Research on the Construction of Risk Early Warning Mechanism in Enterprise Financial Management

Financial data acquisition can use batch acquisition, real-time acquisition, manual acquisition and other ways to obtain data.

The data acquisition layer not only focuses on data collection, but also needs to pre-process the data to ensure data quality. Common types of data preprocessing include:

Missing value processing: filling in missing values or deleting incomplete data.

Outlier detection: detecting outliers in the data through methods such as box plots and standard deviation.

Data standardisation and normalisation: ensuring comparability of data with different measures, e.g. normalising all financial indicators to the [0, 1] interval.

Noisy Data Removal: Remove noise from data through smoothing and clustering methods.

Commonly used preprocessing tools: Pandas and Scikit-learn libraries in Python, data cleaning packages in R, preprocessing toolboxes in MATLAB, etc.

The data collection layer needs to store and manage the collected financial data. In this study, MySQL is proposed to be used for storing the structured data.

During data storage, it is ensured that the data is not lost during the collection and storage process. Ensure that the data is secured by allowing only authorised users to access sensitive data. Protect the privacy of financial data by encrypting the data during transmission and storage.

By scientifically designing the financial risk data collection layer, companies can ensure the high quality and timeliness of financial data, providing a solid data foundation for subsequent financial risk early warning and response. This not only improves the predictive accuracy of the risk warning model, but also effectively reduces the financial risk of the enterprise.

The Z-value model is a statistical model used for financial risk assessment and is widely used to predict the risk of bankruptcy of an enterprise. Proposed by American economist Edward Altman in 1968, the Z-value model was originally developed to help identify potentially financially distressed companies by analysing the financial data of US manufacturing companies. The model generates a Z-value through a weighted combination of multiple financial indicators, with the lower the Z-value, the higher the likelihood of a company's bankruptcy.The Z-value model is one of the earliest bankruptcy prediction models, and one of the most classic and widely used financial early warning tools. Depending on the size of the Z-value, a company can be categorised into safe, warning and insolvency zones, thus helping investors, managers and banks, among others, to assess the financial health of a company.

The dataset of Z-value model is shown in equation (1): (1){ (X,Z)∣Xi∈□n,i=1,2,…,5 }

The prediction function of the Z-value model is shown in the following equation: (2)Z=1.2×X1+1.4×X2+3.3×X3+0.6×X4+0.999×X5

Where X₁ = total working capital assets/total assets, X₂ = retained earnings/total assets, X₃ = EBITDA/total assets, X₄ = market value of shareholders' equity/total liabilities, and X₅ = sales revenue/total assets. When Z>2.99, the enterprise is financially sound, when 1.81<Z<2.99, the enterprise is financially risky, and when Z<1.81, the enterprise is financially high risk.

The goal of early warning in corporate financial risk analysis is to identify and predict financial crises or other major problems that a firm may face in the future so that appropriate interventions can be taken to avoid or mitigate potential financial losses. In this context, Logistic regression model, as a dichotomous statistical learning method, is widely used in the study of corporate financial risk early warning. It analyses a firm's financial data to predict whether it is likely to experience bankruptcy, default or other financial risk events. Logistic Regression (LR) is a statistical method used to solve binary classification problems. Unlike linear regression, the output of logistic regression is a probability value that indicates the likelihood of an event occurring. It is often used to predict whether an event will occur (e.g., whether a firm will experience a financial crisis) and outputs the probability of the event occurring based on given financial characteristics.

In financial risk early warning, the objective is to predict whether a firm will experience a financial risk event based on a set of financial indicators (e.g., profitability, current ratio, debt ratio, etc.) Logistic regression models help predict whether a firm is in a high-risk situation by fitting the relationship between input characteristics in the financial data and the probability of the firm experiencing a financial crisis.

Logistic regression models are based on the conversion of probabilities into log odds form as shown in equation (3): (3)ln(P1−P)=β0+β1X1+β2X2+…+βnXn

The prediction function is shown in equation (4): (4)P(Y=1∣X)=11+e−(β0+β1X1+β2X2+…+βnXn)

Where X_i is the financial indicator and β_i is the regression coefficient. β_i can be estimated by maximising the log-likelihood function as shown in the following equation: (5)L(β)=∑i=1m[ yilnP(Y=1∣Xi)+(1−yi)ln(1−P(Y=1∣Xi)) ]

The optimal regression coefficients can be found by maximising L(β).

