Next Article in Journal
Analysis of a Signalized Intersection with Dynamic Use of the Left-Turn Lane for Opposite through Traffic
Next Article in Special Issue
Determining the Increase in a Building’s Appreciation Rate Due to a Reconstruction
Previous Article in Journal
Assessing Climate Change Impacts and Adaptation Options for Farm Performance Using Bio-Economic Models in Southwestern France
Previous Article in Special Issue
Financial Analyses and Corporate Evaluation on Sustainable Ability to Generate Excess Profit
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:

Bankruptcy or Success? The Effective Prediction of a Company’s Financial Development Using LSTM

Institute of Technology and Business in České Budějovice, School of Expertness and Valuation, Okružní 517/10, 370 01 České Budějovice, Czech Republic
Author to whom correspondence should be addressed.
Sustainability 2020, 12(18), 7529;
Submission received: 20 August 2020 / Revised: 8 September 2020 / Accepted: 11 September 2020 / Published: 12 September 2020
(This article belongs to the Special Issue Company Assessment: Basis of Its Sustainable Development)


There is no doubt that the issue of making a good prediction about a company’s possible failure is very important, as well as complicated. A number of models have been created for this very purpose, of which one, the long short-term memory (LSTM) model, holds a unique position in that it generates very good results. The objective of this contribution is to create a methodology for the identification of a company failure (bankruptcy) using artificial neural networks (hereinafter referred to as “NN”) with at least one long short-term memory (LSTM) layer. A bankruptcy model was created using deep learning, for which at least one layer of LSTM was used for the construction of the NN. For the purposes of this contribution, Wolfram’s Mathematica 13 (Wolfram Research, Champaign, Illinois) software was used. The research results show that LSTM NN can be used as a tool for predicting company failure. The objective of the contribution was achieved, since the model of a NN was developed, which is able to predict the future development of a company operating in the manufacturing sector in the Czech Republic. It can be applied to small, medium-sized and manufacturing companies alike, as well as used by financial institutions, investors, or auditors as an alternative for evaluating the financial health of companies in a given field. The model is flexible and can therefore be trained according to a different dataset or environment.

1. Introduction

What are the future prospects of a company? Will it survive potential financial distress? Will it show positive development or is it heading towards bankruptcy? According to Tang et al. [1], Kliestik et al. [2], or Kliestik et al. [3], these are key questions that financial institutions must ask themselves prior to making decisions. Horak and Krulicky [4] stated that a good prediction is equally important for the strategic and operational decision-making of owners, management, and other stakeholders. However, as Antunes et al. [5] or Machova and Marecek [6] pointed out, the problem of predicting the potential failure of a company is extraordinarily complicated, especially during financial crises. Another complication is the necessity to localize standardized models because a model’s ability to predict bankruptcy is dependent on the specifics of a country, including its socio-economic [7] and legal [8] environments. Horak and Machova [9] stated that in spite of this, and within the context of global economics, especially in the case of financial instability, there is a clear need for generally valid models that, according to Alaminos et al. [10], surpass regionally localized predictive systems. As Alaka et al. [11] and Eysenck et al. [12] stated, current research in this field is focused on two statistical tools (multiple discriminant analysis and logistic regression) and six artificial intelligence tools (support vector machines, casuistic reasoning, decision trees, genetic algorithms, rough sets and, in particular, artificial neural networks). Their application is logical, especially due to the fact that, according to the results of an extensive study by Barboza et al. [13] and Horak et al. [14], those models that use machine learning are, on average, 10% more accurate than traditional models created on the basis of statistical methods. Another logical reason for the ever-increasing application of models based on machine learning is their rapid development and improvement in many other areas [15]. These modern methods go beyond financial statements, thereby taking other elements into account. Many studies have shown that the techniques used in the field of artificial intelligence can be a suitable alternative for traditional statistical methods used for predicting the financial health of a company because they do not apply past assumptions on the distribution of background data or the structure of the relationships between the variables involved. While statistical models require the researcher to specify the functional relationships between dependent and independent variables, non-parametric techniques enable the data to identify the functional relationships between the variables in the model. Within this context, the LSTM (long short-term memory) model holds a special position in that it produces extremely good results in the area of dynamic counting, which is typical, for example, for current electricity distribution networks, as well as for the economic environment [16]. It is for this reason that deep learning LSTM models are used, for example, by predictive models intended for high-frequency trading in financial markets [17] or for the prediction of future trends and the prices of housing [18]. The application of LSTM simply produces better results than vector regression or back propagation neural networks. Chebeir et al. [19], as well as Liu et al. [20], achieved very good results when applying the LSTM method to the assessment and optimization of the profitability of projects dealing with the extraction of shale gas. It is therefore surprising that there are not many bankruptcy models based on this exceptionally progressive method.
The objective of this study is to create a method of failure (bankruptcy) prediction that uses artificial neural networks (hereinafter referred to as “NN”) with at least one long short-term memory (LSTM) layer. The following research question was, therefore, formulated: Are neural networks containing LSTM suitable for predicting potential company bankruptcy?

