Online Appendix:

*
The Four Equation New Keynesian Model
Eric Sims Jing Cynthia Wu
Notre Dame and NBER Notre Dame and NBER
Ji Zhang
Tsinghua PBCSF

The online appendix contains the following sections to supplement Sims et al. (2021):

A: The Full Non-Linear Model: Here we show the full set of non-linear equilibrium conditions
underlying our linearized four equation model.

B: Details of the Linearized Model: Here we show the full set of linearized equilibrium condi-
tions of the model, and discuss how to reduce these down to the four equations presented in the text.

C: Determinacy of Interest Rate Rules: Here we derive conditions on the parameters of policy
rules necessary for a unique rational expectations equilibrium.

D: Potential, Flexible Price, and Efficient Output: Here we derive expressions for potential,
flexible price, and efficient output in our four equation model.

E: Model Calibration: Here we provide more detail on the baseline calibration of our four equa-
tion model.

F: Robustness: Here we compute impulse responses to shocks under a couple of different specifi-
cations of our four equation model.

G: Optimal Monetary Policy: Here we derive optimal equilibrium conditions for a central bank
operating under discretion and provide formal proofs for some of the theorems and lemmas in the
paper.

H: Additional Results: Here we show a few additional quantitative results related to optimal
policy.

* This material is based upon work supported by the National Science Foundation under Grant No. SES-
1949107. Correspondence: [email protected], [email protected], [email protected].
A The Full Non-Linear Model
This appendix describes the full set of non-linear equilibrium conditions of the model.
The optimality condition for retail firms may be re-written in stationary terms by defining
p∗,t = P∗,t /Pt , x1,t = X1,t /Pt , and x2,t = X2,t /Pt−1 :
 x1,t
p∗,t = (A.1)
 − 1 x2,t
x1,t = pm,t Yt + φ Et Λt,t+1 Πt+1 x1,t+1 (A.2)
x2,t = Yt + φ Et Λt,t+1 Π−1
t x2,t+1 (A.3)
The aggregate inflation rate evolves according to:

1−
1 = (1 − φ)p∗,t + φΠ−1
t (A.4)
Price dispersion evolves according to:

vtp = (1 − φ)p−  p
∗,t + φΠt vt−1 (A.5)
Define lowercase variables as real values of nominal bonds, i.e. bt = Bt /Pt . The balance sheet
condition of the FI may be written:

Qt bFt I + ret = st + XtF I (A.6)

The leverage constraint of the FI may be written:

Qt bFt I ≤ Θt X̄ F I (A.7)
The central bank’s balance sheet can be written:

Qt bcb
t = ret (A.8)
Similarly, the market-clearing condition for long term bonds in real terms is:

bt = bFt I + bcb
t (A.9)
The auxiliary QEt variable is just the real value of the central bank’s long bond portfolio:

QEt = Qt bcb
t (A.10)
Under our assumption on the transfer from parent to child, the consumption of the child may
be written:

Cb,t = Qt bt (A.11)
At and Θt obey stationary AR(1) processes, where the non-stochastic steady state value of
productivity is normalized to unity and Θ denotes the non-stochastic steady state value of leverage.

ln At = ρA ln At−1 + sA εA,t (A.12)

ln Θt = (1 − ρθ ) ln Θ + ρθ ln Θt−1 + sθ εθ,t (A.13)
To close the model, it is necessary to specify rules for the policy rate and the central bank’s

1
bond holdings. For the analysis in Subsection 2.3, we assume that the policy rate is set according
to a Taylor rule and that the central bank’s bond holdings obey an exogenous AR(1) process:

ln Rts = (1 − ρr ) ln Rs + (1 − ρr ) φπ (ln Πt − ln Π) + φx (ln Yt − ln Yt∗ ) + sr εr,t

 
(A.14)

ln QEt = (1 − ρq ) ln QE + ρq ln QEt−1 + sq εq,t (A.15)

In (A.14)-(A.15), Rs and QE denote the non-stochastic steady state values of the policy rate and
the central bank’s balance sheet, respectively.
Yt∗ can be found by solving the full system of equations assuming price flexibility (φ = 0) where
the credit shock is constant, i.e. Θt = Θ. Taking Yt∗ as given, the output gap is then:

ln Xt = ln Yt − ln Yt∗ (A.16)
The optimality conditions for the parent household, (2.5)-(2.7); the definition of the return on
the long bond, (2.12); the optimality conditions for the child, (2.13)-(2.14); the optimality conditions
for the FI, (2.19)-(2.20); the balance sheet condition for the FI and the leverage constraint re-written
in real terms, (A.6)-(A.7); the labor demand condition for the wholesale firm, (2.27); the optimal-
ity condition for optimal price-setting for retailers, re-written in stationary form, (A.1)-(A.3); the
market-clearing condition and aggregate production function, (2.30)-(2.31); the central bank’s pol-
icy rule for Rtre , (A.14); the central bank’s balance sheet and definition of the QEt variable, (A.8)
and (A.10); the central bank’s QE rule, (A.15); the consumption of the child, (A.11); the bond
market-clearing condition, (A.9); the evolution of inflation and price dispersion, (A.4)-(A.5); the
definition of theoutput gap, (A.16); and the exogenous processes (A.12)-(A.3) constitute twenty-
seven variables, L s b FI re
t , Ct , wt , Λt−1,t , Rt , Πt , Λb,t−1,t , Rt , Qt , bt , Θt , ret , st , Ωt , Rt , p∗,t , x1,t , x2,t , pm,t , Yt , At , Cb,t ,
p
vt , bcb,t , bt , QEt , Xt in twenty-seven equations.

