Macro-Financial Linkages in Norway:
An Analysis of Non-Performing Loans and Oil Price Shocks
Mikkel von der Fehr1
Final version: August 13th, 2017
Abstract
In this paper, we assess how an oil price shock propagates through the Norwegian economy and what role the banking sector plays during such a shock. First, we investigate the macroeconomic determinants of Non-Performing Loans (NPLs) using Dynamic Panel Data models, and confirm that adverse macroeconomic developments are associated with rising NPLs. Second, we investigate the feedback between NPLs and their macroeconomic determinants in several panel vector autoregressive (panel VAR) models, analysing on a regional and business model level. Our findings indicate that, although the oil price is not a significant determinant of the NPLs, an oil price shock has a clear impact on the Norwegian economy, and banks’ NPLs, especially so for the commercial banks. Furthermore, we identify a key risk factor for the Norwegian banks, and, due to the observed macro-financial linkages, the overall economy, in the development of the housing market. Our results support the assumption that disturbances in the banking systems lead to unwanted economic consequences for the real economy.
Keywords: Macroeconomic Shocks, Banks, Non-Performing Loans, Panel Vector Autoregression. JEL classifications: C63, E44, E51, G21, G28, G32, Q43.
1 First and foremost, I would like to thank dr. Beetsma for insightful discussion and guidance throughout the writing of this
paper – by challenging my thoughts and ideas I have been given the opportunity to write a more complete and thorough paper, and it has ultimately made me a better economist. Likewise, I would like to thank the University of Amsterdam, and especially dr. Romp, dr. Houben and dr. Mavromatis, for providing an excellent master’s programme in economics that encourages intellectual curiosity and academic excellence. Furthermore, I would like to thank Sissel Krossøy, Assistant Director at the Norwegian Banks’ Guarantee Fund, for providing the dataset of the NPLs for all its members – without it this paper would not have been possible – and moreover, for answering all my inquiries thereafter. Last, but not least, I am thankful for all the helpful comments on earlier versions and, at times, the much-needed moral support from Irina Isakova, and, of course, my mother, Marianne von der Fehr, for encouraging and supporting me to achieve my goals.
2
Statement of Originality
This document is written by Student Mikkel von der Fehr 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 completion of the work, not for the contents.
3 Table of Contents
1 Introduction & Research Question ... 4
2 Literature Review ... 7
3 Data Description ... 11
3.1 Variable Description & Data Sources ... 11
3.2 Stylised Facts about the Dataset ... 11
4 Methodology ... 19
4.1 Specification of the Dynamic Panel Data Model ... 19
4.2 Specification of the Panel VAR Model ... 22
5 Results & Analysis ... 24
5.1 Results & Analysis of the Dynamic Panel Data Model ... 24
5.2 Results of the Panel VAR Model ... 29
5.2.1 Regionality & Business Model ... 34
5.2.2 Regional Results & Analysis ... 34
5.2.3 Business Model Results & Analysis ... 36
5.2.4 Robustness of the Panel VAR ... 36
6 Conclusion & Policy Implications ... 38
7 References ... 41
8 Appendix ... 44
8.1 Dynamic Panel Data, Baseline Model ... 44
8.2 Dynamic Panel Data, Commercial Banks ... 45
8.3 Panel VAR, Baseline Model ... 47
8.4 Panel VAR, Region East ... 49
8.5 Panel VAR, Region West... 52
8.6 Panel VAR, Commercial Banks ... 54
8.7 Panel VAR, Savings Banks ... 57
8.8 Panel VAR, Different Ordering ... 60
4
1 Introduction & Research Question
Norway’s petroleum era started more than 50 years ago, with Phillips’ discovery of Ekofisk, one of the largest offshore oil fields ever discovered. The petroleum activities on the Norwegian Continental Shelf (NCS) have played a key role in the development of today’s welfare state. When the first licenses were awarded in the mid-1960’s, few realised the potential and impact of the industry that would come to dominate the Norwegian economy, as seen in the chart of the industry’s share of state’s total value creation, investments, exports and revenues below (Ministry of Petroleum and Energy, 2017). The substantial fall in the oil price since the autumn of 2014 serves as a reminder of the inherent risks of having such a dominant industry (Olsen, 2017).
Figure 1: Macroeconomic indicators for the petroleum sector, 1971-2016
Source: norskpetroleum.com
As a small open economy, Norway is exposed to the global economy. With moderate growth among its main trading partners, low oil price and persistently low interest rates, the Norwegian economy faces challenging times (Olsen, 2017). However, the relationship between oil prices and the Norwegian economy may not be quite as clear-cut as it seems; A low oil price might have a positive effect for the global economy, as some believe that there is a difference in saving behaviour2 between
oil importers and oil exporters, and therefore, as Norway is a small open economy, the oil price could have a positive impact on it (Obstfeld, Milesi-Ferretti, & Arezki, 2016).
2 Consumers in oil importing regions, such as Europe, are believed to have a higher marginal disposition to consume their
income than those in oil exporting countries, such as Saudi Arabia.
0.0 10.0 20.0 30.0 40.0 50.0 60.0 (%)
Share of GDP Share of Investments
5
Figure 2: Brent3 oil price, in USD/barrel
Source: eia.gov
The economic developments in Norway are reflected in the reduced activity levels in the oil sector and in the oil-related industries. The steep fall in the oil price since autumn 2014 (see Figure 2) has brought market impairment in the profits and financial positions of the companies in the petroleum sector – oil companies themselves, as well as, for example, those involved in rig supply and seismology. Moreover, several of the largest Norwegian banks have considerable exposure to these industries. However, Norwegian banks appear to be strong and resilient to the negative shock of the oil price. The direct exposure to oil-related industries is low, for instance oil-related loans account only for around 5% of total loans of the 16 largest banks. At the end of 2015, the banking sector had an exposure to oil-related industries around NOK 200bn (EUR 22.67bn). Furthermore, Norwegian banks’ loan losses have been low for a long time, yet, should the downturn in the oil price be persistent, the losses may increase considerably. With a persistently low oil price, the oil-related industries might find themselves in the same situation as the shipping companies did during the aftermath of the financial crisis of 2008; Norwegian banks’ cumulative losses on loans to the shipping industry in the period 2008–2014 were equal to approximately 10% of loans to this industry (Norges Bank, 2016).
The developments discussed above invite further investigation. How does the oil price propagate through the financial system and feed back into the real economy? By modelling and analysing the macro-financial linkages, the interaction between the financial sector and the domestic economy, this paper aims to answer the following question:
3 Brent Oil, or Brent Blend, is the most widely used marker of all, accounting for roughly two-thirds of all crude contracts
around the world. “Brent” refers to oil from four different fields in the North Sea: Brent, Forties, Oseberg and Ekofisk. 0 20 40 60 80 100 120 140 (U SD )
6
How does an oil price shock propagate through the economy and what is the role of the banking sector in this regard?
To answer the question, we follow the common ‘dual’ approach of the related literature. Firstly, the factors that explain non-performing loans4 (NPLs) are identified using a Dynamic Panel Data model
whilst controlling for bank-specific effects, and, secondly, the interactions between the oil price, NPLs and other macroeconomic variables are uncovered by a panel vector autoregression model. The two approaches, besides being common in the literature, are complementary, where the results of the panel vector autoregression are a useful robustness check of the Dynamic Panel Data results. To our knowledge, this paper represents the first attempt to assess the macro-financial linkages for the Norwegian economy using macroeconomic and bank-level data.
When interpreting the results of this paper, a range of caveats should be kept in mind. First, analyses based on historical data might not account for the effects of recent policy changes, such as the changes to the lending policies of the Norwegian banks. Second, with data spanning the period of 2002-2016, a sufficient number of cycles, such as the oil price and financial, might not be captured. Lastly, there is considerable parametric uncertainty regarding the estimated relationship between macroeconomic shocks and NPL ratios (Miyajima, 2016).
The remainder of the paper is organised as follows. Section 2 provides a brief overview of the relevant literature. In Section 3 an overview of the data is provided, as well as some stylised facts about the dataset. Section 4 describes the methodology, with a subsection for each of the models. Following the sections on methodology and dataset, Section 5 covers the estimation and results, and consequently analysis of the results. Section 6 concludes the paper and brings a brief discussion of the implications of the results on future policy.
