Enterprise financial dilemma prediction method and system based on target driven feature selection

Abstract

The invention provides an enterprise financial dilemma prediction method and system based on target-driven feature selection, a storage medium and electronic equipment, and relates to the field of integrated learning. The invention provides an integrated prediction model based on supervised feature selection for enterprise financial dilemma prediction, and models the construction process of the integrated prediction model into a combinatorial optimization model, so as to maximize the integrated prediction performance of the model and the individual prediction performance of the base model at the same time from the perspective of feature selection. Compared with a traditional model, the model provided by the invention not only can select a group of optimal feature subsets for each base model, but also provides a novel combination mechanism, and can maximize the performance of feature selection in the aspect of improving the overall performance of the model. In addition, the model increases the attention to the base model in the aspect of avoiding recognizing a dilemma enterprise as a healthy enterprise when setting a specific optimization target, and also allows the integrated prediction performance of the model to be continuously optimized by continuously increasing the number of the base models.

Description

Technical Field

[0001] The present invention relates to the field of integrated learning, and in particular to a method, system, storage medium and electronic device for predicting enterprise financial distress based on target-driven feature selection. Background Art

[0002] Ensemble learning is a method of completing learning tasks by building and combining multiple learners. Its core idea is to combine multiple weak learners into a strong learner, thereby improving the generalization ability and prediction accuracy of the overall model. At present, ensemble learning technology has been widely used in the field of financial distress prediction. It provides a lot of valuable reference information for financial decision-making by integrating the advantages of multiple prediction models.

[0003] It is worth noting that the performance of the integrated prediction model mainly depends on the diversity and quality of the selected base models, and combining feature selection and integrated learning can often produce a better prediction model. On the one hand, feature selection can eliminate irrelevant or redundant features, thereby reducing noise and interference in the model training process; on the other hand, when constructing the base model, if different feature subsets can be used to train different base models, it will inevitably increase the diversity between the base models and improve the prediction effect and generalization ability of the integrated prediction model. To this end, existing studies have constructed many integrated prediction models for corporate financial distress based on feature selection.

[0004] Although many studies have confirmed that introducing feature selection into the construction process of ensemble learning models can greatly improve the model's adaptability to new samples, the two are still relatively independent in the field of financial distress prediction. In most cases, existing studies have regarded feature selection as a step in the data preprocessing stage, which is used to screen out financial features that are significantly different between distressed and healthy companies, and then use the selected feature subsets to build an ensemble prediction model. Although this method can also ensure the prediction performance of the model to a certain extent, it ignores the potential correlation between different feature combinations and the model's prediction performance. In addition, when constructing a prediction model for corporate financial distress, existing studies also lack the consideration of whether the model will identify distressed companies as healthy companies, resulting in many companies often not taking effective measures to avoid risks in time and eventually on the verge of bankruptcy.

[0005] In the problem of predicting corporate financial distress, the actual risk of predicting a distressed enterprise as a healthy enterprise is far greater than identifying a healthy enterprise as a distressed enterprise. Therefore, it is necessary to adopt an effective mechanism to ensure that the integrated prediction model can avoid identifying distressed enterprises as healthy enterprises. Summary of the invention

[0006] 1. Technical issues to be resolved

[0007] In view of the deficiencies of the prior art, the present invention provides a method, system, storage medium and electronic device for predicting corporate financial distress based on target-driven feature selection, which solves the technical problem of ensuring that the integrated prediction model can avoid identifying distressed enterprises as healthy enterprises.

[0008] (II) Technical solution

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0010] A method for predicting corporate financial distress based on target-driven feature selection, comprising:

[0011] Collect financial data of multiple companies and label data on whether they are troubled companies to build the original data set;

[0012] Preprocessing the original data set to obtain a financial feature set; and dividing the preprocessed original data set into a training set and a validation set based on the financial feature set;

[0013] Training a base model based on the training set, and constructing a combined optimization model with the goal of maximizing the integrated prediction performance of the base model on the validation set;

[0014] Based on the training set, a benchmark model is constructed using a preset learning algorithm to initialize the first base model of the integrated prediction model;

[0015] Based on the first base model, the combined optimization model is iteratively solved by a feature selection method based on a genetic algorithm, and optimal feature subsets corresponding to other base models are sequentially searched for to obtain an integrated prediction model composed of multiple base models;

[0016] The financial data of the enterprise to be predicted is preprocessed and used as the input of the integrated prediction model. The individual prediction results of each base model are integrated using an average synthesis strategy to obtain the final integrated prediction result of the enterprise's financial distress.

