Event Details

CMAF forum with Yves Sagaert

Abstract: In the era of open data, the availability of vast numbers of potentially informative covariates has transformed aggregate-demand modelling. However, the resulting high-dimensional covariate spaces, often characterized by strong multicollinearity and small estimation samples, significantly degrade the performance of traditional variable selection methods. While Lasso regression remains a popular solution, its inherent instability in the presence of correlated predictors undermines the reliability and perceived credibility of business forecasts. This seminar presents a novel framework for addressing these limitations by structuring input variables into semantically coherent groups prior to selection. We discuss semantic representations by leveraging Semantic Bidirectional Encoder Representations from Transformers (SBERT) to capture domain-specific context, and compare this with statistical clustering and grouping on meta-data such as covariates popularity. Using an empirical case study, we evaluate these approaches across a multi-dimensional performance matrix: predictive accuracy, decision-oriented utility, and model interpretability. The findings reveal a critical trade-off: while factor-based approaches excel in aggregate forecasting for tactical planning, they lack the explainability required for user trust and perform poorly on disaggregated data. Conversely, semantic integration via SBERT yields more consistent improvements and significantly enhances model transparency.