Support vector machines (SVMs) are another way to build financial warning models by constructing hyperplanes for classification. Financial data can be classified into “healthy” and “crisis” categories using support vector machines. Generally, in financial risk early warning research, the data set of support vector machine is shown in equation (6): (6){ (xi,yi)∣xi∈□n,yi∈{−1,1},i=1,2,…,m }

The objective of a support vector machine is to find a hyperplane to maximise the classification interval and the optimisation problem is: (7)min12‖w‖2

The constraints are: (8)yi(wTxi+b)≥1,∀i

The classification decision function of the support vector machine is: (9)f(X)=wTX+b

where w is the weight vector and b is the bias. According to the sign of f(X) can determine whether the enterprise has financial risk.

In the field of enterprise financial risk analysis and early warning, the use of advanced machine learning technology has become a prevalent trend for the prediction of enterprise financial status. Artificial neural network, as a powerful machine learning tool, its ability in nonlinear pattern recognition, feature learning and data prediction makes it widely used in enterprise financial risk prediction. By simulating the working principle of biological neural networks, artificial neural networks are able to automatically learn and perform complex pattern recognition from corporate financial data, identify potential financial risks, and provide effective early warning signals for decision makers.

Artificial neural network is a computational model that performs information processing by simulating a biological nervous system (e.g., the way neurons are connected in the human brain). It performs data processing and pattern recognition through connections and transmission between neurons at multiple levels. Compared with traditional statistical methods (e.g. regression analysis), artificial neural network has strong adaptive learning ability, and is able to automatically extract information from a large amount of data without explicit rules, as well as make predictions and classifications.

In financial risk analysis and early warning, the data set of artificial neural network is shown in equation (10): (10){ (xi,yi)∣xi∈□n,yi∈{−1,1},i=1,2,…,m }

Where x_i. is the financial indicators of the enterprise, such as current ratio, gearing ratio, etc. y_i. is the output, y_i=1 when there is financial risk and y_i=0 when there is no financial risk.

The prediction function of artificial neural network is shown in the following equation: (11)y^=f(∑i=1nwiXi+b)

Where f is the activation function (Sigmoid function), w_i. is the weights, X_i is the input features and b is the bias. When the prediction function predicts the result closer to 1 then the crisis probability is greater.

The data transfer function between the input layer and the hidden layer is: (12)zj(1)=∑i=1nwji(1)xi+bj(1),hj(1)=f(zj(1)),j=1,2,…,H

where wji(1) is the weight between the input layer and the hidden layer, bj(1) is the bias between the input layer and the hidden layer. zj(1) is the weighted input of the jth neuron of the hidden layer, f is the activation function, hj(1) is the output of the jth neuron of the hidden layer, and H is the number of neurons in the hidden layer.

The data transfer function between the input layer and the hidden layer is: (13)zk(m+1)=∑i=1Hwkj(m+1)hj(m)+bk(m+1),y^k=f(zk(m+1))

where wkj(m+1) is the weight between the hidden layer and the output layer, and bk(m+1) is the bias between the hidden layer and the output layer. zk(m+1) is the weighted input of the kth neuron of the output layer, and y^k is the predicted value of the output layer.

To minimise the loss function of the artificial neural network, gradient descent algorithm is used: (14)θ←θ−η∂L∂θ

where θ is a neural network parameter including weights w and bias b. η is the learning rate and ∂L∂θ is the gradient of the loss function with respect to the parameter. L is the loss function, which can be obtained from the following equation: (15)L=1m∑i=1m[ yilny^i+(1−yi)ln(1−y^i) ]

The activation function used in this study is the Sigmoid function as shown in the following equation: (16)f=11+e−z

Where z is the weighted input to the hidden or output layer of the neural network.

Since the output layer in this study has only one node, the Sigmoid activation function is used to calculate the probability of each category as shown in the following equation: (17)y^=11+e−z

If the prediction result is y^≥0.5 , then y^=1 . Conversely, y^=0 .

The fundamental objective of financial early warning is to anticipate the potential risks associated with enterprises through an analysis of their financial data. Subsequently, the financial data of a company from 2008 to 2015 are selected as the model input data for the purpose of comparing the prediction results of the Z-value model, logistic regression model, support vector machine, and artificial neural network in order to identify the most suitable financial risk analysis model for this design proposal. The financial data of a company from 2008 to 2015 are presented in Table 1.

The data in Table 1 was analysed using Models 1-4 and the predictions are shown in Table 2.

