The Impact of Cash Holding Decisions on Firm Performance in the IT Industry

Abstract

This study examines the relationship between corporate cash holdings and firm performance within the IT industry, which is characterized by intense competition and rapid technological advancements. We propose an integrated framework that combines principal component analysis (PCA), machine learning (ML) algorithms, and Shapley additive explanation (SHAP) values to estimate and interpret model outcomes. Based on 21,051 corporate financial statement data items from 2004 and 2023, the empirical evidence supports an inverted U-shaped relationship between cash holdings and profitability, suggesting that holding either too little or too much cash is suboptimal. Among the tested models, the random forest demonstrates the highest explanatory power (R²) and the lowest prediction errors (RMSE), outperforming the traditional ordinary least squares (OLS) regression by explaining 47% more variance. Our findings provide practical implications for researchers and stakeholders interested in enhancing corporate risk management and performance.

1. Introduction

The global business environment has undergone significant transformations due to the COVID-19 pandemic and the rapid adoption of artificial intelligence (AI) technologies. Advancements in technology have significantly lowered geographic obstacles, facilitating business competition through digital platforms. Firms must strategically navigate uncertainty and enhance their performance in response to these developments. Among the various factors influencing corporate success, cash holding decisions have gained significant attention, as they play a critical role in ensuring operational flexibility, supporting investment opportunities, and mitigating financial risks. However, the relationship between cash holdings and firm performance remains subject to debate, with studies suggesting positive (C. Kim & Bettis, 2014; J. H. Kim et al., 2023), negative (Harford et al., 2008; Louis et al., 2012), or inverted U-shaped associations (Tan & Peng, 2003). These mixed findings suggest that contextual factors—particularly industry characteristics—can influence how cash balances affect firm performance (Deb et al., 2017; Chambers & Cifter, 2022).

Prior research on trade-off theory has shown that firms can benefit from maintaining an optimal capital structure. While increased borrowing allows firms to benefit from the tax deductibility of interest expenses, it also exposes them to a higher risk of bankruptcy. Taken together, the trade-off theory suggests that an optimal debt-to-equity ratio exists that maximizes firm value (Howe & Jain, 2010; Abel, 2018; Pinillos et al., 2025). Despite this evidence, it remains underexplored whether an optimal level of cash holdings exists for information technology (IT) firms, and this study aims to fill this gap.

This paper focuses on the information technology (IT) industry, which occupies a pivotal role in driving technological innovation and economic growth. IT firms face unique challenges, including substantial investments in research and development (R&D) as well as high market volatility caused by rapid innovation cycles and competitive pressures. As such, cash holding strategies have become particularly vital for IT firms, serving as a buffer against uncertainty while enabling investment in growth opportunities (Das, 2015). Pecking order theory argues that managers prefer to use internal funds, such as cash reserves, for investment because they are less costly and less affected by information asymmetry than external sources. (Myers & Majluf, 1984; Y.-R. Chen & Chuang, 2009; Faysal, 2024; Eger & Hermis, 2025). Despite the importance of liquidity management, very few studies specifically examine the IT sector to determine how cash management decisions translate into performance outcomes in the technology-driven environment (Yin & Yin, 2025). To fill this gap, our study undertakes an in-depth analysis of cash holding levels and their impact on firm performance within the IT sector, accounting for key firm-level control variables such as investment, growth, size, leverage, and R&D intensity.

By leveraging firm-level data and employing robust econometric techniques, we aim to identify effective cash management practices for IT companies. This study finds empirical evidence of an inverted U-shaped relationship between cash holdings and firm performance, using 21,051 firm-year observations from U.S.-listed IT firms from 2004 to 2023. The results suggest that an optimal level of cash holdings exists that can maximize IT firms’ profitability.

Furthermore, this study incorporates machine learning techniques to enhance the effectiveness of the estimation approach. The application of machine learning in finance and accounting has become increasingly popular, with growing scholarly and practical interest in leveraging data-driven algorithms to enhance predictive accuracy and decision-making (Gow et al., 2023; Zakaria et al., 2023; Najem et al., 2025; Nguyen Thanh & Phan Huy, 2025). Principal component analysis (PCA) is employed to construct master proxy composite variables for key constructs (profitability, investment, growth, etc.), which mitigates proxy selection bias and enhances the robustness and generalizability of our results. Ordinary least squares (OLS) regression is a commonly used method for analyzing panel data, although interest in applying ML techniques for estimation tasks has been growing (Ozlem & Tan, 2022). Accordingly, we employ OLS regression as a benchmark method and compare its performance against machine learning models, including Decision Tree, Random Forest, Multilayer Perceptron, and Support Vector Machine. In addition to model comparison, we introduce SHAP (SHapley Additive exPlanations) values to interpret the machine learning models’ predictions. SHAP values explain how each feature contributes to the model prediction, an alternative to the traditional coefficient-based interpretation in regression models (Meng et al., 2021; Kori & Gadagin, 2024). SHAP values allow for comparison with the variable importance derived from OLS coefficients. This integrated framework can provide higher predictive power and interpretability.