B Details of the Linearized Model

This Appendix provides details of the linearization of the non-linear model, the equilibrium condi-
tions for which are given in Appendix A. Where possible, lowercase variables denote log deviations
from steady state, e.g. θt = ln Θt − ln Θ. Where the corresponding level variable is already low-
ercase, a “hat” is put atop the relevant variable to denote a log deviation from steady state, e.g.
pbm,t = ln pm,t − ln pm . Variables without a time subscript denote non-stochastic steady state val-
ues. The model is linearized about a steady state with zero trend inflation (i.e. Π = 1) where the
leverage constraint on intermediaries binds. The complete list of linearized equilibrium conditions
are as follows:

χlt = −σct + w
bt (B.1)
λt−1,t = −σ(ct − ct−1 ) (B.2)
0 = Et λt,t+1 + rts − Et πt+1 (B.3)
λb,t−1,t = −σ(cb,t − cb,t−1 ) (B.4)
κ
rtb = b qt − qt−1 (B.5)
R
b
0 = Et λb,t,t+1 + Et rt+1 − Et πt+1 (B.6)

2
 qt + bbFt I = θt (B.7)
 FI
Qb (1 − κ) qt + QbF I bbFt I − κQbF I bbFt−1
I
+ κQbF I πt + re · re

b t = s · sbt (B.8)
Rb b Rs s
Et λt,t+1 − Et πt+1 + Et rt+1 − r = ωt (B.9)
sp sp t
rtre = rts (B.10)
b1,t − x
pb∗,t = x b2,t (B.11)
b1,t = (1 − φβ)b
x pm,t + (1 − φβ)yt + φβ Et λt,t+1 + φβ Et πt+1 + φβ Et x
b1,t+1 (B.12)
b2,t = (1 − φβ)yt + φβ Et λt,t+1 + ( − 1)φβ Et πt+1 + φβ Et x
x b2,t+1 (B.13)
w
bt = pbm,t + at (B.14)
(1 − z)ct + zcb,t = yt (B.15)
vbtp + yt = at + lt (B.16)
vbtp = 0 (B.17)
1−φ
πt = pb∗,t (B.18)
φ
qt + bbcb
t = re
bt (B.19)
FI cb
bbt = b bbF I + b bbcb (B.20)
b t b t
cb,t = qt + bbt (B.21)
qet = ρq qet−1 + sq εq,t (B.22)
at = ρA at−1 + sA εA,t (B.23)
θt = ρθ θt−1 + sθ εθ,t (B.24)
rtre re
 
= ρr rt−1 + (1 − ρr ) φπ πt + φx xt + sr εr,t (B.25)
qet = re
bt (B.26)
xt = yt − yt∗ (B.27)
bt , λt−1,t , rts , πt , λb,t−1,t , rtb , qt , bbFt I , θt , re
b t , sbt , ωt , rtre ,

This is twenty-seven equations in lt , ct , w
b2,t , pbm,t , yt , at , cb,t , vbtp , bbcb

pb∗,t , x
b1,t , x t , bt , qet , xt in twenty-seven variables.
b
The model can be reduced to the equations presented in Subsection 2.1 as follows. First, (B.10)
can be used to eliminate rtre , so that the policy rule may be written as (2.34) in terms of rts .
Second, (B.11)-(B.13) can be combined with (B.18), which yields the textbook New Keynesian
Phillips curve expressed as a function of marginal cost, where γ = (1−φ)(1−φβ) φ :

πt = γ pbm,t + β Et πt+1 (B.28)

Combining (B.1) with (B.14) and (B.16), making note of the fact that vbtp = 0 around a zero
inflation steady state, yields:

pbm,t = χyt − (1 + χ)at + σct (B.29)

3
 Making use of (B.15) allows us to write this as:

χ(1 − z) + σ σz
pbm,t = yt − (1 + χ)at − cb,t (B.30)
1−z 1−z
Combining (B.19)-(B.21) with (B.26) allows us to write:

bF I   bcb
cb,t = qt + bbFt I + qet (B.31)
b b
Defining b̄F I = bF I /b and b̄cb = bcb /b (i.e. the fraction of total bonds held by financial interme-
diaries and the central bank, respectively, in steady state), and making use of the binding leverage
constraint, (B.7), leaves:

cb,t = b̄F I θt + b̄cb qet (B.32)

Plugging (B.32) into (B.30) then gives:

χ(1 − z) + σ σz h F I i
pbm,t = yt − (1 + χ)at − b̄ θt + b̄cb qet (B.33)
1−z 1−z
Define the hypothetical natural rate of output, yt∗ , as the level of output consistent with flexible
prices and no credit market shocks. That is, yt∗ is the level of output consistent with pbm,t = θt =
qet = 0, or:

(1 + χ)(1 − z)
yt∗ = at (B.34)
χ(1 − z) + σ
But then, using (B.27), we can write marginal cost as:

χ(1 − z) + σ σz h F I i
pbm,t = xt − b̄ θt + b̄cb qet (B.35)
1−z 1−z
Plugging (B.35) into (B.28), defining ζ = χ(1−z)+σ
1−z , yields (2.2).
To derive the IS equation, combine (B.2)-(B.4) with (B.6) and (B.15). Doing so yields:
1−z s z b

yt = Et yt+1 − (rt − Et πt+1 ) − Et rt+1 − Et πt+1 (B.36)
σ σ
But from the Euler equation for the impatient household, along with the “full bailout” assump-
tion embodied in (B.21), we can write:

h i
b
Et rt+1 − Et πt+1 = σ [Et cb,t+1 − cb,t ] = σ b̄F I (Et θt+1 − θt ) + b̄cb (Et qet+1 − qet ) (B.37)

