Big data analysis and empirical research on the financing and investment decision of companies after COVID-19 epidemic situation based on deep learning

1 Introduction

At present, the fight against the epidemic has entered a critical period and its duration is still unknown [1]. In the short term, the first to be hit is large and small micro-enterprises, but with the spread of the epidemic, large and medium-sized listed companies with strong short-term risk resistance are bound to be hit. For example, the uncertainty of employees, the interruption of the supply chain and the decline in market demand may lead to difficulties in resuming work, production stagnation, profit decline and even losses; insufficient liquidity will bring credit and debt risks to enterprises. In the short term, large and medium-sized listed companies have a certain ability to resist “hit”, but in the medium and long term, these companies also face tremendous pressure, and once problems occur, they will have a huge impact on economic and social operations [2, 3]. All along, the development of large and medium-sized listed companies has played an important role in the stability of the entire economy and society. On the one hand, as the backbone of China’s economic development, large and medium-sized listed companies have obvious support and pulling effect on economic development; on the other hand, as the “leaders” and “wind vane” of industrial transformation and upgrading, they have a trend in the development of the industry Have a deeper and clearer understanding. The greater the capacity, the greater the responsibility. As the “ballast stone” of economic development, as well as the next step to resume production and lead growth, large and medium-sized listed companies shoulder major social responsibilities and have a long way to go. At the same time, some large and medium-sized listed companies in China have been integrated into the global industrial chain, and their healthy operation will also affect the smooth operation of the global economy. If the suspension of these enterprises will make the world economy feel chilled, it will accelerate the transfer of industrial chains from China, which will have a significant impact on the long-term development of our economy.

At the same time, we should objectively realize that the fight against the epidemic may turn into a stalemate [4–6]. Once this situation occurs, we must seriously consider the pressure on economic development and how to resume production while fighting the epidemic. Intensive and closed production may be a more realistic approach. Large and medium-sized listed companies have an advantage in this regard. Large and medium-sized listed companies will also play a backbone role in recovering growth after the outbreak. Therefore, it is especially critical to pay attention to the impact of the epidemic on large and medium-sized listed companies, especially on the investment and financing of listed companies.

2 Historical data and materials analysis

According to relevant data, in 2016, the total operating revenue of B2B e-commerce platform of relevant enterprises was about 24.55 billion yuan, and the total scale was 21.45 billion yuan as of the third quarter of 2017, reaching 7.55 billion yuan in just three months, an increase of 17.5% year-on-year. In the same quarter of the same year, the amount of online shopping reached 1.44 trillion yuan, and the proportion of B2C continued to expand. Today, with the wide spread of Internet, the development of e-commerce not only plays an important role in the adjustment of industrial economic structure and the transformation of traditional enterprises, but also has an important guiding significance for the expansion of China’s “new economy” growth point after the 19th National Congress. In the future, the development of China and even the global economy is closely related to e-commerce.

Stock was born in the capitalist countries in the early 17th century. The emergence of stock is to solve the contradiction between the shortage of capital supply and demand and the expansion of business operation, and to obtain social financing through public issuance to the market. With the continuous emergence of this kind of joint-stock company established by raising share capital, the behavior of absorbing funds by way of stock develops rapidly, which also makes the development and maturity of stock market an inevitable trend. China’s stock was born in 1984 and entered the pilot period three months later. After more than 30 years of development, the stock has become an important member of the market economy. With the improvement of consumption level and income, more and more people regard stock investment as a part of daily financial management. However, due to various uncertainties in the stock market, it is easy to bring risks to the vast majority of ordinary investors, especially “retail investors”; at the same time, the volatility of the stock market also directly affects the stability and development of the whole national economy. Therefore, if we can establish a decision support model of stock quality evaluation and evaluate the stock quality of e-commerce listed enterprises reasonably, it will not only help the state and CSRC to better manage and check the stock, but also provide decision support for investors.

