https://doi.org/10.31449/inf.v47i8.4884 Informatica 47 (2023) 153–160 153 Research on Financial Risk Prediction and Prevention for Small and Medium-Sized Enterprises - Based on a Neural Network Xiaohui Wang School of Accounting and Finance, Wuxi Vocational Institute of Commerce, Wuxi, Jiangsu 214153, China E-mail: aoxihui728824@yeah.net Keywords: small and medium-sized enterprises, financial risk, neural network, risk prediction Received: May 24, 2023 For companies, timely and accurate risk prediction plays an an essential role in sustaining business growth. In this paper, firstly, the financial risk of small and medium-sized enterprises (SMEs) was simply analyzed. Some financial indicators were selected, and then some of the indicators were eliminated by Mann-Whitney U test and Pearson test. For risk prediction, an improved sparrow search algorithm-back- propagation neural network (ISSA-BPNN) method was designed by optimizing the BPNN with the piecewise linear chaotic map (PWLCM)-improved SSA. Experiments were performed on 82 special treatment (ST) enterprises and 164 non-ST enterprises. The results showed that the BPNN had higher accuracy in risk prediction than methods such as Fisher discriminant analysis; the optimization of the ISSA for the BPNN was reliable as the accuracy and F1 value of the ISSA-BPNN method were 0.9834 and 0.9425, respectively; the prediction was wrong for only one sample out of 20 randomly selected samples. The results demonstrate the reliability and practical applicability of the ISSA-BPNN method. Povzetek: Prispevek analizira finančne nevarnosti za mala in srednja podjetja z uporabo nevronskih mrež. ISSA-BPNN metoda, optimizirana z PWLCM, je v testih pokazala visoko zanesljivost in natančnost. 1 Introduction With the fast advancement of the economy and the continuous improvement of the capital market, it brings opportunities for the development of enterprises, but also new risks and difficulties. For business managers, how to ensure the survival of their enterprises has become a primary issue. Financial risks have a significant impact on the survival of enterprises [1]. The emergence of financial risks will not only seriously affect the future development and survival of enterprises, but also affect the stakeholders of enterprises and even the entire capital market. Therefore, it is particularly important to predict and prevent corporate risks. With the advancement of computers and big data, many new methods for predicting and preventing financial risks have been developed [2]. In this paper, a back-propagation neural network (BPNN)- based prediction method was designed for the financial risk of SMEs, and the effectiveness of the method was proved by experimental analysis. This work provides theoretical support to further improve the management of financial risks and promote the smooth development of enterprises. 2 Related works Studies on the prediction and prevention of financial risks are listed in the following table. Table 1: Literature list of related work Literat ure Indicators Methods Results Ptak- Chmiel ewska et al. [3] Share of net financial surplus in total liabilities, capital ratio, inventory turnover, etc. Gradient boosting, logistic regression , decision trees and neural networks The logistic regression model works best for bankruptcy risk prediction; the use of non- financial indicators has an improving effect on the results of all models. Jaki et al. [4] Relative market value added, cash flow return on sale, etc. Discrimin ant function The market measures were characterized by the highest usefulness level in explaining the bankruptcy risks of the studied companies; a change in the proportion of the division of objects to the learning and testing 154 Informatica 47 (2023) 153–160 X. Wang community at the stage of building and verifying the predictive effectiveness of the discriminant models affects both the general