Benchmarking Tabular Foundation Models for Churn Prediction

Customer churn prediction is one of the most consequential ML-driven classification challenges across industries, where a single percentage-point gain in predictive accuracy translates into measurable retained revenue. Historically addressed by classical ML and, more recently, by deep learning and tree ensembles, churn prediction has lacked a systematic evaluation of the latest tabular foundation models (TFMs) built on in-context learning (ICL), a paradigm in which a model issues predictions from a labelled context in a single forward pass, with no retraining on the target dataset. We present the first cross-industry benchmark of sixteen models, spanning classical baselines, deep learning, tree ensembles, and six distinct TFMs (TabPFN, TabICL, TabDPT, Mitra, LimiX, and ConTextTab) across nine publicly available datasets covering seven industry sectors. TFMs collectively and decisively outperform all classical, deep learning, and tree-ensemble alternatives; TabICL v2, which operates without hardware constraints or subsampling, is the state-of-the-art, ranking in the top three on eight of nine datasets and outperforming XGBoost by up to 9.23\,pp PR-AUC with no dataset-specific retraining; a result that positions TFMs as fully industry-ready tools that businesses across all sectors can embrace today. SHAP-validated, ablation-confirmed actionable insights further identify the dominant churn driver in every industry studied, delivering a fully deployable and explainable end-to-end pipeline.