This paper introduces prediction models, which enhance the capabilities of Recurrent Neural Networks RNN (LSTM and GRU) to predict input and output variables of the Decision-Making Units (DMUs). The results obtained by these models are further used with the Network data envelopment analysis (NDEA) model to assess the efficiency of DMUs. To demonstrate the performance of the models, a case study was conducted by selecting eight countries’ data of sugar industry. The findings show GRU achieves lower RMSE values, indicating better prediction accuracy, with its lowest RMSE compared to LSTM. In terms of MAPE, GRU also performs better, with its value ranging from 8.83 to 30.9, while LSTM’s MAPE values are higher, reaching 78.89. The average accuracy rate to predict efficiency for LSTM is 50.8%, whereas GRU achieves a much higher average accuracy of 81.4%. The results showed that GRU outperformed LSTM comparison with results calculated through proposed three-stage NDEA model. This combined three stage NDEA model with LSTM and GRU helps decision makers to identify inefficiency within the stages of the DMUs so that necessary improvement could be made.