Measuring financial conditions for Argentina

(1)

Measuring Financial Conditions for

Argentina

Andres Rabinovich

August 10, 2016

1

st

Supervisor: dr. W.E. Romp

2

nd

Supervisor: dr. C.A. Stoltenberg

Master in Economics

Monetary Policy, Banking and Regulation

Abstract: A series of financial conditions indices for Argentina are built from a set of 21 variables, through a Principal Components Analysis. The financial conditions indices are used to forecast GDP. The resulting forecast errors are compared with those of forecasts performed by single indicators. Em-pirical results show that these compound financial indices can help predicting GDP, reducing errors up to 4% on average with respect to single indicators.

Keywords: Economics, Argentina, Financial conditions index, GDP fore-casting, Principal components analysis.

(2)

Statement of Originality

This document is written by Andres Rabinovich who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of com-pletion of the work, not for the contents.

(3)

Acknowledgments

My special thanks to my supervisors, especially to dr. W.E. Romp with whom I in-teracted the most while writing this document. With his support, encouragement and challenges he has been a key factor for finishing this thesis.

Many thanks to my parents Delia and Daniel, who always backed me in everything I did but particularly in education matters. To my girlfriend Natal´ı, who let me go to Amsterdam for one year, and for her permanent backing and help regarding English language skills. To my family. To my old friends from Argentina and my new friends from Amsterdam.

Finally, I want to thank Alejandra and Robin for their useful advice and suggestions before handing in this final document.

(4)

1. Introduction

Due to the innovations in information and communication technologies and the global-ization effects, national economies are every day more exposed to suffer financial turbu-lence.

Therefore, it is crucial to have tools for evaluating the financial conditions of an economy in order to know more about its structural financial behavior and its sensitivity towards both, national and international shocks in real and financial variables.

A Financial Conditions Index (FCI) is a useful tool which has been mostly studied for developed economies after the outbreak of the financial crisis in 2007. Additionally, it is of particular interest to analyze how an FCI can fit a small Latin American economy and if its trends are useful to predict the future evolution of economic activity.

Along the present document, a series of FCIs are constructed, to determine if they could be better predictors of GDP growth compared to single indicators in Argentina. As it is demonstrated in this thesis, results are positive and conclusive, showing that forecast errors are successfully reduced when considering FCIs for Argentina.

To calculate FCIs, a dataset was built compiling 21 variables from relevant national and international sources. Among them, a political variable is considered to reflect the high volatility in Argentina’s politics and its effects on financial conditions. The methodology is based on a Principal Components Analysis (PCA) that allows reducing large datasets into a small number of common factors that capture the underlying common structure of the variables.

The thesis is structured as follows: in Section 2, existing literature is analyzed. In Section 3 some economic and historic facts about Argentina are presented for a better understanding of financial conditions in this country. In Section 4 the variables that con-form the FCIs are described as well as the required transcon-formations to make them useful. In Section 5 the methodology is explained in depth and in Section 6 the methodology previously described is used for empirical analysis. Finally, Section 7 concludes.

(6)

2. Literature review

FCIs date back to the study of financial cycles and the belief that they influence the real economy. Even though business cycles overshadowed financial cycles during the post-World War II period (Borio, 2014), the last twenty years turned it back, especially after the series of bubbles and bursts that hacked the financial markets and the whole economy world wide. In this section, the origins of FCIs are described, showing leading and regional experiences.

2.1. Development of FCIs

FCIs found their origin on the Monetary Conditions Indices (MCI) that were firstly developed by the Central Bank of Canada in the early 1990s (Freedman, 1996). An MCI is an index that takes into account interest rates and foreign exchange rate to asses how tight or loose monetary policy is. The idea behind this measure is that interest rates and foreign exchange rates are sensible indicators that can reflect exogenous changes instantly, fitting different transmission channels together to guide policy makers’ decisions.

For more than ten years, MCIs were widely developed based on three main character-istics (Xiong, 2012). First, they show domestic but also foreign effects on the monetary conditions of a country. Second, they provide useful information to policy makers, re-searchers, and public in general. Third, they were leading indicators being able to asses monetary policy decisions timely to offset negative effects.

However, MCIs were largely discontinued as many shortcomings appeared (Osborne-Kinch et al., 2010), among others their imprecision in the short-term, the tendency to misguide markets and the distortion over monetary policy when used as operational tar-gets. The first institution to implement an MCI, The Bank of Canada, stopped calculating it by 2006. Instead, FCIs appeared as an improved tool, incorporating additional vari-ables that account for more complex interactions between monetary policy instruments and macroeconomic indicators.

The concept “financial conditions” refers to the state of financial variables that influence economic behavior and the future state of the economy (Hatzius et al., 2010). Among those variables, there are indicators that affect the financial instruments that determine the economic activity. Such variables include stocks and flows related to interest rates,

(7)

asset prices, exchange rates and volatilities among others.

As defined by Hatzius et al. (2010), the aim of an FCI is to summarize the information about the future state of the economy contained in current financial variables. In other words, it is a tool tentatively capable of predicting macroeconomic indicators like economic activity or inflation through analyzing financial shocks, which is the premise to be proven in the present work.

