Early Warning Performance of Univariate Credit-to-GDP Gaps

We use European and simulated Hungarian data to search for the univariate one-sided credit-to-GDP gap that predicts systemic banking crises most accurately. The credit-to-GDP gaps under review are optimized along four dimensions: (1) definition of outstanding credit, (2) forecasting method for extending credit-to-GDP time series, (3) filtering method and (4) maximum cycle length. Based on European data, we demonstrate that credit-to-GDP gaps calculated with narrow definition of outstanding credit and up to 1-year forecasts of credit-to-GDP outperform other specifications significantly and robustly. Regarding the other two dimensions, the Hodrick–Prescott filter with long cycles (popular in regulatory practice), the Christiano–Fitzgerald filter with medium-term cycles and the wavelet filter with short cycles prove to be the best. All three should be applied to credit-to-GDP time series calculated with narrow credit, and with no credit-to-GDP forecast, except the wavelet filter with short-term forecast. Credit-to-GDP gaps with most informative early warning signals exhibit the highest degree of comovement with the financial cycle, but not the lowest level of endpoint uncertainty. Analysis of Hungarian credit-to-GDP time series extended by ARIMA simulations reinforces the early warning quality of the Hodrick–Prescott credit gap and the wavelet credit gap to a lesser extent.