[0017] Preferably, the combined optimization model is expressed as:

[0018] MAX AUC(C1,C2,…,C r )+Recall(C r )―λsum(F r )(1)

[0019] s C r = Train(D a , F r) (r = 2, 3, …, R) (2)

[0020]

[0021] In formula (1), MAX represents the maximization function; AUC(·) represents the current base model set {C1, C2, …, C r}In the validation set D v The predicted AUC value on ; Recall(·) represents the r-th base model C r On the validation set D v The recall value on ; λ represents the regularization coefficient, and sum(·) represents the feature subset F r The number of features contained in ; R represents the number of base models of the integrated prediction model; in constraint (2), Train(·) represents the model C r In the training set D a The feature subset F r The training process on the corresponding financial data; Constraint (3) F r Represents a feature subset selected from the financial feature set F.

[0022] Preferably, the method of constructing a benchmark model based on the training set using a preset learning algorithm to initialize the first base model of the integrated prediction model comprises:

[0023] Based on the training set D a , use SVM to build the benchmark model C0, and calculate C0 on the validation set D v Individual prediction performance on iperfo(C0);

[0024] Randomly select a feature subset F1 from the financial feature set F and select it from the training set D a Select the corresponding financial data to build the first base model C1, and calculate C1 in the validation set D v Individual prediction performance on iperfo(C1);

[0025] Determine the size relationship between iperfo(C0) and iperfo(C1). If iperfo(C1)>iperfo(C0), retain the first base model C1; otherwise, select a new feature subset F1 to rebuild the first base model C1 until iperfo(C1)>iperfo(C0) is satisfied.

[0026] Preferably, based on the first base model, the combined optimization model is iteratively solved by a feature selection method based on a genetic algorithm, and the optimal feature subsets corresponding to the training of other base models are sequentially searched to obtain an integrated prediction model composed of multiple base models; including:

[0027] The financial feature set F = {f1,f2,…,f M} is encoded into an M-dimensional 0-1 vector, where each position of the vector corresponds to a feature; if the mth position of the vector is 1, it indicates that feature f m Selected to build the base model C r , 0 means f m Not selected;

[0028] Assume that the initial population P consists of B chromosomes {I1,…,I B}, the initial population is recorded as the following binary matrix consisting of B×M elements:

[0029]

[0030] Substitute the first base model C1 into the optimization objective formula (1) to calculate the fitness value of each chromosome;

[0031] Perform selection and crossover on the initial population P in sequence, combine the selected chromosomes with the chromosomes obtained by crossover to obtain a new population, and perform mutation operation on the new population;

[0032] According to the fitness value of each chromosome in the current new population and the current evolutionary generation, judging whether the solution process of the combinatorial optimization model has reached convergence;

[0033] If it has not converged, the initial population P is replaced by the current new population, and the above process is repeated until convergence; if it has converged, the chromosome with the highest fitness value in the current new population is output as the optimal solution F of the combined optimization model r* , combined with the training set D a Building base model C r ;

[0034] Repeat the above process several times to obtain R basis models {C1, C2, …, C R}Combined prediction model EC.

[0035] Preferably, the step of sequentially performing selection and crossover on the initial population P, combining the selected chromosomes with the chromosomes obtained by crossover to obtain a new population, and performing a mutation operation on the new population includes:

[0036] The chromosomes in the initial population P are arranged in non-ascending order according to the fitness value, and after selecting the chromosomes with a preset ratio ranked at the top, two-point crossover operations are performed on them respectively to obtain a set of new chromosomes; the selected chromosomes are combined with the chromosomes obtained by the crossover to obtain a new population, and a random flip mutation operation is performed on it.

[0037] Preferably, the pretreatment comprises:

[0038] The maximum and minimum normalization method is used to normalize the original data set D, and the t-test is used to preliminarily filter the normalized data set to obtain the financial feature set F.

[0039] A corporate financial distress prediction system based on target-driven feature selection, comprising:

[0040] The data collection module is used to collect the financial data of multiple companies and label data on whether they are troubled companies to construct the original data set;

[0041] A data preprocessing module, used to preprocess the original data set to obtain a financial feature set; and based on the financial feature set, divide the preprocessed original data set into a training set and a validation set;

[0042] A model building module, used for training a base model based on the training set, and building a combined optimization model with the goal of maximizing the integrated prediction performance of the base model on the validation set;

[0043] A model initialization module, used to construct a benchmark model based on the training set using a preset learning algorithm to initialize the first base model of the integrated prediction model;

[0044] A model solving module, used for iteratively solving the combined optimization model based on the first base model by using a feature selection method based on a genetic algorithm, and sequentially searching for the optimal feature subsets corresponding to the training of other base models to obtain an integrated prediction model composed of multiple base models;

[0045] The distress prediction module is used to pre-process the financial data of the enterprise to be predicted and use it as the input of the integrated prediction model. It uses the average synthesis strategy to merge the individual prediction results of each base model to obtain the final integrated prediction result of the enterprise's financial distress.

