Projecting Financial Statements with Artificial Intelligence
Abstract
We introduce a novel artificial intelligence framework for projecting the entire set of financial statements. Our approach integrates multi-target learning and chained learning to predict interdependent financial statement items, capturing the intricate relationships across income statement and balance sheet components. Leveraging gradient boosting machines (GBMs) as the base learner, the framework employs a three-step process to optimise chaining sequences and expand feature sets, in order to effectively model inter-item correlations. Empirical validation using out-of-sample predictions for a large sample of U.S. public firms demonstrates the model’s ability to produce accurate and internally consistent financial statement projections. Furthermore, we establish the utility of these projections for applications such as detecting financial irregularities and forecasting earnings. Our methodology can be adapted to other tasks involving the prediction of complex and interdependent outputs.
Co-authors
- Paul Geertsema, Vlerick Business School & KU Leuven
- Helen Lu, Vlerick Business School & KU Leuven