In previous studies, researchers have proposed using machine learning as an alternative to statistical methods for predicting time-series data. Artificial neural networks (ANN), particularly recurrent neural networks (RNN), are frequently employed to predict time-series data and to improve the accuracy of forecasting models. Accurate demand prediction helps streamline supply chain operations, enhance customer satisfaction, and prevent stock shortages. When inaccuracies occur in forecasting demand, it leads to uncertainty in the supply chain and result in inventory levels at all stages. Therefore, it is critical to strive for demand forecasts to maintain an efficient supply chain. This research uses a variety of time series and machine learning techniques to predict outcomes, such, as ARIMA (Autoregressive Integrated Moving Average), VAR (Vector Autoregressive), MLP (Multilayer Perceptron), and RNN (Recurrent Neural Network) focusing specifically on Elman RNN and Jordan RNN models. In this study, the effectiveness of all the methods was assessed by comparing their ability to predict cement demand using the mean absolute percentage error (MAPE) values. The dataset for this analysis was collected from a cement company in Indonesia from January 2015 to September 2023. The findings show that the Jordan RNN approach produced the smallest MAPE values in Bali and the East Nusa Tenggara regions; meanwhile the ARIMA method produced the smallest MAPE in the West Nusa Tenggara area. These outcomes are also affected by the nonlinearity observed in cement demand data for Bali and the East Nusa Tenggara region, which highlights the efficacy of using the Jordan RNN method for forecasting compared to other techniques.