ds + lim Et , (5.18)
Pt t ξt T →∞ ξ t PT
| {z } | {z }
PV of primary surpluses Bubble
where ( Ts − Gs ) refers to the flow of primary surpluses (tax revenue minus govern-
ment expenditures except interest payments) which is zero in our setup. The last term
denotes bubble — the new element resulting from our analysis. When evaluating all
government debts, the SDF declines at a speed of the risk free rate: dξ t /ξ t = −rt dt. We
have shown in section 5.4.3 that a bubble can exist as long as r ≤ g in the steady state.
Assume that Tt − Gt = τaKt and K0 = 1, we can rewrite the steadt-state FTPL

equation as
Z ∞
B0 B0
= e−(r− g)t sdt + lim e−(r− g)T .
P0 T →∞ P0
|0 {z }
=:PVS0,∞
The government can potential “mine” the bubble to relieve its fiscal burdens. The equa-
tion above implies three more general cases (allowing st ̸= 0):
1. s > 0, µ B < 0 : then r > g, PVS0,∞ > 0 and a bubble cannot exist. This is the
"conventional" situation considered in the literature.
8 See Brunnermeier et al. (2020a) for derivations. An additional “convenience yield” term appears on
the right hand side of this equation if money also provides transaction services as in section 5.5.3.
96 / 176
2. s = µ B = 0 : then r = g, PVS0,∞ = 0 and there is a finite positive bubble whose

value exactly equals q B K0 and grows over time at the growth rate/risk-free rate.
3. s < 0, µ B > 0 : then r < g and thus the integral PVS0,∞ = −∞. Yet, q B is still pos-
itive, which is only possible if there is an offsetting infinite positive bubble. These
infinite values do not violate any no-arbitrage condition and are also not other-
wise economically problematic, since the bubble cannot be traded separately from
the claim on surpluses. Both are necessarily bundled in the form of government
Bt
bonds. As long as Pt = q B Kt is determined and finite in equilibrium, the model
remains economically and mathematically sensible despite the infinite values in
the decomposition of the value of government bonds.
In all three cases, the (possible) presence of a bubble grants the government some extra
leeway. Clearly in case 3, the government can run a perpetual deficit, "mine the bubble"
and never has to raise taxes to fully fund all government expenditures. In case 2, the
existence of the bubble is beneficial, because the value of government debt is positive -
allowing agents to share part of their idiosyncratic risk - despite the fact that the present
value of primary surpluses is zero. Even in case 1, government debt is more sustainable
since an unexpected drop of primary surpluses to zero results in a bubble instead of a
total collapse of the value of debt.
Second, the fact that there are multiple equilibria (monetary vs. moneyless) means
that price level cannot be pinned down by (5.18) alone. Indeed, equation (5.14) shows
B0 1 + ϕa
= q0K K0 = (1 − ϑ0 ) .
P0 1 − ϑ0 + ϕρ
The existence of multiple possible paths for {ϑt } means the initial price level P0 is un-
determined. However, by adding off-equilibrium tax backing to money, we can easily
eliminate the “moneyless equilibrium.” Brunnermeier, Merkel and Sannikov (2020a)
show that any sequence of {ϑt } that satisfies the money evaluation (5.17) consists of an
equilibrium, and all equilibria
√ B except for the monetary one implies ϑt → 0. If we fix
ρ+µ −(σ B )2
an arbitrary threshold σ̃ > ϑ > 0 and, whenever ϑt < ϑ, make the fiscal
authority switch to a constant tax rate rule s = τa > 0, then the agents expect that in
97 / 176
equilibrium ϑt can never fall below ϑ so all equilibria but the monetary steady could be
ruled out.
5.5.3 Adding the Medium of Exchange Role of Money
Money (in the narrow sense) has the advantage that it serve as a medium of ex-
change and thereby helps to overcome the double-coincident of wants problem that
pure barter economies face. Note that granting money a medium of exchange role is
another way to eliminate the “moneyless equilibrium.” A simple way to model this
double-coincidence of wants friction is by introducing a transaction cost in the produc-
tion technology (Merkel, 2020). That is, capital produces output

yĩt = 1 − C(vĩt ) + Tt akĩt ,
1 ϵ
where the transaction cost C(v) = ϵv (v/v) is increasing the velocity of household’s
money holdings vĩt := Pt akĩt /mĩt . The transaction cost serves as a pure decision wedge
and is rebated to the households via Tt = C(v∗t ), where v∗t is the equilibrium velocity
chosen by all households.
To not confuse ourselves with transaction cost rebate Tt and seigniorage Tt , we as-
sume Mt ≡ M in this subsection so that Tt = 0. Also, assume that government bonds
are in zero net supply so that Bt = Mt . The modified problem of each household is
∞
Z
−ρt
max E0 e log cĩt dt
{ιĩt ,θtĩ ,cĩt }∞
t =0
0
dnĩt cĩt
s.t. =− dt + 1 − θtĩ drtĩ,K (ιĩt , vĩt ) + θtĩ drtM
nĩt nĩt
n0ĩ given.
For simplicity, assume that the government conducts no monetary policy (µtM = σtM =
98 / 176
0), so
 
1 − C(vĩt ) + C(v∗t ) a − ιĩt qK  qK
drtĩ,K (ιĩt , vĩt ) =  + Φ(ιĩt ) − δ + µt  dt + σt dZtσ̃ + σ̃t dZ̃tĩ ,

qtK

qB qB
drtM = Φ(ι t ) − δ + µt dt + σt dZtσ̃ .
We can rewrite the definition of velocity as
a qtK kĩt a 1 − θtĩ

vĩt = = .
qtK mĩt /Pt qtK θtĩ
As ϵ → ∞, the transaction cost converges to a cash-in-advance constraint vĩt ≤ v̄. Equiv-

alently,
!
a a
K
+ v̄ θtĩ ≥ K .
qt qt
The parameter v̄ > 0 captures the velocity of money holdings and hence governs the
tightness of the CIA constraint.
With the CIA constraint, it is easier to use the HJB approach. For each household,
its value function V (nĩ ; σ̃ ) depends on an idiosyncratic state nĩ and an aggregate state
σ̃. Recall that
 ! 
1−C(vĩt )+C(v∗ ) a−ιĩt
 
cĩt qK qB
dn ĩ /nĩ − + (1 − θtĩ ) qtK
+ Φ(ιĩt ) − δ + µt + θtĩ Φ(ι∗t ) − δ + µt 
 t t  =  nĩt

 dt
dσ̃t  
µ(σ̃t )
  
qK qB
(1 − θtĩ )σt + θtĩ σt ĩ σ̃
(1 − θt )σ̃t  dZt 
+
ν(σ̃t ) 0 dZ̃tĩ
The associated HJB equation is then9

(
ρV = max log c + (∂n V )nµn + (∂σ̃ V )µ(σ̃)
c,θ,ι
9 In the following, we suppress the ĩ superscript for simplicity.
99 / 176
)
1 1
+ (∂nn V )n2 [(σn )2 + (σ̃n )2 ] + (∂σ̃σ̃ V )ν(σ̃)2 + (∂nσ̃ V )nσn ν(σ̃)
2 2
!
a a
s.t. K
+ v̄ θ ≥ K ,
q q
where we define
!
1 − C (v ) + C (v∗ ) a − ι

c K

∗ qB

µ n = − + (1 − θ ) + Φ ( ι ) − δ + µ q
+ θ Φ ( ι ) − δ + µ
n qK
K B
σn = (1 − θ )σq + θσq
σ̃n = (1 − θ )σ̃
The first-order conditions w.r.t. θ is

h
h
K M
i (∂nn V )n qK qB
i
qK qB
E dr − dr /dt = − (1 − θ ) σ + θσ (σ − σ ) + (1 − θ )σ̃ 2
∂n V
!
∂nσ̃ V K B a
− ν(σ̃)(σq − σq ) + λ + v̄
∂n V qK
λ denotes the (time-varying) Lagrangian multiplier on the CIA constraint. In case of

log utility, it can be shown that
1
V (n; σ̃ ) = log(n) + Γ(σ̃).
ρ
Hence,
( ) !
h i
qK qB qK qB a
Et drtK − drtM /dt = (1 − θt )σt + θt σt (σt − σt ) + (1 − θt )σ̃t2 + λt + v̄ .
qtK
This asset pricing equation can be rewritten as the following money evaluation equa-
tion
h i
2
µϑt = ρ − ( 1 − ϑt ) σ̃t2 + λ t vt . (5.19)
100 / 176
Using (5.14) and the market clearing condition θt = ϑt , we can express vt as
1 1 − ϑt + ϕρ
vt = . (5.20)
1 + ϕa ϑt
Equations (5.19) and (5.20) jointly characterize ϑt and vt . In general, we can divide the
equilibrium into two mutually exclusive regimes10 :
• The “I theory regime” where vt < v̄ and λt = 0
• The “monetarist regime” where vt = v̄ and λt ≥ 0
5.6 Exercises
5.6.1 Transfers and the Super-neutrality of Money
What would happen if seignorage is transferred to agents proportionally to their

bond holdings? What if seignorage is transferred to agents proportionally to their net
worth?
Hint: start by modifying return equations (5.2) and (5.4).
5.6.2 Equilibrium Multiplicity in the One-sector Money Model
Consider the benchmark model of this chapter, in which money is a pure bubble. We
had assumed a steady state equilibrium (for q B and qK ) and not admitted the possibility
of changing prices over time. The aim of this problem is to characterize all equilibria of
the model with deterministic price drift.11 To simplify the analysis, assume that agents
have log utility, that there is no aggregate risk (σ = 0) and that there is a fixed money
supply (µ M = σ B = 0). For this problem you may assume (σ̃ )2 is high enough so
money has positive value.
10 See Merkel (2020) for details.

11 There are even more equilibria, if one allows for stochastic price changes.
101 / 176
(a) Find all steady state equilibria, that is equilibria with dqtM = dqtK = 0.
Note: you will find two equilibria here12 , which differ – among other things – in
the price of capital qK . In the following, qK and qK (with qK < qK ) refer to the qK
values in these two steady state equilibria.
(b) Now consider more generally all deterministically drifting equilibria, that is pos-
tulate the price processes
dqtM qB dqtK qK
= µ t dt, = µ t dt. (5.21)
qtM qtK
The goal is to characterize all equilibria with such price paths. Proceed as follows:
1. Use the equilibrium conditions to derive an ODE that describes the evolu-
tion of qK , q B or ϑ (choose one of the three) over time in any equilibrium
satisfying (5.21). The right-hand side of the ODE (in explicit form) should
only contain model parameters and qtK (or qtM or ϑt ), but no other endoge-
nous variables/unkowns (such as the other two of the three variables qK ,
p M or ϑ).
B K
Hint: you need to find a relationship between µq , µq and µϑ or a subset
of these variables; Ito’s formula/time differentiation may be useful here.
2. For each of qK and qK found in part 1, choose at least three initial conditions
qtK0 around those steady states and solve the ODE of part (a) numerically.
Plot the resulting time paths for {qtK }tT=t0 and for {qtM }tT=t0 . What are your
observations? Write down whatever patterns you observe, but in particu-
lar with regard to the following two questions:
i. Can all paths represent an actual equilibrium?

ii. What can happen in the long run? Under which conditions and to
where do paths converge/diverge?
For your numerical experiments use the parameters a = 1, ρ = 0.01, σ̃ =

0.4, ϕ → ∞, take the initial time t0 = 0 and plot at least T = 50 time
12 There is a third one that satisfies all equilibrium equations. But it violates the free disposal condition
qK ≥ 0 and can therefore be ruled out.
102 / 176
periods.
3. Based on your observations from part (b), formulate a proposition and

prove it. The proposition should make an exact characterization of the set
of price pairs (q B , qK ) that can be observed in at least one equilibrium at
at least one point in time. Conditional on observing a pair (qtM0 , qtK0 ) in this
set, the proposition should make an assertion about the long-run behavior,
limt→∞ (qtM , qtK ).
4. In which equilibria does money have a positive value? In which equilibria

does it have a positive value even asymptotically (that is limt→∞ qtM > 0)?
(c) Tax backing of money and equilibrium uniqueness
Now consider a small variation of the setting before. Suppose there is a gov-
ernment that taxes capital income/production. k t units of capital still produce
an output flow ak t dt, but only (1 − τ ) ak t dt are kept by the owner of the capital,
whereas τak t dt is collected by the government. Aggregate tax revenues are thus
τaK̄dt. Suppose the government uses all tax revenues to back the value of money
by either of the following two strategies
(i) it pays a real dividend on money holdings, that is each household receives
a fiscal transfer of consumption goods proportional to its money holdings
(ii) it buys back money and destroys it to shrink the money supply13
Assume that the government keeps a balanced budget at all times. Assume fur-
ther that q B and qK follow generic deterministic processes as in (5.21).
1. Derive the processes for capital gains and dividend yields under both poli-
cies. Show that both policies generate the same return process for money.
2. Modify your solution to part 2 (a) to reflect the new return process for
money.14 Derive again an ODE that describes the evolution of qK (or q B or
ϑ) over time. Plot dqK /dt as a function of qK for τ = 0 and some small
13 This is obviously equivalent to assuming that the government allows households to pay taxes with
money instead of real goods and just disposes the collected tax revenues.
14 Of course, also the return on capital changes due to the tax.
103 / 176
BIBLIOGRAPHY
values τ > 0. As in part 2 (b), use a = 1, ρ = 0.01, σ̃ = 0.4, ϕ → ∞, in

addition let τ ∈ {0.1, 0.2, 0.4} be the “small” positive tax rates.15
3. Comparative statics: show that as τ is marginally increased at τ = 0, the

qK steady state is reduced and the qK steady state is increased. What does
this mean for the associated p values?
Hint: do not actually solve for the steady states in the case τ > 0 by solving
the equation dqK /dt = 0 for qK , because this will be a third order polyno-
mial equation. Instead use an argument based on the implicity function
theorem.
4. Show that (for arbitrarily small taxes τ > 0) the tax backing eliminates all
equilibria other than the steady state in which money has positive value.
Hint: show that all other potential equilibria eventually drift to a point that
violates one of the free disposal conditions qtM , qtK ≥ 0.
5. Suppose a government wants to achieve equilibrium uniqueness, but does

not want to distort the capital allocation by taxing capital.16,17 The only
policy tool of the government is still the capital income tax, but suppose the
government can credibly commit to a state-contingent tax rate τ (q B , qK ) as
a function of q B and qK . Is it possible to design a tax policy that eliminates
all but one equilibrium and satisfies τ = 0 along the (unique) equilibrium
path? Give an example or an impossibility argument.
Bibliography
Aiyagari, S Rao, “Uninsured idiosyncratic risk and aggregate saving,” The Quarterly
Journal of Economics, 1994, 109 (3), 659–684.
15 Not exactly small tax rates, but we want to see something in the plot.
16 Potentially, the government even wants to print money and/or subsidize capital for the same rea-
sons as in Brunnermeier and Sannikov (2016b) and Di Tella (2019).
17 Alternatively, this desire can also come from outside the model by our preferences as model
builders: we may not want the complication of a tax, but want to make an argument why our equi-
librium is not only the limit of a sequence of unique equilibria in richer models, but is actually exactly
the unique equilibrium of a slightly richer model.
104 / 176
BIBLIOGRAPHY
Angeletos, George-Marios, “Uninsured idiosyncratic investment risk and aggregate

saving,” Review of Economic dynamics, 2007, 10 (1), 1–30.
Bewley, Truman, “The optimum quantity of money,” Models of monetary economies,

1980, pp. 169–210.
Brunnermeier, Markus K and Yuliy Sannikov, “The I theory of money,” Technical

Report, National Bureau of Economic Research 2016.
and , “On the optimal inflation rate,” American Economic Review Papers and Proceed-
ings, 2016, 106 (5), 484–89.
, Sebastian A. Merkel, and Yuliy Sannikov, “The Fiscal Theory of Price Level with
a Bubble,” Technical Report, National Bureau of Economic Research 2020.
, , and , “A Safe Asset Perspective for an Integrated Policy Framework,” Working

Paper, 2020.
Di Tella, Sebastian, “Risk Premia and the Real Effects of Money,” Working Paper, 2019.
Diamond, Peter A, “National debt in a neoclassical growth model,” The American Eco-
nomic Review, 1965, 55 (5), 1126–1150.
Merkel, Sebastian A., “The Macro Implications of Narrow Banking: Financial Stability
versus Growth,” Working Paper, 2020.
Samuelson, Paul A, “An exact consumption-loan model of interest with or without the
social contrivance of money,” Journal of political economy, 1958, 66 (6), 467–482.
105 / 176
CHAPTER 6. CASH VERSUS CASHLESS ECONOMY AND THE I THEORY OF MONEY
Chapter 6
Cash versus Cashless Economy and the

