Emerging Economies’ Short-Term Private External Debt As Evidence of Economic Crisis

Oya Ekici

Asst. Prof. Dr.oyaekici@istanbul.edu.tr

Affiliation: Department of Business Administration,Division of Quantitative Methods,

Faculty of Political Sciences, Istanbul University

Address: Istanbul University Center Campus,34452 Beyazit / Eminonu-Istanbul

Karun Nemlioğlu

Prof. Dr.karun@istanbul.edu.tr

Affiliation: Department of Econometrics, Faculty of Economics, Istanbul University

Address: Istanbul University Center Campus, Beyazit / Eminonu-Istanbul

Corresponding author: Oya Ekici

Abstract

Considering both economic stability and crises, the need to analyze the short-term private external debt

(ST-PrED) of a country is notably obvious for proactive crisis management. In the economics literature,

the convention is to monitor the ratio of debt to the country’s Central Bank’s international reserves.

However, the debt itself could act as a main precursor. In this context, we examine Turkey’s ST-PrED

data as representative of an emerging economy. Our methodology is to use a linear growth model to fit

the ST-PrED data and a Bayesian method for model estimation and forecasting strategy. The empirical

findings illustrate the performance of our predictions in capturing unstable terms of the economy. As a

policy implication, we recommend that policy makers place special emphasis on ST-PrED as a potent

indicator of the country’s financial vulnerability. Taking into account the ST-PrED level, to prevent the

contagion effect of a crisis, it is essential to implement policies that are more effective in coordinating

international financial flows and improving liquidity positions. Furthermore, fostering structural reform

and an innovative approach to strengthen the real sector and improve education could create economic

stability and sustainability in terms of debt structure among emerging economies.

Keywords: Short term private external debt, economic crisis, dynamic linear model, linear growth model, Bayesian estimation and forecasting, rolling window analysis, debt indicators, debt related policy implications.

1. Introduction1

In 2007, the subprime mortgage crisis, which started in the US, turned into one of the gravest global

financial crises to have ever hit European and other advanced countries in the years 2008-2009. The crisis

period dragged these countries’ economies into recession. In these advanced countries, as the economic

growth slowed and the real sector shrunk its demand for loans, capital flights to nearby countries

occurred. The period, therefore, could be deemed as a downturn of economic conjuncture. In such a

period, credit booms are typically witnessed in nearby countries. Consequently, short-term private

external debt (ST-PrED) increases in these emerging economies’ markets. In the crisis that occurred

November 2008, these emerging nations eventually failed to refinance this debt. This is particularly

evident with private sector borrowing.

Meantime, The Federal Reserve (FED), and actually any advanced country’s central bank, takes steps to

reduce the interest rate, hoping that low rates would lead to monetary expansion in the economy. This

regulation can increase the amount of idle credit in the market, and as a result, the idle credit eventually

flows to emerging economies to be utilized. However, once the bubble burst in these emerging

economies, they started to suffer from this credit expansion as steps were taken to increase interest rates

in the US. This, in turn, caused monetary tightening in the emerging economies’ domestic markets and

ultimately affected the repayment ability of their debt. As is usual, the higher the debt, the more costly the

repayment. Similarly, the nominal debt value increases (see Birdal (2009) for a broad discussion,

specifically on the US mortgage crisis and its spillover effects).

The similar experiences from periods of crises in emerging economies (e.g., Argentina crisis of 1994,

Mexican Tequila crisis of 1994, Russian crisis of 1998, and Asian debt crisis of 1997) and their aftermath

reveal that these emerging countries were hit by domestic capital flight and the depreciation of their

domestic currencies. The Asian crisis was discussed by Salvatore and Campano (2010) in this context.

Along similar lines, in his study, Krugman (2010) also claimed that emerging economies are typically

damaged by the consequences of their excessive investments financed with external debt and sudden

stops or reversal. Therefore, under dynamic economic interactions, the global financial crisis of 2008-

2

2009 and other crises mentioned above have revealed that ST-PrED plays a prominent role as an indicator

of financial vulnerability and capital transfers have been the primary cause in crisis terms.

