NPLs MODELING IN THE BANKING SECTOR OF BOSNIA AND HERZEGOVINA

Za predmet istraživanja smo izabrali loše kredite u bankarskom sektoru Bosne i Hercegovine. Cilj nam je bio da stvorimo modele za predviđanje njihovog kretanja. Pored raznovrsnih prostih linearnih regresionih modela i analize vremenske serije, koristili smo i alate tehničke analize, što je jedna velika novost u ovoj oblasti. Na loše kredite utiču: dospjela potraživanja, stopa rasta BDP i kamatni raspon. Najvažniji rezultat istraživanja je verifikacija hipoteze da se alati tehničke analize mogu koristiti za predviđanje kretanja loših kredita. Dokazali smo i da različite metodologije za predviđanje i različiti regresioni modeli, sa različitim nezavisnim varijablama, odgovaraju različitim periodima razvoja loših kredita.


Introduction
An economic crisis can be reflected in several ways.In the banking sector the most obvious is the decline in the quality of banking assets.On the territory of the second Yugoslavia, including Bosnia and Herzegovina, they mostly consist of loans.Financial instruments are present in a rather low percentage, hence the main form corrupting banking assets are non-performing loans (hereafter referred to as NPLs).As such, they are not directly visible in the balance sheet.They are reached through the process of loans classification.According to the roughest division, NPLs should include loans in respect of which the debtor defaulted for over 90 days.Debtors' default is the main criterion for classification.NPLs include loans from the categories C (debtors' default over 90 days), D (over 180 days) and E (over 270 days).Loans in the category B qualify as classified assets, but not bad assets, whereas loans in the category A are first-class assets.In technical literature NPLs are referred to as non-performing loans or bad loans.
The subject of our research is NPLs in the banking sector of Bosnia and Herzegovina (hereafter referred to as BSB&H).We bring NPLs in relation to matured receivables (hereafter referred to as MR), interest rate spread (hereafter referred to as IRS) and the growth rate of gross domestic product (GDP g.r).The goal is to develop models for NPLs forecasting, directly based on available data on NPLs, by means of technical analysis and time series analysis, and indirectly by means of MR, IRS, and GDP g.r.We start the research with the hypothesis that the forecasting of NPLs movements requires a combination of different methodologies and models; that the different periods of NPLs development require different models, and that even the technical analysis can be successfully implemented in the process of analyzing and forecasting NPLs.
The first part of the research presents the methodological basis of the paper, which is followed by a brief review of bibliographical references.The third segment features the research results, i.e. the main characteristics of the NPLs series, the implementation of technical analysis in the process of NPLs forecasting, establishing the connection between MR and NPLs and analyzing the predictive capabilities of MR, developing regression models NPLs vs. GDP g.r. and NPLs vs. IRS, as well as the time series analysis of NPLs by means of a secondorder polynomial.In addition to summing up the research results, the conclusion highlights certain dilemmas and potential directions for further research.