Combining (B.37) with (B.36) yields:

1−z s h i
yt = Et yt+1 − (rt − Et πt+1 ) − z b̄F I (Et θt+1 − θt ) + b̄cb (Et qet+1 − qet ) (B.38)
σ
Note that an alternative, and arguably more intuitive, way to write the IS expression is based
on a simple algebraic manipulation of (B.36):
1 s z b

yt = Et yt+1 − (rt − Et πt+1 ) − Et rt+1 − rts (B.39)
σ σ

4
 (B.39) is the familiar IS/Euler equation, written in terms of output rather than the output gap,
b
appended with a term equal to the long-short interest rate spread, i.e. Et rt+1 − rts .
∗
The natural rate of interest, rt , is defined as the real rate consistent with the IS equation
holding at the natural rate of output absent credit shocks. This implies that:
σ
rt∗ = ∗
− yt∗


Et yt+1 (B.40)
1−z
Adding and subtracting yt∗ and Et yt+1
∗ from both sides of (B.38) and re-arranging yields (2.1).
Making use of (B.23), allows one to write an AR(1) process for rt∗ as in (2.36), where ρf = ρA and
A −1)(1+χ)
sf = σ(ρχ(1−z)+σ .

σ(ρA − 1) ∗
rt∗ = yt (B.41)
1−z
Computing the dynamics of xt , πt , and rts does not require keeping track of yt , yt∗ ,qt , rtb , ωt , sbt ,
bbF I , bbt , bbcb,t or cb,t . Given the solution for xt , πt , and rs , the dynamics of these variables can be
t t
computed using the full system, (B.1)-(B.27).

C Determinacy of Interest Rate Rules

The linearized four equation model as laid out in Section 2 – as captured by the IS equation, (2.1),
the Phillips curve, (2.2), and rules for the policy rate, (2.34), and the central bank’s long bond
portfolio, (2.35) – has exactly the same requirements for equilibrium determinacy as the textbook
three equation model when the central bank’s long bond portfolio is exogenous.
For the purposes of examining equilibrium determinacy, one can treat purely exogenous variable
as constant – e.g. set rt∗ = θt = 0,and set qet= 0 since we are thinking of it as exogenous for now.
s 0
Define a vector of variables, zt = πt xt rt−1 . This vector of variables evolves according to:

Et zt+1 = Azt , (C.1)

where:
   
γζ (1−z)ρr
1−z 1−z
φπ (1 − ρr ) − β −1


1+ σ φx (1 − ρr ) + β σ σ
A= − γζ β −1 . (C.2)
 
β 0
φx (1 − ρr ) φπ (1 − ρr ) ρr
The proof of determinacy follows Woodford (2003), Appendix C. The vector zt includes two
s . For a unique rational expectations
forward-looking variables, πt and xt , and one state variable, rt−1
equilibrium, there must be exactly two unstable eigenvalues of A. The characteristic roots of A
satisfy:

P(µ) = µ3 + A2 µ2 + A1 µ + A0 = 0, (C.3)
where:

5
 A0 = − det(A), (C.4)
1
tr(A2 ) − tr(A)2 ,


A1 = (C.5)
2
A2 = − tr(A). (C.6)

For exactly two unstable eigenvalues, the following conditions must be satisfied:

1 + A2 + A1 + A0 > 0, (C.7)
−1 + A2 − A1 + A0 < 0, (C.8)
A20 − A0 A2 + A1 − 1 > 0. (C.9)

(C.8) is automatically satisfied given assumptions on signs of the parameters. A necessary and
sufficient condition for (C.7) and (C.9) being satisfied is that:

1−β
φπ (1 − ρr ) + φx (1 − ρr ) > 1 − ρr , (C.10)
γζ
or, more compactly:

1−β
φπ + φx > 1. (C.11)
γζ
This is the same condition required for equilibrium determinacy in the three equation New
Keynesian model, subject to the caveat that the slope coefficient in the Phillips curve, γζ, is
slightly different in the four equation model. (C.11) requires that the central bank react sufficiently
aggressively to endogenous variables inflation and the output gap. Intuitively, so long as the central
bank’s long bond portfolio is exogenous, in terms of endogenous variables the four equation model
takes exactly the same form as the three equation model, only with slightly different coefficients on
the output gap in the Phillips curve and on the real interest rate in the IS equation. Hence, it should
not be surprising that the condition on the parameters of the interest rate rule for determinacy is
the same as in the three equation model.
When we  consider an endogenous
0 QE rule, as in Section 4, the system of variables expands to
four: zt = πt xt rt−1s qet−1 . The system of endogenous variables may be written:

B1 Et zt+1 = B2 zt , (C.12)
where:
 zγσ cb 
β 0 0 − 1−z b̄
1−z

σ − z b̄cb (1 − ρq )λπ 1 − z b̄cb (1 − ρq )λx − 1−z
σ z b̄cb (1 − ρq ) 
B2 =  , (C.13)
 0 0 1 0 
0 0 0 1
 
1 −γζ 0 0
 0 1 0 0 
B1 =   (1 − ρr )φπ (1 − ρr )φx ρr 0  .
 (C.14)
(1 − ρq )λπ (1 − ρq )λx 0 ρq

6
 We then have:

Et zt+1 = Azt , (C.15)

where:

A = B−1
1 B2 . (C.16)
As discussed in Section 4, we numerically characterize restrictions on policy rule reaction coef-
ficients for equilibrium determinacy (i.e. for there to be exactly two unstable eigenvalues in A).

