Economic Modelling 70 (2018) 78–86

Contents lists available at ScienceDirect

Economic Modelling
journal homepage: www.journals.elsevier.com/economic-modelling

A quarterly Phillips curve for Switzerland using interpolated

data, 1963–2016☆
Rebecca Stuart
Central Bank of Ireland, Ireland

A R T I C L E I N F O A B S T R A C T

Keywords: This paper estimates a quarterly Phillips curve for Switzerland, using interpolated data starting in 1963. Since
Switzerland only annual GDP data are available before 1980, the paper first discusses how to interpolate them using a
Historical statistics multivariate Chow-Lin procedure and by adapting forecast combination methods. The preferred interpolated
Long time series series is then used to estimate the Phillips Curve over a 50-year period. The results indicate two structural breaks
Phillips curve which appear to coincide with shifts in the Swiss National Bank's monetary policy strategy.
JEL number:
E1
E4
N1

1. Introduction work finding and preparing data for estimation.

Economic historians regularly work with imperfect data.1 This paper
studies the case in which the time series in a regression model are
observed at different frequencies. This problem regularly occurs with
aggregate macroeconomic data. Historical data for many important, but
hard to construct series, such as GDP, are available only on an annual
frequency whereas others which are much easier to observe, such as
prices, are available quarterly or even monthly.
Often, the frequency of observation may increase over time with more
recent data sampled more finely, as statistical agencies become more
ambitious in collecting the data. This raises the question of how best to
combine historical data, that may be annual, with recent data that may be
quarterly or perhaps monthly. While researchers interested in the current
functioning of the economy may simply drop the historical data, eco-
nomic historians are critically interested in how the economy has evolved
over time. They therefore do not have this luxury. Indeed, research on
historical macroeconomic time series typically involves considerable
The focus in this paper is on the inflation process. Estimated inflation
equations, such as Phillips curves, ought to be routinely tested for
structural breaks at unknown points in time. In such tests, the beginning
and end of the sample is “trimmed” and the search for breaks is under-
taken on the remaining observations. This means that long time spans of
data are highly desirable. That is often problematic in the case of Phillips
curves for advanced economies since data on real GDP that are necessary
to construct estimates of the output gap are often only available annually
before the 1980s or 1990s.
In this paper, a Phillips Curve for Switzerland is estimated on quar-
terly data, with the aim of studying as long a sample as possible. Being
small and highly open, Switzerland has experienced a number of external
shocks, many related to sharp exchange rate movements arising from the
Swiss franc's traditional role as a safe haven currency.2 The Swiss Na-
tional Bank has also changed the weight it attaches to the exchange rate
in setting monetary policy, most recently by adopting an exchange rate
floor of CHF 1.20 per euro between 2011 and 2015. Thus, a long time

☆
This is a much revised and extended draft of an earlier paper entitled ‘A Quarterly Phillips Curve for Switzerland’. The views expressed are solely my own. This paper is related to an
earlier paper on the Irish Phillips Curve with Stefan Gerlach and Reamonn Lydon. I thank my co-authors for many useful discussions. I also thank Daniel Kaufmann, Ronald Indergand,
participants at a KOF Swiss Economic Institute Research Seminar, at the Swiss Society of Economics and Statistics Annual Conference 2017 and at a University of Basel seminar for helpful
comments. Contact information: Rebecca Stuart, Central Bank of Ireland, New Wapping Street, North Wall Quay, Dublin 1, Ireland, email: rebecca.j.stuart@gmail.com, tel þ353 1 224
4159, website: http://rebeccastuart.net/.
E-mail address: rebecca.j.stuart@gmail.com.
1
For instance, Kaufmann (2016) argues that measurement error in 19th century CPI data makes it difficult to distinguish between periods of inflation and deflation. He goes on to argue
that this has led authors to conclude incorrectly that the pre-WW1 evidence that deflation need not be associated with recessions is flawed.
2
For a discussion of Swiss monetary policy see, for instance, Baltensperger and Kugler (2017), Kugler and Sheldon (2010), Gerlach and Jordan (2012), Bernholz (2007), Jordan and
Peytrignet (2007) and Kohli and Rich (1986).

