Escola de P os-Gradua˘c~ao em Economia - EPGE Funda˘c~ao ...
Transcript of Escola de P os-Gradua˘c~ao em Economia - EPGE Funda˘c~ao ...
Escola de Pos-Graduacao em Economia - EPGE
Fundacao Getulio Vargas
Ensaios em Econometria Aplicada
Tese de submetida a Escola de Pos-Graduacao em Economia da Fundacao
Getulio Vargas como quesito para a obtencao Tıtulo de Doutor em Economia.
Aluno: Rafael Martins de Souza
Orientador: Joao Victor Issler
Rio de Janeiro
2009
Escola de Pos-Graduacao em Economia - EPGE
Fundacao Getulio Vargas
Ensaios em Econometria Aplicada
Tese de submetida a Escola de Pos-Graduacao em Economia da Fundacao
Getulio Vargas como quesito para a obtencao Tıtulo de Doutor em Economia.
Aluno: Rafael Martins de Souza
Banca Examinadora:
Joao Victor Issler (Orientador, EPGE/FGV)
Prof. Marco Antonio Bonomo (EPGE/FGV)
Caio Ibsen Rodrigues de Almeida (EPGE/FGV)
Marcelo C. Medeiros (DE/PUC-Rio)
Paulo Pichetti (EESP/FGV)
Rio de Janeiro
2009
iii
Abstract
This thesis has three chapters. Chapter 1 explores literature about exchange rate pass-through,
approaching both empirical and theoretical issues. In Chapter 2, we formulate an estate space
model for the estimation of the exchange rate pass-through of the Brazilian Real against the US
Dollar, using monthly data from August 1999 to August 2008. The state space approach allows us
to verify some empirical aspects presented by economic literature, such as coefficients inconstancy.
The estimates offer evidence that the pass-through had variation over the observed sample. The
state space approach is also used to test whether some of the “determinants” of pass-through are
related to the exchange rate pass-through variations observed. According to our estimates, the
variance of the exchange rate pass-through, monetary policy and trade flow have influence on the
exchange rate pass-through. The third and last chapter proposes the construction of a coincident
and leading indicator of economic activity in the United States of America. These indicators
are built using a probit state space model to incorporate the deliberations of the NBER Dating
Cycles Committee regarding the state of the economy in the construction of the indexes. The
estimates offer evidence that the NBER Committee weighs the coincident series (employees in non-
agricultural payrolls, industrial production, personal income less transferences and sales) differently
way over time and between recessions. We also had evidence that the number of employees in non-
agricultural payrolls is the most important coincident series used by the NBER to define the periods
of recession in the United States.
iv
Resumo
A tese esta dividida em tres capıtulos. O capıtulo 1 trata de uma revisao de literatura sobre
pass-through, abordando aspectos empıricos e teoricos. O segundo capıtulo trata da estimacao de
um modelo de espaco de estados para estimacao dos pass-through da taxa de cambio no Brasil de
agosto 1999 a agosto 2008. A abordagem espaco de estados permite contemplar alguns aspectos
empıricos apresentados pela literatura economica, tais como a inconstancia dos parametros. As
estimativas ofereceram evidencia de que o pass-through no Brasil variou no perıodo estudado.
Ainda, a abordagem por espaco de estados permite que se estude os“determinantes” (ou variaveis
associadas) do pass-through. Com isto tivemos evidencia de que a variancia da taxa de cambio, a
polıtica monetaria e o fluxo de comercio afetam o pass-through. O terceiro e ultimo artigo da tese
trata da construcao de um indicador coincidente e antecedente da atividade economica nos Estados
Unidos da America. Nele utiliza-se um modelo probit de espaco de estados para incorporar as
decisoes do NBER Dating Cycles Committee na construcao dos ındices. A estimativas ofereceram
evidencia de que o comite do NBER pondera as series coincidentes (total de empregados em
atividades nao agrıcolas, producao industrial, renda pessoal menos transferencias governamentais
e vendas) de maneira diferente ao longo do tempo e entre as recessoes. Tambem evidenciou-se que
a serie coincidente total de empregados em setores nao-agrıcolas e a principal serie considerada
para a definicao dos perıodos de recessao nos Estados Unidos.
v
Agradecimentos
Agradeco,
A Deus, por me permitir caminhar ate aqui, apesar de todas as dificuldades.
Ao meu orientador, Prof. Joao Victor Issler pelo apoio, incentivo e exemplo profissional. Ao
Prof. Pedro Cavalcanti Gomes Ferreira que abriu as portas da EPGE quando eu era, ainda, um
aluno de graduacao para fazer bolsa de iniciacao cientıfica.
A Fundacao Getulio Vargas, a CAPES e ao CNPq pelo suporte financeiro.
Aos amigos de San Diego, Daniel, Daniel Aiex e Eillen. O convıvio com voces foi inesquecıvel.
A gratidao e eterna.
A todos os amigos da EPGE. Em especial ao Jose Diogo, ao Orlando, ao James, ao Flavio,
ao Luiz Felipe, a Amanda, ao Pedro, ao Gustavo, ao Gabriel e ao Hilton. Obrigado por todos os
momentos.
Aos meus amigos de longa data, Aline e Ralph, que, mesmo quando estavam em paıses distantes,
estivem sempre proximos o suficente para me encorajar e me incentivar nos momentos difıcies.
Aos meus novos colegas de trabalho, Luisa e Gustavo, que pela paciencia, incentivo e forca na
reta final.
Aos meus pais, Marlene e Ronaldo, e ao meu irmao Samuel, por todo amor, incentivo, encora-
jamento, participacao... Descrever toda a importancia da nossa famılia e impossıvel. A gratidao e
infinita.
A Mozuca, pela sua paciencia, benevolencia, altruısmo, seu trato carinhoso, sua calma, sua
tolerancia, sua crenca, seu suporte,... Enfim, por todo o seu amor. O Amor nunca falha.
vi
Key words and phrases: Exchange rate pass-through, business cycles, indicators of economic
activity, state space models, Kalman filter.
Palavras-Chave: Pass-through da taxa de cambio, ciclo de negocios, indicadores de atividade
economica, modelos de espaco de estados, filtro de Kalman.
Contents
I Exchange Rate Pass-Through 1
1 A Discussion on Exchange Rate Pass-Through 2
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Empirical Evidences on Exchange Rate Pass-Through . . . . . . . . . . . . . . . . 4
1.3 Pass-Through Determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.1 Macroeconomic Determinants . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.2 Output Gap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.3 Microeconomic Determinants of Pass-Through . . . . . . . . . . . . . . . . 11
1.4 The economic model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2 Pass-Through Estimation in Brazil 15
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Econometric Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.1 Linear state space models under restrictions . . . . . . . . . . . . . . . . . . 16
2.2.2 Model Selection and Inference . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 Econometric Setting and Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . 19
vii
CONTENTS viii
2.3.1 Time Varying Coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.2 Determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
II Coincident and Leading Indexes of Economic Activity 37
3 A State Space Model for Indices of Economic Activity 38
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2.1 Determining a basis for the cyclical components of coincident variables . . . 41
3.2.2 Estimating a structural equation for the unobserved business cycle state . . 43
3.2.3 The iterated extended Kalman filter and smoother . . . . . . . . . . . . . . 49
3.2.4 The Kalman Filter and Smoother Instrumental Variables Index . . . . . . . 51
3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3.1 The Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3.2 The Basis Cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3.3 Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.3.4 Predicting Recessions in Real Time . . . . . . . . . . . . . . . . . . . . . . . 65
3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
List of Figures
2.1 IPA-OG smoothed coefficient of Δ log et, Δ log et−1 and Δ log yt−1. . . . . . . . . . 25
2.2 IPA-OGPA smoothed coefficient of Δ log et, Δ log et−1 and Δ log yt−1. . . . . . . . 26
2.3 IPA-OGPI smoothed coefficient of Δ log et, Δ log et−1 (top), Δ log yt−1 and Δ log yt−6
(bottom). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.4 IPA-OG long run pass-through. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.5 IPA-OGPA long run pass-through. . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.6 IPA-OGPI long run pass-through. . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.7 Tested determinants of pass-through: Monetary policy, (solid line), variance of ex-
change rate (dashed line) and international trade (dotted line). . . . . . . . . . . . 31
3.1 Coincident Series. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.2 Coincident Cycles (growth rate) plot. . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.3 Filtered and Smoothed weights. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.4 Predicted and Smoothed probabilities using data from 1960:06 to 2007:03. . . . . . 61
3.5 Predicted and Smoothed probabilities using data from 1960:06 to 2007:03. . . . . . 63
3.6 Filtered and Smoothed weights. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
ix
LIST OF FIGURES x
3.7 1990 Recession. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.8 2001 Recession. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.9 2007 Recession. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
List of Tables
2.1 Some quality of fit statistics of the adjusted models. . . . . . . . . . . . . . . . . . 23
2.2 Information criteria observed values for incomplete, null and complete pass-through
exchange rate models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 P-values of the tests for null and complete pass-through exchange rate. . . . . . . . 24
2.4 Estimates of the IPA-OG, IPA-OGPA and IPA-OGPI series (p-values between paren-
thesis). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.5 Estimated parameters and corresponding p-values (in parenthesis). . . . . . . . . . 34
2.6 Estimated parameters and corresponding p-values (in parenthesis). . . . . . . . . . 34
3.1 Coincident and leading variables. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.2 Squared canonical correlations and canonical-correlation test. . . . . . . . . . . . . 57
3.3 Descriptive statistics of the filtered (left) and smoothed (right) coefficients. . . . . 60
3.4 Accuracy of estimation based on a cut-off point of 0.5. . . . . . . . . . . . . . . . . 62
3.5 Descriptive statistics of the smoothed weights. . . . . . . . . . . . . . . . . . . . . . 65
3.6 Predicted probabilities associated for period from 2007:08 to 2008:07 . . . . . . . . 69
xi
Part I
Exchange Rate Pass-Through
1
Chapter 1
A Discussion on Exchange Rate
Pass-Through
1.1 Introduction
The exchange rate pass-through degree is the elasticity between exchange rate and domestic prices.
In other words, it is the percentage impact on 1% change in exchange rate into domestic prices.
In an open economy, domestic prices can be affected by external shocks, whether by currency
relative price adjustment or by movements in international supply and demand. The exchange
rate pass-through highlights how sensitive each market is to fluctuations in exchange rate.
The study of exchange rate pass-through has intensified since 1980. The literature focuses on
the behavior of the impact of exchange rate on prices and their determinants. However, the real
motivation for these studies was the study of the Purchase Power Parity Puzzle (PPP). The PPP
2
CHAPTER 1. A DISCUSSION ON EXCHANGE RATE PASS-THROUGH 3
assumption states that all variation in exchange rate is passed-through into prices. The major
conclusion from empirical studies in the last years is that the PPP is not valid in the short run;
therefore, exchange rate pass-through into prices is less than one. However, there is some evidence
in favor of the validity of PPP on long run.
The pass-through estimation could provide a test for the existence of the purchase power parity.
If the PPP were valid in the long term, the pass-through would be complete and the sum of the
pass-through coefficients would have to sum to one. Otherwise, the long run pass-through would
be incomplete and the effect of variations in exchange rate into prices would be restricted.
The importance of exchange rate pass-through has increased since the adoption of the inflation
targeting regime. Fraga, Goldfajn and Minella (2003) have shown that the problem of having a
high exchange rate pass-through degree is that it implies a greater difficulty for attaining inflation
targets. A greater exchange rate pass-through means that the domestic economy is more sensitive
to external shocks, consequently the impact of exogenous shocks into domestic prices is amplified.
Exchange rate pass-through also seens to affect the inflation forecast. According to Goldfajn
and Werlang (2000), the exchange rate pass-through into prices is directly associated with inflation
forecast error. With a smaller pass-through, the domestic economy is more stable and less affected
by external factors. Therefore, a smaller pass-through means that the difference between inflation
expectations and inflation targets is smaller. In other words, a small pass-through generates a
minor inflation forecast error. Consequently, a small pass-through is associated with a major
transparency of inflation path and a minor volatility in price variations in the economy, rising
social welfare and monetary policy efficiency.
The exchange rate pass-through into prices is one of the main drivers to optimal monetary
policy. According to Betts and Devereux (2000), the trade-off between output volatility and
CHAPTER 1. A DISCUSSION ON EXCHANGE RATE PASS-THROUGH 4
inflation volatility is dependent on how sensitive prices are to exchange rate variations. They
begin their argument by explaining that the nature of the trade-off between different exchange
rate regimes is quite different in industrial countries from the trade-off in emerging ones. Using a
DGE Model (Dynamic General Equilibrium Model), they argue that the critical distinction is the
exchange rate pass-through into prices. With very high exchange rate pass-through, policies that
stabilize output require high exchange rate volatility, which implies high inflation volatility. But
with limited or delayed pass-through, this trade-off is less pronounced and a flexible exchange rate
policy that stabilizes output can do so without high inflation volatility.
Another study that emphasizes the importance of exchange rate pass-through in an inflation
targeting regime is that of Fraga, Goldfajn and Minella(2003). They have shown that the problem
of having a high exchange rate pass-through degree is that it implies a greater difficulty for attaining
inflation targets. The larger the exchange rate pass-through, the more sensitive the domestic
economy to external shocks, that is, the impact of exogenous shocks on domestic prices is amplified
by a larger exchange rate pass-through.
