Pooled Information and Model Flexibility: Decomposing Machine-Learning Gains in Corporate Budgeting and Forecasting

Abstract

Corporate financial planning and analysis (FP&A) is being reorganised around machine learning, yet the supporting evidence consists largely of vendor case reports using undefined accuracy measures. This study decomposes the machine-learning advantage into components that prior work bundles together—information recency versus model re-estimation, and pooled estimation versus learner flexibility—measured on corporate income-statement data rather than retail or equity series. An open benchmark from SEC XBRL data covering 2599 non-financial US filers and 118,406 firm-quarters, from 2011 to 2025, evaluates five pre-registered hypotheses under rolling-origin validation with clustered inference. First, the advantage of continuous re-forecasting over annual budgeting derives from information recency rather than parameter updating: supplying current quarterly actuals to a model estimated once at fiscal year start recovers 27 to 58 percent of forecast error, whereas quarterly re-estimation produces no detectable improvement (p = 0.82), a result that holds across the COVID-19 and 2022 inflation regimes. Second, the advantage is an interaction rather than a main effect: a pooled linear model modestly outperforms a Theta benchmark, a flexible learner fitted per firm underperforms it, and a foundation model pooled across unrelated domains underperforms further; only pooling across comparable firms combined with a flexible learner realises the full gain. Third, model ranking reverses under asymmetric budget-variance cost.