We develop machine learning models that incorporate both external (deterministic) and internal (voluntaristic) factors affecting firm failure and survival. Using structured and unstructured data, we empirically investigate the external and internal factors that affect the US manufacturing firms’ business failure. We also examine how the interactions between external shocks and firm responses impact business failure. Our findings indicate that while external factors can significantly impact the likelihood that firms fail, specific management responses to these challenges can effectively mitigate the negative effects and contribute to firm survival.
]]>Corporate failure is a common business and economic phenomenon that fluctuates over time, particularly during economic cycles. Although bankruptcy is not completely synonymous with firm failure, they are interchangeable in many empirical studies of business failure (e.g., Altman et al. 2019;Mai et al. 2019;Thornhill and Amit 2003). Figure 1 illustrates the number of bankruptcy cases of public firms in the Compustat database. We can see that the number of firm failures has a similar pattern as the business cycles, with the peaks around the dot-com bubble crash (2000)(2001)(2002) and the financial crisis (2007)(2008)(2009). The total number of large bankrupt firms between 1997 and 2020 is 540, with 173 (32%) in the manufacturing industry. Rather than simply predicting firm failure, the motivation of this study is to examine the effects of both internal and external factors on the failure and survival of US manufacturing firms.
The COVID-19 pandemic has heightened the focus on business resilience. In evaluating the resilience of firms after the 2008-2009 financial crisis in early 2020, McKinsey and Company (2020) found that firms that balance margins, growth, and optionality (i.e., pay fewer dividends, keep more cash reserves) were more resilient after the 2008 crisis than those that focus on maximizing total shareholder returns. However, based on current data, the impact of the great recession in 2008-2009 on long-term unemployment in the United States was worse than we have experienced because of COVID-19 (Center on Budget and Policy Priorities 2023). The ability of many organizations to pivot to online operations has led to a softer impact from the pandemic. However, with new strains of the virus, supply chain disruptions, changes in workforce participation, and inflationary pressures, the long-term impacts of COVID-19 on the economy, and especially on manufacturing firms remain to be seen (Sanders 2023;Wellener and Hardin 2023).
Changes in Disaster Word Counts
Changes in disaster word counts in Form 10-K's by year. Note: The changes in disaster word counts reflect all kinds of disasters.
the northeast (see Figure 2). Although the total economic cost of these disasters is clear, it is less clear what the impact of natural disasters is on business failure. More specifically, COVID-19 is a special situation compared to other disasters and economic downturns (Crick and Crick 2020;Cortez and Johnston 2020).
The negative impact of economic shocks is amplified by the fact that firms do not have adequate foresighted decision models to plan and prepare for such shocks (Amankwah-Amoah and Zhang 2015). Existing predictive models of business failure and credit loss, widely used by the banking industry, do not provide foresighted decision support for management to help prevent such losses (Brainard 1967). These gaps in the literature have provided the motivation for this research to address the question:
The answer to this question is expected to help address a more general question: How can firms improve their resilience to cope with economic shocks as well as ongoing competitive pressures and changes? The objective of this paper is to contribute to the literature by not only examining business failure over a long period of time through a number of economic cycles, but we also deploy innovative machine learning techniques to develop and incorporate unstructured data into our models to gain a deeper understanding of the exogenous shocks firms encounter as well as how firms can effectively respond to them to survive.
This paper focuses on the manufacturing sector for two reasons.
First, manufacturing is one of the largest sectors in the US economy. 1 This implies that the manufacturing sector is a good proxy for the whole economy. Second, many manufacturing firms are B2B businesses whose supply chains might be more affected by economic shocks like COVID-19 or natural disasters than other industries.
Our findings provide new insights on firm failure and how firms can enhance their resilience. The results of our initial model confirm much of the conventional wisdom about bankruptcy prediction, while our second model specifically examines the role of exogenous factors including natural disasters in business failure. Our third model highlights the role of management actions in firm survival regardless of the exogenous environment. Similar to our first model, the findings of our third model are relatively consistent with the bulk of the management literature about management's role firm performance. Our fourth model integrates all aspects of the complexity of doing business in the real world, in which exogenous factors beyond the firm's control may occur and management is tasked with the challenges of navigating the firm through these unexpected events. Our final model has predictive power of over 86% and provides us with insights about the most important factors that lead to firm failure or survival. The findings of our fourth model indicate that there are important interactions between external factors and management's actions in the outcomes for the firm. New product developments and cost cutting can mitigate the effects of employee disruptions, and changes in investing can mitigate the impacts of disasters. These results support the integrative view and emphasize the need for a broader examination of firm failure beyond the current deterministic and voluntaristic approaches.
