Study on Hierarchical Time Series Forecasting using Machine Learning Models

links:-

Preprint of Research Paper

< Notebook link >

< Dataset link >

Abstract:

Hierarchical forecasting methods are extensively utilized for precise decision-making by providing coherent forecasts across various levels. In the context of supply chains (SC), managers and sales executives frequently require forecasts at different SC levels to facilitate optimal decision-making, taking into account multiple features and dependencies. top-down (TD), Bottom-Up (BU), and Middle-Out (MO) are some of the prevalent hierarchical forecasts (HF) reconciliation approaches. Traditionally, statistical models are employed in HF. However, these static approaches often overlook the dynamic nature of the series during disaggregation, potentially resulting in suboptimal performance, particularly when the investigated time series undergoes substantial changes, such as during promotional sales. This paper addresses this issue by leveraging the dynamic capabilities of machine learning. We propose deploying machine learning (ML) at specific hierarchical levels, followed by reconciliation. Specifically, we implement and evaluate the performance of 17 different machine learning models at each hierarchical level, reconciling them post-forecasting. A comprehensive analysis is conducted using the M5 competition dataset from Kaggle, which exhibits various volatilities. The performance of ML models is then compared to that of traditional statistical models, demonstrating the superiority of ML approaches in managing dynamic time series data.