New publications!



Abstract. The present paper is a review of two of the most widely used approaches to constructing an early warning system for banking crises: the econometric and the signals approach. Building upon the econometric approach, the paper goes on to construct its own early warning system for banking liquidity crises that, through analysing behaviour of chosen indicators, is able to predict a liquidity crisis one year ahead of its likely onset.

Keywords: banks, liquidity crisis, indicators, financial instability, early warning system

JEL Classification: G01, G21

STARKOVA, ALENA (2015) "EARLY WARNING SYSTEM FOR BANKING LIQUIDITY CRISES". Journal of Russian Review (ISSN 2313-1578), VOL. 2(3), 29-35

1. Introduction

Instability in the banking sector, be it through loss of deposits, substantial slumps in lending, bank failures, or one of the many other factors, is a sight not infrequent in both developed and developing economies. The banking sector woes in their turn may soon result in a financial crisis. Therefore, prevention of banking crises plays a pivotal role in maintaining financial stability both domestically and internationally.

A crisis early warning system (EWS) that signals the likelihood of the economy going into a decline within a certain time horizon is one of the measures used to prevent such crises. The present paper aims to construct an econometric approach-based early warning system for liquidity crises that could warn of the event one year ahead of the crunch.

2. Literature review

1990s saw a number of studies that dealt with establishing systems of early warning indicators. These models can be divided into two major groups according to the approach used in their construction: the econometric and the signals models.

The econometric models of crisis early warning systems usually involve constructing a multivariate regression model that would evaluate the correlation between selected indicators (macroeconomic, financial, monetary, etc.) and the probability of a crisis event. This most usually employs bivariate or multivariate logit and probit models, with probit models most often used for predicting currency crises, and logit models for banking, balance of payment, and other types of crises.

Among the first researchers to advocate econometric EWSs were IMF economists Asli Demirgüç-Kunt and Enrica Detragiache. In their 1998 paper The Determinants of Banking Crises in Developing and Developed Counties they focused on the economic environment factors that led to banking instability and, therefore, caused banking crises, and used these factors to construct a multivariate logit model [3]. For many researchers this model became their point of reference, and the present paper, too, will be making use of many of its computations.

Demirgüç-Kunt and Detragiache were not the only researchers seeking to design early warning systems for banking crises. Also working on an econometric logit model for predicting banking crises were economists Davis and Karim. Utilising Demirgüç-Kunt and Detragiache’s basic method, sources, and indicators, they used a much wider sample of 105 countries in the 1979-2003 period [2].

The economist Kasper Lund-Jensen presented his own fixed effect binary response dynamic model for conditional probability of a systemic banking crisis [5]. His paper identifies several important factors directly linked to risk of crisis in the banking sector: banking sector leverage, the credit-to-GDP gap, changes in the banks’ lending premium, equity price growth, the degree of banks’ interconnectedness, and real effective exchange rate appreciation. Lund-Jensen also developed a method of translating his systemic risk estimates into crisis probability signals, a method that provided accurate crisis signals in terms of type I and type II errors.

The basic premise of the signals approach is that the economy behaves differently on the eve of financial crises and that this aberrant behaviour has a recurrent systematic pattern manifested in a broad array of economic and financial indicators. Constructing the model, one must select a threshold or critical value that divides the probability distribution of that indicator into two regions. If the observed outcome for a particular variable falls into the rejection region, that variable is said to be sending a signal. If a variable is often seen sending ‘good’ signals (that is, proving highly efficient), one may expect that the probability of volatility if preceded by a signal (conditional probability) is higher than the unconditional probability.

The signals approach to crisis prognostication was pioneered by Graciela Kaminsky, Saul Lizondo, and Carmen M. Reinhart. It was their 1998 paper Leading Indicators of Currency Crises that brought about a widespread use of the signals model [4].

