Does Multi-Slice Training Improve Bank Customer Churn Prediction? Evidence from Corporate and Retail Transaction Histories

Banks and financial technology providers often train churn models on several historical snapshots of their customer base, a practice known as multi-slice training. Evidence for its benefits rests on few datasets, and widely used random row-wise validation can place copies of one customer record in training and test data. We test the practice under an analysis plan written before any model was trained on the bank data, using public transaction histories of 98,721 corporate clients of Sber (2021–2022) and 50,000 retail clients of T-Bank (2019–2020), with churn defined as three months without transactions. Models trained on the latest slice are compared with models trained on all slices, using out-of-time and customer-disjoint test sets, DeLong tests and the expected maximum profit criterion. Multi-slice training changed the area under the receiver operating characteristic curve by +0.0005 for corporate and +0.0011 for retail clients, within the pre-specified equivalence margin of ±0.005, and raised expected campaign profit by 1.2% for corporate clients only. At equal sample size, older slices performed worse than the latest slice. Random row-wise validation inflated performance markedly only when slices duplicated customer snapshots. In Uzbekistan, Central Bank and Global Findex data show registrations consistent with multi-banking and falling concentration.