Optimized Deep Learning Model to Predict Business Bankruptcy

Companies find themselves using business tools and strategies with emerging technology in a more recurrent way in a very aggressive business environment. However, under these conditions, the generation of value from products and services is the greatest aspiration of the business. Likewise, one of the main strategies applied is the forecasting of situations and events that could compromise subsistence or growth activities. Therefore, it is important to develop and implement technological strategies in crucial activities that determine the reason for the business. In this context, artificial intelligence is a science that has been making proposals and initiatives to companies that are mainly in subsistence or that require greater efficiency and optimization of their processes. The results required by a company are aligned with institutional goals and strategic plans.

Alsheref et al. [1] considered that companies have accurate and critical targeting activities due to lack of consistency in data and misinformation about financial aspects. As a result, micro, small and medium-sized enterprises often do not use traditional financial channels. In this sense, financial difficulties have a negative impact on the development of business health and sustainability of these companies. Similarly, Brenes et al. [2] note that difficulties in credit evaluations for companies lead to higher interest rates and distort the nature of financing. Additionally, most of these companies are at the subsistence stage and are prone to business bankruptcy. In this sense, they cannot access financial services, and an appropriate credit evaluation. For this reason, at the macroeconomic level, most startups do not constitute the important engine for local economic growth, especially when facing crises and possible bankruptcies.

Likewise, financial institutions play an important social role in facilitating access to credit and facing the risks associated with lending money. However, there are several difficulties in financing companies and identifying companies that are at risk of bankruptcy. Credit rating and bankruptcy models have been used by financial institutions to assess creditworthiness and identify potential companies that will lose their status as activity [2]. Companies have difficulties in accessing financing and opening up to other markets, motivating them to generate subsistence activities, which over time can be considered a potential bankruptcy [3].

Calabrese [4] defines the credit risk associated with company loan financing as the probability of defaulting on derivative loans, which can lead to a negative credit rating. This rating, in turn, reflects the potential for business bankruptcy. In this sense, data from financial statements are presented to predict the credit and bankruptcy risk of companies. However, behavioral and performance data allow them to qualify for loans through covenants, but once the loan is obtained, they are likely to be unable to meet the payment obligation. According to the study by Cantarero et al. [5], companies that have been in operation for years and those that are new (some with only a few months of start-up) face a highly competitive, complex, ambiguous, volatile and uncertain market, which is not only local or national, but is directly related to the international market through products, services or price.

Tran et al. [6] emphasize that performance measurement is fundamental in a control system and argue that it is essential to measure the financial aspects of companies both before and after they obtain a loan. They also consider that the rating of financial performance originates the profitability of a company and its continuity in the long term. Muthukumaran and Hariharanath [7] observe that both established companies and new enterprises (some only months old) encounter a highly competitive, complex, ambiguous, volatile, and uncertain market. This challenging environment extends beyond local or national boundaries, directly connecting to the international market through products, services, and pricing.

In this sense, SMEs without financial support cannot invest, running out of cash and ending up in bankruptcy. This result has a significant impact on the entire market chain and financial institutions because it misses the opportunity to obtain a profitable return on investment. For these reasons, the prediction of the risk of company bankruptcy is necessary and essential at all stages. Therefore, it is necessary to develop a predictive learning model to forecast the bankruptcy of companies.

Machine learning models have various algorithms that allow the prediction and forecasting of activities with transcendence for the company. However, treatment with relevant criteria, suitable variables, and consistent data is essential [4]. Several studies require models of credit rating, payment yield and business bankruptcy that provide techniques and determining factors through specific patterns in order to generate predictions according to the selection criteria [8]. Credit rating models for SMEs do not reflect the ability to pay in a timely manner towards an opportune horizon for decision-making. The global financial crisis has shown the importance of having a model to predict default [4].

This article presents an optimized machine learning model to determine the bankruptcy of companies, where various bankruptcy models are required with relevant indicators according to the industry and institutional segment. Likewise, the analysis method according to the optimized model, the results obtained and the conclusions are presented.

3.1 Method

To obtain data, an instrument consisting of 20 questions classified according to the corresponding dimension was applied: Business Environment (10) and Financial Situation (10), these questions were elaborated according to the analysis of the business context. The instrument is specified in Table 1, Likewise, the observation technique was used with the instrument of record cards to compile the accuracy of each algorithm to be used.

