Академический Документы
Профессиональный Документы
Культура Документы
, Dimitrios D. Thomakos
b,1
a
School of Accounting, Economics and Finance, Deakin University, 221 Burwood Highway, Burwood, Melbourne, Victoria 3125, Australia
b
Department of Economics, University of Peloponnese, Tripolis Campus 22100, Greece
Abstract
We show that incorporating the effects of exchange rate pass-through into a model can help in obtaining superior forecasts of
domestic, industry-level inflation. Our analysis is based on a multivariate system of domestic inflation, import prices and
exchange rates that incorporates restrictions from economic theory. These are restrictions on the transmission channels of the
exchange rate pass-through to domestic prices, and are presented as testable hypotheses that lead to model reduction. We
provide the results of various tests, including causality and prior restrictions, which support the underlying economic arguments
and the model we use. The forecasting results for our model suggest that it has a superior performance overall, jointly producing
more accurate forecasts of domestic inflation.
2007 International Institute of Forecasters. Published by Elsevier B.V. All rights reserved.
Keywords: Causality; Direction of change; Econometric model; Exchange rates; Forecasting; Inflation; Model reduction; Pass-through; VAR
model; VARMA model
1. Introduction
We examine whether industry-level inflation fore-
casts of domestic consumer and producer prices can
be improved when we incorporate additional infor-
mation about potential exchange rate pass-through.
Our analysis is performed in a multivariate setting,
and we compare standard forecasting models, such
as linear multivariate ARMA and VARMA models,
with a structural system that incorporates informa-
tion about the transmission of exchange rate effects
to domestic prices. We cover several different indus-
tries in three major economies, Japan, the United
Kingdom and the United States.
The exchange rate pass-through effect is de-
fined as the percent change in local currency
import prices resulting from a 1% change in the
exchange rate between the exporting and importing
countries. Theoretically, this effect can be either full
Available online at www.sciencedirect.com
International Journal of Forecasting 24 (2008) 134150
www.elsevier.com/locate/ijforecast
m
i
(E
i
) 0 charged by the foreign exporters that
depends on the prevailing exchange rate E
i
. An exo-
genous change in the exchange rate leads to a change
in import prices given by:
dlogP
ri
dlogE
i
1
dlog 1 m
i
_ _
dlogE
i
; 1
where we assume that such a change has no effect on
the price charged by the foreign exporters, i.e.,
dlogP
i
/dlogE
i
=0. Depending on whether m
i
changes
or not, a change in the exchange rate may or may not
be reflected in P
ri
: if there is no change in the markup
then there is full pass-through to import prices; if
there is a reduction in the markup (e.g., to maintain
the exporters' competitiveness) then there can be no
change in the domestic import prices. As pointed out
earlier, this is the first channel of short-run exchange
rate pass-through. Here foreign exporters can dis-
criminate among destination markets by price.
Once in the home market, retailers have to bear the
transport and distribution costs of final goods before
selling them to the consumers. Denoting these costs for
the ith good as
i
i
(P
ri
) 0, the change in the total
marginal cost the retailers face from a change in the
exchange rate is given by:
dlogMC
ri
dlogE
i
dlogP
ri
dlogE
i
dlog 1 X
i
dlogE
i
; 2
and again, the effect on the marginal cost for the
retailer depends on the response of import prices, and
the corresponding response of distribution costs, to the
exchange rate change. Now, letting the domestic
markup charged by the retailer to final consumers be
i
(MC
ri
) 0, we find that the effect on the final
price of the ith good is given by:
dlogP
ci
dlogE
i
dlogMC
ri
dlogE
i
dlog 1 g
i
dlogE
i
; 3
or by combining Eqs. (1), (2) and (3):
dlogP
ci
dlogE
i
1
dlog 1 m
i
_ _
dlogE
i
dlog 1 X
i
dlogE
i
dlog 1 g
i
dlogE
i
: 4
Eq. (4) shows the link between the exchange rate
E
i
, domestic prices for retailers of imported goods P
ri
,
and domestic prices for consumers of imported goods
P
ci
. An import price increase driven by exchange rate
depreciation may or may not lead to a proportional
increase in consumer prices, depending on the changes
in the corresponding mark-ups. This is the second
channel of pass-through: the response of domestic
prices to changes in the prices of imports can be
explained by changes in the existing domestic
markups,
i
and
i
, in the way we described
previously. The foreign and domestic markup adjust-
ments may dampen the transmission of the changes in
the exchange rate to import prices and domestic prices.
These adjustments can be a reflection of sluggish
137 P.S. Bhattacharya, D.D. Thomakos / International Journal of Forecasting 24 (2008) 134150
nominal wage adjustments in the foreign country as
well as in the home country.
