Forecasting Product Demand and PET Bottle Return Flows in Closed Loop Supply Chains: A Hybrid Machine Learning Approach for Emerging Markets

Abstract

This study examined whether monthly demand and PET bottle return forecasts improve planning in a closed-loop beverage network in Nigeria. The analysis covers Coke, Fanta, Sprite, Big Cola, and Eva from 2019 to 2024, combining product observations with variables calibrated from the underlying network study. We compare XGBoost, random forest, support vector regression, and long short-term memory (LSTM) with the moment-matching benchmark. Model selection and testing preserve temporal order. XGBoost records the lowest point forecast errors for the five products, reducing average demand mean absolute percentage error from 14.6 to 9.8 percent. For returns, its prediction interval coverage is 78 to 84 percent against a nominal 80 percent, compared with 50 to 62 percent for moment matching. SHAP analysis assigns the largest demand contributions to recent sales, seasonality, and temperature; recovery rate, collection coverage, and reverse logistics cost are prominent for returns. We then use the forecasts in the network optimisation model, retaining its structure and cost coefficients. Expected cost falls by 12 to 18 percent across the tested configurations. The XGBoost and LSTM ensemble gives the largest saving, ₦25.977 million. These results concern the studied network and its calibrated inputs. They support evaluating forecasts through their planning consequences as well as prediction error, with external validation needed before the simulated savings are applied to other networks.