A more integrated approach to constructing a signals-based EWS was introduced by Claudio Borio and Mathias Drehmann. The underlying principles of their model, described in the 2009 paper Assessing the Risk of Banking Crises, bear strong resemblance to Kaminsky and Reinhart’s, even their method of historical data analysis for the identification and timing of banking crises is identical to the latter [1].

Despite the plethora of different models for predicting future crises developed by various researchers, none of them is totally satisfactory. It is not even possible to determine which approach, the signals or the econometric one, is more efficient.

The present paper will be using the econometric approach to prognosticating crises. The matter is that, choosing the significant variables, authors of signals-based EWSs tend to select macroeconomic indices. It may mean that the signals approach is more sensitive to underlying factors of instability hidden behind systemic failures of economic performance, while its econometric rivals are more sensitive to failings in the financial sector. Since we are more interested in the latter (because errors in the financial sector are taken to be more representative of national peculiarities), the econometric model for predicting liquidity crises was considered to be more appropriate.

3. Method

As stated earlier, the present econometric model of liquidity crisis early warning system will be based on the Demirgüç-Kunt and Detragiache model, and the Davis and Karim model derived from it.

The econometric model of early warning system for liquidity crisis in the banking sector will be based on a multivariate logit model
where x will be the indices (outlined below) characteristic of the onset of a liquidity crisis. Finding the values of the coefficients β will require constructing a learning sample containing data from 40 developing countries over the period of 1998 to 2012.

Crisis indicators, which the current model will introduce as independent variables, will include

  1. Macroeconomic indices:
    a. Real GDP growth
    b. Unemployment rate
    c. Inflation rate
  2. Exchange-related indices:
    a. National currency to US dollar exchange rate
    b. Foreign currency assets to total assets ratio
  3. Internal banking sector indices:
    a. Loan interest and deposit rates spread
    b. Real interest rate
  4. Balance sheet banking indices:
    a. Non-performing to total loans ratio (proportion of loans overdue by more than 90 days in total number of loans)
    b. Deposit growth
    c. Liquid assets to current liabilities ratio (liquidity rate)
    d. Volume of created reserves
    e. Liquid reserves to bank assets ratio
  5. Real economy indices:
    a. Household consumption
    b. Household savings

Let us specify why each of these indices can be included in the list of indicators.

Indices such as real GDP, unemployment and inflation rates describe the of internal conditions and may indicate that an economy is experiencing systemic difficulties. GDP is representative of the population’s standard of living. The larger the amount of finance at people’s disposal, the more money will be deposited in bank accounts, and a steady growth means lower risk of premature withdrawals of bank deposits. By contrast, a rise in inflation or unemployment rate means higher risk of premature withdrawals. With prospects of currency depreciation or loss of wage, citizens tend to turn to their savings to complement their income. Real sector indices, like macroeconomic indicators, are representative of how stable an economy is overall.

As was already noted, depreciation of a country’s currency makes investing in it significantly less attractive. Among other things, with subsequent currency appreciation the value of assets (loans) will go down and the banks will receive less funds than they expected.

Trends in various interest rates may also serve as an indicator of crises. An upward real interest rate trend indicates that the bank needs to raise substantial additional funds, which may result from its current underliquidity. The same may be said for a fall in interest rates to deposit rates ratio.
A rise in non-performing loans, along with a rise in provisions, means the banking sector is having trouble with its portfolio of assets, and may signify an increased risk of a bad debts crisis, which will inevitably lead to a run by depositors. A slowdown in deposit growth or a negative growth trend will indicate that a run on the banks is probably at hand. A fall in liquid assets to current liabilities ratio may speak of a mismatch in maturities of assets and liabilities on the balance sheet and a possible further increase in liquidity risk.

The second step in finding the values of the β coefficients is to determine the criteria for identifying crisis years to be used by the model.

The present paper’s author believes that the most telling indicators of a liquidity crisis are deposit growth, non-performing loans ratio, and liquidity rate. It is these indicators that will be used to determine whether an economy is going through a crisis or not. Therefore, this paper treats an economy as going through crisis if any one of these variables has deviated from its long-term trend downwards (for deposit growth and liquidity rate) or upwards (for non-performing loans ratio) by more than 50%. To build a model for predicting a crisis one year ahead of its onset, it is necessary to ignore all time periods from one year after the onset till after the crisis is over.

