Predicting Vendor Performance in Data-Scarce Environments: A Hybrid Deep Learning Approach

In the era of Industry 4.0, the imperative for proactive vendor risk management is predicated upon the precise forecasting of future performance. Nevertheless, the deployment of sophisticated Deep Learning (DL) architectures such as Long Short-Term Memory (LSTM) networks remains markedly constrained in developing economies owing to a profound paucity of historical time-series data. This study interrogates the cold start phenomenon by advancing an original Generative-Predictive Framework that fuses Variational Auto-Encoders (VAEs) with LSTM networks. Applied to a leading entity within the Algerian glass manufacturing sector (Mediterranean Float Glass: MFG), the framework utilises VAEs to reconstruct latent historical performance trajectories (2015–2019) from a limited dataset (2020–2023), thereby affording robust LSTM training for long-horizon forecasting (2026). To substantiate the methodology, a visual trajectory analysis is conducted for three representative vendor archetypes: Stable, Stagnant, and Deteriorating. The findings demonstrate that this hybrid approach efficaciously alleviates data scarcity, producing risk-aware forecasts that empower the buyer (MFG) to transcend reactive evaluation in favour of proactive, strategically informed sourcing.