Our paper contributes to theory by addressing conflicting theories about the causes of business failure. We not only review and discuss these theories, but also provide evidence about their validity by testing the various components of these theories in a systematic way. We also contribute to the academic community by introducing analysis methods using machine learning and including novel visualization methods to assist with the interpretation of our results. Our results are also of value to practitioners who have long been using simplistic bankruptcy prediction models that do not incorporate exogenous factors or management skills and actions sufficiently to provide insight to either external creditors, investors or to managers charged with guiding the firms through challenging times.
The paper is organized as follows: Section 2 discusses the prior literature on business failure and resilience. This is followed by our theoretical framework and hypotheses in Section 3. We then discuss the methodology including variable definitions and data collection in Section 4. Section 5 discusses the findings of our models, followed by discussion and conclusions in Section 6. Finally, Section 7 presents the limitations and opportunities for future research.
2 Related Literature
Business failure has long been studied in the literature of management and business. Many studies use the traditional method of discriminant analysis (e.g., Altman 1968;Beaver 1966;Deakin 1972Deakin , 1984) ) or the logit/probit model (e.g., Kamalirezaei et al. 2020;Ohlson 1980;Zmijewski 1984) to predict business failure/bankruptcy. Other studies use different methods like linear regression (e.g., Collins 1980), or rough set approach (e.g., Dimitras et al. 1999).
We examine both internal and external factors systematically to predict the US manufacturing firms' failure and better understand the impact of, and interactions between, these factors. Several other studies in the literature also studied external (risk) factors (e.g., Everett and Watson 1998), and (internal and external) perceived causes (e.g., Gaskill et al. 1993). Altman et al. (2019) discuss that three common causes of business failure are unexpected international competition, lack of technological innovation, and poorly executed strategic investing decisions. In addition to the traditional financial determinants of business failure, our study incorporates qualitative variables derived from text analysis of unstructured data, which is informed by the work of Mai et al. (2019).
Since the emergence of COVID-19, there have been a few studies to examine how firms are coping with the pandemic as well as the incidence of business failure during the pandemic. For example, Amankwah-Amoah et al. ( 2021) developed a theoretical model describing how firms must adjust to exogenous shocks as well as predictable changes in the business environment to avoid failure. Li et al. (Forthcoming) examined text from earnings calls to determine whether the vulnerability or exposure of the firm to COVID-19, as well as the firm response affected stock returns. Using word embedding, they developed dictionaries describing six exposure variables and four response variables from firm earnings calls to discuss the impact of COVID-19. They also included culture variables developed by Li et al. (2021) to determine whether culture had a significant impact on the firms' stock returns. They found that culture had a significant effect on the ability of firms to respond to COVID-19 related economic exposure to affect stock returns.
Supply chain disruptions were a significant problem encountered during the pandemic (Sanders 2023;Xie et al. 2023); therefore, examining the specific challenges that manufacturing firms face during times of stress, and management actions to cope with them are particularly relevant.
Numerous scholars have addressed the issue of business resilience from different perspectives. Annarelli and Nonino (2016) conducted a comprehensive review of articles between 1990 and 2014 and identified over a thousand articles addressing resilience in the management, engineering, and economics literature, with a steep increase in publications on the topic starting in 2006. Conz and Magnani (2020) conducted a systematic examination of articles published between 2000 and 2017 in the business and management literature, noting a significant increase in publications on the topic after the 2008-2009 financial crisis.
From the business strategy perspective, a firm's resilience plays a vital role in helping firms deal with shocks and uncertainty.
It helps firms to improve financial volatility, sales growth, and survival rates (Ortiz-de-Mandojana and Bansal 2016). Note that the firm's resilience is a dynamic process in which firms absorb, adapt, and recover from shocks and uncertain environments (Conz and Magnani 2020;Li et al. 2019;Ortiz-de-Mandojana and Bansal 2016). As an element of the firm's resilience, supply chain resilience also plays an important role in firms' development, particularly in the manufacturing industry (Karp 2023). Ponomarov and Holcomb (2009) show that the greater the supply chain resilience, the greater the sustainable competitive advantage.
In regard to policy shocks, Kang et al. (2014) show that economic policy uncertainty (shocks) in interaction with firm-level uncertainty depresses firms' investment decisions. Specifically, the policy uncertainty causes significant declines in output, consumption, investment, and employment (Basu and Bundick 2017;Baker et al. 2016). Firm's strategies respond to regulatory uncertainty by participating in policy making and increasing strategic flexibility (Engau and Hoffmann 2009;Shaffer 1995).
There is little research connecting the impact of policy shocks and uncertainty on firm survival.
The business failure literature has been dominated by two competing theories (Mellahi and Wilkinson 2004). Scholars adhering to the deterministic point of view often draw on the industrial organization and organizational ecology literatures. The deterministic view posits that due to volatility and increased levels of uncertainty there is frequently an inability by management to predict or foresee and to react appropriately to these external conditions, thus leading to firm failure. Alternatively, proponents of the voluntaristic point of view take the approach that firm management can have a strong effect on how the firm reacts to external events and may effectively influence the success or failure of the firm within the changing environment. This perspective
The modified business failure framework.
links business failure to internal management inadequacies in dealing with external threats. Mellahi and Wilkinson (2004, 32) develop an integrative framework to combine environmental and ecological factors with organizational and psychological factors in a more comprehensive model to explain organizational failure. Amankwah-Amoah et al. ( 2021) further expanded on the integration of the deterministic and voluntaristic views by developing a model that shows the dynamic and temporal characteristics of the interaction between external factors and internal management responses.