Based on its role in economic crises, conventionally, some debt related ratios have become the main tools

in crisis analysis. Typically, the proportion of external debt to reserves, to exports, or to gross domestic

products (GDP) are explored. These indicators have some disadvantages; for example, in some cases, the

ratio of debt to GDP underestimates the equilibrium debt ratio (for an expatiated discussion of debt

related indicators, see the IMF seminar notice (2000)). More broadly, beyond the dynamics of the

economy itself, political factors, social factors and technical constraints also make economic

developments unpredictable and complicated. Hence, under these uncertainties, investigating other, more

fundamental indicators simultaneously can help somewhat in getting to the bottom of the country’s

economic vulnerability. In addition to conventional debt indicators, we basically argue that the ST-PrED

series itself provides information about a pending economic crisis. Thus, there is an explicit need to

highlight this indicator.

Apparently, for developing countries, ST-PrED series shows an ever-increasing behavioral pattern and

the data behavior fits well with a linear growth model, which is, mainly a special case of the dynamic

linear model (DLM). The DLM framework supports the idea of explaining the ST-PrED with unknown

states, letting the current level change linearly through time while the growth rate changes as well.

In this context, we first present a discussion on the basis of having the ST-PrED data as a precursor of an

economic crisis. We then investigate the empirical determinants of the ST-PrED within the framework of

two forms of the DLM, the local level model and the linear growth model, referred and fitted to Turkey’s

ST-PrED series. Out of sample forecasts are used to compare the models’ performance. Findings show

that the linear growth model fits the data better. The model consistently forecasts the ST-PrED series

along the identified period. Our empirical investigations emphasize the situations when there is a change

in the growth level in the ST-PrED series that corresponds with the crisis terms in the emerging market

economies. Without denying the role of other traditional indicators, based on Turkey’s experiences, our

3

method highlights that the ST-PrED is key to simply understanding, and even foreshadowing, financial

instabilities concerning emerging economies.

2. ST-PrED as a Precursor of Economic Crises in Emerging Economies

In order to understand financial crises in East and Southeast Asia, in Salvatore and Campano’s study

(2010), debt related ratios are examined as warning indicators that are useful in predicting a crisis.

Virtually, the accumulation of the ST-PrED itself can be viewed as a proxy for more sophisticated

measures of crises in emerging economies. To be more precise, this necessitates reviewing what the

external debt is composed of. We first look at term structure of external debt. Then, we discuss the type of

debt to be examined; public or private sector. In defining term structure for external debt, short-term

implies liabilities that are due within a year and long-term implies debt lasting over a year. Fluctuation in

economic stability affects the term structure of external debt. During times of instability, borrowing

maturity shifts to the short-term. For this reason, it is of prime importance to monitor short-term debt in

the context of economic crises. Developing countries’ external debt composition shows that the private

sector plays a dominant role, and is determinative of total external debt (see Appendix A1). Based on the

data, it is not much of an overstatement to make this kind of generalization. As a partial confirmation,

according to a relatively recent World Bank International Debt Statistics report, the figures on the

summary tables, showing the composition of external stock debt and evaluations, point to the increasingly

important role of the private sector in many developing countries (World Bank, IDS 2013; see Appendix

A2). Naturally, there are some developing countries whose dominant borrower sector differs from this.

However, this still requires analysis, as there has been a significant shift in the private sector borrower

type in the last decade largely because of the privatization policies in these developing countries.

Therefore, in contrast to government, the private sector gained steam due to the trend among countries to

pursue a privatization program.

In terms of the trajectories, the ST-PrED data expose similar behaviors among countries. To represent this

conformity, several groups of the developing countries’ data are visually illustrated in Figure 1.

Expectably, the debt behavior does not hold true for low-income countries having dissimilar debt

4

systems. Generally, these economies are not strong and are much less prosperous in their private sectors;

thus, they need more public support for their economic activities. In the global financial world, especially

as financial capital flows accelerate through developing countries, almost all other developing countries’

ST-PrED increased (Kumcu, 2008). On the other hand, this is why in developing countries debt behaviors

are similar in terms of ever-increasing external debt. No doubt, we have to consider the composition and

the structure of the countries’ external debt.

Figure 1. ST-PrED Plots of developing countries (in current US dollars)

When a crisis is experienced, the ST-PrED increases, and as plotted in Figure 1, these dates become

visible as local peaks. Therefore, the issue is how to model such time series data for developing countries.

