Open Access

Julian Stang, Jannik Stoever, Daniel Schnitzler, Jan Joppien, Marc Wegmann, Christina Reuter

• Small- and medium-sized enterprises (SMEs) often face significant data scarcity in procurement planning, which limits the effective use of advanced analytics for forecasting and decision support. This study addresses the data scarcity by operationalizing a zero-inflated Hurdle model through a conditional variational autoencoder to generate realistic synthetic procurement data. The model explicitly represents both the structural mass of on-time deliveries and the asymmetric distribution of early and late deviations using a mixture of Gaussians. Following the cross-industry standard process for data mining, the conditional variational autoencoder was trained and validated on industrial procurement records. The evaluation results demonstrate strong fidelity: the generated distribution of delivery deviations achieved a Wasserstein distance of 1.92 days relative to the real data, with higher accuracy for early deliveries (1.84 days) than for late deliveries (1.93 days). The zero-inflated component reached a precision of 74.41 % and an F0.5-score of 62.86 % when identifying on-time deliveries. These findings confirm that the conditional variational autoencoder effectively captures both structural zeros and multimodal deviations. By enriching limited datasets while preserving domain-specific behavior, the approach enables SMEs to train forecasting models, perform supplier risk simulations, and deploy procurement-focused decision-support tools even in data-scarce environments.