2. Literature Review

Research into LSTM is very extensive. Due to its efficiency in capturing dynamic behavior and good performance in the area of long-term dependencies, it is used for a number of applications [21]. According to [22], it can be stated that the deep architecture of a recurrent neural network (RNN) and its variant of long short-term memory (LSTM) are more accurate than traditional statistical methods in modelling the data of time series.
On the other hand, they do not appear to be the best option for all areas in question. A comparative study of LSTM, the hybrid model of convolutional neural networks, and the multi-head self-attention approach (MHSA) by Xiao et al. [23] showed that MHSA turned out to be 6.5% better for the field of phishing. Similarly, a specially constructed network, CSAN (Crime Situation Awareness Network), surpassed predictive Cony-LSTM spatial temporal models in predicting crime frequency [24]. The extensive comparing of methods for logical relationships intended for fuzzy predictions of time series, namely machine learning, LSTM, and supporting vector machines, were the basis for evaluating supporting vector machines as more efficient than the other two [25]; according to Liu et al. [20], the model of adaptive wavelet transform (AWTM) surpassed the LSTM model in predicting time-frequency analysis of multi-frequency trading on stock markets. On the other hand, the experimental results of the LSTM model, with an added layer, according to Pang et al. [26], showed a high level of prediction accuracy for the composite stock model (56.9%, and 52.4% for individual stocks). Other results unambiguously confirm the efficiency of LSTM, for example, in the case of predicting the volatility of and time jumps in financial series within an analysis of 11 global stock markets [27], or in the case of predicting the development of stock markets according to Ding and Qin [28], whose model of associated LSTM showed a prediction accuracy of more than 95%. Chebeir et al. [19] achieved a similar result when it came to predicting stock volatility using LSTM.
The LSTM model is typically used by researchers for creating models and problem solving, for example in gestures or speech processing. Adeel et al. [29] dealt with the combination of declared excellent results, while Liu et al. [20] dealt with the successful classification of models working with the self-attention mechanism and bidirectional functions (SAMF-BiLSTM). The spam filtering models using the aforementioned method achieved a very high prediction accuracy of 99.44% [30].
Other fields in which LSTM is often used include medicine and the energy industry. The challenge is, for example, to combine the difficult task of predicting production from renewable resources with consumption, where, according to Wang et al. [24], the stability of the network can be ensured by using the LSTM module of interactive parallel prediction. Tian et al. [31] dealt with the successful prediction of the model for optimizing the use of lithium-ion batteries in the field of energy, while Chatterjee and Dethlefs [32] focused on the failures and anomalies in the operation of wind turbines by means of LSTM. Hong et al. [33] pointed to the stability and robustness of this method, as verified by extensive cross-validation and comparative analysis.
The advantages of using LSTM in comparison with traditional model solutions are addressed in the work of Yang et al. [34]. Zhang et al. [35] dealt with the efficiency of LSTM and FFNN (feedforward neural networks), where LSTM was less accurate, but with a significantly longer prediction horizon. In the same field, Yang et al. [34] proved that compared to vector modelling, LSTM showed a higher predictive accuracy, faster response time, and stronger generalization capability. In the case of renewable resources, Correa-Jullian et al. [36] compared the predictive methods based on standard neural networks and LSTM. The comparison showed that LSTM models achieve the lowest RMSE (Root Mean Square Error) error score, lowest standard deviation, and smallest relative error. LSTM’s excellent ability to capture the time dependency of a time series was used by [10] to predict electricity consumption, or short-term grid load. Wei et al. [37] used the combined model of singular spectrum analysis (SSA) and LSTM to predict gas consumption. Somu and Ramamritham [38] dealt with electricity consumption by means of another variant of ISCOA-LSTM based on a specific example. According to the authors, it is a highly effective prediction tool. Increasing energy efficiency through a combination of LSTM and bootstrapping was addressed in general by Zhu et al. [39].
A predictive model of oil production applied to actual production in China achieved an almost 100% success rate [20]. The application of a highly efficient integrated convolutional neural network (CNN) with LSTM to predict China’s future energy mix was addressed by Liu [40].
Extensive research has shown the effectiveness of the application of LSTM for predicting the development of prices. According to Qiao and Yang [41], the hybrid model based on wavelet transformation (WT), sparse autoencoder (SAE), and LSTM shows a high level of accuracy with regards to the prediction of electricity prices in the USA. Another hybrid model, WT-Adam_LSTM, used for predicting the development of electricity prices and verified on the basis of four case studies, was presented by Chang et al. [42].
Due to its progressiveness, the LSTM model has been assessed and developed by a number of authors. The assessment of the functional response of the LSTM and spiking neural network (SNN) revealed the dominance of SNN [43]. Du et al. [44] presented an end-to-end deep learning structure that integrates a conventional coded context vector and a time-of-attention vector for the joint learning of time representation on the basis of the LSTM model. Wang et al. [24] dealt with the problem of long-term dependence in sequence data due to insufficient memory capacity in LSTM cells, solving it using an attention-aware bidirectional multi-residual recurrent neural network (ABMRNN). Punia et al. [45] presented a new predictive method that combines LSTM and random forest (RF), the efficiency of which is compared to other methods, such as neural networks, multiple regression, ARIMAX (Autoregressive Integrated Moving Average with Explanatory Variable), etc.
The effectiveness of LSTM in comparing the performances of LSTM RNN by initialized methods of learning transmission and randomly initialized recurrent neural networks was confirmed by Fong et al. [46]. Another variant called CTS-LSTM for the collective prediction of correlated time series with the aim of improving the predictive accuracy of the model was presented by Wan et al. [47]. Bandara et al. [48] solved the problem of time series’ heterogeneity by introducing the concept of similarity between time series. Park et al. [49] strove to improve classification performance by means of the interaction between the model architecture factors and the dimensions of the dataset characteristics. According to the authors, bidirectional LSTM dominated within the given aspects. Zhang et al. [35] presented a different model of weighted auto regressive long short-term memory (WAR-LSTM), using it to extract representative data from more than one variable. Another evolutionary model of LSTM, focused on the transmission of shared parameters in order to increase the efficiency of prediction in multiple time slots, was created by Liu and Liu [18]. Karim et al. [50] focused on streamlining the solution of complex multidimensional time series classification through the transformation of existing univariate classification models into multivariate ones, while Farzad et al. [51] focused on activation functions, according to which the quantitative results showed that the smallest average error is achieved by means of the Elliott activation function and modifications thereof. Another hybrid model based on exponential smoothing combined with advanced long-term memory networks was presented by Smyl [52].
A section in this contribution is dedicated to the potential application of LSTM to the management of financial risks or to the prediction of company bankruptcy. Despite this, many authors continue to work with traditional bankruptcy models, such as Altman Z-Score, Kralicek Quick Test, IN 99, IN05 [53], hybrid models of classification and regression trees (CART), multivariate adaptive regression spline (MARS) models [54], or evolutionary dynamics and the optimization of the strategy for certain types of interconnected evolutionary games [55]. In other cases, for the verification of statistical bankruptcy models, the Monte Carlo method is used [56]. Alternatively, it is possible to see the bankruptcy model as a multivariate grey prediction problem, for which Hu [57] used genetic algorithms.
There is also the question of the verification of these methods, whereby, based on the assessment of bankruptcy model performance (discriminant analysis, logistic regression, and multilayer perceptron network), decision trees appear to be the most efficient [58], as well as the success rate of the bankruptcy models, which, according to Kubenka and Myskova [59], is lower in the overall comparison of the three methods than the researchers stated.
The application of machine learning is nothing new in the case of bankruptcy models. For bankruptcy prediction, Zhou and Lai [60] effectively used the AdaBoost algorithm combined with imputation, while Kim et al. [61] examined the benefits of deep learning on decision-making in financial risk management. Liu and Liu [18] used LSTM models combined with block-chain technology to increase financial performance and reduce risks. Koudjonou and Rout [62] dealt with the net value of assets by means of an LSTM recurrent neural network and the comparison of single-layer, multi-layer, unidirectional, and bidirectional networks. Bankruptcy models based on LSTM, however, are used relatively little, despite the significant development of this method and its effectiveness, especially in dynamically developing situations and in the case of insufficient data. The exception is a deep learning model presented by Mai et al. [63] for predicting company bankruptcy using text information. A comprehensive database of bankruptcy of 11,827 US companies shows that deep-learning models provide excellent predictions; interestingly, simpler models, such as averaging appear to be more efficient than convolutional neural networks. However, the authors themselves say that in this case, it is the first step in using the highly progressive LSTM method within bankruptcy models. The authors also mention the limitation of data obtained exclusively from financial statements (MD&A–Management Discussion and Analysis) or the difficult interpretation of the results based on the processing of unstructured data within the so-called “black box” of deep-learning models. They therefore recommend further research by means of other deep-learning models, as well as by means of using other tools capable of evaluating the performances of the created models, such as H-measures or Kolmogorov–Smirnov statistics of good results.

3. Materials and Methods

The application part was organized according to the following structure, with further explanations below:
Selection and preparation of the data for the calculation.
Division of the data into training and testing data sets.
Creation of a bankruptcy model by means of an experiment using Mathematica Software.
Generation of NN using LSTM networks and other elementwise layers.
Evaluation of the performance of networks in the training and testing data sets, creation of a confusion matrix characterizing the correct classification of companies into “active” and “in liquidation”.
Description of the best NN and discussion on the success rate of the network.