3. Material and Methods

The model choice depends on the purpose of the analysis in the time series. Here, we need to allow for

changes in the values of the parameters as the time varies. This can be achieved through the DLM. The

DLM, namely, a linear state-space model, presents: sequential observations in time and one parameter

value for each time point (Pole, West, and Harrison, 1994). In the DLM, the series are estimated with

unobservable parameters and their dependent structure. For example, if the income level of an individual

needs to be analyzed, we know that it is under the effects of some unobservable factors, such as the

individual’s intelligence, special talents, and specialty, and these factors should be represented in the

analysis (Lütkepohl, 2005). Similarly, we need to include in the model the unknown states for the ST-

PrED. The DLM enables this. It is a powerful tool that can deal with a wide variety of time series

5

problems in its many expanded forms, as seen in the literature (West, 1995; West and Harrison, 1997;

Harvey, Trimbur, and Van Dijk, 2007; Congdon, 2007). A univariate DLM can be stated with its usual

notation as follows;

y t=Ft' θt +vt v t∼N (0 ,V t )

θt=Gt θt−1+wt w t∼N ( 0 ,W t ) . (3.1)

The model is constituted through two equations as the observations and the states, with the assumption of

v t and w t , respectively, as their errors. In the model above (3.1), y t denotes the observations. F t

represents the vector of explanatory variables. θt is the state vector, which represents the parameter

vector. Gt in the state equation is the known coefficient that defines systematic changes of the state vector

against time. v t and w t are zero mean measurement errors and state innovations.

A DLM, represented in its most general form above, can be evolved into other useful forms: a local level

model and a linear growth model.

A local level model (LLM) is specified as follows, when θt=μt and F t=Gt=1;

y t=μt+v t v t∼N (0 ,V )

μt=μt−1+wt w t∼N (0 ,W ) (3.2)

where the term μt representing the level of the model is referred to as the random walk.

A linear growth model (LGM) takes the form as below, when θt=[αt

β t ] and F=[ 1 0 ], G=[1 10 1] and

η=[σ α2 0

0 σ β2 ];

y t=α t+ε t ε t∼N (0 , σ ε2 )

6

α t=αt−1+ βt−1+η1 ,t η1 , t∼N (0 , σα2 )

β t=βt−1+η2 ,t η2 , t∼N (0 , σ β2 ) . (3.3)

Here, the term of the system, η , is the covariance matrix and ε t , η1 ,t and η2 , t are uncorrelated errors.

Through incorporating β t−1, it becomes possible to represent the growth of the level of debt (Gamerman

and Lopes, 2007). The estimation and interpretation of this then becomes crucial. In our analysis, the

parameter implies a sign of economic crisis. Under this DLM framework, we estimated these two models

while y t presents ST-PrED data. After stating the general structure of the model, we can now see the data

that exemplify emerging economies’ debt behavior.

3.1. Data

The analyzed ST-PrED data are from the period of 1990-2014, amounting to 100 quarterly observations,

with the currency in millions. We consider the data of Turkey, which is representative in terms of an

emerging market economy. A preliminary plot of the data is seen in Figure 2;

Figure 2. ST-PrED of Turkey (in USD millions)

7

In addition, to investigate the conformity between the rising debt path and the crises terms, we examined

Turkey’s past economic crises along with related countries’ main economic crises. The chronology of the

economic crises of Turkey and the other countries are listed for the period explored, specifically in the

scope of relative major events, as below:

Table 1

Chronology of economic crises for Turkey’s economy together with other related countries’ economic crises

1989 - Transition to liberal economy1991 - Gulf crisis1992-1993 - European Monetary System1994 - Financial crisis (Mexico)1997-1998 - Asia and Russia crises1999 - Brazil crisis2001 - Monetary crisis (Nov. 2000 & Feb. 2001)2008 - Global financial crisis (Mortgage)2008-2010 - Automotive industry crisis2009-2014 - European debt crisis2014 - Russian financial crisis