There are other tools also related to FCIs but with different objectives. Among them it is important to mention Financial Stress Indices mainly developed after the last international financial crisis, providing an early warning indicator especially for policy makers (Hakkio and Keeton, 2009).

2.2. Leading experiences

FCIs have been studied by innumerable international organizations, central banks, in-vestments banks and academics. In this subsection, some of the most representative works will be described based on how techniques were developed through time, following Dar-racq Paries et al. (2014).

2.2.1. Simple weighted averages

The first method is based on a weighted average of the variables considered. The calcu-lation of this kind of FCIs relies on the availability of data. Although these FCIs can sum-marize any number of indicators, they are generally used for less than ten (Darracq Paries et al., 2014).

The main difference between this kind of indices is how they assign the weights. The Bloomberg United States Financial Conditions Index (Bloomberg) for example, is based on ten variables for the United States since 1991 on a daily basis. To aggregate the data, the variables are distributed into three groups. A sub-index is created for each of the groups as a simple average, and then the final index is calculated as the simple average of the three sub-indices. Equation 1 shows the standardized value (zi,t) of the i-th variable (xi,t)

in time t, where ¯xi is the average of variable xi. Equations 2 and 3 show the procedures

applied for calculating the sub-indices and the FCI, where j is each of the n variables per sub-index SIi and i represents each of the sub-indices. All indicators xi are standardized

(8)

as shown in Equation 1 and the final FCI (Equation 3) is also standardized. zi,t = xi,t− ¯x σ (1) SIi= n X j=1 zi n (2) F CI = 3 X i=1 SIi 3 (3)

Another way of weighting the variables is through a variance-weighted average. Some examples are the Global Index of Financial Turbulence (ECB, 2009), based on six variables considered for 29 OECD countries since 1994 on a monthly basis and the Financial Stress Indicator (IMF, 2008), based on seven variables considered for 17 developed countries since 1980 on a monthly basis. This technique assigns weights wi =

1 σ2

i

for every indicator xi

with variance σ2_i and then aggregates the data and standardize it (Equation 1). Equation 4 shows the aggregation procedure. In this way, the FCI accounts for the value of variables relative to the impact of their variability.

F CI =

n

X

i=1

xiwi (4)

While this method is technically simple and fast to implement, it relies on the assump-tion of independent variables, which is not true and can cause a double counting when variables are correlated with each other.

2.2.2. VAR models

The second method to be discussed analyses the impact of shocks in the indicators on GDP growth (and/or other macro variables like inflation) to determine each of the relative weights. Because it requires estimating the effects of shocks, it is convenient to work with less than eight variables to obtain robust weights (Darracq Paries et al., 2014).

Goodhart and Hofmann (2000) use a small vector autoregressive model (VAR) to deter-mine the weights of financial variables. Equation 5 shows a simple IS curve and embeds the impact of present and past values of aggregate demand (yt) and of each of the eight

(9)

coef-ficients λj represent the weights of each of the financial variables, for each of the countries.

The series of indicators are innovative in the sense that they include asset prices to reflect changes in wealth through prices in housing and stocks. They are demeaned before the computation and are seasonally adjusted if necessary. Goodhart and Hofmann (2000) cal-culated this for 17 developed countries between 1973 and 1998. Their dataset has missing information for some of the countries but they complete the dataset assuming that a miss-ing value from one country can be proxied by another country with a similar economic background. yt+1 = m X i=0 γiyt−i+ 8 X j=1 λjxj,t+ ηt (5)

Another approach is the one by Swiston (2008), who implements a two-steps technique. First, he applies a similar VAR model to that of Goodhart and Hofmann (2000). However, he constructs impulse response functions to asses the impact of financial shocks on GDP growth. This procedure is justified by the benefits of accounting for the timing in which shocks take place and the impact across variables. Equation 6 captures the effect of shocks on one or more variables (vector of variables X) over the rest of the variables.

Xt+1= m

X

i=0

AiXt−i+ vt (6)

It is important to note that this method relies on the strong assumptions that variables follow a linear relationship and that the errors vt are independent. In reality, financial

indicators show a co-movement and are correlated with each other making this methodol-ogy susceptible of large deviations. Despite that, it is very useful to test or simulate the impact of new policies making it possible to quantify the effects produced.

2.2.3. Principal components analysis

The third and last methodology considered is the principal components analysis. This is a statistical procedure based on capturing the co-movement of the variables to explain them in terms of common factors ranked according to the variance explained. The covari-ance matrix is calculated to find clusters of variables with high correlation which represent those common factors and are ranked according to the eigenvalues, obtaining the principal components. The number of principal components used depends on the approach of the

(10)

author and the explained variance intended to be considered.