Having chosen the criteria for identifying a banking liquidity crisis, we proceed to determining the values of coefficients used by the model. To this end, the log-likelihood function will be employed, as was the case with Demirgüç-Kunt and Detragiache, and Davis and Karim:

The quality of the model built with these coefficients will then be evaluated by testing its efficiency on the data used in its construction.

Since the model’s output is the probability of a crisis for a given year, a threshold value must be determined beyond which one should start contemplating preventive measures. To do so, a threshold of probability must be selected starting from which an economy will be deemed set to experience a crisis in a year’s time. Unlike the previous values, this one will not be calculated by minimizing a total loss function, since parameters like crisis prevention cost (that is, the cost of crisis prevention measures) and crisis recovery costs (that is, the amount of funds expended on compensating for the impact of the crisis) cannot be accurately estimated.

Potential threshold values will be set at 20%, 40%, 60%, and 80%. Each value will be substituted into the model to determine the probability of type I (false positive) and type II (false negative) errors. Since crisis prevention (in case the model returns a false positive) and crisis compensation (in case the model returns a false negative) expenses cannot be accurately evaluated, the threshold value will be decided by overall model accuracy.

As the final test of its operability, the model will be fed the same countries’ data from 2013.

4. Intermediate results

When built around the above variables, the logit model, somewhat surprisingly, failed to prove the supposed significance of many of the original indicators (cf. Appendix 1). Only volume of created provisions and non-performing loans ratio variables were significant at the 5% level, and only household consumption and household savings variables were significant at the 10% level. Among the worst performing predictors of a future crisis were, surprisingly, the liquidity rate and foreign assets to total assets variables.

To improve its quality, adjustments were made to the model: instead of absolute values, first order differences of variables were used (see Appendix 2). This time the interest rate and non-performing to total loans ratio variables were significant at the 5% level, while the inflation rate, loan interest and deposit rates spread, volume of created reserves, and liquid reserves ratio variables were significant at the 10% level. In addition to that, there was a sharp increase in the significance of how the foreign assets to total assets ratio indicator was behaving.

Choosing between the two models, it was decided that Z values should be calculated with β coefficients generated by the second model. The calculations resulted in the following accuracy values:

Threshold values
20% 40% 60% 80%
Events predicted   73.0% 74.3% 79.7% 78.4%
Crisis years predicted 83.3% 33.3% 22.2% 11.1%
Non-crisis years predicted 69.6% 87.5% 98.2% 100.0%

Going by the method of choosing the threshold discussed previously, the ideal threshold value should be set at 60%. In this way the model is able to predict nearly 80% of events. The choice is also deemed well-founded because it enables the model to predict over 98% of non-crisis years, that is, to minimize crisis prevention expenditure in case of a false positive result.

Tested on the 2013 data, the result was 71% accurate. That is, the present model was able to correctly predict 71% of events, while correctly predicting 85% of non-crisis years and 19% of crisis years.

5. Conclusions

  1. According to the present model, the most significant indicators of a liquidity crisis are: interest rate, non-performing assets ratio, inflation rate, loan interest and deposit rates spread, reserve growth, and liquid reserves ratio.
  2. The present model sends out a signal one year ahead of a crisis whenever its probability exceeds 60%. Depending on their aims, individual users may substitute various threshold values: to maximize the number of correctly forecasted crisis years one must choose a 20% threshold, to maximize the number of correctly forecasted non-crisis years – an 80% one.

