Exchange rate prediction is a crucial aspect of financial analysis, particularly for countries with dual currency systems such as Cambodia. The Khmer Riel (KHR) is widely exchanged with the United States Dollar (USD), making accurate prediction models essential for businesses and policymakers. This study explores ensemble learning and deep learning approaches to predict the USD/KHR daily exchange rate using historical data collected from the National Bank of Cambodia (NBC). This study evaluates the performance of four advanced machine learning models—Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost)—for predicting the KHR/USD daily exchange rate. These models were rigorously compared against a traditional ARIMA statistical baseline. The analysis utilized standard time-series evaluation metrics, including Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and R-squared (R2). The results demonstrate that all advanced machine learning models significantly outperform the ARIMA baseline, which was found to be unsuitable for this dataset. Among the experimental models, the deep learning architectures, GRU and LSTM, performed best. They achieved an RMSE of 6.36 KHR and 6.79 KHR, respectively, along with the highest R2 scores of 0.963 and 0.959. While the RF and XGBoost models were also highly competitive, they had a crucial limitation: their inability to extrapolate beyond their training data range (4110 KHR), which resulted in lower accuracy than the deep learning models. This research confirms that both deep learning and tree-based ensemble methods are highly effective at capturing the complex, non-linear dynamics of financial time series data