Fanelli & Straub (2021) Apéndices
Fanelli & Straub (2021) Apéndices
A Calibration details
A.1 Illustrative calibration
We discipline the non-homotheticity in the model using evidence on tradable expenditure
shares across the income distribution from the 2008/09 Brazilian Household Expenditure
Survey (see details in Appendix A.2 below). We assume the poor household corresponds to
the bottom 20% share of the population, µ = 0.2, while the rich household corresponds to
the remainder. We choose χ to match the implied expenditure shares, giving χ = 0.078. On
average poor households spend 52% of their income on tradables, while rich households
spend 36% of their income on tradables. We adjust these numbers for government spending,
which we for simplicity assume is nontradable and financed by the rich. After making
this adjustment, we obtain that the rich spend 25% on tradables out of their gross income.
These shares directly imply α = 0.051 and c = 0.19. Our results do not depend on the
initial net foreign asset position, so for simplicity, we set it to zero in the steady state. This
implies y T = 0.23. We normalize y NT = 1 without loss of generality.
We calibrate the remaining parameters as follows. For the discount rate we pick
ρ = 0.075, corresponding to the average 5yr treasury yield from 2000–2015 plus the average
J.P. Morgan EMBI+Brazil return over the same period. We set σ = 2, a standard value. To
get an idea of the relative size of home compared to foreign intermediation, we note that
Brazil’s private home banks operate balance sheets roughly three times the size of foreign
banks’ subsidiaries in Brazil (Cull et al., 2018). Interpreting balance sheet size as a rough
proxy for portfolio constraints (corresponding to X in our microfoundation), this leads us
to calibrate Γ H /Γ = 1/3.
To calibrate the overall degree of capital immobility Γ, we rely on recent evidence from
Kohlscheen and Andrade (2014), a high-frequency study of Brazilian FX interventions.
They find that the announcement of a US$1 billion purchase of FX swaps leads to a
depreciation of somewhere between 0.10 to 0.50%. The more conservative lower end of
the spectrum therefore roughly corresponds to a 2% exchange rate movement after a 1%
of GDP purchase of FX swaps (with a 1-year maturity), which is consistent with Γ = 9.27
To explore the robustness of our predictions around this value, we also plot the model
responses for values Γ = 1 and Γ = 50 below. Finally, we assume that the equilibrium
without intervention, i.e. with bGt ∗ = 0 for all t, does not involve interest rate spreads,
A-1
Figure 7: Tradable expenditure shares by total expenditure percentile.
100 %
50 %
25 %
0%
1 2 3 4 5 6 7 8 9 10
Total expenditure deciles
∗
i.e. τt = 0 for all t. This requires a sufficiently wide interval [b∗H , b H ]. We choose the
∗
smallest such interval, i.e. b H = −b∗H = 1.7y T .28 This assumption ensures that households
have sufficient access to international asset markets that they do not need to rely on costly
intermediaries.
A.3 Calibration of Γ
We calibrate the degree of capital immobility in our model, Γ, to match evidence from
Chamon et al. (2017). To do this, we simulate an experiment where the planner increases
its reserve position by 1% of GDP, for one year, for various levels of Γ, corresponding to a
1-year long reserve swap. We then determine the level of Γ for which the initial exchange
rate adjustment is equal to 2%. In our case, this level is given by Γ = 9. We plot the time
series of this simple experiment in Figure 8.
28 Any sufficiently wide interval works here. The only effect a wider (but still bounded) interval has are
constant offsetting shifts in the paths of reserves and home intermediaries’ asset positions. In particular, the
paths of interest rate spreads {τt } and the exchange rate are unaffected.
A-2
Figure 8: Effects of a simple one-time reserve swap intervention.
A-3
B.2 The planner’s objective function and concavity
In this section we derive the expression (20) for the objective function V (zt ) and prove that
it is strictly increasing and strictly concave, and satisfies Inada conditions. We also show
Σ(z) is positive and bounded from above.
Derivation of V (z) By substituting consumption choices (4) and (5) into preferences (1),
we see that per period utility of a type-i household is given by
V i = u (αzit )α ( p−
t
1
( 1 − α ) z i 1− α
t )
Using the fact that poor households are hand-to-mouth ztP = p(z)χy N /µ + χy T /µ − c
and the notation that ztR = zt , per-period utilitarian welfare is then up to a multiplicative
constant
V (z) = (1 − µ)u p(z)−(1−α) z + µu p(z)−(1−α) p(z)χµ−1 y N + χµ−1 y T − c
1 1
where A0 ≡ 1 − µ, B0 ≡ µ, C0 ≡ 1 − α, B1 ≡ χ(1−µ)(1−α)
(µc − χy T ), and C1 ≡ (1− µ )
(µc −
A0 C1
B0 B1
χy T ). In particular, A0 , B0 , C0 , B1 , C1 > 0, C0 < 1, B1 > C1 , and C ≤ 1 (since χ < µ).
