Intermittent demand forecasting using Machine Learning

Abstract: Different techniques are used for demand forecasting within the In-dustry such as statistical methods like Croston, ARIMA and exponen- tial smoothing methods, Also, During these days Machine learningtechnologies such as SVM, NN and gradient boosting are also usedfor demand forecasting. Both statistical and Machine learning mod- els are widely used in the Industry for demand forecasting basedon performance, requirements, technical availability and data. In thisresearch, we will be dealing with intermittent demand forecasting,which means the data has a large number of zero values within thesales and this is done in connection with Volvo trucks. We will be ex-perimenting with both statistical and Machine learning approaches toevaluate how both statistical and machine learning models respondto the data that we have. Also, we will be introducing a novel hybridapproach where we combine both statistical and Machine learning models into an ensemble architecture(hybrid modelling) which im-proved the performance or prediction accuracy for Intermittent data.In this, the ensembles will be formed by creating metadata combining inputs and predictions from individual models and using the meta- data to train other machine learning models in-order to get predic-tions. We will be using the data from the Volvo trucks supply chaindivision to conduct our experiments and evaluation of the results.The contribution of this paper is twofold. During our experiments,we found out how each model from statistical as well as machinelearning model fits with the data that we have. We also introduced a novel hybrid approach with ensembles combining both statistical aswell as machine learning in a meta-model architecture. Along withimplementing the hybrid model, we compare the best out of the sta- tistical and machine learning models with our hybrid ensemble ap-proach which proves to reduce the predictions error(lesser error) to 6percentage lesser from the best performing individual model.Secondly,we focused on bringing confidence prediction into intermittent de-mand forecasting. The confidence prediction brings a certainty factorin to the predictions, which is the second major contribution of ourresearch. In our approach of confidence prediction, we came up with a confidence of 95 per cent which shows that we are 95 per cent cer-tain that the value will fall in the range of our prediction interval.And we also made sure that the prediction bandwidth with such alarge confidence interval is not too wide so that it doesn’t affect tothe safety stock maintenance in the warehouse thus can reduce thestock accumulation in case of adverse conditions that affect the sales.