6. References

  1. Borio, C. and Drehmann, M. (2009) “Assessing the risk of banking crises revisited”, BIS Quarterly Review, March 2009
  2. Davis, E.P. and D. Karim, (2008), “Comparing early warning systems for banking crises,” Journal of Financial Stability, 4 (2), 89–120.
  3. Demirgüç-Kunt, A. and E. Detragiache, (1998), “The determinants of banking crises in developing and developed countries,” International Monetary Fund Staff Paper, 45 (1), 81–109
  4. Eichengreen, B. and Rose, AK. (1998), “Staying afloat when the wind shifts: External factors and emerging-market banking crises”, NBER Working Paper Series 1998.
  5. Lund-Jensen, K. (2012), “Monitoring Systemic Risk Based on Dynamic Thresholds”, IMF Working Paper WP/12/159.

7. Appendices

Appendix 1

Model 1: logit, observations 1 to 159
Dependent variable: Crisis
Standard errors - QML

  Coefficient Std. Error z Prob.  
const -22.8086 11.5277 -1.9786  0.04786  **
GDP -0.0919857 0.135535 -0.6787 0.49734  
Inflation -0.0244832 0.0516897 -0.4737 0.63574  
Unemployment -0.0411338 0.0575689 -0.7145 0.47491  
Exchange_rate -0.000150038 0.000290081 -0.5172 0.60500  
Interest_rate 0.0437718 0.0873304 0.5012 0.61622  
Spread_of_rates -0.0844109 0.10353 -0.8153 0.41489  
NPL -0.372395 0.184153 -2.0222 0.04316 **
Consumption 0.212121 0.113444 1.8698 0.06151 *
Savings 0.260769 0.145212 1.7958 0.07253 *
Deposits 4.12874 3.32037 1.2435 0.21370  
Foreighn_assets 0.00136639 0.0112114 0.1219 0.90300  
Liquidity_rate -0.00124636 0.0100808 -0.1236 0.90160  
Reserves 1.15026e-013 0 2.3042 0.02121 **
Liquid_reserves 0.0197005 0.0195555 1.0074 0.31373  


Mean dependent var 0.220183    S.D. dependent var 0.416284
McFadden’s R-squared 0.237065   Adjusted R-squared -0.023994
Log likelihood -43.83692   Akaike info criterion 117.6738
Schwarz criterion 158.0441   Hannan-Quinn criter.  134.0454

f(beta’x) for mean independent var = 0.416
Likelihood ratio: chi-square(14) = 27.2427 [0.0179]

Appendix 2

Model 2: logit, observations 1 to 159
Dependent variable: Crisis
Standard errors – QML

  Coefficient Std. Error z Prob.  
const -1.77579 0.46919 -3.7848 0.00015 ***
GDP -0.0376588 0.205549 -0.1832 0.85463  
Inflation 0.159227 0.0882775 1.8037 0.07128 *
Unemployment 0.290385 0.186652 1.5558 0.11977  
Exchange_rate 0.0021913 0.00258899 0.8464 0.39733  
Interest_rate 0.259595 0.107047 2.4250 0.01531  **
Spread_of_rates -0.460876 0.243956 -1.8892 0.05887 *
NPL -0.223213 0.112613 -1.9821 0.04746 **
Consumption 0.439753 0.507295 0.8669 0.38602  
Savings 0.408848 0.519446 0.7871 0.43123  
Deposits 3.71625 5.7172 0.6500 0.51568  
Foreighn_assets 0.0760201 0.0470772 1.6148 0.10635  
Liquidity_rate -0.00640899 0.0378503 -0.1693 0.86554  
Reserves 3.51435e-013 1.85219e-013 1.8974 0.05778 *
Liquid_reserves -0.133402 0.0714644 -1.8667 0.06195 *


Mean dependent var 0.243243    S.D. dependent var  0.431969
McFadden’s R-squared 0.208802   Adjusted R-squared -0.156567
Log likelihood -32.48219   Akaike info criterion 94.96438
Schwarz criterion 129.5254    Hannan-Quinn criter.  108.7512


f(beta’x) for mean independent var = 0.432
Likelihood ratio: chi-square(14) = 17.1445 [0.2486]


скачать dle 10.2 Авто Тюнинг кузова