https://doi.org/10.1016/j.econmod.2017.10.012
Received 18 July 2017; Received in revised form 12 October 2017; Accepted 22 October 2017
Available online 11 November 2017
0264-9993/© 2017 Elsevier B.V. All rights reserved.
R. Stuart
span of data is required to understand changes that have occurred in the
Swiss inflation process.
However, while monthly consumer prices are available since 1921
and monthly import prices from 1963, consistent Swiss real GDP data are
available on a quarterly basis only since 1980. A second quarterly real
GDP series is available from 1965 to 2013 but it is not compiled
consistently with the more recent data, with the result that the correla-
tion of the two series for the overlapping period is just 0.62. Nonetheless,
estimates of annual real GDP are available since 1851 and a number of
other indicators of quarterly real economic activity are available since
the late 1950s. This raises the possibility that quarterly GDP data might
be interpolated, allowing the researcher to fit a Phillips Curve on data
starting around 1960.
How best to do so is the focus of the first part of the paper. A cubic
spline is used to obtain a baseline interpolated series of quarterly real
GDP starting in 1960Q1 and ending in 2016Q2. This baseline series is
then compared with series obtained using the multivariate interpolation
method proposed by Chow and Lin (1971). This method uses ‘indicator’
series to distribute the annual data over the four quarters of the year.
Next, well-known forecast combination methods are adapted to
combine the interpolated data series optimally.3 The accuracy of the
various interpolation methods is assessed by comparing the growth rate
of the constructed quarterly data with actual quarterly GDP data for the
period after 1980.
The baseline cubic spline interpolation is as strongly correlated with
actual real GDP data after 1980 as interpolations obtained using the
multivariate Chow-Lin method. However, the measure of quarterly real
GDP constructed by optimally combining the four interpolated time se-
ries is marginally more closely correlated, in growth rates, with actual
GDP than those of the baseline and the Chow-Lin interpolations.
The preferred interpolated series of quarterly real GDP is used to
estimate a quarterly Swiss Phillips Curve. Output gaps are first con-
structed using both the interpolated quarterly real GDP series, starting in
1960, and the actual quarterly real GDP data, starting in 1980. The
output gap using the interpolated data is very similar to the output gap
calculated using the actual data over the period 1980–2016 for which
both gaps are available. Indeed, the correlation coefficient of the two
output gaps over this period is 0.93.
Next, the quarterly Phillips Curve is estimated over the sample
1963Q2 to 2016Q3 using OLS. This period covers some important epi-
sodes in recent Swiss economic history, including the oil crises of the
1970s, the period of exceptional appreciation of the Swiss franc around
1980, the great moderation and the global financial crisis in the 2000s.
The period has been studied previously in the literature using annual
data; the quarterly estimates provide an interesting comparison to these
earlier results.4
In line with the previous literature, there appears to be a break in the
mid-1970s (Baltensperger and Jordan (1998)). However, unlike this
literature this break does not coincide with a worsening of the
inflation-output trade-off so much as a reduction in inflation persistence.
A second break in the relationship in the early 1990s is associated with a
reduction in the importance of the output gap, which is also similar to
findings elsewhere (Gerlach (2016)).
Overall, the paper leads to three main conclusions.
First, estimates of the Phillips curve starting in the 1980s using output
gaps computed from the actual and interpolated quarterly real GDP are
remarkably similar. This suggests that the estimates of the Phillips curve
using the interpolated data for the period 1960–1980 are probably
similar to those one would have obtained using actual data if they had
been available. Overall, it seems feasible to interpolate annual GDP for
Switzerland before 1980 and estimate quarterly Phillips curve models for
a much longer period than has been undertaken in the literature.
3
Forecast combination methods are discussed in detail in Timmerman (2006).
4
See Gerlach (2016) and Baltensperger and Jordan (1998).
79
Economic Modelling 70 (2018) 78–86
Fig. 1. Output gap, consumer price inflation and import price inflation, 1960–2016.
Second, the simple benchmark interpolation method using a cubic
spline seems to work well compared to the more refined multivariate
Chow-Lin technique or optimally constructed combined interpolations in
the sense that the correlation of the growth rates of the interpolated se-
ries with actual quarterly real GDP growth are 0.70 using the cubic
spline, 0.70 also using the Chow-Lin method, and 0.74 using the com-
bined interpolations.
Third, tests for structural breaks reveal that there were shifts in the
Phillips Curve in 1974 and 1993. These breaks appear to coincide with
shifts in the monetary policy regime of the Swiss National Bank (SNB).
Specifically, the first break coincides with the introduction of monetary
targets following the floating of the Swiss franc, while the second co-
incides with the effective implementation of medium term targets, which
lagged the actual introduction of these measures,5 and the growing use of
the interest rate as an indicator of the monetary stance.
The paper is organized as follows. In the next Section, the monetary
and economic background during the sample period is discussed. In
Section 3, the baseline and Chow-Lin interpolation methods are evalu-
ated. In Section 4, the method for combining the interpolated series is
outlined and assessed. Section 5 presents the Phillips curve estimates,
and Section 6 concludes.
2. Monetary and economic background
This section outlines the historical background. Fig. 1 shows the
annual output gap, alongside annual changes in inflation and import
prices (note that import prices are measured on the right-hand scale)
since 1960.
From Fig. 1, inflation increased trend-wise during the 1960s. During
this time, the Swiss exchange rate was fixed against the US dollar, and the
authorities had little control over domestic inflation. Indeed, the Bretton
Woods system relied initially on conservative US monetary and fiscal
policy to ensure price stability.6 In the 1960s, the US was shifting towards
a more inflationary path as a consequence of expansionary government
spending, partially related to the war in Vietnam. Switzerland, which was
experiencing strong economic growth, began to suffer large capital in-
flows and imported inflation.
The early 1970s are characterized by a sharp business cycle upswing
and rising inflation. Renewed speculative capital inflows into
Switzerland in early 1971 were followed by the suspension of gold
convertibility by the US later in the year. Nonetheless, Nelson (2007)
5
See Rich (2003) for a discussion.
6
Although not formally a member of the Bretton Woods institutions, Switzerland
effectively participated in the system from 1945 since the currency was fixed to gold.
Indeed, Switzerland only officially became a member of the Bretton Woods institutions in
1992. See Baltensperger and Kugler (2017, p. 94).
R. Stuart Economic Modelling 70 (2018) 78–86