One of the main virtues of this technique is that it allows datasets with a large num-ber of variables but reduce their dimensionality to a relatively low numnum-ber of principal components chosen, accounting for common movements within correlated indicators.1

Angelopoulou et al. (2014) developed an FCI for the Euro area and for selected coun-tries2 for the period 2003-2010. They consider a dataset of 25 variables including prices, quantities, risk premia, volatility, survey data and monetary policy rules. All the variables are flows, so they do not need to be transformed as they are already in first differences. Variables are only normalized and changed sign when an increase in their value do not represent an improvement in financial conditions. PCA is applied setting to 70% the threshold for the share of total variance explained by the principal components consid-ered. This threshold is achieved by taking the three highest eigenvalues for the Euro area FCI. The final index is built as shown in Equation 7 where the principal components (P C) are weighted by their explained variance (γi).

F CI =

n

X

i=1

P Ciγi (7)

In a similar way, Hatzius et al. (2010) construct an FCI based on PCA but with some peculiarities. First, they include a broader number of variables, 45 in total, which are divided into five subindices also including survey data. Second, they allow for missing data for some periods and variables. Non-missing observations are summed iteratively to find the least square solution that provides the principal components. Third, variables are transformed to remove non-stationarity. Depending on the type of data and the results of an augmented Dickey-Fuller test, they are differenced, differenced, or just log-transformed and then standardized as in Equation 1. Fourth, they purge all the variables from the endogenous effect caused by present and past changes in GDP and inflation. To do that, following Equation 8 they regress the vector variables X for each variable j at time t on present and lagged values of GDP (Y ) and inflation rate (π), obtaining the error term (ω) as the exogenous effects of the variables. Fifth, they determine that the results obtained with only the first principal component perform just as well as with two or more

1_{This method will be extensively analyzed in Section 5.1.} 2

(11)

components. Last, Hatzius et al. (2010) investigate long-term forecasts of GDP based on an FCI, obtaining better results than average single financial indicators.

Xj,t = αj + k X i=o βi,jYt−i+ l X i=0 γi,jπt−i+ ωj,t (8)

A study conducted by Brave and Butters (2011) also measures an FCI with a PCA approach in order to forecast economic conditions. The authors use twice as many variables compared to the previous study (100 indicators). They built a weekly index based on datasets with different frequencies, ranging from one week to one-quarter. While this technique adds more information, the FCI is more volatile at the end of the series, until the low-frequency variables are revealed resulting in potential advantages and disadvantages. They clean the data following Hatzius et al. (2010) finding that GDP growth and inflation are enough to remove the macroeconomic events from the variables leaving only their exogenous effects. In order to consider various data frequencies, they recur to a large dynamic factor framework, setting up a Kalman filter and adding extra state variables to deal with missing values. The main characteristic of this methodology is that the authors manage to decompose the variance of the FCI among three groups of variables, allowing to explain changes in the FCI in terms of money markets, debt and equity markets, and banking system.

PCA has the advantage of providing a simpler and more parsimonious description of the covariance structure (Johnson and Wichern, 2002). Additionally, it is a reliable method to summarize large datasets robustly. Among its drawbacks, it cannot shift faces in order to synchronize the movements of the variables in a time perspective.3

2.3. Latin American cases

Most of the FCIs in literature have been developed for the United States and Europe. In Latin America, it is still a matter of work, only Colombia and Mexico had some experiences on this.

Clavijo et al. (2014) developed a Colombian FCI for screening economic development, addressing macro-financial shocks and measuring the contra-cyclical role of monetary

pol-3_{To take face shifting into account, dynamic factor models are necessary, which are mostly used by central}

(12)

icy. Based on a VAR approach like the one specified in Equation 5, a series of weights are determined for different random vectors of variables according to their incidence on GDP growth. Then, through a Bayesian approximation, the best parameters are chosen (αi)

and averaged with equiproportional weights (βi) resulting in parameters wi as shown in

Equation 9. The index is built as the weighted sum of each of the five variables, weighted by wi. Data are retrieved quarterly from 2005 to 2014 in real values and normalized as

described in Equation 1.

wi =

αi+ βi

2 (9)

G´omez et al. (2011) developed another FCI for Colombia as a tool for policymakers to forecast GDP and provide information on macroprudential policy. They consider 21 variables from the most relevant markets, creating an unbalanced panel for the period 1991-2010. Variables are retrieved in real terms or if not, transformed into real terms when possible, accounting also for seasonal effects with an Additive Decomposition Method4and adjusting with Census X12. Non-stationarities are detected through augmented Dickey-Fuller tests and corrected by taking differences and log-differences. After the transforma-tion are performed, variables are normalized following Equatransforma-tion 2 and then regressed as in Hatzius et al. (2010) on GDP (Y ) and inflation (π) to extract the exogenous effect, which is captured by the residuals (ω). The process is shown in Equation 10. PCA is applied in line with the methodology proposed by Hatzius et al. (2010) obtaining a monthly FCI. On the one hand, they found the FCI to perform better than a 3-month leading indicator for economic activity than individual variables and an autoregressive model on GDP growth. On the other hand, after testing the FCI for long-term predictability of GDP (one year ahead), they found it to be a reliable early warning indicator for episodes of financial distress or economic deceleration.

