Forecasting intermittent demand: a comparative approach

In this thesis, a comparative approach between forecasting intermittent demand by using machine learning and by using statistical models is carried out. Models implementations are done with the support of different Python libraries aimed at discovering which model would provide better results. For what concerns the data, an electricity demand dataset is used for building the models, where their generated predictions are compared with the real ones. Moreover, performances are investigated against specifically selected scenarios, where different forecast horizons and different times of the day are considered. It gave us the possibility of analysing how these models would perform over distinct settings, including the ones during an anomalous period. The final results showed the KNeighbors Regressor being the best model, especially in scenarios that consider moments in time of very low demand in a normal week, with an accuracy value of 93%. However, despite being the best result, it is not the most intriguing to consider, as instead discovering how some forecasters perform surprisingly well in scenarios where the anomaly is present is the main interest.