The above approach identifies two different channels
of exchange rate pass-through. One channel is at the
point of entry, and is captured by fluctuations in import
prices at the point of entry. The other channel is one
whichtransmits changes in the exchange rate to domestic
prices (which can be either consumer or producer prices)
through markup adjustments by domestic and foreign
agents. To capture these effects empirically, we need
properly matched industry-level import prices, consumer
prices and producer prices. In addition, this conceptual
framework reflects the assumptions of block causality
and triangularity, as discussed before.
Finally, it is important to note a practical implica-
tion of the above formulation. If these assumptions
about the transmission of an exchange rate change are
correct, then one can trace the impact of such a change
sequentially through import prices to domestic prices.
This is not only testable as a hypothesis (the triangu-
larity mentioned above), it is also useful in making
forecasting scenarios: if one has a properly specified
model, then in principle, it can produce forecasts for
any desired path of the exchange rates.
3.2. Specification, reduction and testing of the
structural system
We convert the conceptual framework of the pre-
vious section into a dynamic econometric specification
with testable restrictions and implications. Following
the previous subsection, and because we are interested
in producing inflation forecasts later on, our analysis is
performed on log-changes of prices. Let us therefore
denote CPI inflation by y
t
1
, PPI inflation by y
t
2
, the
percentage change in import prices by y
t
3
, and the
percentage change in the exchange rate by x
t
, all of
which are expressed as deviations from their sample
means. Consider these variables as being the compo-
nents of the vector z
t
=(y
t
T
, x
t
)
T
, where y
t
(y
t
1
, y
t
2
, y
t
3
)
T
denotes the vector of price changes, and incorporate
them into an unrestricted VAR model of order p as:
z
t
p
j1
F
j
z
tj
u
t
; 5
where u
t
=(
t
T
,
t
)
T
is assumed to be a multivariate
white noise with covariance matrix . The coefficient
matrices
j
and the error covariance matrix can be
expressed in partitioned form as:
F
j
F
yyj
j
xyj
j
T
yxj
u
xxj
_ _
;
S
e
s
eg
s
T
eg
r
g
_ _
: 6
This allows us to easily express several hypotheses and
sub-models that we will consider. The hypothesis of no
block causality from prices to the exchange rate is
given by:
Hypothesis set # 1. H
0
BC
:
yxj
=0 for all j =1, 2, , p,
for the absence of block causality.
Note that the above test, as well as the other
hypothesis tests that follow, are Wald-type tests.
If the hypothesis of no block causality is correct,
then the generic VAR model of Eq. (5) is reduced to a
VARX (VAR with exogenous variables) model that
can be written as:
y
t
p
j1
F
yyj
y
tj
j
xyj
x
tdj
_ _
e
t
; 7
where we allow for the possibility of a delayed
response of d periods for prices from a change in the
exchange rates (note that this is compatible with the
idea of sluggish adjustments mentioned before). We
can, as in Eq. (6), partition the relevant coefficient
matrices as:
F
yyj
11j
12j
13j
21j
22j
23j
31j
32j
33j
_
_
_
_
; j
xyj
xy1j
xy2j
xy3j
_
_
_
_
:
8
The VARX model of Eq. (7) can now incorporate
the two-dimensional approach of Obstfeld (2002), the
dual transmission of exchange rate changes through
import and domestic prices. The model can be further
refined by imposing and testing additional hypotheses.
For example, note that triangularity is given as a set of
restrictions on the dynamics of the model, and that it
requires the absence of short-term feedback from con-
sumer and producer prices to import prices, and si-
milarly from consumer prices to producer prices.
138 P.S. Bhattacharya, D.D. Thomakos / International Journal of Forecasting 24 (2008) 134150
Therefore, triangularity can be expressed as a set of
restrictions as:
Hypothesis set # 2. H
0
TR
: /
klj
=0 for all kNl and
every j =1, 2, , p.
In addition, within the context of the VARX model
and the triangularity restrictions, we can test for the
validity of the unidirectional feedback, i.e., we can per-
form causality-type tests within the triangular structure
as:
Hypothesis set # 3. H
0
TRC,1
: /
12j
=/
13j
=0, H
0
TRC,2
:
/
23j
=0, H
0
TRC,3
: /
xy3j
=0 for all j =1, 2, , p.
The first hypothesis above tests whether domestic
consumer prices are indeed affected by producer and
import prices; the second hypothesis tests whether pro-
ducer prices are affected by import prices, given that
the triangularity assumption is satisfied; and the third
hypothesis tests whether import prices are affected by
the exchange rate, again given that the triangularity
assumption is satisfied.
The sequence of steps that we take in specifying
and forecasting using our structural system can now
be summarized as follows:
Note that we do not compute forecasts from the
estimates of the model in step (b). This step is performed
to generate the additional causality assumptions.