I Theory of Money
In this chapter, we extend the simple model in Chapter 5 to a two-sector setting with
financial frictions, which allows us to explore the role of intermediaries in monetary
transmissions. This Chapter is adapted from Brunnermeier and Sannikov (2016).
Main tools we will use in this Chapter:
• Using the total wealth numeraire to simplify derivations
• The “benchmark asset” approach to portfolio problems
• Various technics learned in Chapter 3
Economic insights from this Chapter:
• Intermediaries as experts of risk managements
• The existence of money has important implications when markets are incomplete
• Inflation risk can complete markets and improve risk sharing
• Monetary policy conditional on shocks can lead to better risk sharing
• The “Paradox of Prudence”
106 / 176
6.1 Cash versus Cashless Economy
6.1.1 A Model with an Intermediary Sector
Households. There is a continuum of households of measure one. Denote the house-

hold sector by h. Each household (h, ĩ ), ĩ ∈ [0, 1] operates a firm that has a CRS produc-
tion technology
yth,ĩ = ah k h,
t .
ĩ
The capital stock of each household evolves according to
dk h, ĩ h i
t
= Φ(ιh,
t
ĩ
) − δ dt + σdZt + σ̃dZ̃tĩ
k h,
t
ĩ
The households face both aggregate risk (dZt ) and idiosyncratic risk (dZ̃tĩ ) in the capital
accumulation process. Incomplete markets make it impossible for the households to
insure each other against the idiosyncratic risk. Nevertheless, they can reduce their
exposure to idiosyncratic risk through the intermediary sector.
Finally, the households have logarithmic utility
∞
Z
E0 e −ρt
log cth,ĩ dt .
0
Intermediaries. There is a continuum of identical intermediaries. Denote the inter-

mediary sector by I. The intermediaries are not involved in production. Instead, they
have competitive advantage in risk management. Each intermediary holds shares of
all individual firms and diversifies the idiosyncratic risk of each firm to φσ̃dZ̃tĩ where
0 < φ < 1. The following figure shows the balance sheets of intermediaries and house-
holds in this economy.
107 / 176
A Firms 𝚤̃ L
A … L
A L A Households 𝚤…
̃ L
Intermediaries A L
A
Inside equity
A L L
Risky claims
Out-Money
Capital Inside
𝑞! 𝐾! equity
Net worth
Risky Debt
claims (real vs. nominal) Money
Debt
(real vs. nominal)
…
"𝜒! "
…
Net worth
Figure 6.1: Balance sheets
The intermediaries take deposits from the households (real or nominal) and invest
in shares of household-operated firms (risky claims). The intermediaries also hold out-
side money issued by the central bank. The intermediaries have limited capacity of
risky claim issuance. The fraction of risk held by intermediaries χtI cannot exceed a
certain threshold χ̄ ≤ 1.
The intermediaries also have logarithmic utility
∞
Z
E0 e −ρt
log ctI,ĩ dt .
0
Monetary Policy. The intermediaries also hold government-issued outside money

(narrow money and government debts). The quantity of outside money follows
dBt
= µtB dt + σtB dZt .
Bt
In this chapter, we use the notion “outside money” instead of “money” or “bond” as in
Chapter 5 as a reminder that intermediaries can create "inside money" (privately issued
debts).
Frictions. There are two frictions in this economy
• The intermediaries can only issue a limited amount of risky claims (χt ≤ χ̄).
108 / 176
• The intermediary can only issue debts to the households (incomplete markets).
Specifically, we will study and contrast two situations
1. Nominal debt issuance (monetary economy)
2. Real debt issuance only (cashless economy)
An equivalent expert-household setting. Note that the model with intermediaries is

equivalent to the following expert-household model that resembles the ones in Chap-
ter 2 and 3. Consider a Basak-Cuoco economy with experts and households where
both groups have the same productivity (ae = ah ), but experts are better at managing
idiosyncratic risk (σ̃e < σ̃ h ).
A Experts L
Out-Money
A Households L
Debt 𝜅!# 𝑞! 𝐾!
(real vs. nominal)
Capital Net worth
𝜅!" 𝑞! 𝐾!
Debt
(real vs. nominal)
Net worth
Figure 6.2: An equivalent model
The experts hold κte fraction of capital and the households hold κth = 1 − κte fraction
of capital. Since ae = ah , allocation of capital has no effect on aggregate productivity,
which is the same as in the model with intermediaries where all capital is owned by
households. It can also be shown that any risk allocation induced by a (κte , κth ) pair can
be replicated in the model with intermediaries by the correct choice of χtI . The skin-
in-the-game constraint on the experts corresponds to the constraint on intermediaries
χtI ≤ χ̄.
109 / 176
6.1.2 Solution Method
As in Chapter 5, let qtK be the real price of capital and qtB = Bt /Pt Kt be the normal-
ized price of outside money. Then the real value of total capital stock is qtK Kt and the
real value of total outside money is qtB Kt . Define the total wealth of the economy as
Nt = (qtK + qtB )Kt and the fraction of nominal weath as ϑt = qtB /(qtK + qtB ).
To study the portfolio problems of intermediaries and households, we first state the
utility maximization problem of agent (i, ĩ ) as
∞
Z
max E0 e −ρt
log ci,t ĩ dt
{ιi,t ĩ ,θti,ĩ ,ci,t ĩ }∞
t =0
0
dni,t ĩ ci,t ĩ
s.t. =− dt + (1 − θti,ĩ )drtK (ιi,t ĩ ) + θti,ĩ drtbm
ni,t ĩ ni,t ĩ
n0i,ĩ given.
Here rtbm is the return on a benchmark asset, which corresponds to either nominal or
real debt in this model. Similar to Chapter 5, the return on capital is1
 
 a − ιi,ĩ qtB

qK qB
drtK (ιi,t ĩ ) = t
+ Φ(ιi,t ĩ ) − δ + µt + µtB + (σt − σtB )σtB dt
 qtK qtK 
!
qK qB h i
+ σ + σt + Kt σtB dZt + 1{i= I } φ + 1{i=h} σ̃dZ̃t . (6.1)
qt
The intermediaries are also subject to the following credit constraint:
(1 − θtI )ηtI
χtI = ≤ χ̄,
1 − ϑt
where ηtI = NtI /Nt is the wealth share of the intermediary sector. To understand this
equation, note that (1 − θtI )ηtI = (1 − ϑt )χtI are two expressions of the same variable —
1 Here we omit the index on the idiosyncratic Brownian motion dZ̃t . This is because the interme-
diaries hold claims of all individual firms, and technicially, their exposure should be ĩ σ̃dZ̃tĩ . We have
R
assumed that the intermediaries can only diversify away a part of the idiosyncratic risk (e.g., Z̃tĩ are
correlated with each other), so ĩ σ̃dZ̃tĩ = φσ̃dZ̃t where Z̃t is a single Brownian motion.
R
110 / 176
the intermediaries’ risky claims as a fraction of the total wealth in the economy.
Equations (5.14) and (5.15) still hold. That is, given ϑt
1 + ϕa 1 + ϕa ( 1 − ϑt ) a − ρ
qtK = (1 − ϑt ) , qtB = ϑt , ιt = . (6.2)
1 − ϑt + ϕρ 1 − ϑt + ϕρ 1 − ϑt + ϕρ
Throughout the rest of this chapter, we will proceed in the Nt -numeraire as we once
did in Section 3.2.2. We postulate that
dϑt
= µϑt dt + σtϑ dZt , [in the Nt -numeraire]
ϑt
dξ th,ĩ
= −rth,ĩ dt − ςh, ĩ h,ĩ ĩ
t dZt − ς̃ t d Z̃t . [in the Nt -numeraire]
ξ th,ĩ
The drift of individual SDFs no longer equal to the risk-free rate. This is because by
Itô’s quotient rule,
ξ h,ĩ /N
µt = −rt − µtN + σtN (σtN + ςh, ĩ
t ) ̸ = −r t . [in consumption numeraire]
Ideally, we would use different symbols for variables denoted in the Nt -numeraire.
However, that would make the notations overly cumbersome, and, as a result, making
this chapter unreadable. Hence, we will continue to use the same notations but keep
in mind that all variables are denominated in the Nt -numeraire hereafter.
In equilibrium, all households have the same SDF, so
rth,ĩ = rth , ςh, ĩ h

t = ςt , ς̃h, ĩ h
t = ς̃ t .
Note ϑt is unchanged under the new numeraire. To solve for risk allocation between
households and intermediaries and ϑt , we once again utilize the price-taking social
planner’s problem from Section 3.2.1. In this model, it corresponds to
h i h i h i
max Et drtN /dt − ς tI χtI + ςht χth σtxK − ς̃ tI φχtI + ς̃ht χth σ̃, s.t. χtI + χth = 1,
χtI ≤χ̄
where drtN is the return of the total wealth in the economy and σtxK is the excess risk
111 / 176
of capital relaive to the benchmark asset. The definition of benchmark asset varys de-
pending on the specific model (e.g., nominal debt vs. real debt). In the Nt -numeraire,
the total wealth is constant at Nt /Nt = 1, so drtN = 0. The first-order condition is then
ς tI σtxK + ς̃ tI φσ̃ ≤ ςht σtxK + ς̃ht σ̃. (6.3)
The FOC holds with equality if intermediaries’ credit constraint is loose (χt < χ̄). Anal-
ogous to Section 3.2.2, the evolution of sectoral wealth shares can be obtained from
martingale asset pricing equations on agents’ portfolio return. In the Nt -numeraire, the
net worth of agent (i, ĩ ), i ∈ { I, h}, ĩ ∈ [0, 1], ni,t ĩ , becomes
ni,t ĩ ni,ĩ Ni
= t i × t := η̃ti,ĩ × ηti ,
Nt Nt Nt |{z} |{z}
within-sector wealth share sector i’s wealth share
R1
where Nti = 0 ni,t ĩ dĩ is the total net worth of sector i. Again, consider two assets
• Asset A: agent (i, ĩ )’s portfolio return in terms of total wealth, that is ni,t ĩ /Nt =
η̃ti,ĩ ηti 2 . The return to this asset is
d(η̃ti,ĩ ηti ) + (ci,t ĩ /Nt )dt dηti dη̃ti,ĩ

= + + ρdt.
η̃ti,ĩ ηti ηi η̃ti,ĩ
Remember that this model is scale-invariant, so all agents within the same sector
make the same consumption-portfolio decisions. Hence, η̃ti,ĩ only moves because
h i
of individual agents’ exposure to idiosyncratic risk, which implies Et dη̃ti,ĩ = 0.
As before, we use the following notation3
dηti ηi ηi dη̃ti,ĩ η̃ i
= µ t dt + σt dZt , = σ̃t dZ̃tĩ . (6.4)
ηi η̃ti,ĩ
2 Note that variables that are ratios of two other quantities (e.g., η̃ti,ĩ , ηti ) are numeraire invariant.
3 In N -numeraore, individual net worth ni,ĩ evloves according to
t t
d(η̃ti,ĩ ηti ) ηi ηi η̃ i
= µt dt + σt dZt + σ̃t dZ̃tĩ .
η̃ti,ĩ ηti
112 / 176
• Asset B: a benchmark asset (bm) that everyone can hold. In this chapter, asset bm
is either real debt or money in terms of total wealth. Note that both these assets
are free of idiosyncratic risk. Denote the return on bm by drtbm = µbm bm
t dt + σt dZt r.
The martingale asset pricing equation is then

ηi ηi η̃ i
µt + ρ − µbm
t = ςit σt − σtbm + ς̃it σ̃t . (6.5)
Take ηti -weighted sum across the two sectors (i ∈ { I, h})

ηi ηi η̃ i
∑ η i µt + ρ − µbm
t =∑ η i ςit σt − σtbm + ∑ η i ς̃it σ̃t . (6.6)
i i i
ηi ηi
Since ∑i ηti = 1, ∑i η i µt = ∑i η i σt = 0. Also, since ηth = 1 − ηtI ,
ηh ηtI ηI
σt = − I
σt .
1 − ηt
As in equation (5.8), log utility implies
ηI ηh ηtI ηI η̃ I η̃ h
ς tI = σt , ςht = σt = − σ , ς̃ tI = σ̃t , ς̃ht = σ̃t . (6.7)
1 − ηtI t
The asset pricing equations can thus be simplified to the following “benchmark asset
evaluation equation”
2 2
ηi η̃ i
ρ − µbm
t = ∑η i
σt +∑η i
σ̃t . (6.8)
i i
| {z }
excess return on Nt | {z }
net worth weighted risk premium
Subtracting (6.6) from (6.5), we get the drift of ηti as

ηi ηi −i η̃ i −i
µt = (1 − ηti ) ςit (σt − σtbm ) − ς− i η
t ( σt − σtbm ) + ς̃it σ̃t − ς̃− i η̃
t σ̃t . (6.9)
where −i = { I, h}\i denotes the other sector in the economy. Since each agent’s port-
folio contains a linear combination between the benchmark asset and capital, her net
113 / 176
worth (in the Nt -numeraire) η̃ti,ĩ ηti has a loading on the aggregate risk of
ηi K
σt = θti σtbm + (1 − θti )σtr = σtbm + (1 − θti )σtxK , (6.10)
χit
where 1 − θti = (1 − ϑt ) is the portfolio weight on capital of agents in sector i. Sim-
ηti
ilarly, since the benchmark asset is idiosyncratic-risk-free, η̃ti,ĩ ηti ’s exposure to idiosyn-
cratic risk depends on the portfolio weight on capital,

(1 − θ i ) φσ̃ if i = I
η̃ i t
σ̃t = . (6.11)
(1 − θ i )σ̃ if i = h
t
6.1.3 The Cashless Economy
In this subsection, we consider a cashless economy where outside money has no

value and intermediaries can only issue real debts. This is exactly the Basak-Cuoco
economy in Chapter 2 with an additional idiosyncratic risk component. Equation (6.2)
tells us that in consumption numeriare
1 + ϕa a−ρ
qtB = ϑt = 0, qtK = , ιt = .
1 + ϕρ 1 + ϕρ
In this economy, the benchmark asset bm is the risk-free real debts issued by interme-
diaries. In the Nt -numeraire, σtbm = −σtN = −σ. Therefore, from (6.1) we know that
!
rK qK qB
σtxK = σt − σtbm = σ + σt + Kt σtB − σtN − (−σtN ) = σ.
qt
qK
Further, notice that since σt = 0,
K qK qtB B
σtr = σ + σt + K
σt − σtN = 0.
qt
114 / 176
Equations (6.9) and (6.10) then imply

 !2 !2 
ηI χtI − ηtI (χtI − ηtI )2 2 χtI 1 − χtI
µt = σ + (1 − ηtI )  φ2 −  σ̃2 , (6.12)
ηtI ηtI (1 − ηtI ) ηtI 1 − ηtI
ηI χtI − ηtI
σt = θtI σtbm = σ, (6.13)
ηtI
so the aggregate risk is not perfectly shared as long as χit ̸= ηti . However, investment
rate ι t and capital price qtK are constant (in consumption numeraire). This is because
the changes in ηtI induced by aggregate shocks are fully absorbed by movements in the
risk-free rate, which can be computed from (6.8).
To conclude the model, we can plug (6.11) and (6.13) into the FOC to the planner’s
problem (6.3) to solve for χit as a function of the state ηtI :
( )
η I (σt2 + σ̃2 )
χtI = min , χ̄ .
σ2 + [(1 − ηtI ) φ2 + ηtI ]σ̃2
The evolution of ηtI is governed by the SDE consisted of (6.12) and (6.13).
6.1.4 The Monetary Economy
In this subsection, we consider a monetary equilibrium where outside money is

positively valued and the intermediaries issue nominal debts (inside money) to the
households. In this setting, the benchmark asset bm is the nominal money (inside and
outside). In the Nt -numeraire, the risk of capital and money is
K qK qtB B
σtr = σ + σt + σt − σtN ,
qtK
qB
σtbm = σ + σt − σtB − σtN
115 / 176
The second equation is derived from the fact that in consumption numeraire drtB =
d(1/Pt ) d(qtB Kt /Bt )
= . Hence, (5.9) implies
1/Pt qtB Kt /Bt
!
qK qB qB σtB − σtϑ
σtxK = σt − σt + 1 + Kt σtB = .
qt 1 − ϑt
qtB qB qK qK
Because Nt = (qtB + qtK )Kt , Itô’s lemma implies σtN = qt +qtK
B (σ + σt ) + qB +t qK (σ + σt )4 .
t t
We can rewrite the risk of money as
" #
qB qB qB qK qK
σtbm = σ + σt − σtB − B t K (σ + σt ) + B t K (σ + σt )
qt + qt qt + qt
= σtϑ − σtB .
Equation (6.10) then implies