Material and methods
The information basis of this paper is financial statistics published by the central bank of Bosnia and Herzegovina, and the net balance sheets of the banking sector of the Republic of Srpska (Banking Agency of the Republic of Srpska) and the banking sector of the Federation of Bosnia and Herzegovina (Banking Agency of the Federation of B&H).
The main methodological research tools are simple linear regression analysis, technical analysis, descriptive statistics and time series analysis.To assess regression models, we use standard methodology [Žižić et al, 1992: 273-317], and the same goes for time series analysis [Žižić et al, 1992: 397-449].In modeling NPLs through a time series, we used a second-order polynomial.
NPLs and MR were not presented as absolute, but relative percentage values; NPLs in relation to total loans, and MR in relation to total receivables in respect of loans, leasing and interest.The analysis encompasses a period of 13 years -from the end of 2000 to Q2 2012.The data on NPLs and ROE for all years are quarterly, except for 2000 and 2001 when they are annual.The time series of MR is at the annual level (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011), just like the time series of GDP g.r., but for a shorter period of time (2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011).IRS (in percentage points/p.p.) is calculated as a difference between active and passive interest rates.Active interest rates are interest rate on short term corporate loans in convertible marks (KM), whereas passive interest rates are interest rates on corporate time and savings deposits in convertible marks (KM).The IRS time series starts in 09/2004, and ends in 06/2012.
Koncept impulsa/momentuma (MOM) potiče iz fizike.Predstavlja stopu po kojoj brzina nekog objekta raste ili opada.Na tržištu akcija indikator impulsa mjeri da li se kretanje cjena akcija ubrzava ili usporava, a mi smo ga upotrjebili za mjerenje promjene LK.IT systems, also the data on a more frequent basis.The NPLs series is a time series, just like the time series of financial assets prices -for instance, stock market prices.Technical analysis is a model used to analyze and predict the stock prices movements.It starts from the assumption that the previous stock prices and traded volumes can be used as a basis to determine the future stock prices, i.e. the trends in price movements.Technical analysis can also be applied to the NPLs series, the relevant differences being taken into account.The first difference is that NPLs are in percentages (i.e. a relative value), whereas the stock prices are an absolute value.The second difference between these series is that in the process of analyzing the prices of financial assets, in addition to the price, we also have trade volumes to take into account.The third difference refers to the frequency of data: financial assets price is either daily or inter-daily, whereas the NPLs data are not frequent at all -they are officially given at a quarterly level.Every piece of data on NPLs is unique, as opposed to the stock price which is broken down into four prices: opening price, closing price, highest price and lowest price.That is the fourth difference.
Due to these differences, we have chosen three rather simple and practical tools of technical analysis: support level, trend line, momentum, MOM, and rate of change, ROC.
Support level and resistance level were developed and applied by Dow Jones.At the stock market the support level is the level of stock prices at which the market does not allow the stock prices to decline any further.At the support level, the bidding orders surpass the selling orders, which causes the prices to stop declining.Given that a drop in NPLs is a positive phenomenon, and the term "support level" has been taken from the financial market where a drop in stock prices is a negative thing, "resistance level" would be a more logical and appropriate name for the level at which resistance to a further decline in NPLs forms.
Trend line is a very simple, but useful tool of technical analysis.The trend line connects either the higher lows or the lower highs.The cross-section of a thereby constructed trend line and the stock price, i.e. in our research the NPLs time series, indicates that -with a high degree of certainty -one can expect a trend change.
The concept of momentum (MOM) comes from physics.It refers to the rate at which the speed of an entity increases or decreases.At the stock market the momentum indicator measures whether the stock prices movements are accelerating or decelerating, and in our research we used it to measure the changes in NPLs.Momentum represents the difference between the first and the last value of the time series over a specific time horizon.In our research we used ROC 3 (the difference between NPLs in the first and the third quarter), ROC 6 (the difference between NPLs in the first and the sixth quarter) and ROC 10 (the difference between NPLs in the first and the tenth quarter).
The price rate of change -ROC measures the intensity of price changes over time.This indicator is based on the concept of momentum, the difference being that it determines the relative/percentage change instead of the absolute one.ROC provides the information about the relative difference between the current and some previous value of NPLs.ROC grows when NPLs grow, and it drops when NPLs drop.We used ROC 3, ROC 6 and ROC 10.The lower the ROC, the higher the probability of NPLs growth, and vice versa.

References review
Studies on the topic of how banking and macroeconomic variables affect non-performing loans in the banking sector are numerous and rather diverse.In a geographically close country (i.e.Romania) there have been studies on the correlation between average interest rates and non-performing loans [Socol, Adela and Iuga, Julia, 2010: 777].The correlation has been confirmed (measured by the Pearson correlation coefficient), along with the existence of "other indirect channels affecting non-performing loans".The main findings in the research of what causes the NPLs growth in Tanzania [Evelyn, 2011: 50] are that the non-performing loans grow because of the non-purposeful utilization of the loans, and that a rigorous monitoring of loans utilization limits the growth of NPLs.This research was conducted by examining 48 top banking officials.The banking sector of Tunisia [Bahrini, 2011: 230] suggests that non-

Frequency Distribution and Descriptive Statistics
The time series commencing in Q4 2000 and ending in Q3 2012 encompasses 41 quarters (Table 1).On average (arithmetic mean) BSB&H functions at the NPLs level of 7.59 %.The NPLs value of 6.1 divides the series in two sections -that is the median.In 3 cases the NPLs level is 5.3% and 3% (modes).The highest NPLs level is 21.2% (Q4 2000), and the lowest 3% (first appearing on 31.12.2007).The spread (i.e. the difference between the maximum and minimum NPLs value) of 18.2 p.p. (percentage points) is interpreted as high variability in the NPLs trends.
The findings about high variability are confirmed by the variation coefficient (standard deviation/arithmetic mean) of 0.56.On average, NPLs deviate from the arithmetic mean of the NPLs series by +/-56%.NPLs distribution is skewed to the right (Graph 1), the third moment is 1.16, while in normal distribution it equals 0. The long tail, i.e. the extreme values of NPLs, is located in the right segment of the distribution.The NPLs distribution is flat (platykurtic).Its fourth moment is 1.38 (in case of normal distribution it equals 3).In terms of both kurtosis and skewedness, NPLs distribution significantly deviates from normal distribution.
The probability that NPLs at the level of BSB&H, or individual banks, will be 3% or below 3% amounts to 7.3%.The highest portion of NPLs, 41.5%, falls into the interval from 3.1% to 6%. 65.8% of NPLs are less than or equal to 9%.22.5% of NPLs fall into the interval from 9.1% to 12%.This interval signifies the transition towards the extremely high values of NPLs.