1.2 Empirical Evidences on Exchange Rate Pass-Through
A main factor in pass-through estimation is the difficulty of using aggregated data and the known
problem of aggregation bias. This factor favors a disaggregating process for prices, and tries to
capture the exchange rate pass-through for each good or each market. Campa and Goldberg (2005)
present results where estimates are better across industries than across countries with aggregate
data. These authors also say that the major source of pass-through variations are competition
issues in each sector. Yang (1996) and Olivei (2002) show that pass-through varies significantly
CHAPTER 1. A DISCUSSION ON EXCHANGE RATE PASS-THROUGH 5
across industries. Menon (1996) supports these findings pointing to the aggregation bias, indicating
that disaggregated data provides more accurate estimates and captures the impact of exchange
rates on commodities prices more precisely.
Campa and Goldberg (2005) and Pollard and Coughlin (2005) follow this trend of disaggregated
estimation of pass-through. Their approach permits a more individual analysis for each market,
relating pass-through with market power and degree of competition. This type of analysis allows
more plausible explanations for aggregated pass-through behavior. Campa and Goldberg (2005)
observe that the US has seen a change in composition of certain industries in its import basket.
Industries with a bigger pass-through, such as energy and raw materials, have shown a decrease
in their share in the US imports basket, reducing the aggregate pass-through. The proportion of
tradable and non-tradable goods is important to analyze the aggregate exchange rate pass-through
because tradables are more sensitive to changes in exchange rate than the non-tradables. Therefore,
the greater the share of tradable goods, the higher the exchange rate pass-through.
Some results about pass-through estimation can be seen in Goldberg and Knetter (1997), where
the exchange rate pass-through to US inflation was approximately 50% after 6 months. Campa
and Goldberg (2005) estimate pass-through for 25 OECD countries. They found a pass-through of
26%, in the short term, and 41% in the long term for the US. The average pass-through estimated
for OECD countries in the short and long run was 61% and 77%, respectively.
Sekine (2006) estimated exchange rate pass-through for six developed countries (United States,
Japan, Germany, United Kingdom, France and Italy) by taking into account their time-varying
natures. The author incorporates that characteristic by allowing permanent shifts in pass-through
parameters. He found that pass-through has declined over time in all major industrial countries
and, in most cases, pass-through did not show the parameter shift envisaged by split sample
CHAPTER 1. A DISCUSSION ON EXCHANGE RATE PASS-THROUGH 6
estimations.
Calvo and Reinhart (2000) have shown that the pass-through degree of emerging countries is
four times greater than that of developed countries. Additionally, these authors calculate that the
variance of inflation compared with the variation of exchange rate is 43% for emerging countries
and 13% for developed ones.
In the case of the Brazilian economy, there are few studies estimating the exchange rate pass-
through. Belaisch (2003) used a VAR specification controlling for petroleum shocks and estimated
the exchange rate pass-through into IPCA approximately 6% after 3 months. He also estimated
other price indeces and found that the pass-through to IPA (34%) was larger than to IGP (27%),
which is larger than IPCA. Carneiro, Monteiro and Wu (2002) used a non-linear estimation for
pass-through into IPCA, in an attempt to capture possible asymmetries in exchange rate variations
into prices. Their estimate was 6, 4%, on average.
Albuquerque and Portugal (2003) used a time varying estimation for the IGP, IPCA and IPA.
They found evidence of time varying pass-through in Brazil, although they used a complicated
period (1980-2002). Their data set was prejudiced by multiple exchange rate regimes, multiple
changes in economic policies and some financial crisis. Their state equation estimates for IPA
were not significant, with the exception of the persistence term. For IPCA, their estimates were
approximately 6% on average. However, since 1995 their exchange rate pass-through to IPCA was
not significantly different from null.
After the estimation of exchange rate pass-through, studies started to explain its behavior and
the reasons for so much variation across countries, across time and across industries. The reasons
could be in the exchange rate pass-through determinants. According to Goldfajn and Werlang
(2000), the exchange rate pass-through varies across countries, in a way that more stable countries
CHAPTER 1. A DISCUSSION ON EXCHANGE RATE PASS-THROUGH 7
exhibit smaller pass-through. Another result is that the exchange rate pass-through changes with
time horizon, reaching its peak at 12 months in the case of Brazil. These authors analyzed four
variables as pass-through determinants: real exchange rate misalignment, initial inflation, output
gap and openness degree. Their results indicate that all variables have important correlations with
exchange rate pass-through, depending on the countries’ characteristics, although real exchange
rate misalignment and inflation environment were the most important.
Inflation is positively correlated to exchange rate pass-through. Empirical evidence suggests
that the larger the inflation persistence, the larger inflation rate. Hence, there is more volatily in
the macroeconomic variables than the exchange rate pass-through.
According to Taylor (2000), a low inflation environment implies a decrease in exchange rate
pass-through. He argues that low and more stable inflation should be associated with less persistent
inflation. Hence, the low inflation and the monetary policy that has delivered it have led to lower
pass-through by a reduction in expected persistence of cost and price movements.
Gagnon and Ihrig (2001) argue that recent adoptions of anti-inflationary policies and the rise in
central bank credibility are important factors to explain the diminishing effects of inflation on the
exchange rate pass-through. When inflation is low and the commitment of the central bank to keep
inflation stable has credibility, the economic agents become less inclined to quickly pass-through
costs variations to prices.
According to Choudri and Hakura (2003) there exists strong evidence of a positive and sig-
nificant relation between average inflation and pass-through. The authors argue that a limited
pass-through gives more freedom for a independent monetary policy, benefiting the implementa-
tion of a inflation targeting regime.
CHAPTER 1. A DISCUSSION ON EXCHANGE RATE PASS-THROUGH 8
1.3 Pass-Through Determinants
Since there are many studies trying to identify the causes of the exchange rate pass-through, we
summarize some the possible exchange “determinants”, proposed by these studies. According to
Menon (1996), Goldfajn and Werlang (2000), Taylor (2000) and Campa and Goldberg (2002), the
main drivers of price sensibility to exchange rate changes can be inferred. From the Macroeconomic
point of view, the pass-through depends on the openness degree of the economy, the output gap,
inflation persistence and real exchange rate misalignments. From the standpoint of disaggregated
analysis, the exchange rate pass-through is associated with the competition degree of each industry
and with a firm’s market power (with the elasticity price-demand).
1.3.1 Macroeconomic Determinants
1.3.2 Output Gap
The output gap is defined by the deviation of a product in relation to its long term value; in other
words, the difference between observed product and the value it was supposed to be according to
its long term trend. The evidence of past studies shows a positive correlation between pass-through
and output gap. The larger the difference between GNP and its potential, the greater the demand
pressure over prices. This fact generates an inflation environment, raising the probability that
firms pass-through changes in costs into prices. Therefore, in an environment where the output
gap is increasing, the exchange rate pass-through’s effect on inflation is intensified.
CHAPTER 1. A DISCUSSION ON EXCHANGE RATE PASS-THROUGH 9
Inflation Environment
According to Goldfajn and Werlang (2000), the variable inflation environment is defined as the
frequency which agents remark their prices based on past inflation. In countries with an inflationary
environment, it is easier for the agents to pass-through cost changes and increase prices. As a result,
the larger the inflationary environment – and the more persistent the inflation – the easier it is
for agents to pass-through exchange rate increases into prices. This reasoning is corroborated by
Taylor(2000), who suggests a correlation between inflation and exchange rate pass-through using
the inflation persistence as a channel of transmission. The model indicates that observed changes
in pass-through, or firms’ market power, are partly originated from changes in the persistence of
expected movements in cost and competitor prices. In this sticky price model, the pass-through
to prices depends on how permanent the increase of cost is. The greater the half-life of a rise in
marginal cost, the more firms will revise prices. For this reason, if exchange rate depreciation is
transitory, firms will pass-through to prices some of this increase in costs. However, the greater the
persistence of exchange rate depreciation, the greater the pass-through will be. Taylor(2000) argues
that persistence in cost changes is related to price stability. Therefore, in a stable environment, the
inflation persistence will be smaller. As a result, the half-life of cost changes will decrease causing
a smaller pass-through.
Openness Degree
The openness degree of an economy depends on the presence of tradable goods, which determine
how sensitive prices are to changes in exchange rates. This degree can be defined as the sum
of imports and exports as a proportion of GNP. In a more open economy, we expect that the
presence of goods more sensitive to exchange rate will be lager, which implies a larger exchange
CHAPTER 1. A DISCUSSION ON EXCHANGE RATE PASS-THROUGH 10
rate pass-through to inflation.
Real Exchange ate Disalignment
According to Goldfajn and Valdes (1999), a real exchange rate over valuated results a mean factor
on future inflation composition. If the real exchange rate is below its long term value, agents make
up the expectation of future devaluations, adjusting relative prices. However, if exchange rate
variation are not adjusted by relative prices, it will imply an increase in internal inflation in relation
to external inflation. As a result, an over valuated real exchange rate implies future depreciations
because the exchange rate is supposed to meet its steady state in the future. The agents will
take on this expectation of future depreciation, amplifying the effect on prices. Consequently, the
exchange rate pass-through will be negative associated with the difference of real exchange rate
and its long run value. The more over valuated the real exchange rate, the greater the expectations
of future devaluations, which will lead to an increase in the prices.
Variance of Exchange Rate
Large movements in exchange rate are associated with a higher exchange rate pass-through to
prices. If the variance of exchange rate is large, then the cost of changing prices decreases and
price-makers have more incentive to pass-through cost changes to prices. The idea is that if the
cost variation is large, then it is easier for the price maker to pass-through this cost changes to
prices.
Changing listed prices entails menu costs, such as the cost of printing new price lists and
the cost of notifying consumers of new prices. In order to justify the cost of raising prices, the
anticipated profit from the price change must exceed the menu cost. If the currency depreciation is
CHAPTER 1. A DISCUSSION ON EXCHANGE RATE PASS-THROUGH 11
substantial and is not likely to reverse itself, the cost of changing prices will be small in proportion
to the profit generated by the higher price. Furthermore, a significant cost increase affecting all
competitors simultaneously reduces the impact of a price hike on a company’s reputation. Price
changes therefore occur more frequently when exchange rate movements are large.
Devereux and Yetman (2002) developed a simple theoretical model of endogenous exchange
rate pass-through. The model ignores many factors that might limit pass-through, and focuses
exclusively on the role of price rigidities. Their main argument was that exchange rate pass-
through is determined by the types of shocks in the economy and the persistence of the shocks.
For a given size of the menu cost of price changes, firms will choose a higher frequency of price
adjustment if the average rate of inflation is higher and the nominal exchange rate is more volatile.
Thus, large movements in exchange rate and an inflationary environment are associated with a
higher exchange rate pass-through.
1.3.3 Microeconomic Determinants of Pass-Through
A main factor to analyze the exchange rate pass-through into disaggregated prices is the degree
of competition on the price setting sector. When the competition increases in an industry, the
market power of firms diminishes and the producers can pass-through less cost change to consumers
without losing market-share. Therefore, in a highly competitive environment, the exchange rate
pass-through will be limited and the producers will absorb cost increases – accepting less mark-
ups – and will not fully pass-through exchange rate variations to prices, with the intention to
protect market-share. Therefore, there is a negative relation between competition and exchange
rate pass-through.
Besides market power, the elasticity price-demand also impacts exchange rate degree. The
CHAPTER 1. A DISCUSSION ON EXCHANGE RATE PASS-THROUGH 12
more elastic the demand, the more consumers will respond to price changes, which implies that
producers have a limited ability to pass-through costs changes. Therefore, the more inelastic the
demand, the more producers will pass-through exchange rate variations into prices. This implies
the existence of a negative correlation between pass-through and elasticity price-demand.
Campa and Goldberg (2005) also argue that the aggregated pass-through have declined because
of the change in composition of certain industries in the import basket. Industries with a larger
pass-through have shown a decrease in their share in the US imports basket. At the same time,
industries with prices that are less sensitive to exchange rate devaluations experience a growth in
market share. The authors give the example of the reduction in the US energy sectors share, which
has an exchange rate pass-through of 70%, and raw materials (pass-through of 64%).
1.4 The economic model
The theoretical framework used to formulate the econometric models in the next chapter is directly
inspired by articles as Olivei (2002), Pollard and Coughlin (2005) and Campa and Goldberg (2002),
among others. The law of one price says that the price of any good, say good x, denoted by a
common currency should be the same in any two markets:
PH = EPF , (1.1)
where H is the home country, F is the foreign country and E is the home currency price of the
foreign currency. Given some costs, such as transportation and barriers to trade, the absolute
version of the law of one price usually does not hold. Instead, another version may hold, for
example:
CHAPTER 1. A DISCUSSION ON EXCHANGE RATE PASS-THROUGH 13
PH = �EPF , (1.2)
where � indicates the deviation from the law of one price.
The model is built assuming that the foreign price of a good x, PF , is determined by the
markup over marginal cost,
PF = Markup ⋅ mc, (1.3)
where mc is the marginal cost.
