Effectiveness of long-short term memory network in financial fraud detection

Abstract

To limit the impact of the financial frauds in capital market, it is essential to create a rigorous and effective financial fraud identification model. Our paper discusses the effectiveness of long-term and short-term memory (LSTM) network in detecting financial fraud. In total, 660 Shanghai and Shenzhen listed companies from 1994 to 2018 have been selected as our research samples. Among them, 165 entities were with records of false financial information disclosure. Different types of research approaches are adopted during two stages of the study. The first stage is data preprocessing, during which Artificial neural networks (ANN) are used to screen important variables. In order to accomplish a high-precision financial fraud detection model, both financial and non-financial variables are included. The second stage is the performance evaluation and classifiers comparison. Various classifiers are used to execute and compare the performance of the model: support vector machine, K-Nearest Neighbor (KNN), random forest, multilayer perceptron and LSTM neural network. The results show that the variables screened by ANN and processed by LSTM neural network have high accuracy in identifying financial statement fraud.