Many estimates of early-warning-system (EWS) models of currency crisis have reported incorrect standard errors because of serial correlation in the context of panel probit regressions. This paper documents the magnitude of the problem, proposes and tests a solution, and applies it to previously published EWS estimates. We find that (1) the uncorrected probit estimates substantially underestimate the true standard errors, by up to a factor of four; (2) a heteroskedasicity- and autocorrelation-corrected (HAC) procedure produces accurate estimates; and (3) most variables from the original models remain significant, though substantially less so than had been previously thought.
increases ( Alexander et al. 2009 ; van Dijk, 2013 ).
Macroprudential policy has an important role in crisis prevention and crisis mitigation. The policy effectiveness, however, hinges on whether the macroprudential tools can target the root causes of economic crises. Crisisprediction models, hence, need to support macroprudential policy. By flagging in advance economic and financial conditions leading to an economic crises, the models can guide policy actions aimed at reducing the crisis likelihood.
This paper proposes unFEAR , an unsupervised feature extraction