Occasional Papers No. 17. Andreea Muraru. Building a financial conditions index for Romania

Occasional Papers No. 17 Andreea Muraru Building a financial conditions index for Romania

OCCASIONAL PAPERS No. 17 July 15

N O T E The views expressed in this paper are those of the author and do not necessarily reflect the views of the National Bank of Romania. All rights reserved. Reproduction for educational and non-commercial purposes is permitted provided that the source is acknowledged. ISSN 1584-867 (online) ISSN 1584-867 (e-pub)

Building a financial conditions index for Romania Andreea Muraru The author would like to thank Dorina Antohi, Ioana Udrea and Răzvan Radu for their suggestions and comments during the drafting of the paper. All remaining errors are the sole responsibility of the author.

Contents Abstract 7 1. Introduction 9. Literature review 1 3. Methodology and data 11 4. Building the financial conditions index 14 4.1. VAR estimation of the FCI 14 4.. Principal component analysis estimation of the FCI 19 4.3. Dynamic factor model estimation of the FCI 5. Conclusions 4 References 6 Appendix 1. Analysing the leading indicator properties of VAR FCI 8 Appendix. Analysing the leading indicator properties of PCA FCI 8 Appendix 3. Analysing the leading indicator properties of monthly PCA FCI 9 Appendix 4. Analysing the leading indicator properties of quarterly DFM FCI 3 Appendix 5. Analysing the leading indicator properties of monthly DFM FCI 31

Abstract As the interaction between the financial sector and the real economy gained higher significance after the outbreak of the economic and financial crisis in the fall of 8, we propose in this paper a financial conditions index (FCI) which, while delivering a comprehensive synthesis of the financial conditions for the Romanian economy, also offers a historical perspective and illustrates the factors that underpin its movement (acting towards loosening or tightening). The indicator was built using three methods: weighted averages starting from the impulse-response functions from a vector auto-regression estimated for the selected financial indicators and GDP, principal component analysis and dynamic factor models (DFM) applied on a large set of financial variables describing the domestic and external financial system. The results show that, regardless of the estimation procedure, the index is an instrument capable of offering a broader picture of the relationship between the financial variables and the real sector, extending/ complementing the information provided by the monetary conditions index; moreover, the FCI proved to be an indicator for near-term economic activity, therefore it can be used in forecasting exercises. Keywords: financial conditions index, principal components, vector auto-regression JEL classification codes: E44, E5, E17 NATIONAL BANK OF ROMANIA 7

July 15 1. Introduction The international financial crisis, which started in 8 and was followed by the sovereign debt crisis in the euro area, revealed the importance of the interaction between the financial system and real economy, as well as the stringent need for its revaluation. Given the often divergent and highly volatile developments of financial variables, one of the approaches suggested in the economic literature was the construction of an aggregate indicator the financial conditions index (FCI), which would describe the relationship between the financial system and economic activity, allow a historical evaluation of financial conditions tightness/looseness and at the same time offer information regarding the near-term evolution of economic activity. This indicator is built by including a wide-ranging data set of different statistical content, including quantities, prices, spreads, volatilities variables which, through their impact on consumption and investment decisions, eventually influence economic activity. As it explores the connection between financial conditions and the real sector, and also provides a comparison of financial conditions throughout time, the FCI can be considered an extension of the monetary conditions index 1. Also, because of its strong connection with economic activity, the FCI is a potentially important element in forecasting near-term GDP. The FCI s information content is dependent on the stability of the relationship between financial and real variables. Hence we built a suite of indices by using different, mutually-complementing and validating methods. By taking this approach, the criticism on the FCI s model dependency was avoided (Gauthier et al., 4). In this study we chose three methods, frequently used in the economic literature, for building the index: weighted averages, using the impulse-response functions from a vector auto-regression (VAR) consisting of all the selected financial variables and GDP, principal component analysis (PCA) and dynamic factor model (DFM). The leading indicator ability of the FCI for GDP was investigated by the use of correlation coefficients, Granger causality tests, and also by examining its capacity to improve the forecasting ability of autoregressive GDP equations by comparing RMSE and MSE. In order to build indices which capture only the impact of financial shocks, we used the residual variables obtained from regressing the data series on GDP in our analyses. The paper is structured as follows: section provides a brief review of the literature, section 3 describes the data and the methodology, in section 4 the index is constructed and analysed in connection with the economic activity, and section 5 concludes. 1 The description of the FCI as completing/extending monetary conditions and of its role in the monetary policy transmission mechanism can be found in Hatzius et al., 1; Goodhart and Hofmann, and 1; Gauthier et al., 4. NATIONAL BANK OF ROMANIA 9

