EARLY WARNING SYSTEMS FOR BANKING

CRISES

E Philip Davis

NIESR and Brunel University

West London

e_philip_davis@msn.com

www.ephilipdavis.com

groups.yahoo.com/group/financial_stability

Course on Financial Instability at the Estonian Central Bank,9-11 December 2009 – Lecture 7

Introduction

• 3 types of models for early warning, logit, signal extraction and binary recursive tree

• We apply the models first to prediction of crises in Asia

• And then outline a new logit approach which predicts banking crises in OECD countries

Early warning systems

• Multivariate logit model uses macroeconomic, institutional and financial variables X as inputs to calculate probability of a banking crisis Y as the output via logistic function estimator. Suitable for answering question “what is the likelihood of a banking crisis occurring in the next t years?”

it

it

X'

X'

itite1

eXF1YobPr

• Non-parametric signal extraction approach tracks individual time series X prior to and during crisis episodes to answer question “is there a signal S of future crisis or not?” If an input variable’s aberrant behaviour can be quantitatively defined whenever that variable moves from tranquil to abnormal activity, a crisis is forewarned.

• { S ij = 1 } = { │ Xi

j │ > │ X*ij │ } or

• { S ij = 0 } = { │ Xi

j │ < │ X*ij │ }

• Binary Recursive Tree (BRT) can be used to answer question “which non-linear variable interactions make an economy more vulnerable to crisis than others?” Argued that liquidity, credit and market risks are all potentially non-linear. Estimator identifies single most important discriminator between crisis and non-crisis episodes across the entire sample, thereby creating two nodes. Nodes are further split into sub-nodes based on the behaviour of splitter variables’ non-linear interactions with previous splitter variables. This generates nodal crisis probabilities and the associated splitter threshold values.

Entire Sample: 72 crises

PARENT NODE

Child Node 2: 20 crises

Child Node 1: 52 crises

Terminal Node 3: 48 crises

Terminal Node 3: 4 crises

Terminal Node 4: 17 crises

Terminal Node 5: 3 crises

Splitter Variable: X1

X1≤ V1* X1>V1

*

Splitter Variable: X2

X2≤ V2*

X2> V2* X3≤ V3

*

Figure 4: Schematic Diagram of Binary Recursive Tree (BRT)

X3≥ V3*

Splitter Variable: X3

Advantages and disadvantages• Logistic models are ideally suited to predicting a binary outcome

(1 = banking crisis, 0 = no banking crisis) using multiple explanatory variables selected on the basis of their theoretical or observed associations with banking crises.

• Logistic approach is also parametric, generating confidence intervals attached to coefficient values and their significance, but logit coefficients are not intuitive to interpret and they do not reflect the threshold effects that may be simultaneously exerted by other variables.

• Signal extraction non parametric and can use high frequency data• Logit approach is the most appropriate for use as a global EWS,

while signal extraction methods are more appropriate for a country-specific EWS (Davis and Karim 2008).

• BRT is able to discover non-linear variable interactions, making it especially applicable to large banking crises datasets where many cross-sections are necessary to generate enough banking crisis observations and numerous factors determine the occurrence of systemic failure.

• In BRT no specific statistical distribution needs be imposed on the explanatory variables. Also not necessary to assume all variables follow identical distributions or that each variable adopts the same distribution across cross-sections.

• Although logistic regression does not require variables to follow any specific distribution, Davis and Karim (2008) showed that standardising variables displaying heterogeneity across countries improved the predictive performance of logit models.

• Logistic regressions are also sensitive to outlier effects, yet it is precisely the non-linear threshold effects exerted by some variables that could generate anomalous values in the data.

• In low risk, stable regimes, variables may conform to a particular distribution which subsequently jumps to a regime of financial instability. Non-parametric BRTs should handle such data patterns better than logistic regressions.

• BRT is extremely intuitive to interpret. The model output is represented as a tree which is successively split at the threshold values of variables that are deemed as important contributors to banking crises.

• Signal extraction is also easier to interpret than logit, but is vulnerable to ignoring multivariate patterns at core of instability

Illustrative results – logit (Asia)Variable Coefficient z-Statistic Coefficient z-Statistic

DCRED(-1) -0.033902 -2.091298 -0.032416 -2.046609 GDPPC(-1) -0.000246 -3.451172 -0.000235 -3.535303 FISCY(-1) 0.010451 0.153806

INFL(-1) -0.037791 -1.212934 RIR(-1) 0.114829 2.528462 0.113567 2.612414

DEPREC(-1) 0.053493 2.724725 0.044323 2.712526 DCREDY(-1) 0.022844 2.971898 0.021231 2.959820

DTT(-1) 0.007492 0.322193

DGDP(-1) -0.261366 -3.853235 -0.276748 -4.192324 M2RES(-1) -0.000549 -2.232728 -0.000536 -2.190088

Expectation-Prediction Evaluation for Binary Specification Equation: IND_STAND Date: 12/03/09 Time: 19:46 Success cutoff: C = 0.25

