1 WP/ 9 /2014 Working Paper FINANCIAL CYCLE OF INDONESIA – POTENTIAL FORWARD LOOKING ANALYSIS Cicilia A. Harun, Aditya Anta Taruna, R. Renanda Nattan, Ndari Surjaningsih Desember, 2014 Kesimpulan, pendapat, dan pandangan yang disampaikan oleh penulis dalam paper ini merupakan kesimpulan, pendapat dan pandangan penulis dan bukan merupakan kesimpulan, pendapat dan pandangan resmi Bank Indonesia.
44
Embed
FINANCIAL CYCLE OF INDONESIA POTENTIAL ......1 Financial Cycle of Indonesia – Potential Forward Looking Analysis* Cicilia A. Harun†, Aditya Anta Taruna‡, R. Renanda Nattan ,
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
WP/ 9 /2014
Working Paper
FINANCIAL CYCLE OF INDONESIA –
POTENTIAL FORWARD LOOKING ANALYSIS
Cicilia A. Harun, Aditya Anta Taruna, R. Renanda Nattan,
Ndari Surjaningsih
Desember, 2014
Kesimpulan, pendapat, dan pandangan yang disampaikan oleh penulis dalam
paper ini merupakan kesimpulan, pendapat dan pandangan penulis dan bukan
merupakan kesimpulan, pendapat dan pandangan resmi Bank Indonesia.
1
Financial Cycle of Indonesia – Potential Forward Looking
Analysis*
Cicilia A. Harun†, Aditya Anta Taruna‡, R. Renanda Nattan§,
Ndari Surjaningsih**
Abstrak
Kebutuhan akan referensi yang baik dalam rangka implementasi
peraturan countercyclical menjadi alasan utama dari penelitian untuk
mengkonstruksi siklus keuangan. Penelitian ini merupakan penelitian
lanjutan Alamsyah et al (2014) yang telah menghasilkan siklus keuangan
Indoneisa dengan mengikuti tata cara pembuatan yang dilakukan oleh
Drehman et al (2012). Siklus keuangan pada penelitian ini akan
diperbaharui dengan penggunaan harga aset dan cara pengolahan data
lanjutan. Untuk meningkatkan kepercayaan dalam penggunaan siklus
keuangan sebagai referensi kebijakan di masa yang akan datang,
penelitan ini akan melakukan forecasting. Hasil dari forecasting
menunjukan bahwa siklus keuangan cukup robust dan cenderung untuk
mengikuti pola masa lalu. Hal ini memberikan dorongan untuk penelitian
terkait karakteristik dari siklus keuangan dan indikator tambahan sebagai
referensi untuk dapat menangkap kemungkinan terjadinya perubahan
struktural di masa yang akan datang.
Keywords: Financial cycle, countercyclical capital buffer,
financial crisis.
JEL Classification: G1, G2, F3
* Pendapat dan kesimpulan dalam paper ini merupakan pendapat penulis dan bukan
merupakan pendapat resmi dari Bank Indonesia. Penulis mengucapkan terima kasih kepada Dadang Muljawan, peneliti ekonomi senior, Departemen Kebijakan
Makroprudensial Bank Indonesia atas kontribusinya dalam memberikan metodologi untuk menganalisa siklus untuk keperluan forecasting. Tabel dan grafik merupakan
hasil pengolahan oleh penulis, terkecuali jika dinyatakan berbeda. † Peneliti Ekonomi Senior, Departemen Kebijakan Makroprudensial, Bank Indonesia,
email: [email protected] ‡ Peneliti Ekonomi, Departemen Kebijakan Makroprudensial, Bank Indonesia, email:
[email protected] § Research Fellow, Departemen Kebijakan Makroprudensial, Bank Indonesia, email:
Financial Cycle of Indonesia – Potential Forward Looking
Analysis††
Cicilia A. Harun‡‡, Aditya Anta Taruna§§, R. Renanda Nattan***,
Ndari Surjaningsih†††
Abstract
The need to have a good reference for implementing countercyclical
measure has been the motive behind research for constructing the
financial cycle. The paper carries over the result done in Alamsyah et al
(2014) that constructed the financial cycle of Indonesia following the steps
done in Drehman et al (2012). The cycle is improved with the inclusion of
asset price and more advanced data treatment. In order to have better
confidence in using the cycle for a reference toward the policy that will be
implemented into the future, the paper also exercises forecasting. The
forecasting result shows that the cycle is quite robust and tends to be
persistently following the pattern formed from the history. This suggests
careful study toward the characteristics of the cycle and additional
indicators as references in order to capture the possibility of structural
break in the future.