We compute forecasts from the unrestricted model of
Eq. (5) in step (a), since the no causality assumptions that
lead to the structural system need not be incorrect. As
we explain in the next section, we also use a more
parsimonious VARMAmodel for computing unrestricted
forecasts. However, we do test the triangularity assump-
tion, as it rests on prior economic motivations, and always
generate forecasts from the VARX model, as in step (c).
7
In the discussion of our results, as well as in the
tables, we use the notationSYSto denote the structural
system as described in this section, and the notation
VAR1 to denote the unrestricted VAR model of Eq. (5).
3.3. Competing forecasting models and forecast
evaluation
We compare the forecasts we generate from either
the SYS or VAR1 models with forecasts that are
computed from a different VAR model and from a
VARMA model. We denote these two models VAR2
and VARMA, respectively. The second VAR model
(VAR2) is an unrestricted VAR that excludes import
prices and the exchange rate, but includes consumer
prices and producer prices, and is further augmented
using interest rates and the unemployment rate. Letting
w
t
=(y
t1
, y
t2
, r
t
, u
t
)
T
denote the vector of variables
entering the second VAR model, we have a similar
representation to that in Eq. (5), i.e.:
w
t
k
j1
P
j
w
tj
v
t
; 9
where v
t
is assumed to be multivariate white noise. Both
unrestricted VAR models of Eqs. (5) and (9) are esti-
mated using nine lags.
8
7
An issue raised by the editor was the treatment of the exogenous
variable, the exchange rate, when generating our inflation forecasts
from the VARX model. Clearly, using the true, out-of-sample values
for the exchange rate would not be legitimate, and another
forecasting model would be needed to generate forecasts of the
exchange rate as well. However, in all of our specifications, the
exchange rate was entered with a lag of order d3, and therefore
we could use, in a proper way, the past values of the exchange rate
when we generated inflation forecasts.
8
Nine lags are chosen, keeping in mind that the existing literature
on the exchange rate pass-through effect (see, for example,
Goldberg & Knetter, 1997; and Campa & Goldberg, 2005) points
out that the long-run timeframe starts with the third quarter, where
the pass-through to local currency import prices seems to be
complete. In that case, there will be more inflationary pressure on
domestic industrial prices coming from the exchange rate channel.
(a) Estimate the unrestricted VAR model of Eq. (5),
perform and report the tests in Hypothesis set
#1, and generate the corresponding forecasts.
(b)Impose the block causality hypothesis for the
exchange rate, and reduce the model to the
VARX of Eq. (7). Perform and report the tests in
Hypothesis set #3, under the triangularity
assumption.
(c) Estimate the VARX model of Eq. (7) and
perform the tests in Hypothesis set #2:
1. If the triangularity assumption is not re-
jected, impose it and re-estimate the model
by seemingly unrelated regression (SUR)
then generate forecasts from the restricted
VARX model.
2. If the triangularity assumption is rejected,
then proceed with the unrestricted VARX
model in generating forecasts.
139 P.S. Bhattacharya, D.D. Thomakos / International Journal of Forecasting 24 (2008) 134150
In addition to these unrestricted VARmodels, we also
employ a more parsimonious VARMA(1,1) model.
9
The class of VARMA models is considerably more
flexible in modeling and forecasting linear multiva-
riate time series, but is also considerably more time-
consuming to estimate. However, a VARMA model
seems to be a very appropriate benchmark for fore-
casting comparisons in this context. The model is given
by:
z
t
A
1
z
t1
x
t
B
1
x
t1
; 10
where
t
is again assumed to be a multivariate white
noise. Since our forecasting evaluation requires multiple
nonlinear estimations for this model, we use the fast,
regression-based method of Spliid (1983) for estimating
the parameter matrices and generatingthe corresponding
forecasts.
Our forecasts are for horizons of h=1, 2, 3 months
and are computed using a rolling window of T
0
observations; they are initially evaluated using their
root mean-squared and mean absolute errors. Denoting
the h step-ahead forecast for the ith price variable (CPI
inflation y
t1
or PPI inflation y
t2
) from the mth
forecasting model by
t+h|t,i
, we compare:
R
i
m
1
T
1
T
tT
0
1
y
thjt;i
m y
th;i
_ _
2
;
_
A
i
m
1
T
1
T
tT
0
1
j y
thjt;i
m y
th;i
j;
11
for the root mean-squared and mean absolute errors
respectively, where T
0
+T
1
=T, with T
0
being the
training sample and T
1
being the evaluation sample.
We then use the Diebold and Mariano (1995) test to
more formally compare the forecasting performances
of the competing models, based on the difference of
their forecasting errors. We employ both the squared
loss function and the absolute value loss function in
computing the DieboldMariano test statistics. Our
discussion of the results is based on the root mean-
squared statistic and the DieboldMariano test using
the squared loss function. The results based on the
mean absolute error statistic and the corresponding
DieboldMariano tests are qualitatively similar and
are available on request.