Additionally, some political disputes or other economic developments may occur in the country such as

the debt overhang effect1 (Erbil and Salman, 2006), the FED decision to increase interest rates, and falling

commodity prices. These direct or indirect developments have an impact on debt. Whatever the triggering

events, in the crisis terms, the ST-PrED data expose the peaks. The data reveal the traces of the intended

economic plans, decisions, and unexpected developments. In general, an economy’s deep structural

problems have an effect in terms of the level of borrowing both in the long and short run, whereas

unexpected developments have an indirect effect in terms of economic exposure in the short run. As most

emerging economies borrow in foreign currency, mostly in US dollars, the ST-PrED becomes more

sensitive to dollar fluctuations in crisis terms, and a rise in the dollar punishes borrowers in emerging

markets (Economist, March 2015). Thus, the ST-PrED data reflect the effect of foreign currency

instabilities. Moreover, the ST-PrED in dollars creates added pressure on the current account deficit as

the domestic currency depreciates against the dollar.

1 The debt stock creates the “debt overhang effect.” The debt overhang indicates that the accumulated debt, acting as a tax on future output, discourages productive investment plans of the private sector (Erbil and Salman, 2006).

8

3.2. Estimation of ST-PrED

We used the Bayesian method to estimate two models (3.2) and (3.3). Since the process has no closed

form, we needed to draw from the distributions rather than an analytical solution. To evaluate the

posteriors, we computed with recursive algorithms by updating the related conditional distributions. In

stating the general algorithm, we followed the same notation used in section 3. We have the state vector

θt defined accordingly to the model, unknown parameters v t and w t for the LLM, the diagonal of the

unknown parameter matrix η=(σ α2 , σ β

2 ) and (σ ε2) for the LGM. Thus, in the DLM, with the parameters

we set before, at time t, we observe y t and generate a sample from a posterior for t = 0,1,.., T. Due to the

similarity of the algorithms, only the steps for LGM are related here: the LGM is slightly more

complicated (See Appendix A3). The Bayesian rule is given as:

P ¿

and a joint posterior distribution as:

P ¿

The method uses raw data and does not require processing the series as differencing. Moreover, even if

the available data are intrinsically limited, the Bayesian estimation still provides reliable results. We

applied the algorithm starting at time t = 0 with the full posterior P ¿ throughout the analysis the Monte

Carlo (MC) sample size is at N = 10000 and the burn-in period is 3000 (1000 for LLM). We fit the data,

respectively, to both the LLM and the LGM. The variance typically is unknown and the assigned prior

distributions for the variance parameters (V, W and σ ε2, η) in both models are inverse gamma distributions

with a high level of uncertainty. For our estimation procedure, the trace plots and ergodic mean plots

show that convergency is achieved for both model parameters, supporting the fact that the simulation

model searches appropriately in the parameter space and the chain works efficiently (see Appendix A4).

3.3. Findings

9

Two types of models are simulated and the parameters are estimated for both. Posterior estimates of the

unknown variance parameters from Markov Chain Monte Carlo (MCMC) sampler are listed in Table 2

with their MC standard errors:

Table 2

Models’ posterior estimates of variance and MC standard errors in brackets

LLMV W

0.010966 10516528(0.000112) (19545)

LGMσ ε

2 σ α2 σ β

2

0.0112 1135730 610880(0.0000927) (64229) (46423)

The parameters of both models have low standard errors. We evaluated both of the models’ forecasting

performance. The forecast results indicated that LGM gives the lower forecast error, displaying with root

means squared error (RMSE) scores in Table 3.

Table 3

RMSE of models LLM and LGM: Forecasting performance for Turkey’s ST-PrED data over 25 years 1990(Q1)-

2014(Q4).

RMSE 1 quarter ahead 4 quarters aheadLLM 13844.97 17271.59LGM 5115.035 11438.77

To assess the models’ performance and their stability over time, we used a rolling window analysis of the

ST-PrED data, and refitted each model 36 times. At first, we started sampling for the period 1990(Q1) to

2005(Q4). Then, we moved the window one quarter ahead to 2014(Q4). As we moved, each time we

forecast out of sample for one less quarter. The RMSE scores point to the LGM’s performance. Similarly,

we ran a rolling window analysis for four quarters ahead, and the RMSE, based on the average across the

four quarter forecast, again sees a better model performance for LGM (for an indication of out of sample

performance, see Appendix A5)

