Neural network model for credit risk assessment based on bank customers’ transaction sequences

Problem statement. Conventional scoring models mainly use application data, credit history, and aggregated customer characteristics, while current-account transaction sequences contain additional information about actual financial behavior. For practical use of this signal, it is necessary to consider model quality on a later time period, computational cost, and robustness to different transaction-history lengths.

Goal. To develop and experimentally evaluate a neural-network credit risk model based on bank customers’ transaction sequences and to determine architectural solutions that provide an acceptable balance between predictive quality on an out-of-time sample and inference cost.

Results. The experiment was conducted on approximately three million credit applications with a temporal train/validation/OOT split of about 2.408/0.268/0.295 million observations. For each application, current-account transactions from the year preceding the scoring date were used; sequence length was limited to between 9 and 550 transactions. The main configuration used a bidirectional GRU, entity embeddings for categorical features, and TF-IDF-weighted word2vec averaging for transaction descriptions. With the internal model’s OOT Gini equal to 51.49, the BiGRU configuration with entity embeddings and TF-IDF-weighted word2vec achieved Gini OOT = 52.1, while adding its score to the internal model produced a Gini uplift of +4.8 percentage points. xLSTM increased Gini OOT to 52.3 and uplift to +4.9 percentage points, but required approximately 50% more inference time. Measured BiGRU inference time was 2.9–3.3 ms per application.

Practical significance. The transaction model can be used as an additional scoring signal in an ensemble with an existing tabular bank model. The results show that sequential representation of transactions can improve the discriminatory power of the ensemble on an OOT sample while retaining inference latency suitable for online scoring. In the conducted comparison, BiGRU provided the most practical balance: the small quality advantage of xLSTM did not compensate for its higher computational cost.