and partial predictive effectiveness of the constructed models Chi et al. [5] None Four traditional and sixteen hybrid models by combinin g conventio nal and modern artificial intelligen ce methods In the LR/MLP hybrid model, the inclusion of LR develops the interpretability and cross- validation capacity of the approach, whereas the use of MLP boosts the prediction ability of the planned method. Ragab et al. [6] Five financial variables and thirteen non- financial variables related to governan ce Logistic regression The results showed that the model with financial variables had a prediction accuracy of 91.7%, whereas models with a combination of financial and non-financial variables related to governance predict with comparatively better accuracy of 92.7 and 93.6%. 3 Financial risk forecast indicators for SMEs 3.1 SME financial risk In order to encourage the development of SMEs, the Shenzhen Stock Exchange has established the SME gathering board. These enterprises have good profitability, high revenue growth rate, and active trading, so they are an important part of the capital market. As of April 2021, the number of listed companies in the SME gathering board reached 1,004, and the total turnover reached 160.187 billion yuan. SMEs are relatively large in number and play an important role in promoting employment and stabilizing society, so it is important for SMEs to predict and prevent financial risks. An effective prediction model is helpful for enterprise managers to identify risks early and avoid financial crises, and it can also help market regulators to monitor enterprise risks and stabilize capital markets. The financial risk of an enterprise can be caused by internal mismanagement, personnel changes, uneven distribution, etc. It may also be affected by market turmoil, natural disasters, etc. When an enterprise has financial risk, there are three specific manifestations: (1) loss, the enterprise's poor operating conditions, reduced profitability, to a loss; (2) repayment: the enterprise's inability to repay its debts; (3) bankruptcy: the execution of bankruptcy liquidation. Financial risk prediction means selecting some financial indicators with the help of technologies such as big data to understand whether there is a possibility of financial risk by analyzing the financial data of enterprises, helping enterprise managers to improve their risk perception and minimize the loss caused by financial risk. For enterprises with abnormal financial status, the Securities and Exchange Commission will add special treatment (ST) or ST* before the stock code; therefore, enterprises marked with ST can be studied as enterprises with financial risks. The research data of this paper was obtained from the GTA database. Eighty-two ST enterprises were selected, and 164 normal enterprises (non-financial industry) were randomly selected as non- ST samples. The financial statements of these samples were collected as the experimental data of this paper. Some ST and non-ST enterprises are shown in Table 2. Table 2: Some sample companies. Stock code 002042 Huafu Fashion 002048 Ningbo Huaxiang 002271 Oriental Yuhong 002299 Sunner Development 002352 SF Holding 002422 Kelun Pharmaceutical 002563 Semir Garment 002739 Wanda Film 002788 Luyan Pharmaceutical 002841 CVTE 002052 *ST Coship 002076 *ST CNlight 002089 *ST Xinhai 002113 ST Tianrun 002200 ST Jiaotou 002259 ST Shengda 002499 *ST Kelin 002535 ST Linzhou Heavy 002569 ST Busen Research on Financial Risk Prediction and Prevention for Small… Informatica 47 (2023) 153–160 155 002700 ST Haoyuan 3.2 Indicator selection Before forecasting financial risk, it is first necessary to select appropriate indicators. Based on the existing literature, the following aspects are considered in this paper. (1) Debt service level: An enterprise that is able to repay principal and interest on time indicates that it is in good health and can give