10

In section 2, we mentioned the importance of the growth parameter β t−1 in terms of tracking the level

changes of ST-PrED. Based on the analysis of this parameter, we have been able to identify the extreme

values that may foreshadow the extraordinary periods in an emerging economy; the values may or may

not be of a large scale. The peaks are not only corresponding to major economic-financial crisis dates. In

addition, some regulations, political decisions, elections, and related countries’ economic situations

appear to explain the behavior of the extraordinary moves during the periods. For example, in the noted

crisis of 2008, the outbreak of the financial crisis in late 2008 provides signals in the second quarter of the

year, according to the rolling estimation of β t−1. As the global financial turmoil arose in 2008, the

contagion effect of the crisis initially was positive for domestic markets of emerging economies seen as

safer harbors for investors. In Turkey’s case, looking at the lower left of Figure 3, following the β t−1

level, the estimation of the ST-PrED growth level reveals a decrease after the second quarter of 2008.

This may be explained by the impact of continuing capital inflows and the ‘good’ state of expectation

(trust in government) in the domestic market (as discussed broadly in Della Posta, 2016). However,

subsequently, after a threshold, the increasing growth parameter reveals the likely effect of a sudden stop

of capital flow. Thus, the economy could not avoid incurring a large current account deficit (that hit a

record level of 75.1 billion dollars in 2011). This implies that the global funds were not utilized in the real

sector (leading to less export volume), and instead used to finance the current account deficit of the

economy. Such an economic atmosphere raises concerns of financial vulnerability and sooner or later will

lead to an economic downturn.

11

Figure 3. ST-PrED (upper left); β t−1 parameter estimation (upper right); rolling window growth component (β t−1)

(lower left); β t−1 estimation with confidence interval (lower right)

If we extend our focus through the inspected period, we see 27 peaks from the rolling analysis of β t−1 in

the second row of Figure 3. These terms can be regarded as foreshadowing the crises along the period.

Against the backdrop of this empirical analysis, as a proxy indicator, the ST-PrED may shed light on the

process that engenders a crisis in emerging market economies, since with this design, the method we have

employed is producing consistently accurate forecasts at all levels.

4. Policy implications

For almost two decades now, emerging markets have been attractive to foreign investors. They appear to

prefer countries such as South Korea, Malaysia, Taiwan, Brazil, Mexico, and Turkey. These developing

economies have become a destination for investors who want to avoid zero interest. The capital flow to

developing countries in July 2016 alone was 24.8 billion dollars.

In Turkey, a major part of capital inflow is utilized by the banking sector. In other words, the capital

inflow is not utilized by the real sector. If the real sector of the economy is not buoyant, more

importantly, if the return of capital is not likely to contribute to the production, and specifically to the

production of high-value-added products, the money remains in the banking sector and this artificially

improves economic growth for the prevailing period. Therefore, the effects of such capital inflow will

relate to the structure of the emerging market economy. Emerging economies need to achieve significant

12

changes in their structures to meet the real sectors’ requirements over the long run. The real sector should

be stronger. Such restructuring should be in ways that raise the marginal productivity of capital and

increase the share of research and development spending in GDP (e.g., for Turkey in 2014 this ratio was

0.94%). In particular, investing in education provides a more challenging but perpetual set of

reforms/adjustments. Thus, innovations and high-value-added production would be conducive to break

through the middle-income trap faced by emerging economies. If we follow the ranking of countries in

the global innovation index, we see that for 2014, Turkey was ranked 54th, however, in 2015, it had

fallen to 58th. Thus, countries like Turkey should be more disposed and decisive in being active in

promoting innovations. Hallett et al. (2015), in their study of the European debt crisis, also highlighted

the role of structural reforms for sustainable external debt, for a robust economy.

Besides implementing structural reforms, decision makers should build a stronger liquidity position in the

country and create less strain on the foreign exchange market. In that sense, it is good for the economy to

have a flexible exchange rate regime to reduce the risk-related liquidity level and to help in preventing

current account imbalances. Looking at the ST-PrED indicator, emerging economies should initiate

regulations around capital adequacy in their banking sectors to decrease banks’ risk exposure using the

intervention of the central banks’ treasury; holding optimal level of reserves would reduce the risk.

