Predictive Modeling of Product Returns for Remanufacturing

As awareness of environmental issues increases, the pressures from the public and policy makers have forced OEMs to consider remanufacturing as the key product design option. In order to make the remanufacturing operations more profitable, forecasting product returns is critical with regards to the uncertainty in quantity and timing. This paper proposes a predictive model selection algorithm to deal with the uncertainty by identifying better predictive models. Unlike other major approaches in literature (distributed lag model or DLM), the predictive model selection algorithm focuses on the predictive power over new or future returns. The proposed algorithm extends the set of candidate models that should be considered: autoregressive integrated moving average or ARIMA (previous returns for future returns), DLM (previous sales for future returns), and mixed model (both previous sales and returns for future returns). The prediction performance measure from holdout samples is used to find a better model among them. The case study of reusable bottles shows that one of the candidate models, ARIMA, can predict better than the DLM depending on the relationships between returns and sales. The univariate model is widely unexplored due to the criticism that the model cannot utilize the previous sales. Another candidate model, mixed model, can provide a chance to find a better predictive model by combining the ARIMA and DLM. The case study also shows that the DLM in the predictive model selection algorithm can provide a good predictive performance when there are relatively strong and static relationships between returns and sales.