Projecting Financial Statements with Chained Machine Learning

Abstract

Financial statements are integrated systems, yet most machine learning models treat line items as independent targets. We propose an integrated framework that forecasts stylized financial statements using chained Gradient Boosting Machines (GBMs). The model predicts line items in sequence, conditioning each item on earlier forecasts, thereby propagating contemporaneous shocks and generating error co-movement that stabilizes articulated aggregates. Using a 60-year sample of US public firms in a rolling out-of-sample design, the chained GBM model dominates random-walk, OLS, and parallel GBM benchmarks, delivering statistically significant reductions in forecast errors and lower forecast-error volatility. We provide two empirical applications: structural deviations between realized statements and model-implied expected statements serve as a leading indicator of accounting irregularities, and the model’s net-income forecasts provide information incremental to analyst GAAP consensus forecasts. Overall, the results show that leveraging accounting structure improves the accuracy and coherence of financial statement projections.

Keywords: artificial intelligence, machine learning, chained learning, financial statement projection, restatement, earnings forecasts