Traditional Machine Learning versus Large Language Models for Predicting Ship Turnaround Times at Brazilian Ports

Abstract

Predicting ship turnaround time at ports is an operationally valuable regression problem on heterogeneous, text-rich tabular data. We ask whether Large Language Models (LLMs) can match purpose-built Machine Learning (ML) on this task, using a leakage-controlled pipeline over seven years of Brazilian (ANTAQ) data. On a fixed 200-berthing test set, with 20 seeds and seed-paired statistics (Wilcoxon, Cohen’s d), we compare five families that read the same cases: classical ML on engineered tabular features plus multilingual sentence embeddings; frontier LLMs (Qwen3-32B, DeepSeek-V3) prompted zero-, 5and 20-shot; a small generative LLM (Qwen2.5-1.5B) fine-tuned on 100–6000 serialized records; an encoder with a regression head fine-tuned on 100–12000 records; and a tabular foundation model (TabFM) reading the raw table in-context. Tuned XGBoost attains R2 ≈ 0.73. Every prompted configuration stays at R2 ≤ 0 and in-context examples do not help; generative fine-tuning barely clears zero, reaching only R2 ≈ 0.10 at 6000 examples. The encoder is the only text regime to rise substantially above R2 = 0, reaching ≈ 0.43 at 12000 examples yet still trailing all ML. In contrast, TabFM, with no task training, matches or beats a data-matched XGBoost at every budget (100–6000; R2 up to ≈ 0.60), leading the low-data regime, though full-data XGBoost still wins and TabFM is far costlier at inference. The predictive signal is tabular, not textual: text-only LLMs miss it, whereas a tabular foundation model recovers it in-context and engineered boosting wins at scale.