Markup is a function of industry-specific factors, �, and the general macroeconomic conditions,
proxied by the exchange rate, E, as follows:
Markup = �E�, (1.4)
where � is the elasticity of the exchange rate. Marginal cost mc is determined by the prices of
substitutes goods and services, PS , the cost of inputs of good x in the producer country, W , and
income, Y , as follows:
mc = P�SW 1Y 2 (1.5)
Rewriting the above equations, we have:
PH = ��E(1+�) ⋅ W 1P�S Y 1 , (1.6)
or, applying the logarithm,
CHAPTER 1. A DISCUSSION ON EXCHANGE RATE PASS-THROUGH 14
logPH = log(��) ⋅ logE(1+�) ⋅ log(P�SW 1Y 2), (1.7)
we get a additive model, as follows:
pH = log(��) + (1 + �)e+ �pS + + 1w + 2y, (1.8)
where the small caps represents variables logs.
Goldberg and Knetter (1997) show that the econometric model specification generated in equa-
tion (1.8) is exactly the same as any other widely accepted theoretical approach to study prices and
exchange rate pass-through, such as the pricing-to-market model presented by Krugman (1997).
The econometric specification and estimation will be the subject of the next chapter.
Chapter 2
Pass-Through Estimation in Brazil
2.1 Introduction
The are few studies estimating the exchange rate pass-through in Brazil. In this chapter we present
a state space model to estimate the exchange rate pass-through in Brazil from August 1999 to
August 2008. The state space framework is suitable to build a econometric model based on the
economic model discussed in section 1.8 that is suitable to address some stylized facts presented
on the literature on exchange rate pass-through.
One of the motivations of the proposed econometric model is that, as argued by Parsley (1995),
stability of exchange rate pass-through is not well tested in common econometric specifications of
pass-through equations. Therefore, the state space formulation is suitable to build linear models
with time varying coefficients, allowing us to fill this gap in the literature. There are recent
contributions using state space models, such as Sekine (2002) and Albuquergue e Portugal (2005),
but they do not cover some key aspects of interest. For example, a importante contribuition of
15
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 16
our study is that we used novel the techniques presented by Pizzinga, Fernandes and Contreras
(2008) and Pizzinga (2009) to estimate the model with constraints in the time varying coefficients
to address the PPP puzzle in this framework.
We go further exploring the proprieties of state space models. We estimated a modified version
of the econometric model to test whether some “determinants” of exchange rate pass-through are
related to variations of the exchange rate pass-through over time. This is possible because we can
specify a movement equation to the time varying coefficients with explanatory variables. Although
our methodology does not allow to claim what is the direction of the casual effects, it offers new
evidence on the so called “determinants” by the literature.
With the purpose of controlling for aggregation bias, we estimate both specifications for different
levels of aggregation of the wholesales Brazilian price index used. The model is estimated for the
IPA-OG series, including its versions for industrial products, the IPA-OGPI, and agricultural
products, IPA-OGPA.
2.2 Econometric Framework
2.2.1 Linear state space models under restrictions
We define a linear Gaussian state space model by the following measurement equation, state equa-
tion and initial state vector:
Yt = Zt�t + dt + "t , "t ∼ NID(0, Ht)
�t+1 = Tt�t + ct + �t , �t ∼ NID(0, Qt)
�1 ∼ N(b1, P1).
(2.1)
The former equation linearly relates the observed time series Yt to the unobserved state �t and
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 17
the latter gives the state evolution through a Markovian structure. The random errors "t and
�t are independent (of each other and of 1), and the system matrices Zt, dt, Ht, Tt, ct and
Qt are deterministic. Notice that dt and ct are generally reserved to the inclusion of exogenous
explanatory variables.
For a given time series of size n and any t,j, ℱj ≡ � (Y1, . . . , Yj), �t∣j ≡ E (�t∣ℱj) and Pt∣j ≡
V ar (�t∣ℱj). The Kalman filtering consists of recursive equations for these first and second order
conditional moments. The formulae and their respective deductions corresponding to predicting
(j = t − 1), filtering (j = t) and smoothing (j = n), as detailed in the estimation of unknown
parameters in the system matrices by (quasi) maximum likelihood, can be found in Harvey (1989)
and Durbin and Koopman (2001).
Now, suppose the following: for each t, At�t = qt, where At is a known k × m fixed matrix
and qt = (qt1, . . . , qtk)′
is a k × 1 observable vector, may be random. Also suppose that qt is
ℱt-measurable. A restricted estimation of this type can be achieved under the restricted Kalman
filtering, presented in Pizzinga and Fernandes (2008) and summarized in the following algorithm:
Let t be an arbitrary time period.
1. Re-write the linear restrictions as
At,1�t,1 +At,2�t,2 = [At,1 At,2](�′t,1, �
′t,2
)′= qt, (2.2)
where At,1 is a k × k full rank matrix.
2. Solve (2.2) for �t,1:
�t,1 = A−1t,1 qt −A−1t,1At,2�t,2. (2.3)
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 18
3. Take (2.3) and replace it in the measurement equation of model (2.1):
Yt = Zt,1�t,1 + Zt,2�t,2 + "t
= Zt,1(A−1t,1 qt −A
−1t,1At,2�t,2
)+ Zt,2�t,2 + "t
= Zt,1A−1t,1 qt − Zt,1A
−1t,1At,2�t,2 + Zt,2�t,2 + "t
⇒ Y ∗t ≡ Yt − Zt,1A−1t,1 qt =(Zt,2 − Zt,1A−1t,1At,2
)�t,2 + "t
≡ Z∗t,1�t,2 + "t.
4. Postulate a transition equation for the unrestricted state vector �t,2 and finally get the
following reduced linear state space model:
Y ∗t = Z∗t,2�t,2 + "t , "t ∼ (0, Ht)
�t+1,2 = Tt,2�t,2 + ct,2 +Rt,2�t,2 , �t,2 ∼ (0, Qt,2)
�1,2 ∼ (a1,2, P1,2).
(2.4)
5. Apply the usual Kalman filter to the model in (2.4) and obtain t,2∣j , for all j ≥ t.
6. Reconstitute the estimates �t,2∣j :
�t,1∣j = A−1t,1 qt −A−1t,1At,2�t,2∣j . (2.5)
As Pizzinga and Fernandes (2008) claim, an interesting feature of this approach is that there is
no need to worry about specifying the state vector equation until the reduced form is achieved in the
4th step of the described algorithm. This avoids any risk of obtaining an augmented measurement
equation that is theoretically inconsistent with the original state equation. Another good property
that should be noted, and that will be used later in this paper, is that the reduced restricted
Kalman filtering enables us to investigate the plausibility of the assumed linear restrictions by
using information criteria (e.g. AIC and BIC).
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 19
2.2.2 Model Selection and Inference
One of the purposes of this study is to identify the most adequate number of lags in the exchange
rate. The hypotheses of completeness (or absence) of exchange rate pass-through will also be
verified. To accomplish this we use the following steps:
1. Diagnostic tests with the (standardized) residuals.
2. Information criteria, such as AIC and BIC.
3. Predictive power by comparing PseudoR2 and MSE measures.
Finally, the statistical significance of the parameters of measurement and state equation will
be tested under a likelihood ratio (LR) testing approach. Since both the reduced and the com-
plete model maintain the standards of good properties of maximum likelihood estimation (cf.
Pagan, 1980), it follows that, asymptotically, LR ≡ 2 [logLMax,Comp − logLMax,Red] ∼ �21, where
logLMax,Red represents the maximum of the log-likelihood for a model with a particular explana-
tory variable dropped from the specification.
2.3 Econometric Setting and Estimation
2.3.1 Time Varying Coefficients
The dependent series are the Wholesale Price Index, Global Supply (IPA-OG), Wholesale Price
Index, Global Supply - Industrial Products (IPA-OGPI) and Wholesale Price Index Global Supply
- Agricultural Products (IPA-OGPA) all created by the Getulio Vargas Foundation in Brazil. The
IPA are the best proxy for a producer price index in Brazil and for this reason they are used in
this work. We also used the controlled consumer price index, IPCA-MP to illustrate how useful
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 20
the Pizzinga and Fernandes (2008) technique can be. The monthly average commercial exchange
rate (bid), the monthly GDP, both published by Banco Central do Brasil (the institution that has
a similar role to the American FED in Brazil) and the American PPI, Industrial Commodities
are the explanatory variables. These variables are chosen because they are good proxies to the
variables presented in equation 1.8. The GDP is proxy for income, the PPI is proxy for the cost of
production in the foreign country. As the indexes analyzed work are aggregates of many different
goods, no substitute index prices was adopted in this study.
The sample has data from August 1999 to August 2008. A longer period would be desirable,
however Brazilian economic history lacks longer periods of economic stability. For example, from
March 1994 to January 1999 Brazil experienced the adoption of the Real Plan to fight high inflation.
Much of the strength of this new plan was set on the fixed exchange rate system. However,
a sequence of international crises in the nineties made Brazil change this regime for a floating
exchange rate with inflation targeting regime in February 1999. As it always takes time for economic
agents to adapt themselves to new environments and since we also need to use some lags in the
exchange rate to correctly specify our model, we decided to estimate the proposed model using
observations since August 1999.
Since we decided to investigate whether exchange rates had a contemporaneous effect on Brazil-
ian wholesale prices, it is necessary to correctly deal with a possible endogeneity between the log
difference of exchange rate and the price indexes. This is done using the results in Kim (2006).
The instrumental variables used for the growth rate of the exchange rate are lags in the growth
rate of the exchange rate itself, lags in growth rate of the American PPI, industrial commodities,
lags in growth rate of Brazilian consumer price index, IPC, and the Brazilian IPA-OG. Among
all these instrumental variables, only lags in the exchange rate were statically significant in the
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 21
auxiliary regressions.
We now present our state space model for the exchange rate pass-through for a given index
price. The model is composed by a measurement equation, namely,
Δ log pt =∑mk=0 �ktΔ log et−k + �(m+1)tΔ log pt−1 + 0 + 1Δ log ppit−1 + 2Δ log yt−1 + "t,
"t ∼ NID(0, �2)
(2.6)
and state equation, as follows:
�t+1 = �t + �t, �t ∼ NID(0, Q). (2.7)
The former equation linearly relates the observed monthly log-variation of the domestic price index
to the log-variation of exchange rate from time t to time t − m and to the American Producer
Price Index, ppi and to a demand variable, yt−1. The coefficients of Δlog et−k in equation (2.8)
are the state coefficients and their dynamics are given in equation (2.9).1 The lagged term pt−1 is
introduced to deal with persistence observed in the inflation indexes. As proposed by Kim (2006),
we added residual terms from the auxiliary regression in the measurement equation to control
for endogeneity. The matrix Qm×m is set diagonal for simplicity. As in Sekine(2006), the estate
equation implies that all shocks have permanent effect on the time varying coefficients. Although,
it seems a oversimplifying assumption, it has many advantages. For example, small variance terms
in matrix Q provides evidence that the constant coefficients is the most adequate formulation.
The exchange rate pass-through literature has many studies arguing that the exchange rate pass-
through is declining over time. Therefore, a stationary moving average formulation in the estate
1Many attempts were made with different lags structures in the explanatory variables. The lag structure adopted
in the presented formulation was the one with best results.
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 22
equation would imply a undesirable mean reversion behavior on the coefficients movement that is
not supported by the literature. Finally, specifications with up to 12 lag exchange rate terms were
tested. An AR(1) formulation for the coefficients would have (at least) 24 more parameters than
the adopted in this study, which would imply in a worthless computational enforce.
As proposed by Kim (2006), we added residual terms from the auxiliary regression in the mea-
surement equation to control for endogeneity. The reducing method from the previous subsection
has been used in order to impose the restrictions of the Purchasing Parity (PPP) or Producer
Currency Pricing (PCP) hypotheses, that is,∑m+1i=0 �it = 1, and of the Local Currency Pricing
(LCP) hypothesis, i.e. null pass-through∑mi=1 �it = 0. The completeness of the exchange rate
passing-through means that all the variation of the exchange rate is passed to the domestic prices.
This is a key question for Economic Theory, since accepting it is implies accepting the PPP hy-
pothesis. On the other hand, the accepting that null exchange rate pass-through model is the
most adequate scenario implies that the exchange rate movements do not have an effect in the
domestic prices, and it follows that the monetary authority need not be concerned with exchange
rate movements to make monetary policy with price indexes.
The proposed model shows a good fit for all IPA-OG cited series, as can be seen in table 2.1,
below. For IPCA-MP the goodness of fit was not good, as expected. Since ICPA-MP is a index
of controlled prices, its movements are determined by political decisions, contracts and other ways
not considered by our economic model. Besides that, the estimation results for this series are
present to illustrate the methodology propose by Pizzinga and Fernandes (2008). For IPA-OGPI
series we included a 6 lags term for the dependent variable to control for autocorrelation pattern
in the residuals.
As a first exercise, we test whether the the null (LCP) and full (PCP or PPP) exchange rate
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 23
Table 2.1: Some quality of fit statistics of the adjusted models.
