A bi-annual forecasting model of currency crises

ABSTRACT

This study proposes a novel approach that combines random forests and discrete wavelet transform (DWT) to construct a bi-annual forecasting model of currency crises. The proposed model can achieve a reasonably high level of accuracy in predicting crises and demonstrates that the DWT of monthly real exchange rates and foreign reserves can serve as reliable predictors. The predicted probability of crises in individual countries is visualized through a map, which indicates that the risk of crises has increased substantially across regions in the second half of 2018.

KEYWORDS:

• Currency crisis
• forecasting model
• wavelet transform
• random forests

JEL CLASSIFICATION:

• F31
• F37
• F40

Acknowledgements

The authors are grateful to the editor and anonymous referee for their helpful comments on an earlier version of the paper.

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

¹ MODWT is computed using the pyramid algorithm proposed by Percival and Walden (2000). The computation is conducted using the ‘Wavelets’ package in R.

² The monthly DWT predictors are updated each period so that ex-post data is not used to construct the predictors. The method of reflection is used to handle the boundary coefficients.

³ Although the original index had interest rate differentials, Kaminsky and Reinhart (1999) removed it from their index because developing countries often adopt interest rate control. Since our sample includes many developing countries, we exclude interest rate differentials from the index. Note also that real exchange rates are used instead of nominal exchange rates to take into account the differences in inflation rates across countries.

⁴ The computation is conducted using the ‘randomForest’ package in R.

⁵ See Table 1 for the list of countries. The United States is excluded because the US dollar is used as a numeraire while the euro area countries are excluded owing to the discontinuity of data.

⁶ To reduce model bias arising from the imbalance in the sizes between the crisis and non-crisis samples, this study uses random sampling to down-sample the majority class, that is, the non-crisis sample.

⁷ In the ten-fold cross validation, the samples are randomly split into 10 folds of roughly equal sizes. One fold serves for evaluating the predictive accuracy and the remaining nine folds are used to build the model. The procedure is repeated 10 times and the average predictive accuracy is computed.

⁸ The sensitivity is one minus Type II error, while Type I error is one minus the specificity. The precision is one minus the false discovery proportion.

⁹ The balanced accuracy for the prediction of 2018H1 is 0.924 while that for 1998H1, 2009H1, and 2015H1, which recorded a large number of crises, is 0.804, 0.750, and 0.767, respectively.

Additional information

Funding

This work was supported by the Japan Society for the Promotion of Science [17H00983,17K18564,18K01610].