|
|
The recent global financial crisis has reinvigorated interest in models capable of identifying the warning signs of crisis. Early warning models are typically based on empirical logistic regressions. However, common regression-based models are unable to capture important nonlinearities and complex interactions between macroeconomic and financial variables that may exist in the run-up to crises.
To address these issues authors use Classification and Regression Tree (CART) methodology and its generalisation, Random Forest (RF) analysis, to model explicitly the non-linear interactions between variables and deal with missing values and outliers, which are usually a problem for regression-based frameworks. The CART and RF frameworks provide crisis thresholds for key variables, thus significantly simplifying the interpretation of the results for decision-makers and non-technical audiences.
This framework has both advantages and disadvantages compared with other common early warning methods. On the one hand, it allows explicitly for the fact that not all crises are alike and accommodates non-linearities by including conditional thresholds. On the other hand, it is a nonparametric approach that cannot estimate the marginal contributions of each explanatory variable or confidence intervals for the estimated thresholds.
Authors apply the CART and RF techniques on an unbalanced panel dataset consisting of 36 advanced countries between 1970 and 2010. They investigate what macroeconomic, financial and structural conditions prevailed in the economies in the periods ahead of banking and currency crises in this period.
Their results suggest that one-to-two years ahead of a banking crisis the net interest rate spread in the banking sector is the key predictor of crisis. The crises are more likely to occur when this spread is low. On the contrary, if net interest rate spreads in the banking sector are high, then a flat or inverted yield curve becomes the crucial predictor of banking crisis. Authors interpret this as evidence that the term spread can be thought of as representing the marginal profitability of bank lending, and compression of the term spread, if occurring at the peak of a banking boom, can be a causal signal of bust. Two-to-three years ahead of a banking crisis, house prices seem to be the most important predictor of crisis onset. As for currency crises, the most powerful predictor one-to-two years ahead is exchange rate overvaluation combined with high domestic short-term interest rates.
Authors also evaluate the importance of country-specific structural factors and international variables in predicting crises. For banking crises, they find that both country structural characteristics and international developments are relevant crisis predictors. Currency crises, however, seem to be driven mainly by country-specific short-term developments.
It should be noted that their results, like the results of any early warning model, are conditioned on the country sample, time span and predictors. As such, the results should be considered mainly as a structured presentation of past experience and should not, without due care, be used for predicting future crises.