Predicting Bankruptcy in Wholesale, Retail, and Motor Vehicle Repair: An AI-ML Perspective

3.1 Data

The list of (241) bankrupt companies in the “wholesale trade, retail trade, and repair of motor” sector is sourced from the Insolvency and Bankruptcy Board of India. The aim of the study is to predict or label a company as bankrupt or otherwise given the financial data of the company. This fits the description of the classification problem, which can be addressed using AI-ML models (Pompe and Feelders, 1997). However, to deploy AI-ML models, the training dataset should contain adequate representations from both positive and negative classes, viz., bankrupt and non-bankrupt companies. The efficacy of AI-ML models to predict bankruptcy risks in the trade services sector a sample comprising 5527 firms from wholesale trade, retail trade, and repair of motor vehicles is considered due to data availability (Table 1). Of these 5527 firms, 241 were bankrupt. Hence, to achieve a balanced dataset, SMOTE technique is used to create an oversample dataset comprising 5286 functional and 5286 bankrupt firms.

The share of bankrupt to non-bankrupt companies is around 50:50 resulting in a dataset that is balanced on both positive and negative classes. This addresses the class imbalance issue which affects the efficacy and accuracy of the AI-ML models dealing with the classification problem⁵. The set of explanatory variables used in this study are given Table 2, they include firm level financial variables and ratios drawn from similar studies in the domain of bankruptcy prediction. Further, we have also tried to predict the bankruptcy in the trade sector by dividing the sample on the basis of different business rules (liquidity, profitability, and firm asset size).

3.2 Methodology - AI-ML models

Generally using the labelled data, the supervisory machine learning models extracts the patterns from the training dataset. Further, supervisory machine learning is divided into two categories of algorithms: regression based and classification-based approaches. Regression based supervisory machine learning approach is used to predict the continuous variables whereas the classification based supervisory machine learning approach is used to predict the dichotomous/categorical variables. In this study, our interest variable is dichotomous in nature. Therefore, we focus on the classification type of supervisory machine learning approaches which are very popular in the bankruptcy prediction literature are Logistic Regression (LR), Random Forests (RF), Naïve Baye (NB), Gradient Boosting (GB), Support Vector Machines (SVM), K-Nearest Neighbours (KNN), Decision Trees (DT), and one popular artificial intelligence technique such as Artificial Neural Networks (ANN or NN). Though, many algorithms are available within the supervisory machine learning category, we employ the aforementioned 8 AI-ML techniques on the basis of their popularity in the bankruptcy prediction literature and their explainability, training and prediction speed and ease of implementation.

Further, in this study, we have chosen to test the efficiency of these 8 AI-ML models, slightly departing from the practice adopted in the literature, wherein the focus is on using a single technique or a couple of techniques. Alaka et al. (2018) did a systematic review of 49 articles for the use of AI-ML models for bankruptcy prediction. The authors note that, of the 49 studies under review, only 30 studies compared the performance of the bankruptcy predictions by the AI-ML models. Further, a few techniques viz., Support Vector Machines, Artificial Neural Networks, are compared more often than others (ibid). Contrary to this, the present study has consistently used the 8 AI-ML models viz., LR, RF, NB, GB, SVM, KNN, DT, and NNs, for bankruptcy predictions in the Indian case. The literature on the use of AI-ML models for classification problems in general and bankruptcy predictions in particular clearly underscores that no single model outperforms others (Kumar and Ravi, 2007; Alaka et al., 2018; and Tanaka et al., 2019). Specifically, with reference to the bankruptcy prediction the performance of the models is found to be influenced by sample size, multicollinearity, underlying statistical distributions, computational ability etc.