!
ηI χtI
σt = σtbm + (1 − θtI )σtxK = − 1 (σtB − σtϑ ).
ηtI
ϑ′ (ηtI ) I η I ηI
Postulate that ϑt = ϑ (ηtI ). By Itô’s lemma, σtϑ = I
ηt σt . Solving for ηtI σt , we get
ϑ ( ηt )
ηI χtI − ηtI
ηtI σt = σtB .
χtI − ηtI ϑ′ (ηtI )ηtI
1+
ηtI ϑ (ηtI )
Both sectors’ balance sheets are perfectly hedged against aggregate risk if σtB = 0. This
qB
result is achieved via inflation risk (σt ), which effectively completes the market for
aggregate risk. As such, a strict inflation targeting might not always be desirable.
ηI
Again, we can plug (6.11) and σt = 0 into the FOC to the planner’s problem (6.3)
to solve for χit as a function of the state ηtI :
( )
ηtI
χtI = min , χ̄ . (6.14)
(1 − ηtI ) φ2 + ηtI
4 The risk of aggregate capital stock is simply σtK = σ.
116 / 176
d(qtB Kt /Bt Nt ) d(ϑt /Bt )

In the Nt -numeraire, the return to money is drtbm = = . So
qtB Kt /Bt Nt ϑt /Bt
µbm
t = µϑt − µtB and (6.8) becomes the following money evaluation equation (σtB = 0)
 !2 !2 
χtI 1 − χtI
ρ − µϑt + µtB = (1 − ϑt )2  φ2 ηtI + (1 − ηtI )  σ̃2 (6.15)
ηtI 1 − ηtI
The drift of ηtI is given by (6.9)

 !2 !2 
ηI χtI 1 − χtI
µt = (1 − ηtI )(1 − ϑt )2  φ2 −  σ̃2 . (6.16)
ηtI 1 − ηtI
To solve the model, postulate that ϑt = ϑ (ηtI , t). By Itô’s lemma,
ηI
ϑt µϑt = ∂t ϑ + (∂η ϑ )ηtI µt .
Plugging in (6.15) and (6.16), we get the following PDE

  !2 !2  
 χtI 1 − χtI

ϑt ρ + µtB − (1 − ϑt )2  φ2 ηtI − (1 − ηtI )  σ̃2 =
 ηtI 1 − ηtI

 !2 !2 
χtI I
1 − χt  2
∂t ϑt + (∂η ϑt )ηtI (1 − ηtI )(1 − ϑt )2  I
φ2 − σ̃ ,
ηt 1 − ηtI
where χtI is given by (6.14). We can start with a terminal condition ϑ (η I , T ) and solve
this PDE backwards in time as in Section 3.2.4.
6.1.5 Numerical Results and Discussions
Numerical Code. Following is main code executing the algorithm. We use the PDE
solver payoff_policy_growth.m in Section 3.2.4
1 % solves chashless vs . monetary model in lecture 6
2 % required script : payoff_policy_growth .m
3 close all ; clear ; clc
4
5 %% parameters
6 a = 0.15; % productivity
117 / 176
7 rho = 0.03; % decay rate

8 sigma = 0.2; % aggregate volatility
9 sigmaIdio = 0.3; % indio volatility
10 phi = 2; % investment function parameter
11 chibar = 0.5; % risk claims upperbound of intermediaries
12 varphi = 2/3; % diversification ability of intermediaries for idio risk
13 delta = 0.03; % decay rate
14 tol = 1e -3; % tolerance
15 muB = 0.01; % monetary policy : constant
16 sigmaB = 0; % monetary policy : constant
17 lambda = 0.5; % lambda = dt /(1+ dt ) , dt is the time step
18
19 %% Grid
20 etaLength = 300;
21 eta = ( linspace (0.001 , chibar , etaLength )) '; % wealth share of intermediaries
22 S = zeros ( etaLength ,1) ;
23 G = zeros ( etaLength ,1) ;
24
25 %% ------------ Cashless Economy -- closed form solution ------------ %%
26 % cashless -- vartheta = 0
27 varthetaCashless = 0;
28 % real price of capital
29 qKCashless = (1+ phi * a) /(1+ phi * rho )* ones ( etaLength ,1) ;
30 % normalized price of outside money
31 qBCashless = zeros ( etaLength ,1) ;
32 % Investment
33 iotaCashless = (a - rho ) /(1+ phi * rho ) * ones ( etaLength ,1) ;
34 % risk share of intermediaries
35 chiCashless = min ( eta *( sigma ^2+ sigmaIdio ^2) ...
36 ./( sigma ^2 + ((1 - eta ) * varphi ^2 + eta )* sigmaIdio ^2) , chibar );
37 % drift of etaI
38 muEtaCashless = ( chiCashless - eta ) .^3* sigma ^2./ eta .^2./(1 - eta ) ...
39 + (1 - eta )* sigmaIdio ^2 ...
40 .*(( chiCashless ./ eta ) .^2* varphi ^2 - ((1 - chiCashless ) ./(1 - eta )) .^2) ;
41 % volatility of etaI
42 sigmaEtaCashless = ( chiCashless - eta ) ./ eta * sigma ;
43 % Risk free rate
44 PhiCashless = log ( qKCashless )/ phi ;
45 riskFreeRateCashless = rho + PhiCashless - delta - chiCashless ./ eta *( sigma ^2+ sigmaIdio
^2) ;
46
47 %% ------------ Monetary Economy -- iterative method ------------ %%
48 % risk share of intermediaries
49 chiMonetary = min ( eta ./((1 - eta )* varphi ^2 + eta ) , chibar );
50 % initial vartheta : start from steady state
51 varthetaSteadyStateMonetary = 1 - sqrt ( rho )/ sigmaIdio / varphi ;
52 varthetaInitMonetary = varthetaSteadyStateMonetary * ones ( length ( eta ) ,1) ;
53 varthetaMonetary = varthetaInitMonetary ;
54
55 for i =1:1500
56 % 1. compute updated coefficients
57 muEtaMonetary = (1 - eta ) .* (1 - varthetaMonetary ) .^2 * sigmaIdio ^2 ...
118 / 176
58 .*(( chiMonetary ./ eta ) .^2* varphi ^2 - ((1 - chiMonetary ) ./(1 - eta )) .^2) ;
59 muVarthetaMonetary = rho + muB - (1 - varthetaMonetary ) .^2 * sigmaIdio ^2 ...
60 .*( varphi ^2* chiMonetary .^2./ eta + (1 - chiMonetary ) .^2./(1 - eta )); % money
valuation equation
61
62 % 2. PDE time step , call payoff_policy_growth .m function
63 MU = muEtaMonetary .* eta ;
64 newVarthetaMonetary = payoff_policy_growth ( eta , muVarthetaMonetary , MU , S , G ,
varthetaMonetary , lambda );
65
66 % 3. check convergence
67 absChangeVartheta = abs ( newVarthetaMonetary - varthetaMonetary )/ lambda *(1 - lambda ) ;
68 relChangeVartheta = absChangeVartheta ./( abs ( newVarthetaMonetary )+ abs ( varthetaMonetary
)) *2;
69 maxRelChange = max ( relChangeVartheta );
70 if maxRelChange < tol
71 break ;
72 end
73
74 % 4. update vartheta
75 varthetaMonetary = newVarthetaMonetary ;
76 end
77 % real price of capital
78 qKMonetary = (1 - varthetaMonetary ) .*(1 + phi *a ) ./(1 - varthetaMonetary + phi * rho ) ;
79 % normalized price of outside money
80 qBMonetary = varthetaMonetary .*(1 + phi * a) ./(1 - varthetaMonetary + phi * rho );
81 % Investment
82 iotaMonetary = ((1 - varthetaMonetary )*a - rho ) ./(1 - varthetaMonetary + phi * rho );
83 % drift of etaI
84 muEtaMonetary = (1 - eta ) .* (1 - varthetaMonetary ) .^2 * sigmaIdio ^2 ...
85 .*(( chiMonetary ./ eta ) .^2* varphi ^2 - ((1 - chiMonetary ) ./(1 - eta )) .^2) ;
86 % volatility of etaI
87 sigmaEtaMonetary = zeros ( etaLength );
88 % Find risk free rate
89 PhiMonetary = log ( qKMonetary )/ phi ;
90 muVarthetaMonetary = rho + muB - (1 - varthetaMonetary ) .^2 * sigmaIdio ^2 ...
91 .*( varphi ^2* chiMonetary .^2./ eta + (1 - chiMonetary ) .^2./(1 - eta ));
92 muP = (1./ varthetaMonetary + 1./(1 - varthetaMonetary + phi * rho ) ) .* muVarthetaMonetary .*
varthetaMonetary ;
93 riskFreeRateMonetary = PhiMonetary - delta + muP - sigma ^2;
Numerical Results The equilibrium for the cashless economy and monetary is as fol-
lows
119 / 176
1.5 0.8 0.6
q K (solid) / q B (dashed)
0.5
0.6
1 0.4
I
0.4 0.3
0.5 0.2
0.2
0.1
0 0 0
0 0.2 0.4 0.6 0 0.2 0.4 0.6 0 0.2 0.4 0.6
I I I
0.12 0.04
0.06
0.11 0.05
0.02
0.04
0.1
I
I
0 0.03
I
I
0.09
0.02
-0.02
0.08
0.01
0.07 -0.04 0
0 0.2 0.4 0.6 0 0.2 0.4 0.6 0 0.2 0.4 0.6
I I I
Figure 6.3: Cashless Economy (blue) vs. Monetary Economy (red)
Parameter Value Parameter Value

ρ 0.05 σ 0.2
a 0.15 σ̃ 0.5
ϕ 2 φ 0.4
δ 0.03 ϕ 2
σB 0 µB 0.01 (monetary economy)
Table 6.1: Parameters for Baseline Model.
An Economy with Both Real and Nominal Debts. Including real debts in the mon-
etary economy does not alter the equilibrium at all5 . This is because the markets are
already complete with respect to aggregate risk if there exist nominal debts. Adding
5 This result relies on the absence of price stickiness.
120 / 176
real debts will not change aggregate risk sharing and can not affect idiosyncratic risk
sharing.
The Role of Financial Frictions. Note that if χ̄ = 1 (i.e, the intermediaries are not con-
strained), then both models (cashless vs. monetary) converge to the one-sector model
in Chapter 5 in the long run. This is because the intermediaries will manage all capital
stock and the wealth distribution between sectors have no impact on risk sharing.
Discontinuity at the Cashless Limit. In this model, the absolute quantity of money Bt
does not affect the equilibrium (although µtB does). However, as we move from Bt > 0
(monetary economy) to Bt = 0 (cashless economy), equilibrium dynamics jump discon-
tinuously (as illustrated in previous sections). This result is seemingly contradictory to
the standard DSGE literature (e.g., Woodford, 2003). The reason behind this discrep-
ancy is that traditionary monetary models consider money as a medium of exchange,
whereas in our model money is a store of value.
6.2 The I Theory of Money
6.2.1 Model Setup
In Section 6.1.4, the intermediary sector is perfectly hedged against aggregate risk
η̃ h
and the idiosyncratic risk that households bear (σ̃t ) is constant over time. These two
results might seem unrealistic in practice, and in order to break them, we need the
intermediaries’ exposure to aggregate risk to differ from the aggregate risk of the econ-
omy. One way of modeling differentiated aggregate risk exposure is to introduce two
production technologies that have different levels aggregate risk.
Specifically, assume each individual firm has two technologies a and b. The two
technologies are Leontieff in the sense that their relative capital share is fixed. In other
words, each firm allocates (1 − ψ̄) fraction of its capital to technology a and ψ̄ fraction
121 / 176
to technology b. The capital accumulation processes for both technologies are
dk h, ĩ h i
a: t
= Φ(ιh,
t
ĩ
) − δ dt + σ a dZt + σ̃dZ̃tĩ ,
k h,
t
ĩ
dk h, ĩ h i
b: t
= Φ(ιh,
t
ĩ
) − δ dt + σb dZt + σ̃dZ̃tĩ .
k h,
t
ĩ
where σ := σb − σ a > 0. Further, assume that intermediaries can only hold risky claims
of technology b. The balance sheets of intermediaries and households are as follows
A Firms 𝚤…
̃ L
A L
A L A Households 𝚤…
̃ L
Intermediaries A L
A L A L
Out-Money Tech. 𝑎 Equity
Inside
Inside Equity
claims
Tech. 𝑏 Net worth

Risky
Nominal
Risky Money
Nominal
claims Debts
Debts
…
…
Net worth
Figure 6.4: Balance sheets
To be clear, capital ownership in the economy is
Tech. a Tech. b held by intermediaries Tech. b held by households Total

1 − ψ̄ χtI ψ̄ − χtI 1
The price-taking planner’s problem in Nt -numeraire is
a
h i b
h i
min (1 − ψ̄)ςht σtxK + ς tI χtI + ςht (ψ̄ − χtI ) σtxK + ς̃ tI φχtI + ς̃ht (1 − χtI ) σ̃
χt ≤χ̄
where
a σtϑ − σ B b σtϑ − σ B
σtxK = −ψ̄σ − , σtxK = (1 − ψ̄)σ − .
1 − ϑt 1 − ϑt
122 / 176
The FOC is (which holds with equality if χtI < χ̄)
b b
ς tI σtxK + ς̃ tI φσ̃ ≤ ςht σtxK + ς̃ht σ̃,
i.e.,
ηI
" # " # " #
ηI σtϑ − σtB χtI −ηtI σt σtϑ − σtB
σt (1 − ψ̄)σ − + (1 − ϑt ) I φσ̃ φσ̃ ≤ (1 − ψ̄)σ −
1 − ϑt ηt 1 − ηtI 1 − ϑt
" #
1 − χtI
+ ( 1 − ϑt ) σ̃ σ̃ (6.17)
1 − ηtI
The rest of the model is the same as before. Since the return on money is
d(qtB Kt /Bt Nt ) d(ϑt /Bt ) h i

drtbm = B
= ϑ B
= (µt − µt ) + σt (σt − σt ) dt + (σtϑ − σtB ) dZt .
B ϑ B
qt Kt /Bt Nt ϑt /Bt | {z } | {z }
bm σt
µbm
t
Equation (6.8) implies

2 2
ηi η̃ i
h i
ρ− (µϑt − µtB ) + σtB (σtϑ − σtB ) = ∑η i
σt +∑η i
σ̃t (6.18)
i i
The drift of ηtI is still given by (6.9), (6.7) and (6.11),

ηI ηI ηh η̃ I η̃ h
µt = (1 − ηtI ) ς tI (σt − σtbm ) − ςht (σt − σtbm ) + ς̃ tI σ̃t − ς̃ht σ̃t ,
  2 
2 2 ηI 2
ηtI σt 

 ηI η̃ I η̃ h ηI
= (1 − ηtI )  σt + σ̃t − − σ̃t B
 − σt (σt − σt ).
ϑ
(6.19)
 
1 − ηtI
The volatility of ηtI is given by (6.10)
ηI b
σt = σtbm + (1 − θtI )σtxK (6.20)
" #
χtI σtϑ − σtB
= σtϑ − σtB + I (1 − ϑt ) (1 − ψ̄)σ − .
ηt 1 − ϑt
123 / 176
ϑ′ (ηtI )ηtI η I
Postulate that ϑt = ϑ (ηtI ). By Itô’s lemma, σtϑ = ϑ (ηtI ) t
σ , so
ηI (1 − ϑt ) χtI (1 − ψ̄)σ + (χtI − ηtI )σtB

ηtI σt = ! . (6.21)
χtI − ηtI −ϑ′ (ηtI )ηtI
1−
ηtI ϑ (ηtI )
So the intermediaries are not perfectly hedged under σtB = 0. Amplification results
from the changes in the price of money relative to capital, ϑ (ηtI ). As long as the inter-
mediaries’ portfolio share of households’ equity (1 − θtI ) is greater than (1 − ϑt ), the
overall capital share (so that χtI > ηtI ), and as long as ϑ′ (η I ) < 0, amplification exists.
qB qK −ϑ′ (ηtI )ηtI
Note that since σϑ = (1 − ϑt )(σt − σt )., we can further decompose as
ϑ (ηtI )
!
−ϑ′ (ηtI )ηtI (qtK )′ (ηtI )ηtI −(qtB )′ (ηtI )ηtI
= ( 1 − ϑt ) + .
ϑ (ηtI ) qtK (ηtI ) qtB (ηtI )
Amplification arises from two spirals: changes in the price of capital qtK , i.e. the
liquidity spiral, and changes in the value of money qtB , the disinflationary spiral. In the
region where intermediaries are undercapitalized (i.e. ηtI is low), negative shocks are
amplified both on the asset sides of intermediary balance sheets, as the price of physi-
cal capital qK (ηtI ) drops following a negative shock, and on the liability sided, through
the Fisher disinflationary spiral, as the value of money q B (ηtI ) rises. Both effects im-
pair the intermediaries’ net worth. Intermediaries’ response to these losses is to shrink
their balance sheets, leading to fire-sales (lowering the price qtK ) and reduction in inside
money (increasing the value of liabilities qtB ). In other words, intermediaries take fewer
deposits, create less inside money, and the money multiplier collapses.6 This again re-
duces their net worth, and so on. The “Paradox of Prudence” emerges. Each individual
intermediary micro-prudent behavior to scale back his risk is macro-imprudent, as it
raises endogenous risk.
Specifically, this feedback effects lead to a geometric series, which has been summed
up in Equation (6.21). Amplification becomes greater as ϑ′ (ηtI ) becomes more negative,
6 In reality, rather than turning savers away, financial intermediaries might still issue demand de-
posits and simply park the proceeds with the central bank as excess reserves.
124 / 176
and as intermediary leverage (1 − θtI ) rises.
Numerical Algorithm. Postulate that ϑt = ϑ (t, ηtI ). By Itô’s lemma,

