Backorder prediction in supply chain management - comparing the contribution of machine learning and LSTM classifier in an industrial context

In this study, the comparative analysis between various machine learning techniques and LSTM (deep learning classifier) are employed to predict demand and backorders based on previous sales, national inventory, the minimum recommended local stock, pipeline inventory, and lead time. Various machine learning algorithms are used for evaluating three-month, six-month, and nine-month data for predicting backorder. The performance metrics considered in study are accuracy, precision, F1-score, AUC values, MAE, RMSE, and log-loss. A dataset from Kaggle's repository has been taken to conduct experiments for backorder prediction the findings of various experiments favour AdaBoost classifier, which has AUC value 82.6% for three-month forecasting. Random forest provides 99.3% AUC value for six-month and 85.65% for nine-month prediction. Conversely, the deep learning (LSTM) classifiers provide AUC values of 89.37%, 91.51% and 80.76% for three-month, six-month, and nine-month prediction processes respectively. LSTM gives AUC, log-loss and MSE values much better compare to other model. In the end, it is found that the bagging classifier provides better results as compared to the single classifier, and is even much better than deep learning (LSTM) classifier, there are not many classifiers that can compete with AdaBoost and random forest in terms of accuracy.