All the aforementioned models have relative strengths and weaknesses stemming from the underlying data and model requirements. Hence, to alleviate the issue relating to data all the models are tested on a single sample to compare the relative performance. While the chosen sample may be inherently favourable for certain models, given the fact that all models face similar training and testing conditions, the results can be fairly compared. Furthermore, the AI-ML models inherently present a trade-off between the result accuracy and transparency, with models like LR and DT offering better transparency than SVM and NN which have higher accuracy. Hence, to be agnostic to the choice between transparency and accuracy, the analytical framework of this study presents the performance metrics for the chosen 8 AI-ML models coherently, leaving the researcher or practitioner to make his or her choice based on the use-case at hand. Given the widespread use of these models in the literature, we omit technical details for the sake of brevity.

3.3 Performance metrics

Literature establishes that the performance of classification models is evaluated through the construction of confusion matrices (Kuhn and Johnson, 2013). These matrices are a cross tabulation of number of actual cases and predicted cases as given below. In general, the positive class refers to the variable of interest. In this case “crisis” period is a positive class, with “non-crisis” period being a negative class. The confusion matrices (Table 3) are then used for computing metrics that enable comparison of the model performance.

Accuracy is the primary metric for assessing AI-ML model performance in classification problems, representing the ratio of correct predictions to total instances. However, it doesn’t account for misclassification errors. To address this, metrics like precision, sensitivity (recall), and specificity are used in Table 4. Precision measures the rate of true positive predictions out of all positive predictions, indicating the model’s ability to avoid false positives. Sensitivity (recall) captures the rate of true positive predictions out of all actual positives, indicating the model’s ability to identify positive instances accurately. The F1-score, the harmonic mean of precision and recall, balances these errors. Specificity measures the rate of true negative predictions out of all actual negatives, akin to sensitivity but for negative instances. AUROC (Area Under Receiver Operating Characteristic Curve) assesses the model’s accuracy in distinguishing between positive and negative classes by plotting sensitivity against 1-specificity. A higher AUROC indicates better model performance. These metrics provide a comprehensive evaluation of AI-ML models beyond simple accuracy.

3.4 Application of business rules business overlay

Although AI-ML models provide very good accuracy rates as compared to traditional econometric models, they fail to provide a convincing causative link between explanatory variables and the predicted variables. One of the key concerns using AI-ML models is that the models often function as a black box, wherein only inputs and outputs are visible to the user (Guidotti et al., 2018). While some AI-ML models do provide some guidance regarding causation they fall short of establishing a formal relationship between explanatory and predicted variables (Freitas, 2014). To this end, to improve the explainability of the models used, this study applies business rules to add context to the predictions made by the AI-ML models. Though this falls short of providing a definitive causal link, it can provide direction of likely impact on the predicted variable given the business rules. Also, financial regulators often stipulate dispensations to mitigate stressed firms to avoid bankruptcy or failure based on differential criteria regarding asset size, profitability, and liquidity positions etc. This allows the regulators to ensure that benefits of such dispensations are utilized by genuine firms under stress and avoid a one-size fits all approach (RBI, 2020, 2023)⁶. Hence, these business rules are framed using the conventional credit risk or investment analysis used by banks and fund houses for selecting or monitoring their investments. The study uses the following three business rules based on liquidity, profitability, and asset size position of the firms.

3.4.1 Liquidity based business rules

One of the early warning signs about financial distress in a firm is mismanagement of liquidity, often resulting in default and distress precipitating in bankruptcies. Hence, bankers traditionally stipulate minimum levels of liquidity parameters to be achieved or maintained by the firms to get credit facilities. To illustrate, firms should have quick and current ratios of minimum 1.00 and 1.33 respectively, which signals that the current assets of the firm are adequately covering the current liabilities (Venkatachalam and Natarajan, 2015). Therefore, bankers and investors are more likely to monitor such liquidity ratios and form an opinion about the firm’s financial health. Hence, the sample data is bifurcated into two sets (A and B) using the liquidity thresholds mentioned above. The firms with quick and current ratio above 1.00 and 1.33 are categorized as firms with healthy liquidity, while those below the liquidity thresholds are categorized as firms with liquidity issues. Subsequently, the AI-ML models are run on samples A (healthy liquidity firms) and B (weak liquidity firms) after removing the liquidity ratios from the explanatory variables. Such an assessment primarily considers the liquidity parameters which are key to decisioning by the banks and investors and then looks at the risks of bankruptcy.