1 − B1
1
These properties alone allow us to prove monotonicity and concavity of V (z). Throughout,
bear in mind that the domain of z is ( B1 , ∞).
After some rearranging, we can write the first derivative of V as
V 0 (z)
= A0 ((1 − C0 )z − C1 )
(z − C1 )C0 (σ−1)−1 z−σ
A0 C1 !1− σ −σ
B0 B1 z − B1
+ B0 ((1 − C0 )z − C1 + C0 B1 ) (31)
1 − CB11 z
| {z }
increasing in σ
To show that V 0 > 0, it is without loss to assume σ = 0, since the under-braced term is
increasing in σ and the second term is necessarily positive,
(1 − C0 )z − C1 + C0 B1 > B1 − C1 > 0.
A-4
With σ = 0, V 0 (z) > 0 is equivalent to (after some algebra)
(1 − C0 ) B1 > 0
which holds in light of C0 ∈ (0, 1) and B1 > 0. This completes the proof that V is strictly
increasing.
Inada conditions for V 0 (z) Straight from (31) we see that limz→ B1 V 0 (z) = ∞. Moreover,
note that for large z
V 0 (z) ∼ zC0 (σ−1)−1 z−σ+1 = z−(1−C0 )σ−C0
where the exponent is always strictly negative since σ ≥ 0 and C0 ∈ (0, 1). This proves that
limz→∞ V 0 (z) = 0.
A0 C1 !1− σ − σ −1
V 00 B0 B1 z − B1
= A0 A + B0 B (32)
z−σ−1 (z − C1 )C0 (σ−1)−2 1 − CB11 z
where
A = −zC0 ((1 − C0 )z − 2C1 ) − σ (C1 − z(1 − C0 ))2
and
A0 C1
!1− σ
− σ −1
B0 B1 z− B1
Notice that B < 0 for any σ ≥ 0 and that A, B both decrease in σ. As C z
1 − B1
1
is increasing in σ, it suffices to show that (32) holds for σ = 0, that is,
A0 C1 ! −1
B0 B1 z − B1
A0 A(σ = 0) + B0 B(σ = 0) < 0.
1 − CB11 z
C1
−((1 − C0 )z − 2C1 ) − ((1 − C0 )(z − B1 ) + 2( B1 − C1 )) < 0
B1 − C1
Notice this expression is decreasing in z, so it suffices to show the result when z = B1 , i.e.
−C0 (1 − C0 ) B1 < 0,
which is true by virtue of C0 ∈ (0, 1) and B1 > 0. This completes the proof that V is strictly
concave.
A-5
Σ(z) is positive, continuous, and bounded from above Using the same notation as
before, Σ(z) can be rewritten as
z
Σ(z) = σ + (1 − α)(1 − σ)
z − C1
1−α
Σ( B1 ) = σ + (1 − σ)
1 − χ (1 − α )
Σ(∞) = σ + (1 − σ)(1 − α) > 0.
We assume parameters are such that Σ( B1 ) > 0 in our baseline model. This is satisfied in
our calibration.
Lemma 1. Assume Γ ∈ (0, ∞), V (z) is an increasing and strictly concave function that satisfies
Inada conditions, and Σ(z) is a positive and continuous function bounded from above, i.e. ∃K0 > 0
such that Σ(z) < K0 ∀z. When {rt∗ } follows the path in (21), any solution {zt , τt } to the planning
problem with objective (19) and constraints (18a) and (18b) satisfies:
1. τ0 = limt→∞ τt = 0
2. τt is continuous in t
3. τt is differentiable at ∆r ∗ = 0
−V 00 (z)z
4. if V 0 (z)
> Σ(z) for all z, then sign(τt ) = −sign(∆r ∗ ) for all t ∈ (0, ∞)
−V 00 (z)z
5. if V 0 (z)
= Σ(z) for all z, τt = 0 for all t ∈ (0, ∞)
−V 00 (z)z
6. if V 0 (z)
< Σ(z) for all z, then sign(τt ) = sign(∆r ∗ ) for all t ∈ (0, ∞)
Proof. First notice that the inequality in (18b) can be relaxed to be ≤. This is because zt can
always be scaled up by a constant factor, which increases welfare while leaving (18a) intact.
First order conditions. The current value Hamiltonian of the planning problem is given
by
Rt ∗ α (1 − µ ) 1 z
H (z, τ, λ, ψ, t) = e 0 (rs − ρ ) ds V (z) − λ z − λ τ2 + ψ (r ∗ + τ − ρ ).
1 − χ (1 − α ) Γ Σ(z)
A-6
This is an optimal control problem with a subsidiary condition, as in Gelfand and Fomin
(1963). The state variable is z and has a free initial value z0 . z has a (continuously dif-
ferentiable) costate ψ. λ > 0 is the multiplier on the resource constraint (18b), and τ is
the control variable. For convenience, we define ψ̃ ≡ ψ Σ(zz) , which is also continuously
Rt ∗
differentiable, and It ≡ e 0 (rs −ρ)ds .