4 Non-performing loans are defined as loan that are non-performing over 90 days as a percentage of total loans at end of
7
2 Literature Review
The topic of macro-financial linkages in academic studies was rather underdeveloped up until the turn of the century; Doing an advanced search in Google Scholar results in 82 articles for the period 1980– 1999, and 2,700 for the period 2000–2016. Although merely anecdotal evidence, the formal research provides support for this finding. The topic of macro-financial linkages, and especially the concept of feedback loops, has become more pronounced following the 2008 Global Financial Crisis, as there was a need to review and reassess the transmission channels and mechanisms for monetary policy due to policy makers and academia facing a new “reality”. This new “reality” has shown that there are frictions in the financial market, and that these frictions can lead to increased costs in the real economy (Tetangco, 2016). One of the earliest attempts of capturing these frictions, or in other words, the two-way linkages between financial stability and macroeconomic performance, can be found in “The financial accelerator in a quantitative business cycle framework” by Bernanke et al. (1999). Bernanke et al. propose a theoretical model with financial acceleration, a process by which adverse shocks are amplified by worsening conditions in the financial markets, with the aim of capturing the links between incomplete financial markets and the real economy. The paper provides an excellent insight into how credit frictions, that are endogenously determined, propagate disturbances and how they ultimately spread to the macroeconomy. Our paper seeks to further investigate the insights provided by Bernanke et al. through empirical exploration of how an oil price shock propagates through the Norwegian real economy, and how such a shock impacts the credit risk through the proxy of NPLs. Thus, the relevant literature for this paper includes academic works on:
A. the assessment of bank asset quality, i.e. the determinants of non-performing loans (NPLs), as a measurement for credit risk in the banking system;
B. the feedback relationship between the financial instability in banking systems and the real economy.
The literature on the determinants of NPLs commonly splits them into two key groups: macroeconomic and bank-specific determinants. The macroeconomic determinants of NPLs include variables such as business cycles, unemployment rates, housing prices and interest rates, and influence the balance sheets of the banks and the capacity of debt-service. The bank-specific determinants of NPLs include size of the banks, business model, operational costs and risk management, and vary across banks. Further review of the literature on the two key groups of determinants of NPLs can be found in Nkusu (2011), Espinoza and Prasad (2010), and Klein (2013).
8
Among the literature covering asset quality assessment, Nkusu (2011) is essential. Here, country-level data spanning 1998–2009 for 26 advanced economies is analysed, and it is concluded that rising NPLs are associated with adverse macroeconomic developments. Furthermore, Nkusu (2011) shows that when NPLs increase sharply, the performance of the macroeconomy slows down and the economic growth halts. This conclusion is further elaborated on in De Bock and Demyantes (2012), where 25 emerging markets during 1996–2010 are analysed. De Bock and Demyantes find evidence of NPLs increasing when there is a decline in economic growth, weakened exchange rate, and deterioration of the terms of trade.
Using data spanning 1998-2011, Klein (2013) analyses the ten largest banks in 16 countries5 in Central,
Eastern and South-Eastern Europe (CESEE). Klein finds using both macroeconomic and bank-specific factors that macroeconomic conditions, particularly GDP growth, unemployment and inflation, are comparatively more important than bank-specific factors in explaining the changes in NPLs across the CESEE region. Likewise, Messai and Jouini (2013), covering Italy, Greece and Spain, conclude that an increase in the growth of GDP, as well as declining unemployment rates, lowers the credit risk. An earlier study by Louzis et al., (2012) which investigates the determinants of NPLs in the Greek banking system, echoes these conclusions as it finds the NPLs to largely be explained by GDP growth, unemployment rate, and the level of public debt.
Miyajima (2016) analyses oil-macro-financial linkages in Saudi Arabia by applying panel econometric frameworks to macroeconomic and bank-level data, which makes this paper highly relevant. Miyajima shows that a decrease in the growth of oil prices and non-oil GDP leads to a slowdown in the credit and deposit growth and higher NPL ratios. Furthermore, the existence of feedback loops between bank balance sheets and economic activity is uncovered, and a decline in the oil prices is proven to have a negative effect on both.
Similarly, Espinoza and Prasad (2010), investigating economies in the Gulf Cooperation Council6, find
that the ratio of NPLs increases as economic growth declines, and that the interest rates, as well as the overall risk aversion, increase consequently. However, there is a caveat to Espinoza and Prasad’s findings; Their dataset predates the financial crisis of 2008, and furthermore, does not include the oil price. Given that the countries in the Gulf Cooperation Council are predominantly oil exporting economies, one can imagine that the oil price plays a major part in determining the NPLs across the
5 Total of 160 banks.
9
region. Alodayani (2016) investigates the effect of the 2014–2015 slump in the oil price on the financial stability in the Gulf Cooperation Council region, making it another highly relevant paper. The results indicate that the oil price, non-oil GDP, interest rate, stock prices, and housing prices are major determinants of NPLs and the overall financial stability, which supports the view that disturbances in the banking systems lead to undesirable consequences for the real economy. Adhering to the same region, Love and Ariss (2013) find that increased capital inflows and stronger GDP growth improve banks’ loan portfolio quality using a panel of Egyptian banks covering the time-period of 1993-2010. Furthermore, the authors find that an increase in lending rate may lead to adverse selection, and decreased quality of the banks’ portfolio. Love and Ariss also find evidence that a larger market share of foreign banks improves loan quality.
Marcucci and Quagliariello (2009) investigate how Italy’s different credit risk regimes affect the credit risks and the business cycles using quarterly data from the period 1989-2005. The authors find that the business cycles’ effect on credit risks is noticeably stronger in times with weakened financial conditions, which supports the idea of a strong link between the financial sector and real economy, as made evident by the financial crisis. In a prior study from 2008, Marcucci and Quagliariello examine the cyclical behaviour of default rates of borrowers on Italian banks. The authors find that the default rates are highly cyclical, with decreases during economic booms and increases during recessions – a rather intuitive result. Furthermore, Marcucci and Quagliariello find that the banks’ capital acts as a feedback channel from the banking sector to the real economy. Similarly, to Marcucci and Quagliariello, Jiménez & Saurina (2006) find, through examining the lending cycle and credit standards, empirical evidence of the fact that the banking sector in economic booms are more lenient in screening and requirements of its borrowers. Furthermore, the authors find evidence of a positive, but quite lagged relationship between credit growth and NPLs.
Sorge (2004) reviews macro stress-testing methodologies, and documents a variety of studies that investigate the feedback effects between credit losses and the macroeconomy. In this paper, Sorge argues that assessing the true nature of the feedback is in general difficult, as the effect is obscured by the direct effect from economic growth to NPLs and the banks’ balance sheets. However, Sorge denotes that a solution to this issue is to identify a supply shock. Drehmann (2008) gives an excellent survey of studies that link the real and financial sector, and the difficulties in modelling the linkages. In the section ‘Macro feedbacks’, Drehmann highlights Jacobson et al. (2005); Albeit not modelling banks explicitly, Jacobsen et al., sets up a panel VAR model that incorporates macroeconomic factors, as well as the likelihood of default for Swedish companies. The authors find evidence of
10
macroeconomic feedbacks, and that these feedbacks have important implications, such as creating conflicts between the objectives of monetary policy and financial stability.
This paper builds upon the emerging trend of the literature where both macroeconomic and bank-level data are used to examine and assess macro-financial linkages. In contrast to Espinoza and Prasad (2010) and Nkusu (2011), where aggregate levels of NPLs where employed and bank-specific effects not accounted for, we propose to estimate the model with bank-level data and introduce fixed effects to account for differences in activity and business models. By basing the methodology on Love and Zicchino (2006), Blundell and Bond (1998), and Canova and Ciccarelli (2013), we specify two models, a Dynamic Panel Data model to identify the determinants of NPLs and a panel vector autoregressive model to assess the feedback effects among macroeconomic and bank-level variables. For further discussion, regarding the usage of two models, see Section 4 on methodology.