Xj,t = αj + 6 X i=o βi,jYt−i+ 3 X i=0 γi,jπt−i+ ωj,t (10)

Finally, Villarreal and Bulos (2015) developed an FCI for the Central Bank of Mexico with 21 indicators for the period 2004-2014. Variables are transformed into real terms (except for interest rates), their seasonal effects are removed and first differences are taken

4

(13)

to make them stationary. Atypical data is transformed when it is outside the range µ ± 3(Q3 − Q1), where µ is the mean and Q3 − Q1 is the interquartile range, replacing them for the maximum or minimum limit depending on the case. After standardizing the variables as in Equation 1, PCA is also performed following Hatzius et al. (2010) and considering only the first principal component, which explains 31% of the total variance. To calculate the final measure, the authors normalize the FCI calculated so that the period of reference coincides with the second quarter of 2011, which is characterized by an inflation rate and a GDP close to a long-term stationary level. Following this procedure, the FCI shows in terms of standard deviations the financial conditions with respect to that reference period. This FCI is found to be very sensitive to financial distress, being useful for identifying the sources of the distress and forecasting GDP.

3. Economic context in Argentina

Before constructing indices for Argentina, it is important to understand some things about Argentina and its economy.

Argentina is a large country located in South America known for its economic paradox of becoming an undeveloped country after being one of the richest economies during the late 19th century and early 20th. It has innumerable natural resources, comparable to Australia and Canada, but because of political reasons and the Argentinian idiosyncrasy, the country deviated largely from the track of developed countries having suffered several economic and political crisis over the last 100 years. The 1929’s great depression probably represents the shift in Argentina’s direction after which the growth path never recovered in a sustainable way until present.

High inflation has always reigned the Argentine economy. After the World War II, only 14 years out of 70 had an inflation lower than 10%. Nevertheless, the worst thing is that there were 16 years with inflation higher than 100%, as illustrated in Figure 1. Most of the inflation periods can be attributed to an excess in monetary funding of fiscal deficit and to the population distrust on governmental economic measures that often cause inflation spirals intensified by expectations.

Interest rates are key variables for the Argentine economy, as the Central Bank use them as a tool to accomplish the mandate of preserving the value of the peso (Argentine currency,

(14)

Figure 1: Inflation rate in Argentina (1945 – 2015, broken scale on Y-axis)

Source: INDEC.

ARS). The peso is highly influenced by expectations about changes in the exchange rate against US dollars, which are close substitutes of pesos during periods of financial distress. Argentines do not fully trust their currency and when they foresee a period of turmoil they easily sell their own currency, often emptying the international reserves of the Central Bank.

Economy and politics work like twins in Argentina. Economic instability causes political instability and the other way round. It is worth noting that since the 1930s there has been 31 presidents, less than three years of mandate on average, with 12 of them set by military coups. In Argentina, political changes represent a key component for financial decisions, unlike other countries where the political environment remains almost fixed between government elections.

During the last 30 years, Argentina has been the scenario of successive economic crisis. There were two recessions every ten years as shown in Figure 2. One of the most important crisis was the hyperinflation in 1989, that resulted in the creation of a new currency pegged to the US dollar and the resignation of President Alfonsin. Another remarkable crisis was the convertibility crisis in 2001 when the Argentine peso devaluated after ten years of a currency peg, which was followed by the resignation of President De La Rua and the ascent of three different acting presidents in the lapse of three months before President Kirchner was elected.

With these facts about Argentina, the reader should be able to understand that Ar-gentina is a volatile country in which the economy is highly influenced by domestic and

(15)

Figure 2: GDP growth in Argentina (1986 – 2013)

Source: INDEC.

foreign financial events and fundamentally by political events. Because of this, in the following section, a big dataset is built in order to account for every measured variable susceptible of affecting financial conditions.

4. Data for an Argentine FCI

Financial variables from the main markets are being considered. They consist in quan-tities, growth rates, ratios and volatilities in order to represent the financial conditions in a complete way. Following the classification of English et al. (2005) and G´omez et al. (2011), variables include:

• Interest rates, reflecting the financing cost for firms and households.

• Exchange rates, measuring economic activity through commercial balance and the perception of risk in a specific country.

• Spreads, showing the scarcity of funds to finance investment or the distrust of mar-kets.

• Asset prices, measuring the wealth of consumers and the cost of capital for firms. They are especially relevant because of their close relationship with expectations. • Credit aggregates, to measure complementary demand and supply factors. • Financial ratios, to show the build-up of risks in credit markets.

• Surveys, to provide useful information about expected changes in monetary policy. According to the availability of information from different sources, 20 variables are

(16)

included. They are described in Table 1 and comprise monthly periods between 07/2004 and 06/2015 forming a balanced panel.

In order to estimate the FCI, the variables follow a process that includes the selection, transformation, purge of endogenous effects and standardization.

First of all, the variables are retrieved from governmental institutions, universities, research institutes and international organizations in order to guarantee the credibility of the index. All the variables correspond to one of the categories previously classified and have a leading nature over financial conditions.