4. Data
An appropriate empirical analysis based on the above
discussion can be performed using suitably disaggre-
gated data. While we do not have access to industry-
level domestic prices of the imported goods, the data we
have include (1) consumer prices (CPI) (which reflect
the prices of final products or final consumer goods), (2)
import prices (IMP) for final goods, intermediate goods
and crude materials, and (3) prices for domestic pro-
ducers (PPI), again for final goods, intermediate goods
and crude materials. With these data, we should be able
to distinguish in terms of end uses, and this allows us to
gain a better understanding of the channels of pass-
through under investigation.
Our choice of data is guided by two considerations.
First, we need to have industry-level price data for
imported goods, as well as matching industry-level
price data for intermediate goods (if available, as in
system II) and matching price data for final goods (can
be for producers goods, as in system III, or for final
consumers goods, as pointed out in system I). Second,
we concentrate on three of the largest active global
economies, Japan, the United Kingdom and the United
States, which are involved in heavy trading across three
different regions. It is important to note that these
countries do have the required data. We have access to
perfectly matching industry-level price data for four-
teen US industries, thirteen UK industries and seven
Japanese industries (and not for all SITC industries).
The monthly data frequency we use is guided by the
question we try to answer in this study: whether we can
improve the forecast performance in the presence of
short-run pass through at the SITC level of industrial
manufacturing goods data. Additional details about the
raw data can be obtained from the authors on request.
For Japan, we have data for import prices, producer
prices, wholesale prices (WPI), trade-weighted ex-
change rates, and short-term interest rates from the
Bank of Japan's website. The seven industries we use
9
We thank the referee for suggesting benchmarking on a
VARMA(1,1) model and a VAR model with endogenous exchange
rates. We have cross-checked the forecast results from the VARMA
(1,1) model with models of orders (1, 2), (2, 1) and (2, 2), and VAR
models with lags 6, 10, 12, and 15. The findings remain
qualitatively similar, and are therefore not reported.
140 P.S. Bhattacharya, D.D. Thomakos / International Journal of Forecasting 24 (2008) 134150
Table 1
Tests for causality: United Kingdom (PPI-based models)
Industry/Lags Block causality for exchange rate Causality in the triangular structure
(ppi, imp)wfx wfx (ppi, imp) Block causality from imp to ppi Block causality from wfx to imp
ppi imp impppi impwfx wfximp
Tobacco
3 0.579 0.003 0.227 0.067 0.495 0.002
6 0.262 0.002 0.004 0.085 0.365 0.011
9 0.184 0.000 0.215 0.290 0.538 0.001
12 0.026 0.000 0.233 0.018 0.552 0.000
Pulp
3 0.858 0.534 0.000 0.494 0.725 0.664
6 0.532 0.329 0.000 0.482 0.520 0.236
9 0.390 0.426 0.000 0.493 0.266 0.392
12 0.043 0.086 0.000 0.208 0.105 0.128
Wood
3 0.546 0.032 0.000 0.146 0.275 0.048
6 0.615 0.012 0.003 0.813 0.225 0.024
9 0.082 0.053 0.006 0.437 0.033 0.049
12 0.032 0.042 0.020 0.329 0.043 0.050
Metal
3 0.815 0.137 0.002 0.361 0.887 0.790
6 0.952 0.473 0.004 0.278 0.961 0.768
9 0.288 0.338 0.016 0.172 0.576 0.791
12 0.235 0.182 0.014 0.270 0.397 0.074
Chemical
3 0.876 0.000 0.003 0.005 0.661 0.000
6 0.972 0.000 0.001 0.016 0.883 0.001
9 0.902 0.000 0.012 0.040 0.842 0.004
12 0.835 0.000 0.043 0.006 0.732 0.013
Org-chem
3 0.285 0.014 0.000 0.300 0.201 0.001
6 0.277 0.001 0.000 0.362 0.365 0.002
9 0.152 0.000 0.004 0.568 0.168 0.004
12 0.037 0.002 0.012 0.362 0.069 0.007
Medi.