creditors the confidence to continue to invest. (2) Profit level: It refers to the profit level of an enterprise by selling products over a period of time; the higher the profit level, the better the efficiency of the enterprise; (3) Cash flow level: It can reflect the actual income of an enterprise. (4) Operating level: It can reflect the efficiency of an enterprise's capital use and its re-production capacity. (5) Development level: It refers to the future development potential of an enterprise. A higher development level of an enterprise indicates that there are fewer problems within the enterprise and lower possibility of risks. Based on the above content, the financial indicators selected for this paper are as follows. Table 3: Financial indicators. Category Number Indicator Debt service level X1 Current ratio X2 Quick ratio X3 Asset-liability ratio X4 Equity ratio X5 Times-interest-earned ratio Profit level X6 Earnings per share X7 Return on assets X8 Operating profit ratio X9 Return on total assets X10 Ratio of profits to cost X11 Net interest rate on total assets X12 Net interest rate on fixed assets Cash flow level X13 Cash content of operating income X14 Cash content of net profit Operating level X15 Business cycle X16 Inventory turnover rate X17 Accounts receivable turnover ratio X18 Total assets turnover ratio Development level X19 Total assets growth rate X20 Growth rate of fixed assets X21 Operating income growth rate 3.3 Indicator processing Table 2 contains 21 indicators that need to be screened again in order to improve the effectiveness of the subsequent prediction. First, the Mann-Whitney U test [7] was performed on the data of ST and non-ST enterprises, and the results are shown in Table 4. Table 4: Results of the Mann-Whitney U test (bolded: p > 0.05). Number Indicator P value X1 Current ratio 0.000 X2 Quick ratio 0.000 X3 Asset-liability ratio 0.000 X4 Equity ratio 0.000 X5 Times-interest-earned ratio 0.000 X6 Earnings per share 0.000 X7 Return on assets 0.312 X8 Operating profit ratio 0.000 X9 Return on total assets 0.267 X10 Ratio of profits to cost 0.000 X11 Net interest rate on total assets 0.000 X12 Net interest rate on fixed assets 0.378 X13 Cash content of operating income 0.055 X14 Cash content of net profit 0.000 X15 Business cycle 0.314 X16 Inventory turnover rate 0.000 X17 Accounts receivable turnover ratio 0.411 X18 Total assets turnover ratio 0.002 X19 Total assets growth rate 0.217 X20 Growth rate of fixed assets 0.126 X21 Operating income growth rate 0.075 According to Table 3, p > 0.05 was used as the criterion to exclude the indicators that did not differ significantly between ST and non-ST enterprises, and the indicators with p < 0.05 were retained, totaling 12. Then, the Pearson correlation coefficient test [8] was performed on these 12 indicators, and the results are shown in Table 5. Table 5: Pearson correlation coefficient test results (bolded: small correlation coefficient). Number Indicator P value X1 Current ratio 0.007 X2 Quick ratio 0.009 X3 Asset-liability ratio 0.055 X4 Equity ratio 0.006 X5 Times-interest-earned ratio -0.116 X6 Earnings per share -0.222 X8 Operating profit ratio -0.125 X10 Ratio of profits to cost 0.002 X11 Net interest rate on total assets -0.123 156 Informatica 47 (2023) 153–160 X. Wang X14 Cash content of net profit -0.068 X16 Inventory turnover rate 0.312 X18 Total assets turnover ratio 0.008 According to Table 4, the correlation coefficients of X1, X2, X4, X10, and X18 were small, and these indicators were considered as less relevant to the existence of financial risk in the enterprise, so they were eliminated. The final indicators obtained for forecasting are shown in Table 6. Table 6: Financial indicators used for risk forecasting. Category Number Indicators Debt service level X1 Asset-liability ratio X2 Times-interest-earned ratio Profit level X3 Earnings per share X4 Operating profit ratio