" 2 #
ηI 1 ηI I ηI
dϑt = ∂t ϑ + (∂η ϑ)µt ηtI + (∂ηη ϑ) σt ηt dt + (∂η ϑ )σt ηtI dZt
2 | {z }
| {z } σtϑ ϑt
µϑt ϑt
Plugging in (6.18), we have

 !
2
∂t ϑ ∂η ϑ 1 ∂ηη ϑηI ηI ∂η ϑ ηI
ρ− + µt ηtI + σt ηtI − µtB + σtB σ η I − σtB 
ϑt ϑt 2 ϑt ϑt t t
2 2
ηi η̃ i
= ∑ η σt
i
+ ∑ η σ̃t
i
i i
In a "PDE" form:
 " 
2 2 # 
ηi η̃ i

ϑ ρ + µtB + (σtB )2 − ∑ ηi σt + ∑ η i σ̃t
i i
 
2
ηI ηI 1 ηI
= ∂t ϑ + (∂η ϑ)ηtI µt + σt σtB + (∂ηη ϑ) σt ηtI
2
ηI ηI ηh η̃ I η̃ h
where µt , σt , σt , σ̃t , σ̃t are given by (6.19), (6.21), (6.7), (6.11). Further, χt can be
obtained as a function of ηtI by solving the planner’s FOC (6.17).
Numerical Analysis. Consider parameter values ρ = 0.05, a = 0.5 σ a = 0.1, σb = 0.2,

σ̃ = 0.6, ϕ = 2, φ = 0.5, ψ̄ = 0.5 and χ̄ = 0.3.
Figure 6.5: Equilibrium allocations of risk χtI .
We start by looking at the allocation of risk. When ηtI drops, the risk premia that
intermediaries demand for equity stakes in b-firms rise, to the point that the households
may be willing to sell less than fraction χ̄ of outside equity.
125 / 176
Figure 6.6: Equilibrium prices of capital qtK and money qtB .
Figure above shows the prices qK (η ) and q B (η ) of capital and money in equilibrium.
At ηtI = 0, the values of qtK and qtB converge to those under the benchmark without
intermediaries (Chapter 5). As ηtI rises, the price of capital rises and the price of money
drops. Money becomes less valuable as ηtI rises mainly because intermediaries create
money, the inside money on the liabilities sides of the intermediaries’ balance sheets is
a perfect substitute to outside money, and intermediaries’ diversify way idiosyncratic
risk away, which lowers the demand for money.
Figure below illustrates the equilibrium dynamics through the drift and volatility
of the state variable ηtI .
Figure 6.7: Equilibrium dynamics.
The drift of ηtI captures the relative risk premia that intermediaries and households
earn on their portfolios relative to money. As intermediaries become undercapitalized,
the price of and return from holding claims on b-firms rises, leading intermediaries to
take on more risk. The opposite happens when intermediaries are overcapitalized - risk
premia decline and the households’ rate of earnings rises. The stochastic steady state
of ηtI is the point where the drift of ηtI equals zero - at that point the earnings rates of
intermediaries and households balance each other out.
The left panel also shows the volatility terms: The dashed red curve shows the fun-
damental portion of the volatility of ηtI . The solid curve depicts the total volatility that
includes the effects of amplification due to the liquidity and the disinflationary spiral.
Amplification becomes prominent when intermediaries are undercapitalized. While
the left panel illustrates dynamics for our baseline parameters, the right panel reduces
fundamental risk parameters to σ a = 0.03 and σb = 0.06. The right panel illustrates the
volatility paradox: endogenous risk persists due to amplification even as fundamen-
tal risk declines. We see that the maximal volatility of ηtI below the steady state stays
126 / 176
roughly constant as fundamental risk declines, i.e. amplification in this model can be
very large.
6.2.2 The Transmission of Shocks
To understand how shocks propagate in this economy, we decompose the transmis-

sion of a negative shock to capital stock in the following steps.
Step 1 Shock Onset
Suppose a nagetive shock dZt < 0 destroys some capital σ̃t Kt dZt in the economy.
Recall the evolution of net worth for each agent is
dni,t ĩ
= −ρdt + (1 − θti,ĩ )drtK (ιi,t ĩ ) + θti,ĩ drtbm .
ni,t ĩ
Since the intermediaries are levered ((1 − θtI ) > 1), the shock reduces their net worth
by a larger percentage than it does to the value of assets owned by intermediaries. As
a result, the leverage of intermediaries rises after a negative shock.
In the meantime, since the households are nominally insured ((1 − θth ) < 1), the
negative shock hits them less than the intermediaries in the sense that the households’
net worth shrinks by a smaller fraction.
Step 2 Deleveraging
As the intermediaries’ net worth declines, they have to reduce the size of their bal-
ance sheets by fire-selling their risky claims. If prices of capital and money are fixed,
pure deleveraging has no real impact.
127 / 176
Step 3 The Liquidity Spiral
As the net worth of both groups falls after the shock, the total demand for capital
contracts. Hence, deleveraging results in lower price of capital and lower investment
rate. Drops in the capital price also hurt intermediaries’ balance sheets (as dRtK depends
qK
on µt ), leading to further deleveraging — which we call the “liquidity spiral”.
Step 4 The Disinflation Spiral
The liability side of the intermediaries’ balance sheets is also affected by shocks.
On the one hand, shrinked intermediary balance sheets mean that less inside money
is being created and total money supply decreases, and therefore, the value of outside
money appreciates (qtB ↑, or disinflation).
On the other hand, intermediaries’ ability to diversify idiosyncratic risk is impaired

after deleveraging, so the households have to bear more risk. The households in turn
demand more safe assets, and increased money supply leads to further disinflation.
6.3 Policy
6.3.1 Fiscal Policy
The government can impose a flow of transfers/taxes on individual agents dτti,ĩ ,

subject to a budget constraint
Z
∑ ĩ
dτti,ĩ = dTt ,
i
where Tt is the government’s seigniorage income defined by (5.3). In the models of

Chapter 5 and 6, we have assumed that the transfers are made to agents proportional
to their capital holdings (see equation 5.4). That is, dτti,ĩ = (ki,t ĩ /Kt )dTt . The govern-
ment could also make the transfers proportional to bond holdings (dτti,ĩ = (ni,t ĩ (1 −
θti,ĩ )/Bt )dTt ) or to net worth (dτti,ĩ = (ni,t ĩ /Nt )dTt ). It can be shown that if transfers are
128 / 176
made to bond holders, then policy based on µtB , σtB has no real impact. If transfers are
based on net worth, the effects of fiscal policy are between the two other cases7 .
Intra-temporal Transfer Policy. If the government can instead choose τti,ĩ in an un-
constrained way (while respecting its budget constraint), it can essentially complete
the markets for idiosyncratic risk and achieve first-best outcomes.
If the government is constrained to make only sector-specific transfers (τti,ĩ = τti , ∀ĩ),
it can effectively control ηti by shifting resources between sectors. This constraint on
the government can be micro-founded by agents’ hidden savings (e.g., Di Tella and
Sannikov, 2016)
Inter-temporal Transfer Policy. Now we focus on bond supply policies (µtB , σtB ). We
focus on the case where seigniorage is rebated to capital holders.
• A policy that uses µtB and transfers (σtB = 0) affect only the drift of ϑt (see eq. 6.18)
• A policy that uses σtB and transfers (µtB = 0) affects the risk of money and nominal
bonds and therefore affects agents’ portfolio choice.
Through Chapter 5 and 6, bonds are a bubble and backed by agents’ demand for nomi-
nal insurance. Hence, all transfers/seigniorage come from bubble mining (i.e., through
bond supply policies, see Section 5.5.2). However, if the government also imposes
taxes, a part of the bond price (qtB ) and its fluctuations will be induced by future taxes
(see eq. 5.18). Thus, the government can generate transfers through changes of future
taxes, which we refer to as “inter-temporal” fiscal policy.
6.3.2 Monetary Policy.
Interest Rate Policy. To study monetary policy, we need to first introduce interests on
bond/money. Previously, bonds pay no interest rates and the fluctuations of its value
solely from inflation (1/Pt ). If we instead assume that bonds pay a nominal rate of it ,
7 Please work out Exercise 5.6.1
129 / 176
equation (5.2) becomes
d(1/Pt ) d(q B Kt /Bt )

drtB = it dt + = it dt + Bt
1/Pt qt Kt /Bt
( )
qB B qB B B qB
= it + Φ(ι t ) − δ + µt − µt + (σt − σt )σt dt + (σt − σtB )dZtσ̃ .
To study monetary policy without fiscal implications, we let σtB = 0, so

qB qB
drtB = it − µtB + Φ(ι t ) − δ + µt dt + σt dZtσ̃ .
For now, we shut off transfers to capital holders by setting τti,ĩ = 0. If interest payments
on bonds are entirely funded by drift in Bt , the government budget constraint implies
it = µtB , so changes in interest rate only affect inflation and have on real impact. Note
that this is a special case of transferring seigniorage to money holders, so variations in
µtB (it ) are neutral.
Long-term Government Bonds. Now we introduce a consol bond, which has face
value FtL and nominal price PtL . The consol bond never matures but pays an interest
rate itL . In this subsection, we denote outside money by Mt (whose nominal price is 1)
and Bt = Mt + PtL FtL . With two bonds, the government budget constraint is
dMt + PtL dFtL = it Mt dt + itL FtL dt.
Define the fraction of bond values that are not in short-term reserves as
PtL FtL
ϑtL = .
Bt
Postulate that the price of a single long-term consol bond follows
dPtL L L
L
= µtP dt + σtP dZt .
Pt
130 / 176
Assume that only intermediaries find it worthwhile to hold consol bonds. The martin-
gale pricing condition is that
h i L η
Et drtL − drtM /dt = σtP σt .
Hence,
L L
drtL = drtM + σtP σt dt + σtP dZt .
η
In the total net worth numeraire, the return on total bond portfolio (Bt /Nt = ϑt ) is
drtB = µϑt dt + σtϑ dZt .
On the other hand, by definition
drtB = (1 − ϑtL )drtM + ϑtL drtL

L L
= drtM + ϑtL (σtP σt dt + σtP dZt ).
η
Therefore, the return on reserves
L L
d rtM = (µϑt − ϑtL σtP σt )dt + (σtϑ − ϑtL σtP )dZt
η
Note that reserves is the benchmark asset in this economy. Similar to equation (6.20),
the volatility of ηtI is
ηI M b L M
σt = σtr + (1 − θtM,I − θtL,I )σtxK + θtL,I (σtr − σtr )
M χtI xK b ϑtL ϑt r L M
= σtr + I
( 1 − ϑ t ) σ t + I
(σt − σtr )
ηt ηt
!
χ I (1 − ϑ ) σ ϑ χtI (1 − ϑt ) + ϑt − ηtI L P L
t
= σtϑ + t ( 1 − ψ̄ ) σ − t
+ ϑt σt .
ηtI 1 − ϑt ηtI
131 / 176
ϑ′ (ηtI )ηtI η I L ( P L )′ (ηtI )ηtI η I

Again, by Itô’s lemma, σtϑ = ϑ (ηtI ) t
σ , σtP = ϑ (ηtI )
σt , so
ηI (1 − ϑt ) χtI (1 − ψ̄)σ
ηtI σt = ! !
χtI − ηtI −ϑ′ (ηtI )ηtI L ( P L )′ (η )η I
t t χtI (1 − ϑt ) + ϑt − ηtI
1− + ϑ t
ηtI ϑ (ηtI ) P L (ηtI ) ηtI
A mitigation term arises due to long-term bonds and policy.
6.3.3 Macroprudential Policy
Macroprudential policy usually takes the form of restrictions on the intermediaries’

leverage. In this model, we simply interpret it as the government directly controlling
the portfolio decisions of intermediaries (and households).
6.3.4 Three Policy Benchmarks
Inflation Targeting.
Removing Endogenous Risk. This benchmark of monetary policy aims at eliminat-

ing endogenous risk. That is, µtB = 0, σtB = σtϑ . Now the FOC gives χt in closed form
!
ηtI
χtI = min , χ̄ (6.22)
ηtI + (1 − ηtI ) φ2 + (1 − ψ̄)2 σ2 /σ̃2
and
ηI χtI
σt = (1 − ϑt )(1 − ψ̄)σ.
ηtI
 !2 !2 !2 
ηI 1 − 2ηtI χtI χtI 1 − χtI
µt = (1 − ηtI )(1 − ϑt )2  (1 − ψ̄)2 σ2 + φ2 σ̃2 − σ̃2 
(1 − ηtI )2 ηtI ηtI 1 − ηtI
The “money valuation equation" becomes

 " 
2 2 #  2

ηi η̃ i ηI 1 ηI I
ϑ ρ + µtB − ∑η i
σt +∑η i
σ̃t = ∂t ϑ + (∂η ϑ ) ηtI µt + (∂ηη ϑ) σt ηt
2
i i
 
132 / 176
We compare the equilibrium with and without (µtB = σtB = 0) monetary policy.
Figure 6.8: Drift and volatility of ηtI without policy (solid) and with (dashed).
The price of money falls since the intermediary sector creates more inside money:
it does not need to absorb as much aggregate risk to do that. As a consequence, the
price of capital rises - there is more demand for capital from both sectors. As Figure 6.8
illustrates, capital is shifted to sector b with policy.
The figure also shows the drift and volatility of ηtI with and without policy. With
policy, the intermediary net worth is lower at the steady state. Consequently, their
leverage is higher.
Ultimately, monetary policy affects the degree of market incompleteness with re-
spect to sharing of aggregate risk, but it cannot disentangle risk and risk-taking. The
allocation of capital, the value of money relative to capital, and earnings rates of sectors
a and b as well as intermediaries are endogenously determined by the risk profiles of
available assets.
Perfect Aggregate Risk Sharing. Finally, we can also set the intermediaries’ aggre-
gate risk exposure to zero. From (6.21), we need µtB = 0 and
(1 − ϑt ) χtI (1 − ψ̄)σ + (χtI − ηtI )σtB = 0.
We obtain an economy with perfect sharing of aggregate risk. In this case the aggre-
gate risk exposures of all households and intermediaries is proportional to σtK , and ηt ,
qtK and qtB have no volatility. Also, since intermediaries can trade aggregate risk freely,
households in sector b issue maximal equity shares χ̄ to intermediaries.
Figure 6.9: Comparison on the degree of aggregate risk sharing.
Figure 6.9 compares prices, allocations and dynamics in the baseline model, un-
der policy that eliminates endogenous risk, and with perfect risk sharing. Equilibrium
133 / 176
BIBLIOGRAPHY
moves further in the direction that it took with the application of policy that removes
endogenous risk. Specifically, the value of money falls, the allocation of risk becomes
more skewed to technology b that intermediaries can diversify, the steady state of ηtI
goes to 0 since intermediaries can fully hedge risks, and the volatility of ηtI becomes 0.
Qualitatively, what makes perfect aggregate risk sharing different is the fact that
the boundary condition without intermediaries no longer plays a role at ηtI = 0. The
absence of crisis dynamics contributes to the significant is the drop in the relative value
of money ϑ (ηtI ).8 Also, leverage of intermediaries rises without bound approaching η =
0 - in normal circumstances this would be impossible due to the rise of endogenous risk,
since endogenous risk is generated by the increase in leverage even in environment
when exogenous shocks are small (but not zero).
It is important to highlight one more time the observation that monetary policy
cannot provide insurance and control risk-taking at the same time. Leverage rises en-
dogenously the more risk sharing becomes possible. Asset allocation, together with
asset prices and risk premia, are also endogenous and dependent on the insurance that
monetary policy provides. Hence, the value of money ϑt falls with perfect risk sharing,
which may be detrimental to welfare as we observed in the model without intermedi-
aries.
These links, which cannot be broken without macroprudential policy, have implica-
tions beyond the stylized elements of our model. In particular, loose monetary policy
can lead to excessive leverage in some sectors, reduced risk premia and, consequently,
bubbles in some asset classes. These can pose significant threat to financial stability.
Also, with incomplete markets, improving risk sharing along some dimensions does
not necessarily lead to higher welfare.
Bibliography
Brunnermeier, Markus K. and Yuliy Sannikov, “The I theory of money,” Technical
8 Infact, we might have to raise the idiosyncratic volatility to make money value in the equilibrium
with perfect aggregate risk sharing.
134 / 176
BIBLIOGRAPHY
Report, National Bureau of Economic Research 2016.
Di Tella, Sebastian and Yuliy Sannikov, “Optimal asset management contracts with
hidden savings,” Technical Report 2016.
Woodford, Michael, Interest and prices: Foundations of a theory of monetary policy, Prince-
ton University Press, 2003.
135 / 176
CHAPTER 7. WELFARE AND OPTIMAL MONETARY POLICY
Chapter 7
Welfare and Optimal Monetary Policy
In this chapter, we study welfare and optimal policy design in the monetary models
of Chapter 5 and 6.
7.1 Welfare Function with Log Utility
Recall that under log utility, agents consume a constant fraction of her net worth:
ci,t ĩ = ρni,t ĩ ,
where we have defined
A(ψt ) − ι t
ni,t ĩ = η̃ti,ĩ ηti Nt = η̃ti,ĩ ηti Kt
ρ
Here A(ψt ) is the aggregate productivity which might depend on capital allocation ψt 1 .
The last equality comes from the market clearing condition

