Exponential smoothing with neural networks: a TBATS-NN hybrid model for hierarchical time series forecasting

Time series analysis is a key aspect in the manufacturing company field, as it is used to extract information from the previous sales of various products in order to forecast plausible future behaviours of the market. This allows the managers to elaborate proper marketing campaigns and organize the supply chain according to the predicted quantities they will sell in order to maximize profit. In particular, hierarchical and grouped time series are a huge part of the sales-related data, as the data can be aggregated in various ways based on the type of product and where they were sold. Analysing these kinds of time series has the added difficulty given by the linear constraints the different levels of the hierarchy have with each other: the sum of the forecasted lower levels of the hierarchy needs to be equal to the predicted upper level they stem from. In this thesis, a practical case of a manufacturing company sales data is used to outline how to handle hierarchical or grouped time series analysis using Python code, starting from the exploratory analysis and data cleaning, and then moving on to how to implement the most used forecasting models for time series. In particular, a new approach to the exponential smoothing and neural network hybrid, not documented in the current literature, is proposed: the TBATS-ANN hybrid. It is then shown how to reconcile the predictions in order to have results that satisfy the already mentioned linear constraint to which all the series are subject. In the end, different kind of performances are confronted in order to get the best model for the problem at hand.