Keywords: Financial cycle, countercyclical capital buffer, financial
crisis.
JEL Classification: G1, G2, F3
†† The opinions and conclusions written in this paper are of the authors and do not reflect
the stance of Bank Indonesia. Authors are grateful for the contribution of Dadang
Muljawan, Senior Economic Researcher of Macroprudential Policy Department, Bank
Indonesia for suggesting the methodologies for analyzing the cycles for forecasting
exercise. Tables and figures are authors’ calculations unless stated differently. ‡‡ Senior Economic Researcher, Macroprudential Policy Department, Bank Indonesia,
email: [email protected] §§ Economic Researcher, Macroprudential Policy Department, Bank Indonesia, email:
[email protected] *** Research Fellow, Macroprudential Policy Department, Bank Indonesia, email:
[email protected] ††† Senior Economic Researcher, Macroprudential Policy Department, Bank Indonesia,
and a set of thresholds to signal banking distress before it materialized as banking
crisis. This combination turned out to be performing quite well as an early warning
exercise for banking crisis.1
Drehman et al (2012) provided a seminal paper on the construction of
‘financial cycle’. The paper delivered a financial cycle that was considered best to
represent the definition in Borio (2012) that is the self-reinforcing interactions
between perceptions of value and risk, attitudes towards risk and financing
constraints, which translate into booms followed by busts. This definition is also
close to the very definition of procyclicality. The length of a full financial cycle is
longer than a full business cycle. It is very likely that a financial cycle pass through
more than one business cycle. This is considered more realistic since the frequency
of financial crises is smaller than the frequency of booms in the economy. Drehman
et al (2012) found that fort the U.S. the length of business cycles is 1 to 8 years, while
it is 8 to 30 years for financial cycles. The peaks of financial cycles are associated
with the events of financial crises. This makes financial cycle one of the important
references for determining the timing of setting the CCB.
1 Early Warning Exercise requires the indicator used to identify the risk of crisis with a lead sufficient to allow the authorities to take remedial actions.
5
Alamsyah et al (2014) has constructed the financial cycle for Indonesia using
narrow and broad credit indicators (credit-to-GDP and credit growth). The paper
found the length of financial cycles in Indonesia (9 to 10 years) to be double the
length of the business cycles. The financial cycle using broad credit follows the
financial cycle using narrow credit. The paper also found that the amplitude of the
financial cycle is smaller after the 1997-1998 crisis. This last conclusion fits the fifth
empirical finding in Drehman et al (2012) that came out of the view in Borio & Lowe
(2002) that the amplitude, length and potential disruptive force of the financial cycle
are closely related to the financial, and possibly also monetary, regimes in place.
Indonesian financial system has significantly evolved after the East Asian crisis in
1998 after going through banking restructuring program and economic reform.
This paper is dedicated mainly for two purposes: 1) to enhance the
construction of financial cycle in Alamsyah et al (2014); and 2) to exercise a forward
looking analysis in order to increase the confidence of using the financial cycle as an
early warning exercise and reference for CCB mechanism. The enhancement of the
cycle construction involve the inclusion of structural breaks treatments to the
constructing indicators as well as the inclusion of the asset price indicators as
Drehman et al (2012) did suggest the minimum set of indicators to be indicators on
credit and asset price. The methodology for constructing the cycle follows closely
Drehman et al (2012), with some country-specific considerations and modification in
the weighting to differentiate the contribution of each indicator into the financial
cycle. The result of the financial cycle is not significantly different from the cycle
generated in Alamsyah et al (2014). The timing of both cycles is similar. The
difference comes in the amplitude that can be caused by the differences of the base
year of the normalization of data. The similarity is actually by construction since the
cycle downplays the influence of asset prices as we decided that credit indicators
should play more role in determining the financial cycle as the banking system still
dominates the Indonesian financial system. The decision is also backed up by the
fact that the asset price indicators included here usually influence the financial cycle
in a higher frequency domain, so that it is likely to be truncated from the cycle as it
is focused on the medium frequency domain.
The second objective of this paper is to provide forecasting exercise in order to
increase the confidence of the macroprudential authority in using the financial cycle
as one of the reference to set the CCB. The construction of the financial cycle is such
that an additional point of observation will alter the entire series of cycles. The
forecasting exercise provides a predictive power to the cycle in order to increase the
6
number of observations into the future so that we can reconstruct the cycle using
the new points and therefore provide more information to decide on the CCB setting.