Li Chunwei collected 16 technical index data of “Dongfeng Motor” in 432 days, used BP modeling to predict the closing price in the next few weeks, and optimized the model by selecting the optimal index combination, which improved the accuracy of stock price prediction from 50% to 67% [7]. Han Wenlei [8] used probabilistic neural network to predict Shenzhen Stock Exchange Index, Shanghai Stock Exchange Index and Shanghai Stock Exchange 180 index, with the prediction accuracy of 82%, 77% and 82%, respectively. Peng Ying used the monthly average index data of Shanghai Composite Index from November 2004 to October 2005 to predict the index of November 2005 [9]. The model combined with GM (1,1), RBF and grey system and neural network was used to predict the index. The prediction errors were 2.58, 7.26 and 0.69 respectively. The accuracy of the combined model was the highest. Wang proved that epcnn’s prediction result is better than TSK system, TSK’s prediction effect is better than BPN, and BPN is better than multiple regression analysis [10]. The error of stock prediction of the four is 0.503, 2.16, 3.86 and 6.85. Liu Lei used the deep learning model to model the five indexes of three stocks and forecast the stock fluctuation at the next moment [11]. At the same time, compared with the results of traditional BP network model, the prediction accuracy of BP model was 60.43%, while that of deep learning model was 62.12%. Wu Shaocong compares and analyzes the results of stock time series prediction by the depth LSTM model and ARIMA model [12–16] and selects a mixed model combining the two to predict the short-term rise and fall trend of 12 stocks. The average absolute error predicted by LSTM model is 0.24, the average error predicted by the conventional ARIMA model is 1.35, while the error predicted by the improved DB-ARIMA model is reduced to 0.49, and the error predicted by the mixed model is about 0.33. The result is better than the first two.

3 Deep learning

Deep learning is a kind of machine learning (ML) method. In essence, it is an ANN with multiple hidden layers. The early ANN model can’t solve the nonlinear problem. Until the discovery of BP algorithm in 1986, the nonlinear learning ability can be trained by the multi hidden layer neural network. Because of the limitation of early data scale, calculation scale and ability, the training and learning of parameters or weights of multi-layer neural network cannot achieve satisfactory results. Therefore, the development of neural network has been limited until recent years. With the rapid development of Internet and the explosion of information data, large-scale trainable data can support the learning and optimization of a large number of multi-layer neural network weights.

3.2 BP neural network algorithm

BP neural network is the simplest deep learning model. BP algorithm is called error back propagation algorithm. In addition to the forward propagation, it also modifies the weight isometrically value through back propagation, so that the network error continues to reduce until it meets the expected output. The traditional BP back propagation process is as follows:

Let the expected output be Z’_k, and define the global error between the expected output and the actual output as L: L = 1 2 ∑ k = 1 m ( z k - z k ′ ) 2(1)

Through the back propagation process, the error is expanded to the hidden layer as follows: L = 1 2 ∑ k = 1 m [ f ( k ) - z k ′ ] 2 = 1 2 ∑ k = 1 m [ f ( ∑ k = 1 t w jk y j + b k ) - z k ′ ] 2(2)

It can be seen from the above formula that the network error is a function of the weights w _ij and w _jk . Therefore, by changing the weights of neurons, the error E can be changed: Δ w ij = - e ∂ L ∂ w ij ( i = 1 . . . m , j = 1 . . . n )(3) Δ w jk = - e ∂ L ∂ w jk ( j = 1 . . . n , k = 1 . . . t )(4)

In the formula, the negative sign represents gradient descent, ɛ represents rate, ɛ ∈ (0,1).

Through the above description of the analysis of the back propagation process, we describe the process of BP algorithm as follows, and Fig. 1 is the flow chart:

OPEN IN VIEWER

Fig. 1

Algorithm flow chart.

Initialize network structure parameters, randomly set connection weight [–1,1], learning rate (0,1) and bias [0,1];

Forward transmission: transfer the data into the network from the first layer to the last layer, calculate the output value according to formula 2.11, and compare the obtained value with the expected value. If there is a deviation, carry out step (3);

In the reverse feedback process, according to the final error, it is transmitted to the first hidden layer, and the connection weight of each neuron is adjusted layer by layer by the following formula

It is calculated that the error of nodes in the same layer is δ_k, and the adjustment formula is: w jk ( 1 + t ) = w jk ( t ) + ed k o j(5)

The threshold value of the network can also be learned according to the weight correction process, assuming that the input signal is 1. Train the above process repeatedly until a satisfactory result is obtained.