Estimated Equation Dep=0 Dep=1 Total P(Dep=1)<=C 80 7 87

P(Dep=1)>C 34 41 75 Total 114 48 162

Correct 80 41 121 % Correct 70.18 85.42 74.69

% Incorrect 29.82 14.58 25.31 Total Gain* 70.18 -14.58 45.06

Percent Gain** 70.18 NA 64.04

Signal extraction - Asia

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

0.5 1 2 3.5 4 4.5 6 8 10 20

Percentile Threshold

NT

SR

GDP grow th Change Terms of Trade Depreciation

Real Interest Rate Inflation Fiscal Surplus/ GDP

M2/Reserves GDP per Capita

Real GDP growth Fiscal surplus/GDP Depreciation % crises correct 10 8 6

% no crises correct 99 98 98 % total correct 65 64 62

BRT - Asia

DGDP <= 4.75

TerminalNode 1

Class Cases %0 7 20.61 27 79.4

DCREDY <= 60.49

TerminalNode 2

Class Cases %0 18 100.01 0 0.0

DCREDY > 60.49

TerminalNode 3

Class Cases %0 16 57.11 12 42.9

DGDP > 4.75

Node 3Class Cases %

0 34 73.91 12 26.1

FISCY <= -1.14

Node 2Class Cases %

0 41 51.31 39 48.8

FISCY > -1.14

TerminalNode 4

Class Cases %0 79 89.81 9 10.2

Node 1Class Cases %

0 120 71.41 48 28.6

Asia % crises correct 46

% no crises correct 90 % total correct 84

Leading indicator selection

Asia

Logit Signal Extraction Tree

Real GDP Growth Real Interest Rate Inflation Fiscal Surplus/ GDP M2/ Foreign Exchange Reserves Real Domestic Credit Growth Real GDP per capita Domestic credit/GDP Depreciation Terms of Trade Current account/GDP External short term debt/GDP

A new model for the OECD• Existing work on early warning systems (EWS) for

banking crises generally omits bank capital, bank liquidity and property prices, despite their relevance to the probability of crisis in the mind of bankers, policymakers and the public. One reason for this neglect is that most work on EWS to date has been for heterogeneous global samples dominated by emerging market crises. For such countries, time series data on bank capital adequacy and property prices are typically absent, while other variables affecting crises may also differ in OECD countries.

• We argue results are misspecified

• Triggers of crisis depend on the type of economy and banking system. In OECD countries with high levels of banking intermediation and developed financial markets, shocks to terms of trade are less important crisis triggers than, say, property price bubbles.

• Also developed economy banking systems are more likely to be regulated in terms of capital adequacy and liquidity ratios

• Accordingly, we estimate logit models of crisis for OECD countries only and find strong effects of capital adequacy, liquidity ratios and property prices, such as to exclude most traditional variables. Our results imply that higher unweighted capital adequacy as well as liquidity ratios has a marked effect on the probability of a banking crisis, implying long run benefits to offset some of the costs that such regulations may impose (e.g. widening of bank spreads).

Methodology and data• Multivariate logit with dependent variable being crisis

probability• Problems of crisis dummies

– Definition of banking crises– Start and end dates ambiguous– Focus on switch date in core results

• Data partitioned to 1980-2006 and 2007 to leave subprime crisis for out-of-sample

• Variables for bank regulation:– Unweighted capital adequacy ratio - ratio of capital and reserves

for all banks to the end of year total assets– Liquidity - ratio of the sum of cash and balances with central