The rest of the paper will be arranged as the following. Chapter 2 will be about
the construction of the financial cycle emphasizing on the differences done in this
paper to enhance the result in Alamsyah et al (2014). The theoretical background of
the determinants of the financial cycle will be discussed in Chapter 3. Chapter 4 is
the forecasting exercise. Finally, Chapter 5 concludes.
7
II. CYCLE CONSTRUCTION
The financial cycle is constructed from individual indicators as suggested in
Drehman et al (2012), which will be treated in order to be used as input to filtering
mechanism. The results of the filtering of all the indicators will be combined to
create a common cycle, which then will be called the financial cycle. As it is done in
Alamsyah et al (2014), there will be two filters discussed in this paper: the
frequency-based filter (FBF) and the turning point analysis (TPA), and therefore
there will be two financial cycles produced using different methods of combining.
However, the two financial cycles should be used to reconfirm each other instead of
contradicting each other.
FBF produces a cycle from which peaks and troughs can be identified. The
identification can be made through visual judgment when plotting the cycle or
through a computer program. On the other hand, TPA only presents position of
peaks and troughs. Both FBF and TPA are able to produce short and medium term
cycle. FBF and TPA are explained in more details bellow.
Frequency-Based Filter
The main idea of this analysis is to isolate a specific range of frequency of
macroeconomics data. FBF analysis makes use a band pass filter which is a
combination of high and low pass filter. Data is first changed from time domain to
frequency domain using Fourier Transformation then the filter process takes place,
passing only frequency higher than the low frequency threshold and lower than the
high frequency threshold.
Frequency threshold means the intended cycle length, with higher frequency
corresponds to lower threshold in time domain and vice versa. According to Comin
and Gertler (2003), who studied the behavior of medium-term macroeconomics for
the US economy, a band pass filter with duration of 5 to 32 quarters is used to isolate
a short-term cycle, which is popularly known as a business cycle. The duration of 32
to 120 quarters is used to isolate a medium-term cycle. Due to the availability of
economics data in Indonesia, the duration of a medium-term cycle is adjusted to 32-
80 quarters2.
The band pass filter employed here is suggested by Christiano and Fitzgerald
(1999) and the data filter is in annual growth rate. Under the assumption that the
2 This is also done in Alamsyah et al 2014.
8
growth rates of macroeconomics series are stationary, the filter thus implies zero
trend (or drift). The frequency-based filter analysis in this paper is done using Eviews.
Turning Point Analysis
TPA yields peaks and troughs of a cycle trough Bry-Boschan (BB) Algorithm.
BB algorithm first identifies potential peaks and troughs which are higher and lower
respectively compared to their surroundings. Then potential peaks and troughs will
be subjected to various tests before final peaks and troughs are established.
In the first step, a potential peak is identified at time t if it obeys the rule (yt −
y(t−i)) > 0, with 𝑖 = (−2, −1, 0, 1, 2) for short term cycle while 𝑖 =
(−4, −3, −2, −1, 0, 1, 2,3,4) for medium term. Similarly, a potential trough occurs at time
t if it obeys the rule (yt − y(t−i)) < 0 with 𝑖 = (−2, −1, 0, 1, 2) for short term cycle while
𝑖 = (−4, −3, −2, −1, 0, 1, 2,3,4) for medium term.
Potential peaks and troughs will then be examined under censoring rules.
Censoring rules ensure that length of a phase (from peak to trough and vice versa)
and a cycle (from peak to peak or from trough to trough) meet the minimum
requirement. For short term cycle, the minimum length for a phase is 2Q and a cycle
is 5Q while for medium term cycle, the minimum length for a phase is 9Q and a cycle
is 20Q. Peaks and troughs resulted from turning point analysis will not change
though new data is added unlike frequency-based filter analysis. Data addition leads
to frequency addition thus alters the output in frequency domain.
Base Year
Normalization forces data to have normal distribution with zero mean and standard
deviation at time data used as base year.