Γ
H is concave in τ, with optimum at τt = 2λ ψ̃t , and the costate equation for z is
zt
ψ̃˙ t = rt∗ ψ̃t + F (zt , It )
Σ(zt )
α (1− µ )
where we defined F (zt , It ) ≡ λ 1−χ(1−α) − It V 0 (zt ). Moreover, since z has a free initial
value, ψ0 = ψ̃0 = 0. From this, it already follows that τ0 = 0 and τt is continuously
differentiable in t.
We can rewrite the optimality conditions directly in terms of τt and z,
Γ zt
τ̇t = rt∗ τt + F (zt , It ) (33a)
2λ Σ(zt )
żt
Σ(zt ) = rt∗ + τt − ρ. (33b)
zt
Observe that after time t = T when the shock has faded and rt∗ = ρ, this is a saddle-path
stable system of stationary ODEs. That is, unless it is converging to a steady state, it leads
to a violation of the budget constraint (since either z → ∞ or z → B1 and both imply τ̇/τ is
bounded from below by a positive number greater than ρ). Thus, the only optimal solution
is the one where zt → z∗ and τt → 0, with z∗ uniquely defined by F (z∗ , IT ) = 0 (this is
possible since the image of V 0 (z) is (0, ∞), see Appendix B.2).
−V 00 (z)z
Sign of τt . Suppose ∆r ∗ > 0 and > Σ(z) (the other cases are exactly analogous).29
V 0 (z)
What can we say about the sign of τt ? Define the positive mapping X (z) > 0 by30
X (z) can be thought of as a measure of inverse marginal utility. In the familiar case where
Σ(z) = σ, for instance, X (z) = zσ . Since X (z) is strictly increasing and differentiable in z, it
admits an increasing and differentiable inverse, Z ( x ). By the implicit function theorem,
Z0 (x)x 1
= .
Z(x) Σ( Z ( x ))
29 Note that this proof applies to any kind of world interest rate shock {rt∗ }, as long as rt∗ = ρ after t = T
and either rt∗ ≥ ρ or rt∗ ≤ ρ for all t < T. Furthermore, note that we have not used any properties of the
foreign interest rate process to prove continuity and τt = 0 and limt→∞ τt = 0 other than rt∗ = ρ ∀t > T.
30 This only pins down X ( z ) up to a multiplicative constant which is irrelevant below.
A-7
Using X (z), we do a variable substitution, from zt to xt ≡ X (zt ) It−1 , so that zt = Z ( xt It ).
This is a convenient substitution, since
Γ Z ( xt It )
τ̇t = rt∗ τt + G( xt , It )
2λ Σ( Z ( xt It ))
Γ Z ( xIt )
τ=− G( x, It ) (34)
rt 2λ Σ( Z ( xIt ))
∗
at time t. The relationship in (34) only crosses zero once, at x = xt∗ , where xt∗ = x ∗ for
t ≥ T, and is negative (positive) for x > xt∗ (x < xt∗ ). Moreover, since It increases over time
(due to ∆r ∗ > 0), xt∗ falls over time as
ΣV ( Z ( xI ))
0 00 0 0
G I ( x, I ) = −V (z) − V (z) Z ( xI ) xI = V (z) −1 > 0 (35)
Σ( Z ( xI ))
−V 00 (z)z
where ΣV (z) ≡ V 0 (z)
. (35) is the case studied in this proof (the other cases are analogous).
Figure 9 illustrates these relationships in a phase diagram (thick black) and its stable
arm (red line) in that case. The green line depicts the shape of the optimal trajectory that
we are trying to pin down mathematically.
In a first step, we show that it can never be the case that τt ≥ 0 and xt > xt∗ for any t > 0.
In Figure 9, this would be a state (τt , xt ) that lies to the top right of the time-t x −locus. In
such a case, for any s > t, both ẋs and τ̇s are positive and bounded away from zero, and
hence the state (τt , xt ) would diverge to ∞. As before, τt would diverge at a rate that is
bounded from below by a positive number greater than rt∗ , violating the budget constraint
(18b).
Second, consider the possibility that for some t > 0,
A-8
x
ẋ = 0
∗
• xt
τ̇ = 0 (t < T )
x T∗
•
τ̇ = 0 (t ≥ T )
τ
Figure 9: Describing the optimal policy in the state space for (τ, x ).
Given xt∗ is decreasing in t, if (τt , xt ) ∈ Xt , then (τt , xt ) ∈ Xs for any s < t as well. In
particular (τt , xt ) ∈ X0 . Given no path satisfying the ODEs can ever enter X0 (that is, X0 ,
is a “source” in the vector field sense), it must hold that (τ0 , x0 ) ∈ intX0 (the interior of
X0 ). This contradicts the fact that τ0 = 0. Together, these two steps prove that τt ≥ 0 is
impossible for any t > 0. Thus, τt < 0 for t > 0. This concludes the proof of Lemma 1.
