Accurate demand forecasting within perishable products characterized by immediate service delivery, short lifespan ingredients, and high intangibility, is imperative. This research highlights the effectiveness of Decision Tree Regression (DT) and Random Forests (RF) that are more robust than traditional statistical methods in predicting aggregate demand of fresh goods with non—stationary and low—dimensional datasets. In this study, DT and RF have been applied to the weekly aggregate demand of beverages in the Espresso category, with the expectation that these machine learning models will outperform traditional statistical methods in terms of accuracy and generalizability. Using the real dataset about weekly demand from a coffee company in Vietnam, we compared the machine learning methods with traditional methods such as Simple Exponential Smoothing (SES), Autoregressive Integrated Moving Average (ARIMA) and Multiple Linear Regression and we found that DT outperforms the baseline, mean model by a large margin, reducing Mean Absolute Error (MAE) by 38.4, Symmetric Mean Absolute Percentage Error (sMAPE) by 14.5% and Relative MAE (RelMAE) by 0.560 in the Espresso beverages.