The exploratory data analysis method was used and was complemented with the model proposed by the authors [1], which determined three stages: data exploration, model selection and implementation. In this sense, 21 variables were determined to be considered for the model, where it was identified that the relevant variables are Bankruptcy, Capital, Letters_to_pay, Access_Credit, Satisfaction_E and Cost effectiveness (Profitability). The 21 variables and types are specified in Table 2.

Table 1. Instrument (survey) for data collection

Table 2. Variables and type

3.2 Data exploration

The study population is determined by the companies in the city of Metropolitan Lima. Which is specified according to the National Institute of Statistics and Informatics - INEI in the technical report on business demographics in Peru IV quarter of 2023, which has as its source the Central Directory of Companies and Establishments (DCEE). In that sense, there are 26,330 companies registered as of the IV quarter of 2023 in Metropolitan Lima. The study sample was determined in 379 companies in the city of Lima, Peru (microenterprises, small companies and medium-sized companies). The sample size has been calculated with a margin of error of 5% and a confidence level of 95% using the finite universe sample calculation formula. Likewise, through an exploration process, 118 records were identified that correspond to microenterprises, 166 to small companies and 95 to medium-sized companies. Table 3 considers the type of company.

Table 3. Total de companies

1.png

Figure 1. Time series and correlations (Payment_Performance with Satisfaction_E and Months_A)

An exploratory data analysis was applied to obtain useful factors about business bankruptcy and identify the main causes of it. Likewise, graphs, time series analysis and correlations were used. Nicodemo and Satorra [13] argue that this type of analysis is advantageous because it focuses on essential aspects of a company's context and utilizes important data to generate forecasts. It also allows incorporating new variables for improving or optimizing the model.

Graphs, time series analysis, and correlations were used. This type of analysis allows us to demonstrate that the data patterns are useful to generate a forecast on the effects of the bankruptcy effect of companies, where identifying the factors is very decisive. Figure 1 states that it has been identified that the variables Performance_Payment and Satisfaction_E are correlated with the variable Months_A, so studying these correlations is necessary to understand the characteristics and impact. Correlations were obtained to identify variables that directly impact business bankruptcy.

Data analysis was used in distributions and line graphs of the variables: Bankruptcy, Months_A, Capital, Plans, If, Qualification, Capacity_E, Satisfaction_E, Access_Credit, Access_markets, Personnel, Bills_To_Pay, Companies, Performance_Payment, Competitors_D, Unforeseen, Claims, Profit_rate, Profitability, Efficiency and Insolvency, where levels of predominance were identified that require correlation towards each resulting factor.

In addition to this, it was obtained that the study variables have significant attributes for the identification of specific characteristics of bankruptcy. In this way, the bankruptcy forecast can be based not only on the general context of the company but also on the internal context of the processes. It is important that the data set used has variables with a specific level of disaggregation and that it allows relationships to be generated with other variables. To guarantee the effectiveness of the chosen model, it is necessary to consider two essential criteria: 1. The characteristics of the data set, and 2. The correlations between the variables. Figure 2 specifies the characteristics of each variable, where variables that represent greater relevance for the context of business bankruptcy are presented. Likewise, it is important to identify the variables that will be worked on in correlations using the defined criteria.

3.3 Model selection

Pratt et al. [14] argue that machine learning algorithms, such as logistic regression, random forests, and neural networks, are well-suited for predicting outcomes using a dataset containing variables that may be correlated. These algorithms are particularly effective in addressing risk identification problems within a specific and detailed context. In the same way, Pandey et al. [15] consider that in order to achieve solid metrics, it is important to implement a machine learning model with characteristics of the context analyzed and that can be aligned with institutional goals. Sulastri and Janssen [16] also stated that determining the metrics to evaluate the results of an algorithm corresponds to having an exact understanding of the problem in its context and clear objectives. In this context, Yahia et al. [17] consider that in order to obtain better results it is necessary to generate optimization processes that include the incorporation of more suitable variables and correlations between them.