Model Pseudo R2 MSE
IPA-OG 0.635 0.616
IPA-OGPA 0.378 3.796
IPA-OGPI 0.711 0.461
IPCA-MP 0.014 2.513
pass-through hypotheses are acceptable for the data, as seen in table 2.3. According to the values
of the AIC and BIC criteria, we have no evidence that both hypothesizes of none and full exchange
rate pass-though are the most adequate for the IPA-OG, IPA-OGPA and IPA-OGPI. Therefore, we
have evidence that there exist a partial exchange rate pass-through in Brazil in the sample period
for these series. The exception is the series IPCA-MP, the monitored consumer prices index. Since
the prices are controlled by the government, we do not expect to have a pass-through greater than
zero for this series, as we are using monthly data. 2. This is confirmed by the results. The model
for IPCA-MP shows some evidence that its exchange rate pass-through is zero in the long run, as
indicated by the AIC and BIC information criteria. This is a excepted result as the government
decisions regarding prices rely more on political aspects than on economic ones.
According to previous findings, the relation between exchange rate changes and inflation seems
to be statistically significant for different prices indexes series analyzed in this study, as can be seen
in Figures 2.1, 2.2 and 2.3. From these figures, we have evidence that coefficients have vared in
Brazil since 1999. Moreover, these figures indicate that this relation is declining over time, which
2In Brazil, the controlled prices, such as rent, public transportation, educational, among others, have a annual
schedule of readjustment. Additionally, the political agenda decides whether some cost raise will passed through
another group of prices (fuels are a good example of this group)
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 24
Table 2.2: Information criteria observed values for incomplete, null and complete pass-through
exchange rate models.
Series Criterium unrestricted no pass-through full pass-through
IPA-OGAIC 2.065 2.461 3.082
BIC 2.338 2.684 3.306
IPA-OGPAAIC 3.987 4.119 5.077
BIC 4.260 4.343 5.300
IPA-OGPIAIC 2.129 3.191 2.940
BIC 2.452 3.464 3.213
IPCA-MPAIC 2.685 2.683 4.049
BIC 2.958 2.907 4.272
Table 2.3: P-values of the tests for null and complete pass-through exchange rate.
suggests that the estimates with constant pass-through coefficients are not valid.
Our results suggest that the contemporaneous effect of dollar variation in the analyzed index
prices variations is greater than zero. Its estimated values are mostly constant over time, but
the lagged effects are varying for the IPA-OG, IPA-OGPA and IPA-OGPI series. Therefore, it
is important to point out that almost all of the exchange rate pass-through verified is due to the
amount of the lagged exchange rate variation passed through to prices. This is reinforced by the fact
that the confidence interval for the smoothed coefficient of lagged exchange rate variation contains
zeros in some periods and does not in others. This may favor the macroeconomic environment
effect over pass-through. Some possible explanations for this are that depending on the credibility
of the central bank, the inflationary environment or the economic growth, the price-maker could
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 25
2000 2001 2002 2003 2004 2005 2006 2007 2008
0.1
0.2
2000 2001 2002 2003 2004 2005 2006 2007 2008
−0.25
0.00
0.25
2000 2001 2002 2003 2004 2005 2006 2007 2008
0.55
0.60
0.65
0.70
Figure 2.1: IPA-OG smoothed coefficient of Δ log et, Δ log et−1 and Δ log yt−1.
change the speed of his price adjustment. These hypotheses are going to be investigated later.
The inclusion of the lagged dependent variable with a time varying parameter helps us inves-
tigate whether there is variability of inflation persistence. For example, there are some authors
that argue that persistence in the inflation rate is greater during high inflation periods. If were
the case, the higher long run pass-through during high inflation periods could be a consequence
of higher persistence. Our estimates show that this is not the case for the series analyzed. Al-
though constant, the persistence is very high and statistical significant, around 0.60, for every
series. The lagged 6 IPA-OGPI term in Figure 2.3 is not significant, however it controls for serial
autocorrelation in the residuals. Therefore, we decide to keep it in the model to avoid inconsistent
estimators.
The estimates shown in figures 2.4, 2.5 and 2.6 are the long run estimates for the pass-through.
They present the same picture seen above: the exchange rate pass-through seems to be declining
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 26
2000 2001 2002 2003 2004 2005 2006 2007 2008
0.00
0.25
0.50
2000 2001 2002 2003 2004 2005 2006 2007 2008
0.0
0.5
2000 2001 2002 2003 2004 2005 2006 2007 2008
0.5
0.6
Figure 2.2: IPA-OGPA smoothed coefficient of Δ log et, Δ log et−1 and Δ log yt−1.
over time. Although there are some periods in which it is not verified (i.e. from 2002 to 2003 and
from 2005 to 2006), its value keeps declining until where we see slightly increase 2008.
The high persistence produces high variations in the long run pass-through. For example, it
reaches values as high as 0.90 during the crisis period of 1999 and 2002 for all studied series. It
highlights the importance of the autoregressive term in the measurement equation for the long run.
The agricultural prices have had a step decline since 1999. Since the two peaks of 1999 and
2002, the long run pass-through has declined and converged to almost 0.20. One of the possible
explanations to this is a increase in the competition.
Industrial prices have a similar behavior, with more intense decline. After reaching values
around 1 in 2002, the exchange rate pass-through estimated values were around 0.1, 10% of the
value in the crisis period. It is important to notice that there was a peak in the decreasing trend
from 2004 to 2005. In 2004, Brazil faced a strong GDP growth and an increase in trade volumes. At
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 27
2000 2005
0.00
0.05
0.10
0.15
0.20
2000 2005
−0.2
0.0
0.2
2000 2005
0.55
0.60
0.65
2000 2005
0.00
0.05
0.10
0.15
Figure 2.3: IPA-OGPI smoothed coefficient of Δ log et, Δ log et−1 (top), Δ log yt−1 and Δ log yt−6
(bottom).
the same time, world demand was growing sharply. This led to an increase in demand of industrial
goods because of a lack of competition. Therefore, cost movements and exchange rate changes
were more easily passed through to prices, including exchange rate changes. That is a possible
reason why the pass-through increased in this period. The central bank was forced to implement
a strong restrictive monetary policy that caused a reversion in inflation expectations and reduced
the exchange rate pass-through.
An important fact is the increase in pass-through in 2003 before the Brazilian elections. The
fear of macroeconomic policy changes caused a decrease in foreign investments. The expectation
was that the exchange rate would be devalued for a long time, consequently, agents anticipated
this expected devaluation and changed their prices. However, they realized that economic policies
would be continued, then the exchange rate decreased to its long term value and agents passed
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 28
2000 2001 2002 2003 2004 2005 2006 2007 2008
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Figure 2.4: IPA-OG long run pass-through.
2000 2001 2002 2003 2004 2005 2006 2007 2008
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Figure 2.5: IPA-OGPA long run pass-through.
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 29
1999 2000 2001 2002 2003 2004 2005 2006 2007 2008
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
Figure 2.6: IPA-OGPI long run pass-through.
through this valuation of domestic currency.
These results strongly reinforce the belief that the pass-through is declining in Brazil for whole-
sale prices. It is important to note that 1999 and 2002 were periods of much domestic uncertainty.
In January 1999, Brazil shifted from fixed to a flexible exchange rate regime and suffered a strong
crisis of credibility. According to the Calvo and Reinhart (2000) both lack of credibility and
volatility of exchange rate are linked to a high exchange rate pass-through.
We tested whether the explanatory variables are significant in the model and we found that none
of the coefficients in the measurement equation were significant at the usual levels of significance.
Likelihood tests were conducted for the constant coefficients in the model. The results in Table
2.4 show that for all coefficients, the null hypothesis that ( i = 0, i = 0, 1, 2), cannot be rejected
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 30
Table 2.4: Estimates of the IPA-OG, IPA-OGPA and IPA-OGPI series (p-values between paren-
thesis).
Model constant vt pibt−1 ppit−1
IPA-OG0.3233 -0.024 0.017 0.082
(0.000) (0.505) (0.249) (0.071)
IPA-OGPA0.475 -0.062 0.027 0.002
(0.021) (0.537) (0.508) (0.986)
IPA-OGPI0.276 -0.023 0.019 0.078
(0.003) (0.512) (0.197) (0.071)
IPCA-MP0.428 -0.040 -0.001 0.085
(0.000) (1.000) (0.962) (0.165)
at the usual significance levels3.
2.3.2 Determinants
Some of the changes in the Brazilian economy appear to have exacerbated fluctuations in exchange
rates. The liberalization of capital flows in the last two decades and the increase in the scale of
cross-border financial transactions have increased exchange rate movements. Currency crises in
emerging market economies are unique examples of high exchange rate volatility. In Brazil, these
large movements in exchange rate may be associated with greater pass-through to prices as seen
in Figure 2.7.
Given the results presented in the previous section, we reformulate the model. Since the
3Here we remark that none correction for the log likelihood ratio significance test were implemented for theresiduals of the auxiliary residuals. So, for its coefficients are used for illustration only.
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 31
2000 2002 2004 2006 2008
0.00
0.05
0.10
0.15
0.20
Time
Figure 2.7: Tested determinants of pass-through: Monetary policy, (solid line), variance of ex-
change rate (dashed line) and international trade (dotted line).
coefficients of the contemporaneous effect of the exchange rate on the wholesale price indexes are
mostly constant, we decide to introduce them with constant coefficients. The same was done to
the persistence coefficients. In addition, all variables statistically null in the former model were
excluded from the actual one. The only exception was the error term to control for endogeneity
of the contemporaneous log difference of exchange rate pass-through. Futhermore, we tried to test
the importance of some determinants of exchange rate pass-through. For this purpose, we made
changes in state equation and we introduced some explanatory variables in the state equation. As
will be clear in the next lines:
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 32
Δ log pt = �1tΔ log et−1 + 0 + 1Δ log e0 + 2Δ log pt−1 + "t, "t ∼ NID(0, �2) (2.8)
�t+1 = �t + dt + �t, �t ∼ NID(0, Q). (2.9)
The former equation linearly relates the observed monthly log-variation of price to the log-variation
of exchange rate from time t to time time t− 1 and to its own value at time t− 1. The coefficient
of Δlog et−1 in equation (2.8) is the state coordinate and its dynamics are given in equation (2.9).
This equation now has the explanatory variable (or “determinant”), dt, with coefficient and an
error term with variance Q.
Guided by the literature presented in the previous sections, we tested four explanatory variables
for the latent exchange rate pass-through coefficients: the difference between the exchange rate
variance of daily log returns from time t to time t − 1, dvdnert; the variation of the ratio of the
inflation expectation and the inflation target set by the central bank from time t to time t − 1,
dpmt; the log difference of the trade flow (given by the sum of exports and imports) divided by
the real GDP from time t to time t − 1, dlflowt; and the log difference of the Brazilian IPCA (a
consumer price index computed by the Brazilian Census Bureau, IBGE) from time t to time t− 1.
We also included one lag for each variable because all these variables are likely to be endogenous.
The coefficient gamma represents the effect of each “determinant” over the dynamic of exchange
rate pass-through.
The monetary policy measure is the change in inflation expectation over the inflation target.
In an inflation targeting regime, the central bank uses one monetary policy rule to accommodate
inflation expectations close to the target set before. In an attempt to identify the success of the
Brazilian Central Bank in stabilizing inflation expectations, we constructed PM as the moving
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 33
average of the difference of inflation expectations of 12 months ahead over the inflation target. To
control for the monetary policy’s forward looking behavior, we set a moving weight, as follows
pmt = (1/12)(Et(�t+12)− �t+12) + (1− j/12)(Et+1(t+ 13)− �t+13), (2.10)
for j = 1, . . . , 12.
Mishkin (2008) stated that the correlation between inflation and the rate of nominal exchange
rate depreciation can indeed be high in an unstable monetary environment in which nominal shocks
fuel both high inflation and exchange rate depreciation. Furthermore, the evidence suggests that
even countries where inflation and exchange rate depreciation appear to be fairly closely linked
over time, have experienced a sizable decline in pass-through following the adoption of improved
monetary policies. To test if the credibility of the Brazilian Central Bank has been decreasing the
exchange rate pass-through, we created a monetary policy variable as the inflation expectation (by
Boletim FOCUS) over the inflation target. The credibility and the efficiency of the monetary policy
tend to hinder the ability of price makers to adjust their prices. In an stable inflation environment,
the agents are less likely to adjust their prices.
Taylor (2000) argues that the exchange rate pass-through has a positive relation to the per-
sistence of costs changes. If the volatility of changes in exchange rates is associated with its
persistence, smaller volatility periods will be followed by a smaller degree of pass-through. The
volatility in exchange rate can represent uncertainty in the economy, where large exchange rate
movements could more be easily passed through to prices.
Our results show a statistical significant association between exchange rate volatility and pass-
through. In periods with high uncertainty, large variations in exchange rate are positively correlated
with pass-through. We obtained evidence that the 12 month variance for exchange rate explains the
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 34
Table 2.5: Estimated parameters and corresponding p-values (in parenthesis).