X5 Return on total assets Cash flow kevel X6 Cash content of net profit Operating level X7 Inventory turnover rate 4 Neural network-based prediction methods The neural network method with good self-learning ability and fault tolerance can effectively process a large amount of data, and has mature applications in industrial control and medical health [9]. Therefore, this paper designed a financial risk prediction of SMEs based on neural networks. BPNN is one of the most widely used neural network methods [10]. Taking a simple three-layer network as an example, the model parameters are assumed to be as follows. Input layer vector: 𝑋 𝑖 =(𝑥 1 ,𝑥 2 ,⋯,𝑥 𝑛 ) Hidden-layer output vector: 𝑌 𝑗 =(𝑦 1 ,𝑦 2 ,⋯,𝑦 𝑚 ) Output-layer vector: 𝑂 𝑘 =(𝑜 1 ,𝑜 2 ,⋯,𝑜 𝑙 ) Input-layer and hidden-layer weights: 𝑤 𝑖𝑗 Hidden-layer and output-layer weights: 𝑣 𝑗𝑘 Then, the output of the implicit layer is written as: 𝑌 𝑗 =𝑓 (∑ 𝑤 𝑖𝑗 𝑛 𝑖 =1 𝑋 𝑖 −𝑏 𝑗 ) , where 𝑏 𝑗 is the hidden-layer threshold. The output of the BPNN is written as: 𝑂 𝑘 = 𝑓 (∑ 𝑣 𝑗𝑘 𝑌 𝑗 −𝑏 𝑘 𝑚 𝑗 =1 ) , where 𝑏 𝑘 is the output-layer threshold. Let the expected output vector be 𝑑 𝑘 , then the error of the model is written as: 𝐸 𝑘 = 1 2 ∑ (𝑑 𝑘 −𝑜 𝑘 ) 2 𝑙 𝑘 =1 . The goal of training is to make the error satisfy the accuracy. The BPNN trains the model by back- propagating the error and continuously adjusting the weights and thresholds. However, the BPNN has the characteristics of slow convergence and easy to fall into local optimum. In this paper, an improved sparrow search algorithm (ISSA) was designed to optimize it, and the ISSA-BPNN method is obtained. The SSA is an algorithm based on the foraging behavior of sparrows [11], which has strong local search ability and fast convergence. Suppose that in a D- dimensional space, there are 𝑛 sparrows, their initial positions is written as: 𝑋 ={𝑥 1,1 ,𝑥 1,2 ,⋯,𝑥 𝑛 ,𝑑 } . In the population, the sparrow searching for food is called the discoverer. At time 𝑡 , its position update is written as: 𝑥 𝑖 ,𝑗 (𝑡 +1)={ 𝑥 𝑖 ,𝑗 (𝑡 )∙𝑒𝑥𝑝 (− 𝑖 𝛼 𝑡 𝑚𝑎𝑥 ),𝑖𝑓 𝑅 <𝑆𝑇 𝑥 𝑖 ,𝑗 (𝑡 )+𝑄𝐿 ,𝑖𝑓 𝑅 ≥𝑆𝑇 , (1) where: α: a random number in [0,1], 𝑅 : an alert value in [0,1], 𝑆𝑇 : a safety value in [0.5,1], 𝑄 : a random number satisfying a normal distribution, 𝐿 : a 1×𝑑 matrix. The followers will follow the discover to compete for food, and if they cannot compete for enough food, they will move to other regions to search. The process is written as: 𝑥 𝑖 ,𝑗 (𝑡 +1)= { 𝑄𝑒𝑥𝑝 ( 𝑥 𝑤 −𝑥 𝑖 ,𝑗 (𝑡 ) 𝑖 2 ),𝑖𝑓 𝑖 > 𝑛 2 𝑥 𝑖 ,𝑗 (𝑡 )+|𝑥 𝑖 ,𝑗 (𝑡 )−𝑥 𝑝 (𝑡 )|∙𝐴 + ∙𝐿 ,𝑜 𝑡 ℎ𝑒𝑟𝑤𝑖𝑠𝑒 , where: 𝑥 𝑤 : the location with the worst fitness, 𝑥 𝑝 : the location with the best fitness, 𝐴 + : 𝐴 + =𝐴 𝑇 (𝐴 𝐴𝑇 ) −1 (𝐴 is a 1×𝑑 matrix, and 𝐴 𝑇 is the transposition of 𝐴 ). When danger is detected, individuals on the outside of the population will move closer to the inside, and individuals on the inside will move closer to their peers. The process is written as: 𝑥 𝑖 ,𝑗 (𝑡 +1)= { 𝑥 𝑏 (𝑡 )+𝛽 ∙|𝑥 𝑖 ,𝑗 (𝑡 )−𝑥 𝑏 (𝑡 )|,𝑖𝑓 𝑓 𝑖 ≠𝑓 𝑏 𝑥 𝑖 ,𝑗 (𝑡 )+𝐾 ∙ |𝑥 𝑖 ,𝑗 (𝑡 )−𝑥 𝑏 (𝑡 )| (𝑓 𝑖 −𝑓 𝑤 ) ,𝑖𝑓 𝑓 𝑖 =𝑓 𝑏 , (2) where: x b : the globally optimal position 𝑓 𝑖 : the fitness of sparrow i, 𝑓 𝑏 : current best fitness, 𝑓 𝑤 : current worst fitness, 𝛽 : a random number satisfying a normal distribution, 𝐾 : a random number in [-1,1] to control the direction of the sparrow's movement. When 𝑓 𝑖 ≠𝑓 𝑏 , it is the process of sparrow approaching from outside to inside, and when 𝑓 𝑖 =𝑓 𝑏 , it is the process of sparrows approaching from the inside to their companions. The SSA obtains the optimal solution by continuously updating the sparrow positions. However, the SSA uses random generation for population initialization, which is not conducive to the diversity of the population, so the SSA was improved by combining chaos mapping. PWLCM [12] was used to achieve the initialization of the population. The formula is: x i = { x i−1 δ ,0≤x i−1 <δ (x i−1 −δ) 0.5−δ ,δ≤x i−1 <0.5 0,x i−1 =0.5 F(1−x i−1 ,δ),0.5