reports that Switzerland was almost the only country in the OECD to
report a positive output gap in 1972, and indeed the Swiss output gap had
been positive since at least 1969. With international markets in turmoil
following the US decision to suspend convertibility, Switzerland attrac-
ted strong inflows, causing the SNB to float the Swiss Franc in 1973.7
The sudden lack of an exchange rate anchor presented a challenge to
the SNB (Bernholz (2007)). In response, it adopted a policy of targeting
M1 growth in 1974, the immediate effect of which was a sharp slowing in
M1 growth.8 The result was an abrupt reversal of the business cycle, as is
evident from Fig. 1. GDP and industrial production declined markedly,
while inflation also declined, albeit with a lag (Baltensperger and Kugler
(2017, p. 111)).
The SNB's success in controlling inflation had a large impact on its
credibility,9 however the cost of this was strong capital inflows in 1977
and 1978. The Swiss Franc appreciated sharply, and industry came under
pressure with reports that “entire production facilities [had been aban-
doned] because of the exchange rate’ (Bernholz (2007, p. 179)). The SNB
became increasingly concerned about the impact of the exchange rate on Fig. 2. Quarter-on-quarter growth rates in real GDP series.
economic activity, and temporarily fixed the Swiss Franc vis-
a-vis the
German Mark in October 1978.10 When it resumed monetary targeting in output gap began to close.
1980, the SNB targeted the monetary base rather than M1.11 The low inflation environment was cemented in January 2000 when a
Following a spike in the early 1980s, inflation declined through much new inflation targeting regime was adopted by the SNB.16 The business
of the mid-1980s, exacerbated by the collapse in oil prices in 1986 cycle downturn associated with the bursting of the dotcom bubble was
(evident in import price growth in Fig. 1).12 The wider global slowdown, followed by an upswing in the mid-2000s, and throughout inflation
accentuated by the US stock market crash of October 1987, raised fears of remained below 5%.
a global slump and led the SNB to maintain a more expansionary policy Following the failure of Lehman Brothers in September 2008, the
than otherwise. As a result, inflation began to rise again, accompanied by downturn in the global economy coupled with low oil prices raised the
an upswing in the business cycle.13 prospect of deflation.17 The Swiss Franc came under upward pressure,
In response, the SNB changed strategy, switching to 5-year, rather particularly following the emergence of the sovereign debt crisis in the
than annual, monetary targets.14 There followed a period, from 1991 to euro area. In response, the SNB introduced an exchange rate floor of 1.20
1993, referred to by Rich (2003, p. 38) as ‘trial and error’. This was partly Swiss Francs to the euro in September 2011, and inflation began to rise.
because the economy was buffeted by external shocks during this time, However, in the context of expectations of a euro area quantitative easing
including the fallout from the reunification of Germany and severe ten- programme and debt sustainability issues in Greece, the SNB removed
sions in the EMS. Overall, GDP growth slowed substantially.15 the exchange rate floor, and inflation declined.
In addition, although the SNB could in theory still control the mon-
etary base, money growth was losing its value as an indicator of stance. 3. Data interpolation
The SNB thus began increasingly looking to the interest rate to judge the
current stance of policy (Rich (2003)). In September 1993, SNB staff were 3.1. GDP data
instructed to provide internal guidelines on the future course of action
(Rich, 2003, p. 42). Thereafter, inflation declined and the negative In this section, annual data on real GDP are interpolated in several
ways and the performance of each method is evaluated. The GDP data in
this study are collected from two sources. Annual GDP data from 1960
7
The financial turmoil had resulted in the Swiss Franc being revalued prior to its are taken from Gerlach (2016), who constructs them by splicing data
floating (Marsten (1993).
8 series from the website Historical Statistics of Switzerland Online.18 To
Although using monetary targets from 1974, the SNB only publicly announced a
target in 1975. See Baltensperger and Kugler (2017). evaluate how well the interpolated series describe the GDP data, they are
9
Nelson (2007, p. 718) reports a senior SNB official as stating that adoption of mon- compared with actual quarterly real GDP data which are available since
etary targets showed that ‘we mean business’. Furthermore, the SNB's ability to regain 1980. The quarterly real GDP data are taken from the website of the Swiss
control over inflation is impressive, particularly compared to other small open economies State Secretariat for Economic Affairs (SECO) for the period 1980
(Nelson (2007)).
10
See Gerlach and Jordan (2012) for a discussion.
to 2016.
11
This move to targeting in the money base reflected the SNB's concerns about its ability In addition, there is an earlier quarterly series, also compiled by
to forecast the multiplier in periods of high exchange rate volatility, see Gerlach and SECO, which is available for the period 1965Q1-2013Q3. However, this
Jordan (2012). series is discontinued and is not estimated consistently with current
12
It has been argued that this increase in inflation in the very early 1980s occurred
quarterly data. The seasonally adjusted, quarter-on-quarter growth rates
because the SNB was too slow in removing the excess liquidity resulting from defending
the currency during the period of exchange rate fixity. See Baltensperger and B€ ohm of these two quarterly series are included in Fig. 2 (solid lines). Two
(1984), Baltensperger (1985), Kugler and Rich (2002) and Rich (2003). points are of note. First, the discontinued series is remarkably volatile in
13
See Baltensperger and Kugler (2017). Gerlach and Jordan (2012) argue that taking the pre-1980 years, recording quarter-on-quarter growth rates in excess
account of credit growth in addition to money growth, would have signaled this issue. of ± 2% on a surprisingly regular basis. Second, the correlation of the
14
Also at this time, the introduction of the Swiss Interbank Clearing system and an
overhaul of the liquidity requirement regulations resulted in a large decline in money
growth rates of the two series in the overlapping period is 0.61. The
growth, and played a part in the move away from annual growth targets. interpolations below are more highly correlated with the post-1980 data.
15
See Kleinewefers Lehner (2007) for a discussion. While this was initially attributed to
cyclical factors, it gave way to a prolonged period of lower growth. Having averaged 2.2%
from 1980 to 1990, growth averaged just 1.1% from 1990 to 2005 (Baltensperger and
16
Kugler, 2017). During the 1990s, a series of reforms were undertaken to make Switzerland The new monetary policy approach sets out a target for CPI inflation of “less than 2
more attractive to business, including reducing public sector barriers, introducing free per cent”, uses of inflation forecasts, and is implemented through a target range for 3-
movement of people across borders and reforming the agricultural and network sectors. month Libor.
17
See Rudolf and Zurlinden (2010) for a discussion of the possible impact of these reforms See Gerlach and Jordan (2012).
18
on the contribution of labour input. See: http://www.fsw.uzh.ch/hstat/nls_rev/overview.php.

80
R. Stuart Economic Modelling 70 (2018) 78–86

Table 1
Interpolated series compared to actual quarterly GDP.

Correlation coefficient Mean absolute error (percent)

Baseline interpolation

Cubic spline 0.70 0.35

Indicator series used*: Chow-Lin interpolations

IP, Man, RS, CS 0.42 0.81

IP, Man, RS 0.43 0.79
IP, Man, CS 0.56 0.59
IP, RS, CS 0.42 0.82
Man, RS, CS 0.42 0.81
IP, Man 0.56 0.59
IP, RS 0.43 0.79
IP, CS 0.57 0.58
Man, RS 0.44 0.78
Man, CS 0.56 0.59
RS, CS 0.34 0.93
Fig. 3. Cross correlation of ± 6 leads and lags of the growth rate in cubic spline inter- IP 0.57 0.46
polated series with the growth rate in actual quarterly GDP, 1980–2016. Man 0.56 0.46
RS 0.35 0.67
CS 0.70 0.34
3.2. Baseline interpolation
Combination methods