This study offers several theoretical and practical contributions. It provides evidence that the relationship between cash holdings and performance in the IT industry follows an inverted U-shape. In other words, holding cash contributes positively to performance up to a certain point but becomes counterproductive beyond the optimal level. This pattern is consistent across both OLS and ML-based analyses, supporting the existence of an optimal cash balance that maximizes IT firm profitability. The findings underscore the significance of industry context in corporate liquidity research, bridging the literature on cash management, risk mitigation, and firm performance.

Our results have implications for stakeholders. Stakeholder theory emphasizes the responsibility of firms to account for diverse stakeholder expectations within their business processes and strategies (Mitchell et al., 2015; Amir et al., 2024). The IT industry is a high-stakes environment, and IT firm leaders tend to be highly ambitious and often exhibit risk-seeking behavior such as new market entry (Situmeang et al., 2016), which might result in insufficient attention to effective risk management. The observed inverted U-shaped relationship suggests that cash should be treated as a strategic resource and must be managed actively to balance growth opportunities and risk exposure. Exploring corporate cash holding decisions is critical because they can contribute to improved risk management for IT firms. This study contributes to the research on corporate risk management and firm performance by utilizing machine learning techniques within the context of the IT industry. As stakeholders become increasingly aware of optimal cash holding decisions, they are better able to call for corporate actions that enhance corporate risk management and drive improved performance.

The rest of the paper is organized as follows. Section 2 develops our research hypothesis. Section 3 explains the regression model and machine learning techniques. Section 4 reports the regression results. Section 5 presents additional empirical results. Section 6 and Section 7 cover the discussion, limitations, and future research opportunities.

2. Hypothesis Development

Based on the resource-based perspective (Wernerfelt, 1984; Toms, 2010; Weigel & Hiebl, 2023; Iliyas & Barca, 2025), firms can achieve superior performance by strategically identifying and acquiring critical resources that enable developing products and services aligned with market demand. Cash is a strategically valuable resource, as it is the most liquid asset and serves as a buffer against future uncertainties (Opler et al., 1999; Dittmar et al., 2003; Ozkan & Ozkan, 2004; Y.-R. Chen & Chuang, 2009; Doshi et al., 2018; Lozano & Yaman, 2020). According to pecking order theory, managers prefer to utilize internal funds to finance investments because they are less expensive and involve fewer information asymmetry problems than external sources (Myers & Majluf, 1984; Y.-R. Chen & Chuang, 2009; Faysal, 2024; Eger & Hermis, 2025).

Cash management strategies are especially important for IT firms because strategic cash reserves ensure they can invest without delay and thus remain competitive (C. Kim & Bettis, 2014; La Rocca & Cambrea, 2019). An IT firm with insufficient cash may struggle to make timely investments given that external financing typically involves higher costs than using internal funds (Hennessy & Whited, 2007; Hirth & Uhrig-Homburg, 2010). Firms anticipating strong future growth opportunities are likely to hold higher cash reserves (Boot & Vladimirov, 2019). Frésard and Salva (2010) demonstrate that U.S.-listed firms with large cash reserves tend to achieve higher market values. Indeed, successful IT firms such as Apple, Alphabet, and Microsoft hold high levels of cash reserves (Theissen et al., 2023). Deb et al. (2017) argue that cash-rich companies can use their cash reserves to acquire key competitors and expand their market share. Cash-rich IT firms can also leverage their cash reserves to support innovation, which enhances corporate value in competitive, research-driven industries.

However, when an IT firm’s cash holdings are too excessive, they may be used inefficiently. One critical concern with high cash holdings is moral hazard, where managers could use the cash reserves to serve their own self-interests rather than the shareholders’ interests (Myers & Rajan, 1998). Thus, firms with more experienced board members tend to hold less cash reserves (MengYun et al., 2021). High levels of cash holdings are often associated with negative consequences, including diminished shareholder value (Lee & Powell, 2011), poor earnings quality (Sun et al., 2012), lower accruals quality (García-Teruel et al., 2009), decreased financial statement comparability (Habib et al., 2017), and greater engagement in aggressive real earnings management (Greiner, 2017). In this sense, excessive cash holdings in IT firms could impair their performance.

In practice, IT firm leaders tend to be highly ambitious and often exhibit risk-seeking behaviors, considering the higher failure rates associated with IT investments (Dewan & Ren, 2011; Xin & Choudhary, 2019; Saldanha et al., 2024). Given their substantial investment in R&D and exposure to volatile markets, driven by constant intense competition and pressure to innovate, IT firms should adopt effective cash management strategies to navigate uncertainty while seizing growth opportunities in a timely manner. Thus, we propose that there is an optimal level of cash holdings that maximizes performance in IT firms, and we present the following hypothesis.

Hypothesis. 

There is an inverted U-shaped relationship between cash holdings and performance in the IT industry.