Occasional Papers No.17. Literature review A great number of studies have been devoted to building financial conditions indices, especially in recent years. FCIs have been developed mainly by central banks and private institutions, but also by academic researchers, with the purpose of assessing aggregate financial conditions and of exploring the connection between financial conditions and economic activity (real economy). A seminal paper is that of Hatzius et al. (1). In that study, the authors construct a financial conditions index and, at the same time, provide a comprehensive inventory of the most well-known indicators, methodological alternatives and clusters of variables that are used in the analyses. They consider that the FCI should incorporate only movements exogenous to economic activity; consequently, they use the residuals obtained by regressing the selected financial variables on GDP. There are a large number of IMF papers dedicated to building and analysing financial conditions indices. The study of Ho and Lu (13) focuses on the Polish economy and builds the FCI in two ways: by using the weighted average approach (based on estimating a VAR) and the principal component approach (PCA) applied on both domestic and external financial variables and assesses the impact of each of the data series considered on the level and movement of the indicator. In Gumata et al. (1), the PCA method in building the FCI for South Africa is complemented by a dynamic approach which makes use of Kalman filtering; the index constructed was found to have good predictive abilities for near-term economic growth. Osorio et al. (11) build their FCI for Asian economies as an average of the two indices computed through PCA and the weighted average approach. Matheson (11) applies a dynamic factor model using the methodology of Doz et al. (11) to extract the FCI from a set of 3 variables for the US economy and 17 for the euro area. Another study dedicated to building an FCI for emerging markets is that of Kara et al. (1). The authors compute the financial conditions index for the Republic of Turkey starting from a VAR model, taking into account the impact of external financial conditions. The central bank of Hungary has also constructed an FCI; the methodology selected was structural VAR 3. Angelopoulou et al. (13) apply principal component analysis on 4 variables, retain three of the components extracted, and then determine the FCI as the sum of the three, weighted with the corresponding share of variance they explain. The indicator is determined for the euro area and for some Member States, and it highlights the 3 Among the first studies dedicated to building an aggregate measure of financial conditions was the one developed by the Bank of Canada (Freedman, 1995), concentrated on obtaining a broad monetary conditions index. Financial conditions indices were developed later (Gauthier et al., 4). The indicator and its evolution are periodically presented in the Trends in Lending report, available on the central bank s website. 1 NATIONAL BANK OF ROMANIA

July 15 heterogeneity of financial conditions inside the euro area, both before and after the emergence of the economic and financial crisis. The Deutsche Bank economists (Hooper et al., 14) construct the FCI from 38 variables, weighted with their time-varying average correlation with GDP growth. In their study they also compare various FCIs, based on different data sets and estimated with different procedures, showing that in spite of their uneven content and distinct estimation methodologies, the indicators behave similarly in time 4. Koop and Korobilis (13) adopt a more econometrically challenging approach and build a financial conditions index by using dynamic model selection (DMS) and dynamic model averaging (DMA) among a large number of TVP-FAVARs. The results they obtain suggest better performance of their indicator especially in times of higher financial stress. All the literature on the index reveals that regardless of the chosen estimation procedure, FCIs do succeed in characterising the prevalent financial conditions at different moments and in describing their evolution through time, but also in improving the near-term economic activity forecasts. 3. Methodology and data In the economic literature, the FCI is most frequently built as a weighted average of financial variables:, (1) where are the standardised financial variables used in the analysis and the assigned weights 5. In this paper we chose to estimate the coefficients needed for building the FCI by three different, complementary methods: vector auto-regression, principal components analysis, and dynamic factor models. 4 5 The latest index constructed by Deutsche Bank, the one built by Bloomberg Kansas City Fed, and the USMPF FCI previously developed by Deutsche Bank. When the index is built starting from the VAR estimation, will be the cumulated impulse response of GDP; when applying the PCA, will be the product between the first three eigenvectors (a number of three principal components will be selected) of the covariance matrix and the quantity of information they preserve; in the case of the DFM, if we consider to be the three factors, will be proportional to their variance. NATIONAL BANK OF ROMANIA 11

Occasional Papers No.17 Vector auto-regression The advantage of this method is that it takes into account the interdependence between the financial data series and between those variables and economic activity. The real GDP quarterly growth rate together with the selected financial variables are included in the VAR model and the impulse-response functions are computed. We made use of generalised impulses, therefore the problem of ordering the variables when orthogonalising shocks is avoided. The weighted average approach is based on this technique, the FCI being built, in our study, by weighting the initial standardised variables with the GDP impulse-response cumulated over two quarters to a one-standard deviation shock in the variables considered. Principal components analysis This data analysis technique has the appealing advantage of dimension reduction. Essentially, the information in the initial data series is rewritten, starting from their covariance matrix, in the form of independent variables (linear combinations of the original data series), named principal components, which have among others the characteristic of decreasing variability. The first component has the property of preserving the highest amount of information, the second the next highest and so on, allowing, therefore, that a smaller number of variables account for a sufficiently large share of the variance. Hence, while losing a minimum amount of information, the initial data set can be synthesised in a smaller number of variables which can be further used in the analyses. The retained principal components are then weighted with the share of variance they explain, and the FCI is determined as their sum. As the goal of the analysis is to obtain a financial conditions index that captures the exogenous movements of financial variables, the data series are at first purged from current and past GDP influence. Dynamic factor model estimation of the FCI The previous two methods are complemented by the dynamic approach which has the advantage of accounting for the changing relations between the economic variables. Applying factor analysis on the variables allows for identifying the common component in the data and isolating it from the idiosyncratic shocks.. () The dynamic part of the model assumes that follows an auto-regressive process of order p:. (3) 1 NATIONAL BANK OF ROMANIA

