Forecast reconciliation is a post forecasting process that ensures coherence when dealing with linear constraints time series, aligning predictions at different levels of aggregation. Exploring the cross-temporal framework, this thesis compares different linear and machine learning-based solutions with two empirical applications: Citi Bike rental demand and Energy Load. While both datasets share the same temporal structure, they differ in cross sectional hierarchy and industry context, allowing for a broader evaluation of reconciliation methods. Key concepts in this study include evaluating different reconciliation strategies by exploring various covariance matrix structures, refining residual handling in linear and machine learning based models, and ensuring consistency in non-negativity constraints across all reconciliation methods. The results indicate that, in contrast to prior findings, linear reconciliation methods consistently outperformed ML-based approaches. Among the linear models, heuristic and optimal linear reconciliation approaches demonstrated the highest forecast accuracy. On the other hand, Random Forest remained the strongest ML-based reconciliation method. Sensitivity analyses revealed that including a wide range of temporal aggregation levels generally improves accuracy, reinforcing the value of comprehensive reconciliation structures. The choice between compact and complete feature matrices had a notable impact on ML reconciliation performance, with its effects varying across hierarchical levels. These findings highlight the importance of methodological choices in reconciliation and their influence on forecast accuracy. Future work could explore reconciliation using neural network-based ML algorithms, automated feature selection techniques, and online learning approaches for dynamic model updates and improved computational efficiency.

Forecast reconciliation is a post forecasting process that ensures coherence when dealing with linear constraints time series, aligning predictions at different levels of aggregation. Exploring the cross-temporal framework, this thesis compares different linear and machine learning-based solutions with two empirical applications: Citi Bike rental demand and Energy Load. While both datasets share the same temporal structure, they differ in cross sectional hierarchy and industry context, allowing for a broader evaluation of reconciliation methods. Key concepts in this study include evaluating different reconciliation strategies by exploring various covariance matrix structures, refining residual handling in linear and machine learning based models, and ensuring consistency in non-negativity constraints across all reconciliation methods. The results indicate that, in contrast to prior findings, linear reconciliation methods consistently outperformed ML-based approaches. Among the linear models, heuristic and optimal linear reconciliation approaches demonstrated the highest forecast accuracy. On the other hand, Random Forest remained the strongest ML-based reconciliation method. Sensitivity analyses revealed that including a wide range of temporal aggregation levels generally improves accuracy, reinforcing the value of comprehensive reconciliation structures. The choice between compact and complete feature matrices had a notable impact on ML reconciliation performance, with its effects varying across hierarchical levels. These findings highlight the importance of methodological choices in reconciliation and their influence on forecast accuracy. Future work could explore reconciliation using neural network-based ML algorithms, automated feature selection techniques, and online learning approaches for dynamic model updates and improved computational efficiency.

Linear and machine learning cross-temporal forecast reconciliation: An empirical investigation

GHAMARI, ROYA
2024/2025

Abstract

Forecast reconciliation is a post forecasting process that ensures coherence when dealing with linear constraints time series, aligning predictions at different levels of aggregation. Exploring the cross-temporal framework, this thesis compares different linear and machine learning-based solutions with two empirical applications: Citi Bike rental demand and Energy Load. While both datasets share the same temporal structure, they differ in cross sectional hierarchy and industry context, allowing for a broader evaluation of reconciliation methods. Key concepts in this study include evaluating different reconciliation strategies by exploring various covariance matrix structures, refining residual handling in linear and machine learning based models, and ensuring consistency in non-negativity constraints across all reconciliation methods. The results indicate that, in contrast to prior findings, linear reconciliation methods consistently outperformed ML-based approaches. Among the linear models, heuristic and optimal linear reconciliation approaches demonstrated the highest forecast accuracy. On the other hand, Random Forest remained the strongest ML-based reconciliation method. Sensitivity analyses revealed that including a wide range of temporal aggregation levels generally improves accuracy, reinforcing the value of comprehensive reconciliation structures. The choice between compact and complete feature matrices had a notable impact on ML reconciliation performance, with its effects varying across hierarchical levels. These findings highlight the importance of methodological choices in reconciliation and their influence on forecast accuracy. Future work could explore reconciliation using neural network-based ML algorithms, automated feature selection techniques, and online learning approaches for dynamic model updates and improved computational efficiency.
2024
Linear and machine learning cross-temporal forecast reconciliation: An empirical investigation
Forecast reconciliation is a post forecasting process that ensures coherence when dealing with linear constraints time series, aligning predictions at different levels of aggregation. Exploring the cross-temporal framework, this thesis compares different linear and machine learning-based solutions with two empirical applications: Citi Bike rental demand and Energy Load. While both datasets share the same temporal structure, they differ in cross sectional hierarchy and industry context, allowing for a broader evaluation of reconciliation methods. Key concepts in this study include evaluating different reconciliation strategies by exploring various covariance matrix structures, refining residual handling in linear and machine learning based models, and ensuring consistency in non-negativity constraints across all reconciliation methods. The results indicate that, in contrast to prior findings, linear reconciliation methods consistently outperformed ML-based approaches. Among the linear models, heuristic and optimal linear reconciliation approaches demonstrated the highest forecast accuracy. On the other hand, Random Forest remained the strongest ML-based reconciliation method. Sensitivity analyses revealed that including a wide range of temporal aggregation levels generally improves accuracy, reinforcing the value of comprehensive reconciliation structures. The choice between compact and complete feature matrices had a notable impact on ML reconciliation performance, with its effects varying across hierarchical levels. These findings highlight the importance of methodological choices in reconciliation and their influence on forecast accuracy. Future work could explore reconciliation using neural network-based ML algorithms, automated feature selection techniques, and online learning approaches for dynamic model updates and improved computational efficiency.
Forecasting
Machine Learning
Reconciliation
Cross-temporal
Grouped Time Series
File in questo prodotto:
File Dimensione Formato  
Ghamari_Roya.pdf

accesso aperto

Dimensione 4.33 MB
Formato Adobe PDF
4.33 MB Adobe PDF Visualizza/Apri

The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12608/84782