3. Empirical Models

3.1. Regression Model

To test whether firms in the IT sector have optimal levels of cash holdings, we estimate the following regression model:

$${\mathrm{P}\mathrm{R}\mathrm{O}\mathrm{F}\mathrm{I}\mathrm{T}}_{\mathrm{i},\mathrm{t}}=\mathrm{\alpha }+{\mathrm{\beta }}_1{\mathrm{C}\mathrm{A}\mathrm{S}\mathrm{H}}_{\mathrm{i},\mathrm{t}}+{\mathrm{\beta }}_2{\mathrm{C}\mathrm{A}\mathrm{S}\mathrm{H}\_\mathrm{S}\mathrm{Q}\mathrm{U}\mathrm{A}\mathrm{R}\mathrm{E}\mathrm{D}}_{\mathrm{i},\mathrm{t}}+\mathrm{\lambda }{\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}\mathrm{l}\mathrm{s}}_{\mathrm{i},\mathrm{t}}+\mathrm{Y}\mathrm{e}\mathrm{a}\mathrm{r}+{\mathrm{\varepsilon }}_{\mathrm{i},\mathrm{t}}$$

(1)

where, for firm i and year t, PROFIT is Earnings Before Interest and Taxes (EBIT) divided by total assets. CASH is a cash ratio, measured as cash and marketable securities divided by the book value of total assets. CASH_SQUARED is the squared value of the cash ratio. We expect the coefficient (β₁) of CASH to be significantly positive and the coefficient (β₂) of CASH_SQUARED to be significantly negative, indicating an inverted U-shaped relationship between cash holdings and profitability within the IT sector. Control variables (Controls) include current assets divided by current liabilities (LIQUIDITY), tangible fixed assets divided by total assets (INVESTMENT), sales growth (GROWTH), total assets (SIZE), total debts divided by total assets (LEVERAGE), and R&D expenditure divided by Total Sales (R&D INTENSITY). We control for year fixed effects (Year) in our regression. Standard errors are robust to both firm-level clustering and heteroscedasticity.

3.2. Machine Learning Techniques

Ordinary Least Squares (OLS) regression is a widely used method for analyzing panel data due to its simplicity, interpretability, and established theoretical framework. However, OLS has its limitations, particularly in modeling nonlinear relationships, handling high-dimensional data, or capturing complex interactions among variables (Ozlem & Tan, 2022; Kayakus et al., 2023).

Machine learning (ML) algorithms offer a flexible and powerful alternative, especially in scenarios where traditional linear methods may underperform. ML techniques are well suited for handling multicollinearity and can better model nonlinear patterns (Lundberg & Lee, 2017; Tellez Gaytan et al., 2022; Mahmood et al., 2025). Because the performance of ML algorithms can vary by dataset and context, it is important to compare a range of models. In this study, we evaluate four widely adopted regression-based ML models: Decision Tree, Random Forest, Multilayer Perceptron, and Support Vector Machine.

Decision Tree is a supervised machine learning technique that constructs a tree-like structure to predict discrete or continuous outcomes. It recursively partitions the dataset based on feature values that minimize prediction error at each node. Although intuitive, easy to interpret, and computationally efficient, Decision Trees may suffer from overfitting, particularly in the presence of noisy data (Mohamed et al., 2012; C. Zhang et al., 2019). We employ the Reduced Error Pruning Tree (REPTree) variant in this study, as it is known for its fast regression work and reasonable accuracy (Shubho et al., 2019).

Random Forest is an ensemble technique that aggregates predictions from multiple decision trees to enhance overall model accuracy and stability (S. Zhang, 2024). During training, the mean predictions of several trees are aggregated to reduce variance, noise, and outliers (Mousa et al., 2022). Although more computationally intensive than a single decision tree, it performs well on datasets with complex, nonlinear relationships and interactions.

Multilayer Perceptron is a type of artificial neural network consisting of an input layer, one or more hidden layers, and an output layer. The network is fully connected, and it learns by adjusting connection weights between neurons to minimize the difference between its predictions and the actual outcomes—a process known as backpropagation (Mubarek & Adalı, 2017; Abedin et al., 2019). While Multilayer Perceptron requires careful tuning of parameters (e.g., number of layers, neurons, learning rate) and significant computational resources, they are highly flexible and effective in modeling complex nonlinear structures (Wu et al., 2022).

Support Vector Machine is a supervised learning algorithm used for both classification and regression problems. It identifies a hyperplane or decision function that minimizes prediction error while maximizing generalization capacity (C. Zhang et al., 2019; Ozlem & Tan, 2022). Support Vector Machine is particularly well suited for high-dimensional data and can capture nonlinear relationships. However, it is computationally intensive for large datasets and requires meticulous parameter tuning (Zahariev et al., 2022; Kayakus et al., 2023).

For data training and validation, we preprocess the dataset by removing missing values and outliers. To reduce variance and prevent overfitting, we apply 10-fold cross-validation, a popular technique to evaluate model performance (Jones et al., 2017). In this method, the dataset is divided into 10 equal subsets. The model is trained on nine folds and validated on the remaining one-fold, rotating the validation set across all folds. Although the dataset is structured as a firm-year panel, our primary objective is not to perform forecasting across time but rather to assess the relative explanatory power of different models (e.g., OLS versus machine-learning algorithms) using firm characteristics. Moreover, the use of 10-fold cross-validation is consistent with prior research applying machine-learning models to panel data in corporate finance and accounting (e.g., Chan et al., 2022; Hunt et al., 2022). To verify robustness, we repeated the random fold assignment using multiple random seeds and obtained consistent results. We believe that the standard 10-fold cross-validation provides a reasonable and widely accepted approach for evaluating model performance in this study without introducing bias from the panel structure of the data.

Finally, we assess model performance using three standard regression metrics: R² (Coefficient of Determination), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE). These metrics are widely used for an assessment of the models’ predictive accuracy. The hyperparameter configurations employed in this analysis are provided in Appendix A.

