Navigating the complexities of the foreign exchange market remains a significant challenge for investors and risk managers worldwide. This study addresses this challenge by developing a Deep Learning-based Multilayer Perceptron (DL MLP) model to classify daily multi-currency exchange rate returns (USD, EUR, GBP, CHF, JPY, CNY) from 2010 to 2024. The framework categorizes market movements into three distinct states: appreciation (good returns), depreciation (bad returns), and stability (no returns), transforming these categorical outcomes into a numerical format suitable for neural network processing via one-hot encoding. Powered by a backpropagation algorithm, the model demonstrates exceptional predictive capability, achieving a 91% classification accuracy and near-perfect AUC ROC scores, thereby identifying optimal safe-haven currencies with high reliability. The empirical findings offer a powerful data-driven tool for forex investors seeking to refine their strategic asset allocation and for global risk managers aiming to enhance their hedging strategies against currency volatility.