Estimation of fluid saturation is an important step in dynamic reservoir characterization. Machine learning techniques have been increasingly used in recent years for reservoir saturation prediction workflows. However, most of these studies require input parameters derived from cores, petrophysical logs, or seismic data, which may not always be readily available. Additionally, very few studies incorporate the production data, which is an important reflection of the dynamic reservoir properties and also typically the most frequently and reliably measured quantity throughout the life of a field. In this research, the random forest ensemble machine learning algorithm is implemented that uses the field-wide production and injection data (both measured at the surface) as the only input parameters to predict the time-lapse oil saturation profiles at well locations. The algorithm is optimized using feature selection based on feature importance score and Pearson correlation coefficient, in combination with geophysical domain-knowledge. The workflow is demonstrated using the actual field data from a structurally complex, heterogeneous, and heavily faulted offshore reservoir. The random forest model captures the trends from three and a half years of historical field production, injection, and simulated saturation data to predict future time-lapse oil saturation profiles at four deviated well locations with over 90% R-square, less than 6% Root Mean Square Error, and less than 7% Mean Absolute Percentage Error, in each case.