July 15 For estimating this model, we applied the two-step procedure of Doz, Giannone and Reichlin (11). This method uses the results obtained from the PCA 6 (the first step) and subsequently applies the Kalman filter and smoother (the second step when introducing the dynamics of the common factors). The data series Starting from the existing literature, but also taking into account the particulars of the domestic economy, we chose different categories of variables representative for the local financial system as well as for the external environment, including prices, quantities, and measures of volatility 7. Another category of indicators used in the literature (Swiston, 8; Hatzius, 1; Kara et al., 1; Ho and Lu, 13), which could not be included in this analysis because of its short time span, is credit standards and terms. Its inclusion in the index is intended in future estimation exercises. The variables we selected for this analysis are the following: The real growth rate of credit to the private sector the choice of this variable is motivated by the importance of the banking sector in financing economic activity. When performing the PCA, we used two separate variables, one depicting the leudenominated loans and one for the euro component 8, while in the VAR estimation, given the relatively short sample and the need to preserve as many degrees of freedom as possible, we kept only one variable for total loans to the private sector. The interest rate spread for loans this variable incorporates the risk premium. It was calculated as the difference between the average interest rate on new lei loans to households and non-financial corporations and the ROBOR 3M rate, and between the similar rate on the euro component and EURIBOR 3M respectively. When applying the VAR methodology, we used only one variable representing the spread of the interest rate for total credit 9, computed as a weighted average with the volume of loans. The change in real ROBOR 3M 1 this variable incorporates the effect of monetary policy. The choice of expressing the variable in real terms is justified by the fact that economic agents make investment and consumption decisions starting from the real interest rate. The change in the real effective exchange rate (REER) 11 reveals the impact of the exchange rate on economic growth. The selection of the effective exchange rate is motivated by its strong relation to economic activity (net export). 6 7 8 9 There are numerous studies which prove that principal components are consistent estimators of true latent factors (Stock, Watson, ; Forni, Hallin, Lippi, Reichlin, ). A description of the way in which these variables influence financial conditions can be found in Angelopoulou et al. (13) and Hatzius et al. (1). Expressed in euro. The two spreads are positively correlated throughout the investigated interval. 1 Ex-post measure. 11 The source of this variable is the European Commission. The deflator is the HICP, and the number of partner countries is 37. NATIONAL BANK OF ROMANIA 13

Occasional Papers No.17 The EURIBOR-OIS spread shows the liquidity conditions on the euro area money market (consequently the credit risk, as parent banks of local subsidiaries access financing on the euro-area market). The VIX index describes the risk aversion by measuring the short-term expected volatility based on S&P 5 options. Since its introduction, this index has been considered a barometer for investor sentiment, being frequently named the fear index. 4. Building the financial conditions index In what follows, the financial conditions index is estimated by applying the three methods described in the previous section. The time span covered by the analysis is 5 Q1 14 Q3 and the index is determined with a quarterly frequency by all three methods; when principal component analysis and dynamic factor models are used, we also constructed a monthly index. 4.1. VAR ESTIMATION OF THE FCI Vector-autoregression is the most widely used method for estimating the FCI because, as previously mentioned, it allows us to quantify the impact of shocks on GDP while taking into account the interactions in the data structure. In estimating our VAR, we used a number of six variables, with a quarterly frequency (the real growth rate of credit to the private sector, the interest rate spread for loans, the change in real ROBOR 3M, the change in the REER, the EURIBOR-OIS spread, and the VIX index) 1. Similarly to Kara et al. (1), the generalised impulse-response functions were used, as they do not require any pre-ordering of the data 13. The impulse-response functions are described in Chart 1. In most cases, the maximum response is registered after two periods, hence the weight determined as a two-quarter cumulative impulse-response attached to the variables in building the FCI. 1 The variables are stationary according to ADF and Phillips-Perron tests (most of the series are growth rates, so already in first differences). 13 This manner of computing impulse-responses was developed by Pesaran and Shin (1998) building on Koop et al. (1996) for removing the shortcomings regarding the dependency on the order of the variables when the shocks to the VAR model are orthogonalised using a Cholesky decomposition. 14 NATIONAL BANK OF ROMANIA

July 15 Chart 1. Impulse-response functions Response to Generalized One S.D. Innovations ± S.E. 1.5 Response of GDP_qoq to GDP_qoq 1.5 Response of GDP_qoq to SPREAD_TOTAL 1.5 Response of GDP_qoq to ROBOR 3M_REAL_qoq 1. 1. 1..5.5.5... -.5 -.5 -.5-1. -1. -1. -1.5 1 3 4 5 6 7 8 9 1-1.5 1 3 4 5 6 7 8 9 1-1.5 1 3 4 5 6 7 8 9 1 1.5 Response of GDP_qoq to REER_qoq 1.5 Response of GDP_qoq to EURIBOR_OIS 1.5 Response of GDP_qoq to VIX 1. 1. 1..5.5.5... -.5 -.5 -.5-1. -1. -1. -1.5 1 3 4 5 6 7 8 9 1-1.5 1 3 4 5 6 7 8 9 1-1.5 1 3 4 5 6 7 8 9 1 1.5 Response of GDP_qoq to CREDIT_qoq 1..5. -.5-1. -1.5 1 3 4 5 6 7 8 9 1 Souce: NBR, NIS, Bloomberg, European Commission, author's calculations As Chart 1 shows, the response of GDP to the shocks in the selected financial variables is in accordance with economic theory, with the exception of the response to a one standard deviation of the real effective exchange rate, which proves to be statistically not significant. This behaviour denotes the fact that the contrary effects of the REER on economic activity, through the two specific channels (foreign trade and wealth and balance sheet effect), neutralise each other. Such a result is not singular using the same methodology for estimating the FCI for Poland produces a similar result (Ho and Lu, 13). After extracting the cumulative impulse-responses, the FCI has been computed according to the formula:, (4) where, the attached weight, is the accumulated two quarters impulse-response of GDP to a one standard deviation shock of variable. NATIONAL BANK OF ROMANIA 15