𝐼 =𝑥𝑡 − �̅�
𝜎
In term of index range we can rewrite normalization formula as function of maximum
and minimum, so the formula will evolve to:
𝐼 =𝑥𝑡 − 𝑥𝑚𝑖𝑛
𝑥𝑚𝑎𝑥 − 𝑥𝑚𝑖𝑛
In the case we are going to compare to only a certain time in data, base year
commonly only uses one time as base year then 𝑥𝑚𝑖𝑛 equals to zero, which comply to
normal distribution function perquisite, and 𝜎 = 𝑥𝑚𝑎𝑥 − 𝑥𝑚𝑖𝑛 equals to the data value
at time used as base year. Further, the standard deviation in normal distribution (𝜎)
9
can be calculated as 𝜎 = 𝑥𝑚𝑎𝑥 (where 𝑥𝑚𝑖𝑛 = 0). Applying all mathematical
manipulation into normalization, base year can be calculated using:
𝐼 =𝑥𝑡
𝑥
�̅� = 𝑥𝑚𝑖𝑛 = 0
The mathematical expression shows that base year method forces data to have
zero mean at time used as base year and standard deviation at time used as base
year. Judging the philosophy of base year, forcing data has to change “real mean” to
“normal distribution mean” (zero), we have to be sure that a significant alternation
to mean happen in the data before determine which time used as base.
2.1. Data, Indicators and Treatments
According to Aikman et al (2010) financial cycle can be illustrated from credit
cycle composed only by credit. While Minsky (1982), Kindleberger (2000), and
Claessens et al (2011) suggested financial cycle to be represented by the combination
of property prices and credit. Drehman et al (2012) and Borio (2012) suggested the
minimum indicators used in a financial cycle are credit representing funding risk
and asset price representing price and risk perception. Drehman et al (2012) used
five financial variables: (i) credit to private, non-financial sector, (ii) the ratio of credit
to GDP, (iii) equity prices, (iv) residential property prices, (v) an index of aggregate
asset prices. Referencing to the study, financial cycle in Indonesia will be composed
by those variables yet certain adjustments are to be made due to data availability.
The inclusion of asset price indicators in this paper is the first enhancement from
the construction of financial cycle done in Alamsyah et al (2014).
The indicators used to represent the financial cycle in Indonesia are broad
credit (BC), ratio of broad credit to GDP (BC/GDP), Jakarta Composite Index (JCI),
and Jakarta Property Index (JAKPROP). The definition of BC follows Alamsyah et al
(2012). JCI constitutes equity prices while JAKPROP proxies residential property
prices. Indonesia residential property prices use more than one base year with
different number of cities surveyed thus converting the data to one common base
year is not possible. JAKPROP represents the prices of the stocks of the companies
in property sector, which is considered a good proxy for the movement of property
price in Indonesia. Business cycle is commonly represented by GDP. The table below
summarizes variables used for financial and business cycle.
10
Tabel 1. The Constructing Indicators of the Financial Cycle
Source : Bank Indonesia, Bloomberg, OJK
Data are recorded quarterly and available from 1993Q1 until 2014Q1. Broad
credit is preferred to narrow credit because government foreign debt and outstanding
corporate bond are major sources for credit in Indonesia3. Banking credit to GDP is
varying around 30% in Indonesia, which strengthens the reason to use Broad Credit
indicator in this case. JAKPROP is the composite stock price of listed property
companies in Indonesia introduced in 1996. Data for JAKPROP with a number of
companies in property sector from 1993Q2 to 1995Q4 is calculated using the formula
below
𝑗𝑎𝑘𝑝𝑟𝑜𝑝𝑡 = ∑𝑝𝑡𝑖 × 𝑐𝑎𝑝𝑖𝑡𝑎𝑙𝑖𝑠𝑎𝑡𝑖𝑜𝑛𝑡𝑖
𝑡𝑜𝑡𝑎𝑙 𝑚𝑎𝑟𝑘𝑒𝑡 𝑐𝑎𝑝𝑖𝑡𝑎𝑙𝑖𝑠𝑎𝑡𝑖𝑜𝑛𝑡
𝑛
𝑖=1
with 𝑝𝑡𝑖 is the stock price of property company i at t. However, the formula above
cannot be used to calculate for JAKPROP in 1993Q1 since the raw data is not
available. The data for JAKPROP in 1993Q2-1993Q4 is constructed by extending the
JAKPROP using all the stocks of property companies.
Structural break analysis
All data values needs to be normalized using a base year of a point in time to
ensure comparability of the units. Drehman et al (2012) used 1985 Q1 as the point
of reference to normalize his data since 1985 Q1 is the financial liberalization in the
western world. The point of reference is determined such that it is the point in time
which data characteristics is altered, called a structural break. In Indonesia case,
the point of time for each indicator is detected using both Quandt-Andrew Test and
Chow Test for possible structural breaks as shown in a table below.