3 0.440 0.002 0.521 0.334 0.385 0.731
6 0.301 0.008 0.387 0.740 0.716 0.433
9 0.033 0.000 0.391 0.478 0.408 0.232
12 0.053 0.010 0.048 0.590 0.336 0.360
Plastics
3 0.967 0.019 0.000 0.726 0.898 0.120
6 0.943 0.028 0.001 0.279 0.959 0.023
9 0.917 0.040 0.033 0.303 0.855 0.037
12 0.479 0.011 0.001 0.049 0.431 0.050
Tex-fabs
3 0.421 0.014 0.014 0.337 0.458 0.009
6 0.866 0.037 0.032 0.188 0.784 0.085
9 0.824 0.207 0.072 0.061 0.659 0.293
12 0.752 0.227 0.044 0.052 0.343 0.526
Iron
3 0.536 0.345 0.001 0.237 0.936 0.991
6 0.630 0.102 0.009 0.216 0.812 0.986
9 0.209 0.088 0.056 0.064 0.865 0.960
12 0.276 0.014 0.254 0.207 0.898 0.600
(continued on next page)
141 P.S. Bhattacharya, D.D. Thomakos / International Journal of Forecasting 24 (2008) 134150
are chemicals; foodstuffs and feedstuff; machinery and
equipment; metals and related products; petroleum,
coal and natural gas; textiles and wood; and lumber and
related products. The sample period is from 01/1971 to
12/2002, which contains T=384 observations. The
trade-weighted exchange rate is defined as the Japanese
Yen vis--vis weighted index of Japan's major trading
partners' currencies. The civilian unemployment rate
series, obtained from the International Labor Organi-
zation's (ILO) website, starts from 01/1976 and ends at
12/2002, reducing the sample to T=324 observations.
For the United Kingdom, we have data on import
prices, producer prices and the civilian unemployment
rate from National Statistics Online. The thirteen indus-
tries we use are tobacco; pulp and waste paper; wood
and cork; metal ores; chemicals; organic chemicals;
medicinal products; plastics; textile fabrics; iron and
steel; non-ferrous metals; machinery and transport
equipment; and electrical machinery. The trade-weight-
ed exchange rate and average discount rates for Treasury
bills are downloaded from the Bank of England's
website. For all UK data series, the sample period is
from 01/1991 to 12/2002, giving a total of T=144
observations.
For the United States, we have data on import prices,
producer prices and corresponding consumer prices. In
addition, we again use a trade-weighted exchange rate,
defined as the US dollar/weighted average of 27
countries' currencies, plus the overall civilian unem-
ployment rate, the federal funds rate and the 3-month
Treasury bill rate. All data series are from the Bureau of
Labor Statistics and the Federal Reserve Bank of Saint
Louis online databases. The fourteen industries we use
are meat and meat preparations; vegetables, fruit and
nuts, fresh or dried; beverages and tobacco; beverages;
mineral fuels; lubricants and related materials; chemi-
cals and related products, n.e.s. (not essentially
specified); organic chemicals; inorganic chemicals;
rubber manufactures, n.e.s.; non-metallic mineral man-
ufactures; metalworking machinery; electrical machin-
ery and equipment; furniture and parts thereof; and
articles of apparel and clothing accessories. The sample
period varies across industries and data series, with the
earliest monthly series starting at 01/1989. The number
of available observations thus varies between T=108
and T=168 observations by the series and model used.
For our forecast evaluations, we set the evaluation
sample to T=48 months for thirty industries across all
Table 1 (continued)
Industry/Lags Block causality for exchange rate Causality in the triangular structure
(ppi, imp)wfx wfx (ppi, imp) Block causality from imp to ppi Block causality from wfx to imp
ppi imp impppi impwfx wfximp
Nonfme
3 0.457 0.631 0.000 0.398 0.455 0.973
6 0.487 0.393 0.001 0.662 0.442 0.867
9 0.051 0.010 0.000 0.314 0.399 0.306
12 0.006 0.010 0.008 0.205 0.127 0.361
Mach
3 0.457 0.000 0.463 0.224 0.231 0.000
6 0.373 0.000 0.895 0.248 0.119 0.000
9 0.517 0.000 0.672 0.093 0.190 0.000
12 0.222 0.001 0.916 0.461 0.147 0.000
El-mach
3 0.575 0.125 0.003 0.426 0.731 0.049
6 0.555 0.007 0.001 0.013 0.412 0.011
9 0.066 0.009 0.004 0.018 0.194 0.006
12 0.001 0.005 0.004 0.013 0.086 0.028
xy denotes x does not cause y; the numbers in the cells are p-values of the corresponding test; wpi stands for the wholesale price index;
wfx stands for the trade-weighted exchange rate; imp stands for the import price; and ppi denotes the producer price. In the table, Pulp
denotes pulp and waste paper, Wood denotes wood and cork, Metal stands for metal ores, Chemical stands for chemicals, Org-chem
denotes organic chemicals, Medi. denotes medicinal products, Tex-fabs denotes textile fabrics, Iron stands for iron and steel, Nonfme
stands for non-ferrous metals, Mach denotes machinery and transport equipment, and El-mach denotes electrical machinery.