D Potential, Flexible Price, and Efficient Output

As noted above, we define potential output, yt∗ , as the level of output consistent with flexible prices
and no credit market shocks. It is given by (B.34). One could alternatively define flexible price
output, ytf , as the level of output consistent with flexible prices only. This would satisfy:

(1 − z)(1 + χ) σz h i
ytf = at + b̄F I θt + b̄cb qet (D.1)
χ(1 − z) + σ χ(1 − z) + σ
The difference between (D.1) and (B.34) is that credit market disturbances – both θt as well as
qet – impact flexible price output but not potential output. If we were to instead define the output
gap as xft = yt − ytf , the Phillips curve representation would be:

πt = γζxft + β Et πt+1 (D.2)

In other words, written in terms of flexible price output, the Phillips curve, (D.2), features no en-
dogenous cost-push wedge related to credit market disturbances. The cost-push wedge comes from
defining potential output as not featuring credit market disturbances in addition to price flexibility.
Credit market shocks drive a wedge between yt∗ and ytf , in a similar way to how markup shocks
are often used to justify a wedge between flexible price and efficient output – and a corresponding
cost-push wedge in the Phillips curve – in the standard three equation model.
To be clear, (D.2) and (2.2) are equivalent ways to describe the supply-side of the model. If
one prefers to work with (D.2), one simply needs to write the IS equation and exogenous processes
in terms of xft , ytf and rtf instead of xt , yt∗ , and rt∗ . We prefer working with the specification in
the text, using our definition of potential output, because it aligns closer to the notion of efficient
output, which is ultimately what is relevant for welfare. Our model features two time-varying
welfare-relevant distortions, not one as in the standard New Keynesian model.
We can motivate our notion of potential output, yt∗ , as coinciding with efficient output under a
particular welfare criterion. Suppose that the planner’s objective is to maximize the utility of the
parent, (2.3), subject to perfectly smoothing the child’s consumption at some constant level, C̄b .
Absent credit market disturbances, even with price stickiness the full bailout assumption would
result in perfectly smooth consumption for the child. Formally, the planner’s problem would be:
∞ 1+χ
" 1−σ #
X C t+j − 1 L t+j
max Et βj −ψ (D.3)
Ct ,Lt 1−σ 1+χ
j=0

s.t.

Ct ≤ At Lt − C̄b (D.4)

7
 With no endogenous states, the planner’s problem is effectively static. The first order condition
is:

ψLχt = Ct−σ At (D.5)

Defining Yte = Ct + C̄b = At Lt , and linearizing about the steady state in which C̄b /Y e = z, one
obtains:

(1 + χ)(1 − z)
yte = at (D.6)
χ(1 − z) + σ
The linearized expression for yte in (D.6) is identical to that for yt∗ , (B.34). In the non-linearized
model, Yt∗ would differ from Yte by a constant owing to the markup of price over marginal cost,
though this constant drops out in the linearization. This steady state distortion could be eliminated
by appealing to a constant Pigouvian subsidy to labor, as is commonplace in the New Keynesian
literature on optimal monetary policy.

E Model Calibration
The parameters of the model are calibrated as follows. The unit of time is a quarter. We assume
a zero trend inflation rate, so Π = 1. This implies that steady state price dispersion is v p = 1 and
the steady state relative reset price is p∗ = 1. We set  = 11, which implies a steady state price
markup of ten percent. The discount factor of the parent is set to β = 0.995, which together with
Π = 1 implies a steady state short term rate of 200 basis points at an annualized frequency (i.e.
Rs = 1.005). We then target a steady state spread of the return on the long bond over the short
term bond of 200 basis points at an annualized frequency, which implies βb = 0.99 and Rb = 1.01.
We set κ = 1 − 40−1 , implying a ten year duration of the long bond. Together with Rb , this implies
a steady state value of Q.
The coefficient of relative risk aversion, σ, and the inverse Frisch elasticity, χ, are both set to 1.
We target a steady state share of child consumption, z = Cb /Y , of one-third. We then pick ψ to
normalize steady state labor input to unity. Together, these parameters imply a value of the steady
state transfer from parent to child, X b . We assume that the Calvo parameter is φ = 0.75, implying
a mean duration between price changes of one year. We assume that the size of the central bank’s
balance sheet is 10 percent of steady state output, i.e. QE = 0.1 × Y . We pick a steady state target
of the risk-weighted leverage ratio of Θ = 5. This then implies a value of the steady state equity
transfer from the parent to the FI, X F I .
For the exercises in Subsection 2.3, we assume that the Taylor rule parameters are ρr = 0.8,
φπ = 1.5, and φx = 0. The autoregressive parameter of the QE process, ρq , is also set to 0.8.
The autoregressive parameters for productivity and the credit shock are also both set to 0.8. This
implies, as shown below in Appendix B, that the AR parameter in the natural rate process is also
0.8.