The first interpolation method used is a cubic spline, a common
interpolation method that does not use any information except that in the
annual real GDP series itself. This is considered the baseline
interpolation.
The cubic spline is applied to the change in the log level (or annual
growth rate) of real annual GDP to obtain a quarterly series.19 It is
implemented here by requiring that the fourth quarter growth rate in the
interpolated series is equal to the growth rate of the annual series. Before
evaluating this baseline interpolation, cross-correlations of the interpo-
lated quarterly growth rates with actual data on quarterly GDP growth
for the period 1980–2016 are first computed to determine whether the
interpolated series leads or lags the actual data. Fig. 3 indicates that the
data interpolated using the cubic spline leads the actual real GDP data by
three quarters. The analysis below is therefore conducted using a three-
quarter lead of the series interpolated with the cubic spline method.
Next, the correlation coefficient and the mean absolute error (MAE)
of the growth rate of the interpolated quarterly series with the actual
quarterly real GDP data are computed over the period 1980–2016.20 As is
evident from the first row of Table 1, for this interpolation, the correla-
tion coefficient is 0.70 and the MAE is 0.35 per cent.
3.3. The Chow-Lin method
As previously noted, the cubic spline method of interpolation does not
use any information beyond what is already present in the annual data.
Next, information from various ‘indicator series’ is introduced into the
interpolation procedure using the method proposed by Chow and Lin
(1971).21 This method is a regression-based interpolation technique that
relates higher-frequency indicator series, zt , observed in quarter t, to a
lower-frequency benchmark series, yk , observed in year k to obtain a
quarterly interpolated series, xt . In the first step, the values of yk are
regressed on the annualized values of the indicator series, denoted by zk :
yk ¼ βzk þ ek
19
Other simple interpolation methods, such as linear interpolation and quadratic
interpolation were also considered for the baseline, but these perform less well in the
evaluation tests than the cubic spline, so they are omitted here.
20
Since the cubic spline reports year-on-year growth rates, the series is divided by four
in order to obtain a quarter-on-quarter growth rate.
21
Other methods which also include information from indicator series, such as the
Denton method and variations on the Chow-Lin method which use a Litterman prior over
the variance-covariance matrix, provide somewhat different individual series in-
terpolations compared to those presented in this section, but do not improve on the
combination of interpolations presented in the next section.
Preferred combination 0.74 0.33
Note: * IP ¼ industrial production, Man ¼ manufacturing output, RS ¼ retail sales, CS ¼ car
sales.
The estimated coefficient, b
β, is then applied to the quarterly obser-
vations of the indicator series, zt , to obtain a preliminary quarterly
interpolated series, xt* :
x*t ¼ b
βzt
In the second step, the final interpolated series, xt , is obtained by
distributing the difference between the annual value of yk , and the sum of
values of xt* in each year across the quarters of the year using an AR(1)
process. In this step, the distribution is subject to the following condition:
X
4k
xt ¼ yk
t¼1k
where xt contains observations for each of Q1k , Q2k , Q3k and Q4k , for all
k. This final condition ensures that the sum of the quarterly log changes
in the interpolated series, xt , sum to the annual log change of the
actual data.
There are two important implications of this method. First, the
movement of the indicator series is only transferred to the interpolated
series if the annualized growth rate in indicator series and the growth
rate in the low frequency variable are correlated. Second, the method
assumes that the linear relationship observed in the regression of the low
frequency variable on the annualized indicator variable also holds be-
tween the quarterly series, zt and the true, but unobserved, quarterly
values of yk . As will be seen, this need not be the case.
3.4. Multivariate Chow-Lin interpolations
The indicators are obtained from the OECD Main Economic Indicators
database. Specifically, the growth rates of quarterly data on industrial
production, manufacturing output, retail sales, and car sales are used as
indicators.22 These data are all available in levels from at least the first
quarter of 1959.23 The indicator series are chosen because they are likely
to be correlated with economic activity and because they are available
22
The unemployment rate is another possible indicator series. However, in Switzerland,
the unemployment rate was exceptionally low and stable for a large part of the sample
period. It therefore had little relation to real GDP growth, and is therefore not used here.
23
All indicator series are seasonally adjusted.