11
3 Data Description
We retrieve annualised bank-level financial data for all 126 members of The Norwegian Banks’ Guarantee Fund in a period of 15-years, from 2002 to 2016. All Norwegian banks must be members of the fund, while it is voluntary for branches of foreign banks. Branches of foreign banks who opt into the fund have so-called “topping up" membership, i.e. they are covered by an equivalent domestic fund or scheme. However, due to different legislation the Norwegian Banks’ Guarantee Fund covers up to NOK 2 million, or around €210 000, which is above the equivalent €100 000 for European banks, and as such the branches “top up” the excess (The Norwegian Banks’ Guarantee Fund, 2017). The data is processed and classified based on business model and localisation of the individual bank.
Furthermore, we collect macroeconomic data on Norway for the same period from a variety of sources, including Statistics Norway, Norges Bank (Norway's central bank), OECD, and U.S. Energy Information Administration (EIA). Our variables of interest are based on the approach of related literature; however, our variables are considered on the fit with the characteristics of the Norwegian economy. Following similar literature, oil price, non-oil GDP, interest rate, credit growth, stock prices and housing prices are all considered.
3.1 Variable Description & Data Sources
A brief description of the variables, timespan, definition, unit and information sources can be found in Table 1Error! Reference source not found. on the next page.
3.2 Stylised Facts about the Dataset
After removing banks that have incomplete time series, due to their incorporation after 2002, as well as banks that pose NPLs of either one or zero7, the dataset is left with NPL ratios for 103 banks. The
regional split is as follows: Region East:8 46 banks; Region West:9 24 banks; Trøndelag10 17 banks;
Region North:11 7 banks; Region South:12 9 banks. There are 8 commercial banks and 95 savings banks
in the dataset. A quantitative breakdown of the regionality and business model of the bank can be seen in Table 2 and a graphical breakdown can be found in Figure 3.
7 Done due to computational issues.
8 Region East: Telemark, Buskerud, Hedmark, Oppland, Akershus, Oslo, Vestfold, Østfold.
9 Region West: Møre og Romsdal, Sogn og Fjordane, Hordaland, Rogaland.
10 Trøndelag: Nord-Trøndelag, Sør-Trøndelag.
11 Region North: Finnmark, Troms, Nordland.
12
Table 1: Variable description and data sources
Variable Time Definition Unit Description Source Website Date
NPL 2002 - 2016 Non-performing loans
Ratio Non-performing loans, i.e.
non-performing over 90 days as a percentage of total loans. The Norwegian Banks’ Guarantee Fund - 23.03.17 OPG 2002 - 2016 Inter-national oil price U.S. Dollar
Brent Oil Price - Spot U.S. Energy
Information Administration (EIA) http://bit.ly/ 2fluXA2 24.03.17 NOG 2002 - 2016
Non-oil GDP % Domestic production
activity except for
exploration of oil and natural gas pipelines and ocean transport. Statistics Norway http://bit.ly/ 2vblJMq 24.03.17 IRG 2002 - 2016 Key policy rate
% Key policy rate set by the
Norwegian central bank, Norges Bank
Norges Bank http://bit.ly/
2wyZ3TX
24.03.17
CRG 2002
- 2016
Gross loans % Change in gross loans Statistics
Norway http://bit.ly/ 2vblJMq 24.03.17 SPG 2002 - 2016 Stock price index
Index Development of the stock
market price index
OECD http://bit.ly/ 2vwjN1N 24.03.17 HPG 2002 - 2016 Housing price index
Index Development of the value
of the total housing stock
Statistics Norway
http://bit.ly/ 2vblJMq
24.03.17
Table 2: Breakdown of Norwegian banks based on location and business model
Region Number of Banks Share Commercial Savings
East 46 45% 6 40 West 24 23% 1 23 Trøndelag 17 17% 1 16 North 7 7% 0 7 South 9 9% 0 9 Total 103 100% 8 95
13
Figure 3: Graphical breakdown of Norwegian banks based on the regionality
Source: The Norwegian Banks’ Guarantee Fund, own calculations
The NPLs of the Norwegian banks, using 2002 as the baseline year, have on average decreased with 1.20 percentage points over the time period of the dataset. Region West has experienced the largest decline in NPLs during 2002-2016, with a drop of 1.55 percentage points, with Region North following closely at 1.50 percentage points. The lowest decrease in NPLs over the whole period can be found in Region South, where the NPLs have dropped with 0.40 percentage points, however, this is not a surprising finding as this region experiences the least variation in NPLs. Year over year, we observe that the largest decrease in NPLs happens in Trøndelag, with a drop of close to one percentage point between 2002 and 2003. The largest increase in NPLs is a two-way split between Region East and Region North, both from 2007 to 2008, coinciding with the Global Financial Crisis, with a rise of almost 0.50 percentage point. Besides the two aforementioned regions, all other regions experience their largest increase during the Global Financial Crisis, however the increase is lagged with one year, from 2008 to 2009, with an upturn of about 0.30 percentage points for all regions. A numerical overview, as well as a graphical representation, of the regional NPLs can be found in Table 3Error! Reference source not found., and Figure 4, respectively.
Table 3: Overview of average NPLs sorted by region (2002-2016)
Regions ‘02 ‘03 ‘04 ‘05 ‘06 ‘07 ‘08 ‘09 ‘10 ‘11 ‘12 ‘13 ‘14 ‘15 ‘16 East 2.10 1.90 1.27 1.03 1.00 0.85 1.32 1.53 1.46 1.35 1.28 1.10 1.05 0.77 0.61 West 1.59 1.31 0.97 0.75 0.67 0.58 0.82 1.08 1.15 1.00 0.88 0.96 0.79 0.77 0.65 Trøndelag 2.12 1.17 0.72 0.96 0.58 0.54 0.79 0.71 0.64 0.52 0.53 0.56 0.59 0.61 0.59 North 1.59 1.77 1.58 1.64 1.10 0.73 1.21 1.46 0.74 0.74 0.71 1.02 0.60 0.66 0.61 South 0.04 0.88 0.58 0.51 0.49 0.43 0.53 0.82 0.91 0.79 0.68 0.75 0.60 0.55 0.53
Source: The Norwegian Banks’ Guarantee Fund, own calculations Region East Region West Trøndelag Region North Region South
14
Figure 4: Overview of average NPLs sorted by regional location (2002-2016)13
Source: The Norwegian Banks’ Guarantee Fund, own calculations
Analysing the development of individual banks, we find that the largest annual increase in NPLs is experienced by Storebrand Bank ASA, a commercial bank located in Region East, with an increase of about 6.50 percentage points between 2002 and 2003. This finding is indicative of a general trend with commercial banks; their NPLs are on average, accounting for the entire time period, 0.75 percentage points higher when compared to NPLs of savings banks. In fact, there is not a singular instance where the NPLs of the savings banks are, on average, above the NPLs of the commercial banks. Furthermore, this trend holds true when looking at the data for the individual banks, as Hjelmeland Sparebank, a savings bank located in Region West, poses the largest increase for any savings bank, with an increase of 3.70 percentage points between 2007 and 2008. The opposite relationship also holds true, with Santander Consumer Bank AS, a commercial bank located in Region East, experiencing a drop in NPLs of 7 percentage points between 2004 and 2005, whereas Gildeskål Sparebank, a savings bank located in Region North, experienced a drop in NPLs of slightly above 3 percentage points between 2009 and 2010. A numerical overview, as well as a graphical representation, of the average NPLs based on the business model can be found in Table 4, and Figure 5, respectively.
Table 4: Overview of average NPLs by business model (2002-2016)
B. Model ‘02 ‘03 ‘04 ‘05 ‘06 ‘07 ‘08 ‘09 ‘10 ‘11 ‘12 ‘13 ‘14 ‘15 ‘16
Com. 4.05 3.93 2.41 1.03 1.01 0.85 1.57 2.45 1.98 1.37 1.18 1.18 1.09 0.97 0.84
Sav. 1.64 1.37 0.96 0.97 0.80 0.66 1.01 1.11 1.12 1.03 0.97 0.90 0.82 0.69 0.59
Source: The Norwegian Banks’ Guarantee Fund, own calculations
13 All numbers in %. 0.0 0.5 1.0 1.5 2.0 2.5 (%)
15
Figure 5: Overview of average NPLs by business model (2002-2016)
Source: The Norwegian Banks’ Guarantee Fund, own calculations
The non-oil GDP has fallen from 1.50 percent in 2002 to 0.90 percent in 2016, with a negative cumulative change over the entire period of around 0.60 percentage points. The period 2004 to 2005 saw an average growth of slightly over 5 percent, with a rise of 3.80 percentage points between 2003 and 2004. Following the Global Financial Crisis, the non-oil GDP saw a negative growth of 1.60 percent (2009).