Differentiability with respect to ∆r ∗ . The right hand side of the system (33) is contin-
uously differentiable in ∆r ∗ at ∆r ∗ = 0. By the theorem on differentiable dependence of
ODEs (see, e.g. Theorem 2.16 in Grigorian, 2007), this means that τt and zt are differentiable
in ∆r ∗ for any t ≥ 0.
Application to baseline model Appendix B.2 establishes that V (z) is increasing and
weakly concave, and Σ(z) is positive, continuous and bounded from above, in the model
of Section 2. Therefore, Lemma 1 applies.
A-9
market clearing (12) implies reserves must be
∗
bGt = nfat + b It + b I Ht − b∗Ht (36)
∗
As we assumed that b H is sufficiently large to ensure that an allocation with {τt = 0},
which would lead to a strictly larger net foreign position nfaτt =0 ,31 can be achieved without
reserve position, that is,
∗
max nfaτt =0 ≤ b H
t
Thus,
∗ ∗
bGt = nfat − b H + b It + b I Ht < 0
| {z } | {z }
≤0 <0
The predictions for the exchange rate follow directly from the fact that under the optimal
policy, zt is raised initially relative to the {τt = 0} allocation (i.e. from the fact that ẑt
decreases in t at the optimum in the proof of Lemma 1).
A-10
C.6 Non-fundamental shocks: Proposition 7
R ∞ Rt ∗
Suppose the optimum had a positive present value of costs C ξ = 0 e− 0 ru du τt b It . The
policy τt = 0 then clearly has a lower cost, and in addition, does not distort the consumption
choice by home households, implementing the first best conditional on a zero cost term.
Therefore, the optimum cannot be one with a positive present value of costs.
D Extensions
D.1 Alternative assumptions on intermediaries’ demand
Our baseline model assumed a particularly simple, linear demand schedule of foreign
intermediaries. We now extend our model to allow for more general demand schedules.
Local infinite elasticity. This description of intermediary demand rules out one im-
portant case, namely that where intermediary demand is locally infinitely elastic when
rt − rt∗ = 0. For example, this can occur in our framework if we allow a nonzero mass of
foreign intermediaries to have zero participation costs (still subject to position limits). In
that case, the cost term has a kink at τ = 0, but remains strictly convex. The kink implies
that even very small interventions have a non-negligible cost. This leads the planner to
choose τt = 0 whenever it would have chosen a τt close to zero without the kink, which in
our experiments above means that τt 6= 0 for t in some interval ( T1 , T2 ). Between T1 and
T2 , the economy behaves exactly as in the baseline model.
32 We also conjecture that all our results generalize to the case where g0 ( τ ) τ is globally convex. If τg ( τ ) is
sufficiently non-convex and shocks are large, the planner will start to behave as if the cost term was largely a
fixed cost and therefore intervene more strongly.
A-11
D.2 Model with long-term assets
Deriving the generalized cost term. There are three important equations that change
with long-term assets. First, the consolidated household budget constraint is now given by
α ∞ Z ∞ Z
p(zt )y N + c + ḃ Ht + ḃ∗Ht + πt,s ∂t b Ht,s ds + ∗
πt,s ∂t b∗Ht,s ds = (37)
1−α t t
y T + b Ht,t + b∗Ht,t + rt b Ht + rt∗ b∗Ht + tt + πt .
b Ht + b It + b I Ht + bGt = 0 (39)
Z ∞ Z ∞
nfat ≡ b Ht + b∗Ht ∗
+ bGt + bGt + πt,s (b Ht,s + bGt,s )ds + ∗
πt,s (b∗Ht,s + bGt,s
∗
)ds
t t
Using the market clearing conditions (39) and (40) this further simplifies to
Z ∞
α ˙ t = y T + r ∗ nfat − (rt − r ∗ )(b It + b I Ht +
p(zt )y N + c + nfa t t πt,s (b It,s + b I Ht,s )ds) + πt ,
1−α t
A-12
R∞
and using the definition of profits, πt = (rt − rt∗ )(b I Ht + t πt,s b I Ht,s ds), and the spread,
τt = rt − rt∗ , we arrive at
Z ∞
α ˙ t = y T + r ∗ nfat − τt (b It +
p(zt )y N + c + nfa t πt,s b It,s ds)
1−α t
where
Z ∞
nfa0− ≡ b H0 + b∗H0 ∗
+ bG0 + bG0 + ∗
π0,s (b∗H0,s + b∗I H0,s + bG0,s
∗
− b It,s )ds.