3.1. Data

The source of the data on industrial companies operating in the Czech Republic was the Albertina database. The selected industrial companies fell under section “C” of the CZ NACE (Czech classification of economic activities), specifically groups 10–33.
The data set included 5 consecutive marketing years (from 2014 to 2018). The data set contained those companies able to survive potential financial distress (hereinafter also referred to as “active companies”), as well as companies in liquidation. In total, 5500 companies were included. Data rows (one company and one year representing one data row) with nonsensical data or with a large amount of information missing were excluded.
For the purpose of the analysis, selected items from financial statements were used, specifically the balance sheet and profit and loss statement.
For this analysis, only some of the items were used:
  • AKTIVACELK—total assets, i.e., the result of economic activities carried out in the past. This represents the future economic profit of the company.
  • STALAA—fixed assets, i.e., long-term, fixed, non-current items, including property components, used for company activities over the long run (for more than one year) and consumed over time.
  • OBEZNAA—current assets characterized by the operating cycle, i.e., they are in constant motion and change form. These include money, materials, semi-finished products, unfinished products, finished products, and receivables from customers.
  • KP—short-term receivables with a maturity of less than 1 year, representing the right of the creditor to demand the fulfilment of a certain obligation from the other party. The receivable ceases to exist upon the fulfilment of the obligation.
  • VLASTNIJM—equity, i.e., the company’s resources for financing assets in order to create capital. This primarily concerns the contributions of the founders (owners or partners) to the basic capital of the company and those components arising from the company’s activities.
  • CIZIZDROJE—borrowed capital, i.e., company debts that have to be repaid within a specified period of time. This represents the company’s liabilities towards other entities.
  • KZ—short-term liabilities, i.e., due within 1 year. Together with equity, they ensure the financing of the day-to-day activities of the company. These primarily include bank loans, liabilities to employees and institutions, debts to suppliers or taxes due.
  • V—performance, i.e., the results of company activities which are characterized by the main activity of the company—production. This includes the goods and services used for satisfying demands.
  • SLUZBY—services, i.e., those activities intended to meet human needs or the needs of a company by means of their execution.
  • PRIDHODN—added value, i.e., trademarking, sales, changes in inventory through own activities, or activation reduced by power consumption. This includes both company margin and performance.
  • ON—personnel costs, i.e., gross salaries and the employer’s compulsory social and health insurance contributions for each employee.
  • PROVHOSP—operating results, i.e., the outcomes and products that reflect the ability of a company to transform production factors.
  • NU—interest payable, i.e., the price of borrowed capital.
  • HOSPVZUO—economic result for an accounting period, i.e., from operational, financial, and extraordinary activities.
  • STAV—target situation, i.e., classification as “active” for companies able to survive potential financial distress, and “in liquidation” for companies that will go bankrupt.
Absolute indicators characterize a company from several perspectives. They evaluate the structure of the company´s capital, the price of capital, its technological level (by means of added value), the ability to perform its main activity (that is, to transform production factors into products), and to achieve one of the main goals of its existence (to generate profit). A number of authors (e.g., Altman) have used ratios to create a method for predicting market failure. In such cases, two possible reasons can be identified. Firstly, they anticipated the economic interpretation of their models’ segments, therefore giving their models certain economic rationality. Secondly, they simplified the computing power requirements. This specifically concerns the reduction of future model data. In the model, the NN acts like a black box if the identification of the calculated result is difficult. It is, therefore, not necessary to follow the economic interpretation of each bond of neurons and the propagated signal because the bond as such is insignificant. In addition, it is subsequently not necessary to limit the number of variables. Hardware computing power is able to process a large amount of data during its creation. For the application of the NN generated, a fraction of the computing power is necessary compared to the preparation stage of the NN.
After the aforementioned modifications (removal of nonsensical data, etc.), the data were divided into two datasets: a training and a testing dataset. The training dataset was used to train the neural structure, while the testing dataset was used for the validation of the result. The statistical characteristics of the individual items in the training and testing datasets are presented in Table 1.

3.2. Methods

The bankruptcy model was created using an artificial deep learning neural network. As indicated by the research carried out and the objective of the contribution, at least one layer of LSTM was used for the creation of the NN. For the solution of the problem, Wolfram’s Mathematica software (version 13) was used.
The specific NN structure was determined by an experiment. The individual layers consisted of the following components:
The input layer: 1 × 14 matrix. The matrix consisted of one row that included 14 input continuous variables (AKTIVACELK, STALAA, OBEZNAA, KP, VLASTNIJM, CIZIZDROJE, KZ, V, SLUZBY, PRIDHODN, ON, PROVHOSP, NU, HOSPVZUO).
1st hidden layer: LSTM layer. The output was a 1 × n matrix, whereby the number of matrix elements was part of the experiment. The size of the matrix influenced the predictive ability of the model. A low number of matrix elements could generate a result with a higher level of inaccuracy, whereas a high number of elements could make the model too complex and overfitted (it shows excellent performance parameters for the training dataset, but is totally incapable of generating a proper classification for the testing dataset). The number of matrix elements was at the interval between 5 and 2000 elements.
2nd hidden layer: elementwise layer. The objective was to add a certain degree of non-linearity to the NN. A partial experiment was carried out on the basis of which the suitability of the following functions was tested:
  • Hyperbolic tangent (Tanh),
  • Sinus (Sin),
  • Ramp (referred to as ReLU),
  • Logistic function (logistic sigmoid).
3rd hidden layer: elementwise layer. The objective was to add a certain degree of non-linearity to the NN. As this was the second elementwise layer, the non-linearity was stronger. Also in this case, a partial experiment was carried out. The suitability of the following functions was tested accordingly:
  • Hyperbolic tangent (Tanh),
  • Sinus (Sin),
  • Ramp (referred to as ReLU),
  • Logistic function (logistic sigmoid).
4th hidden layer of neurons: LSTM layer. The output was a 1 × 2 matrix. The size of the matrix was determined by the number of possible results—predicting either “active” or “in liquidation”.
Output layer: two neurons representing a vector with two elements. The vector was subsequently decoded as “active company” or “company in liquidation”.

3.3. Long-Short Term Memory Layer

LSTM is considered to be a specific type of recurrent NN consisting of several components. It is possible to identify the elementwise layer with the logistic sigmoid function and hyperbolic tangent, linear layer, concatenate layer, copy layer, and the transfer of the data in the form of vectors.
The basic LSTM processes are defined as input gate, output gate, forget gate, and memory gate. The state of the cell is defined as follows:
c t = f t c t 1 + i t m t ,
ct: new state of the variable;
ft: forget gate;
ct−1: original state of the variable;
it: input gate;
mt: memory gate.
The input gate is defined as follows:
i t = σ [ W i x x t + W i s s t 1 + b i ] ,
σ: logistic sigmoid;
Wix: input weight in input gate, matrix n × k;
xt: input variable, n × k matrix;
Wis: weight in input gate, n × n matrix;
st−1: previous state;
bi bias: vector size n.
The state is determined by the following formula:
s t = o t T a n h [ c t ] ,
st: state of the variable;
ot: output gate;
Tanh: hyperbolic tangent.
Output gate is represented by the formula below:
o t = σ [ W a x x t + W a s s t 1 + b o ] ,
Wax: input weight in output gate, n × k matrix;
Was: weight in output gate, n × n matrix;
bo bias: vector size n.
Forget gate is an important innovation of LSTM:
f t = σ [ W f x x t + W f s s t 1 + b f ] ,
Wfx: determines the forget gate input weigh, n × k matrix;
Wfs: weight in forget gate, n × n matrix;
bf: vector size n.
The last main process to determine is memory gate:
m t = T a n h [ W m x x t + W m s s t 1 + b m ]
Wmx: input weight in memory gate, n × k matrix;
Wms: weight in memory gate, n × n matrix;
Bm: vector size n.