The interest rate has, using 2002 as the baseline year, fallen around 6.20 percentage points, and was as of 2016 at 0.55 percent. 2008 saw the highest interest rate with 5.32 percent, and it has been decreasing ever since 2011, reaching its lowest point in 2016.
Over the entire period of the dataset, using 2002 as the baseline year, the credit growth has decreased with about 4 percentage points. The maximum increase in credit growth, of 3.25 percentage points, occurred between 2005 and 2006, and the maximum decrease, slightly over 6 percentage points, between 2008 and 2009.
The oil price has, using 2002 as the baseline year, cumulatively increased by close to 113 percent over the period 2002 to 2016, with the highest cumulative growth taking place up until 2012 (≈188%). The largest increase happened in 2004 to 2005 with a growth of almost 43 percent, and the largest decline is slightly above 47 percent, following the drop from 2014 to 2015. There has not been any growth in the oil price year over year since 2012.
Using 2002 as the baseline year, the stock market prices have experienced a steady increase over the whole period from 2002 to 2016, resulting in a cumulative growth above 190 percent. The largest
0.0 1.0 2.0 3.0 4.0 5.0 (%) Commercial Savings
16
year-on-year fall in the prices happened between 2008 and 2009, with a decrease of almost 25 percent, and the largest jump was from 2003 to 2005, with annual increase of about 50 percent. The housing prices have followed a similar pattern as the stock market prices, with a cumulative increase slightly above 90 percent over the entire period. The largest annual rise took place between 2005 and 2006, close to 14 percent, and the largest drop followed two years later, between 2007 and 2008, with a negative growth in the prices slightly above one percent.
In Figure 6 and Figure 7, an overview of the macroeconomic variables, in both levels (oil price, stock price index and housing price index), and percentages (non-oil GDP, interest rate and credit growth) is given.
Figure 6: Overview of the macroeconomic variables that are in levels (2002-2016)
Source: EIA, OECD & Statistics Norway
Figure 7: Overview of the macroeconomic variables that are in percentages (2002-2016)
Source: Statistics Norway & Norges Bank -4.0 0.0 4.0 8.0 12.0 16.0 (%)
Non-Oil GDP Interest Rate Credit Growth 0 50 100 150 200 250
17
In Section 5.1, as well as in Section 5.2, the variables used are both in differences and growth rates. Variables that come in levels in the original dataset, i.e. the oil price, stock index price and housing index price, are transformed into growth rates using the standard procedure of subtracting the past value from the present value, and dividing by the past value. Values that are growth rates are denoted with ‘G’ in the end of the variable name, and kept to three characters, for example the oil price growth is named OPG. The other variables, such as credit growth, were in percentages when obtained, and thus these variables were transformed using the standard method of taking the difference, i.e. subtracting the past value from the present value. A similar naming-scheme to the growth rates is performed, however, here the differences is denoted with ‘d’ for delta – the common mathematical notation – in front of the variable, for example the difference of the non-performing loans becomes dNPL. Summary statistics over the variables, pre- and post-transformation can be found in the Table 5, and Table 6, respectively.
Table 5: Summary Statistics, pre-transformation
Variable Type Observations Mean Std. Dev. Min Max
-NPLt -Percent -1545 -1.029 -0.993 -0.010 -13.750 -OPt -Level -1545 -69.927 -29.041 -24.990 -111.570 -NOt -Percent -1545 -0.025 -0.019 -0.016 -0.057 -IRt -Percent -1545 -0.026 -0.017 -0.006 -0.067 -CRt -Percent -1545 -0.083 -0.032 -0.048 -0.143 -SPt -Level -1545 -101.807 -39.651 -33.100 -157.222 -HPt -Level -1545 -135.022 -34.964 -82.475 -195.475
Source: own calculations
Table 6: Summary Statistics, post-transformation
Variable Type Observations Mean Std. Dev. Min Max
-dNPLt -Difference -1442 -0.001 -0.007 -0.071 -0.065 -OPGt -Growth -1442 -0.080 -0.270 -0.471 -0.426 -dNOt -Difference -1442 -0.000 -0.021 -0.040 -0.038 -dIRt -Difference -1442 -0.004 -0.014 -0.036 -0.016 -dCRt -Difference -1442 -0.003 -0.022 -0.061 -0.033 -SPGt -Growth -1442 -0.137 -0.223 -0.248 -0.529 -HPGt -Growth -1442 -0.064 -0.041 -0.011 -0.137
Source: own calculations
Assuming that the macroeconomic variables included in both models, presented in Section 4.1 & 4.2, can be correlated, pairwise correlations of the transformed variables are computed and analysed. From Table 7 below, we observe that all variables are either weakly or moderately positively correlated (correlation is statistically significant), except for the pairwise correlation between the
18
differenced interest rate and the difference non-oil GDP. Moreover, there is strong positive correlation between the differenced credit growth and the differenced interest rate, the stock price index growth and the differenced credit growth, as well as the housing price index growth – which is strongly correlated both ways. From an economic perspective these relationships seem intuitive, for example a lower interest rate is indicative of a reduced cost of money, which should lead to an increase in credit growth. Moreover, the fact that the stock price index growth and the differenced credit growth are correlated can be explained with yield seeking behaviour in times of credit expansion, in other words, the increase in credit flows into the stock market. Lastly, the pairwise correlation between the housing price index growth and the stock price index growth, and vice versa, is indicative of the strength of the overall economy, and perhaps not a direct relationship. For instance, in a strong economy, the demand for products and services, such as housing and financial investments, increases, which leads to an increase in both variables, making the correlation positive.
Table 7: Pairwise Correlation between macroeconomic variables
OPGt dNOt dIRt dCRt SPGt HPGt
OPGt 1.000 dNOt 0.347*** 1.000 dIRt 0.334*** 0.025 1.000 dCRt 0.475*** 0.313*** 0.580*** 1.000 SPGt 0.496*** 0.702*** 0.275*** 0.641*** 1.000 HPGt 0.263*** 0.653*** 0.364*** 0.599*** 0.756*** 1.000
*, **, and *** denote significance at 10 percent, 5 percent, and 1 percent, respectively.
19
4 Methodology
The methodology follows the common ‘dual’ approach that can be found in the literature (Espinoza & Prasad, 2010; Nkusu, 2011). The aim of this approach is firstly to identify factors that explain NPLs and secondly, to assess the interactions among NPLs and macroeconomic variables. Key determinants of bank NPLs are identified using a Dynamic Panel Data model whilst controlling for bank-specific effects, and the panel vector autoregressive (panel VAR) framework presented in Love and Zicchino’s 2006 paper that is used to identify how variables in the system respond to a shock affecting other variables. The latter approach enables an analysis of the various channels the shock transmits itself, for instance through a regional level or on a bank business model level (Love & Ariss, 2013). The two approaches are complementary, where the panel VAR approach is a useful robustness check of the Dynamic Panel Data model approach (Miyajima, 2017). This complementary relationship is highlighted in the conclusion of Miyajima (2016):
“First, results from a panel multivariate model revealed that lower growth rates of oil prices and non-oil private sector GDP lead to a rise in NPL ratios, representing higher bank solvency risk. Second, results from a panel VAR model suggested that higher weaker macroeconomic conditions (lower growth of oil prices and non-oil private sector GDP) lead to weaker bank balance sheet conditions (higher NPL ratios, lower deposit and credit growth), which feedback to further weaken macroeconomic conditions.”
4.1 Specification of the Dynamic Panel Data Model
As noted in the opening paragraph, the first model identifies the key determinants of the Norwegian bank’s NPLs by employing a Dynamic Panel Data model while controlling for bank-specific effects. Two alternative econometric techniques are considered to estimate the model: (i) Panel Fixed Effect model; and (ii) System GMM Model.