0
In the case of short-term bonds, there was no revaluation term and nfa0− = nfa0 . With this
decomposition in mind, we can rewrite (42),
Z ∞ Z ∞
∗ α
π0,t p(zt )y N + c − y T + τt b It + πt,s ∆b It,s ds = nfa0− (44)
0 1−α t
Optimal policy. Next, we explore the influence of long-term assets on the optimal pol-
icy. To do this, we return to our formulation of intermediary demand from Section 2.1.
Assuming position limits now apply to the total value of an intermediary’s R∞ position
yields a straightforward generalization of intermediary demand (7), b It + t πt,s b It,s ds =
1 ∗
Γ (rt − rt ) . The planner has the possibility of extracting a time-zero valuation gain, given
by second term in (43). Since the planner has commitments at different points in time, the
whole path of home interest rates matters, which makes it difficult to derive general results.
One can nevertheless gain some insight into the solution by approximating the revaluation
A-13
term to first order around τt = 0,
Z ∞ Z ∞ Rt Z ∞ Rs
∗ − ∗ ru∗ du
0 ru du e−
π0,t − π0,t b I0,t dt ≈ e τt t b I0,s dsdt. (45)
0 0 t
R ∞ Rs ∗
Defining ξ 0,t ≡ t e− t ru du b I0,s ds, we see that the problem is approximately the same as
the one with non-fundamental shocks (Section 3.6).
Time consistency. The date-0 revaluation effects that appear in a model with long-term
assets raise an important question, namely whether those could be used to solve the time
inconsistency problem that we identified in Section 3. The idea is that even though assets
are linearly dependent and, hence, redundant, they are revalued differently in response
to policy changes. Thus, if a planner can choose which assets are held in equilibrium, it
can select a combination of assets today whose revaluation effects would “punish” future
planners upon deviations (Lucas and Stokey, 1983).
Clearly, according to our previous results, if the planner were able to choose {b It,s }, the
time inconsistency problem could be resolved by inducing an equivalent sequence of {ξ t,s }
which makes the full-commitment solution {τt } optimal for any future decision maker. In
that sense, the logic in Lucas and Stokey (1983) would fully apply to our model if today’s
planner can influence the maturity structure of the entire country’s liabilities.
This is not the case in our model, however, as the planner cannot directly control
b It,s . Both households and intermediariesare indifferent across all maturities, so that the
R∞ ∗ −π
composition of the revaluation term 0 π0,t 0,t b I0,t dt is indeterminate and cannot
be influenced by the planner alone. This is the reason why the time inconsistency we
identified in Section 3.4 cannot be solved by maturity management—the planner in our
model cannot choose the maturity structure of the entire country.
A-14
Frictionless forward markets: Interventions create CIP deviations. Suppose forward
markets are unconstrained, i.e. any agent can buy them and sell forwards without paying
any participation costs or facing any position limits. Then, it must be that the expectation
hypothesis holds, i.e. forward prices are equal to expected future spot prices
−(1−α)
f t,s = qs ≡ αα (1 − α)(1−α) ps (47)
and thus, due to frictionless forward markets (47), the UIP deviation τt stems entirely from
a CIP deviation. Thus, τt should be interpreted as a CIP deviation.
Frictional forward markets: Interventions create UIP deviations. Now suppose for-
ward markets are also constrained. That is, home agents face a constraint on the total value
of their promises to deliver dollars
Z ∞
∗
b∗Ht − ∗
πt,s d Ht,s ds ∈ [b∗H , b H ], (49)
t
while foreign agents have to pay a participation cost to trade either home bonds b It and/or
currency forwards d It,s and face a constraint on the total value of their promises to deliver
home currency, Z ∞
b It + πt,s f t,s d It,s ds ∈ [− X, X ]. (50)
t
These restrictions capture standard constraints on open cross-currency positions faced by
financial intermediaries.
As before, FX interventions induce UIP deviations given by equation (46). Are they
again reflecting underlying CIP deviations? Consider again the investment strategy above.
Its return is still given by the cross-currency basis (i.e. the CIP deviation) in (48). However,
as f t,s is no longer equal to qs , it does not necessarily equal the UIP deviation. In fact,
R t+∆
q
observe that the investment strategy involves b It = qt and d It,t+∆ = − f t e t ru du .
t,t+∆
Substituting this into (50) shows that the investment strategy involves zero exposure to
home currency and thus can be engaged in arbitrarily by intermediaries.34 Thus, its excess
34 A similar argument works for home agents.
A-15
return and thus the CIP deviation must be equal to zero,
qt R t+∆
ru du
R t+∆
ru∗ du
e t −e t = 0.
f t,t+∆
Thus, in the case of frictional forward markets, the UIP deviations (46) do not reflect
underlying CIP deviations.35
Should the spread in the model should be interpreted as a CIP or a UIP deviation in
the data? Our analysis above emphasizes that to answer this question one must identify
whether agents face higher restrictions shifting funds between home and foreign bond
markets or across currencies. In reality, frictions on both margins exist and, depending
on the situation, one may be more important than the other. For example, UIP deviations
may be more important in emerging markets, which put in place several restrictions on
cross-currency positions (Canales-Kriljenko, 2003), while CIP deviations may be more
relevant in countries close to the zero lower bound (Amador et al., 2020).