As evidenced by the results presented in Table 2, the predictive models, namely the Z-value model, logistic regression, support vector machine, and artificial neural network, have demonstrated consistent accuracy in forecasting outcomes in other years. However, the results of the prediction for the financial data of 2014 demonstrated that the Z-value model, logistic regression model and support vector machine were unable to accurately assess the risk status of the data, whereas the artificial neural network was successful in doing so. This outcome suggests that the artificial neural network is a more appropriate risk analysis model for this design proposal.

The core idea of linear regression model is to represent the relationship between independent variables and dependent variables by constructing a linear equation. In financial data analysis, there may be a certain linear relationship between many key financial indicators (e.g., gearing ratio, operating income, net profit, etc.) and a company's risk monitoring costs. Linear regression identifies this relationship and is used to predict future risk costs by fitting historical data.

The objective of a linear regression model is to identify a set of parameters that minimise the discrepancy between the predicted and actual outcomes. The equation below illustrates the prediction function of the linear regression model: (18)Cost=β0+β1X1+β2X2+…+βnXn

where X_i is the wind direction characteristic variable and β_i is the regression coefficient.

Support vector machine is a supervised learning model that is frequently employed in the context of classification, regression, and anomaly detection. It is based on statistical learning theory and performs data classification or regression prediction by constructing an optimal hyperplane. Support vector machine can not only deal with linear problems effectively, but also can be extended to non-linear problems by kernel function, so it has important applications in complex enterprise risk monitoring cost prediction. The core idea of support vector machine is to maximise the boundary or interval between classes by finding a hyperplane, which makes the data points of different classes to be classified effectively. In the context of regression problems, support vector machines are employed to optimise an objective function with the objective of minimising the discrepancy between the predicted value and the true value. Once this optimisation process has been completed, the resulting model is capable of making accurate predictions.

In enterprise risk monitoring cost prediction, the support vector machine learns and predicts the possible future risk monitoring costs by using input features (e.g., enterprise's financial status, macroeconomic data, industry dynamics, etc.). It is able to handle high-dimensional, non-linear, and noisy data, making it ideal for complex risk prediction tasks.

The prediction function of the support vector machine is shown in the following equation: (19)f(x)=〈w,φ(x)〉+b

where φ(χ) is the kernel mapping, and w and b are the hyperplane normal vector and bias term, respectively.

The objective function of the support vector machine is min(1/2)‖w‖2C∑ Ni=1ξi with the constraints y_i(wx_i + b)≥1–ξ_i, ξ_i≥0. where ξ_i is the slack variable indicating the degree of misclassification. C is the penalty coefficient for balancing the error and spacing.

In enterprise risk monitoring and management, it is crucial to accurately predict risk-related costs. This process requires consideration of a large number of data inputs such as historical financial data, market trends, industry dynamics, and business operations. Whereas conventional risk prediction methodologies typically rely on rule-based statistical models, the advent of big data and machine learning technologies has resulted in an increased utilisation of artificial neural networks, a sophisticated data-driven model, to the forecasting of enterprise risk monitoring costs. Artificial neural networks mimic the workings of biological neural systems, and are able to extract implicit patterns from complex input data, and are especially suitable for dealing with high-dimensional, non-linear and complex data relationships. In cost prediction for enterprise risk monitoring, artificial neural networks are able to make efficient cost predictions based on the historical financial status of the enterprise, industry characteristics and other relevant factors, providing scientific decision support for the enterprise.

The model selected for this programme is shown in Figure 5.

Figure 5.

Structure of artificial neural network

As illustrated in Figure 5, the model comprises the following structure:

Input layer: five nodes, corresponding to five financial features such as the number of employees, annual income, industry risk level, historical risk events and corporate credit score.

Hidden layer: two layers, each containing 10 neurons, with an activation function of ReLU.

Output layer: one node for outputting risk control costs.

The prediction function between the input layer to the hidden layer is:

ReLU function is shown in the following equation: (20)f=max(0,z)

Other formulas are detailed in Eq. (10)-Eq. (17).

Risk control cost prediction plays a crucial role in enterprise management, and enterprises need to adjust their budgets and risk response strategies based on the prediction results. Linear regression model, support vector machine and artificial neural network are commonly used risk control cost prediction models. In order to select an appropriate risk control cost prediction model for this design proposal, the prediction accuracy of these models is compared next.

The risk control cost of a company for 8 years is shown in Table 3.

The data in Table 1 was analysed using Models 1-4 and the predictions are shown in Table 4.

As evidenced by the results presented in Table 4, the linear regression model demonstrates the greatest discrepancy between the predicted and actual costs, whereas the artificial neural network exhibits the least such discrepancy. This result indicates that artificial neural network is more suitable as a risk monitoring cost prediction model for this design solution.