8
 Table E.1: Parameter Values of Full Model

Parameter Value Description (Target)

β 0.995 Discount factor, parent
σ 1 Inverse elasticity of substitution
χ 1 Inverse Frisch elasticity
ψ 1.36 Labor disutility scaling parameter (target L = 1)
βb 0.99 Discount factor, child (target spread of 200 b.p. annualized)
Π 1 Steady state trend inflation
 11 Elasticity of substitution (target markup ten percent)
κ 1 − 40−1 Coupon decay (target duration ten years)
φ 0.75 Calvo price
Θ 5 Steady state risk-weighted leverage
QE 0.1 Steady state central bank bond portfolio
z 0.33 Steady state child share of consumption
Xb 0.33 Steady state parent-child transfer
XF I 0.046 Steady state parent-FI equity transfer
ρr 0.8 Taylor rule smoothing
φπ 1.5 Taylor rule inflation
φx 0 Taylor rule gap
ρA 0.8 AR productivity
ρθ 0.8 AR credit
ρq 0.8 AR QE

Note: this table lists the values of calibrated parameters for the exercises in Subsection 2.3.

F Robustness
F.1 Full Bailout
In the baseline four equation model, we make a “full bailout” assumption that there is a
complete payoff from the parent household, each period, of the outstanding debt of children
(see (2.32)). This significantly simplifies the model in that it makes the child’s consumption
equal to the value of bonds, (2.33).
As noted in the text, this assumption on the transfer is not critical for the quantitative
characteristics of the model. To see this more cleanly, we assume that the transfer from
parent to child is fixed each period, equal to the steady state version of (2.32). This leaves
the steady state of the modified model identical to our baseline model, but does affect some
dynamics. Figure F.1 shows impulse responses to a natural rate shock, Figure F.2 to a
conventional monetary policy shock, and Figure F.3 to a credit/QE shock with and without
the full bailout assumption. The solid lines are responses with the full bailout and are
identical to the responses in the baseline model shown in the text. Dashed blue lines are
responses in the four equation model with a fixed bailout.

9
 Figure F.1: IRFs to Shock to Potential Output, Fixed vs. Full Bailout
0
0.4

Output Gap
-0.2
0.3
Output

0.2 -0.4

0.1
-0.6
0
0 5 10 15 20 0 5 10 15 20
Horizon Horizon

0 0

Interest Rate -0.2

Inflation

-0.5

-0.4
Full Bailout
-1 Fixed Transfer
-0.6
0 5 10 15 20 0 5 10 15 20
Horizon Horizon
Notes: Black solid lines: IRFs to a one percentage point shock to potential output in the baseline four
equation model. Output and the output gap are expressed in percentage points, while the responses of
inflation and the short term interest rate are expressed in annualized percentage points. Blue dashed lines:
IRFs to the same-sized natural rate shock in the four equation model where there is no full bailout from
parent to child.

10
 Figure F.2: IRFs to Policy Shock, Fixed vs. Full Bailout
0.6 0.6

Output Gap
0.4 0.4
Output

0.2 0.2

0 0
0 5 10 15 20 0 5 10 15 20
Horizon Horizon

1 0
Full Bailout
Fixed Transfer
Interest Rate -0.2
Inflation

0.5
-0.4

0 -0.6
0 5 10 15 20 0 5 10 15 20
Horizon Horizon
Notes: Black solid lines: IRFs to a conventional monetary policy shock in the baseline four equation model.
The size and sign of the shock are chosen to generate the same impact response of output as in Figure F.1.
Output and the output gap are expressed in percentage points, while the responses of inflation and the short
term interest rate are expressed in annualized percentage points. Blue dashed lines: IRFs to the same-sized
policy shock in the four equation model where there is no full bailout from parent to child.

In the case of natural rate and monetary policy shocks, the differences for the responses of
aggregate output, inflation, and the interest rate with or without the full bailout assumption
are inconsequential. In comparison to our baseline model, output reacts slightly less to a
natural rate shock on impact without the full bailout assumption; it reacts slightly more to
a monetary policy shock without the full bailout assumption.

11
 Figure F.3: IRFs to Credit/QE Shock, Fixed vs. Full Bailout

0.4 0.4

0.3 0.3

Output Gap
Output

0.2 0.2

0.1 0.1

0 0
0 5 10 15 20 0 5 10 15 20
Horizon Horizon

0.1
0.2 Full Bailout
Fixed Transfer
0.15 Interest Rate
Inflation

0.05
0.1

0.05

0 0

0 5 10 15 20 0 5 10 15 20
Horizon Horizon
Notes: Black solid lines: IRFs to a credit (θt ) or QE (qet ) shock in the baseline four equation model. The
size and sign of the shocks are chosen to generate the same impact response of output as in Figure F.1.
Because the QE and credit shock only differ according to scale in the linearized model (i.e. b̄F I 6= b̄cb ) and
the AR parameters are the same, the normalized impulse responses are identical. Output and the output
gap are expressed in percentage points, while the responses of inflation and the short term interest rate are
expressed in annualized percentage points. Blue dashed lines: IRFs to the same-sized credit/QE shock in
the four equation model where there is no full bailout from parent to child.

Quantitatively, differences are somewhat more noticeable when it comes to a credit or

QE shock. Compared to the baseline model with the full bailout assumption, output and
inflation react less on impact to a credit/QE shock when there is a fixed transfer. The
responses are also less persistent. But qualitatively, they are in-line with our baseline model.

F.2 Parameterization of z
Our linearized four equation model differs from the standard NK model via the parameter z;
when z = 0, the model is equivalent to the standard three equation model. In this subsection,
we show impulse responses to natural rate, monetary policy, and credit market shocks for
different values of z. Responses with our baseline value of z = 0.33 are depicted via solid
black lines; we also show responses when z = 0.167 (dotted black lines) and when z = 0.67

12
(dashed black lines). For point of comparison, the blue dashed lines shows responses in the
three equation model (z = 0).
In response to all three shocks, the bigger z is, the more the responses differ from the
conventional three equation model. Relative to our baseline case (z = 0.33), the impulse
responses with z = 0.167 or z = 0.667 are not very wildly different.