Building on Amankwah-Amoah et al. (2021)'s framework, we propose our modified business failure framework for discussing various strategies and hypotheses that may influence a firm's ability to survive or that lead to failure. Our theoretical model, a modification of earlier integrative models, is illustrated in Figure 3 and is operationalized using a dynamic analysis of the interaction between external factors and management actions, with periodic monitoring, over time.
Our model uses these initial firm states, and then examines the impact of external shocks, and firm responses and actions to predict the firm outcome of failure or survival. Our model consists of the following elements.
• Firm states, a collection of information that represents the status of the firm, such as human resources (e.g., number of employees), financial resources and capital (e.g., cash, inventory, asset and debt and equity levels), as well as prior financial performance experience (e.g., return on assets, earnings before income taxes, Tobin's Q, and sales growth);
• External shocks, including both disasters and changes in market conditions, supply and demand shocks, distresses, and so forth;
• Firm actions, both tactical and strategic, such as expenditures, investments or divestitures, or financing adjustments which are derived from quantitative disclosures from cash flow and income statements (e.g., advertising, R&D, capital expenditure, acquisition, investing, sale of stock, purchase of stock, and borrowing), and firm actions and responses extracted from rhetoric in firm disclosures using textual analysis (e.g., community engagement, cost-cutting, digital transformation, new product development);
• Firm outcomes (i.e., failure or survival).
The firm's initial state reflects the external environment of the firm, including the industry it operates in, and the firm's position in this external environment. Traditional economic theory ties firm productivity to the combination of labor and capital deployed. As summarized by Mellahi and Wilkinson (2004), organizational engineering scholars have identified four factors or characteristics that influence the success or failure of organizations. Two of these factors are related to industry characteristics-population density (Delacroix et al. 1989;Hannan and Freeman 1988;Hannan et al. 1991;Peterson and Koput 1991), and industry life cycle (Agarwal et al. 2002;Balderston 1972). A third factor is organization size (Barnett and Amburgey 1990;D'Aveni and Hambrick, 1988;Wholey et al. 1992). The fourth factor noted is organization age (Baron et al. 1994;Bruderl and Schussler 1990;Fichman and Levinthal 1991;Levinthal 1991;Stinchcombe 1965), which is somewhat endogenous to the model as the age of the firm might be seen more as a result of survival rather than a determinant of failure.
The firm's survival may also depend on the ability to access sources of financing and the opportunity to regulate its sources and costs of capital. With regard to financial risk, high financial leverage and low liquidity can cause business failures when interest rates rise or there are funding shocks, as in 2008-2009. Amankwah-Amoah and Zhang (2015) describe three cases in the airline services industry in which overextension of debt from acquisitions led to business failure. Occasionally, unexpected liabilities can arise from litigation or environmental events, which may lead to filing for bankruptcy (for example, Remington Outdoor Company, Brickley 2020). We therefore propose that high degrees of leverage will be detrimental to firm survival.
As previously noted, McKinsey and Company (2020) found that during the 2008-2009 crisis, firms with higher levels of retained earnings were more resilient. Although there may be some debate about the return of capital to investors in the form of dividends, firms with growth potential are likely to be able to retain earnings within the firm. This may result in more financial flexibility in times of economic stress. Based on prior literature, we propose that larger firms have greater capacity for survival and that higher levels of resources represented by physical as well as human resources can help insulate the firm from failure. This leads to our first hypothesis:
H 1. Firms with greater levels of resources, both capital resources and human resources, and lower financial leverage will have a lower likelihood of failure.
Although firms can control their internal resources, they cannot easily control the external (exogenous) environment. While the general environmental and economic environment may normally be easily anticipated, our model assumes that there may be unexpected exogenous shocks that may cause the firm to fail, or to which the firm will need to react to survive.