3.4.2 Profitability based business rules

Like liquidity ratios, another key early warning indicator that banks and investors look out for monitoring firms is their profitability. Generally, bankers and investors approach firms that are profit making differently from those that are incurring losses in terms of investment strategy. Hence, the sample dataset is bifurcated in two sub-sets (A and B) based on the profitability of the firms viz., profit making (ROA being positive) and loss making (ROA being negative). Subsequently, the AI-ML models are run on samples A (profit making firms) and B (loss making firms) to look out for risks of bankruptcy, beyond profitability.

3.4.3 Asset size based business rules

Notwithstanding the profitability and liquidity status of the companies, another key decision parameter considered by banks and investors is the size of the firm i.e., total assets. The selection and application of credit risk techniques vary depending on the size of the firm. Small firms may be highly vulnerable to macro-economic shocks and pose high risks, while large firms can better withstand such risks, their failure can have very high costs for the banker or investor. Also, in the event of bankruptcy, for larger firms it may take longer to realize the fair value of the stranded assets than compared to smaller firms. Hence, it might be rational for a banker or investor to differentially approach the risk posed by small and large firms. Accordingly, the sample is bifurcated into four categories viz., A, B, C and D based on the asset size of the companies as given Table 5 below. Subsequently, the AI-ML models are run on samples A to D to assess the performance of models and explore the role of various explanatory variables on signalling bankruptcy of firms across asset size categories.

Applying AI-ML models on the bifurcated datasets based on business rules can provides three insights on predicting bankruptcies among manufacturing companies in India. First, it allows the researchers to assess the performance of AI-MLs models of the bifurcated datasets based on business rules and identify the best performing models for each sub-segment. Second, it can identify the key variables to signal the bankruptcy risks beyond the specified business criteria viz., liquidity, profitability, and asset size. Third, it enables discerning not so good companies from the companies that are seemingly good companies on the specified criteria. From an investment risk analysis standpoint, such decision-making insights can be very useful to protect investor interest. The analysis not only offers sharper insights on bankruptcy risks within good performing companies with healthy liquidity and profitability, but also provides a list of variables with high IV values to monitor for picking up the bankruptcy signals. This facilitates investor to apply a differentiated approach to assessing risk across firm categories and better understand business models and risks emanating from them⁷.

3.5 Class imbalance – SMOTE technique

For better bankruptcy prediction, it is crucial that the dataset used for AI-ML models is balanced between positive (bankrupt firms) and negative (non-bankrupt firms) classes. A skewed dataset can lead to higher error rates, as the model may not learn adequately about both classes. While segmenting samples based on business rules offers decision-making insights, it can inadvertently create unbalanced datasets, impacting model performance. To address this, the study employs the Synthetic Minority Oversampling Technique (SMOTE) to generate a balanced dataset (Kim et al., 2015; Le et al., 2018; Veganzones and Séverin, 2018; Ghatasheh et al., 2020; Smiti and Soui, 2020; Tumpach et al., 2020; Alam et al., 2021; Garcia, 2022; Papíková and Papík, 2022; Amirshahi and Lahmiri, 2024). SMOTE, a widely used data preprocessing method, corrects class imbalances by creating synthetic examples from the minority class based on the feature space rather than the data space (Fernández et al., 2018). This ensures the AI-ML models are trained on a balanced dataset, improving their predictive accuracy. The study applies AI-ML models to datasets segmented by business rules and balanced using SMOTE, enhancing the model’s ability to predict bankruptcy accurately.

