The predictive modelling for refined, bleached, and deodorised palm oil (RBDPO) quality was performed with an aim to shape a smart and economically sustainable palm oil refining industry. The aim of this study is to improve the RBDPO predictive modelling framework via the novel use of moving average and moving windows form of prediction. Weighted Moving Average (WMA) data smoothing technique was introduced to investigate the effect of the moving average algorithm by comparing the outcomes with the unsmoothed dataset. The Multiple Least-Squares Regression (MLSR) technique was used for model training and prediction of the RBDPO quality. The models were performed using dynamic windows in an expanding and rolling windows form as well as the conventional static window form of prediction. The prediction improvement was consistently observed in all the methods used and was compared in terms of the Mean Squared Error of Prediction (MSEP). It was revealed that the data smoothing result with WMA reduced 33.73 % of the error compared to the unsmoothed data. By comparing the prediction accuracy of the order of the model, the process fitted the first-order model and gave 16 times error less than the second-order model. The dynamic windows prediction form reflected better precision with an average of 37 % error reduced dynamic form that fulfilled the transient nature of the palm oil refining process, in which the data distribution shifts over time.