Some of the most common exogenous shocks are natural disasters. Natural disasters cause negative effects on firms (Hsu et al. 2018; Huang et al. 2018; Pankratz et al. 2021). The effects include both direct and indirect damages. The indirect costs of natural disasters, particularly large natural disasters, might be larger than the direct costs (National Research Council 1999). For example, natural disasters can cause a severe disruption of the supply chain greater than the direct effects (Inoue and Todo 2019; Xie et al. 2018). Xie et al. (2018) find that dynamic resilience could have reduced business disruption resulting from the Wenchuan earthquake by 47.4% during 2008-2011 and could have shortened the recovery period by 1 year. There is an agreement in the literature in business and management that firm's and supply chain resilience and adaptive policies are the keys for firms to deal with shocks (Boin and van Eeten 2013; Koronis and Ponis 2018). Specifically, Covid-19 has had a worldwide effect on supply chains and business failures (Kalemli-Ozcan et al. 2020; Wang et al. 2020; Amankwah-Amoah et al. 2021). Another common external shock that can be detrimental to the firm's survival is a significant change in the level of competition. Increasing market concentration measured by the Herfindahl-Hirschman index has a negative effect on firm duration, that is, firm survival (Kaniovski and Peneder 2008). Altman et al. (2019) argue that unexpected international competition is one of the common causes of business failure. Numerous scholars have reviewed how firms respond to various types of shocks, such as supply shock (Fasani and Rossi 2018), demand shock (Kee and Krishna 2008; Leduc and Liu 2016; Wen 2006), both supply and demand shock (Guerrieri et al. 2020; Hassan et al. 2020; del Rio-Chanona et al. 2020), and policy shock (Baker et al. 2016; Basu and Bundick 2017; Kang et al. 2014). Demand-side shocks can generate realistic business cycles (Wen 2006) and can affect firms differently depending on their age. In some contexts, there is a relationship between supply and demand shocks. For example, Guerrieri et al. (2020) show that, in the COVID-19 pandemic, Keynesian supply shocks can happen: supply shocks that trigger changes in aggregate demand may be larger than the shocks themselves. Considering the above arguments and recent literature, we propose the following hypothesis:
H 2. The occurrence of exogenous shocks (e.g., natural disasters, demand shocks, liquidity shocks) is positively related to firm failure.
The strategy literature contains many theoretical frameworks describing how management's choices of business unit strategies may lead to greater profitability. Drawing on the concept of contingency theory (Hofer 1975), the underlying assumption is that successful business performance involves selecting an appropriate combination of factors that management controls relating to marketing, production, and investment. Schendel and Hofer (1979) proposed four elements of strategy that management should consider: the scope of product and market matches; the resource deployments and competencies of the firm; competitive advantages; and synergy. Galbraith and Schendel (1983) provide an excellent summary of several seminal typologies of business strategies. The most well-known of these management strategies and archetypes focus on broad management decisions such as a cost focus or product offering differentiation (Porter 1980), market concentration and asset reduction (Hofer and Schendel 1978), and sales and market access maximizing initiatives (Utterback and Abernathy 1975). From a marketing perspective, Woo and Cooper (1981) propose variables such as pricing, promotion, and research and development as important variables for management to control. Based on this body of strategic management and marketing literature, as well as the voluntaristic view previously cited, we propose our third hypothesis focusing on the impact that management actions and decisions can have on the firm's success or failure.
H 3. Management decisions can have a significant influence on a firm success or failure.
Business resilience involves both taking advantage of unexpected opportunities and mitigating the damage from new threats. Individual firms are in dire need of guidance on how to make timely decisions to survive the crisis and foresighted decisions to thrive after the crisis (Altman and Hotchkiss 2010). Morrish and Jones (2020) develop a framework for a post-disaster recovery model. They suggest that management leaders are choosing among resuming, delaying or ceasing the operation of business. Srinivasan et al. ( 2011) examine whether companies should engage in R&D and advertising activities during recessions. They found that firms with greater market share can improve profits followed by higher R&D spending while decreasing profits by an increase in advertising expenses during a recession. When firms are highly leveraged, they are more likely to cut discretionary expenses such as advertising (Grullon and Kanatas 2006). However, if those highly leveraged firms spend more on advertising, this signals their competitive advantage, leading to an increase in profits because they differentiate their products from their competitors during a difficult time (Srinivasan et al. 2011). Corey and Deitch (2011) examine what factors affect business recovery after a catastrophic hurricane and find that firms with greater size and age can recover from a disaster because they tend to have more business locations or capital to endure the loss of damages. In several interviews with small businesses, Runyan (2006) indicates that small businesses that reopen faster experience high market demand due to a lack of competition and an increase in population concentration when people move from destroyed areas and are being evacuated. To respond faster to a crisis, firms that increase capital expenditures can expand their capacity and reopen sooner. In addition, Pan and Qiu (2022) document that firms are more likely to acquire firms located away from the flooding areas to mitigate the disruption caused by the flood.
Allen et al. ( 2022) investigate the impact of natural disasters on community bank liquidity and find that banks' deposits decrease and loans increase after natural disasters. The increased demand for monetary resources for individuals and businesses causes this phenomenon. Addressing liquidity would be the top priority for business survival, and thus, we hypothesize that long-term borrowing can increase the chance of business survival. In addition, Runyan (2006) also documents the challenges faced by small businesses, such as vulnerability to cash flow interruption and lack of access to capital for recovery.
Based on this prior literature, as well as the integrative models of business failure by Mellahi and Wilkinson (2004) and Amankwah-Amoah et al. ( 2021) previously described, our fourth hypothesis focuses on the interactions between external exposures and management responses and actions on firm survival or failure.