Technical Analysis
The trend line (Graph 2, Model 2) obtained by connecting the lower highs intersects with the NPLs series at two points: on 31.12.2007 when NPLs were 3 %, and on 30.09.2008, when NPLs were 3.3%.A shift in the trend is by means of the trend line method identified already in late 2007.The signal is confirmed in Q1 2008, and according to the empirical data it was not evident until Q3 2009, when the NPLs from the initial 3% grew to 4.8 %.Therefore, the trend line method anticipated the growth of NPLs 2 years and 9 months before the change became obvious.The same conclusion, though not that exact as in the case of trend line, is reached by analyzing support levels (Graph Za razliku od ROC, MOM mjeri ubrzanje/ usporenje kretanja kao razliku u vrijednostima LK, a ne kao količnik vrijednosti LK.Za MOM 3 od nula, u trenutku kada je LK 3% i unutar nivoa otpora/podrške, MOM 6 i MOM 10 su -0,8 i -1,9 (Grafikon 5, Model 5).Oni padaju ispod 1 tj.ispod 2. To su ključne vrijednosti MOM, kada je LK unutar intervala otpora, koje najavljuju promjenu trenda LK, i to u ovome slučaju čak godinu dana ranije prije nego što do očigledne promjene trenda i dođe (30.09.2009.g.).

Matured Receivables
According to a very rough division, receivables in respect of loans are classified into: receivables in respect of loan principal, receivables in respect of leasing, and matured receivables.In 2000 MR was 13.74 % (Graph 6).6).Limits to growth are a familiar concept in social and other sciences.Everything develops, grows/declines and reaches its maximum, or minimum, which is, respectively, followed by a decline, or growth.If there is a limit to growth, there must be a limit to decline.

Gross Domestic Product
The most obvious manifestation of the increased production capacities is the GDP g.r.The growth is equivalent to the economic expansion.It takes place in the circumstances of intensified bank lending.What characterizes this stage in the economic cycle is the low level of NPLs and a low share of debtors in default.On the other hand, what marks a cyclical downswing is the deterioration of the credit portfolio and the migration of loans from higher into lower categories of assets.In the period 2003-2011 GDP g.r.ranged from 13.4% to -2.9 % (Table 2) and NPLs from 3.0% to 11.8%.The time series is spoiled by the data for 2009 -the level of NPLs does not correspond to the GDP g.r.After the time series get revised, by omitting the data for 2009, the regression model is very precise (Graph 8, Model 9); 93% of the NPLs variability is explained by means of the variations in GDP g.r.The correlation between variables, inverse as expected, is rather strong, the correlation c o e f f i c i e n t a m o u n t i n g to -0.98.The growth/decline of GDP by one p.p. increases/decreases NPLs by 0.788 p.p.According to the simple linear regression model, GDP g.r. of 0% corresponds to the NPLs level of 13.2.If the GDP g.r. is -5 %, the NPLs are 17.1%, whereas the GDP g.r. of 10% corresponds to the NPLs of 5.3 (Table 4).Higher precision in describing the correlation between the variables can be achieved by using quarterly data on GDP g.r, but the official statistics does not publish such data.However, even on the basis of the expected GDP g.r, according to the characteristics of the constructed regression model, one can determine the end-of-the-year values of NPLs.The model is highly reliable; p values for the assessment of the model's coefficients approach zero (Table 3).