Occasional Papers No.17 The index computed in this manner strongly correlates with the real quarterly growth rate of GDP, the highest correlation coefficient (.8) being obtained for FCI leading GDP by one quarter. The chart below depicts real GDP quarterly growth rates and the FCI a downward/ upward movement of the FCI shows tighter/more relaxed financial conditions. Chart. GDP real quarterly growth rate and FCI 6 % 4 - -4-6 -8-1 Correlation coefficient:.8 5Q1 5Q 5Q3 5Q4 6Q1 6Q 6Q3 6Q4 7Q1 7Q 7Q3 7Q4 8Q1 8Q 8Q3 8Q4 9Q1 9Q 9Q3 9Q4 1Q1 1Q 1Q3 1Q4 11Q1 11Q 11Q3 11Q4 1Q1 1Q 1Q3 1Q4 13Q1 13Q 13Q3 13Q4 14Q1 14Q 14Q3 14Q4 FCI_VAR (-1) GDP_qoq Souce: NBR, NIS, Bloomberg, European Commission, author's calculations The effectiveness of the FCI in providing information regarding the near term evolution of GDP was explored by applying a Granger causality test and also by including the newly-built index in an autoregressive equation for GDP. As the tables in Appendix 1 show, a one-way causality can be detected from FCI to GDP, and including FCI with a lag of 1 in an autoregressive GDP equation improves both the adjusted determination coefficient, R-squared, and the equation s forecasting ability (measured by comparing RMSE and MSE). The FCI offers, first and foremost, useful elements for policy analysis. Knowing the contribution of all variables in building the indicator, it can be decomposed in its constituent elements and, therefore, the impact of each of the variables on financial conditions can be quantified and examined at different moments. Chart 3 illustrates the significant influence of external variables on the evolution of financial conditions in Romania. Among the factors specific to the local financial sector, the most important seems to be the real growth rate of credit to private sector, especially after the start of the global financial crisis. Looking at the evolution of the FCI, a period of relatively loose financial conditions can be distinguished in the first part of the chart and two moments of more pronounced tightening later, each followed by episodes of modest improvement and volatility. 16 NATIONAL BANK OF ROMANIA

July 15 Chart 3. FCI factor decomposition 6 pp % 6 Loosening Tightening 4 - -4-6 -8-1 5Q1 5Q 5Q3 5Q4 6Q1 6Q 6Q3 6Q4 7Q1 7Q 7Q3 7Q4 8Q1 8Q 8Q3 8Q4 9Q1 9Q 9Q3 9Q4 1Q1 1Q 1Q3 1Q4 11Q1 11Q 11Q3 11Q4 1Q1 1Q 1Q3 1Q4 13Q1 13Q 13Q3 13Q4 14Q1 14Q 14Q3 SPREAD ROBOR 3M REER EURIBOR-OIS VIX loans FCI (rhs) Souce: NBR, NIS, Bloomberg, European Commission, author's calculations 4 - -4-6 -8-1 According to this indicator, the financial conditions were the loosest in 6, such an evolution being determined, on one hand, by developments in some variables specific to the domestic financial sector loans and interest rate spread. Credit to the private sector increased substantially, especially due to the faster dynamics of the lei component (which reached a historical high of 95.6 percent in July 6), on account of both demand 14 and supply 15 ; the annual dynamics of the lei component was higher than that of the foreign currency-denominated one, because of the measures taken in 5 by the central bank for discouraging forex loans 16. On the other hand, the external environment also acted towards relaxing the financial conditions, the EURIBOR-OIS spread signalling favourable financing conditions on the external markets, and the VIX index the absence of stress factors. Towards tightening the financial conditions acted the ROBOR 3M rate, in accordance with the higher restrictiveness of the monetary policy stance of the period, and implicitly the leaning-against-the-wind behaviour of monetary policy. As Chart 4 shows, the NBR continued the tightening cycle of monetary policy, started at the end of 5, by raising the monetary policy rate by 1.5 percentage points (from 7.5 to 8.75 percent) and also by increasing the MRR ratio on both leu- and foreign currencydenominated liabilities of credit institutions, on account of the increasing risks to disinflation triggered by the rise and persistence of excess demand. 14 Stimulated by the high households revenues and their positive outlook, likely to foster consumption as well as by companies favourable financial situation and expectations of ongoing economic growth (NBR, Annual Report, 6). 15 Spurred by the tougher competition between banks, especially for the household sector, the improved perception on clients creditworthiness, as well as by the increase of foreign financing of credit institutions (NBR, Annual Report, 6; Inflation reports, 6). 16 In 5, NBR gradually raised the minimum reserve ratio on banks foreign-exchange-denominated liabilities with residual maturities longer than two years to 3 percent, and at the end of the year it was increased to 35 percent. At the same time, the NBR adopted administrative and prudential measures (NBR Norms No. 11/5 for containing the concentration of exposures from forex loans, in effect from September 6, 5 and NBR Norms No. 1/5 aimed at limiting the risk attached to household loans). NATIONAL BANK OF ROMANIA 17

Occasional Papers No.17 In 8, the FCI started to decrease, registering an abrupt decline in the 4 th quarter in the context of the global financial crisis break-out rapidly corrected afterwards. As the decomposition illustrates, the worsening of financial conditions was mainly due to external influences, with the significantly higher global uncertainty and the amplified tensions on the euro area financial market being reflected by the evolution of the VIX index and of the EURIBOR-OIS spread. The ROBOR rate also contributed to diminishing the FCI the amplitude of its contribution being nonetheless comparable to previous quarters, when, acting in accordance with the leaning-against-thewind principle, the central bank continued to increase the monetary policy rate for offsetting the inflationary pressures of excess demand 17 as well as the risks concerning the deterioration of inflation expectations as a consequence of previous supply-side shocks (increase of administered prices, rise in food and oil prices) 18. In the 4 th quarter, the NBR discontinued the ascending path of the monetary policy rate, with the upward movement and the high volatility of interest rates on the interbank market being generated by the change in the net liquidity to deficit (after approximately a decade of surplus), also on account of external factors. In 11, the financial conditions index registered another episode of decline, but of a lower magnitude, also driven mainly by external evolutions, as international financial market sentiment worsened in response to the heightening of the sovereign debt crisis and fears regarding its possible spread, also on account of the downgrade of the sovereign rating of Portugal and Ireland to below investment grade and the worsening situation in Greece. The unfavourable impact of the latter was more strongly perceived by the Romanian economy, given the higher presence of banks with Greek capital on the domestic market compared to other markets in the region. Chart 3 also illustrates that the evolution of the stock of credit has systematically acted towards the deterioration of financial conditions, the feeble lending being the result of both demand and supply factors 19. The subsequent evolution of the FCI continued to receive and reflect especially the influence of the external environment, the decomposition of the index illustrating movements such as: (i) the contribution of the EURIBOR-OIS spread turned positive in 1 Q3, after the speech of the ECB President ( the ECB is ready to do whatever it takes to preserve the euro ); (ii) the contribution of the VIX fluctuated and entered negative territory in May 13, as the Fed announced the possibility of tapering out the financial asset purchase programme, but also in the first and in the third quarter of 14, against the background of the dispersion onto Central and Eastern-European financial markets of the effects of Turkey s worsening situation and of the rising geopolitical tensions respectively. 17 The annual dynamics of GDP grew to 8. percent and 9.3 percent respectively in the first two quarters of 8, according to data available in real time. 18 For details see the NBR's, Annual Report 8 and the Inflation Reports from the same year. 19 The latter reflecting the constraining effects of the consistent growth of nonperforming loans, as well as the cross-border deleveraging process. 18 NATIONAL BANK OF ROMANIA