4. Empirical Results

4.1. Data and Descriptive Statistics

Corporate financial statement data were obtained from the Compustat annual database. The North American Industry Classification System (NAICS) is the standard employed by federal agencies in the US to classify industries according to their primary business activities (The United States Census Bureau, 2025).

Table 1. Selected IT Industries Based on NAICS Classification.

NAICS Codes	Industries
334	Computer and Electronic Product Manufacturing
51121	Software Publishers
51321	Software Publishers
517	Telecommunications
518	Data Processing, Hosting, and Related Services
519	Other Information Services
54151	Computer Systems Design and Related Services

presents IT industries selected based on the NAICS classification codes, whereby we identified the following: computer and electronic product manufacturing (NAICS 334); software publishers (NAICS 51121 and 51321); telecommunications (NAICS 517); data processing, hosting, and related services (NAICS 518); other information services (NAICS 519); and computer systems design and related services (NAICS 54151). Our sample comprises 21,051 firm-year observations with non-missing data across all variables from 2004 to 2023.

Table 2. Descriptive Statistics.

	N	25th Percentile	Mean	Median	75th Percentile	Standard Deviation
PROFIT	21,051	−0.115	−0.311	0.027	0.088	4.105
CASH	21,051	0.097	0.291	0.237	0.443	0.229
CASH_SQUARED	21,051	0.009	0.137	0.056	0.196	0.182
LIQUIDITY	21,051	1.169	3.043	1.969	3.460	5.412
INVESTMENT	21,051	0.037	0.140	0.082	0.175	0.158
GROWTH	21,051	−0.046	1.421	0.080	0.257	70.821
SIZE	21,051	37.573	5684.242	270.810	1727.153	25,421.916
LEVERAGE	21,051	0.003	0.552	0.131	0.335	9.241
R&D INTENSITY	21,051	0.011	1.234	0.094	0.195	28.665

Note: For each proxy, one representative variable is selected as follows. PROFIT is EBIT divided by total assets. CASH is a cash ratio, measured as cash and marketable securities divided by the book value of total assets. CASH_SQUARED is the squared value of the cash ratio, equal to cash ratio × cash ratio. LIQUIDITY is current assets divided by current liabilities. INVESTMENT is tangible fixed assets divided by total assets. GROWTH is sales growth. SIZE is total assets, expressed in millions. LEVERAGE is total debts divided by total assets. R&D INTENSITY is R&D expenditure divided by Total Sales.

presents descriptive statistics for the variables used in our analysis. For each proxy, one representative variable is selected, defined as follows. PROFIT is EBIT divided by total assets. CASH is a cash ratio, calculated as cash and marketable securities divided by the book value of total assets. CASH_SQUARED is the squared value of the cash ratio (i.e., CASH × CASH). LIQUIDITY is current assets divided by current liabilities. INVESTMENT is tangible fixed assets divided by total assets. GROWTH is the annual sales growth rate. SIZE is total assets, expressed in millions. LEVERAGE is total debts divided by total assets. R&D INTENSITY is research and development expenditure divided by Total Sales.

Table 3. Pearson Correlation Matrix.

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
PROFIT (1)	1.000
CASH (2)	0.017	1.000
CASH_SQUARED (3)	0.004	0.952 *	1.000
LIQUIDITY (4)	0.034 *	0.337 *	0.354 *	1.000
INVESTMENT (5)	−0.022 *	−0.317 *	−0.286 *	−0.119 *	1.000
GROWTH (6)	−0.002	−0.012	−0.008	−0.006	−0.008	1.000
SIZE (7)	0.022 *	−0.127 *	−0.106 *	−0.058 *	0.150 *	−0.004	1.000
LEVERAGE (8)	−0.525 *	−0.029 *	−0.022 *	−0.026 *	0.033 *	0.000	−0.007	1.000
R&D INTENSITY (9)	−0.030 *	0.061 *	0.078 *	0.043 *	−0.006	−0.001	−0.009	0.014	1.000

Note: * denotes significance at the 0.01 level. The definitions of variables are described in Table 2.

presents the Pearson correlation coefficients for the variables included in the regression model. The correlations between the cash ratio variables (CASH and CASH_SQUARED) and profitability (PROFIT) are found to be statistically insignificant. However, it is important to note that the Pearson correlation coefficient does not capture nonlinear relationships (Trustorff et al., 2011; Melo et al., 2020).

4.2. Multivariate Results

Table 4. Cash Holdings and Profitability.

	PROFIT
	Coef.	t-Value
CASH	1.7240	2.270 **
CASH_SQUARED	−2.4247	−2.995 ***
LIQUIDITY	0.0208	3.103 ***
INVESTMENT	−0.1179	−0.370
GROWTH	−0.0001	−0.433
SIZE	0.0001	5.100 ***
LEVERAGE	−0.2322	−7.630 ***
R&D INTENSITY	−0.0030	−3.760 ***
Year fixed effects	Yes
N	21,051
Adj. R-sq	0.277

Note: *, **, and *** denote significance at the 10%, 5%, and 1% (two-sided) level, respectively. Standard errors are robust to both clustering at the firm level and heteroscedasticity. The definitions of variables are described in Table 2.

reports the results from pooled OLS regression analysis, based on Equation (1), with profitability (PROFIT) as a dependent variable. The key explanatory variables are the cash ratio (CASH) and its squared term (CASH_SQUARED). We found that (i) the coefficient of CASH is significantly positive and (ii) the coefficient of CASH_SQUARED is significantly negative. These findings support the presence of an inverted U-shaped relationship between cash holdings and profitability in the IT sector, which is consistent with our hypothesis.

