Prediction of Bankruptcy using LSTM Architecture with Time Series Data for Accounting

Advanced support of credit risk management, auditing, and regulatory oversight can be performed through prior and more dependable bankruptcy forecasting. Classic models (e.g. Altman Z-score, logistic regression) describe linear associations at single-point cross-sectional snapshots, but reallife distress dynamics can be observed as sequences of variability infirm financial trends. In this paper, we introduce a long shortterm memory (LSTM)-based end-to-end architecture forecasting default/bankruptcy using multivariate accounting time series. Our contributions are: (i) principled pipeline to convert raw financial statements into labelled aligned daily financial flows with leakage-free labels; (ii) LSTM classifier with temporal attention and calibration to yield well-behaved probabilities; (iii) rigorous evaluation with time support in train and split times, cost-sensitive learning, and back testing; (iv) interpretability through gradient-based and perturbation methods to help understand which time points and features the models use to make decisions. In a representational panel of firms, it performs better than robust benchmarks (logistic/GBM/TCN) in terms of recall and precision, and earlywarning lead time and is consistent among macro regimes. However, when looking at In terms of AUC scores, the other models developed using deep learning performed poorly compared to the median performance among all machine learning models. Nevertheless, the LSTM model is useful for forecasting bankruptcy in imbalanced data sets clearly performs better than other machine learning models.

Ҳали таржима қилинмаган