H 4. The interaction between timely managerial actions in response to external events and exposures by modifying operating activities and adjusting resources and obligations will be positively correlated with firm survival.
Methods-Model, Data, and Variables
The operational version of our theoretical model and hypotheses is illustrated in Figure 4.
In order to study business failure and resilience and test our hypotheses, we use a combination of traditional and machine learning techniques to build a logistic regression model. The logistic model (or the logit model) is a linear binomial classification method that models the probability of business failure as a function of independent variables:
where ÿ 0 , … , ÿ ÿ are coefficients and ý 0 , … ., ý ÿ are dependent variables. After obtaining the logistic model, we will use the model to test our hypotheses and study how a variety of factors influence business failure.
Because we are modeling a complex economic phenomenon over a long period of time, our study draws data from a variety Illustration of the four hypotheses. of sources. The data are obtained from SEC filings on the EDGAR database and processed by Professor Bill McDonald and Compustat from US public companies from 1995 to 2021.foot_2 We match the records from different sources using firm CIK numbers and years. Since a firm may not have records in all three data sources for the required variables, after merging the data sources, we obtain 478 bankrupt firm-year observations and 155,645 non-bankrupt firm-year observations.
In this study, we focus specifically on US manufacturing firms.
In particular, manufacturing firms also have inventory, debt, and employment data, providing the best data set for examining the issue of firm failure or resilience. In the dataset, we have 167 bankrupt manufacturing firm-year observations and 59,018 non-bankrupt manufacturing firm-year observations. Furthermore, we remove some outliers (e.g., firms with negative sales and less than 1 million assets and 5 employees), resulting in 149 bankrupt manufacturing firm-year observations and 52,872 non-bankrupt manufacturing firm-year observations. Because our model incorporates a large number of variables, there are many missing values in the dataset, which results in the elimination of firm-year data. After the elimination of missing values, we have 42 bankrupt firm-year observations and 18,297 non-bankrupt firm-year observations.
Appendix A details the definition of each measurement variable included in the model, organized in the following categories, which parallel our hypotheses.
Firm's initial states-firm resources, capabilities, and capital structure: Our first hypothesis focuses on the impact of the firm's resources and the initial environmental state at the start of the time series analysis. Therefore, our first set of state variables is quantitative variables that reflect the industry the firm operates in, and the assets and resources it has to generate profits for owners and to pay obligations. These variables mirror the traditional production function of labor and capital. Following Mai et al. (2019), we start with traditional quantitative operational accounting resources measurements, including cash, working capital, inventory, and retained earnings, as well as debt levels and financial ratios. We also include resources that are not on the balance sheet, such as the number of employees. To represent the firm's initial state of performance, we also include the prior year performance variables for return on assets, pretax return on assets, market value, and sales growth (ROA, EBITAT, Tobin's Q, and SalesGrowth). We created five industry subclassifications using SIC codes and the final consumer and manufacturing processes.
Exogenous shock variables are measured in three categories.
Our first exogenous variables are control variables typically used in academic accounting research (e.g., Kubick et al. 2015;Cen et al. 2018), including GDP, inflation, and interest rates. We reflect the industry environment by calculating the Herfindahl-Hirschman Index for each firm year to measure industry concentration/competition within the three 3-digit SIC codes.
The second category is business shocks, categorized into six events that may affect businesses in an economic downturn or other unexpected events. These include disruptions in business operations, demand disruptions, layoffs of employees, liquidity shocks, lockdowns, and supply chain shocks. These variables are derived from the 10K text analysis using dictionaries drawn from Li et al. (2021), which are detailed in Appendix 12.2. We define the value of the firm's exposure to each exogenous shock as 1 if the firm's word counts are in the 75th percentile of the sample in year t, and -1 if the firm's word counts are in the 25th percentile of the sample in year t. Firms with word counts between these high and low levels were counted as 0, reflecting that the firm's exposure was moderate compared to other firm years.
Text variables measuring natural disasters: We also measure shocks from natural disasters. These were derived using the Form 10-Ks text analysis and a word count dictionary of specific disaster-related words such as flood, fire, storm, hurricane, and so forth. With regard to natural disasters, we used terminologies from several sources, such as the National Oceanic and Atmospheric Administration (NOAA) database and the Emergency Events Database (EM-DAT), to construct a dictionary of natural disasters. We then count the number of words in Form 10-Ks related to this natural disaster word dictionary. In 10-Ks, firms will disclose their risk metrics and describe potential impacts of natural disasters repeatedly over the years. To mitigate the duplicate word counts, we take a change in word counts of disasters from year t -1 to year t. The complete word list is reported in Appendix B1.
