Hierarchical time series forecasting: a comparative study applied to insurance sales data

The field of time series forecasting has witnessed substantial advancements in recent years, driven by the need for more accurate and efficient predictions across various domains. Hierarchical time series forecasting, in particular, has emerged as a critical component in many decision-making processes within enterprises. Indeed, in today’s market, companies are constantly manging to achieve their strategic goals, and forecasting provides important informations which assist business decisions. In this context, this thesis explores and compares several forecasting models, ranging from classical statistical methods to advanced machine learning techniques, to identify the most effective approach for forecasting a complex dataset of hierarchical time series representing the insurance policy sales of Facile company. The findings of this study reveal distinct advantages and limitations associated with each model. Classical statistical SARIMAX models exhibit limited performance due to their stringent assumptions. LightGBM, a gradient boosting framework, demonstrates remarkable predictive power, particularly in handling numerous time series with diverse exogenous variables. However, given its non-autoregressive nature, it struggles on capturing local and global trends. On the other hand, DeepAR, a recurrent neural network model, and N-BEATSx, a deep learning method based on neural basis expansion analysis, strike a balance between capturing temporal dependencies and effectively leveraging exogenous features. Since it outperformed all other tested models, N-BEATSx emerged as the perfect candidate to be integrated into a production pipeline, providing daily forecasts to be consulted by Facile’s insurance business unit team. In addition, more potential applications of the model, including other business units such as mortgages, telecommunications, and energy, are ahead. In essence, this thesis underscores the significance of hierarchical time series forecasting and offers a comparison and evaluation of forecasting techniques which can help facilitate improved decision-making processes.