3.4. Elementwise Layer

The elementwise layer is a single layer of neurons that takes n inputs from the previous layer. It adds non-linearity in the calculation and transfers n inputs to another layer of the NN.
As part of the research presented in this contribution, non-linearity was added in the form of testing in the 2nd and 3rd hidden layers using the following functions:
Hyperbolic tangent (Tanh):
f ( x ) = tanh x = e x e x e x + e x .
Sinus (Sin):
f ( x ) = S i n ( x ) .
Ramp (referred to as ReLU):
R ( x ) : R R 0 + .
Logistic function (logistic sigmoid):
f ( x ) = 1 1 + e x .

3.5. Evaluation of Network Performance

Within the experiment mentioned above, 1000 neural networks were generated, with differing sizes of the vector from the first LSTM layer and the activation function of the 2nd and 3rd layer of NN.
The evaluation of the networks was based on:
  • The performance of the individual networks in the training and testing datasets.
  • The confusion matrix characterizing the correct classification of companies into “active” and “in liquidation”. The confusion matrix was created for both the training and testing datasets.
The best NN is subsequently described in detail in terms of its characteristics. However, if the network showed signs of overfitting, the second most successful network (in terms of its parameters) was then used.

4. Results

Within the experiment, a total of 1000 NNs were generated and their performance in the training and testing datasets compared. The highest possible performance was sought in all datasets and all parameters (“active”, “in liquidation”) and, at the same time, similar in the training and testing datasets. Table 2 shows the neural networks with the best performance parameters.
The network structure was as indicated above. The value 14 indicates the number of neurons in the input layer. It is a 1 × n matrix, i.e., a vector of variables. The second value in the NN structure indicates the number of elements of the vector (1 × n matrix) of the new state of the variables from the LSTM layer, i.e., the output of the first hidden layer of the NN. What follows in succession are the activation function of the second hidden layer of neurons, the activation function of the third hidden layer of neurons, the output of the second LSTM layer, i.e., a vector (or 1 × 2 matrix), and the final state, i.e., whether the company was “active” or “in liquidation”. The training performance provided information on the accuracy of the classification determined under the training dataset, and the testing performance provided information on the accuracy of the classification determined under the testing (validation) dataset.
A neural structure was sought of which the performance was as close to 1 as possible in all datasets. It is important for the results that the performance value is, ideally, the same for all datasets and groups.
On the basis of this evaluation, the network identified as ID 1 appears to be the best. Its overall performance was 0.978 for the training dataset and 0.971 for the testing dataset. However, there is a big problem with regards to predicting companies that will go bankrupt. This is due to the fact that the decision to enter into liquidation is often not based on the management or the owners’ rational decision on the economic and financial sustainability of a company. Under such circumstances, it may happen that a company that could be able to operate ceases its activities anyway. Obviously, when predicting failure, the evaluator or the model must remove a large amount of noise. In the case of NN ID 1, the performance with regards to predicting failure was almost 0.743 for the training dataset and 0.752 for the testing dataset (i.e., even more than for the training dataset). The NN therefore suffered from overfitting, showing excellent performance parameters, but zero applicability, as stated above. As a result, other successful networks were retained and other successful NNs analyzed.
Figure 1 shows the structure of NN ID 1.
The figure shows 14 neurons in the input layer. It was a 1 × 14 matrix that characterizes the input data of a company (asset structure, capital structure, price of capital, technological level, the ability to carry out its own activities and make profits). There was also a LSTM layer, the output of which was a 1 × 940 matrix (or a vector with 940 elements). To the individual matrix elements, non-linearity was added in the subsequent two layers. In both elementwise layers, it concerned a function of the hyperbolic tangent. The following layer was the LSTM layer, the output of which was a 1 × 2 matrix (or vector with 2 elements), from which the result was derived, i.e., the company was either “active” or “in liquidation”. In terms of the internal functioning of the whole NN, both LSTM layers appear to be of interest. The elementwise layer only represented a certain mechanical element that changed the distribution of the signal in another NN layer. The inner structure of the first LSTM layer (the first inner layer of the NN) is presented in Table 3.
The structure of the LSTM layer shows the distribution of the information in the layer, mainly the relationship between the input data and the output vector with 940 elements. The structure of the second LSTM layer (the fourth hidden layer of the NN) is similarly presented in Table 4.
If the inserted non-linearity were left aside and the whole process of data transformation was simplified, it turned out that 14 data on a company entered into the NN. The data were analyzed and their combinations expressed as 940 values, which were subsequently analyzed (more precisely, their combinations were analyzed) and reduced to two target values expressing the probability of the company being classified as “active” or “in liquidation”. At the end of the NN, there was a decoder that determined the assumed state of the company on the basis of probability.