Optimal policy. The remainder of the analysis carries over almost verbatim. Indeed, in
our deterministic economy and with commitment, the only difference is that initial wealth
depends on the realization of the exchange rate at t = 0,36
Z ∞
Rt
− 0 rs∗ ds α 1 2 −(1−α)
e p(zt )y N + c − y T + τt dt = p0 nfa0 .
0 1−α Γ
If initial wealth were zero, the planning problem would be exactly the same as before.
Otherwise, there is a well-known incentive to inflate away the value of debt (Fischer, 1983;
Calvo, 1988). This “optimal revaluation” problem is somewhat similar to the “smooth-
exchange-rate” problem studied in Section 3.2 in that the planner directly cares about the
initial value of the exchange rate. For example, if the country is a debtor, the planner
would trade off depreciating the currency at t = 0 to decrease the burden of debt with the
large carry-trade costs this kind of intervention entails. As in Section 3.2, this would entail
interest rate spreads τt which jump at t = 0 and then monotonically decrease over time.
Aside from the valuation effect at time zero, or whenever the planner re-optimizes, the
analysis in Section 3 carries over unaltered.
The role of uncertainty. Since our model is deterministic, any UIP deviation is costly—it
reflects foreign intermediaries taking the opposite position and making riskless profits. In
a richer model with uncertainty, UIP deviations may also occur due to risk and not due
to imperfect arbitrage. Suppose that y Nt is stochastic and intermediaries value wealth in
35 Note that, if τt > 0, the home currency forward is cheap i.e. f t,t+∆ > qt+∆ and vice versa when τt < 0.
That is, when τt > 0, there is excess demand of both home currency assets and home currency forwards by
foreigners. On the flip side, home households would like to borrow more in dollar bonds or sell dollars
forward (i.e. buy home currency forward).
36 We assume agents are not trading forwards for this derivation. Otherwise, we would also have an
additional revaluation term owing to the fact that forwards are long-term assets, as in Section 4.2.
A-16
different states of the world according to some exogenous stochastic discount factor m∗ . In
this case, the intertemporal budget constraint becomes37
Z ∞ Rt ∗ α
− r ds 1 −(1−α)
Em ∗ e 0 s 2
p(zt )y Nt + c − y T + τt dt = p0 nfa0
0 1 − α Γ
where Em∗ is the risk neutral measure implied by the intermediaries and the wedge is
given by
1 Em∗ dqt
τt = rt − − rt∗ .
qt dt
In other words, only the part unjustified by risk is costly—a point also stressed by Amador
et al. (2020). Thus, one should be careful when computing the costs of FX interventions
from UIP deviations in the data.
More subtly, one may expect Γ to depend on policy via the exchange rate when there
are financial frictions in currency markets, as in Gabaix and Maggiori (2015) and Itskhoki
and Mukhin (2019). In particular, since the arbitrage is no longer riskless, foreigners may
endogenously choose the size of their position X depending on exchange rate volatility.
This may give rise to additional forces that are not present in our model. For example, the
planner may “noise up” the exchange rate to increase Γ. In this case, while our results
still illustrate one of the main forces at play, they are no longer a full description of the
optimal policy, which would take into account its effect on Γ. While we think this is a very
promising line of research, it is outside the scope of this paper.
Assume wages are perfectly rigid and normalized to set the home currency price of the
nontradable good equal to 1, and denote the nominal exchange rate by et .38 This implies
that nontradable output y Nt is determined by
et−1 y Nt = (1 − α)zt .
37 Note that given market incompleteness, one would also need to keep track of sequential budget con-
straints in the planning problem.
38 We use the convention that lower values of e reflect a more appreciated home currency.
t
A-17
f
As is clear from (13), y Nt exactly coincides with the “flexible-price output level” y N ≡ y N ,
f
where y N satisfies v0 (y N )y N = 1 − α, when et = et ≡ p(zt ) = (1 − α)y− 1
N zt . Thus, the
µ = χ = c = 0 version of our baseline model can be nested by assuming monetary policy
f
is implementing et = et at all times,39 capturing the idea that the output gap objective
takes priority over the exchange rate objective in that model.
In this section, we explore the polar opposite: we assume that—for some unmodeled
reason—the monetary authority has some exchange rate objective et . To make it stark, we
assume a fixed exchange rate regime, et ≡ e = 1, and ask: how can the planner use FX
interventions to regain some monetary independence and mitigate the impact on the home
economy? Examples of interventions of this sort arguably include recent interventions by
Euro neighbors like Denmark, Switzerland, or the Czech Republic, which have tried to
fend off appreciations and at the same time avoid being pushed into the zero, or effective,
lower bound for interest rates.