142 P.S. Bhattacharya, D.D. Thomakos / International Journal of Forecasting 24 (2008) 134150
Table 2
Tests for causality: United States (PPI-based models)
Industry/Lags Block causality for exchange rate Causality in the triangular structure
(ppi, imp)wfx wfx(ppi, imp) Block causality from imp to ppi Block causality from wfx to imp
ppi imp impppi impwfx wfximp
Meat
3 0.929 0.914 0.032 0.690 0.845 0.894
6 0.929 0.908 0.167 0.834 0.851 0.934
9 0.751 0.793 0.036 0.962 0.798 0.967
12 0.656 0.474 0.026 0.312 0.700 0.989
Fruit
3 0.611 0.294 0.010 0.068 0.296 0.247
6 0.352 0.760 0.074 0.387 0.111 0.743
9 0.354 0.718 0.113 0.140 0.213 0.658
12 0.516 0.787 0.170 0.163 0.352 0.687
Beve
3 0.019 0.745 0.833 0.902 0.064 0.933
6 0.070 0.208 0.178 0.923 0.104 0.863
9 0.013 0.075 0.521 0.209 0.061 0.959
12 0.000 0.015 0.880 0.472 0.006 0.873
Chemical
3 0.204 0.000 0.000 0.807 0.238 0.000
6 0.221 0.008 0.004 0.348 0.391 0.001
9 0.288 0.033 0.003 0.058 0.398 0.013
12 0.141 0.004 0.000 0.012 0.260 0.015
Org-chem
3 0.047 0.196 0.259 0.047 0.035 0.081
6 0.023 0.104 0.216 0.122 0.060 0.038
9 0.059 0.024 0.415 0.070 0.150 0.054
12 0.061 0.001 0.207 0.010 0.148 0.029
Inor-chem
3 0.038 0.073 0.418 0.123 0.012 0.044
6 0.015 0.471 0.023 0.064 0.022 0.199
9 0.002 0.439 0.036 0.147 0.026 0.423
12 0.001 0.406 0.076 0.060 0.006 0.535
Rubber
3 0.807 0.006 0.005 0.713 0.785 0.005
6 0.228 0.019 0.023 0.638 0.815 0.021
9 0.158 0.088 0.067 0.634 0.617 0.058
12 0.246 0.087 0.044 0.394 0.682 0.075
Nm-min
3 0.502 0.011 0.476 0.388 0.584 0.004
6 0.202 0.273 0.194 0.377 0.503 0.049
9 0.006 0.484 0.465 0.530 0.375 0.134
12 0.003 0.166 0.039 0.696 0.141 0.313
Me-mach
3 0.845 0.000 0.222 0.821 0.864 0.000
6 0.101 0.000 0.352 0.766 0.817 0.000
9 0.001 0.000 0.209 0.333 0.285 0.000
12 0.001 0.000 0.280 0.063 0.446 0.000
El-mach
3 0.015 0.023 0.216 0.425 0.026 0.018
6 0.010 0.072 0.805 0.248 0.005 0.033
9 0.030 0.053 0.899 0.027 0.022 0.056
12 0.003 0.012 0.123 0.024 0.004 0.063
(continued on next page)
143 P.S. Bhattacharya, D.D. Thomakos / International Journal of Forecasting 24 (2008) 134150
three countries, and to T
1
=44 months for four indus-
tries in the United States. Our choice of estimation and
evaluation samples is such that, for the monthly data we
use, it allows a reasonable sample for both estimating
the parameters and excluding outliers (which could
not be accurately forecasted anyhow) from the
evaluation periods.
10
5. Discussion of results
5.1. Causality tests and model reduction
An important premise in our set-up is the assump-
tion of non-causality of the exchange rate with respect
to the other variables in our system, as well as the
assumption of the triangular structure (based on the
two-dimensional approach of Obstfeld, 2002). In this
section, we discuss some results from the corres-
ponding tests for the United Kingdom and the United
States for PPI-based models; the full set of test results,
including those for Japan, is available on request. These
results are presented in Tables 1 and 2. As all of the tests
are based on a VAR model, we present the results using
four different lags, 3, 6, 9 and 12 months. Each of the
tables has 6 columns; of particular interest are columns
1 and 2 (tests of no block causality) and columns 5 and
6 (tests of no pairwise causality between exchange rates
and import prices within the triangular structure). The
other columns of the tables provide additional tests for
the triangular structure.
Our results can be summarized as follows. For most
of the industries in all three countries, we simulta-
neously find evidence of causality from the exchange
rate to the price variables (column 2), and no significant
evidence of causality from the price variables to the
exchange rate (column 1). There is either a feedback or
a mutual absence of causality in very few industries. We
can claim that, overall, the hypothesis of no block
causality, H
0
BC
:
yxj
=0, for all j =1, 2, , p, is mostly
accepted. This is an important finding, as it allows us to
move on from the unrestricted VAR model of Eq. (5) to
the restricted VARX model of Eq. (7). For a large
proportion of industries (about 70%), we have evidence
of no block causality from the price variables to the
exchange rate, and the tests suggest that our use of the
VARX model appears to be appropriate.
