In this study, we integrated a decomposition technique viz. seasonal trend decomposition procedure based on loess (STL) with an efficient recurrent neural network-based forecasting technique, i.e. long short-term memory (LSTM) and developed an ensemble hybrid model called STL-LSTM for a non-stationary, nonlinear and seasonal agricultural price series. First, the STL technique is used to decompose the original price series into the seasonal, trend and remainder components. Then, an LSTM network with a single hidden layer is constructed to forecast these components individually. Finally, the prediction results of all components are aggregated to formulate an ensemble output for the original agricultural price series. The hybrid model captures the temporal patterns of a complex time series effectively through analysis of the simple decomposed components. The study further compared the price forecasting ability of the developed STL-LSTM model with the other potential models using monthly price series of potato for two major markets of India. The empirical results demonstrated the superiority of the developed hybrid model over the other models in terms of root mean square error (RMSE) and mean absolute percentage error (MAPE). Moreover, the accuracy of the forecasts obtained by all the models is also evaluated using the Diebold-Mariano test. All criteria show that the STL-LSTM based model has a clear advantage over the other models.