In a first step, we ask which allocations can be implemented by central bank policies.40
Fortunately, it is straightforward to show that, in fact, when stated in terms of {zt , rt }, the
same implementability conditions as in Proposition 1 continue to hold in this economy,
just with µ = χ = c = 0. The reason is that aggregate (as well as individual) income from
nontradables is still equal to p(zt )y N , even though it is now entirely caused by a larger
output quantity y Nt and no price response due to sticky prices.
By contrast, the objective function changes. Noting that nt = y Nt and following the
same steps as before, we find the planning problem to be
Z ∞
max e−ρt {log zt − v((1 − α)zt )} dt (52)
{zt ,τt } 0
A-18
χy T , and a fraction µ2 = µ − µ1 of poor households (type 2) work in the nontradable sector,
earning χpt y N . All households share the same utility function (1), where for simplicity we
focus on log preferences, σ = 1. The demands of both types of poor households are given
by
P1
µ1 c Tt = αχy T
P2
µ2 c Tt = αχpt y N
and similarly for nontradable goods. Thus, market clearing for nontradable goods is
which simplifies to
1 1−α
pt = ((1 − µ)zt + χy T ) .
y N 1 − (1 − α ) χ
This is the same as (13), setting the subsistence level to zero, c = 0.
The key distinction to the model in Section 2 emerges in the utility function. In this
model, the planner’s utility is given by, up to a constant
It is easy to see that the implementability conditions (18) are unchanged, and that V (z) is
V 00 (z)z
increasing, concave and satisfies − V 0 (z) > 1 when there are not too many poor households,
µ1 < 1 − α and µ2 < α. Thus, our results in Section 3 continue to hold in this environment.
A-19
where y H is home’s endowment of the home good and y F is home’s endowment of the
−(1−α)
foreign good. All other objects are as in Section 2. We denote by qt ≡ pt the country’s
real exchange rate, following the convention that high values correspond to depreciated
exchange rates. In this environment, home households’ own a nontrivial share of the home
good and exhibit home bias in their preferences. Together, these two assumptions are
essential in generating the terms of trade management motive in our environment.
Maximizing utility subject to budget constraint (53) yields the following Euler equation,
c˙t q˙t
= rt − ρ + . (54)
ct qt
pt c Ht = (1 − α)zt (55)
c Ft = αzt .
c∗t
c∗Ht = α (56)
pt
z˙t
= rt∗ + τt − ρ
zt
Z ∞
−
Rt
rs∗ ds 1 2
e 0 α(zt − 1) − y F + τt dt = nfa0
0 Γ
This is of the same form as the planning problem in Section 4.1: a strictly increasing and
concave objective function, an Euler equation, and a present value budget constraint that is
linear in zt (Σ is constant and equal to one). Thus, our results in Section 3 continue to hold
in this environment.
D.7 Learning
To investigate the effect of learning among intermediaries about profitable carry-trade
opportunities, we repeat the shock experiment of Section 3, assuming that Γ declines over
A-20
time, that is,
Γt = Γ0 χt/5
where we set χ = 0.25. This corresponds to a fall in Γ by 75% every 5 years, or in other
words, the number of active intermediaries quadruples during the period of the shock. The
results of the experiment can be seen in Figure 10. The paths with and without learning
are relatively close overall. They are particularly close for the real exchange rate, which,
perhaps surprisingly, is achieved by a more backloaded path of interest rate spreads in the
case with learning. The reason for this apparent paradox is that smoothing spreads over
time becomes more desirable when Γ is low. Clearly, reserves also rise with a lower Γ.
Overall, the qualitative and quantitative insights are preserved even if there is learning by
intermediaries.
against world net private borrowing −b∗H , as we do in Figure 12. Clearly, they have to
be equal in equilibrium. When central banks keep a fixed stock of reserves bG ∗ (e.g. at
A-21
Figure 10: The effect of learning.
0.4 % 0.4 %
0.2 % 0.2 %
0% 0%
0 2 4 6 8 10 0 2 4 6 8 10
years years
Net foreign assets (NFA) / GDP NFA excl. reserves / GDP
0% 0%
−0.2 %
−0.1 %
−0.4 %
−0.2 %
−0.6 %
−0.3 % −0.8 %
0 2 4 6 8 10 0 2 4 6 8 10
years years
No intervention No learning Learning
Note. This figure illustrates how optimal FX interventions change when intermediaries learn about the
interventions and the associated carry-trade opportunities over time, implying that Γ falls over time. The
black line shows an equilibrium without interventions. The green line shows optimal interventions with
Γ = 9. The red line shows the optimum where we assume an exponentially decaying Γt , falling from by
75% every five years. Intuitively, this corresponds to a scenario where the number of active intermediaries
quadruples during the period of the shock.
