Text Extraction and Classification for Automated Balance Sheet Data Entry from Form 1120

Data entry is a routine task requiring minimal skill and excess time for professionals in financial fields. Auditors, accountants, and others lose time entering data that could be better utilized performing financial analysis. In an effort to reduce this obstacle, tools for automated classification of financial data were tested. These tools extract financial data, classify, and insert the data into a spreadsheet. The specific scope of this research paper is classifying Form 1120 balance sheet items into current and noncurrent assets and liabilities. This small-scale approach can be expanded to automate the process of extracting all data from balance sheet schedules and sorting it into a spreadsheet. Previous research in classification has rarely been applied to this domain. In order to develop a method with accuracy, multiple classifiers were tested. A logistic regression model, linear support vector machine, and polynomial support vector machine were all found to be the most accurate tools for determining classification. The immediate implications of these tests are that time spent on balance sheet data entry can be greatly reduced through automation with accurate results. Future work would be necessary to build an application that utilizes this classification to its fullest potential. Results show that our proposed classification models had an average maximum prediction rate of up to 97.6% and a mean prediction rate of 88.8%, and with the proposed human intervention for classifying the terms that are most likely to be incorrect (i.e., terms for which the highest-performing models did not agree on the classification), all terms were able to be correctly classified. With a larger training dataset, the predictions would better generalize to many possible categories from a balance sheet schedule, and displays the effective nature of this research for classifying terms from balance sheet schedules.