LLMs Complement Tabular Models for Industrial Car Retrofit Prediction

This research investigates the application of Large Language Models (LLMs) to tabular data, specifically in the context of industrial car retrofit prediction, where semantic content is limited. The study compares traditional tabular machine learning baselines with three LLM-based strategies: embedding features, direct prompted classification, and an ML+LLM stacking approach. The findings indicate that classical tree ensembles continue to be the strongest standalone models for tasks like binary occurrence prediction, multi-class classification, and regression on this type of structured operational data. However, the LLM experiments revealed nuanced results. While direct prompting of LLMs performed poorly when semantic signal was removed (e.g., via hashing), LLM-generated embeddings proved useful, and a hybrid stacking approach combining ML with LLMs yielded the best manually built multi-class model. This suggests that LLMs are more effective as complementary components rather than direct replacements for robust tabular baselines in privacy-constrained industrial settings.