Obstfeld et al. (2008) argue the need for sufficient reserves in emerging economies, and reserve adequacy

should be judged relative to M2 money supply. On the other hand, we know that over the last two

decades, most of the emerging economies learned from their previous crisis experiences and became

more resilient to financial instability. For this reason, the 2008 global crisis was not as severe for

emerging economies with solid monetary policy management, the formation of a stronger private sector,

and deeper domestic debt market.

Hence, in looking at the experiences in the emerging market economies, the recent global financial crises

put greater emphasis on the importance of short-term external debt as an indicator of financial

vulnerability. In fact, following the trajectory of the ST-PrED will provide information that could

alleviate a crisis. As stated in the IMF study (2000): a high debt level may suggest that a country is on an

13

unsustainable path. In other respects, it is not easy to compare debt indicators of non-homogenous

countries and there is a limited amount of empirical work on macro debt indicators as reported by the

IMF (2000). Thus, to understand the relationship between debt and instability in emerging economies,

concentrating on the analysis of the debt itself offers an alternative methodological tool. In sum, we know

that policy makers should move rapidly, and be proactive in alleviating the effects of a financial crisis.

This points to the importance of monitoring short–term debt, and being cognizant of the indispensability

of developing and implementing debt strategies consistent with long-term debt sustainability.

5. Conclusion

In the current era of financial globalization, with convenient access to information, the international

mobility of capital is an economic reality. This makes ST-PrED a prominent issue and a fundamental

dynamic in emerging economies. With this fact as our motivation, we discussed the consideration of ST-

PrED as a precursor to economic crises in emerging economies. Based on this assumption, we proposed a

DLM estimation, with the LGM fitting the ST-PrED data better. We estimated the models by recursive

Bayesian updating to develop a reliable, consistent prediction for the emerging economies. Policy makers

can apply these predictions to monitor initial signs of financial (generally economic) crises and to manage

debt. Regarding the crisis terms in Turkey as a representative emerging economy, we saw a consistency

with real economic life experiences. Once the model is estimated, the method can be used in determining

the crisis potential of the economy. In most macroeconomic-financial analysis, the efforts focus on

foreseeing the crisis. Although it may be too complicated to clearly diagnose a crisis, many potential

indicators can be followed simultaneously and evaluated. In this study, we discuss one indicator: the ST-

PrED data. Emerging economies like Turkey should also manage external vulnerabilities that arise from

external liabilities in the banking sector. Above all, in a globalized world, to create a more stable,

sustainable, and predictable economy, even if debt management is the main issue, authorities need to

tailor their policies to foster investments in structural reforms and innovation projects in education and

technology.

14

References

- Birdal M. (2009). “Bir Krizin Anatomisi: Abd Mortgage Piyasasının Kurumsal Yapısı Ve Krizin

Dinamikleri,” Econ Anadolu International Conference in Economics, Eskişehir, Türkiye, pp.1-1

- Carter, C. and Kohn, R. (1994). “On Gibbs sampling for state space models,” Biometrika, 81, 541–553.

- Congdon, P. (2007). Bayesian Statistical Modeling, West Sussex, England, John Wiley and Sons.

- Della Posta, P. (2016). “Currency and external debt crises: A unifying framework,” Journal of Policy

Modeling, Vol. 38, Issue 4, 723–736.

- Erbil, C. and Salman, F. (2006). “Revealing Turkey’s public debt burden: A transparent payments

approach,” Journal of Policy Modeling, 28(7), 825–835.

- Frühwirth-Schnatter, S. (1994). “Data Augmentation and Dynamic Linear Models,” Journal of Time

Series Analysis, Vol.15.

- Gamerman, D. and Lopes, H. (2007). Markov Chain Monte Carlo: Stochastic: Simulation for

Bayesian Inference, Chapman&Hall/CRC.

- Hallett, A.H. and Oliva, J.C.M. (2015). “The importance of trade and capital imbalances in the

European debt crisis,” Journal of Policy Modeling, Elselvier, Vol. 37, Issue 2, 229–252.

- Harvey, A. C., Trimbur, T. M. and Van Dijk, H. K. (2007). “Trends and cycles in economic time series:

A Bayesian approach,” Journal of Econometrics, Elselvier, Vol. 140.