The Dynamic Panel Data model is estimated using Panel Fixed Effect, as it removes unobserved heterogeneity across the banks. However, Panel Fixed Effect has a limitation when the lagged dependent variable is included, in other words, the fixed effect model suffers from dynamic panel bias, often referred to as Nickell bias, a result of correlation between the error term and the lagged dependent variable. This issue is avoided by estimating using the GMM model.
Employing the econometric technique System GMM outlined in Blundell and Bond (1998), the issue of dynamic panel bias is avoided. In their paper, Blundell and Bond focus their efforts on estimating a Dynamic Panel Data model using two alternative linear estimators, so that the properties of the
20
standard first differenced Generalised Method of Moments (GMM) estimator can be improved. This is done as “Dynamic Panel Data models where the autoregressive parameter is moderately large and the number of time series observations is moderately small, the (…) linear GMM estimator obtained after first differencing has been found to have large finite sample bias and poor precision in simulation studies”.
Following from similar research, there is reason to believe that the NPL ratios exhibit strong autocorrelation, in other words the NPLs will be correlated with a lag of themselves, an issue the Dynamic System GMM resolves. Furthermore, the issue of having an unbalanced panel, where the time dimension (T) is short relative to its cross-sectional dimension (N), is also solved by using System GMM. Consequently, System GMM is employed and estimated in this paper, following the techniques provided by Roodman (2007).
The determinants of the NPL ratios are estimated using the following Dynamic Panel Data specification for bank 𝑖 in year 𝑡:
∆NPLi,t= NPLi,t−1+ ∑ α2jMacroj,t−1
j
+ ∑ α3kBankk,i,t−1
k
+ α4dummy08/09+ θi+ εi,t (1)
Where ∆NPLi,t is the difference between the share of non-performing loans to total loans in the current
period and the previous period, with the lagged regressor ∆NPLi,t-1 capturing the persistence that is
commonly found in the literature. The growth rate transformation of the NPLs ensures that the dependent variable spans the interval (-∞; +∞), as opposed to between 0 and 1, and that the dependent variable is distributed symmetrically. For a graphical representation of this see Figure 8Error! Reference source not found..
21
Figure 8: Density distribution, NPL
Source: The Norwegian Banks’ Guarantee Fund, own calculations
Macroj,t-1 represents the macroeconomic variables lagged by one period. The variables (k = 1,2,3,4,5)
are as follows: oil price growth, non-oil GDP, interest rate, stock and housing index prices growth, where the oil price, stock and housing index prices are growth rates, and the non-oil GDP and interest rate are differences.
Bankk,i,t-1 represents real credit growth (k = 1) lagged by one period, where the credit growth is in
difference.
A dummy variable, dummy08/09, is introduced to control for the events of the Global Financial Crisis, as
there is a potential effect of increasing NPL rations stemming from the events of the Global Financial Crisis. Accordingly, θi and εi represent the usual (bank) fixed effects and random errors, where the
errors are independent and identically distributed (i.d.d.).
To address concerns regarding the presence of unit roots in our time-series, we conduct panel unit root tests. The Fisher Augmented Dickey-Fuller (ADF) and Fisher Phillips-Perron (PP), where the null hypothesis is that all series are non-stationary and the alternative hypothesis is that at least one of the series in the panel is stationary, does not require a balanced sample. Furthermore, we report the inverse normal Z statistic, as it offers the best trade-off between size and power (Choi, 2001). The results can be seen in Table 8 below.
Table 8: Fisher Panel Unit Root Tests
Variable Augmented Dickey-Fuller Phillips-Perron
dNPL -35.238*** -35.238***
22 dNO -25.559*** -25.559*** dIR -21.194*** -21.194*** dCR -7.749*** -7.749*** SPG -10.661*** -10.661*** HPG -17.865*** -17.865***
Source: own calculations
dNPL is the delta of the NPLs in percent; OPG is the growth rate of the oil price; dNO is the differenced
non-oil GDP; dIR is the differenced interest rate; dCR is the differenced credit growth; SPG is the growth rate of the stock prices index; HPG is the growth rate of the housing price index. *, **, and *** denote significance at 10 percent, 5 percent, and 1 percent, respectively
The Fisher ADF and Fisher PP tests reject the presence of unit roots for all variables, and we consider all variables as stationary based on the ADF test results and include them as is in the model.
4.2 Specification of the Panel VAR Model
As noted in the opening paragraph, the second model identifies how variables in the system respond to shock affecting other variables by employing a panel VAR framework. The previous section specified a Dynamic Panel Data model that considered the unidirectional effects of macroeconomic shocks on the bank’s NPL ratios. However, by employing a panel VAR the spillback from the bank’s NPL ratios and other macroeconomic variables can be captured. The model and its specification builds upon Love and Zicchino (2006). Here, Love and Zicchino “apply vector autoregression (VAR) to firm-level panel data from 36 countries to study the dynamic relationship between firms’ financial conditions and investment”. In Love and Ariss (2013), it is shown that employing a panel VAR model on a singular country can be beneficial in the estimation of the transmission channel of the shock on the banking sector, as it can determine how, dependent on business model and nationality, it transmits itself. We build upon the business model aspect and introduce a regionality level to account for which regions get hit first by a shock in the oil price. It is reasonable to believe that there should be an uptick in NPLs of banks closer to the production areas during a shock, compared to those further away. However, the largest banks in Norway – those that can facilitate larger investments – are located far from oil production. The analysis in this paper is limited to cover two regions, Region East and West, and two business models, Commercial and Savings Banks, as most of the oil, and oil related, industry is located on the west coast and the majority of the Commercial Banks are located in Region East.
The panel VAR model will uncover the macro-financial links between the banking system and the real economy through Impulse Response Functions (IRFs), a process where the reaction of one variable to
23
the movements in another variable in the system is shown graphically, while holding all other shocks equal to zero. In other words, all possible interactions between the variables in the model are considered simultaneously. As with the Dynamic Panel Data model, the panel VAR model is estimated with a system-based GMM following approach of Areliano & Bover (1995).
To identify shocks, Choleski decomposition is employed, a process where the variables are stacked accordingly to explore how the macroeconomic shock is believed to propagate throughout the economy. In our paper, the macroeconomic shock, the oil price growth, is stacked at the top. The interest rate growth is stacked below, with the growth rate of the non-oil GDP and credit beneath. Finally, the bank level variable, the NPLs, is stacked under the macro-level variables. The potential role of the housing price index growth in creating macro-financial feedback effects is explored in Section 5.2.4, as a measure of checking the robustness of the key findings.
The specification of the panel VAR model is as follows:
𝑌𝑖,𝑡= 𝛽0+ 𝛽1𝑌𝑖,𝑡−1+ 𝜆𝑖+ 𝑒𝑖,𝑡 (2)
Yi,t represents a vector of macroeconomic and bank-level variables at time t, where i = 1, ... , N and t =
1, ... , T, with Yi,t-1 being the lagged operator. For more information about the stacking of the vector
see the paragraph about Choleski decomposition above.
λi represents bank fixed effect. By using fixed effects, one of the advantages of panel VAR – accounting
for individual cross-sectional unit specificity (here: bank) in the level of the variables – can be employed and the response of the banks’ credit channel to macroeconomic shock can be isolated, which allows for unobserved heterogeneity.
ei,t represents the usual random errors, where the errors are independent and identically distributed
24 5 Results & Analysis
In this chapter, we present the estimation of the Dynamic Panel Data model and the panel vector autoregressive model.
5.1 Results & Analysis of the Dynamic Panel Data Model
In this section, we present the estimation of the Dynamic Panel Data model (equation (1)), using fixed effects and System GMM, the latter being proposed by Blundell and Bond (1998), and Arellano and Bover (1995). By closely following Roodman (2007), we aim to limit the lag depth and reduce the number of instruments, as a measure of avoiding overfitting our model.