banks and securities for all banks over the end of year total assets

Table of crises in sampleBG CN DK FN FR GE IT JP NL NW SP SD UK US

1980 0 0 0 0 0 0 0 0 0 0 0 0 0 01981 0 0 0 0 0 0 0 0 0 0 0 0 0 01982 0 0 0 0 0 0 0 0 0 0 0 0 0 01983 0 1 0 0 0 0 0 0 0 0 0 0 0 01984 0 0 0 0 0 0 0 0 0 0 0 0 1 01985 0 0 0 0 0 0 0 0 0 0 0 0 0 01986 0 0 0 0 0 0 0 0 0 0 0 0 0 01987 0 0 1 0 0 0 0 0 0 0 0 0 0 01988 0 0 0 0 0 0 0 0 0 0 0 0 0 11989 0 0 0 0 0 0 0 0 0 0 0 0 0 01990 0 0 0 0 0 0 1 0 0 1 0 0 0 01991 0 0 0 1 0 0 0 1 0 0 0 1 1 01992 0 0 0 0 0 0 0 0 0 0 0 0 0 01993 0 0 0 0 0 0 0 0 0 0 0 0 0 01994 0 0 0 0 1 0 0 0 0 0 0 0 0 01995 0 0 0 0 0 0 0 0 0 0 0 0 1 01996 0 0 0 0 0 0 0 0 0 0 0 0 0 01997 0 0 0 0 0 0 0 0 0 0 0 0 0 01998 0 0 0 0 0 0 0 0 0 0 0 0 0 01999 0 0 0 0 0 0 0 0 0 0 0 0 0 02000 0 0 0 0 0 0 0 0 0 0 0 0 0 02001 0 0 0 0 0 0 0 0 0 0 0 0 0 02002 0 0 0 0 0 0 0 0 0 0 0 0 0 02003 0 0 0 0 0 0 0 0 0 0 0 0 0 02004 0 0 0 0 0 0 0 0 0 0 0 0 0 02005 0 0 0 0 0 0 0 0 0 0 0 0 0 02006 0 0 0 0 0 0 0 0 0 0 0 0 0 02007 0 0 0 0 0 0 0 0 0 0 0 0 1 1

Box 1: List of Variables (with variable key)

1. Real GDP Growth (%) (YG) 2. Real Interest Rate (%) (RIR) 3. Inflation (%) (INFL) 4. Fiscal Surplus/ GDP (%) (BB) 5. M2/ Foreign Exchange Reserves (%) (M2RES)

Variables used in previous studies:

Demirguc-Kunt and Detragiache (2005);

Davis and Karim (2008). 6. Real Domestic Credit Growth (%) (DCG) 7. Liquidity ratio (%) (LIQ) 8. Unweighted capital adequacy ratio (%) (LEV)

Variables introduced in this study.

9. Real Property Price Growth (%) (RHPG)

it

it

X'

X'

itite1

eXF1YobPr

(1)

n

i

T

titeititeite XFYXFYLLog

1 1

'1log1'log

Table 2: The General To Specific Approach

LIQ(-1 )-0 .118 (-3.55)

-0.124 (-3 .55)

-0 .137 ( -3 .64)

-0 .135 ( -3.55)

-0 .135 ( -3 .45)

-0.144 (-3 .39)

-0.147 (-3.25)

LEV(-1)-0 .333 (-2.85)

-0.239 (-1 .90)

-0 .315 ( -2 .24)

-0 .247 ( -1.64)

-0 .271 ( -1 .67)

-0.280 (-1 .72)

-0.273 (-1.62)

R HPG(-3)0 .113 (2.8)

0.113 (2.87)

0.104 (2 .67)

0 .100 (2 .59)

0.104 (2.67)

0.108 (2.76)

0 .110 (2.67)

D CG(-1) --0.099 (-1 .82)

-0 .10 ( -1 .97)

-0.10 ( -1.86)

-0 .10 ( -1 .99)

-0 .13 (-1 .98)

-0.13 (-1.98)

R IR (-1) - -0.084 (1 .37)

0 .085 (1 .40)

0.165 (1.41)

0.173 (1.46)

0 .166 (1.30)

M 2RES(-1) - - --0.00 (-1.0)

-0 .00 ( -1.0)

-0 .00 ( -1 .1)

-0.00 (-1 .1)

INFL(-1) - - - --0 .13 ( -0.8)

-0 .14 ( -0 .8)

-0.13 (-0 .7)

YG(-1) - - - - -0.116 (0.65)

0 .125 (0.66)

BB(-1) - - - - - --0.013 (-0 .1)

Note: estimation period 1980-2006; t-statistics in parentheses; LIQ-liquidity ratio, LEV- unweighted capital

adequacy ratio, YG-real GDP growth, RPHG-real house price inflation, BB-budget balance to GDP ratio,

DCG-domestic credit growth, M2RES-M2 to reserves ratio, RIR-real interest rates, DEP-depreciation, INFL-

inflation.

Table 3: Comparing the Effects of Sample Period on Estimation Results

1 980-2006 19 80-2 007

L IQ-0.11 8 (-3. 55)

-0.13 (-4.1 )

LE V-0.33 3 (-2. 85)

-0 .261 (-2. 51)

P HG0.113 (2. 8)

0.10 6 (2. 79)

E s t im at ion period

log

p(crisis)-1

p(crisis) = - 0.333 LEV(-1) – 0.118 LIQ(-1) + 0.113 RHPG(-3) (3)

(-2.85) (-3.55) (2.8)