Ct = A(ψt ) − ι t Kt = ρNt .
1 In Chapter 5 and 6, we simply had A(ψt ) = a.
136 / 176
CHAPTER 7. WELFARE AND OPTIMAL MONETARY POLICY
In sum,
ci,t ĩ = η̃ti,ĩ ηti Nt = η̃ti,ĩ ηti A(ψt ) − ι t Kt

Under log utility, the welfare of agent (i, ĩ ) is
∞ ∞ ∞
Z Z Z
E0 −ρt
log ci,t ĩ dt = E0 −ρt
log A(ψt ) − ι t dt + E0 −ρt

e e e log Kt dt
0 0 0
Z ∞ Z ∞
−ρt i −ρt i,ĩ
+ E0 e log ηt dt + E0 e log η̃t dt .
0 0
Suppose dXt /Xt = µtX dt + σtX dZt + σ̃tX dZ̃t . By Itô’s lemma,

1 X 2 1 X 2
d log Xt = µtX − (σt ) − (σ̃t ) dt + σtX dZt + σ̃tX dZ̃t .
2 2
In intergral form,
"Z #
∞ ∞
Z 2 1 2
1 1 1
E0 e−ρt log ( Xt ) dt = log ( X0 ) + E0 e−ρt µtX − σX − σ̃ X dt .
0 ρ ρ 0 2 t 2 t
Recall (6.4), we can then write the welfare function according to

"Z #
∞ ∞
Z
log K0 1 1 K 2
E0 e−ρt log Kt dt = + E0 e−ρt Φ (ι t ) − δ − σ dt ,
0 ρ ρ 0 2 t
 
∞ ∞
2 !
log η0i
Z
1 1
Z
−ρt ηi ηi
E0 e = log ηti dt
+ E0  e−ρt µt − σt dt ,
0 ρ ρ 0 2
" #
log η0i,ĩ 1 2
Z ∞ Z ∞
−ρt i,ĩ 1 η̃ i
E0 e log η̃t dt = − E0 e−ρt σt dt .
0 ρ ρ 0 2
137 / 176
Part III
International Models
138 / 176
CHAPTER 8. A SYMMETRIC INTERNATIONAL MODEL WITH RUNS / SUDDEN STOPS
Chapter 8
A Symmetric International Model with

Runs / Sudden Stops
In previous chapters, we have solved models with two-type asymmetric sectors and
visited the essential macro-finance solution techniques. Armed with modeling skills
and math techniques, we put our knowledge into practice. In this chapter, we imple-
ment the new methods into a symmetric international model with runs and sudden
stops. This chapter is mainly based on Brunnermeier and Sannikov (2015).
Financial market, pecuniary externalities and terms of trade hedge. Financial mar-
kets play a vital role in the allocation of physical capital (resources) and risk sharing.
In complete markets, capital allocation and risk sharing can be treated as two inde-
pendent problems and pecuniary externalities have no welfare effects, with only second
order welfare effects. Therefore, first best economy is achieved. Under incomplete mar-
kets with frictions, the allocation of capital and risk are interlinked, i.e., F (κ, χ) ≤ 0. In
such a second best world pecuniary externalities have welfare effects.
Pecuniary externalities can arise due to exogenous credit constraint or and incom-
plete market. In most models inefficiencies arise because the price move tightens an
exogenously imposed collateral constraint (see Greenwald and Stiglitz (1986); Dávila
and Korinek (2017)). Our model in this chapter falls in a second strand of literature in
which pecuniary externalities lead to constrained inefficient outcomes due to incom-
plete market. It follows the general equilibrium literature (Stiglitz (1982),Geanakoplos
139 / 176
and Polemarchakis (1986)). Newbery and Stiglitz (1984)’s incomplete markets setting
shares the feature that free trade destroys the automatic hedge that price movements
provide. However, their paper is about free trade of goods rather than open capital ac-
count, while out analysis also highlights that partially completing the market can lead
to inferior outcomes.
The terms of trade hedge argument is formally introduced by Cole and Obstfeld (1991)
and refers to an automatic insurance system. Although productive inefficiency limits
economic growth, it comes with a favorable side effect: the price of its output good
rises, i.e., the terms of trade improve. Symmetrically, after a positive shock the terms
of trade worsen. We will see that the "terms of trade hedge" is powerful in our interna-
tional model.
International Macro Interpretation. For a long time “Washington Consensus”1 held

the view that free trade (intratemporal flow of goods/services) and full capital account
liberalization (free finance, intertemporal flow of capital) are conducive to higher eco-
nomic growth and welfare. Recently, the IMF took on a more balanced view that in a
second best world, liberalizing only some markets might be harmful. Especially, the
buildup of persistent capital flow imbalances in form of short-term debt, referred to as
hot money, increases the risk of financial instability. To avoid sudden reversals it might
be desirable to “manage” capital flows. Capital controls should be part of the macro-
prudential regulation.
An essential question about "Washington Consensus" is when full capital account

liberalization can reduce welfare while capital controls can improve welfare. To answer
this question, the sudden stop and terms of trade hedge are well studied in this chapter.
1. The inflow of short-term hot money allows the poorer country to boost its pro-
duction capacity, and also acts as a palliative when a country suffers a negative
shock. However, if another series of negative shocks arrive, the country experi-
ences a sudden stop of fundings and conditions deteriorate quickly. Fear of future
deterioration leads to fire sales. Unfortunately, three liquidity mismatchs amplify
the bad situation, which include
1 See the history in the Wikipedia: “Washington Consensus”.
140 / 176
• Technological illiquidity due to the adjustment costs2 .
• Market illiquidity due to low redeployability across border and and produc-
tion specificity.3
• Funding illiquidity due to short-term funding.4
In addition, the country becomes vulnerable to Sunspot Runs. A jump to a disaster

equilibrium leads to a discrete drop in the price of capital, which abruptly erodes
a country’s wealth. In short, better capital allocation financed with hot money
brings lower stability.
2. The terms of trade hedge can be undermined when (i) industry’s output is not
easily substitutable, and (2) lack of strong competitors in other countries5 .
In this chapter, we set up a symmetric international model in section 8.1 following

framework in previous chapters, solve and contrast the equilibrium under three levels
of market completeness in section 8.2. In section 8.3, we conduct an explicit welfare
analysis to see the performance of capital controls under incomplete market. Finally in
section 8.4 we study sudden runs and stops under incomplete market with international
debt market.
8.1 Model Setup
Environment. We develop a simple baseline model of a global economy that is popu-

lated by agents who live in two different countries, A and B. Both types of agents have
the same preferences and can own capital. They can also both produce the two con-
sumption goods a and b6 . Like in the classical Ricardian trade model agents in country
A have a comparative advantage in producing product a, while agents in country B are
2 Recall it is captured by the concavity of investment function Φ
3 We ignore market illiquidity in this chapter.
4 Some other types of capital flow may also matters, including foreign direct investment, portfolio
flows(equity) and long-term debt. Here we focus on short-term debt.

5 eg: (1) natural resources, like oil and copper for chile, (2) Bananas for Ecuador, computer hard drives
for Thailand
6 For simplicity, we use I to indicate country, I ∈ { A, B }, and i for good, i ∈ { a, b } in this chapter.
141 / 176
better at producing product b. There are no trade barriers for the two output goods a
and b as well as for the input good, physical capital.
Technology. Capital can be used to produce goods a or b, which can be combined

to produce the composite good. The composite good can be consumed, or used for
investment to produce new capital. Goods a (say "apple") and goods b (say "banana")
are imperfect substitutes and when quantities yta and ybt of goods a and b are combined
they make a total quantity
" # s
s −1
1 a s−s 1 1 b s−s 1
yt = y + y ,
2 t 2 t
of composite good. s ∈ [0, ∞) is the elasticity of substitution between goods a and b.7
The composite good serves as numeraire and its price is normalized to one.
Agents in different countries have comparative advantages. From k t unit of capital

an agent in country A can produce good a at rate ak t and good b at rate of only ak t ,
where a > a ≥ 0.8 Symmetrically, an individual in country B can produce good b at
rate ak t and good a at rate ak t .
We denote the aggregate amount of world capital at time t ∈ [0, ∞) by Kt . Define

the fraction of world capital used by agents in country A to produce good a as κtAa , the
fraction used by agents in country B to produce good b as κtBb , etc., so that
κtAa + κtAb + κtBa + κtBb = 1
Then the total world supply of goods a and b is given by

Yta = aκtAa + aκtBa Kt and Ytb = aκtBb + aκtAb Kt ,
7 Note for s = ∞ both goods are perfect substitutes, for s = 0 there is no substitutability à la Leontieff,
while for s = 1 the substitutability corresponds to that of a Cobb-Douglas.
8 If a = 0, then country A will not use capital to produce good b, because it is a strictly (weakly if
there is no country B) dominant strategy to produce good a.
142 / 176
and the total world supply of composite good writes

" # s
s −1 s −1
1 a
s −1 1 s
Yt = Y s
+ Ytb , (8.1)
2 t 2
hence the prices of goods a and b in terms of the composite good
  1   1
!1/s ! s −1 s −1 ! s −1 s −1
s s s
1 Yt 1 1 1 Ytb 1 s −1  aκtBb + aκtAb
Pta = =  + = 1 + ,
 
2 Yta 2 2 2 Yta 2 aκtAa + aκtBa
 
  1   1
!1/s ! s −1 s −1 ! s −1 s −1
s s s
1 Yt 1 1 1 Yta 1 s −1  aκtAa + aκtBa
Ptb = =  + = 1 + .
 
2 Ytb 2 2 2 Ytb 2 aκtBb + aκtAb
 
(8.2)
Then the terms of trade writes

!1/s !1/s
Pta Ytb aκtBb + aκtAb
= = . (8.3)
Ptb Yta aκtAa + aκtBa
Capital employed in country I (I ∈ { A, B}) evolves according to
dk tI
I
= (Φ(ι tI ) − δ)dt + σ I dZtI .
kt
The concavity of investment function Φ(ι) captures technological illiquidity, i.e., the ad-
justment cost due to converting output to new capital and vice versa. The country spe-
cific Brownian motions dZtA , dZtB are exogenous and independent.9 These two shocks
affect both (i) the global capital stock Kt as well as (ii) the relative wealth shares ηt . In
other words, shocks have redistributive consequences.10 This allows us to provide a
different interpretation of the shock structure.
9 Examples of such shocks are the discovery of new resources or natural catastrophes like earthquakes
and tsunamis.
10 Example: Apple vs. Samsung lawsuit.[Add some details later...]
143 / 176
Preferences. All agents in the country A or B have a CRRA utility function with same
discount rate ρ
"Z #
∞ ( c i )1− γ
E0 e−ρt t dt .
0 1−γ
Market structures. We study the international model under three settings – complete
markets, open credit account and closed credit account. Figure 8.1 reveals market struc-
tures for the three settings. In section 8.2.1 we postulate stochastic processes and ana-
lyze the benchmark case – economy under complete market and no frictions. Then we
move to open capital account with debt in section 8.2.2 and closed capital account in
Market structures
8.2.3. Finally we conduct a formal welfare analysis in section 8.3.
Trade Finance
Markets Output Physical Debt Equity

𝑦𝑦 𝑎𝑎 , 𝑦𝑦 𝑏𝑏 capital 𝐾𝐾
Complete Markets X X X X
Full integration/First Best
Open credit account X X X
(equity home bias)
Closed credit account
X X
Brunnermeier & Sannikov
Add taxes/capital controls

intratemporal intertemporal
13
Figure 8.1: Markets structures of three settings
8.2 Contrasting Different Degrees of Market Complete-

ness
8.2.1 Complete Markets
In this section an economy without frictions under complete markets is studied.

First, we postulate essential processes, which will also be used in incomplete market
144 / 176
settings. Second, we solve for the equilibrium following special techniques of chapter
3. Finally, we conclude the essential characteristics of equilibrium of the benchmark
model.
Step 0 Postulate price processes and obtain return processes
All agents can trade physical capital in a fully liquid international market, which is
affected by shocks dZtA and dZtB . For now the equilibrium market price of capital per
unit qt is measured in the composite good. We postulate that qt follows
dqt q qA qB
= µt dt + σt dZtA + σt dZtB
qt
The return from capital depends on the identity of the agent holding it and the good
that it is used to produce. The capital gains from capital are given by d(qt k tI )/(qt k tI ).
The dividend yield from capital – after reinvesting output at a rate ι tI – is given by
( aPti − ι tI )/qt when it is used to produce good i = a, b with comparative advantage,
and by ( aPti − ι tI )/qt otherwise. For example, when an agent of type A uses capital to
produce good a, he earns the return of
Dividend Yield Capital Gain

z }| { z }| {
aPta − ι tA d(qt k tA )
drtAa = dt +
qt qt k tA (8.4)
" #
a
aPt − ι tA
q qA qA qB
= + µt + Φ(ι tA ) − δ + σ A σt dt + σ A + σt dZtA + σt dZtB .
qt
Similarly when agent A uses capital to produce good b, he earns

" #
aPtb − ι tA q A qA

qA

qB
drtAb = A
+ µt + Φ(ι t ) − δ + σ σt A
dt + σ + σt dZtA + σt dZtB . (8.5)
qt
Similar equations hold for agents in country B.
We then postulate that SDF (ξ tI = e−ρt [u′ (ctI )/u′ (c0I )]) follows
dξ tA /ξ tA = −rtA,F dt − ς tAA dZtA − ς tAB dZtB , (8.6)
145 / 176
dξ tB /ξ tB = −rtB,F dt − ς BA A BB B
t dZt − ς t dZt , (8.7)
where rtI,F is risk-free rate in country I. Note now SDF depends on shocks in both
country – for example, shock in country B can affect capital price qt and consequently
affect the motion of ξ tA .
Step 1 For given SDF processes, derive equilibrium conditions
Optimal Reinvestment rate ι. Like before, the optimal investment rate, which maxi-
mizes returns, is always given by the first-order condition (Tobins’ q)
Φ′ (ι tI ) = 1/qt .
Asset pricing equations. Recall the martingale approach discussed in Chapter 2, for
agents A, the return of any asset has to satisfy the following property: if wealth ϵt is
invested in asset X, so that dϵt /ϵt = drtX , then ξ tA ϵt must be a martingale if the portfolio
position of X is positive, i.e., the drift of ξ tA ϵt must be zero. On the other hand, if the
portfolio position of X is zero, then ξ tA ϵt must be a supermartingale11 , i.e., the drift of
ξ tA ϵt is negative.
Therefore, then pricing condition for capital used to produce good a is
Expected return Risk premium,−Cov[dr Aa ,dξ A /ξ A ]

z h }| i { Riskz}|{
free rate z }| t t t
{
Aa A,F AA qA A AB qB
Et drt /dt − rt = ςt σt + σ + ς t σt . (8.8)
Likewise, for capital used to produce good b,
Expected return Risk premium,−Cov[dr Aa ,dξ A /ξ A ]

z h }| i { Riskz}|{
free rate z }| t t t
{
Ab A,F AA qA A AB qB
Et drt /dt − rt ≤ ςt σt + σ + ς t σt . (8.9)
Similar equations also hold for agents B.
h i h i
11 Martingale: Et ξ tA+s ϵt+s = ξ tA ϵt ; supermartingale: Et ξ tA+s ϵt+s ≤ ξ tA ϵt .
146 / 176
First Best Benchmark. In a complete markets setting capital allocation and risk shar-
ing can be treated as two independent problems. And perfect specialization and perfect
risk sharing realize.
Perfect specialization includes
1. Investment rate equalization: according to Tobin’s q, now ι tA = ι tB .
2. Full specialization: for simplicity, we assume a symmetric economy, in which σ A =