Raza et al. [18] highlight that logistic regression and random forest algorithms are particularly well-suited for incorporating dynamic market information and variables related to the business context. This means that they are operational and suitable for estimating investment probabilities and forecasts of various types, in such a way that data sources are open to any information related to the probability of bankruptcy and potential for loss. For example: stock market information, particularities of the sector in which the activity is carried out, regulation, ease of access to credit, business bankruptcy, etc. In the same way, Muthukumaran and Hariharanath [7] state that deep learning is particularly effective in applications involving Small and Medium-sized Enterprises (SMEs), thanks to its versatility and capability to facilitate numerous training iterations. The main objective of the technique is to determine the financial status of SMEs, which contains the design of the selection of characteristics based on the optimization algorithm and neural network, which is used for data classification. Metrics are necessary and play an important role in defining financial firms' profitability and productivity.

To select a machine learning model, the objectives of the study must be aligned with the activities to be developed. It is also important to have the detail of the general and specific context variables. In this sense, for the selection of the model, application tests were carried out on the dataset with 3 machine learning models that respond to the objectives of the study. These are: Logistic regression, random forests and neural networks.

Calabrese [4] points out that logistic regression is useful for carrying out analysis activities and prediction processes that allow reducing costs that mean a setback or delay, thereby improving and generating efficiency. Companies to obtain better profitability results and business bankruptcy forecasts require knowing variables and patterns appropriate to their context. They also consider that the essential characteristics of this algorithm are: Identify the question, collect historical data and train the model. Zhang et al. [12] specify that they are important because the structure considers various trees with decision aspects. Likewise, the ease of application and flexibility has been successfully implemented, as it handles classification and regression problems at a general and specific level. Likewise, decision trees tend to have better power in the result when the variables are correlated. Muñoz-Cancino et al. [8] consider that neural networks are ideal for solutions that generate forecasts because they generate correlations from the identification of variables, which are necessary to establish focus points to achieve effective training and generate the precise learning to obtain ideal results. In this sense, neural networks consider training data to learn and promote improved accuracy over time. In this context, the proposed optimization model considers that training is an indispensable process to improve the results of the metrics.

2.jpg

Figure 2. The characteristics of the data sets

Table 4. Algorithm comparison

3.png

Own elaboration

Figure 3. Optimized model architecture

Table 5. Metrics

Various algorithms were compared for the treatment of the specific problem of prediction. The comparison is specified in Table 4.

The metrics were chosen according to the experience generated with research resulting from the similar context on the predictions of a particular problem. This choice of metrics is specified in Table 5. Likewise, it is specified that the chosen algorithms respond to the score generated after the evaluation carried out in Table 4, which is aligned with the background of the topic addressed. It was identified that the neural network is the algorithm that has the highest score and that it is in 3 antecedents on prediction, then the algorithms of random forests and logistic regression have 1 and 2 antecedents of prediction respectively, and 4 of punctuation. The background information chosen is the product of the research carried out as a state of the art, which considers references in contexts similar to the subject treated with similar metrics.

3.4 Model optimization

The study variables have significant attributes for the identification of specific characteristics of bankruptcy. In this way, the bankruptcy forecast can be based not only on the general context of the company but also on the internal context of the processes. Sulastri and Janssen [16] consider that generating suitable variables is essential to generate optimal results, thereby ensuring adequate prediction and performance. It is important that the dataset used has variables with a specific level of disaggregation and that it allows generating relationships with other variables. To guarantee the effectiveness of the chosen model, it is necessary to consider two essential criteria: 1. The characteristics of the data set, and 2. The correlations between the variables.

The optimized model was developed based on the model proposed by Alsheref et al. [1] who determined three stages: data exploration, algorithm selection and implementation, in this way it was possible to identify the best machine learning algorithm. In this context, an additional stage was incorporated: Optimization of the model, which included the identification of factors (relevant variables) and the correlations between them. Likewise, it gave greater relevance to the training process, where it was considered as a key step for the optimization of the values of the selected algorithm.

In this sense, 21 variables that were worked on in the data exploration stage were considered. For the algorithm selection stage, 3 algorithms were considered, where the neural network algorithm was the one that obtained results with the highest percentages in the established metrics. For the model optimization stage, it was identified that the relevant variables were: Performance_Payment, Capital, Letters_for_pay, Credit_Access, Satisfaction_E and Profitability. The proposed model considered exploratory data analysis to obtain useful factors on business bankruptcy and identify the main causes of bankruptcy. Likewise, graphs, time series analysis and correlations were used. This type of analysis allows us to demonstrate that data patterns are useful to generate a forecast on the effects of the bankruptcy effect of companies, where the identification of factors is very decisive. It has been identified that the Performance_Payment and Satisfaction_E variables are correlated with the Meses_A variable, so the study of these correlations is necessary to know characteristics and impact. Correlations were obtained to identify variables that have a direct impact on business bankruptcy. Figure 3 indicates the architecture of the optimized model.

