Research Work

Abstract: [Work in progress]

Abstract: Accurate probabilistic forecasts are vital for large-scale operational tasks such as inventory management, where decision-making under uncertainty depends on quantifying risks effectively. Although global Gradient Boosted Decision Trees (GBDTs) have gained popularity for their strong point forecast accuracy in such settings, their potential for reliable prediction intervals remains underexplored. Existing prediction interval estimators — such as quantile regression, empirical methods, and conformal predictions — often do not fully leverage GBDTs' advantages. This paper argues that bootstrapping is better suited for global GBDTs as it aligns with their autoregressive nature and internal mechanisms, providing local corrections for patterns overlooked during global optimization. We propose enhancements to conventional bootstrapping to address limitations like neglected temporal dependencies and non-representative error sampling. Specifically, we refine bootstrapping with structured residual sampling, validation-based error selection, magnitude-specific binning, and kernel density estimation for quantile extraction. We evaluate these modifications on three demand forecasting case studies in the food retail sector in Flanders, Belgium, and demonstrate that our modified bootstrapping consistently outperforms established baselines in prediction interval accuracy and downstream inventory performance. Additionally, we provide recommendations for specific modifications under different dataset characteristics and analyze the correlation between point forecast and prediction interval errors. We further show that our methods do not have constraining computational requirements. The significance of our findings lies in the enhancement of GBDTs, a forecasting tool adopted widely by practitioners, with a flexible prediction interval estimator that efficiently improves predictive reliability across diverse scenarios.

Abstract: Information sharing in supply chains, between a retailer and a supplier, has been researched extensively. Much of the relevant literature has focused on the theoretical impact of this collaboration, with several studies arguing for its benefits. These range from counteracting the bullwhip effect to reducing the likelihood of stock-outs. Despite this, most empirical work has focused on improving the accuracy of supplier-level forecasting, often in binary supply chains. In this work we propose a novel information sharing framework that can directly map arbitrary large supply chains, and benefit all echelons. Motivated by advances in hierarchical forecasting, and particularly cross-temporal approaches, we propose organising the demand series of different echelons into hierarchical nodes. This structure allows to represent the propagation of the demand across the supply chain, enhancing the potential for modelling any benefits of information sharing. Building on the hierarchical forecasting literature we propose the combination of base forecasts, which are produced independently for each hierarchical node, with a multi-task deep neural network architecture. In contrast with conventional cross-temporal reconciliation methodologies we produce base forecasts for the temporal levels that matches the frequency of the specific cross-sectional level. The advantage of the proposed approach is that it allows to flexibly map the different objectives across the supply chain, which is a non-trivial problem for conventional hierarchical forecasting. We demonstrate its efficacy using a real case study with three supply chain levels, and provide evidence of the benefit of both the multi-task training and the hierarchical structure.

Abstract: Forecasting student performance can help to identify students at risk and aids in recommending actions to improve their learning outcomes. That often involves elaborate machine learning pipelines. These tend to use large feature sets including behavioural data from learning management systems or demographic information. However, this complexity can lead to inaccurate predictions when concept drift occurs, or when a large number of features are used with a limited sample size. We investigate the performance of different machine learning pipelines on a data set with change in study behavior during the Covid-19 period. We demonstrate that (i) LASSO, a shrinkage estimator that reduces complexity and overfitting, outperforms several machine learning models under these circumstances, (ii) a linear regression relying on only two handcrafted features achieves higher accuracy and substantially less predictive bias than commonly used, more complex models with large feature sets. Due to their simplicity, these models can serve as a benchmark for future studies and a fallback model when substantial concept or covariate drift is encountered.

Abstract: While the interest for artificial intelligence (AI) has exponentially grown in the last years, many companies find it difficult to implement AI or machine learning (ML). One of the most common issues is the lack of historical data. However, in the context of demand forecasting, a better estimate of future demand can result in taking better business decisions. In this paper we investigate how we can use the hidden structures that are present in the limited historical data via reviewing the data in a temporal and cross-sectional setting. this allows for more data points, which enables the use of machine learning techniques. We propose methodology combining Naive models and global modelling, with coherent reconciliation. The results indicate an improvement of 26.5\% in forecast accuracy over seasonal Naive.

Abstract: In time series analysis and forecasting the identification of an appropriate model remains a challenging task. Model misspecification can lead to erroneous forecasts and insights. The use of multiple views of the same time series by constructing temporally aggregate levels has been proposed as a way to overcome the model specification and selection uncertainty, with ample empirical evidence of forecast accuracy gains. Temporal Hierarchies is the most popular approach to achieve this, which itself is based on research in hierarchical forecasting. Although there has been substantial progress in this literature, the vast majority of methods rely on a restricted linear combination of different model outputs across the hierarchy. We investigate the use of deep learning to augment temporal hierarchies, relaxing the classical restrictions. Specifically we look at deep learning for the generation of all the base forecasts, the hierarchical reconciliation, and an end-to-end method that embeds all steps in a single neural network. We inspect the performance of the proposed methods when applied to individual time series, or with global training across complete sets of series. We further investigate the requirements in terms of series set size, illustrating the conditions where deep learning temporal hierarchies outperform conventional temporal hierarchies.