The results from the tests within the triangular
structure for the domestic price and import variables
are mixed (columns 3 and 4), while the results about
the causality between import prices and exchange rates
(columns 5 and 6) are more informative and robust.
For about 40% of all industries, we simultaneously
find evidence of no causality from import prices to the
exchange rate, but causality from the exchange rate to
import prices.
Table 2 (continued)
Industry/Lags Block causality for exchange rate Causality in the triangular structure
(ppi, imp)wfx wfx(ppi, imp) Block causality from imp to ppi Block causality from wfx to imp
ppi imp impppi impwfx wfximp
Furni.
3 0.408 0.939 0.534 0.913 0.207 0.875
6 0.428 0.001 0.337 0.977 0.285 0.004
9 0.386 0.004 0.783 0.928 0.171 0.003
12 0.658 0.000 0.716 0.739 0.348 0.000
xy denotes x does not cause y; the numbers in the cells are p-values of the corresponding test; cpi stands for the consumer price index; wfx
stands for the trade-weighted exchange rate; imp stands for the import price; and ppi denotes the producer price. In the table, Meat stands
for meat and meat preparations, Fruit stands for vegetables, fruit and nuts, fresh or dried, Beve stands for beverages, Chemical denotes
chemicals and related products, not essentially specified, Org-chem stands for organic chemicals, Inor-chem stands for inorganic chemicals,
Rubber stands for rubber manufactures, not essentially specified, Nm-min stands for non-metallic mineral manufactures, Me-mach denotes
metalworking machinery, El-mach stands for electrical machinery and equipment, and Furni denotes furniture and parts thereof.
10
Here we consider outliers as those observations that are about
three sample standard deviations away from the corresponding
sample mean. We do note in passing that, based on this definition,
the selected evaluation samples show suggestive evidence of
underlying normality for the series. A complete list of descriptive
statistics is available on request.
144 P.S. Bhattacharya, D.D. Thomakos / International Journal of Forecasting 24 (2008) 134150
Overall, the results of the causality analysis are
supportive of our use of prior economic information in
our system and of the corresponding model reduction
we obtain from turning the unrestricted VAR into the
triangular, restricted VARX model.
5.2. Comparison of forecasting performance
While the results on causality testing and model
reduction are supportive of our proposed approach, we
still need to see how our structural system fares out-
of-sample. Tables 310 contain all of the out-of-sample
evaluation results. We have grouped them by country:
Tables 3 and 4 for Japan, Tables 5 and 6 for the United
Kingdom, and Tables 710 for the United States. We
consider our findings quite interesting, both from an
economic perspective and from a purely forecasting
perspective.
If we compare the models based on their RMSE
values, then our system is either the best or second best
forecasting model in almost all cases considered. In
addition, it is more frequently the best model as the
forecasting horizon increases. This could potentially be
attributed to the appropriate or efficient use of the
economic structure we impose on our system. For the
1 month ahead forecasts, our system is the best (second
Table 4
DieboldMariano tests using a squared loss function for Japan
Industry Monthly rolling forecasts
VARMA VAR1 VAR2 VARMA VAR1 VAR2 VARMA VAR1 VAR2
One month ahead Two months ahead Three months ahead
Chemical 0.55 1.24 0.18 0.81 2.17
0.03
Food 0.64 1.77
Mach 0.82 0.24 1.36 0.02 1.02 0.60 0.11 0.80 1.26
Metal 1.06 1.69
0.57
Petrol 0.22 0.43 0.21 0.24 1.36 1.75
0.66 1.91
1.68
0.41
Wood 0.26 0.47 0.29 0.41 1.50
1.60
1.66
0.00
Chemical 0.34 0.14 0.95 0.99 0.53 0.84 0.31 0.30 0.38
Org-chem 0.25 0.63 1.28 0.25 0.71 0.41 0.36 1.49
0.87
Medi. 0.14 1.06 2.72
0.15 1.59
0.93
Tex-fabs 0.21 0.32 1.49
0.88
Mach 0.39 0.61 0.69 0.44 1.16 0.74 0.89 0.58 0.01
El-mach 0.38 0.77 2.75
0.43 1.75
0.94
Cell entries are the values of the test statistic;
,
and
denote the 99%, 95% and 90% levels of significance, respectively; and the squared
loss function is used to calculate the DieboldMariano test statistic. All results are reported in comparison with the SYS results: a negative number
indicates superior performance of the SYS, a positive number indicates superior performance of the competing model. For model and industry
details, see the notes of Tables 1 and 3, respectively.