σ B = σ. Then given the efficient allocation of capital to the production of the two
goods, κtAa = κtBb = 1/2 and κtAb = κtBa = 0 – agents A specialize in producing
output good a and agents B only produce output good b.
3. Output equalization: Since marginal cost of goods a and b are identical, marginal
product of goods a and b in producing output-index Y must be the same. Hence
the social planner must also choose output equalization Yta = Ytb , and the com-
posite output is Yt = aKt /2.
A social planner that assigns Pareto weights (λ A , λ B ) must divide the consumption
stream according to the shares
(λ A )γ (λ B )γ
CtA = Ct and CtB = Ct , (8.10)
(λ A )γ + (λ B )γ (λ A )γ + (λ B )γ
where Ct is aggregate global consumptions.12
Besides, perfect risk sharing realizes. As two countries have symmetric response to
Brownian shocks dZtA and dZtB , we can aggregate them into a single standard Brownian
√
shock dZt = (dZtA + dZtB )/ 2. Now the global capital evolutes as
σ dZtA + dZtB
dKt = Φ (ι t ) − δ Kt dt + √ Kt

√ (8.11)
2 2
| {z }
:=dZt
Proposition 8.1. With complete markets, the market outcome leads to the first-best allocation
with full specialization, κtAa = κtBb = 1/2. The risk-free rate r F and the price of capital are
12 By FOC: u′ (CtA )/u′ (CtB ) = λ A /λ B .
147 / 176
time-invariant and given by
γ ( γ + 1) σ 2 a/2 − ι
r F = ρ + γ(Φ(ι) − δ) − and q= . (8.12)
4 rF + γ 2
2σ − (Φ(ι) − δ)
The time-invariant wealth shares (η, 1 − η ) are also the consumption shares. The corresponding
Pareto weights of the centralized economy can be found from (8.10).
Proof. In Brunnermeier and Sannikov (2015), Proposition 8.1 is proofed by solving so-
cial planner’s problem via HJBs. Here we derive (8.12) via the Macro-finance tech-
niques in Chapter 3.
Time-invariant risk-free rate. Recall CRRA utility function u′ (c) = c−γ and SDF ξ tI =
e−ρt [u′ (CtI )/u′ (C0I )], I ∈ { A, B}. As consumption fractions are constants, according to
(8.10), CtA /Kt is also a constant, hence we have
!−γ −γ !−γ
CtA CtB

−ρt −ρt Kt −ρt
ξ tA =e =e =e = ξ tB
C0A K0 C0B
Plug in (8.11) and use Itô’s lemma,

( )
dξ tA dξ tB γ ( γ + 1) σ 2 γσ
−ρ − γ Φ (ι t ) − δ +

= = dt − √ dZt .
ξ tA ξ tB 4 2
| {z }
=Et [dξ tI /ξ tI ]/dt
Compared with the postulated SDF in (8.6), we have risk-free rate and risk premium
γ ( γ + 1) σ 2 γσ
rtF = ρ + γ[Φ(ι t ) − δ] − , ςt = √ . (8.13)
4 2
Time-invariant price. In complete market investment opportunities ωt is time-invariant,

hence recall this special case of (3.14)
" #
ct γ−1 F ς2t
= ρ+ rt − ρ + . (8.14)
nt γ 2γ
148 / 176
Recall the aggregate consumption Ct = CtA + CtB = ( a − ι t )Kt and aggregate net work
Nt = qt Kt . Plugging in the risk-free rate (8.13), we have
Nt a − ιt a − ιt
qt = = = F γ 2 . (8.15)
Kt Ct /Nt rt + 2 σ − [Φ(ι t ) − δ]
d
Note time-invariant q is given by the well-known Gordon growth formula p = r− g ,
where the numerator is the dividend rate a − ι and the denominator is the required
return on capital r F + γ2 σ2 minus the growth rate of capital Φ(ι) − δ.
The first-best benchmark economy which emerges in a setting with complete mar-
kets is particularly simple. Perfect capital allocation and perfect risk sharing realize.
Networth shares are equal to consumption shares, which depend on Pareto weights
(λ A , λ B ). Prices are constant and independent of shocks. The size of economy essen-
√
tially scales up or down depending on the total shock dZt / 2. Besides, the elasticity
of substitution, s, has no impact on prices.
8.2.2 Open Capital Account for Debt
In this section, we analyze the equilibrium in an economy in which agents in both

countries can borrow through risk-free debt. Markets are not complete since agents can-
not issue equity claims to foreigners (equity home bias). We follow the Macro-finance
techniques in Chapter 3 to solve the equilibrium.
Step 0 Postulate price processes and obtain return processes
Same as Step 0 in Section 8.2.
Step 1 For given SDF processes, solve for individual equilibrium conditions
Price-taking planner’s problem. We solve for capital and risk allocations via the "price-
taking planner’s problem". Recall the general form of the problem with one exogenous
149 / 176
shock dZt :
(Weighted) Risk Premium

 z  }|  {
 h i 
∑
K
max Et drtK (κt ) /dt −  ςit χit  σr
× |{z} s.t. Financial Friction.
{κt ,χt } | i ={e,h}

{z } Risk
Expected Return | {z }
(Weighted) Price of Risk
With two brownian shocks in open capital account for debt setting, the price-taking
planner’s problem can be written as
 A-risk Premium
B-risk Premium 
 h i }|
z
{ z }| {
max Et drtK (κt ) /dt − ς tAA χtAA + ς BA BA
t χt σ N A − ς tAB χtAB + ς BB
× |{z} t χt
BB
σ NB
× |{z}
{κt ,χt }  | {z } | {z } A-risk | {z } B-risk 
Expected Return (Weighted) Price of Risk (Weighted) Price of Risk
qA qB qA qB
κtA (σ A + σt ) AB κtA σt κtB σt κtB (σ B + σt )
s.t. χtAA = BA BB
, χt = N A , χt = NB , χt = .
σtN A σt σt σtNB
| {z }
Financial frictions: no outside equity issuance
As postulated before, κt = (κtAa , κtAb , κtBa , κtBb ) denotes capital fractions, with κ I =
κ Ia + κ Ib the whole capital allocated to agents I ∈ { A, B}. Let χt = (χtAA , χtAB , χtBA , χtBB )
denotes risk allocations, with χ I J the risk share of agents I loading on brownian dZtJ .
Note shocks dZtJ 1) affects capital of agents I ( I = J ) by κtI σ I , and 2) induces price
volatility σqJ , which works for both agents A and B. Hence the volatility of world net
qJ
worth affected by brownian dZtJ 13 is given by σtN J (κt ) = κtJ σ J + σt . As in current set-
ting there is no outside equity issuance, the risk allocations is fully captured by capital
allocations, as the friction friction conditions show.
Recall (8.4) and (8.5), the overall return to capital14 :

" #
∑ I ∑i κtIi a Ii Pti − ι t
∑ ∑ + Φ(ι t ) − δ + µt + ∑ σ I σt dt
q qI
drtK (κt ) = κtIi drtIi =
I ∈{ A,B} i ∈{ a,b}
qt I
+ volatility terms.
13 For simplicity, we call it J-risk

14 a Aa = a Bb = a, a Ab = a Ba = a.
150 / 176
Hence, the planner’s problem can be written as

 
∑ I ∑i κtIi a Ii Pti − ιt
 
max − ς tAA χtAA + ς BA BA
t χt × σtN A − ς tAB χtAB + ς BB BB
t χt × σtNB
{κt ,χt }  qt 
qA qB qA qB
κtA (σ A + σt ) AB κtA σt κtB σt κtB (σ B + σt )
s.t. χtAA = BA BB
, χt = N A , χt = NB , χt =
σtN A σt σt σtNB
qA qB
σtN A (κt ) = κtA σ A + σt , σtNB (κt ) = κtB σ B + σt
κ Ii ∈ [0, 1], κtAa + κtAb + κtBa + κtBb = 1,
Pta , Ptb satisfy (8.2)
(8.16)
It’s fairly complicated to solve the problem directly with regard to κ Aa , κ Ab , κ Ba , κ Bb .

For simplicity, we first solve for the static intermediate goods market equilibrium to
derive the effective capital productivity in composite good A(κ A ). After that, the social
planner’s problem is simplified with only one variable κ A .
Taking price P a , Pb and κ A as given, the capital allocation problem for country A is
given by
max aκ Aa P a + aκ Ab Pb ,
{κ Aa ,κ Ab }
s.t. κ Aa + κ Ab = κ A , κ Aa , κ Ab ≥ 0.
and similar problem for country B writes
max aκ Ba P a + aκ Bb Pb ,
{κ Ba ,κ Bb }
s.t. κ Ba + κ Bb = 1 − κ A , κ Ba , κ Bb ≥ 0.
Solving above problems respectively, the combining FOC is as follows:
151 / 176
Both only produce goods ; Full specialization Both only produce goods .
Country A: ! <= = ! < , ! <? = 0 Country A: ! <? = ! < , ! <= = 0 Country A: ! <? = ! < , ! <= = 0
Country B: ! @= = ! @ , ! @? = 0 Country B: ! @= = ! @ , ! @? = 0 Country B: ! @? = ! @ , ! @= = 0
Region I Region II Region III

A? /A=
0 ./.8 ./.
8 1
Country A: ! <? + ! <= = ! < , 0 < ! <= < ! < Country A: ! <? = ! < , ! <= = 0
Country B: ! @= = ! @ , ! @? = 0 Country B: ! @? + ! @= = ! @ , 0 < ! @= < ! <
Figure 8.2: FOCs of capital allocation problem for country A and B (given P a , Pb , κ A ).
Substituting the FOCs in to the market-clearing price ratio (8.3), we find a/a ≤
3
P a /Pb ≤ a/a. Region I and Region III are forbidden15 , and in Region II (without end-
points) full specialization realized. Note in the left endpoint country A produce both
goods a and b16 while country B specializes in good b, and in the right endpoint a
symmetric condition appears.
Then we derive the capital productivity case by case. When a/a < P a /Pb < a/a,
according to (8.3),
!1/s
Pa 1 − κA as A as
= ⇒ < κ < , (8.17)
Pb κA as + as as + as
and κ Ii , P a , Pb can be written in κ A :
κ Aa = κ A , κ Bb = 1 − κ A , κ Ab = 0, κ Ba = 0,
  1   1
s ! s −1 s −1 s ! s −1 s −1
s s
a 1 s −1  1 − κA b 1 s −1 κA
P = 1 + , P = 1 + .
  
2 κA 2 1 − κA
 
15 Intuitively,in Region I P a /Pb ≤ a/a, i.e., the price of good a is relatively much lower. But if firms
in country A and B both only produce good b, the price of good a should be much higher.
16 Hence the price ratio, i.e. terms of trade hedge, remains a constant.
152 / 176
Plug the capital allocation into CES production function (8.1) and now the "effective"
capital productivity in composite good follows
" # s
s −1 s −1 s −1
Yt (κ A ) 1 s 1 s
A(κ A ) := =a κA + 1 − κA (8.18)
Kt 2 2
In the left endpoint of Region II (P a /Pb = a/a), "rich" country A shares a larger capital
as
fraction, i.e., κ A ≥ as + as
, and participates in both goods production. According to (8.3),
the κ Ii , P a , Pb can be written in κ A :
!1/s
Pa a(1 − κ A ) + a(κ A − κ Aa ) a
= =
Pb aκ Aa a
as (1 − κ A ) + as−1 aκ A a s κ A − a s (1 − κ A )
⇒ κ Aa = , κ Ab =
a ( a s −1 + a s −1 ) a ( a s −1 + a s −1 )
⇒ κ Bb = 1 − κ A , κ Ba = 0,
s 1 s 1
a 1 1 s −1 s −1 s −1 s −1 1 1 s −1 s −1 s −1 s −1
⇒P = a +a , Pb = a +a
a 2 a 2
now the effective capital productivity in composite good writes,

s
Yt (κ A ) 1
A 1 s −1 s −1 s −1 s −1 a A A
A(κ ) := = a +a (1 − κ ) + κ (8.19)
Kt 2 a
Symmetrically, we obtain the effective capital productivity in the right endpoint, i.e.
as
κA ≤ s
a + as
:
s
Yt (κ A ) 1
A 1 s −1 s −1 s −1 s −1 A a A
A(κ ) := = a +a (1 − κ ) + κ (8.20)
Kt 2 a
Figure ?? plots the effective productivity A(κ A ) under different elasticity of sub-
stitution s, where the red circles depict the switching points for three regions17 . The
international system achieves higher aggregate productivity under higher substitution
elasticity. While under perfect substitution case, i.e., s = ∞, good a and good b can be
regarded as one good and the productivity stays at a constant a/2. Note all the curves
17 We will see that it corresponds to the (η a , η b ) in the step 4.
153 / 176
pass through the same point (1/2, a/2), as we have showed in the previous complete
market case (perfect specialization).
0.08
0.07
0.06
0.05
0.04
0.03
s = 0.1 s = 10
s = 1.01 s=
0.02
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Figure 8.3: Effective productivity in composite good A(κ A ). The red circles depict the
switching points for three regions. a = .14, a = .04
Substitute the effective productivity into social planner’s problem, then the FOC
w.r.t. κ A writes:
A ′ (κ A )
− (σ A + σqA )ς AA + (σ B + σqB )ς BB
qt
κ A σ A + σqA A B
AB qB (1 − κ ) σ + σ
qB
+ ς BA σqA − ς σ
(1 − κ A )σ B + σqB κ A σ A + σqA
σ A σ B + σqB σ A + σqA σ B A B qB A
AB qB A σ σ + σ σ + σ σ
qA B
− ς BA σqA (1 − κ A ) 2 + ς σ κ 2 =0

(8.21)
154 / 176
Step 2 Value functions
Value function and BSDE. Same as section 3.2.3, we can derive the prices of risk ςs
from value function.
Aggregate capital Kt follows the process of dKt /Kt = (Φ(ι t ) − δ)dt + κtA σ A dZtA +
κtB σ B dZtB , according to Itô’s lemma,
d ( K t )1− γ

1 A A 2 B B 2
= (1 − γ)(Φ(ι t ) − δ) − γ(1 − γ)[(κt σ ) + (κt σ ) ] dt
( K t )1− γ 2
+ (1 − γ)κtA σ A dZtA + (1 − γ)κtB σ B dZtB
Then apply the Itô’s product rule to the value function (3.16),
dVtI

vA 1
I
= µ t + ( 1 − γ )( Φ ( ι t ) − δ ) − γ(1 − γ)[(κtA σ A )2 + (κtB σ B )2 ]
Vt 2
A B
i
+(1 − γ)(κ A σ A σtv + κ B σ B σtv ) dt + volatility terms.
Recall (3.15), now the drift of VtA equals
I CI
µV = ρ I − tI
Nt

vI 1 A A 2 B B 2 A A vA B B vB
= µt + (1 − γ)(Φ(ι t ) − δ) − γ(1 − γ)[(κt σ ) + (κt σ ) ] + (1 − γ)(κ σ σt + κ σ σt )
2
This gives us the following backward stochastic differential equation (BSDE)

"
dvtI I CtI 1
I
= ρ − I
− (1 − γ)(Φ(ι t ) − δ) + γ(1 − γ)[(κtA σ A )2 + (κtB σ B )2 ]
vt Nt 2 (8.22)
A B
i I I
−(1 − γ)(κ A σ A σtv + κ B σ B σtv ) dt + σtv A dZtA + σtv B dZtA .
Like in Chapter 3, we will use the BSDE to derive PDE in step 4.
Prices of risk ς t . The optimal consumption condition (3.18) now writes

!1/γ
CtI ηtI qt
= ,
Kt vtI
155 / 176
where ηtI is the networth share of country I ( I ∈ { A, B}). Applying Itô’s quotient rule
and comparing the volatility terms (note there are two brownians), the prices of risk for
agents A writes
AA AA ηA A qA
ς tAA = γσtC = −σtv + σt + σt + γκtA σ A ,
(8.23)
AB AB ηAB qB
ς tAB = γσtC = −σtv + σt + σt + γκtA σ B ,
where σt· J ( J ∈ { A, B}) donate the volatility of variables loading on Brownian dZtJ , and
ς tI J ( I, J ∈ { A, B}) is the price of risk of agents I loading on brownian dZtJ . The appear-
ance of κtA is due to aggregate capital evolution dKt = Φ (ι t ) − δ Kt dt + σ A κtA Kt dZtA +

