We propose a highly optimized latent factor representation of the yield curve obtained by training a variational autoencoder (VAE) to curve data from multiple currencies. A curious byproduct of such training is a "world map of latent space" where neighbors have similar curve shapes, and distant...