5. Discussion

A NN was obtained that, at first sight, is able to predict, with a high probability, the future development of a company operating in the manufacturing sector in the Czech Republic. The results described clearly show the structure of the network and the method of data processing in the network. Both the NN and the background data are available for calculation for the validation of the results and for practical application. However, it is necessary to consider the practical or theoretical benefits of the NN obtained and its applicability in practice.
The theoretical benefit of this contribution consists in the possibility to apply LSTM NN as a tool for predicting bankruptcy. It was verified and proved that this type of recurrent NN is able to process and analyze data on a company, as well as produce a result. In terms of the theoretical benefit, instead of the results of the model, it is necessary to procedurally monitor (mathematically in this case) whether it is possible to process the data and obtain a meaningful result. The NN structure could be further processed and adapted to the required outputs. It is also possible to train the NN (and the partial weights of the NN) so that it is possible to obtain the correct results in the required structure.
The practical application of the NN with the LSTM layers was confirmed mainly by the confusion matrix for the training and testing datasets. Figure 2 shows the confusion matrix for the training dataset.
The confusion matrix for the training dataset shows that the NN appears to be very successful in predicting the ability to overcome potential financial distress. In 2878 cases, the result was predicted correctly, with only 64 errors. For the same dataset, it predicted bankruptcy for 1089 companies, with 377 errors. This is the aforementioned noise arising from the fact that a number of companies cease their activities without being forced to do so by their financial and economic results. Despite this, this represents an excellent result. Figure 3 presents the results for the testing dataset.
The second confusion matrix also produced excellent results. Of the 487 active companies, the NN was able to identify 473 companies able to survive potential financial distress; of the 262 companies identified as going bankrupt, the NN identified 182.
By default, the successful prediction rate was higher than 50% in those situations where it was not a coincidence. In the case of NN ID 1, the overall successful prediction rate was higher than 97% for all companies and 75% for those going bankrupt. It can therefore be concluded that the NN with the LSTM layer (NN ID 1 in particular) is applicable in practice. It is possible to compare the results with other studies. For example, Mihalovic [64] dealt with the application of the original Altman Index Z-Score. The authors focused on the financial statements of 373 Greek companies in the years 1999–2006. The results of their studies indicated that the success rate of the model was 52% two years before bankruptcy and 66% one year before bankruptcy. It should also be mentioned that the authors used the market value of equity in the index; the results, therefore, differed in the individual years and also according to the actual situation of the financial markets. The issue of predictive models was also addressed by Mihalovic [64], who focused on predictive bankruptcy models for a total of 236 Slovak companies. In the study, the author primarily compared the overall predictive performance of two models, the first based on discriminant analysis and the second on logistic regression. The results of the research showed that the model based on the logit function provided more accurate results, and that the most important factors that prevent the failure of a company are short-term assets, short-term liabilities, net income, and total assets. Lin [65] examined the predictive power of the four most commonly used models of financial distress. On the basis of his study, he created reliable predictive models related to the bankruptcy of public industrial companies in Taiwan, specifically, logit, probit, and artificial neural network models. The author concluded that the aforementioned models are able to generalize and show higher predictive accuracy. It also showed that the probit model has the most stable and best performance. Unvan and Tatlidil [66] dealt with the comparison of models that could be applied in bank investigations and the supervisory process for detecting banks with serious problems. The dataset consisted of 70 Turkish banks and included information on their financial situation, as well as data on their capital adequacy, liquidity, asset quality, cost and return structure, and profitability. Using variable methods of choosing financial data, the most important financial characteristics were determined and subsequently used as independent variables to create probit and logit models. Finally, these models were compared with the selected best models with the best predictive power. Jifi [67] evaluated the accuracy and power of conventional credibility and bankruptcy models. For the purposes of the evaluation, companies operating in the construction field in the Czech Republic, which went bankrupt within a period of 5 years, were selected. For each of the companies, the evaluation was carried out by means of the following models: Kralicek Quick test, the plausibility index, Rudolf Doucha’s balance analysis, Grünwald’s index, D-score, Aspect Global Rating of the Altman model, Taffler’s model, Springate score, the Zmijewski X-Score model, and all variants of the IN index. The overall evaluation was subsequently based on the success rate of the individual models. The research results revealed that the most successful model for predicting bankruptcy is the Aspect Global Rating with a success rate of 99%, followed by Zmijewski (95%). Given the specific features of the recent financial crisis, Iturriaga and Sanz [68] created a model of NN to study the bankruptcy of American banks. Their research combined multilayer perceptrons and self-organizing maps, therefore providing a tool that displays the probability of failure up to three years in advance. On the basis of the failures of US banks between May 2012 and December 2013, the authors created a model for detecting failure and a tool for assessing banking risk in the short, medium, and long term. This model was able to detect 96.15% of failures, therefore overcoming the traditional models of bankruptcy prediction. Bateni and Asghari [69] predicted bankruptcy using techniques for predicting logit and genetic algorithms. The study compared the performance of predictive models on the basis of the data obtained from 174 bankrupt and non-bankrupt Iranian companies listed on the Tehran Stock Exchange in the years 2006–2014. The research results showed that the genetic model in training and testing samples achieved 95% and 93.5% accuracy, respectively, while the logit model achieved only 77% and 75% accuracy, respectively. The results show that these two models are able to predict bankruptcy, while, in this respect, the model of genetic algorithms is more accurate than the logit model.
The ability to predict the future state of companies creates the potential to apply NN in practice. However, there is a problem with the complex structure of the NN. It cannot be recalculated or programmed in a different environment than Wolfram’s Mathematica software or in the form of C++ code or Java. Managers or financial managers do not have such knowledge. Even this contribution only presents selected characteristics of the NN ID 1 and is not able to capture the whole structure of the best NN. Taking into account, for example, the Altman Z-Score [70] and Zeta [71] models, their advantage is in the fact that users are able to implement them themselves with minimum requirements in terms of their knowledge of mathematics or deep knowledge of the issue of company financial management. The problem is how to present the NN to the public so that it is applicable even for a layman. The solution could be the implementation of the NN in an application for company evaluation or the creation of a user-friendly interface. Both alternatives require the use of information and communication technologies, and the users’ remote access to the NN.
It is, therefore, possible to formulate the answers to the research question about whether neural networks containing LSTM are suitable for predicting the potential bankruptcy of a company.
Yes, neural networks containing a LSTM layer are suitable for predicting the potential bankruptcy of a company. However, there is a problem if a layman wants to use the model. For laymen, the model will only be applicable when it is accessible through a user-friendly interface.

6. Conclusions

It is clear that NNs are currently not only able to solve a number of tasks in the economic sphere but are also more efficient in doing so than models created using conventional statistical methods (e.g., using logistic regression). So far, there have been a number of concepts of NN imitating the actual biological neural structure. Researchers have already moved on from basic NN (e.g., multilayer perceptron NN or generalized regression NN, deep learning). Deep learning networks have the potential to solve relatively complex tasks. This is also evident from the NN ID 1 that arose out of the aforementioned research into company failure. This NN is able to predict the future development of a company operating in the manufacturing sector in the Czech Republic. The NN ID 1 is flexible and can be trained on different datasets for different environments (temporally, spatially, and materially different). The objective of the research and this contribution has therefore been achieved.
However, the result was limited by the transferability of the NN and its application for professional and lay public. Although the result of the NN application was clear and easy to interpret, the model is not easily graspable for professionals and laymen with a poor command of ICT, as it is too complex.
From the above, further direction of research follows. The factual side can “only” be adjusted in order to improve the network’s performance. However, the formal side, in other words, the simple presentation of the model and its easy applicability, must be solved.
The limitations of the application of the proposed method for predicting market failure lie mainly in the requirement for available data on a company. The problem with this is the difference in accounting methods applied in specific countries to individual items. This shortcoming could either be solved by preparing the data prior to processing or through NN overfitting. A significant limitation is the difficult work with the created NN. It can only be used by laymen if the network runs in the background and the user has a user-friendly environment, for example, in the form of a thin client operated by means of a web browser. The resulting NN will be more easily accessible to an expert who has a command of NNs, and especially of the Mathematica software environment, or who is able to program in cascading languages. Nevertheless, even this shortcoming can be solved, as most small- and medium-sized companies and all large companies use websites and programmers who can apply the method (once or by means of a simple thin client), since the NN representing the method is being freely distributed by the authors—the link is given in Section 4: Results.

Author Contributions

Conceptualization, M.V. and J.V.; methodology, M.V.; software, J.V. and P.S.; validation, J.V. and P.S.; formal analysis, P.S.; investigation, M.V. and J.V.; resources, P.S.; data curation, J.V.; writing—original draft preparation, M.V.; writing—review and editing, J.V. and P.S.; visualization, J.V.; supervision, M.V.; project administration, M.V. and P.S. All authors have read and agreed to the published version of the manuscript.


This research received no external funding.

Conflicts of Interest

The authors declare no conflict of interest.


