The carry trade strategy is a well-known and time-tested investment strategy that arbitrages the difference in interest rates offered by different currencies. However, the carry trade exposes investors to political risks and currency crises, which are historically difficult to predict because currency crises are often a self-fulfilling prophecy. This dissertation investigates the efficacy of a Gaussian Mixture Variational Auto-Encoder (GMVAE) in forecasting panel data of currency exchange rates relative to the U.S. Dollar. The non-linear projection capabilities of the GMVAE allow more nuanced regime clustering compared to conventional econometric models while the one-step optimization process spanning both projection and clustering stages enables rapid updates. We find that the GMVAE is able to dynamically respond to new circumstances, making it perform better for risk-sensitive investment strategies.