[0046] A storage medium stores a computer program for predicting enterprise financial distress based on target-driven feature selection, wherein the computer program enables a computer to execute the enterprise financial distress prediction method as described above.

[0047] An electronic device, comprising:

[0048] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include instructions for executing the enterprise financial distress prediction method as described above.

[0049] (III) Beneficial effects

[0050] The present invention provides a method, system, storage medium and electronic device for predicting corporate financial distress based on target-driven feature selection. Compared with the prior art, the present invention has the following beneficial effects:

[0051] The present invention proposes an integrated prediction model based on supervised feature selection for the prediction of corporate financial distress, and models the construction process of the integrated prediction model into a combined optimization model, aiming to maximize the integrated prediction performance of the model and the individual prediction performance of the base model from the perspective of feature selection. Compared with the traditional integrated learning model based on feature selection, the proposed model can not only select a set of optimal feature subsets for each base model, but also provide a new combination mechanism that can maximize the performance of feature selection in improving the overall performance of the model. In addition, considering the particularity of the problem of predicting corporate financial distress, the proposed model increases the focus on the base model in avoiding identifying distressed enterprises as healthy enterprises when setting specific optimization goals, and also allows the integrated prediction performance of the model to be continuously optimized by continuously increasing the number of base models. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0053] Figure 1 An overall flow chart of a method for predicting corporate financial distress based on target-driven feature selection provided by an embodiment of the present invention;

[0054] Figure 2 A block diagram of a method for predicting corporate financial distress based on target-driven feature selection provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] The embodiments of the present application solve the technical problem of ensuring that the integrated prediction model can avoid identifying distressed enterprises as healthy enterprises by providing an enterprise financial distress prediction method, system, storage medium and electronic device based on target-driven feature selection.

[0057] The technical solution in the embodiment of the present application is to solve the above technical problems, and the overall idea is as follows:

[0058] like Figure 1 As shown, the enterprise financial distress integrated prediction model based on evolutionary feature selection proposed in the embodiment of the present invention includes three main stages:

[0059] The first stage is responsible for data collection and processing (data partitioning), and divides the processed data set into two non-overlapping subsets, namely the training set and the validation set.

[0060] The second stage is responsible for constructing an integrated prediction model for financial distress (model construction), with the goal of maximizing the integrated prediction performance of the base model on the validation set, and using a genetic algorithm to continuously search for a set of optimal feature subsets for each base model.

[0061] The third stage is responsible for generating integrated prediction results (distress prediction). According to the optimal feature subset selected when training each base model, the individual prediction results of each base model for the financial status of the enterprise to be predicted are obtained, and then the average synthesis strategy is used to obtain the final integrated prediction results.

[0062] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0063] Embodiment 1:

[0064] like Figure 2 As shown, an embodiment of the present invention provides a method for predicting corporate financial distress based on target-driven feature selection, including:

[0065] S1. Collect financial data of multiple companies and label data on whether they are distressed companies to construct the original data set;

[0066] S2, preprocessing the original data set to obtain a financial feature set; and based on the financial feature set, dividing the preprocessed original data set into a training set and a validation set;

[0067] S3, training a base model based on the training set, and constructing a combined optimization model with the goal of maximizing the integrated prediction performance of the base model on the validation set;

[0068] S4. Based on the training set, a benchmark model is constructed using a preset learning algorithm to initialize a first base model;

[0069] S5. Based on the first base model, the combined optimization model is iteratively solved by a feature selection method based on a genetic algorithm, and optimal feature subsets corresponding to other base models are sequentially searched for to obtain an integrated prediction model composed of multiple base models;

[0070] S6. Preprocess the financial data of the enterprise to be predicted and use it as the input of the integrated prediction model. Use the average synthesis strategy to merge the individual prediction results of each base model to obtain the final integrated prediction result of the enterprise's financial distress.

[0071] The embodiments of the present invention can not only improve the prediction performance of corporate financial distress, but also effectively reduce the probability of prediction errors in identifying distressed enterprises as healthy enterprises, thereby providing better decision support for corporate financial early warning.

[0072] like Figure 1 As shown, Figure 1 An overall flow chart of the enterprise financial distress prediction method based on target-driven feature selection provided by an embodiment of the present invention is given.

[0073] Next, we will combine Figure 1 The following are the steps of the above solution:

[0074] First, the embodiment of the present invention performs Figure 1 The data division stage shown in the figure, please refer to the contents of steps S1 to S2 for details:

[0075] In step S1, the financial data of multiple companies and label data on whether they are troubled companies are collected to construct an original data set.

[0076] In this step, based on the preset financial characteristics framework, we collect N companies x n The original data set D that can be collected (n=1,…,N) (including distressed enterprises and healthy enterprises).

[0077] Since the prediction of corporate financial distress is a typical binary classification problem, each sample in the original data set D consists of two parts. The first part represents the financial performance of the enterprise in terms of financial characteristics in a certain year, and the second part indicates whether the enterprise is in financial distress.