  1. Tang, Y.; Ji, J.; Zhu, Y.; Gao, S.; Tang, Z.; Todo, Y. A differential evolution-oriented pruning neural network model for bankruptcy prediction. Complexity 2019, 2019, 1–21. [Google Scholar] [CrossRef] [Green Version]
  2. Kliestik, T.; Misankova, M.; Valaskova, K.; Svabova, L. Bankruptcy prevention: New effort to reflect on legal and social changes. Sci. Eng. Ethics 2018, 24, 791–803. [Google Scholar] [CrossRef] [PubMed]
  3. Kliestik, T.; Vrbka, J.; Rowland, Z. Bankruptcy prediction in Visegrad group countries using multiple discriminant analysis. Equilib. Q. J. Econ. Econ. Policy 2018, 13, 569–593. [Google Scholar] [CrossRef]
  4. Horak, J.; Krulicky, T. Comparison of exponential time series alignment and time series alignment using artificial neural networks by example of prediction of future development of stock prices of a specific company. In Proceedings of the SHS Web of Conferences: Innovative Economic Symposium 2018—Milestones and Trends of World Economy (IES2018), Beijing, China, 8–9 November 2018. [Google Scholar] [CrossRef]
  5. Antunes, F.; Ribeiro, B.; Pereira, F. Probabilistic modeling and visualization for bankruptcy prediction. Appl. Soft Comput. 2017, 60, 831–843. [Google Scholar] [CrossRef] [Green Version]
  6. Machova, V.; Marecek, J. Estimation of the development of Czech Koruna to Chinese Yuan exchange rate using artificial neural networks. In Proceedings of the SHS Web of Conferences: Innovative Economic Symposium 2018—Milestones and Trends of World Economy (IES2018), Beijing, China, 8–9 November 2018. [Google Scholar] [CrossRef]
  7. Gavurova, B.; Packova, M.; Misankova, M.; Smrcka, L. Predictive potential and risks of selected bankruptcy prediction models in the Slovak company environment. J. Bus. Econ. Manag. 2017, 18, 1156–1173. [Google Scholar] [CrossRef] [Green Version]
  8. Nakajima, M. Assessing bankruptcy reform in a model with temptation and equilibrium default. J. Public Econ. 2017, 145, 42–64. [Google Scholar] [CrossRef] [Green Version]
  9. Horak, J.; Machova, V. Comparison of neural networks and regression time series on forecasting development of US imports from the PRC. Littera Scr. 2019, 12, 22–36. [Google Scholar]
  10. Alaminos, D.; Del Castillo, A.; Fernández, M.A.; Ponti, G. A global model for bankruptcy prediction. PLoS ONE 2016, 11. [Google Scholar] [CrossRef] [Green Version]
  11. Alaka, H.A.; Oyedele, L.O.; Owolabi, H.A.; Kumar, V.; Ajayi, S.O.; Akinade, O.O.; Bilal, M. Systematic review of bankruptcy prediction models: Towards a framework for tool selection. Expert Syst. Appl. 2018, 94, 164–184. [Google Scholar] [CrossRef]
  12. Eysenck, G.; Kovalova, E.; Machova, V.; Konecny, V. Big data analytics processes in industrial internet of things systems: Sensing and computing technologies, machine learning techniques, and autonomous decision-making algorithms. J. Self-Gov. Manag. Econ. 2019, 7, 28–34. [Google Scholar] [CrossRef] [Green Version]
  13. Barboza, F.; Kimura, H.; Altman, E. Machine learning models and bankruptcy prediction. Expert Syst. Appl. 2017, 83, 405–417. [Google Scholar] [CrossRef]
  14. Horák, J.; Vrbka, J.; Šuleř, P. Support vector machine methods and artificial neural networks used for the development of bankruptcy prediction models and their comparison. J. Risk Financ. Manag. 2020, 13, 3390. [Google Scholar] [CrossRef] [Green Version]
  15. Vrbka, J.; Rowland, Z. Using artificial intelligence in company management. In Digital Age: Chances, Challenges and Future, 1st ed.; Lecture Notes in Networks and Systems; Ashmarina, S.I., Vochozka, M., Mantulenko, V.V., Eds.; Springer: Cham, Switzerland, 2020; pp. 422–429. [Google Scholar] [CrossRef]
  16. Zheng, C.; Wang, S.; Liu, Y.; Liu, C.; Xie, W.; Fang, C.; Liu, S. A novel equivalent model of active distribution networks based on LSTM. IEEE Trans. Neural Netw. Learn. Syst. 2019, 30, 2611–2624. [Google Scholar] [CrossRef] [PubMed]
  17. Rundo, F. Deep LSTM with reinforcement learning layer for financial trend prediction in FX high frequency trading systems. Appl. Sci. 2019, 9, 4460. [Google Scholar] [CrossRef] [Green Version]
  18. Liu, R.; Liu, L. Predicting housing price in China based on long short-term memory incorporating modified genetic algorithm. Soft Comput. 2019, 23, 11829–11838. [Google Scholar] [CrossRef]
  19. Chebeir, J.; Asala, H.; Manee, V.; Gupta, I.; Romagnoli, J.A. Data driven techno-economic framework for the development of shale gas resources. J. Nat. Gas Sci. Eng. 2019, 72, 103007. [Google Scholar] [CrossRef]
  20. Liu, W.; Liu, W.D.; Gu, J. Forecasting oil production using ensemble empirical model decomposition based Long Short-Term Memory neural network. J. Pet. Sci. Eng. 2020, 189, 107013. [Google Scholar] [CrossRef]
  21. Karevan, Z.; Suykens, J.A.K. Transductive LSTM for time-series prediction: An application to weather forecasting. Neural Netw. 2020, 125, 1–9. [Google Scholar] [CrossRef]
  22. Sagheer, A.; Kotb, M. Unsupervised pre-training of a deep LSTM-based stacked autoencoder for multivariate time series forecasting problems. Sci. Rep. 2019, 9, 19038. [Google Scholar] [CrossRef] [Green Version]
  23. Xiao, X.; Zhang, D.; Hu, G.; Jiang, Y.; Xia, S. A Convolutional neural network and multi-head self-attention combined approach for detecting phishing websites. Neural Netw. 2020, 125, 303–312. [Google Scholar] [CrossRef]
  24. Wang, B.; Zhang, L.; Ma, H.; Wang, H.; Wan, S. Parallel LSTM-Based regional integrated energy system multienergy source-load information interactive energy prediction. Complexity 2019, 2019, 1–13. [Google Scholar] [CrossRef] [Green Version]
  25. Panigrahi, S.; Behera, S.H. A study on leading machine learning techniques for high order fuzzy time series forecasting. Eng. Appl. Artif. Intell. 2020, 87, 103245. [Google Scholar] [CrossRef]
  26. Pang, X.; Zhou, Y.; Wang, P.; Lin, W.; Chang, W. An innovative neural network approach for stock market prediction. J. Supercomput. 2020, 76, 2098–2118. [Google Scholar] [CrossRef]
  27. Au Yeung, J.F.K.; Wei, Z.; Chan, K.Y.; Lau, H.Y.K.; Yiu, K.C. Jump detection in financial time series using machine learning algorithms. Soft Comput. 2020, 24, 1789–1801. [Google Scholar] [CrossRef]