Series dvdnert dvdnert−1 dpmt dpmt−1 dflowt dflowt−1
IPA-OG2.230 2.273 2.977 −0.912 −0.304 −0.328
(0.001) (0.000) (0.051) (0.493) (0.265) (0.223)
IPA-OGPA3.493 3.687 4.097 −1.027 −0.416 −0.427
(0.002) (0.002) (0.231) (0.757) (0.028) (0.028)
IPA-OGPI1.717 1.646 −5.246 −5.247 −0.143 −0.198
(0.033) (0.045) (0.036) (0.036) (0.665) (0.551)
Table 2.6: Estimated parameters and corresponding p-values (in parenthesis).
Series dlntradeflow/GDPt dlntradeflow/GDPt−1 dlnIPCAt dlnIPCAt−1
IPA-OG−0.178 −0.179 −0.005 −0.005
(0.551) (0.569) (0.484) (0.423)
IPA-OGPA−0.535 −0.565 −0.007 −0.006
(0.033) (0.026) (0.322) (0.291)
IPA-OGPI−0.224 −0.136 −0.001 −0.002
(0.447) (0.666) (0.858) (0.817)
pass-through dynamics for all price indexes, more strongly agricultural prices. The trade openness
variable is only significant for exchange rate pass-through to agricultural prices. If an economy
is more open to foreign goods, it will face more competition and market power of producers will
decrease. For the whole IPA and for industrial products, the increase in imports share is less
pronounced and the effect over pass-through decline is not statistically significant. In the case of
agricultural prices, the sharp increase in imports and exports increased the competition level in
CHAPTER 2. PASS-THROUGH ESTIMATION IN BRAZIL 35
the sector and reduced the propensity to pass-through cost changes to prices.
For the monetary policy variable, we found mixed results. For agricultural prices, the monetary
policy did not explain the pass-through. However, the exchange rate pass-through to industrial
prices was negatively associated with the monetary policy, where a large misalignment between
inflation expectations and the target is related to a smaller pass-through. We found that credibility
and a smaller deviation of expectation over the inflation target decrease the incentives to readjust
prices for the IPA only.
The variable dlnIPCA was introduced as an attempt to capture the inflationary environment.
With this variable we did not obtain evidence that the inflation environment affects the exchange
rate pass-through, as shown by table 2.6.
2.4 Conclusions
In this paper we estimated the evolution of the exchange rate pass-through for some wholesale
indexes prices in Brazil with a Gaussian state space model.
Using our formulation we were able to investigate some important aspects as endogeneity be-
tween exchange rate pass-through and the indexes prices, aggregation effects and persistence vari-
ation over time. We were also able to investigate the significance of some possible “determinants”
of exchange rate pass-through.
The estimates shown suggest that the the short run and long run exchange rate pass-through
are declining over time. Around 2002, the presidential election year President Lula ran for office,
the short run pass-through had risen to approximates one. Since then, the short run pass-through
has followed its decreasing trend and stabilized in 2008.
Other results about inflation persistence show that estimated coefficient of the lagged dependent
variable are constant over time, indicating that the persistence of inflation is not varying. This
implies that if the long run pass-through changes over time, it is be due to the variation of the
short run pass-through.
We did not find strong evidence that there are important endogeneity from Brazilian wholesale
price indexes on exchange rate.
The data for the wholesales indexes does not support the null and the full exchange rate pass-
through hypotheses. This reinforces the belief that there exists a positive, although incomplete,
exchange rate pass-through in Brazil. For illustration propose, we estimated our model to a
Brazilian price index for monitored prices. In this case the estimates confirmed our previous belief
that there is no exchange rate pass-through for monitored prices.
Finally, we motivated and tested the importance of set exchange rate pass-through determinants
suggested by the literature on exchange rate pass-through. We obtained strong evidence that the
variance of exchange rate causes a greater pass-through to prices. We also obtained evidence that
some variables are able to explain the pass-through of some index prices but not of others. For
example, we found evidence that adjusting monetary policy led to a reduction in the pass-through
to industrial prices but not to agricultural prices. On the other hand, an increase in trade flow
results in a decrease the pass-through to agricultural prices but not to industrial prices. We had
no evidence that the inflationary environment was able to cause pass-through.
Part II
Coincident and Leading Indexes of
Economic Activity
37
Chapter 3
A State Space Model for Indices
of Economic Activity
3.1 Introduction
Traditionally, business-cycle research has focused on sophisticated econometric models aiming to
capture the main features of either GDP or of the four coincident variables that the NBER is
said to follow (employment, industrial production, income and sales) to estimate coincident and
leading indices of economic activity, establish business-cycle turning points, as well as to estimate
their respective probability of occurrence; see Stock and Watson (1988a, 1988b, 1989, 1991, 1993a),
Hamilton (1989), Kim and Nelson (1998), Harding and Pagan (2003), Hamilton (2003), and Chau-
vet and Piger (2008), inter-alia. Arguably, these models mis a key variable that should be included
in them – the NBER decisions on U.S. turning points as determined by its business-cycle dating
38
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 39
committee. Although this information is usually available with a considerable lag, there is no
reason not to include it ex-post on econometric models. This point was forcefully made in Issler
and Vahid (2006).
There has been a recent trend of incorporating NBER dating-committee decisions into different
business cycle econometric models. Although some of these contributions are independent, they all
recognize that one should not discard the informational content of these decisions when constructing
econometric models; see Birchenhall et al. (1999), Dueker (2005), Issler and Vahid, and Chauvet
and Hamilton (2006). A key aspect of the NBER dating committee is that there are some changes in
its members through time. Additionally, shocks hitting the economy affect GDP and key economic
variables that the NBER is said to follow in a different manner, either happens because these
shocks vary across time (i.e., supply shocks in one recession and demand shocks in another) or
because some of these relationships are indeed not stable. Thus, in building econometric models
using the NBER-committee decisions we should consider the possibility of time-varying weights in
econometric relationships.
Our first original contribution is to propose a state-space model with time-variable weights
using the decisions to construct coincident and leading indices of economic activity for the U.S.
economy. Our model is a probit regression of NBER decisions on the coincident series, where
instrumental-variable techniques are needed to consistently estimate time-varying weights of this
index. In estimation, we apply the extended iterated Kalman filter and use the Rivers and Voung
(1988) procedure to correct for simultaneity. Also, we account for the fact that NBER decisions
on whether there is or not a recession at time t is made well into the future, i.e., in time t + ℎ,
ℎ > 0.
We use canonical-correlation analysis to extract the cycles from the coincident variables. With
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 40
these cycles, we run the instrumental-variable state-space probit regression mentioned above. Using
the probit regression we are able to estimate the time-varying coefficients related to each coincident
series: the so called kalman Filter Coincident Index (KFCI) and Kalman Smoother Coincident
Index (KSCI). Here, we integrate state-space research with the probit regression and instrumental-
variable techniques which proposes a unique algorithm to estimate a simple coincident index with
a good track record vis-a-vis NBER decisions. We employ the fact that the NBER Business Cycle
Dating Committee uses information available at time t + ℎ, where ℎ > 0, to decide whether the
economy is in recession or not at time t. This generates an MA(ℎ) structure in the error term of
the probit regression.
Despite the fact that the NBER decisions use future information, the econometrician cannot do
the same if he/she is interested in building models that are useful in real time. Because we want
our model to be useful in real time, our techniques use only lagged information to forecast current
and future variables in our model. This leads us to our second original contribution which is shows
that our real-time model performs well in predicting the probability of a recession in a real-time
setting. We illustrate the model’s predictive ability with data from the past three recessions: 1990,
2001, and 2007.
In an out-of-sample exercise, the parameters of the model are estimated using information up
to 1 year prior to each of the last 3 recessions. Variable weights are predicted into the future. Using
the real time value of the coincident series we then estimate the probability of the NBER declaring
a recession in real time. For all three recessions we show that our model is useful if the objective
is to have a reliable real-time estimate of the NBER decisions. For example, when using a cutoff
of 50% for probabilities, our filtered estimates predict 84.93% of the recession periods correctly.
We then present evidence that the weights of the coincident series of employment is very high,
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 41
greater than 0.5, in our sample period. This agrees with a memo from the Business Cycle Dating
Committee (Hall et al., 2002) which states that “employment is probably the single most reliable
indicator [of recessions]”.
Finally, based on our coincident index estimate, we construct an optimal variable-weight leading
index of economic activity using the results from the canonical-correlation analysis. The optimal
variable-weight leading index has an important role in our method. It extracts the cycles from the
coincident series that have common features with the “business cycle”, where by cycle we mean the
information that can be linearly predicted from the past. This step is important for two reasons: it
allows us to separate signal and noise in state-space estimation and it allows combining the present
and past by linking coincident and leading indexes, respectively.
3.2 The model
3.2.1 Determining a basis for the cyclical components of coincident vari-
ables
One of the innovations proposed by Issler and Vahid (2006) was the use of the statistical technique
in canonical correlation analysis (Hotelling, 1935 and 1936) to create a coincident and a leading
indices of economic activity. As stated before, the canonical correlation analysis is important for
the extraction of the non-cyclical features from the coincident series that could introduce noise
in state-space estimation (Chauvet, 1998); thus, allowing the combination of present and past
when linking coincident and leading indices, respectively. Also the canonical correlation analysis
is suitable in dealing with possible asymmetric cycles in coincident series. As the use of canonical
correlation analyses were well motivated by Issler and Vahid’s work, we will only mention some
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 42
essential points for understanding the remaining of paper.
Canonical analysis maximizes correlations among all possible linear combination in both sets
of the coincident and leading series. Denote the set of coincident variables with the vector xt =
(x1t, x2t, x3t, x4t)′ and the set of m (m ≥ 4) “predictors” by the vector zt (this includes lags of
xt as well as lags of the leading variables). Canonical-correlation analysis transforms xt into four
independent linear combinations A(xt) = (�′1xt, �′2xt, �
′3xt, �
′4xt) with the property that �1xt
is the linear combination of xt which is the most linearly predictable of zt, �2xt is the second
most predictable linear combination of xt from zt after controlling for �1, and so on. These linear
combinations are uncorrelated with each other and they are restricted to have unit variances, so
that they can be uniquely idenfified up to sign change. By-products of this analysis are four linear
combinations of zt = ( ′1zt, ′2zt,
′3zt,
′4zt) with the property that ′1zt is the linear combination
of zt, which has the highest squared correlation with �′1xt, for i = 1, 2, 3, 4. Again, the elements
of G(zt) are uncorrelated with each other, and they are uniquely identified up to a sign change
with the additional restriction that all four have unit variances. The regression R2s between
�ixt and ′izt zt for i = 1, 2, 3, 4 which we denote by the squared canonical correlations between
(r2l , r22, r
23, r
24) xt and zt.
We can use a simple statistical test procedure to examine whether the smallest canonical cor-
relation (or a group of canonical correlations) is statistically equal to zero. The likelihood ratio
test statistic for the null hypothesis that there are k significant cycles (i.e., there are 4 − k zero
canonical correlations) is
LR = −T4∑
i=k+1
ln(1− �2i )
which has an asymptotic �2 distribution with (4 − k)(m − k) degrees of freedom (see Anderson,
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 43
1984). It is customary to use (T − m) instead of T in the above statistic to improve its finite
sample performance. If the null is not rejected, then the linear combinations corresponding to the
statistically insignificant canonical correlations cannot be predicted from the past and therefore,
can be dropped from the set of basis cycles. In this case, we conclude that all cyclical behaviors in
the four coincident series can be written in terms of less than four basis cycles. Hence, the use of
linear combinations of xts that are not associated with a zero canonical correlation is equivalent
to using only the cyclical components of the coincident series. Any linear combination of the
significant basis cycles is a linear combination of coincident variables, which is convenient for our
purposes as it implies that our coincident index will be a linear combination of the coincident
variables themselves.
3.2.2 Estimating a structural equation for the unobserved business cycle
state
After we obtained the cycles estimates, we combine them to construct our coincident index of
economic activity. In building a model with this aim we consider a specification that accounts for
some technical difficulties. The data we would have liked to work with, the state of economy, is
not observable. Instead, we have a dummy variable, the NBER indicator, that signalizes whether
the economy is in a recession or not at time t to the NBER Business Cycle Dating Committee’s
best knowledge. A important characteristic of the NBER indicator is that the NBER Business
Cycle Dating Committee uses information available up to time t+ℎ, where ℎ is a positive integer,
to make its decision about the state of the economy at time t. These specific characteristics of the
available data brings technical difficulties that we explore in the next paragraphs.
There are key assumptions that enables us to estimate the coincident index. We start by the
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 44
following:
Assumption 1 There exists a time-varying linear index of (the cyclical parts of) the coincident
series that has the exact same correlation pattern with past information as the unobserved state of
the economy.
Although the index which has the same correlation pattern with the past only involves the
significant basis cycles (i.e., will not involve white noise combinations of the coincident series),
these basis cycles are themselves (time-varying) linear combinations of coincident series. Hence,
the index is ultimately a linear combination of coincident series.
Let yt denote the unobserved state of the economy and {c1t, c2t, c3t} denotes the significant
basis cycles of the coincident series at time t. Assumption 1 clearly implies that there must be a
time varying linear combination of yt and {c1t, c2t, c3t} that is unpredictable from the information
before time t. That is,
E(yt − �0 − �1tc1t − �2tc2t − �3tc3t∣It−1) = 0, (3.1)
where It−1 = (xt−1,xt−2, . . . ; zt−1, zt−2, . . .) is the set of all observed values of the coincident and
leading variables until time t− 1.
