By Oliver Binz
Recent research suggests that profitability forecasts often fail to outperform simple benchmarks. New evidence shows that this shortcoming stems not from accounting analysis itself, but from how it is applied. By combining structured accounting frameworks with modern machine learning, profitability forecasting becomes more accurate, informative, and economically meaningful.
For decades, financial statement analysis has faced a troubling conclusion: detailed accounting-based profitability forecasts frequently fail to outperform simple rules of thumb. In many empirical studies, a naïve assumption, such as that next year’s profitability will closely resemble this year’s, matches or exceeds the performance of more sophisticated models built from financial ratios.
If correct, this finding would call into question a central pillar of fundamental analysis. A closer look, however, suggests that the problem lies not in accounting itself, but in the statistical tools traditionally used to implement it.
The Promise (and Frustration) of Profitability Decomposition
Decomposing profitability into its underlying drivers has long been a cornerstone of financial analysis. By separating operating performance from financing effects, and margins from asset efficiency, analysts can better understand what is driving profitability and whether earnings are likely to persist. In simple terms, this approach breaks profitability into its key building blocks, such as operating performance, efficiency, and financing effects, to understand what is really driving results.
A common approach in financial analysis is to break profitability into its underlying drivers to better understand performance. While intuitively appealing, this approach has often failed to improve forecasts in practice, leading many to question how useful detailed financial analysis really is.
The Missing Ingredient: Nonlinearity
The core issue lies in the assumption of linearity.
Profitability dynamics are inherently nonlinear. Financial leverage enhances returns only when operating performance exceeds borrowing costs. Margins and asset turnover interact differently across industries and business models. Small changes in one component can have vastly different implications depending on the level of another.
Linear models struggle to capture this complexity. They impose constant, additive relationships even when economic intuition suggests otherwise. As a result, much of the information embedded in financial statements remains unused.
This is where modern machine learning techniques become relevant.
Machine Learning, with Discipline
Rather than abandoning structure in favor of opaque “black box” prediction, a more productive approach combines established accounting frameworks with machine learning methods designed to capture nonlinear and interactive relationships.
Gradient-boosted regression trees provide such a tool. They allow the data to reveal complex interactions among familiar accounting drivers, while remaining anchored in accounting logic. The result is not an indiscriminate search across thousands of variables, but a disciplined model that learns how profitability components work together in practice.
Using more than sixty years of firm-level data, out-of-sample forecasts of return on common equity were generated and compared with standard benchmarks. The results show that machine learning improves forecast accuracy relative to both random-walk models and linear regressions, especially when paired with detailed profitability decomposition. The largest gains come from reducing large forecasting errors where traditional models perform poorly.
What Actually Improves Forecasts
A structured framework also makes it possible to draw practical conclusions about how analysts should use financial statements.
First, detail matters only when used appropriately. Breaking profitability into finer components improves forecasts only when nonlinear estimation is applied. Under linear models, additional detail can actually reduce accuracy, helping explain why earlier studies reached pessimistic conclusions.
Second, not all earnings components are equally informative. Forecast performance improves when attention is focused on core, recurring items, while transitory or unusual components are downweighted. This aligns with long-standing analytical intuition, but the evidence shows that the benefits are tangible.
Third, history matters, but only to a point. Incorporating one to three years of past financial data improves forecast accuracy by capturing firm-specific dynamics and business cycles. Beyond that horizon, the benefits diminish as firms evolve and business models change.
Once this structured, nonlinear approach is in place, adding industry classifications or macroeconomic indicators contributes little additional forecasting power. Much of that information is already embedded in financial statements themselves.
Why Investors and Analysts Should Care
Improved forecasts are only valuable if they contain information not already fully reflected in market prices or analyst expectations. To assess this, the relationship between forecasted profitability and future stock returns was examined.
The results are economically meaningful. Even after controlling for standard asset-pricing factors and consensus analyst forecasts, profitability predictions remain strongly related to subsequent returns. Firms with greater forecasted improvements in profitability experience significantly higher future stock performance.
The forecasts also predict future changes in profitability beyond what analysts anticipate. This suggests that structured machine learning extracts information from financial statements that markets and analysts do not fully incorporate.
Why Structure Still Matters in an AI World
Much of today’s enthusiasm for AI in finance emphasizes scale, with more data, more predictors, fewer assumptions. While powerful, this approach often sacrifices interpretability, particularly in accounting, where variables are tightly linked by design.
A structured approach offers an alternative path. By combining accounting-based frameworks with machine learning, it is possible to achieve both predictive accuracy and economic insight. Accounting structure grounds the model, while machine learning captures relationships that linear tools cannot.
This balance is essential for decision-makers who must understand, explain, and act on forecasts, not merely compute them.
What This Means for Business Leaders
For executives, the implications are practical rather than technical. Forecasting accuracy depends less on adopting ever more data and more on using the right analytical tools for the complexity of modern business models. Financial statements already contain rich strategic information, but much of it remains underutilized when linear metrics dominate planning and performance reviews. Leaders who combine accounting discipline with advanced analytics are better positioned to anticipate turning points, identify hidden risks, allocate capital more effectively, and challenge overly confident consensus forecasts before markets do.
Rethinking the Role of Fundamental Analysis
The broader implication is clear. The perceived failure of accounting-based profitability forecasting is not a failure of accounting, but a failure of the tools used to analyze it.
When methods capable of capturing the nonlinear reality of business performance are applied, financial statement analysis proves both relevant and powerful. Artificial intelligence does not replace fundamental analysis; it enhances it.
The future of profitability forecasting lies not in choosing between structure and prediction, but in combining the two.
Oliver Binz