A-22
Figure 11: Best response to the world interest rate and the FX intervention policy of others.
V best response τ BR
r1∗ = r ∗ (τ = 0)
45◦ line
r2∗ = r ∗ (τ Nash )
V (r1∗ , τ )
best response
τ ∗ (r ∗ (τ ))
V (r2∗ , τ )
τ τ
τ ∗ (r1∗ ) τ ∗ (r2∗ )
Note. The left panel illustrates that lower world interest r ∗ reduce an individual country’s welfare and lead to
larger FX interventions (higher interest rate spreads τ). The right panel illustrates that this induces a strategic
complementarity between countries’ FX interventions.
r∗
r∗
A-23
zero), independent of world interest rates r ∗ , and the world sees an increase in private
saving, interest rates fall somewhat and net private saving remains unchanged (left panel in
Figure 12). When central banks, however, implement their best response policies, increasing
reserves precisely as interest rates fall, their savings supply curve becomes downward
sloping and any shift in net private borrowing now has amplified effects on world interest
rates and private borrowing (right panel in Figure 12).
The optimal choice of τ has to satisfy the necessary first order condition
Strategic complementarity. We prove that sign(τ ) = sign(1 − (1 + r ∗ ) β). The result for
∗ follows from the fact that for τ > 0,
bG
∗
bG = nfa + b∗H + b It > 0 (60)
| {z } |{z}
≥0 >0
In this part and the next, we focus on the case η0 = η1 . We can thus drop the subindex
“t” on Vt (z) without risk of confusion. Consider the case β(1 + r ∗ ) > 1 and suppose τ ≥ 0.
A-24
This immediately implies z1 > z0 . Since V 0 (z)z is decreasing in z,43
V 0 ( z1 ) z1 < V 0 ( z0 ) z0 ≤ V 0 ( z0 ) z0 (1 + τ ).
Equation (59) then implies τ < 0, a contradiction. Hence, τ < 0. Using analogous
arguments one can establish that β(1 + r ∗ ) < 1 and β(1 + r ∗ ) = 1 imply τ > 0 and τ = 0,
respectively, completing the proof.
Negative externality. To see when changes in r ∗ could actually hurt a country, we com-
pute the derivative of the indirect utility V with respect to 1 + r ∗ . After some algebra, we
find44
∂V dV (z0 )
0 ∗
∗ V ( z1 ) z1 (r )
=− β (1 + r ) 0 −1 . (61)
∂ (1 + r ∗ ) d (1 + r ∗ ) V ( z0 ) z0 (r ∗ )
∂V
The first interesting observation is that ∂(1+r ∗ )
=0 if and only if r ∗ = β−1 − 1. At that
V 0 ( z1 ) z1 (r ∗ )
point, when r ∗ moves slightly, the term B ≡ ∗
β(1 + r ) V 0 (z ) z (r∗ ) changes as follows
0 0
(after some algebra),
d log B
| ∗ −1 = 2 − σ V
d log(1 + r ∗ ) r = β −1
−V 00 (z)z ∂V
where σV ≡ V 0 (z)
. This term is negative if and only if σV > 2, which implies ∂ (1+r ∗ )
changes sign at r ∗ =
β−1 − 1, from positive to negative. Thus, V has a local maximum at
r ∗ = β−1 − 1, implying that changes in interest rates decrease welfare. Conversely, ∂(1∂+Vr∗ )
has a local minimum if and only if σV < 2. This concludes this proof.
∗1 1 2
z0 (τ, r ) = η0 − τ (62)
η Γ
1
z1 (τ, r ∗ ) = η1 (63)
η
From (62) and (63), the world interest rate is pinned down by
1 z1 1 η1
1 + r∗ = = . (64)
β (1 + τ ) z0 β(1 + τ ) η0 − Γ1 τ 2
A-25
We can substitute out (64) in (59) to obtain
− B1 B0 ((z − B1 ) + ( B1 − c1 ))2 + C1 C0 (z − B1 )2
C1
B1 B0 > > C1 C0 .
C0
Taken together, there exists a unique solution (τ, z0 ) to the system of equations (62) and
(65), and, hence, a unique Nash equilibrium.
Other results. The logic in the preceding paragraph establishes that, when η1 < η0 , the
interest rate spread τ is strictly positive in the Nash equilibrium. Applying (60) implies that
∗ > 0 in equilibrium, but since nfa = 0 due to symmetry, this can only work if there are
bG
private inflows, b∗H + b It > 0. Finally, all countries’ welfare is lower, as is immediate from
(62), the only difference between the Nash equilibrium allocation and a τ = 0 allocation is
that z0 is lower, reducing welfare. Since τ 6= 0, intermediaries make profits in the Nash
equilibrium allocation, while they do not when τ = 0.
45 Thisfollows from the implicit function theorem as the left hand side strictly increases in z0 and decreases
in τ while the right hand side strictly increases in τ.
A-26