Rewriting Eq. (3.1), we obtain
E(yt∣It−1) = E(�0 + �1tc1t + �2tc2t + �3tc3t∣It−1). (3.2)
Assumption 2 �∗it and cit are independent given It−1.
Assumption 2 suggests that the way the NBER makes its decisions is not cyclical nor does it
depend on business cycles.
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 45
Using Assumption 2, we have
E(yt∣It−1) = �0 +E(�1t∣It−1)E(c1t∣It−1)+E(�2t∣It−1)E(c2t∣It−1)+E(�3t∣It−1)E(c3t∣It−1). (3.3)
It is important to remember that we do not observe E(cit∣It−1), but cit at time t, i = 1, . . . , 3.
Therefore, Equation 3.3 is rewritten as follows:
E(yt∣It−1) = �0 + E(�1t∣It−1)c1t + E(�2t∣It−1)c2t + E(�3t∣It−1)c3t + !t, (3.4)
where E(!t∣It−1) = 0 and !t is obviously correlated with cit, i = 1, . . . , 3.
As previously mentioned, y∗i is not observable. Instead we only observe the NBER indicator.
The NBER indicator is set to one when, to the best knowledge of the NBER Dating Cycle Com-
mittee at time t+ℎ, the economy was in recession in time t. That is, the indicator that makes use
of information available at time t+ ℎ is bellow a critical value:
NBERt =
⎧⎨⎩1, if E(yt∣It+ℎ) < 0,
0, otherwise.
(3.5)
As we can always write
E(yt∣It+ℎ) = E(yt∣It−1) + �t + �t+1 + . . .+ �t+ℎ, (3.6)
using equation 3.4, we obtain
E(yt∣It+ℎ) = �0+E(�1t∣It−1)c1t+E(�2t∣It−1)c2t+E(�3t∣It−1)c3t+!t+�t+�t+1+. . .+�t+ℎ, (3.7)
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 46
where �t+i is the “surprise” associated with the new information arriving in period t + i. ut =
�t + �t+1 + . . . + �t+ℎ is unforeseeable given information at time t − 1 and imposes a “forward”
MA(ℎ) structure in the model.
As mentioned in the introduction, we believe that the NBER Dating Cycles Committee either
changes the weights of the coincident cycles according to the its composition, its members cu-
mulated knowledge over time or the political environment. These components are not stored in
the set of past values of the coincident and leading variables. Therefore, consider the following
assumptions:
Assumption 3 The time-varying coefficients �it, i = 1, 2, 3, do not depend on the past values of
the coincident and leading series stored in It−1.
Assumption 4 The changing weights mechanism acts over the weights of the basis cycles accord-
ing to the following law of movement:
�it = �it−1 + "it,
where "i ∼ N(0, �2i ) are independent white noise error terms, i = 1, 2, 3.
Assumption 4 deserves special attention. It postulates that the weights movement is given by a
random walk process, which has strong persistence. As these movements are caused by the events
previously cited, we believe it is a reasonable way to describe the way the Committee weights each
cycle over time. Actually, none of those motivations are likely to be transient (even the political
motivations are likely to act for more than a couple of months). By the econometric standpoint, it
allows a direct comparison with a model with constant coefficients. If the constant coefficients are
the correct way to deal with the weights of the coincident variables, the variance of innovations "it
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 47
will be close to zero and, therefore, the coefficients will be mostly constant.1
This formulation, leads us to a state space econometric model for business cycle. A adequate
way to deal with this framework is the Kalman filter, as it will be described later. However, before
describing the Kalman filter, there are some issues we have to deal with.
As pointed out in early paragraphs, c1t, . . . , c3t are correlated to !t. Moreover, the cycles
c1t, . . . , ckt are not observed, but estimated by canonical correlation analysis and modeled as cit =
�i( ′zt) + �it. These factors generate endogeneity in our model. Using the ideas of Rivers and
Vuong (1988) the correlation is modeled as follows:
⎛⎜⎜⎝ !t
�t
⎞⎟⎟⎠ ∼ N⎛⎜⎜⎝0,
⎡⎢⎢⎣ �2! �′!�
��! ��
⎤⎥⎥⎦⎞⎟⎟⎠ (3.8)
where the �it, i = 1, . . . , k, are collected in to a vector �t, �t and ′izt for i = 1, . . . , k come from
the canonical-correlation analysis, ��� is a k × k diagonal variance-covariance matrix of �t, and
��! is a k × 1 vector of covariances between ut and �t. Joint normality of !t and �t implies that
!t = �′t� + �t, (3.9)
where � = Σ−1����u, �t ∼ N(0, �2u − �′�uΣ−1����u) and �t is independent of �t. Substituting, for !t
in equation (3.7), we obtain
E(yt∣It+ℎ) = �0 +E(�1t∣It−1)c1t + . . .+E(�3t∣It−1)ckt + �t + �t+1 + . . .+ �t+ℎ + �′t� + �t, (3.10)
1Besides all these considerations, we obtained some estimates adopting a auto-regressive model for the movement
equation of the coefficients. The results were far worse than those obtained with a random walk process which
motivate us to consider only the simpler model.
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 48
Notice that, by construction, all the regressors in (3.10) are uncorrelated with the error term �t.
The normalization �2� = 1 make all the parameters identifiable. Since this transformation eliminates
the correlation between the error term and the regressors, we no longer have the endogeneity
problem.
We still have some estimation issues to deal with. First, there are many parameters and states
to be estimated what possibly would be too demanding computationally. Kim (2006) pointed
out this problem and adapted a two stage procedure to a Gaussian state space model with good
results. Second, as explained before, yt is not observed. Instead, we have the NBER indicator. As
the observable data is a dichotomic variable, we have to use limited information model. This is
done replacing the the Gaussian model by a probit one and, therefore, we have to estimate a state
space probit model under presence of endogeneity. Rivers and Vuong (1988) proposed a suitable
way of estimating probit models with endogenous regressors. They prove their procedure produces
strongly consistent estimates and their estimators are asymptotically normally distributed.
Adopting a probit model, the new equation to be estimated is given as follows:
Pr(NBERt = 1) = Φ(−(�0+E(�1t∣It−1)c1t+. . .+E(�3t∣It−1)c3t+�t+�t+1+. . .+�t+ℎ+�1v1t+�2v2t+�3v3t)).
(3.11)
Considering all drawbacks from the previous paragraphs, we use the ideas developed by Kim
(2006) and Rivers and Vuong (1988) to estimate the time varying coefficients, as shortly presented
in the following steps:
1. Regress cit, i = 1, . . . , k, on zt to get �it and ��, a consistent estimate of ��.
2. From �it i = 1, . . . , k , form �t and then obtain estimates of �t = (�0, �′) denoted by � and
of the states (�1t, �2t, �3t, �t, . . . , �t+ℎ).
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 49
3.2.3 The iterated extended Kalman filter and smoother
As we are interested in investigating the time varying structure of the coefficients, it is natural
to use a state space model to estimate �1t, �2t and �3t. State space models are well presented
in many textbooks,such as Durbin and Koopman (2002), Harvey (1989), among others. Among
many interesting aspects, one special feature of state space models is their ability to deal with
correlation patterns in the error term. This aspect will be used in our work to address the MA(ℎ)
structure in the residuals.
The Kalman filter, from to Kalman (1960), is a direct way to estimate state space models. It is
widely used to estimate Gaussian state space models. However, as we have a limited information
dependent variable, we must to use a probit state space model. Some methods of estimation of
non-Gaussian state space methods rely on Monte Carlo procedures. The method we decided to
use is the iterated extended Kalman filter and smoother, which is very well presented by Klein
(2003). Iterated extended Kalman filter and smoother is based on a Taylor expansion of the
probit equation, that, as an approximation, has less accuracy than other methods not based on
aproximations. Nonetheless, our results are evidence that we are able to predict the states of the
economy well. For instance, Klein (2003, chapter 4) compared the Markov Chain Monte Carlo
(MCMC) method with the iterated extended Kalman filter to estimate a state space model for
simulated data. He concluded that there were no important differences in favor of the MCMC
method.
The results of the Taylor expansion of equation (3.11), give us the following approximating
equation:
ˇNBERt = �( t)−1{NBERt − Φ( t)}+ t (3.12)
where � is the Gaussian density function, Φ is the Gaussian cumulative distribution function, and
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 50
t is the predicted value for NBERt, based on a initial guess (or some previous filtered value, if
iteration process was ran at least once) on the vector of states, as we clarify later. The measurement
equation of the approximating Gaussian state space process can now be written:
ˇNBERt = Z∗t �∗t + �t, �t ∼ N(0, Vt) (3.13)
were Z∗t = (1 v1t v2t v3t c1t c2t c3t 1 0 . . . 0) is a 1×7+ℎ vector, �∗t′ = (�0 �1 �2 �3 �1t �2t �3t �
′t)′
is a (7 + ℎ)× 1 vector,
�t =
⎛⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎝
ˇNBERt
�t + . . .+ �t−ℎ+2
�t + . . .+ �t−ℎ+3
...
�t
⎞⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎠(3.14)
is a ℎ× 1 vector and variance Vt = exp t(1 + exp t)−2�( t)
−2.
We write equation 3.13 using the notation and formulation proposed by Durbin and Koopman
(2002, p. 46 and p. 54) to address the MA(ℎ) structure in the model. The state equation is defined
as follows:
�∗t+1 = T ∗�∗t + �t, �t ∼ N(0, Q). (3.15)
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 51
where
T ∗ =
⎡⎢⎢⎣ I7 0
0 T
⎤⎥⎥⎦ , T =
⎛⎜⎜⎝ 0ℎ−1 Iℎ−1
0 0′ℎ−1
⎞⎟⎟⎠ , "t =
⎛⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎜⎝
07
"1
"2
"3
�t+1
...
�t+1
⎞⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎟⎠
, (3.16)
Q =
⎛⎜⎜⎜⎜⎜⎜⎝04×4 0 0
0 Σ" 0
0 0 Σ�
⎞⎟⎟⎟⎟⎟⎟⎠ , Σ" =
⎛⎜⎜⎜⎜⎜⎜⎝�2"1 �"1"2 �"1"3
�"2"1 �2"2 �"2"3
�"3"1 �"3"2 �2"3
⎞⎟⎟⎟⎟⎟⎟⎠ and Σ� = �2�Iℎ, (3.17)
noting that Ik is an identity k×k, 04×4 and 0k is a k×1 vector. The state space model is completed
with a initial distribution to the state vector: �∗t0 ∼ N(a0, Q0).
Finally, to obtain the desired estimates, we run the Kalman filter and smoother estimates using
the equations present above. The set of information that will feed the Kalman filter and smoother
will be {It, Dt}, where Dt is the set of all deliberations of the Committee stored until time t.
Again, the details of filtering ans smoothing procedures are well described by the textbooks cited
above.
3.2.4 The Kalman Filter and Smoother Instrumental Variables Index
In this section, we present the products of this paper. Using the state space approach, we gen-
erate two indexes. The first index is the Kalman Filter Instrumental Variable Coincident Index
(KFIVCI), namely,
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 52
KFIV CIt =
[E(�1t∣It, Dt) E(�2t∣It, Dt) E(�3t∣It, Dt)
] [�′1xt �′2xt �′3xt
], (3.18)
if n, the sample size is greater than t. For periods in which there is not a NBER Business
Cycles Committee deliberation (i.e., time n+q) there is available information about the coincident
variables the KFIVCI is written as
KFIV CIn+q =
[E(�1,n+q∣In, Dn) E(�2,n+q∣In, Dn) E(�3,n+q∣In, Dn)
] [�′1xn+q �′2xn+q �′3xn+q
].
(3.19)
However, we may be presented with a case where we want to use all information available at
time n to revise our previous estimates of the index. In this case we apply the Kalman Smoother
to obtain the K alman Smoother Instrumental Variable Coincident Index (KSIV CI), namely,
KSIV CIt =
[E(�1t∣In, Dn) E(�2t∣In, Dn) E(�3t∣In, Dn)
] [�′1xt �′2xt �′3xt
]. (3.20)
The probabilities associated to each of these indexes are straightforwardly computed using
Equation (3.11). The only modification we applied here is the entering of the cyclical components
of Equation (3.11) and their respective coefficients into the model. The outputs of this procedure
are the K alman Filter Instrumental Variables Probabilities (KFIVP) and the K alman Smoother
Instrumental Variables Probabilities (KSIVP). Additionally we use the same ideas and revisions
for creating the probabilities as those used in the index construction.
Another important output as a result of this methodology is a tool to analyze how the NBER
Business Cycles Dating Committee members make their decision through time. We extract at
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 53
each time t the weight of each coincident variable. For example, the K alman Filter Instrumental
Variables Weights, KFIVW , which are as follows:
KFIVWit =
[E(�1t∣It, Dt) E(�2t∣It, Dt) E(�3t∣It, Dt)
]⎡⎢⎢⎢⎢⎢⎢⎣�1i
�2i
�3i
⎤⎥⎥⎥⎥⎥⎥⎦
[E(�1t∣It, Dt) E(�2t∣It, Dt) E(�3t∣It, Dt)
]⎡⎢⎢⎢⎢⎢⎢⎣�11 �12 �13 �14
�21 �22 �23 �24
�31 �32 �33 �34
⎤⎥⎥⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
1
1
1
1
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦
.