Interest Rate Spread
The growth of NPLs is accompanied by the widening of the IRS.In order to compensate for the losses in their credit portfolio, banks must attract additional deposit potential.Deposits, like all other goods, have a price.The supply of deposits increases as the passive interest rates grow, which is why a growth of NPLs is, on average, accompanied by a permanently higher level of passive interest rates.Their increase subsequently brings up the prices of the banking aggregates -i.e. interest rates on loans.This is the theoretical process of harmonizing the prices of bank resources and aggregates when the NPLs grow.
Yet, the theory and practice are not always harmonized; the processes in the banking books do not necessarily follow theoretical assumptions or empirical patterns.Since the outbreak of the global economic and financial crisis, IRS in the BSB&H have not been incessantly revised according to the theoretical/practical model of harmonizing active and passive interest rates.In the period since October 2008 (bankruptcy of the US investment bank Lehman Brothers) until the end of 2012, we have not managed to establish, for the entire analyzed time interval, a unique direct linear correlation between the growth of NPLs and IRS.
From Q1 2009 to Q2 2010, the growth of NPLs and the growth of IRS were concurrent.The growth of NPLs by 1 p.p. increased the IRS by 0.26 p.p. (Graph 9).In the model with a time lag by one quarter, the strength of the correlation is identical (R 2 =0.9), with a note worth mentioning that it is still closer to reality.If the expected IRS is 5 p.p. NPLs will be 7.45%.This is one way to interpret the inverse function/ model.Based on the same model (Graph 10), we can also draw the following conclusion: if we isolated a direct linear correlation between NPLs and IRS, and if IRS is 5%, and NPLs substantially deviate from 7.45%, then the NPLs are underestimated/overestimated.It is taken that NPLs cannot be negative.This is why, for the IRS values below 2.87, the inverse model is useless when it comes to NPLs forecasting.The linear regression model, in addition to the high correlation coefficient, also has the excellent estimated values of the model's coefficients.P values are 0.000257 and 0.0038 (Table 5).In Model 11 (Table 6) we rejected the hypothesis that coefficients equal zero with the probability of error amounting to 1.7 %, and 0.3 %.Modeliranje LK preko vremenske serije LK smo modelirali, kao polinom drugog reda, i to za dva perioda.U prvom periodu, on obuhvata pad LK (Q2 2003 -Q3 2008), polinom objašnjava 98% varijabiliteta LK (grafikon 11, Model 14).Produžetak/ekstrapolacija modela daje buduće vrijednosti koje značajno odstupaju od stvarnih/empirijskih vrijednosti LK.Tako je u Q3 2012 teorijska/modelirana vrijednost LK 5,36%, a stvarna 12,7%.Polinom je odlična teorijska aproksimacija LK tokom njegovog pada (standardna greška modela 0,08), ali on nije model podesan za dalje predviđanje LK.
The praise directed at Model 12 ends when we try to use it for reaching the high values of NPLs.IRS of 10 corresponds to NPLs of only 8%.Although it encompasses a long time period, this model is not economically justified beyond it.IRS is not the only variable determining NPLs.
Again, a seemingly excellent model, with high R 2 , higher than in the previous model
Već u 2004.g. opadajući trend DP slabi, promjena u odnosu na 2003.g. je -0,6 p.p. U reach the local maximum.Stopping the growth of NPLs fits into the limits-to-growth concept and supports the principle that every phenomenon, after a period of intense growth, enters a stage of slower growth, end of growth, and decline.
In the banking systems characterized by a rapid decline of NPLs, preceded and/or followed by a credit expansion, a second-order polynomial is the best theoretical approximation of NPLs trends.The same goes for the second-order polynomial in the period of strong NPLs growth.Moreover, it is a useful tool for short-term prediction of NPLs.Adopting a rather rigorous and extremely optimistic assumption that the NPLs will follow the created polynomial model even in the medium term (Graph 12), in 2016 forecasts the NPLs in the interval between 5% and 3%.It is in that interval that resistance should be formed to further decline of NPLs.After the decline in NPLs stops, the secondorder polynomial is no longer suitable for further prediction of the NPLs movements, unless history repeats itself in the absolutely exact form, which is either an impossible event, or a highly unlikely event, or an event repeated in the distant past.The optimism of this model comes as a consequence of its structure.Its independent variable is the time itself, i.e. the passage of time, and the fundamental banking or economic variables may be, but are typically not represented and featured.The polynomials are divided by the NPLs stagnation period, i.e. the time in which resistance to the further decline of NPLs is formed.Formation of the resistance level, a technical analysis tool, enables the division of the time series and the formation of the polynomials, which, each in its own section, excellently approximate the NPLs movements.The resistance level signifies the end of one trend and the beginning of another trend.The concept of technical analysis is, overall, intertwined with the time series analysis.