Table 5
Root mean-squared errors (RMSEs) for United Kingdom
Industry Monthly rolling forecasts
VARMA SYS VAR1 VAR2 VARMA SYS VAR1 VAR2 VARMA SYS VAR1 VAR2
One month ahead Two months ahead Three months ahead
Tobacco 14.03 13.56 16.57 18.66 13.44 13.59 15.30 18.14 13.24 13.61 15.45 18.01
Pulp 29.34 28.56 29.89 30.09 29.21 29.61 30.34 30.55 30.27 32.45 32.02 32.74
Wood 7.10 8.34 9.06 8.96 8.08 7.96 9.21 9.12 8.28 7.97 8.76 8.64
Metal 17.94 18.07 19.62 19.63 17.80 18.32 19.42 18.88 18.15 18.37 20.27 18.26
Chemical 2.87 2.79 3.08 3.58 2.82 2.84 3.06 2.94 3.23 2.89 3.24 3.20
Org-chem 13.05 13.60 14.21 14.72 14.33 13.82 14.50 14.09 13.15 13.93 15.67 14.66
Medi. 13.29 14.43 15.19 16.48 13.29 14.48 15.97 15.23 13.34 14.51 16.08 15.10
Plastics 17.67 20.04 19.47 17.15 19.59 21.91 22.22 18.26 21.82 21.43 24.25 19.65
Tex-fabs 2.26 2.31 2.34 2.68 2.33 2.18 2.37 2.55 2.20 2.18 2.45 2.60
Iron 12.91 13.00 12.17 14.09 13.80 13.78 14.50 15.08 14.16 13.83 15.31 15.47
Nonfme 25.81 25.94 28.01 26.57 25.98 24.99 28.47 26.96 26.23 24.88 28.20 26.37
Mach 1.77 1.67 2.10 2.12 1.78 1.67 2.20 2.12 1.69 1.67 2.09 1.99
El-mach 2.82 2.55 2.59 3.27 2.72 2.39 2.55 3.19 2.48 2.41 2.49 2.96
For model and industry details, see the notes of Tables 1 and 3, respectively.
146 P.S. Bhattacharya, D.D. Thomakos / International Journal of Forecasting 24 (2008) 134150
The use of the DieboldMariano test for comparing the
forecasting models is appropriate in our context, but the
small evaluation sample we use may not provide
trusted inferences. A casual look at the corresponding
tables reveals that our system mostly performs
significantly better than the competing VAR models,
while it is on a par with the competing VARMA model.
There is no doubt that we would have preferred to have
had our system also outperform the VARMA model in
terms of the DieboldMariano test results, and it is
quite possible that a larger evaluation sample could
have produced such a favorable outcome for our
system. Be that as it may, the weight of evidence still
favors us: the test does not indicate that the VARMA
model is better than our system (the two models are
equivalent), while our system is better in forecasting
the direction of inflation (results not reported here) and
very frequently has smaller RMSE values.
What do these results say about the usefulness of our
approach? From a purely forecasting point of view, we
find that our restricted VARX model does provide
RMSE forecasting improvements over the unrestricted
VARmodels of, either Eq. (5) or Eq. (9), while it is on a
par with the VARMA model of Eq. (10). This result on
its own would be additional evidence on the ability of
low-order, parsimoniously specified time series models
to forecast well. However, taking the economic
arguments and the choice of evaluation measure into
account provides us with a different result. A properly
specified system that incorporates restrictions from
economic theory produces RMSE forecasts that are as
good as or better than those of a properly specified,
atheoretic time series model; such a system should be
of use to any practitioner that works in forecasting
industry-level inflation. Also, as noted at the end of
Section 3.1, one can use our system for preparing
Table 8
DieboldMariano tests using squared loss function for the United States (CPI-based models)
Industry Monthly rolling forecasts
VARMA VAR1 VAR2 VARMA VAR1 VAR2 VARMA VAR1 VAR2
One month ahead Two months ahead Three months ahead
Mfl 1.12 0.65 1.51
0.58 1.76
2.23
3.76
0.18 2.53
4.81
1.81
2.47
4.67
Appa 2.51
3.29
4.19
4.65
3.15
3.74
0.83 4.58
6.20
2.13
0.79 1.54
2.11
1.81
0.38 2.81
2.03
2.53
1.93
1.90
3.70
0.23 2.16
3.27
0.65 1.81
2.16
1.40
Inor-chem 0.35 1.05 0.10 0.17 1.25 0.49 1.65
1.15 0.15
Rubber 0.26 1.89
1.86
0.41 1.96
2.21
1.10 1.72
1.66
Nm-min 1.55
2.43
4.11
1.68
1.89
3.55
0.66 1.92
3.01
1.78
0.15 1.14
Furni 1.06 2.94
3.65