(3.21)
The Kalman Smoother Instrumental Variables Weights, KSIVW , is built using the same method,
but using the smoothed coefficients
KSIVWit =
[E(�1t∣In) E(�2t∣In) E(�3t∣In)
]⎡⎢⎢⎢⎢⎢⎢⎣�1i
�2i
�3i
⎤⎥⎥⎥⎥⎥⎥⎦
[E(�1t∣It) E(�2t∣It) E(�3t∣It)
]⎡⎢⎢⎢⎢⎢⎢⎣�11 �12 �13 �14
�21 �22 �23 �24
�31 �32 �33 �34
⎤⎥⎥⎥⎥⎥⎥⎦
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
1
1
1
1
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦
. (3.22)
Note that at every time t the weights are set to sum one. Therefore, we can check whether there
is difference in the importance of each coincident variable in the NBER decision at a given time t.
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 54
As it will be explored in the following section, this exercise is in which interesting in under-
standing the way the NBER Committee made their decision regarding the beginning and the end
of some recessions.
3.3 Results
3.3.1 The Data
The data used in this work are monthly observations from 1960:01 to 2008:07. The four coincident
variables, “Employment” (Et), “Industrial Production” (Yt), “Income” (It) and “Sales” (St)), and
the leading series are defined in Table 3.3.1. The growth rates of the coincident series are plotted
in Figure 3.1 (the shaded areas show the recessions as claimed by NBER). Figure 3.1 shows that
the employment and industrial production show the strongest cyclical behavior.
3.3.2 The Basis Cycles
The basis cycles were found using the same procedure used by Issler and Vahid (2006). In this
section we will be describing the results found. Conditional on a VAR(4), we calculated the
canonical correlations between the coincident series (Δ lnEt, Δ lnYt, Δ ln It, Δ lnSt) and the
respective conditioning set, which comprised of four lags (Δ lnEt, Δ lnYt, Δ ln It, Δ lnSt) and
four lags of the leading series. All the leading series were transformed by taking the log difference,
however the two interest rate series were transformed by taking the first difference.
The canonical-correlation tests results in Table 3.3.2 indicate that only the last of the four
basis cycles obtained is not correlated to the last linear combination obtained from the leading
variables. Only the null hypothesis which stated that the 4th correlation coefficient associated
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 55
Time
Em
ploy
ees
1960 1970 1980 1990 2000 2010
−0.
03−
0.01
0.01
0.03
Time
Indu
stria
l Pro
duct
ion
1960 1970 1980 1990 2000 2010
−0.
03−
0.01
0.01
0.03
Time
Inco
me
1960 1970 1980 1990 2000 2010
−0.
010
0.00
00.
010
0.02
0
Time
Sal
es
1960 1970 1980 1990 2000 2010
−0.
03−
0.01
0.01
0.03
Figure 3.1: Coincident Series.
with the related basis cycle is rejected. All other hypotheses are not rejected.
The obtained business cycles are given as follows:
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 56
Table 3.1: Coincident and leading variables.
Series definition
Coincident series
Employess in non-agricultural payrolls
Industrial production
Personal income less transferences
Manufacturing trade and Sales
Leading series
MFG Unfilled Orders: Durable Goods, Ind. Total
Manufacturing & Trade Inventory: Total
NPOHUA by Building Permits in PIP
Industrial Production, Durable Consumer Goods
10-Year Treasury Constant Maturity Rate
Interest rate spread: 10-Year - 3 months Treasury Constant Maturity Rate
Nominal Weighted Exchange Rate
All Employees: Service-Providing Industries
Number Unemployed for Less than 5 Weeks
[c1t c2t c3t
]=
[Δ lnEi Δ lnYi Δ ln Ii Δ lnSi
]×
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
21.68 10.38 7.75
−0.56 −6.44 −4.32
0.66 −1.27 1.77
−1.14 3.64 −2.82
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦(3.23)
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 57
Table 3.2: Squared canonical correlations and canonical-correlation test.
Sq. Canonical correlations Degrees of freedom �2j and all smaller �2j = 0
�2j P-values (df corrected test)
0.5534 208 0.0000
0.2454 153 0.0000
0.2026 100 0.0000
0.0973 49 0.1511
From the coefficients of equation 3.23 we notice that the employment series has the greatest
weight in all three relevant cycles (more so in the first cycle). This reinforces our comprehension
that employment is the most cyclical variable among the coincident variables. We could use
the basis cycles (or a combination of the cycles) as a coincident index of economic activity. In
Figure 3.2 we can observe that the ups and downs of the first three cycles (mainly the first cycle)
have a similar pattern as the NBER official dating of recessions. So, as the first three cycles have
cyclical behavior, it is natural to seek a linear combination of the three that provides the “optimal”
prediction. As mentioned before, Issler and Vahid (2006) used a probit regression to extract all
information contained in the NBER Business Cycle Committee on the construction of the their
index. In their work they noticed that weights of each coincident variables were likely to vary over
time. So in the next section we present the results of our time varying probit model.
3.3.3 Estimates
The time varying coefficients estimated are presented in Figure 3.3. Table 3.3.3 shows that there are
variations in the coefficients of the basis cycles. For example, the coefficient of the first dependent
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 58
Time
c1
1960 1970 1980 1990 2000 2010
−0.
10.
2
Time
c2
1960 1970 1980 1990 2000 2010
−0.
150.
10
Time
c3
1960 1970 1980 1990 2000 2010
−0.
150.
10
Time
c4
1960 1970 1980 1990 2000 2010
−0.
10.
2
Figure 3.2: Coincident Cycles (growth rate) plot.
variable varies from 64.98 to 71.31, which yields a variation of 9.47%. Figure 3.3 maps the variations
that occur. It is easy to notice that most of the change happened durring a period of recession.
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 59
Some of these more prominent changes can be seen in the recessions during the 1970’s, 1980 and
2001.
Time
C1
slop
e
1960 1970 1980 1990 2000 2010
5565
7585
Time
C2
slop
e
1960 1970 1980 1990 2000 2010
−20
020
Time
C3
slop
e
1960 1970 1980 1990 2000 2010
3040
5060
Figure 3.3: Filtered and Smoothed weights.
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 60
Table 3.3: Descriptive statistics of the filtered (left) and smoothed (right) coefficients.
Statistic c1 c2 c3
Min 64.98 67.54 -13.37 -9.76 38.62 41.36
First Quant. 69.01 69.23 -7.05 -7.02 40.92 41.76
Median 70.08 70.34 -5.45 -5.23 42.36 42.11
Mean 69.88 69.90 -5.76 -5.96 42.60 42.65
Third Quant. -0.05 70.64 -3.53 -4.75 43.60 43.45
Maximum 71.31 70.97 -0.52 -4.22 48.42 45.49
We can compare the filtered and smoothed probabilities obtained with our model to the official
recessions periods proclaimed by NBER Business Cycle Committee. The results are shown in
Figure 3.4. We decided to use a probability cutoff point of 0.5 claim whether the economy is in
recession or not. As it is a probability, this is a intuitive classification rule. If we decided to choose a
optimal cutoff point, we would have to choose a loss function, what is itself arbitrary. For example,
Kamisnky and Reinhart (1998) and Ito and Yabu (2006) use the noise-to-signal ratio. Adopting a
procedure like that would generate different optimal cutoff points for each probabilities estimation
method, which is not desirable. In this Figure we observe that all three estimated probabilities are
likely to show values greater than 0.5 in periods officially dated as recessions. Initially, there was a
draw back in our method: we are more likely to claim false recessions, as our model presents more
indications of recession in periods when there is not a recession.
Using a cutoff point of 0.5 in the estimated probabilities to claim recession and expansion,
we can compare the adjustments of our model to that of Issler and Vahid (2006). Comparing
our smoothed probabilities to theirs’ (which is fair as they use all the data to estimate their
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 61
Filt
ered
IVC
I
1960 1970 1980 1990 2000 2010
0.0
0.4
0.8
Sm
ooth
ed IV
CI
1960 1970 1980 1990 2000 2010
0.0
0.4
0.8
Time
Issl
er
1960 1970 1980 1990 2000 2010
0.0
0.4
0.8
Figure 3.4: Predicted and Smoothed probabilities using data from 1960:06 to 2007:03.
probabilities) we see in Table 3.3.3 that our model has a better performance in dating recessions.
Actually, our model is able to correctly date 92.59% of the recessions, compared to 69.13% in Issler
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 62
and Vahid (2006) model. Our filtered probability, which uses only the information available at time
t to predict the state of economy at time t, has the same performance 92.59%. Better performance
during recessions has a drawback of a worse fit during expansions. Issler and Vahid (2006) model
correctly dates 97.98% of the recession periods compared to 84.31% of our smoothed probabilities
and 83.90% of our filtered probabilities.
Table 3.4: Accuracy of estimation based on a cut-off point of 0.5.
Estimators State of Economy
Overall Recessions Expansions
KFIVCI 85.12 92.59 83.90
KSIVCI 85.47 92.59 84.31
IVCI 93.94 69.13 97.98
There are a few important things to note about these difference. Figure 3.5 confirms the good
performance of our KFIVIC and KSIVIC. Both indexes are likely to show negative growth rates
at the dated recessions. And, more important, although the index cut the origin line in periods in
which there is not recession, the magnitude of the growth rate of the indexes in these situations
are not comparable to the magnitudes when there is really a recession. This may be a suggestion
that the indexes are more reliable than the probabilities.
Figure 3.6 shows how the coincident variables weights vary over time. As it was reported by
other studies, the employment series seems to be the most important one to determine the status
of the economy. Most of time, its smoothed weight varies around 0.55 but a careful look at its
behavior reveals that its coefficients are likely to increase at the recession periods archiving its
larger value just after the recession period in each recession period “neighborhood”. This behavior
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 63
Time
KF
IVC
I
1960 1970 1980 1990 2000 2010
−0.
015
0.00
0
Time
KS
IVC
I
1960 1970 1980 1990 2000 2010
−0.
010
0.00
5
Figure 3.5: Predicted and Smoothed probabilities using data from 1960:06 to 2007:03.
occurs at all recessions but the last in 2001. Actually, the 2001 recession is the one that the model
presents worse prediction.
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 64
Time
Em
ploy
men
t
1960 1980 2000
0.51
0.53
0.55
0.57
Time
Indu
stria
l Pro
duct
ion
wei
ghts
1960 1980 2000
0.24
80.
252
0.25
6
Time
Per
sona
l Inc
ome
wei
ghts
1960 1980 2000
0.22
890.
2291
0.22
93
Time
Sal
es w
eigh
ts
1960 1980 2000
−0.
06−
0.02
Figure 3.6: Filtered and Smoothed weights.
The opposite behavior occurs with the Industrial Production weights, which varies around a
value close to 0.25. It seems to decrease during recessions periods, except the last recession. The
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 65
same behavior is presented by sales, which has coefficients that vary around a value close to 0.1.
The sales weights vary mostly around 0.23 and personal income varies around −0.05, therefore,
suggesting that they are not significantly considered by the Business Cycle Committee. Table 3.3.3
shows some descriptive statistics from the time-varying coefficients.
Table 3.5: Descriptive statistics of the smoothed weights.
Statistic Employees Ind. Production Personal Income Sales
Min 0.5121 0.5317 0.2472 0.2504 0.2289 0.2291 -0.06529 -0.04366
First Quant. 0.5450 0.5457 0.2518 0.2522 0.2291 0.2292 -0.04710 -0.04049
Median 0.5534 0.5550 0.2530 0.2535 0.2292 0.2292 -0.03559 -0.03766
Mean 0.5518 0.5513 0.2528 0.2530 0.2291 0.2292 -0.03380 -0.03346
Third Quant. 0.5635 0.5575 0.2544 0.2538 0.2292 0.2292 -0.02582 -0.02713
Maximum 0.5794 0.5603 0.2566 0.2542 0.2293 0.2292 0.01177 -0.01117
The changes in the weights could explain why our model has a better performance during
recessions. It is exactly during these periods that the weights change. For example, there is a
group of economists that claim that the 2001 recession lasted longer than the period that was
claimed by the NBER Committee. Actually, at that time, the model suggested that the committee
decreased the weights in the three main coincident variables.
3.3.4 Predicting Recessions in Real Time
As a final exercise, we show how this model can be useful in predicting recessions. We consider
the last three recessions. So, we estimate the parameters of the model using information up to 1
year prior to each recession. As the cycles weights are random walks, their best prediction will be
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 66
given by their values at the end of the estimation period. Using this estimates and the “real time”
value of the coincident series we can estimate the probability of the NBER declaring a recession
in real time – at least with the minimal lag of 2 months for data disclosure.
Figure 3.7 presents the results for the recession during the 1990’s. The model is able to strongly
indicate the occurrence of recessions. However, it indicated the start of the recession before its
occurrence and estimated the recessions to last longer than it was declared by the NBER.
Time
pred
icted
prob
aliliti
es
1989.0 1989.5 1990.0 1990.5 1991.0 1991.5
0.00.2
0.40.6
0.81.0
Figure 3.7: 1990 Recession.