Traditional stock or stock market analysis mainly evaluates stock prices, stock market fluctuations, etc. More often, it evaluates the general trend of stock price changes over a period of time. The methods include basic analysis, technical analysis, statistical analysis, etc. The traditional methods have greatly promoted the development of stock market research in the past. In recent years, with the continuous progress of computer technology, more and more methods have been applied to stock analysis, while more and more researches have been made on stocks using in-depth neural networks. It is evolved from the traditional artificial neural network. It is an artificial neural network with multiple hidden layers. The idea originates from the multi-level processing mechanism of human brain for input information. Deep neural network can learn the most significant characteristics of the original data through layer-by-layer training, which can well solve the problem of classification. Moreover, through layer-by-layer initialization, the training difficulty of neural network can be greatly reduced. The stock market is a complex dynamic system, containing a large amount of transaction data. The characteristics of neural network structure make it very suitable for analyzing stock market problems.

Input index data into the network model. Taking the 18 indicators of the largest shareholder’s shareholding ratio, listing time, 2014–2016 earnings per share, 2014–2016 dividend allotment, 2014–2016 asset liability ratio, whether to restructure in three years, number of legal proceedings, total share capital and circulating shares, issuance price, social responsibility report and key pollutant discharge units as input variables, 311 group sample data is used for training, 191 group sample data is used for testing, and the evaluation output is stock category. Table 1 shows the original data.

4.1 The establishment of LSTM neural network evaluation and prediction model

The specific steps of evaluation and prediction model modeling are as follows,

Firstly, the network structure and prediction accuracy are determined based on sample data and experience.

Generally, there are two types of sample set: training sample and test sample.

Select the appropriate network algorithm to train the model, so that the model can best fit the data of training period, this paper uses LSTM algorithm to train the network.

Test the trained network with the test set data. If the result is ideal, use the model to evaluate the test set data; otherwise, adjust the model parameters and repeat the previous operation until the ideal output is obtained.

4.1.1 Structure design of prediction model

For the neural network model, its network structure design can be said to be relatively “free”. In addition to the input and output layers need to be set according to data characteristics and objectives, the design of other values or parameters is basically based on experience, or through experiments to compare, without fixed procedures and steps. The level series of neural network is generally set according to the previous experience. A large number of studies and experiments show that the network with a small number of layers has strong generalization power, is easy to understand and extract features, and is conducive to implementation; However, the more the number of levels, the smaller the error of the network will be, which can improve the accuracy and complexity as well as increase the training time and speed of the model, which is likely to lead to over-fitting. Therefore, the number of levels can be set according to the effect to be achieved by the model. Generally, for a model with moderate data volume and high accuracy, a three-layer network can be selected. Andrey Nikolaevich Kolmogorov has proved that networks with more than three layers can fit any non-linear function and are also the networks most used in multi-layer networks. In this paper, the 10-layer depth LSTM network is used to build the stock quality evaluation prediction model.

The input and output layer parameters only need to be set according to the characteristics of index data. The input node matches the feature dimension, and the output node matches the target dimension. The sample contains 502 stocks, 18 data for each stock. Therefore, the number of input layer nodes m is set to 18, the last stock type is the format data of output data, and the number of output layer nodes n is set to 3.

The setting of the number of nodes in the hidden layer is the key and difficult point of the whole network structure design. There is no unified analytical expression to express it. The number of nodes has an important impact on the whole model effect. On the one hand, it affects the calculation ability of the network, on the other hand, it also affects the ability to approach the objective function. The more the number of nodes in the middle layer, the stronger the representational ability, which shows that the data storage ability of the network is strong, each node processes less data, and the accuracy is high; But once the number of neurons exceeds a certain degree, it will not only increase the learning time, but also affect the realization of the computer, similar to the problem of network level progression setting. Therefore, specific data need to be specifically analyzed and selected according to the characteristics of experimental data. Researchers usually determine the number of neurons according to previous experience or through continuous experiments. The commonly used determination method is called trial and error method, that is, several optional values are preset, then GridSearch algorithm is used to traverse the preset optional values, and the optimal value is selected by comparing the evaluation results. The trial-and-error method requires several basic formulas, and the calculated value can be used as the initial value of the number of hidden layer nodes. The specific formula is a s follows: s = log₂ m; s = mn; s = m + n + a (s, m, n are the number of hidden layer, input and output layer nodes respectively, a is the constant; the number of input nodes finally selected through the experiment is 18; the number of hidden layer nodes is 50; the number of output layer nodes is 3.

4.1.2 Selection of activation function

From the second chapter, we have a general understanding of the types of activation functions. S and T are S-type functions when the neural network is small. Tanh performs better than Sigmoid function, and Relu is the most widely used activation function.

