An Efficient Framework for Predicting Future Retail Sales Using Ensemble DNN-BiLSTM Technique

Forecasting retail sales often requires various number of products from different stores. Existing deep or machine learning techniques fall short of producing accurate classification results because of overfitting and two-class problem that affects the performance of evaluation parameters like precision, recall, accuracy and F-measure. Hence there is a need for an efficient prediction framework that addresses the existing problems. This work proposes an efficient framework for predicting retail sales using an ensemble DNN-BiLSTM framework. We suggest creating a base forecaster pool that includes both individual and pooled forecasting techniques for developing this ensemble approach to forecasting retail sales. Instead of focusing on finding the best individual technique, we suggest finding the optimal combination of forecasts. Classification Accuracy, Precision, Recall, and F-measure performance metrics of the experiment utilizing the proposed ensemble approach DNN + BiLSTM surpass the current DNN, CNN, and LSTM classifiers by 98.3%, 98.1%, 97.8%, and 97.94%, respectively.