[0078] In step S2, the original data set is preprocessed to obtain a financial feature set; and based on the financial feature set, the preprocessed original data set is divided into a training set and a validation set.

[0079] In this step, considering that each financial feature has different meanings and units, the Max-Min Normalization method is first used to normalize the original data set D to eliminate the adverse effects of the dimension on model construction. The t-test is then used to perform a preliminary filtering of the obtained data set to select financial features with significant differences between distressed enterprises and healthy enterprises, thereby ensuring the predictive performance of the proposed model.

[0080] Suppose the selected financial feature set is F = {f1,f2,…,f M}, under the financial feature set, the original data set D is divided into a training set and validation set Satisfies N1+N2=N.

[0081] in, and Respectively represent enterprises and The financial feature vector of and and and For the corresponding label, indicate the enterprise and Are you in financial trouble? Let y = 0, which means the enterprise is a healthy enterprise and its financial situation is not in trouble; y = 1, which means the enterprise is a troubled enterprise and its financial situation is in trouble.

[0082] Secondly, the embodiment of the present invention performs Figure 1 The model building stage shown in , see steps S3 to S5 for details:

[0083] In step S3, a base model is trained based on the training set, and a combined optimization model is constructed with the goal of maximizing the integrated prediction performance of the base model on the validation set.

[0084] The embodiment of the present invention designs an integrated prediction model construction process based on evolutionary feature selection, which can simultaneously consider the integrated prediction performance of the model and the individual prediction performance of the base model; and taking into account the particularity of the problem of corporate financial distress prediction, when setting specific optimization goals, increased attention is paid to the base model in avoiding identifying distressed enterprises as healthy enterprises.

[0085] Specifically, in this step, based on the training set D a and validation set D v , by solving the following combination optimization model to build an integrated prediction model EC = {C1,…,C R The rth base model C in r (r=2,…,R):

[0086] MAX AUC(C1,C2,…,C r )+Recall(C r )―λsum(F r )(1)

[0087] s C r = Train(D a, F r ) (r = 2, 3, …, R) (2)

[0088]

[0089] In formula (1), MAX represents the maximization function; AUC(·) represents the current base model set {C1, C2, …, C r}In the validation set D v The predicted AUC value on ; Recall(·) represents the r-th base model C r On the validation set D v The recall value on ; λ represents the regularization coefficient, and sum(·) represents the feature subset F r The number of features contained in ; R represents the number of base models of the integrated prediction model; in constraint (2), Train(·) represents the model C r In the training set D a The feature subset F r The training process on the corresponding financial data; Constraint (3) F r Represents a feature subset selected from the financial feature set F.

[0090] It should be noted that the above combined optimization model is only effective when there is at least one base model, otherwise AUC(·) will be meaningless.

[0091] In step S4, a baseline model is constructed based on the training set using a preset learning algorithm to initialize the first base model.

[0092] It is understandable that solving the above combined optimization model requires constructing an initial base model C1. To achieve this goal, the embodiment of the present invention selects SVM as the base learning algorithm of the proposed model.

[0093] Specifically, this step includes:

[0094] Before building the base model C1, based on the training set D a , use SVM to build the benchmark model C0, and calculate C0 on the validation set D v Individual prediction performance on iperfo(C0);

[0095] After that, a feature subset F1 is randomly selected from the financial feature set F and from the training set D a Select the corresponding financial data to build the first base model C1, and calculate C1 in the validation set D v Individual prediction performance on iperfo(C1);

[0096] Finally, the size relationship between iperfo(C0) and iperfo(C1) is determined. If iperfo(C1)>iperfo(C0), the first base model C1 is retained; otherwise, a new feature subset F1 is selected to reconstruct the first base model C1 until iperfo(C1)>iperfo(C0) is satisfied.

[0097] The purpose of setting the above conditions is to ensure the effectiveness of the feature selection method, so that a better model can be trained using fewer features.

[0098] In step S5, based on the first base model, the combined optimization model is iteratively solved using a feature selection method based on a genetic algorithm, and the optimal feature subsets corresponding to the training of other base models are searched in turn to obtain an integrated prediction model composed of multiple base models.

[0099] In order to generate a model that can simultaneously consider the integrated prediction performance of the model and the individual prediction performance of the base model, and increase the model's focus on avoiding identifying troubled enterprises as healthy enterprises, based on the generated base model C1, the embodiment of the present invention uses a feature selection method based on a genetic algorithm (GA) to solve the combined optimization model in formulas (1)-(3).

[0100] Specifically, this step includes:

[0101] The financial feature set F = {f1,f2,…,f M} is encoded into an M-dimensional 0-1 vector, where each position of the vector corresponds to a feature; if the mth position of the vector is 1, it indicates that feature f m Selected to build the base model C r , 0 means f m Not selected.