From our results, presented in Table 9 and Table 10, we see that NPLs in Norway appear to be primarily driven by the housing market, interest rate and credit growth, results which are intuitive given the environment the banks operate in. The fact that the housing market is a significant determinant for NPLs in Norway is especially unsurprising given the breakdown of loans to key customer groups presented in Figure 9 below. Moreover, we find a strong relationship between NPLs and the its lagged value, i.e. auto-correlation, estimated to be between -0.20 and -0.16, with the former, the Fixed Effects estimate, suffering from a downward Nickell bias and the latter, the System GMM estimate, being upward biased. Furthermore, we note that the oil price growth (OPG), the non-oil GDP (dNO), and stock price index growth (SPG) have little to no effect on determining the NPLs of Norwegian banks, with none of the specifications yielding significant estimates for these particular variables. The time dummy (dum0809) for the Global Financial Crisis, is found to be significant across all specifications, with the implied result that the NPLs, albeit marginally, increased during 2008 and 2009. In order to check the robustness of our findings, a variable is dropped in each of the estimations, as to show that the estimate of the determinants of the NPLs is not sensitive to the exact specification (Varian, 2014), and considering the results in Table 9 and Table 10 we observe that this is the case, deeming our model to be robust.
25
Figure 9: Loans, Key Customer Groups in Percentages
Source: The Financial Supervisory Authority of Norway
Running the same specifications for the commercial banks, presented in Table A3-6 in the Appendix, we observe a three to five times increase in the coefficients of the oil price growth, meaning that the NPLs of the commercial banks are more closely linked with the oil price compared to the banking system overall. Although the coefficients are still close to zero, it indicates a plausible scenario where the larger commercial banks finance the more capital-intensive projects, as projects in the oil industry, leaving it more exposed to the oil price as its debtors belong to said industry. Furthermore, we find increases in the coefficients of the interest rate growth, housing price index and credit growth. However, none of the model specifications, neither Fixed Effects, nor System GMM, for the commercial banks are significant, most likely due to the low number of commercial banks (8), and thus the importance of the results should not be overstated.
A more granular numerical analysis indicates that a one-percentage point increase in NPLs in the previous period leads to a statistically significant decrease in NPLs in the current period of around 0.21 percentage points across all fixed effects specifications, and around 0.16 percentage points for the GMM specifications. The relationship of the NPLs exerts mean reverting behaviour, in other words when there is intense growth in NPLs in period one, the NPLs will tend to be closer to the average in the second period, and vice versa. The intuition behind this relationship is that the banks that accrue NPLs over time, or quickly over one period, will, due to the overhanging danger of defaulting on their obligations, be incentivised to lower their amount of NPLs, the same intuition can be made for the debtors’ behaviour. Although not significant, we find that a one-percentage point decline in oil price growth in the previous period results in an increase in NPLs by a meagre 0.001 percentage points. The signs of the explanatory variables are intuitive, as the part of the banks’ loan portfolio that consists of
26
loans directly, or related, to the oil industry, and on the margin of defaulting, will not meet its obligations and the loan will be classified as non-performing, leading to growth in the NPLs. We also find that, dependent on the specification, a one-percentage point increase in housing price index growth in the previous period leads to a statistically significant increase in the NPLs between 0.01 and 0.02 percentage points. The dynamics of this relationship is intuitive, as the burden of debt increases, due to the increase of the housing price index, the debtors, especially those that lend on the margin of their debt-service capabilities, will have difficulties facing their principals and interest payments. Given the loan portfolios of Norwegian banks (see Figure 9 above), this explanation seems plausible. In all specifications, except for FE4 and GMM4, we find that one-percentage point increase in the credit growth in the previous period leads to a statistically significant decline of about 0.03 percentage points in the NPLs. The intuition behind this relationship is straightforward, and in line with economic theory; lenient banks create more credit. Moreover, in times where banks are exceedingly lenient, there will be a rapid growth in credit, and as credit tightens, the debtor who lends on the margin during the lenient times, faces problems servicing its debt.
Lastly, as noted in opening paragraph of this section, we find that during the financial crisis of 2008 and 2009, the NPLs statistically significantly increased, albeit marginally, with about 0.003 percentage points across all specifications, a self-explanatory relationship. However, the size of the explanatory variable entices further examination. During the Global Financial Crisis, the Norwegian government, led by the then Prime Minister Jens Stoltenberg, an economist from the University of Oslo, pursued classical countercyclical policy, by lowering the interest rates from 5.75 percent to 1.25 percent and expanded the fiscal stimulus to upkeep the demand. As a result, the unemployment was kept low, at around 3 percent, and GDP grew during 2008 and 2009, leaving Norway quite unscathed (Organisation for Economic Co-operation and Development, 2010). With this in mind, the figure of 0.003 percentage points seems probable.
Looking at the results as a whole, there is a clear indication that the Norwegian banking system is seemingly robust, however, there exists a clear risk from the housing market. The result echoes the sentiment of the Norwegian regulators, as well as IMF (Norges Bank, 2016) (International Monetary Fund, 2016).
The Arellano-Bond AR(1) test for autocorrelation of the residuals, noted as AR(1) in Table 10, rejects the hypothesis of the errors not being autocorrelated, meaning that the errors are autocorrelated. This result is to be expected, as taking the difference generates first-order autocorrelation. The p-value of the Arellano-Bond AR(2) test, noted as AR(2) in Table 10, are all above the common limit of 5
27
percent (0.05), and thus, the hypothesis of the errors in the levels equation being uncorrelated is also not rejected. As with the results of the Arellano-Bond AR(1) test, this result is also to be expected, as it is a necessary assumption that ensures that the Arellano-Bond specifications and the orthogonality conditions are correct.
In all GMM specifications the number of instruments is kept below 10314. One could question whether
the number of instruments is too high, a question that is eloquently explored in Roodman (2009). In this paper Roodman discusses the difficulties of short-panel econometrics and explores the pitfall of "Type I errors", commonly known as false positives, as system GMM is prone to generate results that are simultaneously invalid yet appear completely valid. It is important to note that the number of instruments are not rulebound, in other words there are not a ruleset that declares the appropriate number of instruments for a given dataset. However, Roodman offers the advice that the number of instruments should be appropriate to, and categorically not outnumber, the number of data observations. For each GMM specification there are 1442 observations, and thus we qualitatively deem the 103 instruments to be suitable. Expanding on the instruments suitability, the Hansen-test of overidentifying restrictions, noted as Hansen in Table 10, is performed and p-values are given. The results, taking the p-values at face value, suggest that the instruments are appropriate, and that we should keep our null hypothesis that all overidentifying restrictions are jointly valid, in other words the instruments are not overidentifying.
Table 9: Fixed Effects, lag depth 1, Baseline Model
FE1 FE2 FE3 FE4 FE5 FE6 FE7
dNPL(-1) -0.214*** -0.214*** -0.211*** -0.206*** -0.213*** -0.212*** -0.208*** OPG(-1) -0.000 -0.001 -0.001 -0.001 -0.001 -0.001 dNO(-1) -0.019 -0.021 -0.018 -0.021 -0.004 -0.007 dIR(-1) -0.062*** -0.063*** -0.062*** -0.062*** -0.063*** -0.033 SPG(-1) -0.001 -0.001 -0.001 -0.001 -0.001 -0.002 HPG(-1) -0.019** -0.021** -0.015** -0.020*** -0.018** -0.013** dCR(-1) -0.037** -0.039*** -0.030** -0.012 -0.038*** -0.026** dum0809 -0.003*** -0.003*** -0.003*** -0.004*** -0.003*** -0.003*** -0.004*** _cons -0.002*** -0.002*** -0.002*** -0.003*** -0.002*** -0.001*** -0.002*** # of observations -1339 -1339 -1339 -1339 -1339 -1339 -1339
Source: own calculations
The dependent variable is dNPL, the differenced NPLs, where NPLs are defined as non-performing loans, i.e. non-performing over 90 days, as a percentage of total loans. dNPL(-1) is the lag of the
28
dependent variable. OPG(-1) is the annualised growth rate of the oil price lagged with one period. dNO(-1) is the differenced non-oil GDP lagged with one period. dIR(-1) and dCR(-1) is the lagged differenced of the interest rate and credit growth, respectively. SPG(-1) and HPG(-1) the growth rate of the stock and house price index lagged with one period. The variable dum0809 is included to control for the Global Financial Crisis, and _cons is the constant term. *, **, and *** denote significance at 10 percent, 5 percent, and 1 percent, respectively.