Table 2. Structural Break Candidates
3 Alamsyah et al 2014 provides a discussion on the comparison of using narrow credit and broad credit data for Indonesia case.
Variables Details Source
Broad Credit Nominal is sum of:
1. Narrow Credit Bank Indonesia (SPI)
2. Government Foreign Debt Bank Indonesia (DSTA)
3. Outstanding Corporate Bond CEIC-OJK
Broad Credit Nominal same as above
GDP Nominal Bank Indonesia (PPDI)
JCI Jakarta Composite Index Bloomberg
Jakprop Jakarta Propery Index Bloomberg
GDP Real Bank Indonesia (PPDI)
BC
BC / GDP
11
* significance at 0% ≤ 𝛼 ≤ 1%
** significance at 1% < 𝛼 ≤ 5%
*** significance at 5% < 𝛼 ≤ 10%
Despite being proved to be one of the structural breaks in JCI and JAKPROP,
2009 Q2 will not be used as one of the base years used because the levels of
significance of 2009Q2 in Quandt-Andrew Test and Chow Test are higher or equal to
10%. Authors decided to use four potential structural breaks: 1) 1998 Q3; 2) 2004
Q2; 3) 2011 Q1; and 4) 2007 Q4. All of the structural breaks will be used as the point
of reference to normalize all the data. Comparisons and observations will be made to
judge which of the individual structural breaks will be used. The use of structural
break analysis also follows the suggestion of Drehman & Tsatsaronis (2014) about
the use of the analysis for Indonesia data.
The data of every indicator has to go through series of treatments before it is
analyzed using both frequency-based filter and turning point analysis. The various
treatments are listed below, however not necessarily applied to all indicators used in
constructing the financial cycle.
1. Seasonal Adjustment (SA): SA is applied on data level of all variables using
Eviews.
2. Logarithm (log): Log is applied to all variables except for ratio variables.
3. Normalization: The point of time used to pivot the data is one of the structural
dates of the variables.
4. Taking the growth
Data input to both band pass filter and turning point analysis is growth data.
Should the data have been in log, annual growth can be approximated by differencing
four quarterly data. On the other hand, common growth formula is applied. Various
sets of treatments can be arranged from the above list and choosing the most suitable
series of data treatment is crucial in capturing the natural characteristics of the data
12
and more importantly should not create pseudo or unnatural characteristics to the
data. There are two sets of series of data treatments that have met the criteria
mentioned earlier.
First procedure: Data Level SA Log Normalization Differencing
(annualized)
Second procedure: Data level Growth Normalization
In Alamsyah et al (2014), the first procedure is used to process the data.
However, in this paper the second procedure is employed in processing the data as
we believe that data in growth do not have to be converted into log as they were
assumed to be stationary as the treatment used by Drehman et al (2012) and Comin
& Gertler (2003). Bearing in mind that the purpose of producing financial cycle is to
capture the perception of the people about the economy, applying SA will only
eliminate the seasonal outlier which can mean the people’ response on occasions.
Thus in this paper, SA will not be applied.
After normalization, data is ready for input to both frequency-based filter and
turning point analysis, as illustrated below. Under FBF, output of band pass filter
will be processed under BB algorithm for consistency checking. Output of BB
algorithm under this analysis will not be the same as under TPA. FBF produces a
cycle with peaks and troughs while TPA can only deliver peaks and troughs.
Frequency-based analysis: data (normalization) band pass filter cycle
Turning point analysis: data (normalization) bry boschan peaks and
troughs
Concordance Index (CI)
A selection among variables is needed to decide which variables will be used
to compose the financial cycle. Variables that do not co-move with the most potential
variable will cancel out the potential peaks and troughs of the financial cycle while
variables that co-move will reinforce the potential peaks and troughs of the financial
cycle. An index from Harding and Pagan (2006) called concordance index can
measure the co-moving degree of a variable toward another variable. This index does
not only measure the linearity of two variables but also the cyclicality thus it is
completely different from correlation. The index has a range value of 0% to 100%
with increasing index indicating better co-movement between two variables.
Before CI between two variables can be calculated, each variable has to
undergo FBF or TPA to obtain peaks and troughs of its cycle. An expansion phase is
defined to be an area ranging from after a trough to a peak and a contraction is an
13
area starting from after a peak until a trough. The concordance index, 𝐶𝐼𝑥,𝑦 between