The demand is often forecasted using econometric (regression) or statistical forecasting methods. However, most of these methods lack the ability to model both temporal (linear and nonlinear) and covariates-based variations in a demand series simultaneously. In this context, a novel forecasting model is proposed that combines a state-of-the-art sequence modeling method and a machine learning method in an ensemble model. The proposed model can handle both types of variations in demand data, and thus, enhances forecasts’ accuracy. A big sample of 4235 demand series consisting of structured and unstructured data (could be referred to as “big data”) related to packaged food products is used for experimentation. Data contain point-of-sales, promotion, weather, regional economy, internet media, and economic activity index related variables. Some of these variables and their combinations is probably used for the first time in a demand forecasting model. The forecasting results are evaluated through multiple error metrics (i.e. mean error, mean absolute error, mean squared error), and it has been observed that proposed method outperformed the benchmarking methods. A demand sensing algorithm is also proposed to forecast demand in real-time.