Data Mining to Detect Fraud Patterns in a Taxpayer’s Financial Statement

Abstract

The application of machine learning in the analysis of financial statements is a relatively underexplored area compared to mainstream data mining fields, such as natural language processing (NLP) and image analysis, yet it holds significant potential. This study investigates the use of advanced linear regression techniques to identify patterns in taxpayers’ financial statements, employing a conceptual approach that combines both vertical and horizontal financial statement analysis methods. Using financial statement data reported to the Indonesian Tax Administration and historical taxation audit records,this study determines the presence of identifiable patterns. This study applies linear regression to financial statement account values to measure changes over the years and uses yearly account values to create unique data points representing each entity. A clustering method is then employed to group entities with similar patterns. The findings indicate that the proposed method can effectively analyse how entities report their financial statements over time and cluster them based on the likelihood of committing fraud, as inferred from historical audit records. These patterns are validated by instances of underpayment or overpayment of corporate income taxes identified during tax audits. By examining the clustering results, the study reveals that certain clusters accurately align with labelled patterns, correctly identifying 2 of 3 labels. The comparison between unsupervised clustering and labelled criteria demonstrates a significant fitness probability.