3.5 Implementation

Yao et al. [19] indicate that the implementation stage must specify the choice of the algorithm that has better metrics and that is adapted to the requirements of the business context according to the recorded data. Likewise, Zou et al. [20] consider that the neural network model meets the necessary characteristics to achieve the proposed objectives to generate predictions on a large number of variables and data. In this sense, correlations were established between the relevant variables that have been necessary to achieve better results. To achieve optimal results with optimization characteristics, it is necessary that the selected algorithm has obtained results greater than the results of the other algorithms subject to evaluation. Likewise, it is essential to consider that the training process is a key factor for the learning of the algorithm; all of this will guarantee satisfactory, efficient and truthful results.

It is specified that the model optimization neural network algorithm exceeded the metrics of the other algorithms considered in the present research where an accuracy of 97%, accuracy of 97%, sensitivity of 99% and f1-score of 98% was obtained. It is also specified that these results are similar to the research carried out by Calabrese [4] that in the study: "Contagion effects of UK small business failures: A spatial hierarchical autoregressive model for binary data", it obtained 4 main metrics: precision (85%), accuracy (90%), sensitivity (92%) and f1-score (89%). These metrics made it possible to assess credit and bankruptcy risk, which was designed specifically for SMEs. Similarly, results similar to the study: "Deep Learning-Based Corporate Performance Prediction Model Considering Technical Capability" by the authors of the authors were obtained [21], who proposed a model based on an artificial neural network that predicts stock prices using economic and financial indices as predictors. They considered 4 main metrics: precision (94.5%), accuracy (92.7%), sensitivity (93.5%) and f1-score (94%). They keep similar results with the aim of predicting the financial performance of companies using an artificial neural network with backpropagation algorithms.

Regarding the general objective of the research, it was necessary to develop a machine learning model to improve the prediction of company bankruptcy, which initially through the 3 stages proposed by Alsheref et al. [1]: data exploration, algorithm selection and implementation, after which an additional stage was incorporated: model optimization. This allowed the results of the metrics to improve. These results have a similar behavior with what was specified in the study: "Credit risk prediction of SMEs in supply chain finance by fusing demographic and behavioral data", proposed by Zhang et al. [12] who specify the importance of the financial index and transaction data to predict the credit and bankruptcy risk of SMEs with an improved random forest model, where they considered the improvement of the model based on the respective training, for this reason they obtained the metrics: precision (88%), accuracy (89%), sensitivity (90%) and f1-score (88%). Similar results were presented in the study: "Explainable Machine Learning in Credit Risk Management", proposed by Bussmann et al. [3], which requires the treatment of five artificial neural network models for the prediction of credit risk, bankruptcy risk and payment yield. It was shown that the deep neural network performs more accurately in predicting credit risk than other models. The proposal was a deep learning model to support decision-making in financial risk management, and experiments with trade data. They considered 4 main metrics: precision (94%), accuracy (94%), sensitivity (93%) and f1-score (94%). This study demonstrates the superiority of deep learning over other machine learning and rule-based techniques.

The optimized model has considered representative variables according to the context of the company, all of which has allowed us to improve the results. The results obtained have allowed us to know and predict business bankruptcy based on specific and representative variables assumed in the company based on its context. It is important that future work is oriented towards the identification or incorporation of new variables that improve the training process, with respect to the number of events that require learning. All of this will allow obtaining results with greater accuracy, precision and less time.

Companies have a large percentage of money and workforce losses in formalization, continuity, subsistence and credit search processes because not all companies continue in operation after 6 or 12 months. Likewise, business bankruptcy, motivated by various factors, is very significant in small and medium-sized companies. Likewise, it is important to specify that the optimized model has obtained results with great solvency and with this, variables are presented that the company needs to know to strengthen its own critical processes. In this sense, the optimized model has made it possible to identify the variables that generate the possible bankruptcy of the company and the subsequent treatment.