To provide a clearer understanding of economic significance, we calculate the turning point (−β₁/2β₂) of the quadratic cash–performance relationship. As indicated in Equation (1), β₁ and β₂ correspond to the coefficients of CASH and CASH_SQUARED, respectively. Using the coefficients reported in Table 4, the turning point is calculated as −1.7240/2 (−2.4247), yielding an approximate value of 0.356. Given that the 95% confidence interval for the estimate ranges from 0.264 to 0.447, the true turning point is highly likely to be located within this range. These empirical results carry managerial implications, suggesting that cash holdings outside this range are likely to be suboptimal for IT firms. Specifically, IT firms with cash and marketable securities below 26.4% of total assets need to consider increasing them, whereas IT firms exceeding 44.7% need to consider reducing them.

5. Additional Tests

5.1. One-Year-Forward Dependent Variable

As an endogeneity remedy, we use a one-year-forward dependent variable for the regression described in Equation (1).

Table 5. One-Year-Forward Dependent Variable.

	One-Year-Forward PROFIT
	Coef.	t-Value
CASH	2.0576	2.173 **
CASH_SQUARED	−3.2437	−2.459 **
LIQUIDITY	0.0161	2.151 **
INVESTMENT	−0.0132	−0.029
GROWTH	−0.0001	−0.612
SIZE	0.0001	2.968 ***
LEVERAGE	−0.2763	−4.164 ***
R&D INTENSITY	−0.0051	−2.143 **
Year fixed effects	Yes
N	18,281
Adj. R-sq	0.090

Note: *, **, and *** denote significance at the 10%, 5%, and 1% (two-sided) level, respectively. Standard errors are robust to both clustering at the firm level and heteroscedasticity. The dependent variable is the one-year-forward PROFIT, calculated as one-year-forward EBIT divided by one-year-forward total assets. The definitions of other variables are described in Table 2.

shows that the coefficient of CASH remains significantly positive, while the coefficient of CASH_SQUARED remains significantly negative, suggesting that the effect of cash holding decisions on subsequent profitability is significant over the next year. Our results collectively suggest an inverted U-shaped relationship between cash holdings and subsequent profitability, thereby supporting our hypothesis.

The estimated turning point is around 0.317, with the 95% confidence interval spanning from 0.264 to 0.371. This implies that IT firms holding less than 26.4% of total assets in cash and marketable securities need to consider accumulating more, whereas those above 37.1% need to consider trimming them to improve next year’s performance.

5.2. Subsample Analysis: Big IT Firms

As a robustness test, we perform a subsample analysis using a firm-fixed effects regression. First, we rank all observations into four groups based on the magnitude of firm size in each year.

Table 6. Subsample Analysis: Big IT Firms.

	One-Year-Forward PROFIT
	Coef.	t-Value
CASH	0.1207	2.903 ***
CASH_SQUARED	−0.1284	−1.776 *
LIQUIDITY	−0.0013	−1.065
INVESTMENT	0.0324	1.146
GROWTH	0.0014	0.962
SIZE	−0.0001	−1.810 *
LEVERAGE	0.0216	0.911
R&D INTENSITY	−0.3007	−4.909 ***
Year fixed effects	Yes
Firm fixed effects	Yes
N	5258
Adj. R-sq	0.040

Note: *, **, and *** denote significance at the 10%, 5%, and 1% (two-sided) level, respectively. The dependent variable is the one-year-forward PROFIT, calculated as one-year-forward EBIT divided by one-year-forward total assets. The definitions of other variables are described in Table 2.

presents the empirical results for firms in the top quartile of size in each year. Our empirical analyses—using a fixed effects regression model that controls for all time-invariant variables—reaffirm that the coefficient for CASH is positive and significant, while the coefficient for CASH_SQUARED is negative and significant. Interestingly, the inverted U-shaped relationship becomes insignificant for firms in the bottom quartile of size each year, although the results are not tabulated. Our findings suggest that large IT firms can enhance their performance by managing cash holdings at an optimal level.

The turning point is estimated at 0.470, with a 95% confidence interval from 0.188 to 0.752. Big IT firms with cash and marketable securities below 18.8% of total assets may improve future performance by increasing cash reserves, while those above 75.2% may need to decrease them.

5.3. Effect of Recessionary Periods

Firms faced significant economic shocks during recessionary periods such as the global financial crisis in 2007–2010 and the COVID-19 pandemic in 2020. To reinforce the robustness of the empirical results, we perform additional analyses focusing on these recession periods, all of which lie within our sample period. The findings reported in

Table 7. Effect of Recessionary Periods.