4.1 Performance of AI-ML models on full sample

The efficacy of AI-ML models to predict bankruptcy risks in trade services sector a sample of comprising 5527 firms from wholesale trade, retail trade, and repair of motor vehicle is considered. Of these 5527 firms, 241 were bankrupt. Hence, to achieve a balanced dataset, SMOTE technique is used to create an oversample dataset comprising 5286 functional and 5286 bankrupt firms. Foremost, the correlation matrix of the given in Figure 1 indicates that sparse correlation amongst the explanatory variables, underscoring their usefulness in signalling bankruptcy risks. Subsequently, like in the case of manufacturing and construction firms, we deploy the same 8 AI-ML models on the full sample and followed by the testing on the sub-samples which are bifurcated on the basis of liquidity, profitability, asset size business rules. Further, the testing of AI-ML models in this chapter follows the same methodologies adopted for manufacturing and construction firms. Hence, for brevity, the extended discussions on model performance are not presented in this chapter. The focus is limited to identify key models and the set of explanatory variables with high IV values and the results are presented hereunder.

Figure 1. Correlation matrix of numeric features (trade services sector firms).

The performance metrics of the AI-ML models tested on the full sample of firms is given the Table 6. The models boast an average accuracy of around 80 per cent indicating that AI-ML models can be used for predicting bankruptcy risks in the trade services sector. Also, the AUROC scores of models is around 0.83 indicating reasonable discriminatory power of the models. Further, as compared to the performance of AI-ML models for manufacturing and construction firms, the accuracy and discriminatory power of the models is lower in case of trade service firms. However, the performance of random forest and neural network models in case of trade service firms stands out compared to other models. Other models like gradient boosting, k-nearest neighbours, decision trees also register decent performance levels, next only to random forest and neural network models in this sector. Furthermore, the usefulness of the models in discerning both the functional and bankrupt firms is also good. The F1-score for random forest model is 0.96 while that of the neural network model is 0.88 indicating balanced prediction performance.

On the basis the information value (IV) or weight of evidence which are available in Table 7, for trade services firms, the top-5 variables with high IV are interest coverage (PBIT_INT), return on assets (ROA), debt to total assets (D_TA), profit to total assets (PAT_TA), and working capital to total assets (WC_TA). The explanatory variables wise, the information value is provided in Table 7 below. The indicators with high IV can perform the role of early warning indicators as they contain relatively higher information about the impending bankruptcy risks than other explanatory variables. Interestingly for trade service firms, the working capital to total assets is a key variable with high IV value. Working capital is more relevant for trading firms as they depend on stock in trade and try to optimize creditors and debtors to maximize their revenues. A typical trade service firm can be thought of as moving stocks in trade, purchasing and/or selling on credit. Such mismatches in sale realizations may necessitate higher working capital requirements.

4.2 Performance of AI-ML models on bifurcated sample - Business rules

As observed in case of manufacturing and construction firms, deployment of AI-ML models on the sub-samples bifurcated based on business rules (viz., liquidity, profitability, and asset size) yield interesting insights. Specifically, in terms of IVs of variables, the sub-samples have revealed differential relative impact of explanatory variables to signal bankruptcy risks. Hence, a similar exercise is carried out for the trade service firms. The overall sample is bifurcated into sub-samples using liquidity, profitability, and asset sized based business rules. Further, using SMOTE technique, the sub-samples are balanced. The business rule wise performance metrics of the AI-ML models is presented hereunder.

4.2.1 Performance of AI-ML models on bifurcated sample (liquidity ratios)

The performance metrics of the AI-ML models on the sub-samples created using liquidity-based business rules are given in Table 8 and Table 9. The accuracy rates of some of the AI-ML models viz., random forest, gradient boosting, neural network in predicting bankruptcy risks for both firms with and without liquidity problems are above 85 percent. Also, the AUROC scores of these models are above 0.90 indicating strong discriminatory power of the models. The F1-scores of the models are also around 0.90 suggesting a balanced performance of the models. Interestingly, the IVs of the explanatory variables for the companies with and without liquidity issues vary divergently. For companies without liquidity issues, the profit margin, total loans to total assets, asset turnover, debt to total assets, cash flows to sales are the top 5 explanatory variables with higher IV values in Table 10.

