A natural language processing approach to improve demand forecasting in long supply chains

Information sharing is one of the established approaches to improve demand forecasting and reduce the bullwhip effect, but it is infeasible to do so effectively in a long supply chain. Using the polystyrene industry as a case study, this thesis explores the usage of modern natural language processing (NLP) techniques in a deep learning model, known as NEMO, to forecast the demand of a commodity — without requiring downstream companies to share information. In addition, this thesis compares the effectiveness of such an approach with other non-deep learning approaches, specifically an ARIMA model and a gradient boosting model, XGBoost, to demand forecasting. All three models returned large forecast errors. However, NEMO tracked the volatility of actual data better than the ARIMA model. NEMO also had better success in predicting demand than the XGBoost model, returning approximately 20% better Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) scores. This result suggests that NEMO can be improved with better data, but other issues, such as legality of text mining, need to be considered and addressed before NEMO can be used in day-to-day operations. In its current form, NEMO can be used alongside other forecasting models and provide invaluable information about upcoming demand volatility.