July 15 Chart 4. FCI, monetary policy rate, annual inflation rate 4 - -4-6 -8-1 % FCI_VAR FCI_PCA (rhs) 1-1 - -3-4 -5 16 14 1 1 8 6 4 % 5Q1 5Q 5Q3 5Q4 6Q1 6Q 6Q3 6Q4 7Q1 7Q 7Q3 7Q4 8Q1 8Q 8Q3 8Q4 9Q1 9Q 9Q3 9Q4 1Q1 1Q 1Q3 1Q4 11Q1 11Q 11Q3 11Q4 1Q1 1Q 1Q3 1Q4 13Q1 13Q 13Q3 13Q4 14Q1 14Q 14Q3 annual inflation rate monetary policy rate Souce: NBR, NIS, Bloomberg, European Commission, author's calculations 4.. PRINCIPAL COMPONENT ANALYSIS ESTIMATION OF THE FCI This type of analysis is more permissive with the number of financial variables that can be used, therefore we chose to include the growth rate of credit and the interest rate spread separately for the two components, lei and euro. Besides the variables described in section III, in this analysis we also included the CDS spread for Romania, as a measure of the risk premium. The PCA is a purely statistical method and does not interpret the data series in connection to GDP, hence it can also be applied on higher frequency data. As mentioned before, in order to obtain an FCI exogenous to GDP, the influence of economic activity was eliminated from the variables (Hatzius et al., 1; Ho and Lu, 13, Paries et al., 14):, (5) where =1 n, the number of variables, represents the financial variables considered, is GDP and the residual variable represents the purged financial variables, uncorrelated with the GDP. NATIONAL BANK OF ROMANIA 19

Occasional Papers No.17 Further on, the PCA is applied on residuals as follows:, (6) where is the number of principal components, are the weights attributed to the exogenous financial variables, representing the eigenvectors of covariance matrix. Before applying the PCA, the variables underwent some transformation: they were standardised so as to avoid the problem of different measurement units, and some of them (spreads and volatilities) had their sign reversed so that an increase in the variable represents an improvement in financial conditions. Thus, the index can be interpreted as in the previous case: the increase in its value denotes looser financial conditions (prone to favour economic growth), while its decrease is a sign of heightened financial pressure (which could prove to be a drag on GDP growth). The decision regarding the number of principal components was made starting from the quantity of information retained. A single component would not have been appropriate, as the proportion of variance it describes is of just 44 percent. A number of three components, nevertheless, retain more than 7 percent of the initial information in the data. Table 1. PCA eigenvalues Number Value Difference Proportion Cumulative value Cumulative proportion 1 3.93.38.44 3.93.44 1.54.5.17 5.47.61 3 1.4.1.1 6.51.7 4.83.1.9 7.35.81 5.71.31.8 8.6.9 6.41.17.5 8.47.95 7.4.1.3 8.71.97 8.3.17.3 8.94.99 9.6 -.1 9. 1. Following Angelopoulou et al. (13), we built the FCI as an average of the first three principal components weighted with the proportion of explained variance. Consequently, (7) The manner in which the PCA is applied in this paper follows closely the methodology in Angelopoulou et al. (13). NATIONAL BANK OF ROMANIA

July 15 Chart 5. GDP and FCI PCA 6 4 % GDP_qoq FCI_PCA (-1) (rhs) 3 1 - -1-4 - -6-8 Correlation coefficient:.6 5Q1 5Q 5Q3 5Q4 6Q1 6Q 6Q3 6Q4 7Q1 7Q 7Q3 7Q4 8Q1 8Q 8Q3 8Q4 9Q1 9Q 9Q3 9Q4 1Q1 1Q 1Q3 1Q4 11Q1 11Q 11Q3 11Q4 1Q1 1Q 1Q3 1Q4 13Q1 13Q 13Q3 13Q4 14Q1 14Q 14Q3 14Q4-3 -4 Souce: NBR, NIS, Bloomberg, European Commission, author's calculations Even though it is highly correlated with the FCI obtained from the VAR estimation (Chart 4), the FCI PCA has a lower correlation with GDP, especially due to its movement in the first part of the interval (Chart 5). Nevertheless, its leading indicator properties, investigated with the same procedure as in the previous case, are still noticeable one-way Granger causality from FCI to GDP and improvement of equation forecasting performance (Appendix ). The PCA estimation was also implemented on monthly data the financial condition index is represented in Chart 6 (eigenvalues are presented in Appendix 3). The index was computed as a weighted average of the first three principal components, the weights being the corresponding proportion of explained variance:. (8) Chart 6. Monthly FCI PCA FCI PCA monthly frequency 6 % FCI and GDP 1.5 1 4 1..5. -1 - -4 -.5-1. - -3-4 5M1 5M8 6M3 6M1 7M5 7M1 8M7 9M 9M9 1M4 1M11 11M6 1M1 1M8 13M3 13M1 14M5-6 -8-1 5Q1 5Q3 6Q1 6Q3 7Q1 7Q3 8Q1 8Q3 9Q1 9Q3 1Q1 1Q3 11Q1 11Q3 1Q1 1Q3 13Q1 13Q3 14Q1 14Q3 GDP_qoq Correlation coefficient:.73 FCI_PCA_quarterly average (-1) (rhs) -1.5 -. -.5 Souce: NBR, NIS, Bloomberg, European Commission, author's calculations NATIONAL BANK OF ROMANIA 1