81
R. Stuart
over the entire sample period. However, the correlation with annual real
GDP growth differs across the series, from 0.76 for retail sales to
approximately 0.55 for both industrial production and manufacturing
output, and 0.13 for car sales.
A priori, it is not clear which combination of these series will perform
best. Therefore, all possible combinations of these indicators are esti-
mated using the Chow-Lin method. In total, 15 specifications, including
univariate ones, are estimated.24
As with the cubic spline interpolation, to evaluate the series inter-
polated using each of these indicators, comparisons are made with the
quarterly real GDP data available from 1980 onwards.25 The middle
panel of Table 1 shows the correlation of the growth rate of the series
interpolated using the Chow-Lin procedure with actual quarterly GDP
growth post-1980. Three results are of note:
 First, the series interpolated using just retail sales has the lowest
correlation with quarterly real GDP growth, despite the fact that retail
sales is most highly correlated with annual real GDP growth.
 Second, and similarly, only the series interpolated using just car sales,
which has a low correlation with real GDP on an annual basis, per-
forms as well as the cubic spline (correlation coefficient ¼ 0.70).
Indeed, this is also the only series that has a marginally smaller MAE
than the cubic spline (0.34 per cent compared to 0.35 per cent).
 Third, none of the multivariate interpolations perform better than the
univariate interpolations.
What is causing these three results? Consider the correlations in
Table 1: when retail sales and car sales are used, the interpolated series is
very similar to that obtained using only retail sales (the correlation co-
efficients of these interpolated series with actual quarterly GDP growth
are 0.34 and 0.35). Essentially, the low correlation of annualized growth
in car sales with annual real GDP growth results in it having a low
weighting in the interpolation process. Thus, the comparatively good
performance of the series interpolated using just car sales arises, not from
the correlation of the annualized series with the annual GDP, but from
the distribution of the quarterly changes thereafter.
In a similar fashion, the comparatively poor performance of the series
interpolated using just retail sales calls into question the assumption in
the Chow-Lin procedure that the linear relationship observed in the
regression of the low frequency variable on the annualized indicator
variable also holds between the quarterly series.
Overall, it seems that the disappointing performance of the multi-
variate Chow-Lin procedure is due to a combination of the low weight
placed on car sales in the annual regressions, and the high weight placed
on retail sales in these regressions. The best performing multivariate
interpolations are those which load heavily onto industrial production
and manufacturing output. However, since these series are relatively
similar, there is little additional information to be gained from combining
them. As a result, these multivariate interpolations have similar corre-
lation coefficients with actual GDP growth as the univariate in-
terpolations using the same variables (0.56–0.57).
24
It is also possible that there is a common factor underlying all series which is strongly
correlated with GDP growth. If so, the principal components of the four series might be a
good additional indicator series. However, this proves not to be the case: the performance
of the series interpolated using the first principal component is worse than that of the
series interpolated using car sales.
25
Cross-correlations of leads and lags of each of the interpolated series against the actual
quarterly real GDP growth (similar to Fig. 3, but not shown here for brevity) indicate that
the correlation is highest contemporaneously in all instances, except the series interpo-
lated using only car sales, which appears to lead actual GDP growth by two quarters.
Therefore, the analysis below is conducted using the contemporaneous interpolated series
in all cases except the series interpolated using only car sales, for which a two-quarter lead
is used.
82
Economic Modelling 70 (2018) 78–86
4. Combining interpolated series and final assessment
Since the multivariate Chow-Lin procedure does not improve the
performance of the interpolation, in this section forecast combination
methods are applied to the four series interpolated using the individual
indicators (the last four rows of the middle panel of Table 1). Combi-
nation methods are employed when forecasters have several competing
models among which they cannot necessarily identify the best. The
literature indicates that even the simplest of combination methods can
improve predictions. While this paper studies interpolations rather than
forecasts, forecast combination methods can easily be adapted to the
current case.
4.1. Combination methods
Combination methods involve computing a weighted average of the
possible models or forecasts. The literature proposes various methods to
calculate the required weights, which are generally selected by
comparing the output of the models with the actual data over some
‘reference’ period. The Appendix provides details of the results using four
different combination methods; however, here only the most successful
method based on the evaluation criteria used previously is discussed.
Specifically, the coefficients from a regression of the interpolated series
on the actual data are used as weights, as discussed in Granger and
Ramanthan (1984):
X
i
yt ¼ ωi xit þ εt
0
Where yt is the actual data at time t, xit are the i:th interpolated series and
P
ωi is the estimated parameters. The sum of the coefficients, i0 ωi , is
constrained to equal one. This results in the series interpolated using car
sales receiving the highest weight, 0.69, the series interpolated using
manufacturing output receiving a weight of 0.15, the series interpolated
using industrial production receiving a weight of 0.09 and the series
interpolated using retail sales receiving a weight of 0.06.
The bottom panel of Table 1 presents the correlation and MAE of this
combined series with actual real GDP. At 0.74, the pairwise correlation of
this combined interpolation outperforms that of both the baseline cubic
spline, for which the correlation is 0.70, and the best-performing Chow-
Lin interpolation (using just car sales), for which the correlation is also
0.70. Considering the mean absolute error of the interpolated series, the
combined interpolation is 0.01 percentage points more accurate than the
series interpolated using car sales alone and 0.02 percentage points more
accurate than the cubic spline (0.33 per cent compared to 0.34 per cent
and 0.35 per cent, respectively).
4.2. Final assessment
It is possible that the indicator series may be more (or less) strongly
correlated with actual GDP after 1980 than before. For instance, a higher
portion of the population owned cars in the 2000s than in the 1960s, and
thus car sales may be a poorer indicator series in the earlier period. It is of
course impossible to know whether the relationship between the indi-
cator series and real GDP changed over time since the latter series are
only observed from 1980 onward.
However, the baseline cubic interpolation, which only uses the in-
formation in the annual series itself, should presumably be equally close
to the actual real GDP series before and after 1980. It is instructive
therefore to compare the optimally combined interpolated series with the
baseline interpolation. Strikingly, the correlation coefficient of the two
series is 0.96 in the post-1980 period and 0.97 in the pre-1980 period.
Thus, insofar as the cubic spline is considered an equally good interpo-
lation of the real GDP data across the full period, it appears that this is
true also for the combined interpolation.
R. Stuart Economic Modelling 70 (2018) 78–86

Table 2
Phillips curve estimates comparing actual and interpolated GDP output gaps, 1980Q1-
2016Q3.

(1) (2) (3)

Constant 0.15*** 0.15*** 0.15***

(0.04) (0.4) (0.04)
First lag inflation 0.20*** 0.18*** 0.18**
(0.07) (0.07) (0.07)
Second lag inflation 0.19** 0.18*** 0.18**
(0.07) (0.07) (0.07)
Third lag inflation 0.16** 0.16** 0.16*
(0.07) (0.07) (0.07)
Fourth lag inflation 0.14** 0.14* 0.14*
(0.06) (0.06) (0.06)
Import prices 0.19*** 0.19*** 0.19***
(0.02) (0.02) (0.02)
First lag actual GDP output gap 0.08*** 0.01
(0.03) (0.07)
First lag interpolated GDP output gap 0.10*** 0.11
(0.03) (0.08)
Fig. 4. Output gaps based on actual and interpolated data. R-squared 0.64 0.64 0.64
Durbin-Watson Statistic 1.69 1.71 1.71

4.3. Final series
The dashed line in Fig. 2 (above) is the quarter-on-quarter growth rate
of the preferred final series.26 This is shown along with the actual data
post-1980 and the discontinued quarterly series which is available from
1965 to 2013. Overall, the interpolated series is less volatile than the
other two series. Compared to the post-1980 data, the lower volatility of
the interpolated series means that it does not capture, for instance, the
depth of the financial crisis in 2008/9, but it does capture the trend in the
actual data well. Compared to the discontinued series, the interpolated
data are significantly less volatile: while following a similar trend, the
growth rate of the interpolated is always within a range of ± 2% (with the
exception of the very first quarter). This seems reasonable for quarter-on-
quarter data; in contrast, the discontinued series regularly records
growth rates in excess of ± 4%.27
Therefore, the combined interpolation series is used in the below
application to the Swiss Phillips Curve. Nonetheless, bearing in mind the
possibility that there may be measurement error in the actual real GDP
series, the analysis is first conducted for both the interpolated series and
the actual series over the period 1980Q1-2016Q3.
5. Estimating a Swiss Phillips curve, 1963–2016
Note: standard errors in parenthesis. ***/*/* indicate significance at the 1%, 2.5% and 5%
levels.
X
4
π t ¼ β1 þ αi π ti þ β2 Gapt1 þ β3 Impt þ εt
i¼1
Where π t is the quarterly inflation rate, Gapt1 is the lagged output gap,
Impt is the quarterly percentage change of import prices and εt is the
error. Inflation and import price data are taken from the Swiss National
Bank's online database.30 Import price data are only available from the
start of 1963.
The model is first estimated using data starting in 1980Q2. The results
for the two different output gaps are presented in the first two columns of
Table 2. In both cases, the coefficients are remarkably similar. In
particular, the coefficient on the interpolated output gap is 0.10, and that
on the actual gap is 0.08.31 In addition, if both gaps are included in ‘horse
race’, the parameter on the interpolated gap has a positive sign, while the
parameter on the actual output gap has a negative sign, although both are
insignificant (p-values ¼ 0.17 and 0.88, respectively) (third column of
Table 2). As a result, the interpolated output gap is used for the full
sample in the estimation of the Phillips Curve below.

