financial data may receive different labels as only one of them experienced a crisis in the near term. A supervised learner would try to separate both countries even though from a vulnerability perspective both countries belong to the same class. We explain the biased label problem in detail below.
unFEAR avoids the biased label problem using unsupervised learning to find clusters using information in the distribution of the economic and financial data. Rather than working with the raw data unFEAR leverages on the use of autoencoders to reduce the dimensionality of the
We introduce unFEAR, Unsupervised Feature Extraction Clustering, to identify economic crisis regimes. Given labeled crisis and non-crisis episodes and the corresponding features values, unFEAR uses unsupervised representation learning and a novel mode contrastive autoencoder to group episodes into time-invariant non-overlapping clusters, each of which could be identified with a different regime. The likelihood that a country may experience an econmic crisis could be set equal to its cluster crisis frequency. Moreover, unFEAR could serve as a first step towards developing cluster-specific crisis prediction models tailored to each crisis regime.
6. Missing data imputation using an autoencoder
7. A Boosted Autoencoder
8. The Mode Constrastive Autoencoder
9. Time detrended data clusters: K-mean clusters and time periods
10. Time detrended data clusters: country and crisis presence
11. Time detrended balanced data: clusters and crisis/non-crisis observations
12. Scree plot for cluster selection
13. Mode Contrastive Autoencoder: clusters and crisis data points
14. Mode Contrastive Autoencoder: residuals
1. Country list
2. Crisis clusters: empirical and shadow crisis