Occasional Papers No.17 In order to investigate the leading indicator properties of the monthly FCI, we computed quarterly averages of its values and investigated the same statistics as in the previous two cases. The results are similar to the former ones, FCI Granger causing quarterly GDP real growth rate and also improving the forecasting ability of GDP s equation. Still, the criticism of non-exogeneity with respect to GDP can be raised about this version of the index 1. 4.3. DYNAMIC FACTOR MODEL ESTIMATION OF THE FCI Starting from the study of Matheson (11), the financial conditions index was also built by applying the two-step procedure developed by Giannone et al. (4), Giannone et al. (8) and whose consistency properties were analysed in Doz et al. (11). The two-step approach implies using the PCA and the Kalman filter: in the first stage, the principal components are extracted from the data, and in the second step, the dynamics of the common factors as well their heteroscedasticity are taken into account. After applying this procedure (by means of the MATLAB code made available by one of the authors ) 3 both on quarterly and monthly data, we obtained an index with a pattern similar to the one computed with the PCA. In building the FCI, we made use of the previous information, and selected a number of three principal components which were used in the second step for obtaining the dynamic factors through Kalman filtering and smoothing. In the autoregressive structure of the factors we used one lag. For computing the FCI, the three factors were weighted proportional to their variability, as a proxy for the quantity of the total variance they explain.. (9) Unsurprisingly, this version of the FCI also proves to have leading indicator properties for economic activity. Just as in the previous cases, the strongest correlation is registered for FCI leading one quarter. Granger causality tests are in line with previous findings, GDP does not Granger-cause FCI, whereas the hypothesis that FCI does not Granger-cause GDP can be discarded. At the same time, RMSE, MSE and adjusted R-squared favour the introduction of the FCI in any equation of the GDP (see Appendix 4). 1 The higher value of the correlation coefficient in this case is due to the fact that the variables, having monthly frequency, could not be regressed on GDP before applying the PCA. If the FCI built as a quarterly average is purged from the influence of GDP, the correlation coefficient with economic activity diminishes to.57 (see Appendix 3). This drawback of the indicator is compensated by its monthly availability, independent of the release of quarterly data on economic activity. This estimation procedure is widely used for computing dynamic factors. See, for example, Koop and Korobilis, who include it in their DMS and DMA models (Koop, Korobilis, 13). 3 Available at http://homepages.ulb.ac.be/~dgiannon/. NATIONAL BANK OF ROMANIA

July 15 Chart 7. GDP and FCI DFM 6 4 - -4 % 3 1-1 - -6-8 Correlation coefficient:.53-3 -4 5Q1 5Q3 6Q1 6Q3 7Q1 7Q3 8Q1 8Q3 9Q1 9Q3 1Q1 1Q3 11Q1 11Q3 1Q1 1Q3 13Q1 13Q3 14Q1 14Q3 GDP_qoq FCI_DFM(-1) (rhs) FCI_PCA(-1) (rhs) Souce: NBR, NIS, Bloomberg, European Commission, author's calculations The DFM estimation was also performed on monthly data the higher frequency index is illustrated in Chart 8. Just like in the case of the monthly FCI PCA described in the section above, monthly averages were computed and analysed in connection with GDP, the results supporting the same leading indicator abilities of the FCI (see Appendix 5). Chart 8. FCI DFM monthly frequency 3 FCI DFM monthly frequency 6 FCI and GDP.5 4 1.5 1.75. -1 - -.75 - -4-1.5-3 -4-5 5M1 5M8 6M3 6M1 7M5 7M1 8M7 9M 9M9 1M4 1M11 11M6 1M1 1M8 13M3 13M1 14M5 Souce: NBR, NIS, Bloomberg, European Commission, author's calculations -6-8 5Q1 5Q4 6Q3 7Q 8Q1 8Q4 9Q3 1Q 11Q1 11Q4 1Q3 13Q 14Q1 14Q4 GDP_qoq Correlation coefficient:.74 FCI_DFM_quarterly average(-1) (rhs) -.5-3. The correlation coefficient between GDP and the quarterly averages of monthly FCI is also higher than the one between GDP and FCI computed on quarterly data. Nevertheless, as in the case of the monthly FCI computed by applying the PCA, this index is also subject to the non-exogeneity critique 4. 4 After isolating GDP s influence, the correlation coefficient diminishes to.58 (Appendix 5). The leading indicator ability of the index is, nevertheless, preserved. NATIONAL BANK OF ROMANIA 3