Marginal effect of 1% rise in variable on crisis probability

LIQ LEV RHPGBG -0.17 -0.49 0.17CN -0.22 -0.61 0.21DK -0.05 -0.14 0.05FN -0.23 -0.65 0.22FR -0.78 -2.17 0.74GE -0.23 -0.65 0.22IT -0.17 -0.46 0.16JP -0.38 -1.05 0.36NL -0.56 -1.57 0.53NW -0.33 -0.91 0.31SD -0.12 -0.34 0.12SP -0.08 -0.24 0.08UK -1.19 -3.32 1.13US -0.08 -0.22 0.07

Crisis probabilities

0.00

0.20

0.40

0.60

0.80

1.00

Probability Crisis

In sample prediction

Total

Calls Crises

Aftermath

of the

Crises

False

Calls Timing of False Calls relative to Crisis Onset

BG 0 0 0 0

CN 6 1 1 4 next year

DK 0 0 0 0

FN 10 1 1 8 next year

FR 14 1 0 13

GE 4 0 0 4

IT 7 0 2 5 2nd and 3rd years

JP 15 1 6 8 Next 7 years, with a break on the 4th year

NL 18 0 0 18

NW 14 1 2 11 next 2 years

SD 6 1 1 4 next year

SP 2 0 0 2

UK 20 2 0 18

US 0 0 0 0

total 116 8 13 95

Out of sample predictions2007 2008 definition1 definition2

BG X X X XCN - - -DK - -FN - XFR X X X XGE - - - -IT X - XJP - -NL X - X XNW X XSD - - -SP X X XUK X X X XUS - - - -

Country elimination tests

LIQ(-1)-0.118 (-3.55)

-0.143 (-2.99)

-0.125 (-3.55)

-0.111 (-3.28)

-0.119 (-3.29)

-0.124 (-3.59)

-0.121 (-3.5)

-0.115 (-3.41)

LEV(-1)-0.333 (-2.85)

-0.3 (-1.78)

-0.339 (-2.79)

-0.344 (-2.94)

-0.349 (-2.86)

-0.282 (-2.38)

-0.293 (-2.43)

-0.343 (-2.87)

PHG(-3)0.113 (2.8)

0.152 (3.44)

0.119 (2.82)

0.111 (2.74)

0.118 (2.76)

0.089 (2.04)

0.083 (1.84)

0.107 (2.58)

Final panel

US not included

Japan not included

US and Japan not included

UK not included

Norway not

included

Finland not

included

Sweden not

included

Alternative crisis dates

LIQ(-1)-0.118 (-3.55)

-0.119 (-3.56)

-0.12 (-3.58)

LEV(-1)-0.333 (-2.85)

-0.332 (-2.85)

-0.317 (-2.73)

PHG(-3)0.113 (2.8)

0.113 (2.8)

0.104 (2.56)

Japanese crisis at

1992

US crisis at 1984

Final version

Aftermath elimination and subprime runup

LIQ(-1)-0.118 (-3.55)

-0.111 (-3.48)

LEV(-1)-0.333 (-2.85)

-0.329 (-2.91)

PHG(-3)0.113 (2.8)

0.111 (2.74)

Final version

Aftermath of the Crisis

LIQ(-1)-0.118 (-3.55)

-0.128 (-3.4)

LEV(-1)-0.333 (-2.85)

-0.241 (-1.94)

RHPG(-3)0.113 (2.8)

0.106 (2.85)

LIQ(-1)b --0.029 (-0.34)

LEV(-1)b --0.045 (-0.19)

RHPG(-3)b -0.006 (0.05)

Final version

1980-2007 estimation with break

Further lags and systemic crises

LIQ (-2) -0.104 (-3.27)

LEV (-2) -0.385 (-3.22)

PHG (-3) 0.119 (3.00)

LIQ (-1) -0.121 (-2.49)

LEV (-1) -0.768 (-3.59)

PHG (-3) 0.235 (3.71)

Conclusions• 3 approaches complementary• Traditional approaches fruitful for EMEs such as Asia but

not for OECD countries• Found relevance of bank capital, liquidity and property

prices absent from traditional EWS, exclude traditional variables

• Can predict crises out of sample and specification is robust• Warrants policy focus on bank regulation – of capital,

liquidity but also of terms of mortgages loans• Also supports measures to reduce procyclicality, adjusting

capital or provisions countercyclically – and use of simple leverage ratio as well as risk weighted capital adequacy

References• Davis, E P and D Karim (2008a), "Comparing early

warning systems for banking crises", Journal of Financial Stability, 4, 89-120

• Davis E P and Karim D (2008b), "Could early warnings systems have helped to predict the subprime crisis?", National Institute Economic Review, 206, 25-37 and Brunel University Economics and Finance Working Paper No 08-27

• Barrell R, Davis E P, Karim D and Liadze I (2009), "Bank Regulation, Property Prices And Early Warning Systems For Banking Crises In OECD Countries", NIESR Discussion Paper No. 330