Demand Forecasting using SARIMA, Prophet and NNAR

Abstract

Long-term planning is a difficult but at the same time crucial process for companies. In fact, a correct forecast can drastically change the fate of a company, thus leading it to success. Today these forecasts are made using increasingly complex algorithms such as Autoregressive Integrated Moving Average, Vector Autoregression, neural network and many other methods. Drawing on the scientific literature of forecasting based on machine learning, this thesis aims to evaluate the ability of 3 models (Seasonal Autoregressive Integrated Moving Average, Prophet and Neural Network Autoregressive) in predicting the value of the quantity ordered for the month of June 2023. In more detail, the objective is to create 3 predictive models for a company whose business is to offer printing services. The data made available are on the orders received in the time frame from January 1, 2018 to June 30, 2023. The practice followed for the development of the thesis includes an initial exploratory analysis to understand whether or not changes in the data are needed and to highlight any patterns. The next stage is the implementation of models with the appropriate technical specifications to optimise performance. Final step, the evaluation of model results on the basis of metrics shared with the company and in particular in this case on the basis of RMSE, MAPE and Relative error. The models showed good but not excellent performance, which led to the decision to assess the inherent uncertainty in them by calculating the coverage probability. Most of the time the results are not good and show that the models are characterised by a certain degree of uncertainty.