  28. Ding, G.; Qin, L. Study on the prediction of stock price based on the associated network model of LSTM. Int. J. Mach. Learn. Cybern. 2019, 11, 1307–1317. [Google Scholar] [CrossRef] [Green Version]
  29. Adeel, A.; Gogate, M.; Hussain, A. Contextual deep learning-based audio-visual switching for speech enhancement in real-world environments. Inf. Fusion 2020, 59, 163–170. [Google Scholar] [CrossRef] [Green Version]
  30. Roy, P.K.; Singh, J.P.; Banerjee, S. Deep learning to filter SMS Spam. Future Gener. Comput. Syst. 2020, 102, 524–533. [Google Scholar] [CrossRef]
  31. Tian, Y.; Lai, R.; Li, X.; Xiang, L.; Tian, J. A combined method for state-of-charge estimation for lithium-ion batteries using a long short-term memory network and an adaptive cubature Kalman filter. Appl. Energy 2020, 265. [Google Scholar] [CrossRef]
  32. Chatterjee, J.; Dethlefs, N. Deep learning with knowledge transfer for explainable anomaly prediction in wind turbines. Wind Energy 2020, 23, 1693–1710. [Google Scholar] [CrossRef]
  33. Hong, J.; Wang, Z.; Chen, W.; Yao, Y. Synchronous multi-parameter prediction of battery systems on electric vehicles using long short-term memory networks. Appl. Energy 2019, 254. [Google Scholar] [CrossRef]
  34. Yang, G.; Wang, Y.; Li, X. Prediction of the NO emissions from thermal power plant using long-short term memory neural network. Energy 2020, 192. [Google Scholar] [CrossRef]
  35. Zhang, X.; Zou, Y.; Li, S.; Xu, S. A weighted auto regressive LSTM based approach for chemical processes modeling. Neurocomputing 2019, 367, 64–74. [Google Scholar] [CrossRef]
  36. Correa-Jullian, C.; Cardemil, J.M.; López Droguett, E.; Behzad, M. Assessment of deep learning techniques for prognosis of solar thermal systems. Renew. Energy 2020, 145, 2178–2191. [Google Scholar] [CrossRef]
  37. Wei, N.; Li, C.; Peng, X.; Li, Y.; Zeng, F. Daily natural gas consumption forecasting via the application of a novel hybrid model. Appl. Energy 2019, 250, 358–368. [Google Scholar] [CrossRef]
  38. Somu, N.; Ramamritham, K. A hybrid model for building energy consumption forecasting using long short term memory networks. Appl. Energy 2020, 261, 114131. [Google Scholar] [CrossRef]
  39. Zhu, X.; Zeng, B.; Dong, H.; Liu, J. An interval-prediction based robust optimization approach for energy-hub operation scheduling considering flexible ramping products. Energy 2020, 194. [Google Scholar] [CrossRef]
  40. Liu, Y. Novel volatility forecasting using deep learning–Long Short Term Memory Recurrent Neural Networks. Expert Syst. Appl. 2019, 132, 99–109. [Google Scholar] [CrossRef]
  41. Qiao, W.; Yang, Z. Forecast the electricity price of U.S. using a wavelet transform-based hybrid model. Energy 2020, 193, 116704. [Google Scholar] [CrossRef]
  42. Chang, Z.; Zhang, Y.; Chen, W. Electricity price prediction based on hybrid model of ADAM optimized LSTM neural network and wavelet transform. Energy 2019, 187, 115804. [Google Scholar] [CrossRef]
  43. Xie, X.; Liu, G.; Cai, Q.; Sun, G.; Zhang, M.; Qu, H. An end-to-end functional spiking model for sequential feature learning. Knowl. Based Syst. 2020, 195, 105643. [Google Scholar] [CrossRef]
  44. Du, S.; Li, T.; Yang, Y.; Horng, S.J. Multivariate time series forecasting via attention-based encoder–decoder framework. Neurocomputing 2020, 388, 269–279. [Google Scholar] [CrossRef]
  45. Punia, S.; Nikolopoulos, K.; Singh, S.P.; Madaan, J.K.; Litsiou, K. Deep learning with long short-term memory networks and random forests for demand forecasting in multi-channel retail. Int. J. Prod. Res. 2020, 58, 4964–4979. [Google Scholar] [CrossRef]
  46. Fong, I.; Li, H.T.; Fong, S.; Wong, R.K.; Tallón-Ballesteros, A.J. Predicting concentration levels of air pollutants by transfer learning and recurrent neural network. Knowl. Based Syst. 2020, 192, 105622. [Google Scholar] [CrossRef]
  47. Wan, H.; Guo, S.; Yin, K.; Liang, X.; Lin, Y. CTS-LSTM: LSTM-based neural networks for correlated time series prediction. Knowl. Based Syst. 2020, 191, 105239. [Google Scholar] [CrossRef]
  48. Bandara, K.; Bergmeir, C.; Smyl, S. Forecasting across time series databases using recurrent neural networks on groups of similar series: A clustering approach. Expert Syst. Appl. 2020, 140, 112896. [Google Scholar] [CrossRef] [Green Version]
  49. Park, H.; Song, M.; Shin, K.S. Deep learning models and datasets for aspect term sentiment classification: Implementing holistic recurrent attention on target-dependent memories. Knowl. Based Syst. 2020, 187, 104825. [Google Scholar] [CrossRef]
  50. Karim, F.; Majumdar, S.; Darabi, H.; Harford, S. Multivariate LSTM-FCNs for time series classification. Neural Netw. 2019, 116, 237–245. [Google Scholar] [CrossRef] [Green Version]
  51. Farzad, A.; Hmashayekhi, H.; Hassanpour, H. A comparative performance analysis of different activation functions in LSTM networks for classification. Neural Comput. Appl. 2019, 31, 2507–2521. [Google Scholar] [CrossRef]
  52. Smyl, S. A hybrid method of exponential smoothing and recurrent neural networks for time series forecasting. Int. J. Forecast. 2020, 36, 75–85. [Google Scholar] [CrossRef]
  53. Schönfeld, J.; Kudej, M.; Smrcka, L. Financial health of enterprises introducing safeguard procedure based on bankruptcy models. J. Co. Econ. Manag. 2018, 19, 692–705. [Google Scholar] [CrossRef] [Green Version]
  54. Affes, Z.; Hentati-Kaffel, R. Forecast bankruptcy using a blend of clustering and MARS model: Case of US banks. Ann. Oper. Res. 2019, 281, 27–64. [Google Scholar] [CrossRef] [Green Version]
  55. Fu, S.; Li, H.; Zhao, G. Modelling and strategy optimisation for a kind of networked evolutionary games with memories under the bankruptcy mechanism. Int. J. Control 2017, 91, 1104–1117. [Google Scholar] [CrossRef]
  56. Sant’anna, P.H.C. Testing for Uncorrelated Residuals in Dynamic Count Models with an Application to Corporate Bankruptcy. J. Co. Econ. Stat. 2017, 35, 349–358. [Google Scholar] [CrossRef] [Green Version]
  57. Hu, Y. A multivariate grey prediction model with grey relational analysis for bankruptcy prediction problems. Soft Comput. 2020, 24, 4259–4268. [Google Scholar] [CrossRef]
  58. Nyitrai, T.; Virag, M. The effects of handling outliers on the performance of bankruptcy prediction models. Socio-Econ. Plan. Sci. 2019, 67, 34–42. [Google Scholar] [CrossRef]