[0102] Assume that the initial population P consists of B chromosomes {I1,…,I B}, the initial population is recorded as the following binary matrix consisting of B×M elements:

[0103]

[0104] It can be understood that in the above matrix, each row represents a chromosome and each column corresponds to a feature. The fitness of each chromosome is calculated using the optimization objective formula (1).

[0105] Next, the first base model C1 is substituted into the optimization objective formula (1) to calculate the fitness value of each chromosome.

[0106] The initial population P is selected and crossed in turn, the selected chromosomes are combined with the chromosomes obtained by crossover to obtain a new population, and the new population is mutated. The specific operation process is as follows:

[0107] The chromosomes in the initial population P are arranged in non-ascending order according to the size of the fitness value, and after selecting the chromosomes with a preset ratio ranked at the top, two-point crossover operations are performed on them respectively to obtain a set of new chromosomes; the selected chromosomes are combined with the chromosomes obtained by the crossover to obtain a new population, and a random flip mutation operation is performed on it (that is, some chromosome genes are randomly selected and changed from 1 to 0 or from 0 to 1) to obtain the current new population.

[0108] Afterwards, it is determined whether the solution process of the combinatorial optimization model has reached convergence according to the fitness value of each chromosome in the current new population and the current evolutionary generation.

[0109] If it has not converged, the initial population P is replaced by the current new population, and the above process is repeated until convergence; if it has converged, the chromosome with the highest fitness value in the current new population is output as the optimal solution F of the combined optimization model r* , combined with the training set D a Building base model C r ;

[0110] Repeat the above process several times to obtain R basis models {C1, C2, …, C R}Combined prediction model EC.

[0111] Finally, the embodiment of the present invention performs Figure 1 The dilemma prediction stage shown in FIG. 1 is shown in FIG. 1 , and specifically refers to the content in step S6:

[0112] In step S6, the financial data of the enterprise to be predicted is preprocessed and used as the input of the integrated prediction model. The individual prediction results of each base model are fused using an average synthesis strategy to obtain the final integrated prediction result of the enterprise's financial distress.

[0113] As we can see from the previous text, in the financial feature set F = {f1,f2,…,f M}, each base model in the integrated prediction model EC is constructed by different feature subsets. When using EC to predict the financial status of an enterprise, we first need to solve the optimal solution F obtained by solving formulas (1)-(3). r* (r=1,…,R) is used to train other corresponding optimal feature subsets in turn, and then input into the corresponding base model C r To obtain individual forecast results of the company's financial status.

[0114] Taking the base model C3 as an example, let F 3*={f1,f4,f M-1 ,f M}, then for the enterprise whose financial status is to be predicted, it is necessary to predict the 3* Based on the financial performance of a company, use C3 to predict whether it will encounter financial difficulties in the future.

[0115] In this step, the same preprocessing method is used to process the financial data of the enterprise to be predicted and use it as the input of the integrated prediction model. After obtaining the prediction results of R base models for the target enterprise in turn, the average synthesis strategy is used to merge the individual prediction results of each base model to obtain the final integrated prediction result of the enterprise's financial distress.

[0116] So far, the embodiment of the present invention has completed the entire process of the enterprise financial distress prediction method based on target-driven feature selection.

[0117] In order to help understand the superiority of the enterprise financial distress prediction method proposed in the embodiment of the present invention, the performance of the proposed model is verified using real financial data of listed companies, and the experimental results are compared with mainstream prediction methods, highlighting the advantages of the proposed model. For details, see the following example:

[0118] The following empirical study is conducted based on real data of Chinese listed companies collected from the China Economic and Financial Research Database (CSMAR) to test the effectiveness of the proposed model.

[0119] Scholars at home and abroad have not been able to give a unified definition of corporate financial distress. Domestic scholars are accustomed to considering companies that have been treated specially (ST) due to abnormal financial operations as companies in financial distress. For this reason, this example also uses ST as a distinction standard, and regards listed companies in Shanghai and Shenzhen Stock Exchanges that have been ST from 2012 to 2022 as positive samples of financial distress, with the category label recorded as 1, and the remaining listed companies that have not been ST as paired negative samples, with the category label recorded as 0. After eliminating samples of companies with incomplete information, a total of 3,969 listed company samples were obtained, of which 399 were ST companies and 3,570 were non-ST companies. Assume that the year when the listed company was ST is the Tth year. According to the lag of financial distress, the financial characteristic data of these companies in the T-3th year are collected to predict the financial status of the company in the Tth year. That is, if a company is ST in 2014, the company's financial data for 2011 (T-3) will be collected.