Conclusion
Using different methodologies, we have developed 15 models (Table 9) for predicting NPLs.The hypotheses of our research have been proven: 1) predicting NPLs requires a combination of various methodologies 2) different forecasting models are suitable for different periods of NPLs development 3) technical analysis is one of the methods/models for forecasting NPLs.
The NPLs time series is extremely variable.In the period 2000-2012, the minimal level of NPLs was 3%.The maximum value of NPLs was 21.2% -the long tail is formed in the right section of the distribution.Probability that NPLs will take extremely low values, below 3%, and extremely high values, above 12.1%, amounts to 7.3% and 12.2% respectively (Model 1).The most frequent (41.5%) values of NPLs are between 3.1% and 6%.The arithmetic mean of the NPLs series is 7.59%, and the median is 6.1%.

Literatura / References
The logic behind all the applied diagnostic tools concerning NPLs trend changes is the rule that, just like there is a limit to growth, there must be a limit to decline.In B&H the limit to decline of NPLs is 3% (2007 and 2008), and for MR 2% (2007).This has been indicated by the financial statistics of BSB&H.
The correlation between GDP g.r. and NPLs is very strong (Model 9).GDP g.r.explains 93% of the NPLs variability, and p value of the model's coefficient equals zero.Model 9 approaches the ideal model.
The assessment of coefficients in the model with a time lag (NPLs=-0.52*GDPg.r.(-1)+10.23,Model 10) is excellent (p value equals zero), but the model is still weaker than the previous one, given that its determination coefficient is lower and amounts to 0.744.Model 10 is partially usable.
NPLs were put in correlation with IRS.Model 11 is the inverse model IR=ƒ(NPLs).Its R 2 is high (0.9), and the highest p value is low, i.e. 0.018.The linear functional dependence between NPLs and IRS is very high and reliable, but only for a small, limited period (Q1 2009-Q2 2010) during which there is an inverse correlation between the variables.Model 11 is partially usable for predicting NPLs, meaning that it is only usable during the period in which the trends between NPLs and IRS are, or are expected to be, inverse.The standard error of the model is 0.74.
Model 12 has an excessively high p value in the assessment of the model's coefficients.It cannot be used for forecasting NPLs.Its variation with a time lag, Model 13, has a high R 2 and a low p value at 0.022.Based on its elements/specification, it can only be used to forecast lower/low values of NPLs, because the high values of NPLs require extremely high, practically impossible, IRS values.
With a further development of the marketoriented model of the banking system, implying a reduced oligopoly and monopoly rent, we expect the establishment of a higher and longer inverse correlation between the NPLs and the IRS.
Model 14 and Model 15 are second-order polynomials.The first is an extraordinary approximation of the NPLs trends during their decline (R 2 =0.98), and the second during their growth.Under the assumption that there are no hidden losses in the balance sheets of the BSB&H, and that the variable of time contains in itself the basic economic variables, Model 15 may predict the NPLs values with a high level of reliability (R 2 =0.97).
The detected rules concerning the trends in values of the observed variables, especially the correlation strength, must undergo the process of additional scientific verification.The development of banking business will either confirm or deny the hypotheses established and proven in this paper.Presented results are just the beginning of a deeper and more precise understanding of the relation among the key variables in the BSB&H and the possibility of their prediction.
Further research of NPLs can be directed at building multiple linear and non-linear regression models, describing even more precisely the cause and effect relations between NPLs and banking/macroeconomic variables.
Since 2004 the momentum of the declining MR trend has weakened.After a rapid decline, in 2002, by 5.3 p.p. compared to 2001, in 2004 MR decreased by only 0.6 p.p.A similar standstill in the decline of MR is visible in the subsequent three years (until 2007), whereas in 2008 the annual drop in MR amounted to just -0.1 p.p. Having reached the limit to their decline, for the first time in 9 years MR started to grow.The annual growth of MR in 2009, 2010, and 2011 was 1.5 p.p., 2.1 p.p. and 3.5 p.p. respectively.According to the annual growth/drop of MR, the increase in MR which occurred for the first time in 2009 could have been anticipated/foreseen already based on the slowed down decline of MR in 2004 (Model