5.2. Half a century of the Swiss Phillips curve

5.1. Calculating and comparing the output gaps

Next, the Phillips Curve for the full sample is estimated using the same
Next, output gaps are calculated using both the preferred interpolated
specification as above, and the interpolated output gap. Since import
data and the actual GDP data in log-levels using a Hodrick Prescott filter,
prices are only available from 1963 onwards, the sample period is
with lambda of 1600, which is standard for quarterly data.28 The two
1963Q2-2016Q3. The results are presented in the first column of Table 3.
output gaps, referred to for simplicity as the ‘interpolated’ and the
However, a Bai-Perron multiple break-point test indicates that there are
‘actual’ gaps, are shown in Fig. 4; it is clear that both gaps follow a similar
two breaks in the regression: in the first quarter of 1974 and the third
pattern, although the gap calculated using interpolated data is smoother
quarter of 1993.32 Re-estimating the Phillips curve for each of these sub-
than that computed on actual data.29
periods provides further insight into the shifts driving policy. The results
To consider in more detail the effect of using interpolated data for
are reported in Columns 2, 3 and 4 of Table 3. It is interesting to consider
GDP, estimates of the Phillips curve using the two measures of the gap are
these results in greater detail since the breaks appear to coincide with
compared over the period for which actual data are available. A trian-
shifts in monetary policy.
gular Phillips Curve specification is estimated:
In relation to the first break, the shift between the first two subperiods

26 30
These data are available upon request from the author. The data are seasonally adjusted using the U.S. Census Bureau's X-13 seasonal
27
Indeed, the correlation of the interpolated series with the discontinued series is 0.55, adjustment method. The data are available monthly and are converted into a quarterly
compared to 0.74 with the post-1980 data. frequency by taking the value for the final month in each quarter.
28 31
To calculate a level for the interpolated series (which is in log changes), actual GDP in Indeed, if it is assumed that the output gap measured using the actual GDP series is
1980Q1 is used as a base quarter. However, the base makes little difference, since the filter subject to end-point bias, and data for 1980 are dropped from the sample, then the co-
is insensitive to levels. efficient on the output gap is the same in both equations (0.08).
29 32
In 1980 the two output gaps move somewhat differently, which may be the result of The test is conducted with 15% trimming and for a 5% significance level. The breaks
‘end point bias’ in the output gap computed using actual real GDP, to which Hodrick are identified using the repartition method, although sequential method also identifies
Prescott filtered data are subject. Nonetheless, overall, the two output gaps are closely two breaks: 1974Q1 (the same as the repartition method) and 1993Q4 (compared to
correlated (ρ ¼ 0.93). 1993Q3 using repartition method).