Different from sigmoid function and tanh function, relu function is characterized by gradient unsaturation. The gradient calculation formula is: 1 x > 0. Therefore, in Backpropagation, the parameters can be adjusted quickly and the gradient vanishing process can be alleviated, so the calculation speed is fast. In the forward training, Relu does not need to calculate the index but only set the threshold value. This function can speed up the network computing speed. Therefore, Relu function is used in this paper.

Relu activation function (The Rectified Linear Unit) expression is: f ( x ) = max ( 0 , x )(6)

Figure 2 below shows the basic graph of Relu function.

OPEN IN VIEWER

Fig. 2

Relu activation function.

4.1.3 Other relevant parameter settings

There is great flexibility in setting the learning step size. In this paper, LSTM neural network algorithm is adopted, and the learning step length is updated automatically in training, so the initial value can be set according to previous experience. In this paper, the learning step is initialized to 0.001.

The error accuracy should be set according to the specific situation of the model, whether it requires high accuracy or strong generalization ability. The former can set the accuracy smaller, the latter can be appropriately larger, and the accuracy can be adjusted according to the results in the model training; in addition, the number of sample data and model neurons have an impact on the accuracy setting of the error. The more the training data, the larger the initial error, and the lower the training accuracy of the model. Because the larger the scale of the original data, the more variables in the model, the more variables lead to large errors. Set the error accuracy here to 0.01. The termination conditions of model training are mainly designed according to two aspects: first, when the model error meets the training accuracy requirements; second, when the model training time is too long or too many times, the process should be terminated. However, the first case shows that the training of network model is successful, while the second case shows that the artificial termination of network training due to time and number of times shows that the model does not meet the requirements of precision setting, so the model should be trained again. The key factor of unsuccessful network training is the unreasonable setting of initial connection weight, which makes the minimum value of model training deviate from the initial state greatly, or the excessive extreme value in the model may lead to too long training time. In this paper, the maximum number of training is set to 1000. When this value is exceeded, the time is too long and should be stopped.

5 Simulation and implementation of evaluation results

6 Analysis of experimental results

According to the training results of the model test set, the model fitting error is about 0.08, which proves that the model is effective. The relative average error between the output value and the actual evaluation result is about 0.25, which improves the accuracy compared with the previous stock prediction. Due to many uncontrollable factors in the stock market, the general stock evaluation can not reach a high precision.

The results data after discretization is shown in Table 3.

From the above results, we can see that the average error of stock market prediction by scholars is about 0.36. The accuracy of the experimental model results is improved, so the experimental results show the accuracy and rationality of the stock quality evaluation model. Compared with the traditional single evaluation index, the index system established in this paper can basically reflect the influencing factors of all aspects of stock quality. Through the analysis of the impact indicators of stock quality, the quality, accuracy and comprehensiveness of the stock are improved to some extent. Investors do not need to carry out a series of complex calculations. They can see the quality of the stock intuitively according to the indicator system model, so as to consider investment. In addition, when using ANN to build evaluation model, attention should be paid to:

The rise of domestic stock market is late, the regularity is not obvious enough, there is no sound market system to guarantee, and the securities investment market needs sound mechanism and standard rules to ensure the healthy operation. But China’s stock market is affected by many uncontrollable factors. At that time, political, economic and social factors will cause stock market fluctuations, especially policy changes or emergencies. At present, the evaluation and prediction model in this paper cannot solve these problems, and it needs further study by scholars in the future.

For a long time, the stock market attracts a large number of investors with its high risk and high reward. At the same time, the stock market, as the “alarm” of market economy, attracts the attention of the government, investors and the general public. Therefore, the research on it has great theoretical and Application value. Since the birth of the stock market, there has been no lack of research on its fluctuation law and investment potential. There are many traditional research methods, such as fundamentals, technical level, sequence analysis, etc. because of the complexity and volatility of the stock market, it is difficult to build a simple and accurate model.

This paper expounds the structural design of the network model, including the design of the network hierarchy, the selection of the number of neuron nodes in each hierarchy, the determination of initial weights and the setting of other relevant parameters, and finally designs the structure of the stock quality evaluation model. The python language is used to show the results of data pre-processing and the training process of sample data. Through the training of the data processed by the two methods, the optimal network classification model is selected. Finally, the experimental results are analyzed and summarized, and the decision-making suggestions are given.