  59. Kubenka, M.; Myskova, R. Obvious and hidden features of corporate default in bankruptcy models. J. Bus. Econ. Manag. 2019, 20, 368–383. [Google Scholar] [CrossRef] [Green Version]
  60. Zhou, L.; Lai, K.K. AdaBoost models for corporate bankruptcy prediction with missing data. Comput. Econ. 2017, 50, 69–94. [Google Scholar] [CrossRef]
  61. Kim, A.; Yang, Y.; Lessmann, S.; Ma, T.; Sung, M.; Johnson, J.E.V. Can deep learning predict risky retail investors? A case study in financial risk behavior forecasting. Eur. J. Oper. Res. 2020, 283, 217–234. [Google Scholar] [CrossRef] [Green Version]
  62. Koudjonou, K.M.; Rout, M. A stateless deep learning framework to predict net asset value. Neural Comput. Appl. 2020, 32, 1–19. [Google Scholar] [CrossRef]
  63. Mai, F.; Tian, S.; Lee, C.; Ma, L. Deep learning models for bankruptcy prediction using textual disclosures. Eur. J. Oper. Res. 2019, 274, 743–758. [Google Scholar] [CrossRef]
  64. Mihalovic, M. Performance comparison of multiple discriminant analysis and Logit models in bankruptcy prediction. Econ. Sociol. 2016, 9, 101. [Google Scholar] [CrossRef] [PubMed]
  65. Lin, T. A cross model study of corporate financial distress prediction in Taiwan: Multiple discriminant analysis, logit, probit and neural networks models. Neurocomputing 2009, 72, 3507–3516. [Google Scholar] [CrossRef]
  66. Unvan, Y.A.; Tatlidil, H. A comparative analysis of Turkish bank failures using logit, probit and discriminant models. Philippines 2013, 22, 281–302. [Google Scholar]
  67. Lunacek, J. Selection of bankruptcy prediction model for the construction industry—A case study from the Czech Republic. In Proceedings of the 25th International Company Information Management Association Conference—Innovation Vision 2020: From Regional Development Sustainability to Global Economic Growth, IBIMA, Amsterdam, The Netherlands, 7–8 May 2015; pp. 2627–2637. [Google Scholar]
  68. Iturriaga, F.J.L.; Sanz, I.P. Bankruptcy visualization and prediction using neural networks: A study of US commercial banks. Expert Syst. Appl. 2015, 42, 2857–2869. [Google Scholar] [CrossRef]
  69. Bateni, L.; Asghari, F. Bankruptcy prediction using logit and genetic algorithm models: A comparative analysis. Comput. Econ. Comput. Econ. 2020, 55, 335–348. [Google Scholar] [CrossRef]
  70. Altman, E.I. Financial ratios, discriminant analysis and the prediction of corporate bankruptcy. J. Financ. 1968, 23, 589–609. [Google Scholar] [CrossRef]
  71. Altman, E.I.; Haldeman, R.G.; Narayanan, P. ZETA analysis: A new model to identify bankruptcy risk of corporations. J. Bank. Financ. 1977, 1, 29–51. [Google Scholar] [CrossRef]
Figure 1. Structure of neural network (NN) ID 1 (14-940-Tanh-Tanh-2-1).
Figure 1. Structure of neural network (NN) ID 1 (14-940-Tanh-Tanh-2-1).
Sustainability 12 07529 g001
Figure 2. Confusion matrix for the training dataset.
Figure 2. Confusion matrix for the training dataset.
Sustainability 12 07529 g002
Figure 3. Confusion matrix for the testing dataset.
Figure 3. Confusion matrix for the testing dataset.
Sustainability 12 07529 g003
Table 1. Statistical characteristics of the datasets.
Table 1. Statistical characteristics of the datasets.
MinimumMaximumMeanStandard DeviationMinimumMaximumMeanStandard Deviation
Total Assets06,963,568101,794.3421,183.6062,924,684193,629.32,402,017
Fixed Assets−19,0654,558,81649,269.92251,519.4030,832,57690,484.691,204,674
Current Assets02,391,85551,801.29182,982.9032,066,562102,624.71,222,082
Short Term Receivables−1271,947,96424,162.22106,237.4027,683,66865,334.361,039,648
Borrowed Capital05,024,08950,269.76256,551.2−219,605,51361,351.15469,600.1
Short Term Liabilities04,906,38231,905.76204,347.7−967,970,38641,665.4366,137.2
Added Value−7788824,34526,734.3976,957.12−20,2843,457,58330,516.86167,441.7
Personnel Costs0463,68517,940.9246,808.62−602,665,23319,125.81111,284.5
Operating Result−150,310852,9917316.55644,552.34−847,530914,0634213.2763,193.46
Interest Payable0187,218744.45547525.506052,798494.32652912.81
Economic Result for Accounting Period (+/−)−245,135700,0926354.04143,196.58−645,493673,1795049.78653,302.32
Source: Own processing according to Albertina database.
Table 2. NN with best performance parameters.
Table 2. NN with best performance parameters.
ID NNNeural NetworkTraining PerformanceTest Performance
Source: Authors.
Table 3. Inner structure of the first long short-term memory (LSTM) layer of NN ID 1.
Table 3. Inner structure of the first long short-term memory (LSTM) layer of NN ID 1.
FieldType of OutputSize of Output
Input gate input weightsmatrix940 × 14
Input gate state weightsmatrix940 × 940
Input gate biasesvector940
Output gate input weightsmatrix940 × 14
Output gate state weightsmatrix940 × 940
Output gate biasesvector940
Forget gate input weightsmatrix940 × 14
Forget gate state weightsmatrix940 × 940
Forget gate biasesvector940
Memory gate input weightsmatrix940 × 14
Memory gate state weightsmatrix940 × 940
Memory gate biasesvector940
Source: Authors.
Table 4. Inner structure of the second LSTM layer of NN ID 1.
Table 4. Inner structure of the second LSTM layer of NN ID 1.
FieldType of OutputSize of Output
Input gate input weightsmatrix2 × 940
Input gate state weightsmatrix2 × 2
Input gate biasesvector2
Output gate input weightsmatrix2 × 940
Output gate state weightsmatrix2 × 2
Output gate biasesvector2
Forget gate input weightsmatrix2 × 940
Forget gate state weightsmatrix2 × 2
Forget gate biasesvector2
Memory gate input weightsmatrix2 × 940
Memory gate state weightsmatrix2 × 2
Memory gate biasesvector2
Source: Authors.

Share and Cite

MDPI and ACS Style

Vochozka, M.; Vrbka, J.; Suler, P. Bankruptcy or Success? The Effective Prediction of a Company’s Financial Development Using LSTM. Sustainability 2020, 12, 7529.

AMA Style

Vochozka M, Vrbka J, Suler P. Bankruptcy or Success? The Effective Prediction of a Company’s Financial Development Using LSTM. Sustainability. 2020; 12(18):7529.

Chicago/Turabian Style

Vochozka, Marek, Jaromir Vrbka, and Petr Suler. 2020. "Bankruptcy or Success? The Effective Prediction of a Company’s Financial Development Using LSTM" Sustainability 12, no. 18: 7529.

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop