This paper explores the use of feedforward multilayer perceptron (MLP) neural networks for predicting the allocation of process-based cost components across 2000 ATM units. Cost structures from 2021 were used to train models on volume- and value-based drivers, and predictions were generated for dataset from 2022. Both sets of data were gathered during ATM Cost Project© which relied on cost analysis of the ATM networks in Poland. Three MLP architectures were tested under two training strategies: with early stopping (validation) and without. Evaluation was conducted at both aggregated and detailed levels. Results demonstrate that MLPs can effectively learn complex cost distribution patterns and generalize to new data with high accuracy. Stratified analysis of MAPE and SMAPE across cost brackets reveals strong dependency between error behavior and true cost magnitudes. Validation-based models typically performed better in lower value ranges, while non-validation models showed improved performance in higher brackets. The study highlights the importance of stratified error reporting and demonstrates that training strategies and model depth must be adapted to data structure and prediction goals.