Figure 3.8 generally shows the same picture before. Some remarks should be made here.
The model is able to indicate the recession in a timely manner. The model does indicate that
the economy was in a recession during the entire recession period declared by the NBER. However,
it also show that the recession actually started one month prior to that declared by the NBER.
The model shows some early signs the recession was going to start before the NBER officially
CHAPTER 3. A STATE SPACE MODEL FOR INDICES OF ECONOMIC ACTIVITY 67
claimed the recession started. And, more important, is clearly suggests that the 2001 recession
should last longer than it lasted according to the NBER deliberations.
Time
pred
icted
prob
aliliti
es
2000.5 2001.0 2001.5 2002.0
0.20.4
0.60.8
1.0
Figure 3.8: 2001 Recession.
Finally, Figure 3.9 presents the results for the last recession. The model is able to indicate
the recession in a timely manner. The model does indicate that the economy was in a recession
during the entire recession period declared by the NBER. However, it also show that the recession
actually started one month prior to that declared by the NBER. For comparison proposes, the
Issler and Vahid (2006) probabilities computed using the entire sample period did not indicate any
recessions for all last recession periods. Table 3.3.4 presents the estimated probabilities.
Time
pred
icted
prob
aliliti
es
2007.0 2007.5 2008.0 2008.5
0.20.4
0.60.8
Figure 3.9: 2007 Recession.
3.4 Conclusion
The goal of this paper is to introduce a time varying structure on the IVCI and IVLI proposed
by Issler and Vahid since they suggest that found evidence that the weights of the coincident
variables are likely to vary over time. To do this we propose a state space probit model. We
also take advantage of the flexibility of the state space formulation to incorporate the time series
structure that exists on the residuals. With this model we targeted: the endogeneity of the model
and treated the dynamical structure of the residuals in an adequate manner.
With this structure we observed variations in the weights. For example, for the employment
series the weights vary from 0.96 to 0.52. We also had evidence that our index is able to predict the
recessions better. Our filtered probabilities predict 84.93% of the recessions, compared to 68.49%
obtained by Issler and Vahid.
Table 3.6: Predicted probabilities associated for period from 2007:08 to 2008:07
Year and month Predicted Probability of Recession
2007:08 0.2603
2007:09 0.3120
2007:10 0.1233
2007:11 0.5202
2007:12 0.8150
2008:01 0.7654
2008:02 0.9554
2008:03 0.8726
2008:04 0.5316
2008:05 0.6169
2008:06 0.5706
2008:07 0.6247
Bibliography
[1] Albuquerque, C., Portugal, M. (2005) Pass-Through from Exchange Rate to Prices
in Brazil: An Analysis using time-varying parameters for the 1980-2002 period.
Revista de Economıa, Montevideu - Uruguay, v. 12, n.1, p. 17-54.
[2] Akaike, H., 1976. Canonical correlation analysis of time series and the use of an
information criterion. In: Mehra, R.K., Lainiotis, D.G. (Eds.), System Identification:
Advances and Case Studies. Academic Press, New York, pp. 27 - 96.
[3] Belaisch, A. (2003) Exchange rate pass-through in Brazil, IMF Working Papers,
n.141, July.
[4] Betts, C., Devereux, M.B. (2000) Exchange Rate Dynamics in a Model of Pricing-
to-Market, Journal of International Economics 50(1), 215-244.
[5] Birchenhall, C.R., Jessen, D. R., Osborn, D. R., Simpson, P., 1999. Predicting U.S.
Business-Cycle Regimes. Journal of Business and Economic Statistics, Vol. 17, No.
3, 313-323.
[6] Brockwell, P. J. and Davis, R. A. (1991) Time Series: Theory and Methods. 2nd
edition. Springer-Verlag.
70
[7] Brockwell, P. J. and Davis, R. A. (2003) Introduction to Time Series and Forecasting.
2nd edition. Springer-Verlag.
[8] Burns, A. F., Mitchel, W. C., 1946. Measuring Business Cycles. National Bureau of
Economic Research, New York.
[9] Campa, J.M., Goldberg, L.S. (2005) Exchange Rate Pass-through into Import Prices:
A Macro or Micro Phenomenon?, Review of Economics and Statistics 87, 4, 679-690.
[10] Calvo, G., Reinhart, C. (2000) Fixing for your life, NBER Working Paper n.8006,
November.
[11] Carneiro, D., Monteiro, A., Wu, T.(2002) Mecanismos nao lineares de repasse cam-
bial para o IPCA, PUC-RJ, Working Paper n.462.
[12] Chauvet, M., 1998. An Economic characterization of business cycle dynamics with
factor structure and regime switching. International Economic Review 39, 969-996.
[13] Chauvet, M., Hamilton, J. D., 2006. Dating Business Cycle Turning Points in Real
Time. In Nonlinear Time Series Analysis of Business Cycles, edited by Costas Milas,
Philip Rothman, and Dick van Dijik. Vol. 276, Contributions to Economic Analysis.
Amsterdam: Elsevier Science and Technology.
[14] Chauvet, M., Piger, jeremy, 2008. A Comparison of the Real-Time Performance
of Business Cycle Dating Methods, Journal of Business and Economic Statistics,
American Statistician Association, 26, 42-29.
[15] Choudri, E., Hakura, D., (2006) Exchange Rate Pass-Through to Domestic Prices:
Does the Inflationary Environment Matter, Journal of International Money and
Finance, 25, 614-39.
[16] Corsetti, G., Dedola, L. (2002) Macroeconomics of International Price Discrimina-
tion, European Central Bank Working Paper No. 176.
[17] Devereux, Michael B., Yetman, James, (2002) Price Setting and Exchange Rate
Pass-Through: Theory and Evidence, HKIMR Working Paper n. 22.
[18] Dornbusch, R. (1987) Exchange Rate and Prices, The American Economic Review
77.
[19] Duarte, A. J. M. A., Issler, J. V., Spacov A. D., 2004. Indicadores coincidentes de
atividade economica e uma cronologia de recessoes para o Brasil, Ensaios Economi-
cos, no. 527.
[20] Durbin, J. and Koopman, S.J., (2000) Time Series Analysis by State Space Methods.
Oxford.
[21] Dueker, M., 2005. Dynamic Forecast of Qualitative Variables: A Qual VAR Model
of U.S. Recessions, Journal of Business and Economic Statistics, Vol. 23, 2005.
[22] Durbin, J., Koopman, S.J., 2001. Time Series Analysis by State Space Methods.
Oxford University Press. Oxford.
[23] Feenstra, R.C., Gagnon, J.E., Knetter, M.M. (1996) Market Share and Exchange
Rate Pass-through in World Automobile Trade, Journal of International Economics
40(1/2) 187-207.
[24] Fraga, A., Goldfajn, I., Minella, A. (2003) Inflation Targeting in Emerging Market
Economies, NBER Working Paper 10.019.
[25] Gagnon, J. E., Ihrig, J. E. (2001) Monetary Policy and Exchanger Rate Pass-
Through, Federal Reserve Bank International Finance Discussion Paper, No. 704.
[26] Goldberg, L. (2004) Industry Specific Exchange Rates for the United States, Federal
Reserve Bank of New York Economic Policy Review 10(1), 1-16.
[27] Goldberg, P.K., Knetter, M.M. (1997) Goods Prices and Exchange Rates: What
Have We Learned?, Journal of Economic Literature 35(3), 1243-1272.
[28] Goldfajn, I., Valdes, R. (1999) The Aftermath of Appreciations, PUC-RIO, Working
Paper No. 396.
[29] Goldfajn, I., Werlang, S.R.C. (2000) The Pass-through from Depreciation to Infla-
tion: A Panel Study, Banco Central do Brasil Working Paper No. 5.
[30] Gron, A., Swenson, D.L. (1996) Incomplete Exchange-Rate Pass-through and Im-
perfect Competition: The Effect of Local Production, American Economic Review
Papers and Proceedings 86(2), 71-76.
[31] Hamilton, J. D. (1994) Time Series Analysis. Princeton University Press.
[32] Hamilton, J. D., 1989. A New Approach to the Economic Analysis of Nonstationary
Time Series and the Business Cycle. Econometrica 57, 357-384.
[33] Hamilton, J. D., 2003. What is an Oil Shock? Journal of Econometrics, 113, pp.
363-398.
[34] Harding, D., Pagan, A., 2003. A comparison of Two Business Cycles Dating Meth-
ods, Journal of Econometric Dynamics and Control 27, 1681-1690.
[35] Harvey, A. C. (1989) Forecasting, Structural Time Series Models and The Kalman
Filter. Cambridge University Press.
[36] Harvey, A. C. (1993) Time Series Models. 2nd edition. Harverster Wheatsheaf. Press.
[37] Hotelling, H., 1935. The most predictable criterion. Journal of Educational Psychol-
ogy 26, 139-142.
[38] Hotelling, H., 1936. Relations between two sets of variates. Biometrika 28, 321-377.
[39] Issler, J. V., Vahid, F., 2006. The Missing Link: Using the NBER Recession Indica-
tor to Construct Coincident and Leading Indices of Economic Activity, Journal of
Econometrics, Vol. 91, no. 1, pp. 281-303.
[40] Johnston, R. A., Wichern, D. W., 2001. Aplied Multivariate Statistical Analysis.
Pearson Education, New York. Fith edition.
[41] Kalman, R. E., 1960. A new approach to linear filtering and prediciton problems,
Journal of Basic Engeneering, Transactions ASME. Series D 82: 35-45.
[42] Kim, Chang-Jin, Nelson, Charles R., 1998. State-Space Models with Regime Switch-
ing: Classical and Gibbs-sampling Approaches with Applications, MIT Press.
[43] Kim, Chang-Jin, 2006. Time-varying parameter models with endogenous regressors,
Economics Letters, 91, 21-26.
[44] Klein, B. M., 2003. State Space Models for Exponential Family Data, PhD Thesis.
Department of Statistics, University of Southern Denmark.
[45] Krugman, P.R. (1987), “Princing to market when the exchange rate changes.
In: Arndt, S.W., Richardson, J.D. (Eds.), Real-Financial Linkages among Open
Economies. MIT Press, Cambridge.
[46] McCarty, J. (2000) Pass-through of Exchange Rate and Import Prices to Domestic
Inflation in Some Industrialized Economies, BIS Working Paper No. 79.
[47] Menon, J. (1996) Exchange Rate Pass-through, Journal of Economic Surveys, 9 (2),
197-231.
[48] Olivei, G.P. (2002) Exchange Rates and the Prices of Manufacturing Products Im-
ported into the United States, New England Economic Review, First Quarter, 3-18.
[49] Pagan, A. 1980. “Some Identification and Estimation Results for Regression Models
with Stocastically Varying Coefficients.” Journal of Econometrics 13:341-63.
[50] Parsley, D. (1995) Anticipated future shocks and exchange rate pass-through in the
presence of reputation, International Review of Economics 4(2).
[51] Pizzinga, A. H. and Fernandes, C., Contreras, S. (2008), Restricted Kalman filtering
revisited. Journal of Econometrics, v. 144, p. 428-429.
[52] Pizzinga, A. (2009), Futher Investigation into Restricted Kalman Filtering. Statistics
& Probability Letters, v. 79, p. 264-269.
[53] Pollard, P., Coughlin, C. (2005) Pass-through Estimates and the Choice of an Ex-
change Rate Index, St. Louis FED Working Paper.
[54] Rivers, D., Vuong, Q. H., 1988. Limited Information Estimators and Exogeneity
Tests for Simultaneous Probit Models, Journal of Econometrics, 39, 347-366.
[55] Sekine, T. (2006) Time-Varying Exchange Rate Pass-Through: Experience of Some
Industrial Countries, BIS Working Paper.
[56] Stock, J., Watson, M., 1988a. A New Approach to Leading Economic Indicators,
mimeo, Harvard University, Kennedy School of Government.
[57] Stock, J., Watson, M., 1988b. A Probability Model of the Coincident Economic
Indicator, NBER Working Paper # 2772.
[58] Stock, J., Watson, M., 1989. New Indexes of Leading and Coincident Economic
Indicators, NBER Macroeconomics Annual, pp. 351-95.
[59] Stock, J., Watson, M., 1991. A Probability Model of the Coincident Economic Indica-
tors, in “Leading Economic Indicators: New Approaches and Forecasting Records”,
K. Lahiri and G. Moore, Eds. New York: Cambridge University Press.
[60] Stock, J., Watson, M., 1993a. A Procedure for Predicting Recessions with Leading
Indicators: Econometric Issues and Recent Experiences. In Stock, J. H., Watson, M.
W. (Eds.), New Research on Business Cycles, Indicators and Forecasting. University
of Chicago Press, Chicago.
[61] Stock, J., Watson, M., 1993b. New Research on Business Cycles, Indicators and
Forecasting. University of Chicago Press, Chicago.
[62] Taylor, J. (2000) Low Inflation, Pass-through and the Pricing Power of Firms, Eu-
ropean Economic Review 44, 1389-1408.
[63] Yang, J. (1996) Exchange Rate Pass-through in U.S. Manufacturing Industries, Re-
view of Economics and Statistics 79(1), 95-104.