Table 10: System GMM, lag-depth 1, Baseline Model15
GMM1 GMM2 GMM3 GMM4 GMM5 GMM6 GMM7 dNPL(-1) -0.161** -0.162** -0.159** -0.154** -0.161** -0.160** -0.156** OPG(-1) -0.000 -0.001 -0.001 -0.001 -0.001 -0.001 dNO(-1) -0.017 -0.019 -0.016 -0.018 -0.002 -0.005 dIR(-1) -0.058*** -0.059*** -0.057*** -0.058*** -0.059*** -0.031 SPG(-1) -0.001 -0.001 -0.001 0.000 0.001 -0.001 HPG(-1) -0.019** -0.021** -0.015** -0.020** -0.018** -0.013** dCR(-1) -0.034** -0.037** -0.028** -0.011 -0.035*** -0.023* dum0809 -0.003*** -0.003*** -0.003*** -0.004*** -0.003*** -0.003*** -0.004*** _cons -0.002*** -0.002*** -0.002*** -0.003*** -0.002*** -0.001*** -0.002*** # of observations -1339 -1339 -1339 -1339 -1339 -1339 -1339 # of instruments -103 -103 -103 -103 -103 -103 -103 Lag depth, GMM -1 -1 -1 -1 -1 -1 -1 P-values AR(1) -0.000 -0.000 -0.000 -0.000 -0.000 -0.000 -0.000 AR(2) -0.880 -0.873 -0.911 -0.980 -0.885 -0.973 -0.977 Hansen -0.417 -0.426 -0.402 -0.285 -0.444 -0.400 -0.328
Source: own calculations
The dependent variable is dNPL, the differenced NPLs, where NPLs are defined as non-performing loans, i.e. non-performing over 90 days, as a percentage of total loans. dNPL(-1) is the lag of the dependent variable. OPG(-1) is the annualised growth rate of the oil price lagged with one period. dNO(-1) is the differenced non-oil GDP lagged with one period. dIR(-1) and dCR(-1) is the lagged differenced of the interest rate and credit growth, respectively. SPG(-1) and HPG(-1) the growth rate of the stock and house price index lagged with one period. The variable dum0809 is included to control for the Global Financial Crisis, and _cons is the constant term. # of instruments and Lag depth, GMM, show the number of instruments and the lag-depth of the estimation. The P-values of AR(1), AR(2) and Hansen are included to test for autocorrelation of the residuals, the correlation in the errors in
29
the levels equation, and overidentifying restrictions. *, **, and *** denote significance at 10 percent, 5 percent, and 1 percent, respectively.
5.2 Results of the Panel VAR Model
In this section, we present the estimation of the panel VAR framework, equation (2), using the Stata codes ‘pvar2’ by dr. Ryan Decker16 of the University of Maryland. The ‘pvar2’ codes are a revision of
Love’s ‘pvar’ codes, which was written by Love for Love and Zicchino (2006). The key focus, and aim of this section, is to explore how the various macroeconomic shocks affect the bank-level variable. A pivotal feature of the panel VAR framework is that it allows for all variables in the system to affect each other. In other words, the framework accounts for all possible interactions between the various variables simultaneously.
As noted in section 4.2, the ordering of the variables in the baseline panel VAR specification considers three macroeconomic aggregates, and two bank-level variables, with the ordering being as follows {OPG dIR dNO dCR NPL}17. In Table 11, descriptive statistics of the key variables are presented. Table 11: Descriptive statistics of the variables deployed in panel VAR model
Variable Observations Mean Std. Dev. Min Max
OPGt -1442 -0.080 -0.270 -0.471 -0.426
dIRt -1442 -0.004 -0.014 -0.036 -0.016
dNOt -1442 -0.000 -0.021 -0.040 -0.038
dCRt -1442 -0.003 -0.022 -0.061 -0.033
NPLt -1442 -0.010 -0.009 -0.000 -0.126
Source: own calculations
From Figure 10 below, the impulse response functions of the baseline model, we can see the impact of a negative shock in the oil price growth of one standard deviation18, as well as the interaction
between the real and financial factors. Qualitatively, we see that as the oil price decreases, from period 0 to period 2, the interest rate, non-oil GDP, and credit growth follow suit, although at a slower pace. From its initial shock, the NPLs rise as the oil price drops, which is an intuitive and straightforward relationship. The following periods, period 2 to period 4, see an increase in all variables expect for the NPLs, until all variables slowly decrease back to their pre-shock levels after 15
16 For further information, and download, see: http://econweb.umd.edu/~decker/code.html
17 Oil price growth, interest rate, non-oil GDP, credit growth, dNPL
30
periods. Quantitatively, we find that, after normalising the standard deviation to one percent19, the
implication of the oil price growth shock is an immediate decrease in the non-oil GDP of about 0.04 percentage point, as well as an instantaneous increase in NPLs close to 0.01 percentage points. Whereas, the interest rate and credit growth experience a lagged decrease, period 2 instead of period 1, of around 0.20 percentage points. The long-term effects of the oil price growth shock for the non-oil GDP is a decrease close to 0.03 percentage points, with the interest rate and credit growth also decreasing, however at smaller scale, with about 0.03 and 0.01 percentage points respectively.
Figure 10: Impulse response functions to shocks, Baseline Model
Source: own calculations
From Table A7 in the Appendix, a feedback effect, within the bank balance sheets is also observed; a one percentage increase in the NPLs leads to an, albeit heavily lagged, spike in the credit growth of about 0.21 percentage points. However, the long-term effect on the credit growth, following the one percentage rise in the NPLs, is a drop close to 0.03 percentage points. Moreover, a one percentage
19 Calculated as follows: impulse response magnitude of a given variable divided by the standard deviation of the variable
that is shocked. For example, the impulse response magnitude of the interest rate divided by the standard deviation of the oil price growth is equal to a increase in the interest rate of about 0.03 percentage points (0.01/0.27=0.03)
0.0000 0.0500 0.1000 0.1500 0.2000 0.2500 O PG 0 2 4 6 8 10 12 14 s OPG shock -0.1000 -0.0500 0.0000 0.0500 O PG 0 2 4 6 8 10 12 14 s dIR shock -0.0500 0.0000 0.0500 0.1000 O PG 0 2 4 6 8 10 12 14 s dNO shock -0.0600 -0.0400 -0.0200 0.0000 0.0200 O PG 0 2 4 6 8 10 12 14 s dCR shock -0.0200 0.0000 0.0200 0.0400 0.0600 0.0800 O PG 0 2 4 6 8 10 12 14 s NPL shock -0.0050 0.0000 0.0050 0.0100 dI R 0 2 4 6 8 10 12 14 s OPG shock -0.0020 0.0000 0.0020 0.0040 0.0060 0.0080 dI R 0 2 4 6 8 10 12 14 s dIR shock 0.0000 0.0010 0.0020 0.0030 0.0040 0.0050 dI R 0 2 4 6 8 10 12 14 s dNO shock -0.0010 0.0000 0.0010 0.0020 0.0030 dI R 0 2 4 6 8 10 12 14 s dCR shock -0.0020 -0.0010 0.0000 0.0010 0.0020 0.0030 dI R 0 2 4 6 8 10 12 14 s NPL shock -0.0100 -0.0050 0.0000 0.0050 dN O 0 2 4 6 8 10 12 14 s OPG shock -0.0100 -0.0050 0.0000 0.0050 dN O 0 2 4 6 8 10 12 14 s dIR shock -0.0050 0.0000 0.0050 0.0100 0.0150 dN O 0 2 4 6 8 10 12 14 s dNO shock -0.0040 -0.0020 0.0000 0.0020 0.0040 dN O 0 2 4 6 8 10 12 14 s dCR shock -0.0040 -0.0020 0.0000 0.0020 dN O 0 2 4 6 8 10 12 14 s NPL shock -0.0060 -0.0040 -0.0020 0.0000 0.0020 0.0040 dC R 0 2 4 6 8 10 12 14 s OPG shock -0.0100 -0.0050 0.0000 0.0050 0.0100 dC R 0 2 4 6 8 10 12 14 s dIR shock -0.0050 0.0000 0.0050 0.0100 0.0150 dC R 0 2 4 6 8 10 12 14 s dNO shock -0.0050 0.0000 0.0050 0.0100 dC R 0 2 4 6 8 10 12 14 s dCR shock -0.0040 -0.0020 0.0000 0.0020 dC R 0 2 4 6 8 10 12 14 s NPL shock -0.0005 0.0000 0.0005 0.0010 0.0015 N PL 0 2 4 6 8 10 12 14 s OPG shock -0.0004 -0.0002 0.0000 0.0002 0.0004 0.0006 N PL 0 2 4 6 8 10 12 14 s dIR shock -0.0010 -0.0005 0.0000 0.0005 N PL 0 2 4 6 8 10 12 14 s dNO shock -0.0010 -0.0005 0.0000 0.0005 N PL 0 2 4 6 8 10 12 14 s dCR shock 0.0000 0.0020 0.0040 0.0060 N PL 0 2 4 6 8 10 12 14 s NPL shock
Errors are 5% on each side generated by Monte-Carlo with 500 reps
31
increase in the interest rate leads to an almost immediate decrease in the credit growth of slightly above 0.60 percentage points, with the long-term effect being close to a drop of 0.50 percentage points.