	PROFIT
	Coef.	t-Value
CASH	1.6279	2.587 ***
CASH_SQUARED	−2.7366	−2.565 **
LIQUIDITY	0.0248	1.928 *
INVESTMENT	−0.9977	−1.463
GROWTH	−0.0002	−1.907 *
SIZE	0.0001	3.480 ***
LEVERAGE	−0.2260	−6.207 ***
R&D INTENSITY	−0.0006	−1.326
Year fixed effects	Yes
N	5479
Adj. R-sq	0.613

Note: *, **, and *** denote significance at the 10%, 5%, and 1% (two-sided) level, respectively. Standard errors are robust to both clustering at the firm level and heteroscedasticity. The definitions of variables are described in Table 2.

reinforce the consistency of our main results. Table 7 presents a positive and significant coefficient for CASH and a negative and significant coefficient for CASH_SQUARED, providing evidence that is consistent with our hypothesis.

Our results indicate a turning point of 0.297, with a 95% confidence interval ranging from 0.212 to 0.383. This suggests that during economic downturns, maintaining cash and marketable securities between 21.2% and 38.3% of total assets may help IT firms optimize performance. Overall, the empirical results remain robust even during periods of economic shock.

5.4. Principal Component Analysis

The existing literature offers multiple proxy measures for each variable of interest. For instance, in the regression test above, we use earnings before interest and taxes (EBIT) divided by total assets as a proxy for PROFIT. However, other profitability metrics exist, such as net income divided by equity or gross profit divided by sales (see Appendix B). Prior studies show that the selection of proxy can substantially influence empirical results, meaning that findings derived from one proxy do not guarantee the same results when a different proxy is used to represent the same underlying construct (Jolliffe & Cadima, 2016). As such, testing multiple proxies is essential for robustness, but analyzing each one separately would be complex and time-consuming.

To address this issue, we adopt a master proxy approach using Principal Component Analysis (PCA), following the method introduced by Mahmood et al. (2025). PCA is a statistical technique used to reduce the dimensionality of a dataset while retaining as much of its original variance as possible. It transforms raw data into a new set of uncorrelated variables, called principal components (PCs). The first principal component (PC1) captures the greatest variance, the second (PC2) captures the next highest variance, and so forth. We applied PCA separately to the proxies for each conceptual variable to extract principal components. Because PCs capture the shared variance among all proxies, this method enhances the generalizability of our results compared to using any single proxy. In subsequent analyses, we use the first principal component (PC1) as the master proxy, as it captures the largest proportion of variance and offers a comprehensive summary of the underlying construct.

Table 8. Master Proxy Variances.

Variables	Number of Proxies	Variance Explained by PC1
Profitability	10	39.7%
Liquidity	2	57.7%
Investment	2	51.1%
Growth	3	36.4%
Size	5	67.6%
Leverage	3	48.8%

Note: The definitions of proxies are provided in Appendix D. PC1 denotes the first principal component as the master proxy.

presents the proportion of variance explained by each master proxy. For example, the PC1 derived from ten profitability proxies explains approximately 40 percent of the total variance—equivalent to capturing the information in four distinct proxies. Similarly, PC1s for investment, growth, size, and leverage explain 51.1 percent, 36.4 percent, 67.6 percent, and 48.8 percent of the variance in their respective proxies. These values suggest that the master proxies capture a substantial amount of information, providing a reliable foundation for our empirical analyses. Appendix C presents the scree plots and principal component loadings used in the analysis.

5.5. Cash Holdings and Profitability Using Master Proxies

Using the master proxies, we estimate the following regression model:

$${\mathrm{P}\mathrm{R}\mathrm{O}\mathrm{F}\mathrm{I}\mathrm{T}\_\mathrm{M}}_{\mathrm{i},\mathrm{t}}=\mathrm{\alpha }+{\mathrm{\beta }}_1{\mathrm{C}\mathrm{A}\mathrm{S}\mathrm{H}}_{\mathrm{i},\mathrm{t}}+{\mathrm{\beta }}_2{\mathrm{C}\mathrm{A}\mathrm{S}\mathrm{H}\_\mathrm{S}\mathrm{Q}\mathrm{U}\mathrm{A}\mathrm{R}\mathrm{E}\mathrm{D}}_{\mathrm{i},\mathrm{t}}+\mathrm{\lambda }{\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}\mathrm{l}\mathrm{s}}_{\mathrm{i},\mathrm{t}}+\mathrm{Y}\mathrm{e}\mathrm{a}\mathrm{r}+{\mathrm{\varepsilon }}_{\mathrm{i},\mathrm{t}}$$

(2)

where, for firm i and year t, PROFIT_M represents the master proxy derived from ten profitability variables. CASH denotes the cash ratio, calculated as cash and marketable securities divided by the book value of total assets. CASH_SQUARED is the square of the cash ratio. We expect the coefficient (β₁) of CASH to be significantly positive and the coefficient (β₂) of CASH_SQUARED to be significantly negative, indicating an inverted U-shaped relationship between cash holdings and profitability in the IT industry. Control variables (Controls) include the master proxies for two liquidity variables (LIQUIDITY_M), two fixed asset investment variables (INVESTMENT_M), three growth variables (GROWTH_M), five size variables (SIZE_M), three financial leverage variables (LEVERAGE_M), and research and development intensity (R&D INTENSITY). Full definitions of these variables are provided in Appendix D. In addition, we control for year fixed effects (Year) in our regression. Standard errors are robust to both clustering at the firm level and heteroscedasticity.

Table 9. Cash Holdings and Profitability Using Master Proxies.

