Anonymous OEM (complex asset manufacturer)

In complex asset manufacturing — sectors such as aerospace and defense where a single finished product may depend on thousands of components across hundreds of first-tier suppliers — supply chain disruptions carry disproportionate consequences. Late deliveries from even one critical supplier can halt assembly lines, trigger penalty clauses, and compress delivery schedules that are already measured in months. This anonymous OEM had access to years of historical operational data on supplier performance but lacked the analytical infrastructure to turn that data into forward-looking risk signals. Without reliable disruption prediction, procurement and operations teams could only react after a late order materialized, leaving them with limited options and significant downstream production risk.

Researchers from the University of Cambridge, working with the anonymous OEM, developed a three-phase machine learning methodology applied directly to the company's historical supplier transaction data. The first phase focused on exploratory feature engineering — systematically evaluating which attributes of past supplier behavior could serve as leading indicators of disruption. A critical innovation was the construction of an agility feature derived from historical delivery variance patterns, capturing each supplier's demonstrated ability to recover from past schedule slippage. The second phase developed a domain-aligned performance metric calibrated to the OEM's operational priorities rather than generic classification benchmarks. The third phase ran a systematic experimental design across multiple ML algorithm families and hyperparameter configurations, ensuring model selection was evidence-driven. No specific commercial vendor was identified; the methodology was developed as an academic-industrial collaboration.

The best-performing model achieved 80% accuracy in predicting late orders across first-tier suppliers, outperforming all baseline approaches tested during the experimental phase. Ablation analysis confirmed that the engineered agility feature was the single largest contributor to model performance — models without it fell materially short of the 80% threshold. Key outcomes:

• 80% late-order prediction accuracy on the final algorithm, versus lower accuracy on all baseline models
• Agility feature identified as the highest-impact predictor, validating the role of domain knowledge in feature construction
• Systematic experimental design across algorithm types and parameters demonstrated more reliable model selection than heuristic approaches

The study also established a replicable methodology that other complex-asset manufacturers could adapt to their own historical datasets.

• Domain expertise is a first-class input: engineering a supplier agility feature from historical delivery variance outperformed raw data approaches — supply chain knowledge should inform feature design from the outset.
• Metric alignment determines algorithm selection: a performance metric calibrated to actual business goals (minimizing operational disruption) selects for different models than generic accuracy scores.
• Systematic experimentation beats intuition: testing across algorithm families and hyperparameter grids produced more defensible and higher-performing results than heuristic model picking.
• Historical operational data has latent predictive value: organizations sitting on years of supplier transaction records likely have sufficient signal for disruption forecasting without new data collection infrastructure.