Figure F.4: IRFs to Shock to Potential Output, Different Values of z

0.6 0

Output Gap
0.4 -0.2
Output

0.2 -0.4

0 -0.6
0 5 10 15 20 0 5 10 15 20
Horizon Horizon

0 0

-0.5
Interest Rate
Inflation

-1 -0.5
z = 0.33
z=0
-1.5 z = 0.17
z = 0.67
-2 -1
0 5 10 15 20 0 5 10 15 20
Horizon Horizon
Notes: Black solid lines: IRFs to a one percentage point shock to potential output in the baseline four
equation model with z = 0.33. Output and the output gap are expressed in percentage points, while the
responses of inflation and the short term interest rate are expressed in annualized percentage points. Blue
dashed lines: IRFs to the same-sized natural rate shock in the three equation model (equivalent to the four
equation model with z = 0). Black dotted lines: responses in the four equation model with z = 0.167. Black
dashed lines: responses in the four equation model with z = 0.667.

13
 Figure F.5: IRFs to Shock to Policy Shock, Different Values of z
0.6 0.6

Output Gap
0.4 0.4
Output

0.2 0.2

0 0
0 5 10 15 20 0 5 10 15 20
Horizon Horizon

1 0
z = 0.33
z=0
Interest Rate -0.2 z = 0.17
Inflation

z = 0.67
0.5
-0.4

-0.6
0
0 5 10 15 20 0 5 10 15 20
Horizon Horizon
Notes: Black solid lines: IRFs to a conventional monetary policy shock in the baseline four equation model
with z = 0.33. The size and sign of the shocks are chosen to generate the same impact response of output
as in Figure F.4 when z = 0.33. Output and the output gap are expressed in percentage points, while the
responses of inflation and the short term interest rate are expressed in annualized percentage points. Blue
dashed lines: IRFs to the same-sized natural rate shock in the three equation model (equivalent to the four
equation model with z = 0). Black dotted lines: responses in the four equation model with z = 0.167. Black
dashed lines: responses in the four equation model with z = 0.667.

14
 Figure F.6: IRFs to Shock to Credit/QE Shock, Different Values of z

0.4 0.4

Output Gap
0.3 0.3
Output

0.2 0.2

0.1 0.1

0 0
0 5 10 15 20 0 5 10 15 20
Horizon Horizon

0.3
0.1 z = 0.33
z=0
0.2 Interest Rate z = 0.17
Inflation

z = 0.67
0.05
0.1

0 0
0 5 10 15 20 0 5 10 15 20
Horizon Horizon
Notes: Black solid lines: IRFs to a credit (θt ) or QE (qet ) shock in the baseline four equation model with
z = 0.33. The size and sign of the shocks are chosen to generate the same impact response of output as in
Figure F.4 when z = 0.33. Because the QE and credit shock only differ according to scale in the linearized
model (i.e. b̄F I 6= b̄cb ) and the AR parameters are the same, the normalized impulse responses are identical.
Output and the output gap are expressed in percentage points, while the responses of inflation and the short
term interest rate are expressed in annualized percentage points. Blue dashed lines: IRFs to the same-sized
natural rate shock in the four equation model where there is no full bailout from parent to child.

G Optimal Monetary Policy

We consider optimal monetary policy under discretion. The central bank picks instruments,
rts and qet , and endogenous variables, πt and xt , to minimize (3.1) subject to the IS equation,
(2.1), the Phillips Curve, (2.2), a ZLB constraint on the policy rate, and a potential constraint
on the use of balance sheet policies. Formally:

min µx2t + πt2

rts ,qet ,πt ,xt

s.t.
1−z s ∗
h
FI cb
i
xt = Et xt+1 − (rt − Et πt+1 − rt ) − z b̄ (Et θt+1 − θt ) + b̄ (Et qet+1 − qet ) , (G.1)
σ

15
 zγσ h F I i
πt = γζxt − b̄ θt + b̄cb qet + β Et πt+1 , (G.2)
1−z
rts = 0, (G.3)
qet = 0. (G.4)
(G.3) only holds when the policy rate is constrained by the ZLB (i.e. the policy rate is
fixed), as in Section 3.2, while (G.4) only binds when QE is unavailable, as in Section 3.3.
In solving the problem under discretion, the policymaker treats expectations of all future
variables as given.
A Lagrangian is:

 i
1−z s ∗
h
L= µx2t +πt2 +λ1,t xt − Et xt+1 + FI cb
(rt − Et πt+1 − rt ) + z b̄ (Et θt+1 − θt ) + b̄ (Et qet+1 − qet )
σ
 
zγσ h F I i
+ λ2,t πt − γζxt + cb
b̄ θt + b̄ qet − β Et πt+1 − λ3,t rts − λ4,t qet .
1−z

The first order conditions are:

∂L 1−z
s
= λ1,t − λ3,t = 0, (G.5)
∂rt σ
∂L cb zγσ b̄cb
= −λ1,t z b̄ + λ2,t − λ4,t = 0, (G.6)
∂qet 1−z
∂L
= 2πt + λ2,t = 0, (G.7)
∂πt
∂L
= 2µxt + λ1,t − γζλ2,t = 0. (G.8)
∂xt

G.1 Both Instruments Available

When both instruments are available, we have λ3,t = λ4,t = 0. From (G.5), this means that
λ1,t = 0. But then from (G.6), λ1,t = λ4,t = 0 implies that λ2,t = 0. But then (G.7) then
requires that πt = 0. Then, from (G.8), we see that xt = 0.