Second, variables are tested for a unit-root through an augmented Dickey-Fuller test. When the null hypothesis cannot be rejected at the 95% level, transformations are applied to make them stationary. Depending on the variables, they are analyzed and specific trans-formations are applied, after which they are re-tested and transformed again if necessary5. The possible transformations are:

• First difference: for rates and surveys6_.

• First log difference: for quantities.

• Levels: for spreads, volatilities and indices.

Table 1 specifies all the transformations applied to each variable.

Third, following Hatzius et al. (2010) variables are purged from their endogenous effects. They are regressed on the present and past values of GDP and inflation, and then residuals are taken as the exogenous components of each of the variables.

Re-expressing Equation 10, Equation 11 shows the procedure. Xi,t is the variable i in

time t, αi is the intercept for variable i, Yt−j and πt−j are the real GDP and the inflation

rate in period t − j, βj,i and γj,i are the coefficients for Y and π respectively, and vi,t is

the residual for variable i in time t.

Xi,t = αi+ 3 X j=0 βj,iYt−j+ 3 X j=0 γj,iπt−j+ vi,t (11)

Fourth and last, transformed variables are standardized to have mean zero and unit variance.

5

If variables remain non-stationary after a first difference, a second difference is calculated.

6_{When survey data represents an index, it is not necessary to take the first difference and levels are}

(17)

Table 1: Variables and transformations

Variable Source Type Transformation

1 Market capitalization World Federation of Exchanges

Quantity First log difference

2 Merval (stocks index based on liquidity)

Buenos Aires Stock Exchange Market

3 Burcap (stocks index based on

capitalization)

Buenos Aires Stock Exchange Market

4 Multilateral exchange rate

Central Bank of Argentina

Rate First difference

5 TED spread FED St. Louis Spread Level

6 VIX FED St. Louis Volatility Level

7 Turnover velocity checkings accounts

8 Badlar (corporate deposits interest rate ARS)

9 Badlar (Corporate deposits interest rate USD)

10 10-year mortgage loans interest rate

11 International reserves USD

12 Money supply M1 Central Bank of Argentina

13 Money supply M2 Central Bank of Argentina

(18)

Table 1: Variables and transformations

Variable Source Type Transformation

14 Share of deposits and liabilities subject to legal requirements

15 Share of effective liquidity

16 3-month LEBAC (Central Bank of Argentina’s notes) volatility Central Bank of Argentina Volatility Level 17 3-month LEBAC (Central Bank of Argentina’s notes) average implicit rate of return

18 Call money median rate

19 Call money volatility Central Bank of Argentina

Volatility Level

20 Call money volume operated

21 Government confidence index

Universidad Torcuato Di Tella

Survey Level

The dataset is graphed in Appendix 1, showing the variables before and after the trans-formations and their respective exogenous effects captured by the error term in Equation 11.

(19)

5. Methodology

The present section describes the procedures for calculating the FCI index through PCA and analyzing it as a predictor of GDP.

5.1. Principal component analysis

In order to determine the weights for the index, a principal component analysis is implemented. The method captures the underlying common structure of a dataset with highly correlated variables reducing them into few components that explain most of the common variance.

First, the previously determined exogenous effects of variables (vt,i) are compiled in a

matrix defined as V where Vi is the i-th resulting variable (Equation 12).

V =         v1, v2, . . . , vn         =         v1,1 v1,2 . . . v1,n v2,1 v2,2 . . . v2,n .. . ... . .. ... vt,1 vt,2 . . . vt,n         (12)

The target is finding a linear function α0₁V with maximum variance. After that, finding another function α0₂V , uncorrelated with the previous one and also with maximum vari-ance, and then continuing iteratively until α0_kV as shown in Equation 13. Each of those linear combinations is a principal component zk where k < n.

α0_kV = αk,1v1+ αk,2v2+ · · · + αk,nvn= zk (13)

Moreover, αk is the eigenvector of the covariance matrix Σ, corresponding to the k-th

largest eigenvalue λk (Jolliffe, 2002) as defined in Equation 14.

Σ ∗ αk= λk∗ αk (14)

At this point, it is important to specify the number of components to consider. Regard-ing the literature review, three different types of FCI will be calculated. One index with only one component, following Hatzius et al. (2010). Another with as many components

(20)

as eigenvalues larger than one, following Rencher (1995). The last one, with as much prin-cipal components as to account for 70% of the explained variance of the original dataset, following Angelopoulou et al. (2014).

Assuming that m components are retained, Z results in the matrix of principal compo-nents with the m first principal compocompo-nents in its columns as stated in Equation 15.

Z = [z1, z2, . . . , zm] (15)

In order to aggregate all the information from the principal components, loadings ωi

are calculated regarding the variance (γj) that the j-th principal component explains in a

dataset with n variables (Equations 16 and 17).

γj =

λj

n (16)

ωi = γ1∗ zn,1+ γ2∗ zn,2+ · · · + γm∗ zn,m (17)

Finally, the FCI in period t is the resulting sum of the exogenous effect of the variables vi,t weighted by the loadings ωi as shown in Equation 18.