Occasional Papers No.17 The indicators can be used individually in analyses, validating and complementing each other, but there are also studies in which their average is preferred (for example Osorio et al., 11) 5. The chart below illustrates the indices that have the highest correlation with economic activity. Chart 9. FCI and average FCI 1-1 - -3-4 5Q1 5Q 5Q3 5Q4 6Q1 6Q 6Q3 6Q4 7Q1 7Q 7Q3 7Q4 8Q1 8Q 8Q3 8Q4 9Q1 9Q 9Q3 9Q4 1Q1 1Q 1Q3 1Q4 11Q1 11Q 11Q3 11Q4 1Q1 1Q 1Q3 1Q4 13Q1 13Q 13Q3 13Q4 14Q1 14Q 14Q3 6 3-3 -6-9 -1 FCI_PCA monthly FCI_DFM monthly FCI_VAR (rhs) average FCI FCI_VAR FCI_PCA monthly FCI_DFM monthly FCI_VAR 1 FCI_PCA monthly.76 1 FCI_DFM monthly.75.99 1 Souce: NBR, NIS, Bloomberg, European Commission, author's calculations 5. Conclusions The purpose of this paper was to construct a financial conditions index relevant for the Romanian economy. Starting from the data widely used in the economic literature on the topic, but also paying attention to local specificities, a number of nine financial variables (six in the case of the VAR method) were selected, including indicators describing both domestic and external financial conditions containing quantities, prices, and volatilities. The index was obtained by applying three different methods (weighted average starting from the impulse-response functions of the VAR estimated for the selected financial variables and GDP, principal components analysis and dynamic factor model), which complement and validate each other. They were applied on quarterly data, but also on monthly frequency data when the estimation procedure allowed. The evaluation of the FCI s effectiveness as an indicator for economic activity showed that, regardless of the estimation procedure, the index has leading indicator abilities for GDP, the correlation with economic activity being the highest when the indicator is considered with a lag of one quarter; Granger tests suggest a one-way causality 5 Other authors apply the PCA on their set of financial conditions indicator for obtaining an aggregate version of the FCI (Aramonte et al., 13). 4 NATIONAL BANK OF ROMANIA

July 15 from FCI to GDP, and the MSE, RMSE and R-squared show that the inclusion of FCI in the autoregressive equations of GDP improves their forecast performance. To the same extent, the construction of the aggregate indicator enables the analysis of overall financial conditions, while its decomposition facilitates the identification of influences exerted, throughout time, by the variables specific to the domestic and external financial sectors. From this perspective, the FCI can be considered a complementary indicator to the monetary policy index, being a useful instrument for monetary policy analyses. Nevertheless, the FCI cannot always precisely identify the direction of the evolution of economic activity, as GDP growth can also be affected by exogenous factors (for example productivity shocks). Moreover, the good indicator properties of the index are conditional on the stability of the relationship between real and financial variables and of the allocated weights therefore, given their changing nature, the index needs to be periodically reassessed so as to capture the most relevant combination of indicators, both in terms of their selection and of the assigned weights, as the relationships between the variables change. Moreover, the use of new estimation methods could add extra information in time. NATIONAL BANK OF ROMANIA 5

Occasional Papers No.17 References Angelopoulou, E., Balfoussia, H., Gibson, H. Aramonte, S., Rosen, S., Schindler, J.W. Doz, C., Giannone, D., Reichlin, L,. Freedman, C. Forni, M., Hallin, M. Lippi, M., Reichlin, L. Gauthier, C., Graham, C., Liu, Y. Giannone, D., Reichlin, L., Sala, L. Giannone, D., Reichlin, L., Small, D. Goodhart, C., Hofmann, B. Guichard, S., Turner, D. Gumata, N., Klein, N., Ndou, E. Hatzius, J., Hooper, P., Mishkin, F.S., Schoenholtz, K.L., Watson, M.W. Building a Financial Conditions Index for the Euro Area and Selected Euro Area Countries. What does it tell us about the crisis?, ECB Working Paper Series, No. 1541, May 13 Assessing and Combining Financial Conditions Indices, Federal Reserve Board, Staff working papers in the Finance and Economics Discussion Series, 13 A Two-Step Estimator for Large Approximate Dynamic Factor Models Based on Kalman Filtering, Journal of Econometrics, No. 164, pp. 188 5, 11 The Role of Monetary Conditions and the Monetary Conditions Index in the Conduct of Policy, Bank of Canada Review, 1995 The Generalized Dynamic-Factor Model: Identification and Estimation, The Review of Economics and Statistics, Vol. 8, No. 4, pp. 54-545, November Financial Conditions Indices for Canada, Bank of Canada Working Paper, 4-, June 4 Monetary Policy in Real Time, in: Gertler, Mark, Rogoff, Kenneth (Eds.), NBER Macroeconomics Annual, pp. 161, 4 Nowcasting: The Real-Time Informational Content of Macroeconomic data, Journal of Monetary Economics, No. 55(4), pp. 665 676, 8 Asset Prices, Financial Conditions, and the Transmission of Monetary Policy, Paper presented at the conference on Asset Prices, Exchange Rates and Monetary Policy, Stanford University, March 3, 1 Financial Variables and the Conduct of Monetary Policy, Sveriges Riskbank Working Paper, No. 1, Quantifying the Effect of Financial Conditions on US Activity, OECD, Economics Department Working Papers, No. 635, September 8 A Financial Conditions Index for South Africa, IMF Working Papers, August 1 Financial Conditions Indices: A Fresh Look after the Financial Crisis, NBER Working Paper Series, No. 1615, July 1 6 NATIONAL BANK OF ROMANIA