83
R. Stuart Economic Modelling 70 (2018) 78–86

Table 3 This break is in line with findings in Gerlach (2016) for Switzerland,
Estimates of Phillips Curve using interpolated GDP output gap, 1963Q2 – 2016Q3. and a more general finding internationally of a reduction in the signifi-
(1) (2) (3) (4) cance of the output gap during this period. One hypothesis is that if Swiss
1963q2 - 1963q2 - 1974q2 – 1993q4- monetary policy began to systematically offset movements in the output
2016q3 1974q1 1993q3 2016q3
gap to stabilize inflation at this time, then shocks to the output gap would
Constant 0.21*** 0.69** 0.68*** 0.17*** lead to serially uncorrelated fluctuations of inflation around its mean.
(0.05) (0.20) (0.11) (0.05) Since import price shocks, which stem from both exchange rate move-
First lag inflation 0.12 0.24* 0.10 0.04
(0.06) (0.12) (0.09) (0.12)
ments and changes in the foreign currency prices of Swiss imports,
Second lag inflation 0.24*** 0.37** 0.02 0.09 impact on Swiss inflation instantaneously, monetary policy was unable to
(0.06) (0.15) (0.07) (0.12) offset such shocks.
Third lag inflation 0.25*** 0.60*** 0.13* 0.01
(0.06) (0.17) (0.06) (0.11)
6. Conclusions
Fourth lag inflation 0.01 0.41** 0.03 0.01
(0.06) (0.17) (0.06) (0.11)
Import prices 0.14*** 0.12*** 0.14*** 0.18*** This paper addresses an issue frequently faced by economic histo-
(0.02) (0.03) (0.02) (0.03) rians: how to deal with data for which the frequency of observations
First lag interpolated 0.10*** 0.21*** 0.24*** 0.05 changes over time. Specifically, economic historians must often deal with
GDP output gap (0.03) (0.07) (0.04) (0.04)
R-squared 0.63 0.62 0.65 0.50
the question of how best to combine historical data, that may be annual,
Durbin-Watson Statistic 1.78 1.81 1.94 1.75 with recent data that may be quarterly or perhaps monthly.
Several methods were used to obtain an interpolated quarterly series
Note: standard errors in parenthesis. ***/*/* indicate significance at the 1%, 2.5% and 5%
levels. for Swiss real GDP since 1960. In the first instance, a cubic spline was
used to obtain a baseline interpolation. This was then compared with
several series interpolated using a multivariate Chow-Lin method.
Finally, forecast combination methods were adapted to combine the
seems to be in the level of inertia, which is strong in the first sub-period, interpolated series in an optimal manner. All series were evaluated
but almost entirely missing in the second. In the period 1963–1974, the against the actual real GDP data that are available since 1980.
inflation rates in the previous four quarters all have a significant impact Having obtained a preferred interpolated series computed using the
on the current inflation rate (although signs of the coefficients vary, the forecast combination methods, the paper next used this series in an
cumulative impact is 0.31). In the subperiod 1974–1993, only the third application to the Swiss Phillips Curve. First, comparisons are made be-
lag of inflation is significant, and the sum of the parameters is just 0.22. tween the output gap obtained using the preferred interpolated series and
The timing of this break (1974Q1) coincides with the introduction of that obtained from the actual real quarterly GDP since 1980. Second, the
monetary targets. Prior to this (and the floating of the Swiss Franc in gap obtained using the interpolated data was used in the estimation of
1973), the fixed exchange rate regime left the authorities in Switzerland the Phillips curve over the full sample period, 1963Q2 to 2016Q3. One
with limited control over inflation. The adoption of monetary targets reason economic historians use long sample periods, is to understand
signaled the SNB's preference for price stability and was followed by a how the economy has evolved over time. In this respect, Bai-Perron tests
reduction in inflation. for structural breaks at unknown points in time indicate two breaks in the
It is interesting that a shift in the Phillips curve relationship is also Swiss Phillips Curve relationship, one in the mid-1970s and one in the
identified at this time in studies using annual data by Baltensperger and early 1990s. The timing of these breaks was considered in detail.
Jordan (1998) and Gerlach (2016). However, in contrast to my results, The main findings of the paper are as follows. First, the series inter-
these studies find that the break is associated with a worsening of the polated using the cubic spline compares favorably with those interpo-
output-inflation trade-off in the 1970s compared to previous periods. The lated using the Chow-Lin method and those obtained using the optimal
results in Table 3 also indicate a worsening of this trade-off in the 1970s, combination methods. However, a series computed using the optimally
although the effect is marginal: a 1 percentage point increase in the combined method has a marginally higher correlation (0.74) with actual
output gap raises inflation by 0.21 basis points in the first subperiod, and GDP growth than either the cubic spline or the Chow-Lin interpolations
by 0.24 basis points in the second subperiod. One is tempted to conclude (both 0.70).
that the use of quarterly data may allow clearer estimates of the Phillips Second, the analysis suggests that the output gap obtained using the
curve than annual data. interpolated series and that obtained using the actual quarterly real GDP
Nonetheless, my results here fit well with the hypothesis that the shift series provide very similar estimates of the Phillips Curve in the period
in monetary policy in 1973/74 coincided with the break in the Phillips after 1980. From this it is inferred that Phillips curve estimates over the
curve. The shift to floating exchange rates gave the SNB more control period 1963–1980 using the interpolated data are likely to be similar to
over inflation, and the announcement of monetary targets should have those one would have obtained using actual data if they had been
driven expectations for money growth and therefore inflation in the available. Overall, the paper shows that it is possible to use interpolated
coming year. real GDP data to estimate quarterly Phillips curve models for Switzerland
Turning to the second break, Column 4 of Table 3 indicates that, in over the period since 1963, a significantly longer period than has pre-
the final subperiod, lagged inflation and the output gap are insignificant. viously been examined in the literature.
This break coincides with the end of the period referred to as ‘trial and Finally, the paper argues that the structural breaks identified by the
error’ by Rich (2003) following the introduction of medium-term mon- Bai Perron breakpoint tests appear to correspond to shifts in Swiss
etary targets. From September 1993, the Governing Board of the SNB monetary policy. Specifically, the first break, identified in 1974, coin-
benefited from guidance on the evolution of the interest rate, which had cided with the floating of the Swiss Franc at the end of the Bretton Woods
become the dominant measure of the monetary stance. In the subsequent regime. The second break, identified in 1993, coincides with the end of
years, the usefulness of money as an indicator of monetary stance the period of ‘trial and error’33 after the introduction of medium term
declined further, until inflation targeting was formally introduced money targets by the SNB and the growing use of the interest rate as an
in 2000. indicator of the monetary stance.

33
Rich (2003).

84
R. Stuart Economic Modelling 70 (2018) 78–86

Appendix. Forecast combination methods applied to interpolations

In this Appendix, four different combination methods are considered, one of which is discussed in the paper. The first method is an equal weighted
combination:

1 Xm
by ct ¼ xit
m i¼1

c
Where b
y t is the combined interpolated series at time t, and xit , is the i:th of m interpolated series. As such, in this case, each series is given the
same weight:

1
ω¼
m
While this method has been used with some success in the literature,34 by giving all series equal weight, the relative performance of the interpolated
series compared to the actual data is ignored. The other three methods take account of the performance of the series using different methods, and are
discussed extensively in Timmerman (2006). The first is a rank-based weighting scheme. While there are several ways to specify such a scheme, the one
proposed by Aiolfi and Timmerman (2006) is used here. This scheme sets the weights to be inversely proportional to the models' rank:
, !
1
X
m
1
ωi ¼ R i R i
i¼1

Where R 1i is the inverse of the rank of the i:th of m interpolated series, and ωi are the weights applied to each of i interpolated series and the combined
series. Here the models are ranked by their mean squared error in the period 1980–2016.
The second uses the inverse of the mean square errors of the interpolated series, MSEi , as weights, as proposed in Stock and Watson (2001):

ð1=MSEi Þ
ωi ¼ Pm
i¼1 ð1=MSEi Þ

The final method, which is discussed in the paper, uses the coefficients from a regression of the interpolated series on the actual data as weights, as
proposed by Granger and Ramanthan (1984):

yt ¼ ωi xit þ εt

Where the sum of the coefficients, ωi , which are also the weights, is constrained to equal one.35
In each of these final three cases, the combined series is then calculated as:

X
m
by ct ¼ ωi xit
i¼1

Table 1a shows the weights calculated for each series as calculated using these methods. Interestingly, while all measures weight car sales most
heavily, the weighting is much larger using the regression-based method (0.69) compared to the MSE method (0.35) and the ranking method (0.48).
Notably, in almost all cases, the remaining variables are weighted most heavily using the equal-weighting method. Using the regression-based weighting
scheme and the ranking method, manufacturing output gets the second largest weighting (0.15 and 0.24, respectively), and industrial output the third
largest (0.09 and 0.16, respectively). This ordering is reversed using the MSE method, with industrial production (0.26) slightly more heavily weighted
than manufacturing output (0.25). Finally, retail sales are weighed least heavily using both methods, although the weightings using the MSE-based
method and the ranking method (0.12 and 0.11, respectively) are almost twice than of the regression-based method (0.06).

