Summary

This thesis delves into the intersection of advanced machine learning and financial accounting to address practical challenges in producing accurate financial reports, detecting accounting frauds, and improving earnings forecasts. Conventional methodologies for analysing structured and unstructured financial data demand substantial time and resources. Leveraging the potential of advanced machine learning techniques, this study establishes a bedrock for efficient and effective solutions.
In Chapter 2, a framework of semi-supervised machine learning is introduced for identifying anomalies within extensive transaction-level datasets, thereby enhancing the precision of financial reports. Chapter 3 focuses on the utilization of contextual language learning models to detect accounting frauds in textual content, outperforming benchmarks and identifying more fraudulent firms. In Chapter 4, a stack ensemble approach integrates hard and soft accounting data for improved earnings forecasting, surpassing existing models.

These chapters not only contribute to bridging the gap between the realms of machine learning and financial accounting but also emphasize the significance of human validation and domain expertise in refining model outcomes. By combining cutting-edge algorithms with established practices, this research aims to benefit both researchers and practitioners, extending efficient, accurate, and timely solutions to challenges in the financial accounting domain.