The impact of the feedback effect within the bank balance sheets can also be found to have an impact on the economic activity; A one percentage point increase in credit growth leads to a 0.11 percentage point decline in non-oil GDP. Likewise, a one percentage point rise in the NPLs decreases the non-oil GDP by 0.32 percentage points, with the long-term effect being a decrease in the non-oil GDP of about 0.01 percentage points.
The dynamics of the findings are economically intuitive, and, moreover, the signs are as expected and in line with the findings in section 5.1. Furthermore, the results outline the macro-financial linkages of the Norwegian economy; The increase in NPLs arises after the drop in the oil price growth and decline in the non-oil GDP. Moreover, the results show that in the long term both the interest rate and credit growth decline, as a response to the shock in the oil price growth. Investigating the slowdown in credit growth more closely indicates that it leads to weaker growth of the non-oil GDP in the short term, which subsequently leads to an increase in the NPLs. Intuitively, the drop in the oil price leads to a slowdown in the economy, both on- and offshore, and as a response, the interest rate is lowered. The lowered interest rate leads in turn to an increase in credit growth, however, over time, as the oil price normalises again, there is a tightening in credit and subsequent rise in NPLs.
From the variance decomposition, reported in Table 12 - Table 15, the shocks presented in the impulse-response functions, Figure 10, can be easily quantified, as the decomposition gives an indication of the amount of information each variable contributes to explain the other variables in the panel vector autoregression. In other words, the variance decomposition determines how much of the variance in the forecast errors of each of the variables can be explained by the exogenous shocks to the other variables. The variance decomposition of credit growth, reported in Table 12, indicates that oil price shock explains about 14.5 percent of the variation in the credit growth, while the interest rate and non-oil GDP explains almost 30 and 27 percent respectively of the variation. The variance decomposition of the interest rate growth, reported in Source: own calculations
Table 13, indicates that oil price shock explains 41 percent of interest rate variation, while non-oil GDP
and credit growth explain 11.6 percent and 5 percent of the interest rate variation. In Table 14, depicting the variance decomposition of non-oil GDP, we find that that oil price shock explains slightly over 28 percent of non-oil GDP variation, whilst the interest rate growth explains almost 22 percent
32
of the variation. Lastly, the variance decomposition of NPLs, presented in Table 15, indicates that oil price shock explains almost 4 percent of variation in NPLs, while the non-oil GDP, credit growth and interest rate explain close to 3 percent of the NPL variation.
Table 12: The forecast error variance decomposition of the credit growth Credit Growth
Time OPGt dIRt dNOt dCRt NPLt
1 2.6% 27.9% 11.6% 57.9% 0.0% 2 3.8% 14.2% 45.4% 36.4% 0.2% 3 9.9% 26.7% 33.1% 29.2% 1.0% 4 10.9% 30.4% 29.7% 26.1% 3.0% 5 13.0% 30.9% 27.8% 25.3% 2.9% 6 13.7% 30.6% 27.5% 25.3% 3.0% 7 14.4% 29.6% 27.4% 25.4% 3.3% 8 14.5% 29.2% 27.2% 25.4% 3.8% 9 14.4% 29.2% 27.0% 25.2% 4.2% 10 14.3% 29.4% 26.9% 25.1% 4.4% 11 14.3% 29.5% 26.8% 25.0% 4.4% 12 14.4% 29.5% 26.8% 25.0% 4.4% 13 14.4% 29.4% 26.8% 25.0% 4.4% 14 14.4% 29.4% 26.8% 25.0% 4.4% 15 14.4% 29.4% 26.7% 25.0% 4.4%
Source: own calculations
Table 13: The forecast error variance decomposition of the interest rate Interest Rate
Time OPGt dIRt dNOt dCRt NPLt
1 45.7% 54.3% 0.0% 0.0% 0.0% 2 38.1% 44.7% 13.1% 1.1% 3.0% 3 43.2% 38.6% 11.3% 4.3% 2.6% 4 42.2% 37.8% 11.4% 4.8% 3.9% 5 41.6% 37.5% 11.8% 4.9% 4.3% 6 41.2% 37.9% 11.6% 4.9% 4.4% 7 41.1% 38.0% 11.6% 4.9% 4.5% 8 41.1% 38.0% 11.6% 4.9% 4.4% 9 41.1% 37.9% 11.6% 5.0% 4.5% 10 41.1% 37.8% 11.6% 5.0% 4.5% 11 41.0% 37.8% 11.6% 5.0% 4.5% 12 41.0% 37.8% 11.6% 5.0% 4.5% 13 41.0% 37.8% 11.6% 5.0% 4.5% 14 41.0% 37.8% 11.6% 5.0% 4.5% 15 41.0% 37.8% 11.6% 5.0% 4.5%
33
Table 14: The forecast error variance decomposition of the non-oil GDP Non-Oil GDP
Time OPGt dIRt dNOt dCRt NPLt
1 4.9% 1.5% 93.6% 0.0% 0.0% 2 25.0% 22.4% 49.7% 2.7% 0.2% 3 27.2% 22.4% 45.9% 2.5% 2.0% 4 29.0% 21.7% 43.8% 3.5% 1.9% 5 28.6% 21.7% 43.5% 3.9% 2.3% 6 28.6% 21.3% 43.3% 4.3% 2.5% 7 28.5% 21.3% 42.9% 4.4% 2.8% 8 28.4% 21.4% 42.8% 4.4% 3.0% 9 28.3% 21.5% 42.6% 4.4% 3.0% 10 28.3% 21.6% 42.6% 4.4% 3.1% 11 28.4% 21.6% 42.6% 4.5% 3.0% 12 28.4% 21.6% 42.5% 4.5% 3.1% 13 28.4% 21.6% 42.5% 4.5% 3.1% 14 28.3% 21.6% 42.5% 4.5% 3.1% 15 28.3% 21.6% 42.5% 4.5% 3.1%
Source: own calculations
Table 15: The forecast error variance decomposition of the non-performing loans Non-Performing Loans
Time OPGt dIRt dNOt dCRt NPLt
1 0.7% 0.0% 1.3% 0.5% 97.5% 2 2.9% 0.0% 2.2% 1.7% 93.2% 3 3.6% 0.2% 2.7% 2.5% 91.0% 4 3.7% 0.6% 3.0% 2.8% 89.8% 5 3.6% 1.2% 3.1% 2.9% 89.2% 6 3.6% 1.6% 3.1% 2.9% 88.8% 7 3.7% 1.9% 3.0% 2.9% 88.5% 8 3.8% 1.9% 3.1% 2.9% 88.3% 9 3.8% 1.9% 3.1% 3.0% 88.2% 10 3.8% 1.9% 3.1% 3.0% 88.1% 11 3.8% 1.9% 3.1% 3.0% 88.1% 12 3.8% 2.0% 3.1% 3.0% 88.1% 13 3.8% 2.0% 3.1% 3.0% 88.1% 14 3.9% 2.0% 3.1% 3.0% 88.1% 15 3.9% 2.0% 3.1% 3.0% 88.0%