Quantitative variables: To measure firm responses to economic shocks, we first drew quantitative variables from the financial statements (Mai et al. 2019). Market strategy variables from the firm's financial data include advertising and R&D. For long-term asset and liability responses, we included capital expenditures, acquisition spending, investing, sales of property plant and equipment, sales of stock, purchases of stock, borrowing, and repayment. We assume that the decisions made in the current year will take effect in the future. Therefore, quantitative variables for decisions are the change values of those accounting items from year t to year t + 1, scaled by lag total assets.
To qualitatively measure firm responses and actions, we drew from four response variables created by Li et al. (2021). Using their word dictionary detailed in Appendix B3, we created word counts of the four responses from the Form 10-Ks: community engagement, cost cutting, digital transformation, and new product development. We defined each item as 1 if the firm's word counts are in the 75th percentile of the sample in year t, and -1 if the firm's word counts are in the 25th percentile of the sample in year t. Firms with word counts between these high and low levels were counted as 0, reflecting that the firm's response was moderate compared to that from upper and lower percentile firms.
Appendix A defines the detailed measurement and variable names we used to operationalize our theoretical model. Our key dependent variable is Failure, defined by the code of reason for a firm being deleted in Compustat. We defined Failure as a binary variable equal to 1 when the code is either 02 for bankruptcy or 03 for liquidation, and 0 otherwise. This means firms that were deleted from Compustat due to acquisition, merger, leveraged buyout, or other non-failure reasons are not considered business failures. Furthermore, our action variables are developed as the change values of accounting items. To avoid missing values in the future, business failure is 1 if firms are removed from Compustat in year t + 1, and 0 otherwise.
Table 1 shows the summary statistics of all variables for the main logistic model in this study. The mean of Failure is relatively low, implying that a bankruptcy is a rare case. The mean of Disaster is 0.261, and the median is 0, suggesting that firms that disclose disaster words in 10-Ks are highly skewed to the left.
Even though Disaster at the 75th percentile is 1, the majority of the sample is not normally subject to severe disaster events. The means of response and exposure variables are not close to zero because we measure them before we merge the text sample with the Compustat dataset. The most significant exposure in our sample is the supply chain exposure (SC_exposure), as the mean is 0.452, while the least mentioned exposure is liquidity exposure (LQ_exposure), as the mean is -0.016. For responses, new product development (NP_response) is frequently captured in the manufacturing sample, as the mean is 0.216, while the least frequently mentioned response is community engagement (CE_response), as the mean is 0.129.
Our initial data set of variables drawn from Mai et al. (2019) includes many variables that are highly correlated. However, when interpreting the meaning of such models, the inclusion of highly correlated variables can present a challenge. Similar to multiple regression, the correlation between the related variables may result in offsetting positive and negative coefficients, which may not allow an accurate interpretation of the model. Therefore, we use a systematic process to build a more parsimonious model and select meaningful independent variables to obtain the best model fit and to provide interpretable coefficients. The detailed procedure is provided in Appendix C.
We use the logistic model to fit the data. The accuracy is the classic criterion to evaluate how well the model fits the data. However, our data set is highly imbalanced (i.e., 42 bankrupt and 18,339 non-bankrupt firm-year observations). A more appropriate criterion is the balanced accuracy, defined as the average of the true positive rate (i.e., percentage of correctly predicted bankrupt observations) and the true negative rate (i.e., percentage of correctly predicted non-bankrupt observations). For example, if we had a model that simply outputs surviving firms regardless of the independent variables, we would have an accuracy of 18,339/(42 + 18,339), which is almost 100%, while the balanced accuracy is close to (0% + 100%)/2 = 50%. Therefore, the balanced accuracy is a more appropriate performance criterion for our imbalanced dataset.
To maximize the balanced accuracy, we employ balanced logistic regression, which assigns different weights to bankrupt and nonbankrupt observations that are inversely proportional to their percentages in the data. In our case, the weight of failure observations is about 400 times that of surviving observations.
We will present detailed performance measures in the form of the confusion matrix, which tells us how many failed and surviving observations are predicted correctly.
Based on the predictive model we have developed, we can gain insights about the exposures and responses that may lead to firm failure. For each of our hypotheses, we have created a logistic regression model using machine learning techniques to predict bankruptcy. After completing the variable selection described in Appendix C, for each hypothesis, we present a dendrogram for each model reflecting a hierarchical cluster analysis illustrating the relationships between independent variables. We also performed a permutation feature importance analysis for each model. This technique allows us to quantify the importance of each independent variable in the predictive power of the model. For each independent variable, we randomly shuffle it 500 times and compute the average reduction in the accuracy as its permutation importance (Molnar 2022;Fisher et al. 2019). In
Confusion matrix (balanced accuracy = 79.2%).
Actual Surviving 14,609 3688 Failed 9 33
this way, we can determine how much of the predictive power is attributable to each variable.
