This is a past event
The seminar will be held remotely via MS Teams on January 31st from 3pm - 4:30pm.
Join Dr Minjoo Kim from the University of Liverpool as they discuss their research on Detecting Accounting Fraud in China.
Abstract: The research problem: We apply ensemble learning to raw financial and corporate governance information to predict accounting fraud in China.
Motivation or theoretical reasoning: Detecting accounting fraud has significant implications for all participants in the financial market. Given China's financial markets' vast size and relatively immature nature, identifying accounting fraud in this context becomes even more crucial. While numerous empirical studies have underscored the importance of financial information in detecting accounting fraud, there remains a paucity of research concerning the critical role of corporate governance information. The primary reason behind this gap is the qualitative nature of corporate governance data. Machine learning emerges as an ideal tool to effectively process the substantial volume of quantitative and qualitative information. Specifically, the RUSBoost algorithm, an ensemble learning model incorporating random undersampling, presents a viable solution to address imbalanced data issues associated with accounting fraud detection.
The test hypotheses: Hypothesis 1. Ensemble learning outperforms single-learner-based machine learning in accounting fraud prediction. Hypothesis 2. Undersampling is better suited to imbalanced accounting fraud data than simple random sampling. Hypothesis 3. Augmenting raw financial information with corporate governance information leads to enhanced accounting fraud detection.
Target population: Financial institutions, accounting and auditing organisations, Chinese listed companies, regulators, and scholars in accounting fraud.
Adopted methodology: Decision Tree (DT), K-Nearest Neighbour (KNN), Adaptive Boosting (AdaBoost), and Random Undersampling Boosting (RUSBoost). Analyses: We collect fraud and non-fraud cases, raw financial and corporate governance variables from 2007 to 2019. We train machine learning models using data from 2007 to 2014 as the training set. We optimize the hyperparameters of the models using a sample of 2015 as the validation set. We then employ the trained models to predict accounting fraud for 2017. We recursively repeat this exercise until we predict accounting fraud in 2019. We evaluate the prediction performance of the DT, AdaBoost-DT, and RUSBoost-DTs using the AUC score over the prediction period of 2017-2019. We also employ the KNN algorithm as the base learner for RUSBoost, i.e., RUSBoost-KNN, and compare its performance with RUSBoost-DT.
Findings: First, ensemble learning models outperform single machine learning models in accounting fraud prediction. Second, ensemble learning models utilising undersampling techniques yield more accurate accounting fraud predictions than ones relying on simple random sampling. Third, incorporating raw financial and corporate governance information significantly enhances the prediction performance of ensemble learning models. Last, using DT as the base learner in RUSBoost yields superior and more stable prediction performance than KNN.
- Speaker
- Dr Minjoo Kim
- Hosted by
- University of Aberdeen
- Contact
-
No booking required.