Table 1a
Weightings used in combination methods

Industrial production Manufacturing output Retail sales Car sales

Equal weighting 0.25 0.25 0.25 0.25

Ranking method 0.24 0.16 0.12 0.48
Inverse-MSE-weights 0.26 0.25 0.11 0.39
Regression-based weights 0.09 0.15 0.06 0.69

34
See for instance De Pooter et al. (2010).
35
Constraining the coefficients to equal one guarantees that the combined forecast is unbiased, and that the errors from the combined forecast are not serially correlated, as they could be
if yt is persistent but the forecast errors from the models are not serially correlated. See Diebold (1988) for a discussion.

85
R. Stuart Economic Modelling 70 (2018) 78–86

Fig. 1a. Interpolations using combined methods.

Results

Fig. 1a shows the combined interpolated series. The series move quite closely together although some, for instance the equally weighted combi-
nation, are more volatile than others. Table 2a shows the correlation between these series and the actual quarterly data over the period since 1980.
Interestingly, based on the pairwise correlation, the combination methods using equal weights (correlation coefficient 0.66) displays a higher corre-
lation than the ranking method (0.62), and only a marginally lower correlation than that based on MSE (0.70).
In contrast, the regression-based combination is more successful with a correlation coefficient of 0.74, somewhat higher than any other interpolated
series. Furthermore, when considering the mean absolute error of the interpolated series, the regression-based combination is more accurate than the
other series.

Table 2a
Combined interpolations, comparison with actual quarterly GDP.

1980–2016 Correlation coefficient Mean absolute error

Equal weighting 0.66 0.38

Ranking method 0.62 0.38
Inverse-MSE-weight 0.70 0.35
Regression-based weights 0.74 0.33

References Jordan, Thomas, Peytrignet, Michel, 2007. The path of interest rate management and
inflation forecasts. In: Swiss National Bank. Neue Zürcher Zeitung Publishing, Zurich,
pp. 1907–2007.
Aiolfi, Marco, Timmerman, Allan, 2006. Persistence of forecasting performance and
Kaufmann, Daniel, 2016. Is Deflation Costly after All? Evidence from Noisy Historical
combination strategies. J. Econ. 135 (1–2), 31–53.
Data. KOF Working Papers, No. 421, November.
Baltensperger, Ernst, 1985. Disinflation – the Swiss experience. Z. Wirtsch.
Kleinewefers Lehner, Anne, 2007. Economic Trends. In: Swiss National Bank. Neue
Sozialwissenschaften 105 (2/3).
Zürcher Zeitung Publishing, Zurich, pp. 1907–2007.
Baltensperger, Ernst, B€ohm, P., 1984. Geldmengenpolitik und Inflationskontrolle. In:
Kohli, Ulrich, Rich, Georg, 1986. Monetary control: the Swiss experience. Cato J. 5 (3),
M€oglichkeiten, Kosten, flankierende Massnahmen, Diessenhofen.
911–926. Winter.
Baltensperger, Ernst, Jordan, Thomas, 1998. Are there shifts in the output-inflation trade-
Kugler, Peter, Rich, Georg, 2002. Monetary policy under low interest rates: the experience
off? The case of Switzerland. Swiss J. Econ. Stat. 134 (2), 121–132.
of Switzerland in the late 1970s. Schweiz. Z. Volkswirsch. Stat. 138 (3).
Baltensperger, Ernst, Kugler, Peter, 2017. Swiss Monetary History since the Early 19th
Kugler, Peter, Sheldon, George, 2010. Unemployment and monetary policy in
Century. Cambridge University Press.
Switzerland. Swiss J. Econ. Stat. 146 (1), 185–208.
Bernholz, Peter, 2007. From 1945 to 1982: the transition from inward exchange controls
Marsten, Richard C., 1993. Interest rate differentials under Bretton Woods and the post-
to money supply management under floating exchange rates. In: Swiss National Bank.
Bretton Woods float: the effects of capital controls and exchange rate risk. In:
Neue Zürcher Zeitung Publishing, Zurich, pp. 1907–2007.
Bordo, Michael D., Eichengreen, Barry (Eds.), A Retrospective on the Bretton Woods
Chow, Gregory C., Lin, An-loh, 1971. Best linear unbiased interpolation, distribution, and
System: Lessons for International Monetary Reform. University of Chicago Press.
extrapolation of time series by related series. Rev. Econ. Stat. 53 (4), 372–375.
Nelson, Edward, 2007. Ireland and Switzerland: the jagged edges of the Great Inflation.
November.
Eur. Econ. Rev. 52, 700–732.
De Pooter, Michiel, Ravazzolo, Francesco, van Dijk, Dick, 2010. Term structure
Rich, Georg, 2003. Swiss Monetary Targeting 1974-1996: the Role of Internal Policy
forecasting using macro factors and forecast combination. Board of Governors of the
Analysis. ECB Working Paper Series, 236, June.
Federal Reserve System. International Finance Discussion Papers, Number 993,
Rudolf, Barbara, Zurlinden, Mathias, 2010. Productivity and economic growth in
January.
Switzerland 1991-2006. Swiss J. Econ. Stat. III (3), 577–600.
Diebold, Francis X., 1988. Serial correlation and the combination of forecasts. J. Bus.
Stock, James, Watson, Mark W., 2001. A comparison of linear and nonlinear univariate
Econ. Stat. 6, 105–111.
models for forecasting macroeconomic time series. In: Engle, Robert F.,
Gerlach, Stefan, 2016. The Output-inflation Trade-off in Switzerland, 1916-2015. CEPR
White, Halbert (Eds.), Festschrift in Honour of Clive Granger. Cambridge University
Discussion Paper, DP11714, December.
Press, Cambridge, pp. 1–44.
Gerlach, Stefan, Jordan, Thomas, 2012. Tactics and strategy in monetary policy:
Timmerman, Allan, 2006. Forecast combinations. In: Elliot, Graham, Granger, Clive W.J.,
Benjamin Friedman's thinking and the Swiss National Bank. Int. J. Central Bank. 8
Timmerman, Allan (Eds.), Handbook of Economic Forecasting, North-Holland,
(S1), 23–56.
pp. 135–196. Chapter 4.
Granger, Clive W.J., Ramanthan, Ramachandra, 1984. Improved methods of combined
forecasting. J. Forecast. 3, 197–204.

86