	PROFIT_M
	Coef.	t-Value
CASH	0.0502	2.438 **
CASH_SQUARED	−0.0460	−2.380 **
LIQUIDITY_M	0.0068	2.363 **
INVESTMENT_M	−0.0047	−3.574 ***
GROWTH_M	0.0013	2.184 **
SIZE_M	0.0210	6.512 ***
LEVERAGE_M	−0.2958	−2.423 **
R&D INTENSITY	0.0001	2.101 **
Year fixed effects	Yes
N	21,051
Adj. R-sq	0.136

Note: *, **, and *** denote significance at the 10%, 5%, and 1% (two-sided) level, respectively. Standard errors are robust to both clustering at the firm level and heteroscedasticity. The definitions of variables are described in Appendix D.

presents the empirical results from Equation (2), employing master proxies. After controlling for these variables, the coefficient for CASH remains significantly positive, and the coefficient for CASH_SQUARED remains significantly negative. These results align with our earlier findings and further confirm the robustness of the hypothesized inverted U-shaped relationship.

5.6. Machine Learning Model Results

Next, we compare the predictive performance of OLS regression with several machine learning models. A higher R² indicates that the model explains more variance, while lower MAE (Mean Absolute Error) and RMSE (Root Mean Square Error) values suggest closer alignment between predicted and actual values. As shown in

Table 10. Comparison of Model Performances.

Models	MAE	RMSE	R2
Random Forest	0.0084	0.0723	0.6090
Neural Network	0.0128	0.0989	0.4293
Decision Tree	0.0140	0.1010	0.2174
Support Vector Machine	0.0128	0.1101	0.1207
OLS Regression	0.0824	1.0098	0.1360

Note: To assess the model performances, Mean Absolute Error (MAE), and Root Mean Square Error (RMSE), and R² are computed.

, the random forest model outperforms all others, achieving the highest R² and the lowest MAE and RMSE, indicating superior predictive performance. Compared to traditional OLS model, the random forest explains 47 percent more variance, highlighting a substantial improvement in predictive performance.

While OLS regression provides coefficients to interpret the direction and magnitude of each variable’s effect, tree-based and other nonlinear machine learning models such as Decision Tree, Random Forest, and Neural Networks do not yield interpretable coefficients. To address this, we compute SHAP (SHapley Additive exPlanations) values, which offer a theoretically grounded and interpretable measure of each feature’s contribution to the model’s predictions (Merrick & Taly, 2020; Meng et al., 2021; Nohara et al., 2022; Rozemberczki et al., 2022; Kori & Gadagin, 2024; M. Li et al., 2024). SHAP values not only reflect the importance of each feature but also indicate whether the feature has a positive or negative impact on the predictions (Acosta-Jiménez et al., 2024).

SHAP values are computed for all individual instances in the Random Forest model, and the mean values are used for global interpretation of the model.

Table 11. Mean SHAP values in Random Forest Model.

	Mean SHAP Value
CASH	0.000140
CASH_SQUARED	−0.000222
LIQUIDITY_M	−0.000064
INVESTMENT_M	−0.000367
GROWTH_M	0.000479
SIZE_M	0.003200
LEVERAGE_M	−0.000068
R&D INTENSITY	0.001593

Note: Dependent variable is the master proxy for ten profitability variables (PROFIT_M). The definitions of variables are described in Appendix D.

shows that the mean SHAP value for CASH is positive and for CASH_SQUARED is negative, once again supporting the presence of an inverted U-shaped relationship between cash holdings and profitability. In Appendix E, we present a SHAP dependency plot with additional discussion to support of an inverted U-shaped relationship.

6. Discussion

Our study explores the role of cash-holding decisions in improving IT firm performance. The IT sector is distinct due to the prevalence of risk-taking behaviors such as new market entry and high investment failure rates. Insufficient cash holding can constrain research and development (R&D) or acquisitions, exposing IT firms to the risk of financial distress and competitive disadvantages, whereas excessive cash may signal poor management, which could result in unnecessary expenditures or reduced shareholder value. Given competitive pressures and market volatility shaped by rapid technological innovation, IT firms should adopt efficient cash holding strategies to manage uncertainty and seize growth opportunities. We examine the impact of cash holdings on firm performance in the IT industry, using a dataset of 21,051 firm-year observations from US-listed IT firms between 2004 and 2023. To address limitations of conventional ordinary least squares (OLS) regression, we incorporate machine learning (ML) techniques and apply principal component analysis (PCA) to construct master proxies for key variables, thereby improving robustness and generalizability.

Our research is inspired by trade-off theory, which posits that firms can maximize value by maintaining an optimal debt-to-equity ratio. Likewise, our empirical analyses identify an inverted U-shaped relationship between cash holdings and profitability in IT firms, indicating that maintaining an optimal cash balance enhances firm performance in the IT sector. This relationship is valid for both conventional OLS regressions and machine learning models that use master proxies. In both specifications, the findings reveal an inverted U-shaped between cash holdings and firm performance indicating that while cash serves as a buffer against uncertainty and supports innovative projects, excessive cash retention may be associated with inefficient resource allocation or moral hazard. Even after controlling for key firm characteristics such as liquidity, investment, growth, size, leverage, and R&D intensity, the results consistently point to diminishing returns to cash. Overall, the evidence suggests that IT firm managers should pursue balanced cash-management strategies to optimize performance.