σ B κtB Kt dZtB .
Similarly we have the prices of risk for agents B
BA BA ηB A qA
ς BA
t = γσtC = −σtv + σt + σt + γκtB σ A ,
BB BB ηB B qB
(8.24)
ς BB C
t = γσt = −σtv + σt + σt + γκtB σ B .
Step 3 Evolution of state variable ηt
Like Chapter 3, Markov equilibrium is a map from histories of shocks to equilibrium

prices.
map from histories { ZsA , ZsB , s ∈ [0, t]} to price qt , investment rate (ι tA , ι tB ), price of risk ςit and
capital/risk allocation (κtAa , κtAb , κtBa , κtBb ), (χtAa , χtAa , χtAa , χtAa ), such that
(2) all markets, for capital, equity and consumption goods, clear.
η
Drift of ηt . Similarly, we compute the drift of wealth shares µt using the methods of
change-of-numeraire and martingale approach. Here we change the numeraire from
consumption goods to the total wealth in the economy Nt = NtA + NtB . Like (3.10), the
156 / 176
martingale asset pricing formula for agents A can be written as
ηA CtA ηA A ηAB
µt + − rtF = (ς AA − σ N A ) σ + (ς tAB − σtNB ) σt (8.25)
NtA | t {z t } t | {z }
price of risk price of risk
under Nt numeraire under Nt numeraire
loading on dZtA loading on dZtB
Aggregate {ηt }-weighted sum of agents A and agents B
=0, as η A +η B =1
z }| {
ηA ηB Ct ηA A ηB A
(ηtA µt + ηtB µt ) + − rtF = ηtA (ς tAA − σtN A )σt + ηtB (ς BA NA
t − σt ) σt (8.26)
Nt
ηAB ηB B
+ ηtA (ς tAB − σtNB )σt + ηtB (ς BB NB
t − σt ) σt .
Subtract (8.26) from (8.25), the drift of ηtA is
ηA ηA A ηB A
µt = (1 − ηtA )(ς tAA − σtN A )σt − (1 − ηtA )(ς BA NA
t − σt ) σt
ηAB ηB B CtA Ct (8.27)
+ (1 − ηtA )(ς tAB − σtNB )σt − (1 − ηtA )(ς BB NB
t − σt ) σt −( A
− ).
Nt Nt
Volatility of ηt . Recall that the wealth share ηtI is numeraire invariant, so we still use
Itô’s quotient rule to solve for its volatility. First focus on country A, since ηtA = NtA /Nt ,
ηA A AA AA AA BA AA BA
σt = σtN − σtN A = σtN − (ηtA σtN + (1 − ηtA )σtN ) = (1 − ηtA )(σtN − σtN ),
ηAB AB AB AB B AB B
σt = σtN − σtNB = σtN − (ηtA σtN + (1 − ηtA )σtN B ) = (1 − ηtA )(σtN − σtN B ),
recall in our setting
AA κtA A qA BA κtB qA
σtN = A
(σ + σt ), σtN = A
σt ,
ηt 1 − ηt
AB κtA qB BB κtB qB
σtN = σ , σtN = (σ B + σt ),
ηtA t 1 − ηtA
hence
ηA A 1 h A A A A A qA
i
σt = ( 1 − η t ) κ t σ + ( κ t − η t ) σt , (8.28)
ηtA
157 / 176
ηAB 1 h A A B A A qB
i
σt = − η t ( 1 − κ t ) σ + ( κ t − η t ) σt . (8.29)
ηtA
Note ηtA + ηtB = 1, hence
A B AA BA AB B
dηtA + dηtB = (ηtA µη + ηtB µη )dt + (ηtA ση + ηtB ση )dZtA + (ηtA ση + ηtB ση B )dZtB = 0
AA A ,A ηB A ηtA ηA A
⇒1. ηtA ση + (1 − ηtA )σ1−η =0 ⇒ σt =− σt
1 − ηtA
AB A ,B ηB B ηtA ηAB
2. ηtA ση + (1 − ηtA )σ1−η =0 ⇒ σt =− σt
1 − ηtA
Amplification. We can use Equation (8.28,8.29) to evaluate the volatilities of η A and q

from κtA and the values of qt (ηtA ) and q′t (ηtA ). Indeed, using Itô’s lemma,
asset-equity ratio
z }| {
qA q′t (ηtA )ηtA η A A (8.28) ηA A κtA /ηtA (1 − ηtA )
σt = σ ===⇒ σt = σA (8.30)
qt (ηtA ) t 1− (κtA − ηtA )/(ηtA )
q′ (ηtA )
| {z } q(ηtA )/ηtA
Leverage: debt-equity ratio
The factor in front of σ A highlights the amplification, which is captured by two

effects.
• Leverage effect: asset-equity ratio κtA /ηt and debt-equity ratio (κtA − ηt )/ηt cap-
tures amplification effect with leverage effect.
• Loss spiral: The denominator captures the loss spiral. If the price of capital is more
sensitive, i.e., q′ (η ) is larger, additional losses arise which amplify the endogenous
risk even further. For constant qt (Frist-best Benchmark), i.e., q′t = 0, the loss spiral
is switched off.
Step 4 Value function iteration and good market clearing
Proposition 8.2 characterizes the evolution of wealth share η A .
Proposition 8.2. In incomplete market with open capital account for debt setting, the state
space can be divided into three regions.
158 / 176
In the middle region, η A ∈ [η a , η b ] full specialization realizes, i.e. (κ Ab = κ Ba = 0). In

the left region [0, η a ), κ Ba > 0, and the right region (η b , 1], κtAb > 0. ηtA follows dηtA /ηtA =
A AA AB
µη dt + ση dZtA + ση dZtB , where the drift and volatility are from (8.27-8.29).
In addition,
Pa
a/a ≤ ≤ a/a
Pb
In the left region the first inequality becomes equality, and the right, the second.
Algorithm. To solve the equilibrium, we take η A as state variable and postulate that
vtI = v I (ηtA , t). By the Multivariate Itô’s fomular18 , it follows
ηA ηA A 2 ηAB 2
dvtI ∂t vtI + (ηtA µt )∂η vtI + 12 [(ηtA σt ) + (ηtA σt ) ]∂ I
ηη vt
= dt
vtI vtI
ηA A ηAB
(ηtA σt )∂η vtI (ηtA σt )∂η vtI
+ dZtA + dZtB .
vtI vtI
Comparing with the BSDE (8.22), we get the growth equation

ηA ηA A 1 ηA A
ηAB ηAB
∂t vtI
+ ηtA µt + (1 − γ)(κtA σt σ A ∂η vtI + [(ηtA σt )2 + (ηtA σt )2 ]∂ηη vtI
+ κtB σt σ B )ηtA
2
CI

1
I
= ρ − (1 − γ)(Φ(ι t ) − δ) + γ(1 − γ)[(κt σ ) + (κt σ ) ] vtI − tI vtI .
A A 2 B B 2
(8.31)
2 Nt
where
ηA ηA A ηB A
µt = (1 − ηtA )(ς tAA − σtN A )σt − (1 − ηtA )(ς BA NA
t − σt ) σt
ηAB ηB B CtA Ct
+ (1 − ηtA )(ς tAB − σtNB )σt − (1 − ηtA )(ς BB NB
t − σt ) σt −( A
− ),
Nt Nt
qA qB
σtN A = κtA σ A + σt , σtNB = (1 − κtA )σ A + σt ,
ηB A ηtA ηA A ηB B ηtA ηAB
σt =− σt , σt =− σt
1 − ηtA 1 − ηtA
ηA A ηAB
To settle the PDE, we use four red variables κtA , qt , σt , σt as intermediate vari-
18 See Yuliy’s "Overview of Stochastic Calculus"
159 / 176
ables, which can be determined by four conditions:
• Good market clearing condition

!1/γ !1/γ
ηtA qt (1 − ηtA )qt
A(κtA ) − ι t = + . (8.32)
vtA vtB
• The amplification equations
κtA 1−κtA
ηtA
(1 − ηtA ) 1−ηtA
(1 − ηtA )
ηA A ηAB
σt = σA, σt =− σB. (8.33)
q′ (η A ) q′ (η A )
1 − [κtA − ηtA ] q(η At ) 1 − [κtA − ηtA ] q(η At )
t t
• FOC condition for the planner’s problem
A ′ (κ A )
− (σ A + σqA )ς AA + (σ B + σqB )ς BB
qt
κ A σ A + σqA A B
AB qB (1 − κ ) σ + σ
qB
+ ς BA σqA − ς σ
σ A σ B + σqB σ A + σqA σ B A B qB A
AB qB A σ σ + σ σ + σ σ
qA B
− ς BA σqA (1 − κ A ) 2
+ ς σ κ 2
=0

(8.34)
The blue terms can be written in red terms:
• Tobin’s q gives us the investment rate ι t :
1
ιt = ( q t − 1). (8.35)
ϕ
• Productivity A(κ A ) is given by (8.18,8.19,8.20) in three regions.
• σqA , σqB are given by Itô’s lemma (8.30)
qA q′t (ηtA )ηtA η A A qB q′t (ηtA )ηtA η A B

σt = σ , σt = σ (8.36)
qt (ηtA ) t qt (ηtA ) t
160 / 176
• The price of capital ς tI J is given by (8.23, 8.24) :
ηI J
I ηI J qJ I (ηtI σt )∂η vtI
ς tI J = −σtv J + σt + σt + γκtI σ J where σtv J = . (8.37)
vtI
• In (3.20) and (3.21), we have obtained the consumption ratios Ct /Nt , CtI /NtI from
the optimal consumption condition:
!1/γ
CtI (ηtI qt )1/γ−1 Ct 1 ηtI qt
NtI
=
(vtI )1/γ
,
Nt
=
qt ∑ vtI
. (8.38)
i
Similarly, the pseudo-code goes as following
1. Start by guessing two functions v A (η A , T ), v B (η A , T ) over a grid of η A
2. Loop over t = { T, T − ∆t, · · · , 0} until changes in v A -functions are small. In

each step, do the following:
(a) Compute ∂η vtA by first-order differences

(b) Start at η A = 0 and solve to η A = 1/2a . Compute the red variables
using three equilibrium conditions (8.32-8.34).
(c) Determine the growth equation (8.31) by plugging in red and blue
items.
(d) make time-step – back in time – and update the vit (t, ·) functions to
vit (t − ∆t, ·)
a Note η A is divided into two regions, where κ A ∈ (0, as /as + as ), ( as /as + as , 1/2)
161 / 176
Step 5 Numerical Results
Equilibrium dynamics. For numerical examples we assume that that all agents have
logarithmic utility, i.e., γ = 1, and that the investment technology is given by Φ(ι) =
1 19
p
ϕ ( 1 + 2ϕι − 1) . The base line parameters are as follows.
ρ a a δ σA σB ϕ s
0.05 0.14 0.04 0.05 0.10 0.10 2 {0.5, 1, ∞}
Note that due to the symmetry of our setting, it is sufficient to characterize the equi-
librium for the wealth shares η A ∈ [0, 0.5]. The equilibrium dynamics are symmetric
about η A = 0.5.
Panel A of Figure 8.4 plots the capital shares for three different levels of elasticity of
substitution. With incomplete markets, as long as η A ≤ 0.5 country A still specializes
in good a. However, as the wealth share η A declines, so does the capital share κ Aa . The
capital share declines slower than the wealth share as firms in country A can borrow,
hence the curve κ Aa stays above the 45-degree line. The level of borrowing depends on
the elasticity of substitution s between two goods. Specially, with perfect substitutes
s = ∞ there is no borrowing20 (dotted magenta line). As η A falls below η a , firms in
country B take participate in producing good a as well. In the special case perfect
substitutes, κ Ba = 0 for any η A .
19 The investment technology has quadratic adjustment costs, i.e., an investment of Φ + ϕ Φ2 generates
2
new capital at rate Φ.
20 For proof, see Corollary 1 in Brunnermeier and Sannikov (2015)
162 / 176
Figure 8.4: Panel A plots the capital shares κ Aa and κ B a, Panel B plots the terms of trade
P a /Pb and Panel C plots the price of physical capital q, as functions of the wealth share
η A , for three different levels of elasticity of substitution: s = 0.5 in dashed blue, s = 1.01
(Cobb-Douglas) in solid black, and s = ∞ in dotted magenta.
Panel B of Figure 8.4 clearly depicts the terms of trade hedge. As the wealth share
of country A drops after a negative shock, output good a becomes more scares and its
price rises – country A receive better terms of trade. Once the ratio P a /Pb rises to a/a,
firms in country B start producing good a and the ratio is capped at a/b. Note that the
terms of trade improvement after a negative shock is sharper when the goods a and b
are worse substitutes. While in the special case perfect substitutes, the terms of trade
hedge vanishes.
Panel C of Figure 8.4 shows the price of physical capital q for the three different
values of s. According to the Tobin’s q condition, a higher capital price q translates to a
higher investment rate. In perfect substitute case the price stands as a constant.
Stationary distribution. Figure 8.5 characterizes the stochastic dynamic of state vari-
able η A . The drift of η A (Panel B) reveals that the system has a basin of attraction at
η A = 0.5. Whenever the system falls below η A = 0.5, the positive drift pushes it back
163 / 176
towards η A = 0.5. Similarly, for η A > 0.5, the system drifts back to η A = 0.5 as the
drift is negative in the range η A ∈ (0.5, 1). In addition, Panel B reveals that the drift
is high for low, but not extremely low, η A values. Competition is depressed and prof-
its (corrected for legacy losses) are high. This empirical effect is called Phoenix miracle.
The Phoenix miracle vanishes for very extreme η A values, when the poor economy is
extremely impaired.
Figure 8.5: Panel A plots the stationary distribution. Panel B the drift, Panel C the
volatility of wealth share η A for three different levels of elasticity of substitution:
s = 0.5 in dashed blue, s = 1.01 (Cobb-Douglas) in solid black, and s = ∞ in dot-
ted magenta.
Like Section 3.3.1, we solve Kolmogorov Forward Equation to get stationary distri-
bution. Panel A plots the stationary distribution for η A ∈ [0, 0.5]. Two features stand
out. First, lower output good substitutability leads to a tighter distribution of wealth
shares. For smaller s the “terms of trade hedge” ensures that countries are better in-
sured against redistributive shocks despite the fact that no risky claims can be traded
in international markets. In the limit as s → ∞ and two goods become perfect substi-
tutes, the stationary distribution becomes degenerate with atoms only at η A = 0 and
164 / 176
η A = 1.21 Second, unlike Cole and Obstfeld (1991) the Cobb-Douglas case does not
lead to full insurance nor to the first best result in our setting. And the stationary dis-
tribution does not degenerate to an atom. There are two reasons for the difference: 1)
in Cole and Obstfeld (1991) a positive productivity shock for country A is like a posi-
tive endowment shock of good a. Supplying more a good worsens country As terms
of trade. In our setting the terms of trade hedge is less pronounced, since the shock
loads on the capital good K. After a positive shock, agents in country A sell off some of
the (composite) capital goods instead of only selling off output good a. 2). in our our
economy capital is persistent and hence a productivity shock has long-lasting implica-
tions, while in Cole and Obstfeld (1991) capital depreciates fully in each period. Our
results show that their resolution of the international diversification puzzle in the case
of Cobb-Douglas preferences is particular to their setting.
8.2.3 Closed Capital Account: Capital Controls
Finally, we consider the case in which the capital account is closed or macropruden-
tial regulation prevents borrowing. Agents in the economy cannot tap in to the interna-
tional debt and equity markets but can still trade goods a and b as well as physical cap-
ital. In this case the wealth shares (η A , η B ) equals to the capital shares (κ A , κ B ). Instead
of solving price-taking planner’s problem (8.16), the capital allocation and risk alloca-
tion can be determined by solving static intermediate goods market equilibrium22 . The
factor allocation can be visualized as following:
21 Recall that the first-best complete-market solution implies a degenerate stationary distribution that
is concentrated at the initial η A .
22 See the step 1 of 8.2.2
165 / 176
E/C
.8 1 − B< .
Terms of trade hedge A? /A= = A? /A= = A? /A= =
. B< .8
! <? = ! < = B < , ! <= = 0 ! @= = ! @ = 1 − B < , ! @? = 0