F CIt= ω1v1,t+ ω2v2,t+ · · · + ωnvn,t (18)

5.2. GDP predictability

Based on the leading ability of indicators, the FCI can be tested as a predictor of economic activity. Following Hatzius et al. (2010), an autoregressive structure can be performed as shown in Equation 19 where real GDP growth (yt+h− yt) is predicted by its

lags and the lags of the FCI (xt) over a horizon of h periods. The regression is performed

for h = 1, 2, 3, 6, 9, 12. yt+h− yt= β0+ py X i=1 φi∆yt+1−i+ px X i=1 ψixt+1−i+ et (19)

Alternatively, x is replaced by the growth of M2 and the MERVAL stock index in order to asses if it is advantageous to use a compound index or not. To assess the predictability, the root-mean-square errors (RMSEs) are computed. To facilitate comparability, the RMSEs

(21)

are expressed as ratios or relative RMSEs where the purely autoregressive model serves as a benchmark.

6. Empirical analysis

Applying the methodology previously described, several FCIs are calculated and then tested as predictors of GDP. The target is to determine whether an index performs better than a single variable.

6.1. FCIs

Two of the FCIs are calculated according to the eigenvalues. The first of them considers only the largest eigenvalue and the other considers all the eigenvalues larger than one, which results in taking the first eight values as Figure 3 shows.

Figure 3: Scree Plot

The third FCI is built based on considering as many eigenvalues as to account for 70% of the variance explained in the dataset. Table 2 shows the components matrix, the cumulative explained variance and the correspondent loadings. Through this methodology, the first eight principal components should be considered, which is the same as taking all the eigenvalues larger than one. Additionally, an FCI for which the eigenvalues account for 50% of the explained variance is considered. Figure 4 depicts the three FCIs.

(22)

Table 2: Variables

Variable 1 2 3 4 5 6 7 8 Weights Weights >50% >70% Market

capitalization

.351 -.084 -.228 -.014 -.272 -.089 .009 .217 -.203 -.097 Merval (stocks index

based on liquidity) .447 -.122 -.199 .055 -.197 -.007 .052 .037 -.114 -.058 Burcap (stocks index based on capitalization) .449 -.118 -.169 .038 -.192 -.037 .052 .054 -.106 -.055 Multilateral exchange rate .155 -.213 -.317 -.124 .302 .029 .079 -.264 -.057 -.175 TED spread -.360 -.121 -.175 -.070 -.275 .068 -.013 .057 -.347 -.275 VIX -.337 -.218 -.244 -.083 -.166 .055 -.133 .002 -.347 -.400 Turnover velocity checkings accounts .036 .157 -.123 -.642 .006 .076 -.004 .079 -.291 -.191 Badlar (corporate deposits interest rate $AR) -.115 .131 -.315 -.003 .111 -.386 .031 -.003 -.048 -.261 Badlar (Corporate deposits interest rate U$S) -.078 -.074 .088 .040 -.079 -.443 .426 -.117 -.024 -.093 10-year mortgage

loans interest rate

-.140 .078 -.085 -.093 -.276 -.312 -.453 -.220 -.225 -.865 International reserves U$S -.002 .193 .136 -.232 -.132 .407 .324 .080 -.067 .443 M1 -.004 .540 -.305 .048 .033 .107 -.004 -.046 .077 .105 M2 .003 .536 -.298 .049 .056 .092 -.011 -.061 .093 .096 Share of deposits and liabilities subject to legal requirements -.052 -.127 -.147 .140 .056 .323 .359 -.095 -.009 .350 Share of effective liquidity .203 -.052 -.247 .058 .443 -.067 -.164 -.029 .185 .016 3-month LEBAC (Central Bank of Argentina’s notes) volatility -.054 -.237 .015 .000 .418 .225 -.281 -.139 .149 .000 3-month LEBAC (Central Bank of Argentina’s notes) average implicit rate of return

(23)

Table 2: Variables

Variable 1 2 3 4 5 6 7 8 Weights Weights >50% >70% Call money median

rate

-.138 .078 -.002 .296 .262 -.065 .046 .659 .275 .735 Call money

volatility

-.233 -.138 -.386 .117 -.025 -.080 .138 .294 -.193 .058 Call money volume

operated .084 -.083 .094 -.505 .150 -.105 -.151 .464 -.127 .040 Government confidence index .117 .252 .311 -.011 .169 -.350 .117 -.061 .302 .127 Cumulative explained variance .176 .293 .394 .464 .530 .595 .654 .710 Figure 4: FCIs

In general, the indices show a good understanding of financial conditions. One of the most important events in the last decade, the financial crisis in 2008, is represented by a sudden fall. Around October 2008 the three FCIs reached values close to two standard

(24)

deviations. Another important fact captured is the good financial conditions lived in Argentina at the end of 2010, after the country canceled its debt with the Paris Club in advance and restructured half of the remaining defaulted debt.

