With　increasingly　fierce　competition　in　other　transportation　markets,　the　customer　churn　in　railway　freight　becomes　a　very　serious　problem.　Facing　the　high　similarity　and　indistinguishability　of　railway　freight　data,　the　customer　churn　prediction　(CCP)　becomes　one　of　the　challenging　tasks　in　this　industry.　In　this　paper,　a　deep　forest-based　model　is　developed　which　can　achieve　better　accuracy　of　churn　predicting　in　railway　freight　customer　and　effectively　separate　the　churners　from　the　non-churners.　Inspired　by　the　layer-by-layer　processing　of　deep　neural　network,　the　cascade　structure　is　adopted　to　the　deep　forest,　in　which　the　input　of　each　layer　includes　the　original　features　and　the　feature　information　processed　by　the　previous　layer.　The　deep　forest　in　this　paper　achieves　the　best　performance　in　railway　freight　data　with　the　accuracy　of　0.78,　the　precision　of　0.75,　the　recall　of　0.66,　the　F1-score　of　0.7　and　the　AUC　value　of　0.86,　which　are　higher　than　those　of　decision　tree,　AdaBoost　and　XGBoost.　At　the　same　time,　this　model　has　smaller　standard　error　and　makes　more　stable　performance.　©　2021,　Springer　Nature　Switzerland　AG.