Machine Learning Methods for Optimal Reconciliation of Hierarchical and Grouped Time Series of Financial Institutions
Name: ALBERSON DA SILVA MIRANDA
Publication date: 29/02/2024
Examining board:
Name |
Role |
|---|---|
| GUILHERME ARMANDO DE ALMEIDA PEREIRA | Advisor |
Summary: In the last decade, hierarchical time series forecasting has experienced substantial growth, characterized by advancements that have significantly improved the accuracy of forecasting models. Recently, machine learning methods have been integrated into the literature on hierarchical time series as a new approach for forecasting reconciliation. This work builds upon these advancements by further exploring the potential of ML methods for optimizing the reconciliation of hierarchical and grouped time series. Moreover, the impact of various training set acquisition strategies, such as in-sample forecasts obtained through rolling origin forecasting, fitted values of reestimated models, and fitted values of base forecast models, as well as alternative crossvalidation strategies, was investigated. To evaluate the proposed methodology, two case studies were carried out. The first study focuses on the Brazilian financial sector, specifically forecasting loan and financing balances for the State Bank of Espírito Santo. The second study uses Australian domestic tourism datasets, which are frequently referenced in hierarchical time series
literature. The proposed methodology was compared with traditional methods for forecasting reconciliation such as bottom-up, top-down and minimum trace. The results show that there is no unique method or strategy that consistently outperforms all others. Nonetheless, the appropriate combination of ML method and strategy can lead to up to a 93% improvement in accuracy
compared to the best-performing analytical reconciliation method.