In Model 1, using only the selected state variables for our initial predictive model, we have achieved 79% accuracy in the bankruptcy prediction (Failure). The permutation feature importance figure indicates that the most important factor in bankruptcy prediction is GDP, followed by sales growth. Liquidity and asset utilization, measured by current ratio (CurrentRatio and CashToCL) and assets per employee (AssetsperEmp), are important, as well as the environmental factors of Inflation and industry competition (HHI). As expected, the firm's equity ratio (Equity Ratio) is also an important factor. We do not see strong support for the proposition that larger firms (Log_Assets) are less likely to go bankrupt, except with regard to the market value of the firm (Q). The sub-industry effects are more easily interpreted in combination with the coefficient diagram inExhibit 1B, which shows that food and textile firms (SubIndustry1) have a higher than average bankruptcy experience, while the other subindustries are below the average of the total sample. Although our model shows factors that predict firm failure fairly accurately, we do not have strong support that firm size has a significant impact on firm failure. Without considering other events, the external environment and liquidity seem to be the primary determinants of firm failure.
Coefficients of independent variables. The coefficient of AssetsPerEmp is 0.00017. It is too small to show on the graph.
Permutation importance of independent variables.
Dendrogram -Hierarchical clustering of independent variables.
Confusion matrix (balanced accuracy = 78.9%).
Actual Surviving 14,479 3818 Failed 9 33
To test Hypothesis 2, we created the analysis inExhibit 2. This model includes the selected state variables from model 1 and adds the exogenous shock measures, including natural disasters and the six exposure variables measured by the 10 K text. As there is a significant correlation between the various exposure variables, following our previously described variable selection process, we retained supply chain exposure (SC_exposure) and employee exposure (EM_exposure) and Disasters as exogenous shocks. The resulting model has a similar predictive accuracy as Model 1, of about 79%. However, the permutation feature importance in Exhibit 2C reveals that GDP is no longer a key determinant of failure and that liquidity, sales growth, and assets per employee are more important. Supply chain exposure has a significant impact on firm failure. Employee exposure is also important, along with profitability (ROA), inflation and firm market value. Although disasters do influence the outcome they are somewhat lower in importance on the variable list.
This model provides moderate support for Hypothesis 2 for the deterministic view that exogenous events are a cause of firm failure.
In Exhibit 3, we create the model to test Hypothesis 3 by including the selected state variables from model 1 and adding the management action variables. Management actions are measured using structured variables from the financial statements and the unstructured response variables from the 10 K text. The model accuracy improves to 84.5%. Although there was some correlation between the text response variables, the signs of the coefficients did not reverse when we removed them one at a time, so we have included all 4 responses in the model. The permutation importance figure shows that liquidity (CashToCL), sales growth in prior year (SaleGrowth_Lag), and profitability are the three most important factors in predicting firm failure. Asset utilization, as measured by inventory to sales (InvtToSale), is also important. Interestingly, the community engagement response (CE_response) variable is also very important. However, the coefficient for this variable is positive, indicating that firms that have more community engagement are more likely to go bankrupt. In this case, we feel this is a correlation and not a causation relationship. As firms begin to experience problems, they are likely tasked to engage in more community dialogue. However, this discussion does not prevent them from failing. The market size of the firm is also important and negatively related to firm failure. The change in financing (Ch_OtherFinancing) and dividends
Coefficients of independent variables.
Permutation importance of independent variables.
Dendrogram -Hierarchical clustering of independent variables.
Confusion matrix (balanced accuracy = 84.5%).
Actual Surviving 14,786 3511 Failed 5 37
(Ch_Div) become key factors in this model, a sign that firms adjust their capital structure to avoid bankruptcy. Inflation and changes in interest expense (Ch_Int) also play an important role in firm failure, as well as changes in acquisitions (Ch_Acq). Examining these key features in light of the correlations, we see that the signs of the coefficients are relatively as expected, although there may be offsetting coefficients in some investing and financing items, which are drawn from the cash flow statement. As expected, cost cutting (CC_response), digitization (DT_response), and new product development (NP_response) are all responses that are negatively correlated with bankruptcy; however, the permutation importance of these variables is relatively minor.
The importance of several of the management action variables indicates that there is support for the voluntaristic view in Hypothesis 3. There are specific management actions and decisions that have a significant impact on firm failure.
In Exhibit 4 we include the selected state variables, as well as the selected exposures, and responses, and add interaction variables between exposures and responses to our model. This model is designed to test the integrative theory in Hypothesis 4. This increases the accuracy to 86.6%. Examining Exhibit 4C, we see the most important factor is profitability (ROA), followed by sub-industry effects (SubIndustry3). Interestingly several of the interaction variables become very important when added to this model. Specifically, the interaction of employee disruption exposures with new product development is very important. Supply chain exposures with community engagement response is also important. We see disasters taking on a more important role in this model, and the interaction of disasters with changes in investments (Disaster*CH_OtherInvesting) becomes very important. Employee exposure with a cost-cutting response is also an important interaction (EM_exposure*CC_response). In this model we see leverage (Leverage) taking a more important role, along with several other important change variables dealing with management's financing and investing decisions. Inflation is also an important external factor as well as the US borrowing rate (USBorrowingRate). Cost-cutting response as a standalone variable is also important.
