EARLY WARNING SYSTEMS FOR BANKING CRISES

1. EARLY WARNING SYSTEMS FOR BANKING CRISES Course on Financial Instability at the Estonian Central Bank, 9-11 December 2009 – Lecture 7 E Philip Davis NIESR and Brunel University West London e_philip_davis@msn.com www.ephilipdavis.com groups.yahoo.com/group/financial_stability

2. 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

3. 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?”

4. 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 } = { │ Xij │ > │ X*ij │ } or • { S ij = 0 } = { │ Xij │ < │ X*ij │ }

5. 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.

6. Figure 4: Schematic Diagram of Binary Recursive Tree (BRT) Entire Sample: 72 crises PARENT NODE X1≤ V1* X1>V1* SplitterVariable: X1 Child Node 1: 52 crises Child Node 2: 20 crises Splitter Variable: X2 Splitter Variable: X3 X2≤ V2* X2> V2* X3≤ V3* X3≥ V3* Terminal Node 3: 48 crises Terminal Node 3: 4 crises Terminal Node 4: 17 crises Terminal Node 5: 3 crises

7. 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).

8. 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.

9. 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

10. Illustrative results – logit (Asia)

11. Signal extraction - Asia

12. BRT - Asia

13. Leading indicator selection

14. 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

15. 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).

16. 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

17. Table of crises in sample

21. Marginal effect of 1% rise in variable on crisis probability

22. Crisis probabilities

23. In sample prediction

24. Out of sample predictions

25. Country elimination tests

26. Alternative crisis dates

27. Aftermath elimination and subprime runup

28. Further lags and systemic crises

29. 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

30. 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