📊 statistics

Working Capital Financing and Firm Performance:A Machine Mislearning Approach - A comment on Mahmood, Ahmed, Hussain &amp; Ben-Zaied (2025)

This comment critiques Mahmood et al. (2025) for methodological errors in their comparison of working capital financing and firm performance, specifically highlighting the misclassification of principal component analysis, the incorrect reporting of linear regression statistics for tree-based models, and the presence of mutually inconsistent estimates that undermine the study's conclusion that machine learning models outperform fixed-effects models.

Original authors: Gwenael Piaser

Published 2026-10-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gwenael Piaser

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of business, companies constantly juggle two competing needs: keeping enough cash on hand to pay their daily bills, and investing that money to grow and earn profits. This balancing act is known as working capital management. For decades, economists have tried to find the perfect formula for this relationship, often assuming that the more efficiently a company manages its short-term finances, the more profitable it becomes. However, the real world is rarely a straight line. Sometimes, having too much cash is just as bad as having too little, creating a curve where an optimal middle ground exists. To understand these complex patterns, researchers have increasingly turned to computer programs called machine learning. These tools are designed to find hidden shapes and connections in massive amounts of data that traditional statistical methods might miss, offering a new way to see how a company's financial decisions actually affect its success.

A recent study by a researcher named Mahmood, Ahmed, Hussain, and Ben-Zaied attempted to apply this modern approach to Chinese manufacturing firms. They gathered data on over 15,000 company-year observations spanning nearly two decades, hoping to prove that their new machine learning models could explain the relationship between working capital and profit better than the standard methods used for years. They claimed their computer models were superior because they fit the data more tightly and produced higher scores of accuracy. However, a sharp-eyed critique by a researcher named Gwenael Piaser has since examined their work and found that the study, while asking a valid question, is built on a foundation of fundamental errors. The critique reveals that the study did not actually use the machine learning tools it claimed to use, and that the numbers presented in the final report are internally contradictory.

The core of the problem lies in how the researchers described their methods. They stated they were using "regression trees" and "random forests," which are specific types of computer algorithms that work by splitting data into smaller and smaller groups to make predictions. These tools do not produce the kind of results the author published. A true tree-based model does not generate a list of signed numbers with standard errors and stars indicating statistical significance, nor does it produce an F-statistic, which is a measure used only for traditional linear equations. Instead, these computer models produce rankings of which variables are most important and estimates of how well the model predicts new data. Piaser points out that the tables in the original study look exactly like the output of a standard, old-fashioned linear regression, not a machine learning algorithm. The presence of manually added squared terms and interaction variables in the study's equations further confirms that the researchers were actually running a traditional linear model, despite labeling it as a "decision tree."

The confusion deepens when looking at the specific numbers reported. The study presents a table titled "Random forest regression algorithm results," yet the numbers inside it are a confusing mix of two different models. For the main variables of interest, the table lists values that match the traditional fixed-effects model perfectly, while the control variables and the overall accuracy scores match the random forest model. It is mathematically impossible for a single model to produce a set of coefficients that belong to one method and an accuracy score that belongs to another. Furthermore, the study reports the exact same coefficient for a key variable three different times, with three completely different values appearing in the text and the tables. This suggests that the data was not processed consistently, or that the final numbers were assembled from different sources without proper verification.

Another significant issue involves the use of a technique called principal component analysis. The author listed this as one of their competing estimation methods, placing it on the same level as the machine learning models. However, principal component analysis is not a method for estimating relationships between variables; it is a tool used to simplify data by combining many different measurements into a single, cleaner summary. It is a step taken before the actual analysis begins, not a way to test a hypothesis. By treating it as a rival to the other models, the study created a category error, comparing a data-cleaning tool against a method for finding answers. Additionally, the study applied a data-washing process called winsorizing twice, first to the raw data and again after combining the variables. This double compression of the data likely artificially smoothed out the extremes, making the models appear more accurate than they truly are by hiding the very difficult-to-predict cases that a robust model should be able to handle.

Ultimately, the critique concludes that the central claim of the original study—that machine learning models explain the variance in firm performance better than traditional methods—is unsupported. The comparison relied on measuring how well the models fit the data they were already trained on, a practice that favors complex models simply because they can memorize the training data rather than learn from it. Without testing the models on new, unseen data, the higher accuracy scores reported are meaningless. The study also failed to provide the raw data or the computer code needed for other scientists to check the work, leaving the inconsistencies unresolvable by anyone other than the original authors. While the question of how working capital affects profitability remains a vital one for the global economy, this particular attempt to answer it using machine learning appears to have mislearned the very tools it sought to employ. The numbers presented do not reflect the output of the advanced algorithms they claim to represent, and the conclusions drawn from them cannot be trusted.