- IMF, 2000, March 23, “Debt- and Reserve-Related Indicators of External Vulnerability” Prepared by

the Policy Development and Review Department in consultation with other Departments,

https://www.imf.org/external/np/pdr/debtres/debtres.pdf

- Krugman, P. (2010). “Crises: The Next Generation,” /sapir_conferences /Krugman.pdf. Last accessed:

01.12.2015.

15

- Kumcu E., Hurriyet Archive, (7 Sept 2008). Available at the website: http://www.hurriyet.com.tr/ozel-

sektorun-dis-borclanmasi-9842702. Last accessed: 01.12.2015.

- Lütkepohl, H. (2005). New Introduction to Multiple Time Series Analysis, Berlin, Springer Verlag.

- Obstfeld M, Shambaugh J. C. and Taylor A. M. (2008). “Financial stability, the trilemma and

international reserves,” NBER Working Paper, no 14217.

- Petris, G., Petrone S. and Campagnoli P. (2009). Dynamic Linear Models with R, New York, Springer.

- Pole, A., West M. and Harrison, J. (1994). Applied Bayesian Forecasting and Time Series Analysis,

Chapman and Hall.

- Salvatore, D. and Campano, F. (2010). “The financial crisis in East Asia—Then and now,” East Asia

Law Review, 121–133.

- West, M. (1995). “Bayesian Inference In Cyclical Component Dynamic Linear Models,” Jasa, Vol.90,

No.432.

- West, M. and Harrison, J. (1997). Bayesian Forecasting and Dynamic Models, New York, Springer,

1997.

- ‘dlm’ R package used for estimations (Author Giovanni Petris, Bayesian and Likelihood Analysis of

Dynamic Linear Models, 2009)

- The Economist, 21 March 2015, (http://www.economist.com/news/leaders/21646749-rise-dollar-will-

punish-borrowers-emerging-markets-mismatch-point?

fb_action_ids=10153257372911037&fb_action_types=og.likes&fb_ref=scn%2Ffb_ec

%2Fmismatch_point)

Acknowledgement

Computations were implemented using R DLM (Petris, 2009). The authors thank Murat Birdal for his

constructive contributions.

16

For Data

- Country Groups’ External Debt Data: World Development Indicators (WDI) and Global Development

Finance (GDF) under the World Bank's Debtor Reporting System (DRS).

http://databank.worldbank.org/ddp/home.do

- Turkey’s External Debt Data: Under Central Bank of The Republic of Turkey (TCMB) Database,

http://evds.tcmb.gov.tr/cbt.html

Appendices

A1. Figure

17

A2. Figure: Composition of Total External Debt Stocks for 2011

18

19

20

A3. Algorithm

The general algorithm for the estimation of LGM, for N number of the Monte Carlo sample size; i) Select an initial value for η=η0, for each j=1,…, N, ii) Sample the state θt

j from P ¿ with FFBS2, iii) Sample η j from P ¿, iv) Repeat steps ii and iii N-1 times to produce a final posterior.

A4. Figure

Figure A4. Plots for convergence assessment of LGM and LLM, respectively: Trace Plots (in the first row), Ergodic Mean Plots (in the second row) of simulated observation and equation variances. A5. Figure2 Forward Filtering Backward Sampling (a sampling method introduced by Carter and Kohn, 1994, and Frühwirth-Schnatter, 1994)

21

To see out of sample performance, we show a plot of 36 steps-ahead forecasts that are computed at the

end of one time sampling in Figure A5. Although there are long forecasting steps moving from the base

period, the forecasted means appear similar to the original data. Confidence bands are a bit narrow.

However, we can say that this results from the strong priors setting.

Figure A5. ST-PrED Data fitted to LGM with 36 steps ahead forecasted values confidence interval (left) - ST-PrED data

fitted to LGM with one-step ahead forecasted values confidence band (right)

Hereby, we visually compare observed and forecasted trajectories on the right of Figure A5 as a final

generic display. This plot shows us that the Bayesian estimation of the LGM is successful in forecasting

the ST-PrED series and indirectly capturing path and crises terms, since in the crises terms the peaks can

be seen on a smoothed line of the Bayesian estimation as well. As a second order DLM, LGM is

appropriate for the data in terms of having a parameter that represents the growth of the series.

22