6.2. Testing the predictive power of the FCIs

The FCIs and the single variables M2 and Merval Stock Index are tested as predictors of GDP following Equation 19. RMSEs are calculated relative to a pure autoregressive model in order to make them comparable, resulting in the values shown in Table 3. The lags of GDP and each specific variable tested were set to py = 4 and px = 3 respectively,

following Goodhart and Hofmann (2000).

Table 3: RMSEs for FCI 1, FCI >50%, FCI >70%, M2 and the Merval Stock Index relative to a pure GDP autoregressive model

Forecast

period FCI 1 FCI >50% FCI >70% M2 Merval Stock Index

h = 1 0.9754 0.9698 0.9643 0.9708 0.9923 h = 2 0.9556 0.9488 0.9374 0.9675 0.9852 h = 3 0.9950 0.9832 0.9828 0.9606 0.9756 h = 6 0.9936 0.9887 0.9886 0.9892 0.9916 h = 9 0.9853 0.9712 0.9730 0.9900 0.9811 h = 12 0.9800 0.9751 0.9731 0.9816 0.9682 Average 0.9808 0.9728 0.9698 0.9766 0.9823

The relative RMSE represents the length of the forecast errors in terms of the pure autoregressive model. All the relative RMSEs calculated are smaller than one, meaning that they are improving on forecasts based only on lagged values of GDP.

The results of the exercise show that on average FCI >50% and FCI >70% perform better than FCI 1 and the single indicators. The best prediction is obtained using the FCI >70% when h = 2, being h = 2 the best forecast interval on average. When forecasting periods are larger than six months, the Merval Stock Index outperforms any other indicator analyzed.

6.3. Analysis for separate subintervals

To asses the explanatory power of the FCIs more thoroughly, two subintervals of 40 months are considered independently: June 2007 to September 2010, which is characterized by an overall financial distress and October 2010 to January 2014, which is a normal period

(25)

on average. Even though it would also be optimal to evaluate a prolonged sub-period of economic wellness, there is not any for Argentina between 2005 and 2015.

With more homogeneous data within each interval, the FCIs are retested as GDP pre-dictors as in the previous section. Table 4 shows the relative RMSEs for each of the sub-periods.

Table 4: RMSEs for FCI 1, FCI >50%, FCI >70%, M2 and the Merval Stock Index relative to a pure GDP autoregressive model considering two subintervals: June 2007 to September 2010 and October 2010 to January 2014

Period 1: June 2007 to September 2010 (period of financial distress) Forecast

h = 1 0.9570 0.9567 0.9611 0.9409 0.8628 h = 2 0.9299 0.9246 0.9302 0.8115 0.8902 h = 3 0.9149 0.9099 0.9053 0.9587 0.9291 h = 6 0.9211 0.9154 0.9091 0.9585 0.9231 h = 9 0.9607 0.9610 0.9606 0.9360 0.9483 h = 12 0.9638 0.9618 0.9623 0.9460 0.8920 Average 0.9412 0.9382 0.9381 0.9253 0.9076

Period 2: October 2010 to January 2014 (normal period) Forecast

h = 1 0.9259 0.8187 0.7976 0.9274 0.9160 h = 2 0.9576 0.7774 0.7864 0.9521 0.9489 h = 3 0.9867 0.9692 0.9637 0.9686 0.9374 h = 6 0.9326 0.8770 0.8764 0.9329 0.9444 h = 9 0.8751 0.8013 0.7884 0.8904 0.8320 h = 12 0.9009 0.8434 0.8504 0.9187 0.8087 Average 0.9298 0.8478 0.8438 0.9317 0.8979

For the first period considered in Table 4, relative RMSEs are slightly reduced on av-erage. Analyzing each predictor separately, it is notable that single indicators outperform FCIs during periods of tense financial conditions. On average, errors are reduced by 4% more compared to the pure autoregressive model when analyzing this subinterval. Among the FCIs, FCI 1 improves its performance reaching similar relative RMSEs to those of FCI >50% and FCI >70%.

Regarding the forecasting power, in period 1 FCIs have a better performance in longer terms, mainly for h = 3 and h = 6. It is important to mention that M2 does a good job predicting GDP for h = 2. Its relative RMSE decrease by almost 20% compared to the pure autoregressive model of GDP.

(26)

For the second period in Table 4, FCI >50% and FCI >70% improve substantially with relative RMSEs around 0.84. This implies that those FCIs reduce forecast errors by more than 15% on average relative to the pure autoregressive model, with peaks of more than 20% for FCI >50% with h = 2 and for FCI >70% with h = 1, 2, 9.

When considering normal periods, FCIs perform better than single indicators, and they also improve with the number of principal components included. In general, because of heterogeneities in the response of variables across time, a decomposition of the time-series into subintervals results in better forecasts.

6.4. Main empirical outcomes

Suggestive evidence shows that FCIs and single indicators like M2 and Merval Stock Index are better predictors of GDP, than GDP itself and its lags. When considering heterogeneous time-series, FCIs are better predictors of GDP than single indicators for forecasted periods of up to two months. Additionally, the FCI >50% and FCI >70% are better predictors of GDP than FCI 1.