When we examine the permutation importance of this model, we can clearly see that it supports the integrative theory of business failure. The external environment and exogenous events, as well as management's responses to those events, all have a significant impact on the firm's failure. This model provides strong support for Hypothesis 4.
In this paper, we have analyzed the business failure of US public manufacturing firms in detail for a period of over 20 years,
Coefficients of independent variables.
Permutation importance of independent variables.
Dendrogram -Hierarchical clustering of independent variables.
including the economic downturn in the early 2000s, and on through the current economic challenges created by the COVID-19 pandemic. Our analysis employs machine learning techniques to incorporate a large variety of factors to improve the predictive accuracy of the models. Unlike prior bankruptcy studies that simply focused on predicting bankruptcy, our analysis first develops an acceptably accurate model and then analyzes the specific factors relating to business failure and the interaction between exogenous factors and management responses. We also deploy innovative text variable measurement techniques from unstructured data to analyze business exposures and management responses derived from rhetoric included in firms' 10-Ks.
Our research has uncovered fresh insights into how firms can improve their resilience and avoid failure. We used four different models to explore various factors that contribute to business failure, with our first model aligning with traditional bankruptcy prediction methods. The second model focused specifically on the impact of external factors, including natural disasters, while the third model examined the importance of management actions in ensuring firm survival regardless of the external environment.
The results of our fourth model, which took into account the complexity of doing business in the real world, showed that a combination of external factors and management actions plays a significant role in determining whether a firm succeeds or fails.
Our final model was able to predict with 86% accuracy the most critical factors that lead to firm failure or survival. We found that there are significant interactions between external factors and management actions, with new product developments helping to mitigate the effects of employee disruptions, changes in investing offsetting the impacts of disasters, and cost-cutting mitigating the impacts of employee disruptions. Our research emphasizes the importance of taking a broader view of firm failure beyond the
Confusion matrix (balanced accuracy = 86.6%).
Actual Surviving 15,119 3178 Failed 4 38 current deterministic and voluntaristic approaches, and supports an integrative view that considers the complex interplay between external factors and management actions.
For manufacturing, construction, and retail firms making or distributing physical products, one of the most difficult operational challenges recently has been addressing supply chain bottlenecks. We now see that disruptions of the supply chain can lead to serious work stoppages and revenue losses in many industries. Our analysis shows that the most effective response to supply disruptions is new product development.
Finally, our analysis shows that debt service costs can play a significant role in firm failure. Firms that want to survive may need to maintain a greater cushion of cash reserves to survive business cycles. That said, investment in new technology and equipment may be needed to meet the challenges of supply chain in-sourcing and to meet new competition, and financing for these investments must be adequately planned.
Our research paper makes a significant contribution to the field of business failure by addressing the current theories surrounding its causes. In addition to reviewing and discussing these theories, we conduct a systematic examination of their various
Coefficients of independent variables. components, providing evidence to determine their validity. We employ logistic regression in terms of a machine learning perspective and less commonly used visualization techniques to assist the interpretation of our findings, thereby contributing to the academic community. Furthermore, our results have practical applications for practitioners, who have traditionally relied on simplistic bankruptcy prediction models that fail to consider external factors or the actions and skills of management. Our research can provide valuable insights for external creditors, investors, and managers tasked with guiding firms through challenging times.
The perennial question of how firms can improve their resilience to survive significant economic shocks, as well as ongoing
Permutation importance of independent variables.
Dendrogram -Hierarchical clustering of independent variables. competitive pressures and challenges, is a key issue that deserves ongoing attention. Our analysis is limited to US-based manufacturing firms. Although this is a large and important sector of the economy, further research on the causes of business failure across more industry sectors may provide additional insights that our data set did not reveal. We also measured management actions through rhetoric in the 10K filings. However, it is widely known that these annual filings contain a lot of <boilerplate= language, and that management has incentives to describe the company and management's actions in the best possible light. Though beyond the scope of this study, there are other sources that might provide additional or more accurate insights about management actions, including quarterly earnings calls, analyst reports, social media, and other sources. Although our study covers financials through the end of 2020, the data for 2021 and 2022, which might reveal more about the impact of the COVID-19 pandemic on business failure, is not yet available. As increasing amounts of data become available and more sophisticated machine learning techniques can be applied, the exploration of the complex mechanisms involved in business failure merits ongoing examination.
Journal of CorporateAccounting & Finance, 2025
For example, manufactured goods exports account for 81.83% of total good exports, see https:// www.nam.org/state-manufacturing-data/2021-united-states-manufacturing-facts/.
Our 10-K filings are obtained from the Notre Dame Software Repository for Accounting and Finance (SRAF) https://sraf.nd.edu.