On the one hand, having insufficient cash can hinder IT firms’ ability to innovate and compete. Pecking order theory suggests that internal funds—such as cash holdings—are prioritized resources for investment because they are less costly than external sources. Without adequate internal funds, essential R&D projects might be delayed, and strategic investments such as product development or timely acquisitions of startups can lose momentum. In fast-moving technology markets, opportunities appear quickly and disappear just as quickly, and firms without sufficient cash might miss critical growth opportunities. Leading IT companies including Apple, Google, and Microsoft have demonstrated the value of maintaining cash buffers to remain agile and resilient. Thus, for IT firms on the left side of the inverted U-curve, increasing cash from low to moderate levels can yield substantial performance improvements.

On the other hand, excessive cash holdings can be detrimental for IT firms. Once liquidity exceeds the necessary level to fund operations and key projects, additional cash might introduce inefficiencies and governance risks. According to the agency theory, excess cash reserves often encourage managerial overconfidence or moral hazard, prompting executives to invest in risky projects or pursue questionable acquisitions simply because the funds are available. Empirical evidence also suggests that excessive cash can reduce shareholder value and compromise earnings quality (Harford et al., 2008; Chang et al., 2018; Farinha et al., 2018). Therefore, for IT firms on the right side of the inverted U-shaped curve, each extra dollar of idle liquidity imposes growing costs, in terms of both forgone opportunities and the agency problems that accompany financial slack. Therefore, IT firms must treat cash as a strategic resource that requires active management. Conducting regular assessments of cash levels against industry peers and internal benchmarks can help firms make informed adjustments to their cash policies in a timely manner.

A second key finding is the superior performance of ML techniques over traditional OLS regression. Among the models tested, the random forest demonstrates the strongest predictive accuracy and best overall fit, outperforming OLS regression by explaining 47% more variance. This result highlights the ability of ML methods to capture interactions and nonlinearities that linear models often overlook. While ML models often provide higher predictive accuracy than traditional methods such as OLS, they tend to be less interpretable. OLS provides clear, coefficient-based insights into how each variable influences the outcome, whereas complex ML models typically function as black boxes, making it difficult to understand how predictions are generated. Our framework offers a more balanced approach by combining the predictive power of ML with interpretable techniques such as SHAP values, which help clarify the contribution of each variable and enhance the model’s overall explainability. The SHAP analysis reinforces our findings. The influence of cash holdings on firm performance follows an inverted U-shape indicating that moderate cash levels enhance performance, whereas excessive holdings diminish it.

Finally, the PCA-based master proxy approach address proxy selection bias (Jolliffe & Cadima, 2016; Mahmood et al., 2025). For example, one master proxy constructed from ten profitability indicators captures nearly 40 percent of the total variance, equivalent to the explanatory power of four distinct proxies. We demonstrate that the master proxies capture a substantial amount of information, providing a reliable foundation for our empirical analyses.

Upon further analysis, we find that our empirical evidence holds during recessionary periods. This result implies that the inverted U-shaped relationship in IT firms can be leveraged as a strategic approach to risk management and performance, even under recessionary conditions. Interestingly, while the inverted U-shaped relationship between cash holdings and profitability becomes insignificant among IT firms in the bottom size quartile each year, it is significant for those in the top quartile. This implies that as an IT firm grows and increases in size, cash holding strategies become more critical to its performance, highlighting the need for managers in larger IT firms to pay closer attention to effective cash holding decisions.

Our research contributes to the literature on stakeholder theory by providing insights into how firms manage cash reserves within an optimal range. For example, our main findings suggest that IT firms with cash and marketable securities below 26.4% of total assets would benefit from increasing them, whereas firms above 44.7% may need to reduce them to optimize performance. As stakeholders gain greater awareness of optimal cash holding decisions, they are better positioned to influence corporate actions that boost risk management and overall performance. These insights are valuable not only for stakeholders but also for managers who design cash management strategies with future investment opportunities in mind. Our findings underscore the need for regulators to acknowledge optimal cash levels and enforce transparent cash disclosure policies, which can enhance trust and communication between corporate managers and stakeholders.

7. Limitations and Future Research

Our paper’s findings must be interpreted cautiously, as it is necessary to acknowledge several limitations. Although we establish a general inverted U-shaped relationship using a combination of quadratic regression, PCA-based master proxies, and comparative ML models, it is recommended that future studies further explore various methodological approaches to provide firm-specific guidance for managers and regulators on optimal cash thresholds and disclosure policies. We encourage future researchers to further investigate this research avenue.

While we utilize data across IT sectors using NAICS classifications, future work could explore whether the relationship varies across distinct industry contexts. For instance, the optimal level of cash holdings in physical capital-intensive industries may differ from that in human capital-intensive industries, where cash is primarily allocated to talent development. Future studies could extend the analysis to additional industry contexts to assess whether the cash–performance displays heterogeneous inverted U-shaped patterns or even alternative forms of nonlinear relationships.

It would also be an interesting avenue for future research to examine firms’ geographical location or cultural and social factors. For example, firms in countries with stronger uncertainty avoidance may hold higher levels of cash due to precautionary motives. Exploring how geographic, cultural, and social factors influence deviations from the optimal range of cash holdings could provide deeper insights into cross-country heterogeneity in corporate cash management practices.