[0120] Since there is no recognized financial characteristic system to predict the financial status of enterprises, we refer to relevant existing studies and select 40 financial characteristics in Table 1 to describe the financial status of each company from six aspects: operating capacity, profitability, development capacity, capital expansion capacity, debt repayment capacity and financial structure. In the process of building each model, the selected features will be screened using t-test to avoid feature redundancy.

[0121] Table 1 Selected financial characteristics

[0122]

[0123]

[0124] The final number of ST companies and non-ST companies collected was 268 and 2148 respectively. In order to avoid the negative impact of data imbalance on the prediction model, this example first selects the financial data of all ST companies, and then randomly selects a non-ST company as a paired negative sample for each ST company according to the 1:1 principle. If a ST company does not have a non-ST company paired with the same year, the ST company will not be selected. Considering the long time span of data collection, some companies may be ST multiple times during this period. In order to avoid the negative impact on the experimental results due to repeated sampling of a company, only the financial data of T-3 years corresponding to the first time these companies were ST will be selected for prediction. The final data set is recorded as D T-3 .

[0125] Generally speaking, financial distress prediction is to predict the probability of a company's future financial crisis based on its current financial situation. T-3 The companies that were ST between 2012 and 2018 and their paired non-ST companies are regarded as training samples, the companies that were ST between 2019 and 2020 and their paired non-ST companies are regarded as verification samples, and the remaining ST companies and their paired non-ST companies are regarded as test samples. The company financial data of the training samples and the verification samples are used to form the training set and the verification set respectively to construct the model proposed in the embodiment of the present invention, and then the financial data corresponding to the test sample is used to test the predictive performance of the model. Since the selection of paired non-ST companies is random, the above experimental process is repeated 30 times, and the predictive performance of the proposed model in the 30 average experiments is regarded as the final predictive performance for comparison with the benchmark method. The relevant parameters of the proposed model in the experiment are shown in Table 2.

[0126] Table 2 Parameter settings of the proposed model

[0127] parameter value Number of base models R 10 Regularization coefficient λ 0.001 SVM kernel function Linear kernel function Population size B 30 Select Ratio 0.5 Crossover probability 0.7 Mutation probability 0.2 Evolution maximum generation MaxG 100 Fitness improvement threshold δ 0.0001

[0128] Given that the label of ST companies is 1 and the label of non-ST companies is 0, there are four possible prediction results for a company whose financial status is to be predicted:

[0129] (1) If a company will encounter financial difficulties in the future (the true label is 1), and the model’s prediction results also show that the company will be in trouble (the predicted label is 1), then the company sample is called a true positive sample. The number of true positive samples is TP.

[0130] (2) If a company is in financial distress in the future (the true label is 1), but the model’s prediction shows that the company will not be in distress (the predicted label is 0), then the company sample is called a false negative sample. The number of false negative samples is denoted as FN.

[0131] (3) If the enterprise will not encounter financial difficulties in the future (the true label is 0), and the model's prediction results also show that the enterprise will not fall into difficulties (the predicted label is 0), then the enterprise sample is called a true negative sample. The number of true negative samples is TN.

[0132] (4) If the enterprise will not encounter financial difficulties in the future (the true label is 0), but the model's prediction results show that the enterprise will be in trouble (the predicted label is 1), then the enterprise sample is called a false positive sample. The number of false positive samples is recorded as FP.

[0133] Based on TP, FN, TN and FP, the following measures are selected to evaluate the performance of the proposed model:

[0134] Accuracy:

[0135] Recall:

[0136] Specificity:

[0137] F1 score value:

[0138] G mean:

[0139] In addition, considering that the area AUC value enclosed by the ROC curve and the coordinates is the most accepted performance evaluation indicator in binary classification problems, this example also considers this measure.

[0140] In order to highlight the financial distress prediction advantages of the proposed model, the following mainstream methods are selected as the benchmark methods for this example.

[0141] Single model prediction method: Based on the training set, four mainstream methods, SVM, DT, ANN and logistic regression (LOG), are selected to build a single financial distress prediction model. Although these four methods are often used for financial distress prediction, existing studies have not given their specific parameter settings. In this example, the scikit-learn package is used to implement these four methods, and the relevant parameters are all set by default.

[0142] Ensemble prediction model: Based on the training set, four ensemble learning methods with superior performance, namely random forest (RF), random subspace (RS), AdaBoost (AB), and XGBoost (XB), are selected to build an ensemble prediction model for financial distress. Among them, RF, RS, and AB are implemented by the scikit-learn package, and XB is implemented by the xgboost package. The number of base models of the four ensemble prediction models is set to 10, and the rest of the parameters are set by default.

[0143] Table 3 shows the average results of 30 repeated experiments, where the best result in each row is bolded.