A
change in the structure of receivables in respect of credit instruments may signify a change in the business strategy.On the other hand, a change in MR and a change in the size of their growth/decline indirectly signify a change in the credit portfolio's quality.By interpreting this change one is able to determine the quality of receivables (NPLs) more precisely than by loan classification.If we leave out rescheduling and refinancing of loans, as an artificial yet legitimate form of reducing matured receivables, then MR, due to the lower discretion in their balancing, serve as a more reliable indicator of loans' quality than NPLs.There is an extremely powerful inverse correlation between MR and NPLs (Graph 7).The model is set up as NPLs=1.55MR(Model 7).According to it, 95% of NPLs variability is explained by the change in MR.The model approaches a perfectly linear one.The change in MR by 1 p.p. increases NPLs by 1.5 p.p.In determining the quality of the credit portfolio, MR can be used as a proxy for NPLs.P values of the model's skewness coefficient approach zero.The model with a time lag (NPLs-2.49+MR(-1)) is a poorer interpretation Graph 5. NPLs and MOM, BSB&H, Model 5 Source: Ibid.
)*GDPg.r.(-1), and R 2 is only 0.136.However, the regression model with a time lag, for the period from 2003-2011, including the year 2009, features the completely satisfactory characteristics: R 2 of 0.744 and p values of the model's coefficients approaching zero.The model's specification is as follows: NPLs=-0.52*GDPg.r.(-1)+10.23(Model 10).The increase of GDP by 1%, which reduces the NPLs by about half a percentage point, is fully justified in economic terms.The model with a time lag results in NPLs of 10.23 at the GDP g.r of 0 %.

Tabela 8 .
Ocjena parametara modela LK=0,929+0,699KR(-1), Q4 2004 -Q3 2008, Model 13 , and low p value of the model's coefficients assessment, obtained by a transformation of the previous model into a model with a time lag (Table 8, Model 13).The main characteristics of the model are high determination coefficient (R 2 =0.87) and p values for the model's coefficients below zero.The problem with this model is that its extrapolation provides complete illogical and impossible values.For instance, the IRS of 13 p.p. corresponds to NPLs of only 10%, and IRS of 13 p.p. is an impossible event.IRS cannot explain the high values of NPLs.The model can only function in the domain of extremely low NPLs.Therefore, it has a limited usability.NPLs Modeling Through a Time SeriesIn the form of a second order polynomial we modeled NPLs for two periods.In the first period covering the decline of NPLs (Q2 2003-Q3 2008), the polynomial explains 98% of the NPLs variability (Graph 11, Model 14).The extension/extrapolation of the model provides future values that substantially deviated from the real/ empirical NPLs values.Thus, in Q3 2012 the theoretical/modeled value of NPLs was 5.36% and the actual one 12.7%.The polynomial is an excellent theoretical approximation of NPLs during their decline (standard error of the model is 0.08), but it is not a model suitable for further forecasting of NPLs.A completely different polynomial (Graph 12, Model 15), but with approximately the same accuracy level as the previous one (R 2 ≈0.97≈0.98)describes the NPLs behavior in the period of growth (Q3 2008-Q3 2012).It can also be used to forecast the NPLs movements in the foreseeable future (standard error of the model is 0.26).Under the assumption of no strong internal and/or external shocks and realistically estimated value of bank assets, at the end of 2013 NPLs should amount to 12.07 %, and at the end of 2014 to 9.82%.A polynomial is a sound mathematical and logical projection of the NPLSs development, because according to the empirical data, in the interval between 12% and 13% NPLs A trend line identified the possibility for a trend change already in Q4 2007 and Q3 2008 (Model 2).The support level, or, in line with the characteristics of the NPLs series, resistance level of 3% was formed in Q2 2007 and lasted until Q1 2009.Based on the formed resistance level, already in the second half of 2007 it could have been known that there would be a change in the trend and growth of NPLs (Model 3).Similar predictive ability is featured by momentum (MOM) and rate of change (ROC).MOM 10 suggests a change in the NPLs trend Graph 12. NPLs during an increase, (Q4 2008-Q3 2012), Model 15* Source: Ibid.Note: * Variable x signifies time.NPLs Poly.(NPLs) Graph 11.NPLs during a decline, (Q3 2003-Q3 2008), Model 14* Source: Ibid.Note: * Variable x signifies time.NPLs Poly.(NPLs) sljedeće tri godine (do 2007.g) godišnji pad DP je stalno ispod -1 p.p., da bi u 2008.g. došao do -0,1 p.p. Nakon toga, u 2009.g., DP rastu za 1,5 p.p. Indicija rasta LK postoji već u u 2007.g. kada je pad DP svega -0,5 p.p. (Model 6) što je za oko dvije trećine manje od -1,3 p.p. za koliko je DP smanjen u 2003.g.Usporenje pada DP je očigledno već u 2004.g.

Table 4 .
NPLs values based on the regression model and the expected GDP g.r.