G.1.1 Proof of Theorem 1

The theorem can be proved by contradiction. Suppose first we can achieve xt = πt = 0 with
qet = 0 for all t. Then the Phillips curve may then be written as:
zγσ F I
0=− b̄ θt , (G.9)
1−z
which does not hold unless θt = 0, which contradicts the assumption. Hence, there is a
contradiction.
Second, suppose we can achieve xt = πt = Et xt+1 = Et πt+1 = 0 with rts = 0 for all t.

16
Then the Phillips curve becomes
zγσ h F I cb
i
0= b̄ θt + b̄ qet . (G.10)
1−z
This requires

b̄F I
qet = − θt . (G.11)
b̄cb
Note that (G.11) is identical to the equilibrium path for qet given in Proposition 2. With
the policy rate fixed and the equilibrium path of QE given in this way, the IS curve becomes
1−z ∗ h
FI cb
i
0= r − z b̄ (Et θt+1 − θt ) + b̄ (Et qet+1 − qet ) . (G.12)
σ t
Further apply (G.11) on both qet and Et qet+1 . This implies
1−z ∗
0= r , (G.13)
σ t
which does not hold unless z = 1 (which we have ruled out) or rt∗ = 0 (which contradicts
the assumption). Hence, we have another contradiction.

G.2 The ZLB

G.2.1 First Order Condition
Next, suppose that the ZLB binds, so λ3,t ≥ 0, but QE is available, so λ4,t = 0. We will have
λ1,t > 0. We can solve for λ1,t in terms of λ2,t from (G.6):
γσ
λ1,t = λ2,t . (G.14)
1−z
Then, from (G.8), we have:

2µ(1 − z)
xt = λ2,t . (G.15)
γζ(1 − z) − γσ
Plug (G.15) into (G.7):

2µ(1 − z)
2πt + xt = 0.
γζ(1 − z) − γσ
Simplifying yields:

µ(1 − z)
πt = − xt , (G.16)
γζ(1 − z) − γσ
as shown in the main text.

17
G.2.2 Proof of Proposition 2
When the ZLB is binding, the IS curve becomes
1−z ∗
h
FI cb
i
xt = Et xt+1 − (− Et πt+1 − rt ) − z b̄ (Et θt+1 − θt ) + b̄ (Et qet+1 − qet ) , (G.17)
σ
and the Phillips curve remains as (2.2). (3.4) holds and describes balance sheet policy, and
the policy rate is fixed, i.e. rts = 0.
Guess
πt = ω1 rt∗ + ω3 θt ,
then (3.4) implies

γζ(1 − z) − γσ
xt = − (ω1 rt∗ + ω3 θt ) ≡ ω2 (ω1 rt∗ + ω3 θt ).
µ(1 − z)

If in period t + 1, the economy exits the ZLB, then both instruments are available, which
implies πt+1 = xt+1 = 0. If in period t + 1, the economy is still at the ZLB, πt+1 =
∗ ∗
ω1 rt+1 + ω3 θt+1 , xt+1 = ω2 (ω1 rt+1 + ω3 θt+1 ). As a result,

Et πt+1 = α(ρf ω1 rt∗ + ρθ ω3 θt ),

Et xt+1 = αω2 (ρf ω1 rt∗ + ρθ ω3 θt ).

We also guess the equilibrium path of QE is:

qet = τ rt∗ + ω4 θt .

If in period t+1, the economy exits the ZLB, then everything goes back to the unconstrained
FI
∗
case: qet+1 = − b̄b̄cb θt+1 . Otherwise, qet+1 = τ rt+1 + ω4 θt+1 . Then,

b̄F I
Et qet+1 = αρf τ rt∗ + αρθ ω4 θt − (1 − α)ρθ θt .
b̄cb
Plugging the guesses into (2.2) and (G.17), we have:
zγσ h F I i
ω1 rt∗ + ω3 θt = γζω2 (ω1 rt∗ + ω3 θt ) − b̄ + b̄cb ω4 θt + b̄cb τ rt∗ + βα(ρf ω1 rt∗ + ρθ ω3 θt ),


1−z
(G.18)
1−z
ω2 (ω1 rt∗ + ω3 θt ) = αω2 (ρf ω1 rt∗ + ρθ ω3 θt ) − [−α(ρf ω1 rt∗ + ρθ ω3 θt ) − rt∗ ]
σ
FI cb
h b̄F I i
−z b̄ (ρθ − 1) θt − z b̄ αρθ ω4 − (1 − α)ρθ cb − ω4 θt
b̄
−z b̄cb (αρf − 1)τ rt∗ . (G.19)

18
 Matching coefficients, we have the following system of equations:
zγσ cb
ω1 = γζω2 ω1 − b̄ τ + βαρf ω1 (G.20)
1−z
1−z
ω2 ω1 = αω2 ρf ω1 + (αρf ω1 + 1) − z b̄cb (αρf − 1)τ (G.21)
σ
zγσ F I
b̄ + b̄cb ω4 + βαρθ ω3


ω3 = γζω2 ω3 − (G.22)
1−z
1−z h b̄F I i
ω2 ω3 = αω2 ρθ ω3 + αρθ ω3 − z b̄F I (ρθ − 1) − z b̄cb αρθ ω4 − (1 − α)ρθ cb − ω(G.23)
4
σ b̄
Solve τ as a function of ω1 in the first two equations