[0144] Table 3. The performance of the nine methods on dataset D T-3 Average prediction performance on

[0145]

[0146] As can be seen from Table 3, the proposed model (PM) has achieved better prediction results than the benchmark methods in many performance measures, especially in the two comprehensive evaluation measures of F1 score and G mean, the performance advantage of PM is very obvious. Although PM's performance is not as good as the benchmark method in the measure of specificity, it can be found that PM is significantly better than all benchmark methods in terms of recall. The relationship between recall and specificity is equivalent to the two types of errors in statistics, that is, identifying distressed enterprises as healthy enterprises and identifying healthy enterprises as distressed enterprises. Although the limited number of samples and the randomness in the sample selection process make it difficult to improve the model's prediction performance in terms of recall and specificity at the same time, considering that the actual risk brought by the first type of error is much greater than the second type of error in the problem of financial distress prediction, the above experimental results further emphasize the effectiveness and applicability of PM in the problem of corporate financial distress prediction.

[0147] Embodiment 2:

[0148] The embodiment of the present invention provides an enterprise financial distress prediction system based on target-driven feature selection, comprising:

[0149] The data collection module is used to collect the financial data of multiple companies and label data on whether they are troubled companies to construct the original data set;

[0150] A data preprocessing module, used to preprocess the original data set to obtain a financial feature set; and based on the financial feature set, divide the preprocessed original data set into a training set and a validation set;

[0151] A model building module, used for training a base model based on the training set, and building a combined optimization model with the goal of maximizing the integrated prediction performance of the base model on the validation set;

[0152] A model initialization module, used to construct a benchmark model based on the training set using a preset learning algorithm to initialize the first base model of the integrated prediction model;

[0153] A model solving module, used for iteratively solving the combined optimization model based on the first base model by using a feature selection method based on a genetic algorithm, and sequentially searching for the optimal feature subsets corresponding to the training of other base models to obtain an integrated prediction model composed of multiple base models;

[0154] The distress prediction module is used to pre-process the financial data of the enterprise to be predicted and use it as the input of the integrated prediction model. It uses the average synthesis strategy to merge the individual prediction results of each base model to obtain the final integrated prediction result of the enterprise's financial distress.

[0155] Embodiment 3:

[0156] An embodiment of the present invention provides a storage medium storing a computer program for predicting corporate financial distress based on target-driven feature selection, wherein the computer program enables a computer to execute the method for predicting corporate financial distress as described in Example 1.

[0157] Embodiment 4:

[0158] An embodiment of the present invention provides an electronic device, including:

[0159] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include methods for executing the enterprise financial distress prediction method as described in Example 1.

[0160] It can be understood that the enterprise financial distress prediction system, storage medium and electronic device based on target-driven feature selection provided in the embodiments of the present invention correspond to the enterprise financial distress prediction method based on target-driven feature selection provided in the embodiments of the present invention, and the explanations, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the enterprise financial distress prediction method, which will not be repeated here.

[0161] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0162] 1. The embodiments of the present invention can not only improve the prediction performance of corporate financial distress, but also effectively reduce the probability of the first type of prediction errors (identifying distressed enterprises as healthy enterprises), and provide better decision support for corporate financial early warning.

[0163] 2. On the one hand, the embodiments of the present invention promote the deep integration of feature selection technology and ensemble learning technology in the field of enterprise financial distress prediction. On the other hand, they also provide a lot of valuable reference information for studying the construction of adaptive ML models from the perspective of feature selection.

[0164] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0165] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting corporate financial distress based on target-driven feature selection, characterized in that: include: Collect financial data of multiple companies and label data on whether they are troubled companies to build the original data set; Preprocessing the original data set to obtain a financial feature set; Based on the financial feature set, the preprocessed original data set is divided into a training set and a validation set; Training a base model based on the training set, and constructing a combined optimization model with the goal of maximizing the integrated prediction performance of the base model on the validation set; Based on the training set, a benchmark model is constructed using a preset learning algorithm to initialize the first base model of the integrated prediction model; Based on the first base model, the combined optimization model is iteratively solved by a feature selection method based on a genetic algorithm, and optimal feature subsets corresponding to other base models are sequentially searched for to obtain an integrated prediction model composed of multiple base models; The financial data of the enterprise to be predicted is preprocessed and used as the input of the integrated prediction model. The individual prediction results of each base model are integrated using an average synthesis strategy to obtain the final integrated prediction result of the enterprise's financial distress.

2. The method for predicting corporate financial distress according to claim 1, characterized in that: The combined optimization model is expressed as: MAX AUC(C1,C2,…,C r )+Recall(C r )―λsum(F r )(1) s.t. C r = Train(D a , F r ) (r = 2, 3, …, R) (2) In formula (1), MAX represents the maximization function; AUC(·) represents the current base model set {C1, C2, …, C r }In the validation set D v The predicted AUC value on ; Recall(·) represents the r-th base model C r On the validation set D v The recall value on ; λ represents the regularization coefficient, and sum(·) represents the feature subset F r The number of features contained in ; R represents the number of base models of the integrated prediction model; in constraint (2), Train(·) represents the model C r In the training set D a The feature subset F r The training process on the corresponding financial data; Constraint (3) F r Represents the feature subset selected from the financial feature set F.