4.2.2 Performance of AI-ML models on bifurcated sample (profitability ratios)

The performance metrics of the AI-ML models on the sub-samples created using profitability-based business rules are given in Table 11 and Table 12. As can be observed from the performance metrics, the average accuracy of the AI-ML models on the sub-samples for profit making and loss-making companies is like that of the overall sample. The average accuracy of AI-MLs for profit making companies is around 84 per cent and for loss making companies the average accuracy is 77 percent. The average AUROC scores of the models are 0.90 for profit making companies and 0.83 for loss making companies. Indicating that AI-ML models have higher discriminatory power to discern bankrupt firms from functional firms in case of profit-making firms than in case of loss-making firms. Also, the average F1-scores of the models follow similar trends between profit- and loss-making firms. Overall, the performance metrics indicate that AI-ML models are performing better in case of profit-making trade service firms than in case of loss-making firms. Notwithstanding the above, the performance metrics of AI-ML models in this case i.e., profitability-based bifurcation is either comparable or better than the levels registered for the overall sample.

The IVs of the explanatory variables for the sub-samples based on profitability business rules is given in Table 13. As in the case of liquidity-based bifurcation, in this case too, the variables with high IVs vary for both the profit making and loss-making firms as compared to the overall sample. As observed earlier, for the overall sample, profit margin, return on asset, debt to total asset have highest information content in signalling bankruptcy. Followed by profit after tax to total asset and working capital to total asset. However, for the profit-making firms, the current ratio, interest margin, profit margin, profit to total assets, and working capital to total assets are the top 5 variables with highest IVs. Also, for the loss-making firms, the growth in retained profit, interest coverage, profit before interest and taxes to sales, revenue growth, and profit to total assets are the top 5 variables with highest IVs. It is interesting to note the differences between the set of top 5 variables for the profit- and loss-making firms and with that of the overall sample. While interest coverage and profit to total assets figure out as variables with high IVs for the overall sample, they also figure out in case of both profit- and loss-making firms. Thus, bifurcating the overall sample into sub-samples level provides useful insights.

4.2.3 Performance of AI-ML models on bifurcated sample (Asset size of the company)

The overall sample is bifurcated into 4 sub-samples based on the asset size of the firms. This enables analysis of the performance of AI-ML models and to glean the relative importance of explanatory variables using IVs in signalling bankruptcy risks across firm sizes. Comparatively the average firm size of trade services firms is lower than that of the manufacturing or construction firms⁸. The performance metrics of the AI-ML models is given in Table 14, Table 15, Table 16, and Table 17. The IVs of the explanatory variables for the four sub-samples are given in Table 18. From the performance metrics, it can be observed that the average accuracy rate of AI-ML models for categories A, B, C and D companies are at 83 percent, 79 percent, 81 percent, and 84 percent respectively, which is greater than the accuracy rate of 80 percent achieved for the overall sample. Likewise, the AUROC scores for the AI-ML models for category A, B, C and D companies are at 0.90, 0.86, 0.87, and 0.88 respectively as compared to AUROC score of 0.84 achieved for the overall sample. This represents an adequate discriminatory power for the models. Further, across categories of companies, random forest model has achieved accuracy rates of 94 percent to 98 percent and the AUROC scores range from 0.99 to 1.00. Thus, outperforming all other models across categories. Furthermore, neural networks have a high accuracy rate in the case of category D companies.

The analysis of IV of explanatory variables across asset size categories of companies reveals interesting insights in the Table 19. At the overall sample level, interest coverage, return on asset, debt to total asset is seen to be the foremost variables with high IV values. Among these variables, profit to total assets is among the top 5 variables with high IV across all firm types. This is followed by ROA which figures in the top 5 variables for firms in categories B, C, and D, while interest coverage is important for firms in categories A, B, and C. In contrast, debt to total assets is among the top 5 variables only in case of firms in categories B and C. For larger firms (A, B), profit margin is more relevant. While for smaller firms in category D, revenue growth along with profit margin are relevant. Variables like profit before interest and taxes to sales, and current ratio also among the top 5 variables with high IV values. Though there are common variables possessing high IVs both at the overall sample and bifurcated sample, it may be prudent for the investor to adopt a segmented approach to capture the bankruptcy risks in an efficient manner.