(1 − z)(γζω2 ω1 + βαρf ω1 − ω1 )
τ =
b̄cb zγσ
αω2 ρf ω1 + 1−zσ
(αρf ω1 + 1) − ω2 ω1
τ = cb
z b̄ (αρf − 1)

The first equation is the same as (3.10). Then,

αω2 ρf ω1 + 1−zσ
(αρf ω1 + 1) − ω2 ω1 (1 − z)(γζω2 ω1 + βαρf ω1 − ω1 )
cb
=
z b̄ (αρf − 1) b̄cb zγσ

Simplify, we get

γ(1 − z)
ω1 =
(αρf − 1)(1 − z)(γζω2 + βαρf − 1) − γσαω2 ρf − γ(1 − z)αρf + γσω2

This is the same as (3.8).

Next, simplify (G.23)
1−z
ω2 ω3 = αω2 ρθ ω3 + αρθ ω3 + z(1 − αρθ )(b̄F I + b̄cb ω4 )
σ
Therefore,

ω2 − αω2 ρθ − 1−z αρθ

b̄F I + b̄cb ω4 = σ
ω3 (G.24)
z(1 − αρθ )

Similarly, rearrange (G.22)

1−z
b̄F I + b̄cb ω4 = (γζω2 + βαρθ − 1) ω3 (G.25)
zγσ

19
Combining (G.24)-(G.25),

ω2 − αω2 ρθ − 1−z
 
σ
αρθ 1−z
− (γζω2 + βαρθ − 1) ω3 = 0
z(1 − αρθ ) zγσ

This has to hold for any parameter specification. Therefore,

ω3 = 0.

Plug back to (G.24), we get

b̄F I
ω4 = − .
b̄cb


G.3 No QE Available
G.3.1 First Order Condition
When QE is unavailable, we have λ4,t > 0, but λ3,t = 0. The later implies from (G.5) that
λ1,t = 0. From (G.7), we have:

λ2,t = −2πt . (G.26)

Plugging into (G.8), we have:

2µxt + 2γζπt = 0, (G.27)

which implies:
µ
πt = − xt , (G.28)
γζ
as shown in the text.

G.3.2 Proof of Proposition 3

When QE policy is not available, (3.11) holds, and qet = 0. The IS Phillips curves can be
written:
1−z s
xt = Et xt+1 − (rt − Et πt+1 − rt∗ ) − z b̄F I (Et θt+1 − θt ) , (G.29)
σ
and
zγσ F I
πt = γζxt − b̄ θt + β Et πt+1 . (G.30)
1−z
Substitute (3.11) into (G.30)

γ 2ζ 2 zγσ F I
πt = − πt + β Et πt+1 − b̄ θt . (G.31)
µ 1−z

20
Solve the above equation forward, we have

πt = ϕθt (G.32)

where
µ zγσ F I
ϕ=− b̄ . (G.33)
γ 2ζ 2 + µ(1 − βρθ ) 1 − z
Combining the solution for πt and (3.11), we obtain

γζ
xt = − ϕθt . (G.34)
µ

Plug the solutions for πt and xt into (G.29):

(1 − ρθ )σz b̄F I
 
s ∗ σ(1 − ρθ ) γζ
rt = rt + ρθ ϕ + ϕ+ θt . (G.35)
1−z µ 1−z

As a result,
σ(1 − ρθ ) γζ (1 − ρθ )σz b̄F I
η = ρθ ϕ + ϕ+ . (G.36)
1−z µ 1−z


H Additional Results
Figure H.1 plots the optimal QE response to a natural rate shock, τ , as a function of α,
which governs the expected duration of a ZLB episode. These results are discussed in Section
3.2 of the text.

21
 Figure H.1: τ As a Function of ZLB Duration, α

-5

-10

-15

-20

-25

-30
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

Notes: This figure plots the response of QE to a natural rate shock under optimal policy when the ZLB binds
for different values of α. The expected duration of the ZLB is 1/(1 − α), so higher values of α correspond to
higher expected ZLB durations. The figure is generated assuming a value of µ = 1.

22
 Figure H.2 plots the optimal QE response to a natural rate shock, τ , as a function of µ,
the relative weight the central bank attaches to the output gap. These results are discussed
in Section 3.2 of the text.

Figure H.2: τ as a Function of µ

-24

-26

-28

-30

-32

-34

-36

-38

-40

-42
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

Notes: This figure plots the response of QE to a natural rate shock under optimal policy when the ZLB
binds, with α = 3/4, for different values of µ, the welfare weight on the output gap.

23
 Figure H.3 plots the optimal interest rate response to a credit shock, η, in a situation in
which QE is unavailable. We show the plot as a function of µ, the weight the central bank
attaches to the fluctuations in the output gap. For low values of µ, η is positive; for large
values of µ, η is negative. These results are discussed in Section 3.3 of the text.

Figure H.3: η As a Function of µ

0.03

0.02

0.01

-0.01

-0.02

-0.03

-0.04

-0.05
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

Notes: This figure plots η as a function of µ, the welfare weight on the output gap, for the case when qet = 0.

24
References
Sims, Eric, Jing Cynthia Wu, and Ji Zhang, “The Four Equation New Keynesian
Model,” Review of Economics and Statistics, 2021. forthcoming.

Woodford, Michael, Interest and Prices: Foundations of a Theory of Monetary Policy,

Princeton, NJ: Princeton University Press, 2003.

25