Estimating profitability decomposition frameworks via machine learning: Implications for earnings forecasting and financial statement analysis
Machine learning with gradient boosted trees can forecast one-year return on equity with up to 7 percent lower errors than simple benchmarks. Linear models with accounting ratios fail to beat a naive forecast. A full four level decomposition cuts errors by 3 percent. The study uses Compustat data from 1964 to 2023. It highlights design choices. Key points include detailed recurring ratios and labeling core and transitory items. Results hold under robustness tests despite historic data limits.
What it examines
This paper uses machine learning, specifically gradient-boosted regression trees, to estimate Nissim and Penman’s nonlinear profitability decomposition framework. By training models on financial ratios from 1964--2023, it forecasts one-year-ahead firm profitability and benchmarks accuracy against random walk and linear models to show improved forecasting performance.
What it concludes
Nonlinear estimation of detailed profitability drivers yields more accurate forecasts than simple or linear models. Focusing on core items, finer decomposition levels, and up to three years of history boosts precision. Applications include improved analyst earnings forecasts, investor trading strategies, and automated financial statement analysis. Future work could test human implementation.
Evidence objects
The study uses gradient-boosted regression trees to model Nissim & Penmans profitability decomposition, cutting oneyearahead ROE forecast errors by up to 7%, with extra 3% gain from full fourlevel decomposition.
key_findings bullet 1 · key_findings · validation V0
Traditional linear modelseven with detailed accounting ratiosfail to outperform a nave benchmark, while combining structured financial framework, high-granularity ratios, core recurring and three-year lags with ML yields significant accuracy improvements.
key_findings bullet 2 · key_findings · validation V0
Using Compustat data from 1964--2023 with a rolling ten-year training window, out-of-sample tests, and robustness checksneural nets, random forests, industry and macro predictorsnote limits: historic bias, overfitting, real-time validation need.
key_findings bullet 3 · key_findings · validation V0
This paper integrates nonlinear machine learning into profitability decomposition frameworks, enhancing traditional $ROE$ forecasting through AI-driven adaptation. It provides a fresh perspective on merging corporate fundamentals and advanced ML for earnings prediction. Though incremental rather than revolutionary, its pragmatic fusion of classic models and $ML$ delivers valuable forecasting accuracy gains.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
We find that nonlinear estimation of profitability decomposition frameworks yields more accurate out-of-sample profitability forecasts than forecasts from both a random walk and linear estimation. The improvements derive from nonlinear estimation and synergies between nonlinear estimation and profitability decomposition frameworks. We analyze three essential financial statement analysis design choices to provide insights for the practice of fundamental analysis and find robust evidence that higher levels of profitability decomposition, focusing on core items, and using up to three years of historical information improve forecast accuracy. We find that our forecasts predict returns and profitability changes before and after controlling for analyst forecasts and common asset pricing factors.
Source row: 733 · abstract type: unknown