. C (1 − B < ) + .8 C B <
<?
! =! =B < < .8 C 1 − B < + .8 CDE .B <
Capital allocation ! @? = ! <? =
.(. CDE + .8 CDE ) ! @= = ! @ = 1 − B < .(. CDE + .8 CDE )
.
8 CB< + . 8 CDE .(1 − B < ) ! <= = ! @? = 0 . CB< + .8 C (1 − B < )
! @= = ! <= = CDE
.(. CDE + .8 CDE ) .(. + .8 CDE )
.C .8 C
B? = B= =
.C + .8 C . C + .8 C
B<
0
1
B < 2 < + 2 4< B < 2 4@ (1 − B < )2 4@ @@ (1 − B < ) 2 @ + 2 4@
Risk allocation * << = , * <@ = F@ , * @< = ,* =
2 F< 2 2 F< 2 F@
Figure 8.6: Capital allocation, risk allocation and terms of trade hedge for country A
and B under capital controls
4
Following similar steps as Section 8.2.2, one can solve the equilibrium under closed
capital account.23 We applied the same parameter values as previous case to ease com-
parison. In Figure 8.7 and 8.8 we stress the difference between outcomes under open
and closed capital accounts.
Panel A of Figure 8.7 shows the difference in capital allocation. For equal wealth
shares, i.e., for η A = 0.5, agents are fully specialized in both cases. However, as η A de-
clines, production in country A falls faster under closed capital account than it does un-
der the open capital account. The reason is that firms in country A can issue short-term
debt to agents in country B under open capital account. Besides, κ Aa is significantly
higher in the case without capital controls. And agents in country B start producing
the good a much sooner at lower η A values without capital controls, i.e. ηopen
a a
< ηclose .
As a result, the economy with open capital accounts exhibits a higher degree of special-
ization than the economy with closed capital accounts.
Panel B shows the difference in the terms of trade, i.e. P a /Pb . Without capital
controls agents in country A can borrow in order to hold a larger capital fraction κ A ,
23 Note now each country has its own risk-free rate.
166 / 176
resulting in greater output of good a. This undermines the terms of trade improvement
that occurs under strict capital controls.
Cutting down on debt financing also pushes down the price of physical capital q, as
shown in Panel C of Figure 8.7. This lowers the net worth. In terms of wealth shares, it
hits the country which is levered further. As pointed out above the loss liquidity spiral
amplifies the initial shock even further.
Figure 8.7: Panel A, B and C contrast the capital shares κ Aa and κ Ba , the terms of trade
P a /Pb and the price of physical capital q of an economy without and with capital con-
trols (black solid versus red dashed curves) assuming an elasticity of substitution of
s = 1.01 (close to Cobb-Douglas)
Figure 8.8 shows the stability profile of an economy without and with capital con-
trol. The debt financing induces much thinner stationary distribution in Panel A, larger
Phoenix miracle effect in Panel B, and higher volatility of η A in Panel C. Overall, debt
financing increases specialization, it leads to better allocation of resources and boosts
economic growth in normal times. However, it comes at the price of reduced economic
stability.
167 / 176
Figure 8.8: Panel A plots the stationary distribution. Panel B the drift, Panel C the
volatility of wealth share η A for the global economy without and with capital controls
(black solid versus red dashed curves) assuming an elasticity of substitution of s = 1.01
(close to Cobb-Douglas)
8.3 Welfare Analysis
As Figure 8.7 and 8.8 show, while capital controls can improve risk sharing through
the terms-of-trade hedges, they can also lower average economic growth. In our incomplete-
market setting, capital controls interact with many pecuniary externalities. In this sec-
tion we conduct a formal welfare analysis to see the performance of capital controls.
Pecuniary Externalities. Multiple pecuniary externalities lead to inefficiency in our

incomplete-market setting. The analysis of pecuniary externalities help us to identify
the source of constrained inefficiency of the market outcome implicitly.
First, pecuniary externalities work through the price of capital q. The debt financing
allows high leverage ratio of a country and consequently increase the sensitivity of the
price q to shocks, as fire sale externality.
Second, pecuniary externalities can be induced by the price of output goods. Firms
168 / 176
in a country can affect goods prices in two ways: 1) when overcapitalized, they de-
cide at what point to start competing with the firms in the other country, 2) when un-
dercapitalized, they set their scale of production choosing the amount of international
borrowing. When a firm decides to start producing the good for which it has less ex-
pertise, it takes output prices as given. This ruins the terms of trade hedge for the firms
in the other county as the ratio P a /Pb reaches a/a or a/a. Note Capital controls have
no impact on this pecuniary externality.
Moreover, another pecuniary externality stems from firms’ borrowing decisions to

scale up their operation. The borrowing after an adverse undermines the terms of
trade hedge. If borrowing is limited, competition is reduced and profit margins rise,
which can lead to higher profits and help firms recapitalize themselves more quickly.
Of course, the dose of the borrowing limits has to be right; limits that are too draconian
are counterproductive. This is the case if higher profit margins do not offset the losses
from the decreased volume, as it is the case when foreign firms enter to compete with
the domestic firms. 24
Welfare Calculations. To figure out the extent of the inefficiency and the welfare im-
pact of various policy measures we have to conduct an explicit welfare analysis. Policy
measures that reduce uncertainty can be welfare enhancing even when they result in
a slightly lower growth rate. Hence, we calculate the value functions reflecting the
discounted future expected utility stream for both types of agents.
Under logarithmic utility, the value functions take the simple form
log nt
V ( n t , ηt ) = + h ( ηt ),
ρ
where the wealth-independent term h(ηt ) depends on the agent’s investment oppor-
tunities, summarized by the state variable ηt . Its exact form depends on the market
24 That is, capital controls can improve risk sharing from terms-of-trade hedges only when specialized
domestic firms do not have viable foreign competitors
169 / 176
frictions. For country A the value function can be decomposed into
log NtA log(ηtA Kt qt ) log ηtA log Kt log q(ηtA )

V A ( NtA , ηtA ) = + h(ηtA ) = + h(ηtA ) = + + + h(ηtA )
ρ ρ ρ ρ ρ
| {z }
:= H (ηtA )
For clarify, we normalize the initial size of the economy to K0 = 1 and also drop out the
second term. Besides, we combine the third and fourth terms to H (η A ) for simplicity.
Proposition 8.325 shows that H (η A ) has to solve a second order ordinary differential
equation.
Proposition 8.3. In the competitive equilibrium, H (η A ) = log(q(η A ))/ρ + h(η A ) satisfies

the following second order differential equation:
ηA ηA A ηAB
A Aµt (σt )2 + (σt )2 Φ(ι) − δ
ρH (η ) = log(ρq(η )) + − +
ρ 2ρ ρ
2 2
κ Aa + κ Ab (σ A )2 + κ Ba + κ Bb (σ B )2 ηA A ′ A
ηA A ηAB
(σt )2 + (σt )2 A 2 ′′ A
− + µt η H (η ) + (η ) H (η )
2ρ 2
For agents B an analogous ODE applies, and their value function is given by V B ( NtB , ηtA ) =
log NtB /ρ + H (1 − ηtA ) in the symmetric case σ A = σ B = σ.
25 See the proof in Brunnermeier and Sannikov (2015)
170 / 176
Figure 8.9: Welfare frontier for agents in country A (x-axis) and agents in country B
(y-axis). Outer dashed-dotted green frontier: the first-best solution. Black frontier: the
market equilibrium frontier without capital controls. Red dashed curve: frontier under
strict capital controls. Circles on the frontiers: the switching points, (η a , η b ). (top-right:
zoom-in profile around η A = 1/2)
Figure 8.9 plots the frontier of value functions. The first best frontier (dashed-
dotted) is the Pareto frontier, hence it is strictly decreasing. The black frontier (incom-
plete markets without capital control) is inward bending when η A close to 0 or 1. That
is, for low enough values of η A an unanticipated wealth transfer from agents in country
B to agents in country A, e.g. in form of a bailout or debt relief program, can make both
types of agents better off (So does the symmetric case.). The intuition for this feature is
that with extreme wealth inequality it would be better to distribute wealth to the poor
country. Firms in the poor country increase their output which lowers their output
price. Agents from the richer country benefit from the lower output price, justifying
the initial wealth transfer.26
Figure 8.9 reveals that capital controls are welfare reducing except for a small range
of η A around 1/2 (see top-right zoom-in profile). To make it clear, we plot the sum of
both countries’ welfare with regard to CDF of η A in Figure 8.10.27 . Figure 8.10 takes the
26 Sinceeach individual rich agent takes prices as given he would be unwilling to do such a transfer,
Only a government can coordinate such a transfer that internalizes the pecuniary externalities.
27 Recall any monotone transformation of η A would have been mathematically an equally good state
171 / 176
relative frequency of η A into account and shows that initial η A values for which capital
controls reduce welfare are rare events. Besides, the values of η A for which the value
frontier Figures 8.9 is inward bending and unanticipated debt relief program should be
enacted are even further down in the extreme tails of the stationary distribution.
Figure 8.10: Sum of both countries’ welfares vs. CDF of η A , s = 1.
8.4 Sudden Stops and Runs
Sudden stops are drastic drops in credit flows. Strictly speaking, we have to intro-
duce jumps in the processes of price and capital to study stops/runs. However, here
we focus on the unanticipated jumps of the price processes, which can happen if the
economy is in a “vulnerable region” within in which multiple equilibria can arise. See
Mendo (2019) for a recent paper that models endogenous jump risk explicitly within a
banking crisis setting.
In our setting, sudden stops can arise (i) when a negative fundamental shock triggers
a larger percentage decline in debt than net worth, and (ii) when a sunspot triggers a
sudden capital price drop.
Fundamental induced sudden stops. Here sudden stops refer to situations where
∂κ Aa κ Aa
leverage ratio κ Aa /η A declines with η A , i.e., when ∂η A
> ηA
. In our model, sudden
variable to use. And as the CDF of η A is uniformly distributed, one can easily integrate the areas be-
tween the curves. Note that since the stationary distribution differs across these three cases, so does the
transformation of η A .
172 / 176
stops occur naturally on the equilibrium path because of amplification of exogenous

shocks. The debt28 amplification equation is
 
κ A
t A
 
( 1 − )
 

"
 ′ A Aa A A
#
A
η t


q(κ Aa −1)
 q ( ηt ) ∂ (κ − η ) η ηt 
σt = 1+ A A
+ A Aa A A A ′ A
σA.
q(ηt )/ηt ∂η κ − η κ −η q ( ηt ) 
1− t A t

 
 
q(ηtA )/ηtA 
 
 ηt
∂κ Aa κ Aa ∂(κ Aa −η A ) η
If ∂η A
> ηA
, the “leverage ratio sensitivity” factor ∂η A κ Aa −η A
is greater than 1.
Whether sudden stops can appear is driven by the market illiquidity of capital q′ (η ),
which in turns depends on the technological illiquidity, ϕ. We study the following three
situations
ϕ Technological Illiquidity Sudden Stops

0 Perfect No
2 Imperfect No
2 × 1{ι≥0} + 100 × 1{ι<0} Downward Rigidity Yes
As we can see from Figure 8.11, sudden stops only happen in the last situation. This
∂κ Aa κ Aa
is because the inability to disinvest creates a kink in q(η A ), which allows ∂η A
> ηA
in
the nearby region.
28 The total debt of country A is measured by qt (κtAa − 1).
173 / 176
Figure 8.11: Fundamental induced sudden stops.
Sunspot induced sudden stops. Sunspots can trigger another type of sudden stops
when market illiquidity is sufficiently high, i.e., when q(η ) is steep enough. Absent any
fundamental shocks, a discontinuous jump wealth share is accompanied by a sunspot
jump in the price of capital. In other words, the run can be entirely self-fulfilling.
Intuitively, when a sunspot triggers a jump in q, each agent A shrinks his balance
sheet by fire-selling assets and repaying his debt. The price drop erodes all the net
wort of all agents in country A. Due to pecuniary externality, all agents from the same
country jointly erode the price capital good with fire-sales and impose negative exter-
nality on another. The spillover effects work through the price qt , which each agent A
takes as given. Note the difference from a classic bank run in which lenders withdraw
funding and cause a crunch in credit supply. Here the decline in total credit is a credit
demand effect. Borrowers cut back since they are worried about downside risks and
about hitting the insolvency boundary condition where net worth is zero.
Formally, when a sunspot induces a price drop from q to q̃, the net work of each
agent drops to N + (q̃ − q)κ Aa K. Hence, wealth share drops to
N + (q̃ − q)κ Aa K ηq + (q̃ − q)κ Aa

η̃ (q̃) = = . (8.39)
q̃K q̃
174 / 176
BIBLIOGRAPHY
If q(η ) is sufficiently steep and the initial debt level q(η )[κ Aa − η ] is sufficiently large,
the hyperbola (8.39) crosses q(η ) at two point – the initial (q, η ) and the second sunspot
equilibrium (q̃, η̃). Therefore, there exist multiple equilibria for a range [η, η ] where the
economy is vulnerable to sudden stops.
Bibliography
Brunnermeier, Markus K. and Yuliy Sannikov, “International Credit Flows and Pecu-
niary Externalities,” American Economic Journal: Macroeconomics, January 2015, 7 (1),
297–338.
Cole, Harold L. and Maurice Obstfeld, “Commodity trade and international risk shar-
ing: How much do financial markets matter?,” Journal of Monetary Economics, 1991,
28 (1), 3 – 24.
Dávila, Eduardo and Anton Korinek, “Pecuniary Externalities in Economies with Fi-
nancial Frictions,” The Review of Economic Studies, 02 2017, 85 (1), 352–395.
Geanakoplos, John and Heraklis Polemarchakis, “Existence, Regularity, and Con-

strained Suboptimality of Competitive Allocations When the Asset Market Is Incom-
plete,” in Walter P. Heller, Ross M. Starr, and David A. Starrett, eds., Uncertainty,
Information and Communication: Essays in Honor of Kenneth J. Arrow, Vol. 3, Cambridge
University Press, 1986, pp. 65–95.
Greenwald, Bruce C. and Joseph E. Stiglitz, “Externalities in Economies with Imper-

fect Information and Incomplete Markets,” The Quarterly Journal of Economics, 1986,
101 (2), 229–264.
Mendo, Fernando, “Risk to Control Risk,” Working Paper, 2019.
Newbery, David M. G. and Joseph E. Stiglitz, “Pareto Inferior Trade,” The Review of
Economic Studies, 01 1984, 51 (1), 1–12.
175 / 176
BIBLIOGRAPHY
Ostry, Jonathan David, Atish R. Ghosh, Karl F Habermeier, Marcos d Chamon, Mah-
vash S Qureshi, and Dennis B. S. Reinhardt, “Capital Inflows; The Role of Con-
trols,” IMF Staff Position Notes 2010/04, International Monetary Fund February
2010.
Stiglitz, Joseph E., “The Inefficiency of the Stock Market Equilibrium,” The Review of
Economic Studies, 1982, 49 (2), 241–261.
176 / 176

ECO529 Notes

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

ECO529 Notes

Uploaded by

Copyright:

Available Formats

Lectures on Macro, Money, and Finance

A Heterogeneous-Agent Continuous-Time Approach1

Markus K. Brunnermeier2 Sebastian Merkel3

August 31, 2021

Part I Real Models with Financial Frictions

2 A Simple Heterogeneous Agents Model 8

3 Endogenous Risk Dynamics 29

4 A Model with Jumps 67

Part II Monetary Models with Aggregate and Idiosyncratic Risk

5 A One-Sector Monetary Model with Idiosyncratic Risk 82

6 Cash versus Cashless Economy and the I Theory of Money 106

7 Welfare and Optimal Monetary Policy 136

Part III International Models

8 A Symmetric International Model with Runs / Sudden Stops 139

8.1 Model Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

i Sector type i (experts, households...)

1.1 The Three Watershed Moments in Macroeconomics

1.2 Fire sales, Pecuniary Externalities, Strategic Comple-

1.3 Continuous-time Modeling with Itô Processes

Real Models with Financial Frictions

A Simple Heterogeneous Agents

2.1 Model Setup

Household. Denote household’s consumption rate by cth . They maximize

Figure 2.1: Balance sheets of experts and households

2.2 Solution Method

The following results hold for geometric Itô processes Xt , Yt where

2.2.0 Postulate aggregates, price processes and obtain return processes

In general, the aggregate capital stock is obtained by

Dividend Yield Capital Gain

where rt is the risk-free rate.6

2.2.1 For given SDF processes, derive individual equilibrium condi-

The first-order condition yields the classic Tobin’s q equation:

For simplicity, we will use a special functional form Φ(ι) = 1

(cet )−γ = ∂n Vte ,

Note the last condition is exactly the Tobin’s q condition.

7 We will verify if in section 2.2.4.

• Optimal investment choice (2.4).

Similarly we can solve the much simpler household’s problem:

• Optimal consumption choice ch = ρnh .

Pontryagin’s stochastic maximum principle. Pontryagin’s maximum principle is a

Stochastic maximum principle. Consider a control problem

dXt = µ( Xt , At )dt + σ( Xt , At )dZt ,

with terminal condition p T = G ′ ( XT ).

The first-order condition w.r.t. cet , θte , ιet are

e−ρt (cet )−γ = ξ te (2.11)

In addition, now the dynamics of costate ξ te follows BSDE (2.10),

then drift of ξ te is given by

so the FOCs (2.5)-(2.6) and (2.11)-(2.12) are equivalent.

The martingale approach. Now we introduce the martingale approach, a powerful

Martingale approach in discrete time.

s.t. θt pt = θt−1 ( pt + dt ) − ct ∀t,

i.e., the process ξ t ptA is a martingale.

Martingale approach in continuous time.

Consider a similar portfolio choice problem in continuous time:

As in (2.2), assume that the SDF follows

Using Itô’s product rule,

Since ξ t xtA follows a martingale, its drift equals zero, i.e.,

Example 1. For risk-free asset, σtA = 0. Hence, rtF = rt .

B be the risk-free debt, the following equation holds for experts14

Plugging in the return equation (2.1), we have

In the implementation of Stochastic Maximum Principle, we keep the stochastic pro-

14 We use µ X and σ X to denote the drift and volatility of variable X.

Martingale approach works best in endowment economy, where the consumption

2.2.2 Evolution of state variable ηt