Abstract
Cash managers make critical daily decisions regarding monetary inflows and outflows. Traditional financial models often rely on the assumptions that daily net cash flow follows a normal distribution, lacks internal correlation, and remains stationary over time. This paper presents a comprehensive empirical study using a real-world dataset from 54 small and medium enterprises (SMEs). By introducing a novel cross-validated test for time-series non-linearity, the research demonstrates that standard statistical assumptions rarely hold true in practice. Furthermore, the study finds that typical data transformations have minimal impact on achieving linearity or normality. The results advocate for a data-driven forecasting strategy, utilizing non-linear models like Random Forests, to provide more accurate expert systems for cash management.
Introduction
For decades, the theoretical foundation of cash management—the process of determining optimal cash balances to minimize transaction costs and maximize investment opportunities—has been dominated by models such as those proposed by Baumol and Miller-Orr. These elegant mathematical frameworks typically assume that daily net cash flow is "well-behaved," following a Gaussian (normal) distribution with stationary properties.
However, these assumptions are seldom tested against the messy reality of corporate finance, particularly for Small and Medium Enterprises (SMEs). Given that SMEs constitute the backbone of the European economy, reliance on flawed statistical assumptions can lead to sub-optimal liquidity decisions. This paper bridges the gap between theoretical elegance and empirical reality by analyzing a massive, longitudinal dataset of Spanish SMEs to determine the true nature of cash flow time series.
Methodology and Dataset
The researchers utilized a dataset comprising 54 Spanish SMEs from the manufacturing and service sectors, covering approximately eight years (2009–2016) and totaling 58,005 observations. This dataset is notably public, providing a rare transparent look into corporate financial behavior.
To evaluate the data, the authors employed a multi-layered statistical approach:
- Standard Statistical Testing: Normality was assessed via Shapiro-Wilk and Lilliefors tests, correlation via Ljung-Box tests, and stationarity through rolling mean and variance analysis.
- Data Transformations: Common techniques, including Box-Cox transformations and outlier trimming, were applied to see if the data could be coerced into a linear or normal state.
- Cross-Validated Non-Linearity Test: The authors proposed a novel test to detect non-linearity by comparing the predictive performance of a Linear Regression model against a Random Forest (non-linear) model using time-series cross-validation.
- Trivial Series: If neither model outperformed a simple historical mean baseline, the series was classified as trivial (dominated by noise).
- Non-linear Series: If the Random Forest significantly outperformed the linear model (validated via a Wilcoxon rank-sum test), the series was classified as non-linear.
The Myth of the "Normal" Distribution
The most striking finding of the study is the total collapse of the normality assumption. In 100% of the 54 companies analyzed, the daily net cash flow failed normality tests. The data consistently exhibited high kurtosis ("peaky" distributions) and significant skewness.
Crucially, the study found that standard data transformations are not a panacea. Even after applying Box-Cox transformations or removing extreme outliers, the data remained stubbornly non-normal and non-linear. This suggests that the underlying structure of cash flow is inherently complex and cannot be "fixed" by simple mathematical adjustments to fit classical models.
Non-linearity and the Impact of Noise
The study revealed that approximately half of the datasets were "trivial," meaning the cash flow was so volatile or random that neither linear nor non-linear models could beat a simple mean. For the remaining non-trivial series, non-linear models frequently outperformed classical linear approaches.
An interesting paradox emerged regarding outliers: removing extreme values actually increased the number of datasets identified as non-linear. This indicates that extreme "noise" often masks the underlying non-linear patterns. When the noise is reduced, the superior predictive power of non-linear tools like Random Forests becomes more apparent, suggesting that these models are better equipped to capture the complex dynamics of corporate liquidity.
Implications for Forecasting
The research confirms that while seasonality (day-of-the-week and day-of-the-month effects) exists, it explains only a small fraction of total variance. Therefore, relying solely on seasonal adjustments is insufficient for accurate forecasting.
The primary takeaway for cash managers is the necessity of a data-driven approach. Because cash flows are frequently non-linear and non-stationary, classical models from the 1950s and 60s are often ill-suited for modern SME financial management. Instead, the implementation of expert systems that utilize machine learning and non-linear forecasting can provide a significant competitive advantage in liquidity management.
Conclusion
This empirical analysis demonstrates that the statistical properties usually assumed in cash management theory—normality, stationarity, and independence—are virtually non-existent in real-world SME data. By moving away from these flawed assumptions and adopting more robust, non-linear forecasting strategies, companies can better navigate the complexities of daily cash flow. Future work should continue to refine these non-linearity tests, integrating outlier treatment directly into the modeling process to further separate financial signal from random noise.
Summary updated