When GDP is forecasted into sub-periods with relatively more homogeneous data, re-sults show that RMSEs are significantly reduced. During times of financial distress, single indicators perform better than FCIs. In contrast, during normal financial conditions FCIs outperform single indicators on average, showing improvements on relative RMSEs mostly for forecasting periods of one, two and nine months.

To sum up, FCIs prove to be important tools for capturing the financial conditions of Argentina and to predict changes in the economic activity, especially for sub-periods with normal financial conditions.

7. Conclusion

Along this document a series of FCIs are built for the Argentine economy. Even though Argentina is a big economy among Latin American countries, these kinds of indices are not part of the standard information provided by the Central Bank or other public entities.

After a thorough literature review, a methodology for building a series of Argentine FCIs is developed based on Principal Components Analysis. The FCIs reflect information from 21 variables providing a tool for forecasting Argentine economic activity. Different

(27)

numbers of principal components are considered, producing several alternative FCIs for predicting GDP growth for different time horizons.

PCA allows managing a large number of variables at the same time that captures their co-movement reducing their dimensionality to a relatively low number of principal com-ponents. Even though this method proved to be successful for the Argentine economy, further research could also consider a dynamic factor model approach to achieve problems of timing between variables and to provide a tool for synchronizing the dataset.

Like most of the literature reviewed, the empirical results support the premise that FCIs can successfully reduce forecast errors while regressing future GDP growth. In general, FCIs perform better than single indicators, reducing relative forecast errors between 2% to 4%. Particularly when considering sub-periods of normal financial conditions, relative errors are reduced by 7% to 15%.

Even though FCIs are good predictors of GDP compared to pure autoregressive models, results are not robust for every subinterval of time. When considering periods of financial distress, single indicators like M2 can outperform FCIs. Those results can be influenced by abnormal political conditions and implications of political circumstances in Argentina and represent a topic for future work.

All in all, this document has shown that FCIs are useful indicators that can fit a small and turbulent Latin American economy like Argentina, being able to predict changes in future economic activity.

(28)

A. Appendix

A.1. Datasets

(29)

(30)

Figure 7: Exogenous effects of the variables (error terms; variables on Y-axis; period in X-axis))

(31)

B. Bibliography

Angelopoulou, E., Balfoussia, H., and Gibson, H. D. Building a financial conditions index for the euro area and selected euro area countries: what does it tell us about the crisis? Economic Modelling, 38:392–403, 2014.

Borio, C. The financial cycle and macroeconomics: What have we learnt? Journal of Banking & Finance, 45:182–198, 2014.

Brave, S. A. and Butters, R. A. Monitoring financial stability: A financial conditions index approach. Economic Perspectives, 35(1):22, 2011.

Clavijo, S., Zuluaga, A. M., and Malag’on, D. El ’indice de condiciones financieras. Enfoque Mercado de Capitales, 84:1–4, 2014.

Darracq Paries, M., Maurin, L., and Moccero, D. Financial conditions index and credit supply shocks for the euro area. 2014.

ECB. Financial stability review. European Central Bank, Frankfurt Am Main, 2009. English, W., Tsatsaronis, K., and Zoli, E. Assessing the predictive power of measures

of financial conditions for macroeconomic variables. In Investigating the relationship between the financial and real economy, pages 228–252. Bank for International Settle-ments, 2005.

Freedman, C. The role of monetary conditions and the monetary conditions index in the conduct of policy. The Transmission of Monetary Policy in Canada, page 81, 1996. G´omez, E., Murcia, A., and Zamudio, N. Financial conditions index: Early and leading

indicator for colombia. Ensayos sobre Pol´ıtica Econ´omica, 29(66):174–221, 2011. Goodhart, C. and Hofmann, B. Financial variables and the conduct of monetary policy.

Technical report, Sveriges Riksbank Working Paper Series, 2000.

Hakkio, C. S. and Keeton, W. R. Financial stress: what is it, how can it be measured, and why does it matter? Economic Review, 5(50):94–96, 2009.

Hatzius, J., Hooper, P., Mishkin, F. S., Schoenholtz, K. L., and Watson, M. W. Financial conditions indexes: A fresh look after the financial crisis. 2010.

(32)

IMF. World economic outlook: Financial stress, downturns, and recoveries October 2008. World Economic Outlook, 2008.

Johnson, R. and Wichern, D. Applied Multivariate Analysis. 5 edition, 2002. Jolliffe, I. Principal component analysis. Wiley, 2002.

Osborne-Kinch, J., Holton, S., et al. A discussion of the monetary condition index. Quar-terly Bulletin, 1:68–80, 2010.

Rencher, A. C. Methods of multivariate analysis. 1995.

Swiston, A. A US financial conditions index: Putting credit where credit is due. Number 8-161. International Monetary Fund, 2008.

Villarreal, T. A. and Bulos, C. R. Estimaci´on de un ´ındice de condiciones financieras para M´exico. Working Paper, 2015-17, 2015.

Xiong, W. Constructing the monetary conditions index for China. Frontiers of Economics in China, 7(3):373–406, 2012.

Measuring financial conditions for Argentina