3. The method for predicting corporate financial distress according to claim 1, characterized in that: Based on the training set, a benchmark model is constructed using a preset learning algorithm to initialize the first base model of the integrated prediction model; including: Based on the training set D a , use SVM to build the benchmark model C0, and calculate C0 on the validation set D v Individual prediction performance on iperfo(C0); Randomly select a feature subset F1 from the financial feature set F and select it from the training set D a Select the corresponding financial data to build the first base model C1, and calculate C1 in the validation set D v Individual prediction performance on iperfo(C1); Determine the size relationship between iperfo(C0) and iperfo(C1). If iperfo(C1)>iperfo(C0), retain the first base model C1; otherwise, select a new feature subset F1 to rebuild the first base model C1 until iperfo(C1)>iperfo(C0) is satisfied.

4. The method for predicting corporate financial distress according to claim 2, characterized in that: Based on the first base model, the combined optimization model is iteratively solved by a feature selection method based on a genetic algorithm, and optimal feature subsets corresponding to other base models are sequentially searched for to obtain an integrated prediction model composed of multiple base models; include: The financial feature set F = {f1,f2,…,f M } is encoded into an M-dimensional 0-1 vector, where each position of the vector corresponds to a feature; if the mth position of the vector is 1, it indicates that feature f m Selected to build the base model C r , 0 means f m Not selected; Assume that the initial population P consists of B chromosomes {I1,…,I B }, the initial population is recorded as the following binary matrix consisting of B×M elements: Substitute the first base model C1 into the optimization objective formula (1) to calculate the fitness value of each chromosome; Perform selection and crossover on the initial population P in sequence, combine the selected chromosomes with the chromosomes obtained by crossover to obtain a new population, and perform mutation operation on the new population; According to the fitness value of each chromosome in the current new population and the current evolutionary generation, judging whether the solution process of the combinatorial optimization model has reached convergence; If it has not converged, the initial population P is replaced by the current new population, and the above process is repeated until convergence; if it has converged, the chromosome with the highest fitness value in the current new population is output as the optimal solution F of the combined optimization model r* , combined with the training set D a Building base model C r ; Repeat the above process several times to obtain R basis models {C1, C2, …, C R }Combined prediction model EC.

5. The method for predicting corporate financial distress according to claim 4, characterized in that: The step of sequentially performing selection and crossover on the initial population P, combining the selected chromosomes with the chromosomes obtained by crossover to obtain a new population, and performing a mutation operation on the new population includes: The chromosomes in the initial population P are arranged in non-ascending order according to the fitness value, and after selecting the chromosomes with a preset ratio ranked at the top, two-point crossover operations are performed on them respectively to obtain a set of new chromosomes; the selected chromosomes are combined with the chromosomes obtained by the crossover to obtain a new population, and a random flip mutation operation is performed on it.

6. The method for predicting corporate financial distress according to any one of claims 1 to 5, characterized in that: The pre-processing comprises: The maximum and minimum normalization method is used to normalize the original data set D, and the t-test is used to preliminarily filter the normalized data set to obtain the financial feature set F.

7. A corporate financial distress prediction system based on target-driven feature selection, characterized in that: include: The data collection module is used to collect the financial data of multiple companies and label data on whether they are troubled companies to construct the original data set; A data preprocessing module, used for preprocessing the original data set to obtain a financial feature set; Based on the financial feature set, the preprocessed original data set is divided into a training set and a validation set; A model building module, used for training a base model based on the training set, and building a combined optimization model with the goal of maximizing the integrated prediction performance of the base model on the validation set; A model initialization module, used to construct a benchmark model based on the training set using a preset learning algorithm to initialize the first base model of the integrated prediction model; A model solving module, used for iteratively solving the combined optimization model based on the first base model by using a feature selection method based on a genetic algorithm, and sequentially searching for the optimal feature subsets corresponding to the training of other base models to obtain an integrated prediction model composed of multiple base models; The distress prediction module is used to pre-process the financial data of the enterprise to be predicted and use it as the input of the integrated prediction model. It uses the average synthesis strategy to merge the individual prediction results of each base model to obtain the final integrated prediction result of the enterprise's financial distress.

8. A storage medium, characterized in that: The computer program for predicting enterprise financial distress based on target-driven feature selection is stored therein, wherein the computer program enables a computer to execute the enterprise financial distress prediction method according to any one of claims 1 to 6.

9. An electronic device, characterized in that: include: one or more processors; Memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs comprising instructions for executing the enterprise financial distress prediction method according to any one of claims 1 to 6.

Images