Technical Summary: A Comment on Mahmood et al. (2025)

Problem Statement
Mahmood et al. (2025) investigate the relationship between short-term working capital financing (WCF) and firm profitability using a balanced panel of Chinese manufacturing firms (2000–2018). The original study hypothesizes a nonlinear, concave (inverted U-shaped) relationship, moderated by ownership, firm size, and leverage. Methodologically, the author claims to compare a standard fixed-effects (FE) panel model against two "machine learning" estimators: a regression tree and a random forest. They conclude that the tree-based models "explain the variance better" based on higher reported $R^2$ and $F$ -statistics.

This comment, authored by G. Piaser, identifies fundamental methodological flaws, internal inconsistencies, and misrepresentations of statistical outputs in the original article. The core problem addressed is not the economic question of WCF and profitability, but the validity of the econometric and machine learning implementation used to answer it.

Methodology and Critique
The comment employs a verification approach, checking the internal consistency of the original article's text, equations, and tables without requiring external data re-estimation. The critique focuses on three primary methodological failures:

Misclassification of Principal Component Analysis (PCA):
The original study lists PCA as a competing estimator alongside regression trees and random forests. Piaser argues this is a category error. PCA is an unsupervised dimension-reduction technique used for feature construction (creating "master proxies" from multiple literature proxies), not a supervised estimator that models relationships between dependent and independent variables. It produces no coefficients, significance levels, or $R^2$ values regarding a target variable. In the original study, PCA serves only as a preprocessing step, making its inclusion in a model comparison table logically invalid.
Misrepresentation of Machine Learning Outputs:
The comment demonstrates that the tables labeled "Decision Tree" and "Random Forest" in the original study report quantities that tree-based methods do not produce. Specifically, the tables list signed coefficients, parametric standard errors, $t$ -statistics, significance stars, and $F$ -statistics.
- Nature of Tree-Based Methods: Regression trees and random forests partition the predictor space into regions and predict constant values within those regions. They do not yield a global parameter vector ( $\beta$ ), nor do they generate parametric standard errors or $F$ -statistics (which are defined for linear models via ANOVA decomposition).
- Evidence of Linear Regression: The presence of manually constructed quadratic terms ( $WCF^2$ ) and interaction terms ( $WCF \times \text{dummy}$ ) confirms a parametric linear specification. Tree-based methods approximate nonlinearity and interactions through recursive splitting, rendering such manual term construction redundant. The reported tables are identified as standard Ordinary Least Squares (OLS) or Fixed-Effects output mislabeled as machine learning results.
Internal Numerical Inconsistencies:
The comment documents severe numerical discrepancies that render the reported results unreproducible:
- Hybrid Columns: Table 13, titled "Random forest regression algorithm results," is shown to be a composite of two different models. The coefficients for the key variables ( $WCF$ and $WCF^2$ ) match the Fixed-Effects estimates, while the control variables, $R^2$ , and $F$ -statistic match the Random Forest estimates. This combination is mathematically impossible for a single fitted model, as $R^2$ is mechanically determined by the specific coefficient vector and the data.
- Conflicting Coefficients: The full-sample coefficient for $WCF$ in the Random Forest model is reported three times with three incompatible values: 0.598 (text), 0.403 (Table 13), and 0.323 (Table 7).
- Data Inconsistencies: Discrepancies exist in sample sizes (15,238 vs. 15,239 observations) and sub-sample counts (e.g., textile industry observations varying between 167 and 679 across tables).

Key Findings and Results
The comment does not offer new empirical results regarding WCF and profitability. Instead, its "results" are the exposure of the original study's methodological incoherence:

Invalid Model Comparison: The central claim that machine learning models outperform fixed-effects models is based on a comparison of in-sample $R^2$ values. The comment argues this is uninformative because flexible tree-based models (if they had been correctly implemented) would mechanically achieve higher in-sample fit than parsimonious linear models. Valid comparison requires out-of-sample or cross-validated error metrics, which are not reported.
Data Processing Flaws: The original study applies double winsorization (first at 5% on raw proxies, then at 10% on the resulting principal components). This procedure compresses data dispersion twice, artificially inflating $R^2$ and attenuating slope estimates, yet no sensitivity analysis is provided.
Reproducibility Failure: Due to the use of proprietary CSMAR data without providing replication datasets, code, or firm identifiers, and given the internal contradictions in the reported numbers, the results cannot be verified or reconstructed.

Significance and Claims
The paper modestly claims to assess the "internal consistency" of the original article's methods, reported statistics, and numerical values. Its significance lies in:

Methodological Correction: Clarifying that PCA is a feature construction tool, not an estimator, and that tree-based methods do not produce linear regression-style coefficients and $F$ -statistics.
Replication Warning: Highlighting that the original study's conclusions are unsupported by the reported evidence due to the mislabeling of linear regression outputs as machine learning results and the presence of contradictory numerical values.
Research Integrity: Emphasizing that without a replication dataset and code, and given the documented inconsistencies, the study's findings regarding the superiority of machine learning approaches in this context cannot be accepted.

The comment concludes that the original article's central claim—that machine learning models provide a more accurate explanation of the data—is invalid on its own terms due to the misuse of statistical objects and the lack of a valid generalization metric.

More like this