July 15 Ho, G., Lu, Y. Hooper, P., Spencer, M., Slok, T. Kara, H., Özlü, P., Ünalmış, D. Koop, G., Korobilis, D. Koop, G., Pesaran, H., Potter, S. Magyar Nemzeti Bank Matheson, T. National Bank of Romania Osorio, C., Runchana, P., Filiz, U. Paries, M. D, Maurin, L., Moccero, D. Pesaran, H., Shin, Y. Stock, J., Watson, M. A Financial Conditions Index for Poland, IMF Working Papers, October 13 Global Economic Perspectives: Drop in Financial Conditions Means Some Softening of US Growth Ahead, Deutsche Bank Securities Inc., October 14 Financial Conditions Indices for the Turkish Economy, CBT Research Notes in Economics, No., November 1 A New Index of Financial Conditions, European Economic Review, 13 Impulse Response Analysis in Nonlinear Multivariate Models, Journal of Econometrics, 74, pp. 119 147, 1996 Trends in Lending Financial Conditions Indices for the United States and Euro Area, IMF Working Papers, April 11 Inflation Reports, Annual Reports A Quantitative Assessment of Financial Conditions in Asia, IMF Working Paper, 11 Financial Conditions Index and Credit Supply Shocks for the Euro Area, ECB Working Paper Series, No. 1644, March 14 Generalized Impulse Response Analysis in Linear Multivariate Models, Economics Letters, 58, pp. 17 9, 1998 Forecasting Using Principal Components from a Large Number of Predictors, Journal of the American Statistical Association, vol. 97, No. 46, pp. 1167-1179,. NATIONAL BANK OF ROMANIA 7

Occasional Papers No.17 Appendix 1 Analysing the leading indicator properties of VAR FCI Table A1.1. Granger causality test Null hypothesis F-stat (probability) GDP qoq does not Granger cause FCI.7 (.4) FCI does not Granger cause GDP qoq 58.4 (.) Table A1.. Forecast performance of GDP FCI OLS Adjusted R-squared.1.7 RMSE 1.58.89 MSE 1.13.75 Appendix Analysing the leading indicator properties of PCA FCI Table A.1. Granger causality test Null hypothesis F-stat (probability) GDP qoq does not Granger cause FCI.3 (.87) FCI does not Granger cause GDP qoq 6.76 (.) Table A.. Forecast performance of GDP FCI OLS Adjusted R-squared.1.54 RMSE 1.58 1.1 MSE 1.13.87 8 NATIONAL BANK OF ROMANIA

July 15 Appendix 3 Analysing the leading indicator properties of monthly PCA FCI Table A3.1. Eigenvalues Number Value Difference Proportion Cumulative value Cumulative proportion 1.87.44.3.87.3.43 1.35.7 5.3.59 3 1.9..1 6.39.71 4.87.17.1 7.6.81 5.7.6.8 7.96.88 6.44.8.5 8.4.93 7.36..4 8.76.97 8.16.8. 8.9.99 9.8 -.1 9. 1. Table A3.. Granger causality test Null hypothesis F-stat (probability) GDP qoq does not Granger cause FCI.3 (.59) FCI does not Granger cause GDP qoq 5.19 (.) GDP qoq does not Granger cause exogenous FCI.3 (.86) Exogenous FCI does not Granger cause GDP qoq 5.76 (.) Table A3.3. Forecast performance of GDP FCI OLS Adjusted R-squared.1.53.53 RMSE 1.58 1.1 1.1 MSE 1.13.8.9 NATIONAL BANK OF ROMANIA 9

Occasional Papers No.17 Chart A3.1. FCI and GDP 6 4 - -4 % 1.5 1..5. -.5-1. -6-8 Correlation coefficient:.57 5Q1 5Q 5Q3 5Q4 6Q1 6Q 6Q3 6Q4 7Q1 7Q 7Q3 7Q4 8Q1 8Q 8Q3 8Q4 9Q1 9Q 9Q3 9Q4 1Q1 1Q 1Q3 1Q4 11Q1 11Q 11Q3 11Q4 1Q1 1Q 1Q3 1Q4 13Q1 13Q 13Q3 13Q4 14Q1 14Q 14Q3 14Q4-1.5 -. GDP_qoq FCI_PCA_quarterly average (-1)(rhs) - exogenous Souce: NBR, NIS, Bloomberg, European Commission, author's calculations Appendix 4 Analysing the leading indicator properties of quarterly DFM FCI Table A4.1. Granger causality test Null hypothesis F-stat (probability) GDP qoq does not Granger cause FCI.15 (.7) FCI does not Granger cause GDP qoq 17.59 (.) Table A4.. Forecast performance of GDP FCI OLS Adjusted R-squared.1.46 RMSE 1.58 1.3 MSE 1.13.95 3 NATIONAL BANK OF ROMANIA

July 15 Appendix 5 Analysing the leading indicator properties of monthly DFM FCI Table A5.1. Granger causality test Null hypothesis F-stat (probability) GDP qoq does not Granger cause FCI 1.1 (.31) FCI does not Granger cause GDP qoq 6.63 (.) GDP qoq does not Granger cause exogenous FCI.3 (.86) Exogenous FCI does not Granger cause GDP qoq 7.8 (.) Table A5.. Forecast performance of GDP FCI OLS Adjusted R-squared.1.54.54 RMSE 1.58 1.8 1. MSE 1.13.8.9 Chart A5.1. FCI and GDP 6 4 - -4 % 1.5 1..5. -.5-1. -6-8 Correlation coefficient:.58 5Q1 5Q 5Q3 5Q4 6Q1 6Q 6Q3 6Q4 7Q1 7Q 7Q3 7Q4 8Q1 8Q 8Q3 8Q4 9Q1 9Q 9Q3 9Q4 1Q1 1Q 1Q3 1Q4 11Q1 11Q 11Q3 11Q4 1Q1 1Q 1Q3 1Q4 13Q1 13Q 13Q3 13Q4 14Q1 14Q 14Q3 14Q4-1.5 -. GDP_qoq FCI_DFM_quarterly average(-1)(rhs) - exogenous Souce: NBR, NIS, Bloomberg, European Commission, author's calculations NATIONAL BANK OF ROMANIA 31

Occasional Papers No.17 3 NATIONAL BANK OF ROMANIA