CN108629675A - A kind of Multiple trees financial alert method - Google Patents

Abstract

The present invention provides a kind of Multiple trees financial alert method.This method includes：Step 1 obtains financial data sample set D to be detected, and the financial data sample set D to be detected includes the data subset corresponding to each data attribute in data attribute subset and the data attribute subset；Step 2 carries out M random sampling according to preset random sampling number M to the financial data sample set D to be detected, extracts M data every time, determines that the sample set that kth time random sampling obtains is training subset D_k, wherein k=1,2,3 ... M；Step 3, using preset noise reduction own coding model in the training subset D_kAcquistion go to school to decision tree T_k；Step 4, according to the decision tree T_kFinancial situation class prediction is carried out to each data subset in financial data sample set D to be detected, the financial situation classification includes that financial situation is good and financial situation is abnormal.The present invention can improve the accuracy predicted financial situation and generalization ability.

Description

A Multi-decision Tree Financial Early Warning Method

technical field

The invention relates to the technical field of data analysis, in particular to a multi-decision tree financial early warning method.

Background technique

In recent years, with the rapid development of my country's market economy and capital market, the competition among domestic enterprises has become increasingly fierce. At the same time, more and more multinational companies and domestic companies are also competing for customers, human resources, commodity and capital suppliers. While the integration of the world economy has brought unprecedented opportunities to enterprises, it has also brought enormous challenges, and the uncertainties faced by enterprises are increasing day by day. With the exposure of the bankruptcies of large group companies such as Lehman Brothers Holdings and WorldCom, it is not uncommon for companies to encounter financial difficulties or even go bankrupt. However, it is a gradual process for an enterprise to fall into financial crisis, and it is not formed in a short period of time, but has a long incubation period. Changes in the financial status of an enterprise can be reflected by certain financial indicators, so financial indicators have the ability to predict the company's future financial status. If the financial situation of the enterprise is predicted by certain technical means, it is of great significance to various stakeholders of the enterprise, including investors: first, it will help the enterprise managers to discover the financial problems of the enterprise in time, solve them, and prevent problems before they happen; It is beneficial to investors' investment decisions and protects their interests.

In order to predict the financial situation timely and accurately, universities, enterprises and investors in our country have successively carried out research on financial early warning methods. At present, there are corresponding financial and police methods. However, most of the existing financial early warning methods in China use single-model forecasting methods, such as neural network models, decision tree models, Logistic regression models, etc., with low accuracy and performance needs to be improved; while a few multi-model forecasting methods The method has low generalization ability.

Contents of the invention

The invention provides a multi-decision tree financial early warning method, which is used to improve the accuracy and generalization ability of financial status prediction.

The invention provides a multi-decision tree financial early warning method, the method comprising:

Step 1. Obtain a financial data sample set D to be detected, the financial data sample set D to be detected includes data attribute subsets and data subsets corresponding to each data attribute in the data attribute subsets;

Step 2. Perform M random sampling on the financial data sample set D to be tested according to the preset random sampling times M, extract M data each time, and determine the sample set obtained by the kth random sampling as the training subset D _k , where k=1,2,3...M;

Step 3, using the preset noise reduction autoencoder model to learn the decision tree T _{k on the training subset D k ;}

Step 4. According to the decision tree T _k , predict the financial status category of each data subset in the financial data sample set D to be detected, and the financial status category includes good financial status and abnormal financial status.

Further, the step 3 specifically includes:

Step 31. Use the preset noise reduction autoencoder model

Encode the training subset D _k in the l-layer neural network to obtain a new training subset Where W ^{(k, 1)} is the weight of the l-th layer neural network, and b ^{(k, l)} is the bias item of the l-th layer neural network.

Step 32, in the new training subset The decision tree model T _k is obtained through learning above.

Further, the weight W ^{(k, 1)} and the bias item b ^{(k, l)} in the step 31 are respectively according to the following formula:

Determined; where, λ is the weight decay parameter, x is the data subset under the specific data attribute, is a subset of data under a specific data attribute after adding noise, represents the output of the output layer, and represent the partial derivatives of the cost function J(W ^(k,l-1) ,b ^(k,l-1) ) with respect to W ^(k,l-1) and b ^(k,l-1) respectively.

Further, the step 32 specifically includes:

Step 321, initialize the generation node node as the root node;

Step 322, if If the financial data samples belong to a specific financial status category or the number of financial data samples is less than a preset threshold, the node is marked as a leaf node of the specific financial status category;

Step 323, if different financial data samples belong to a specific financial status category and the number of financial data samples is not less than the preset threshold, then use the objective function

Search test conditions, where D _t is the data set corresponding to the current node t, D _tj is the data set of the child node j of the current node t, and Ent(D _tj ) is the information entropy of D _tj ;

Step 324, according to test condition will Divide into two subsets and build two child nodes.

Step 325, for each subset, repeat step 322 and step 323.

Further, the step 4 specifically includes:

Using the combined base classifier T ^*

Classify a data subset x under a specific data attribute, where δ(true)=1, δ(false)=0, T _k (x) is the prediction of the decision tree T _k on the data subset x under a specific data attribute.

Further, the subset of data attributes includes: asset-liability ratio, return on net assets, return on assets, net profit rate, total asset turnover, accounts receivable turnover, current asset turnover, main business growth rate , at least one data attribute in total asset growth rate and net profit growth rate.

Technical effect of the present invention:

A multi-decision tree financial early warning method provided by the present invention obtains a plurality of different training subsets by randomly sampling the acquired financial data sample set to be detected, and then encodes a plurality of different training subsets to obtain multiple Different decision trees, and finally predict the subset of data to be detected by combining multiple decision trees. Therefore, the present invention can improve the accuracy and generalization ability of financial status prediction.

Description of drawings

Fig. 1 is a schematic structural diagram of a decision tree provided by an embodiment of the present invention;

Fig. 2 is a schematic flow chart of the multi-decision tree financial early warning method provided by the embodiment of the present invention;

Fig. 3 is a schematic flowchart of a multi-decision tree financial early warning method provided by another embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the present invention Examples, not all examples. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

A decision tree is an easy-to-use nonparametric classifier. It does not require any prior assumptions about the data, the calculation speed is fast, the results are easy to interpret, and it is robust

The composition of the decision tree has four elements: (1) decision node; (2) program branch; (3) state node; (4) probability branch. As shown in Figure 1: The decision tree is generally composed of square nodes, circular nodes, plan branches, probability branches, etc. The square nodes are called decision nodes, and several thin branches are derived from the decision nodes. Each thin branch It represents a plan, called a plan branch; the circular node is called a state node, and several thin branches are drawn from the state node, which represent different natural states, called probability branches. Each probability branch represents a natural state. Mark the content of the objective state and its probability of occurrence on each twig. At the end of the probability branch, indicate the result (benefit value or loss value) achieved by the plan in the natural state. In this way, the tree diagram is expanded from left to right, from simple to complex, forming a tree network diagram.

Fig. 2 is a schematic flowchart of a multi-decision tree financial early warning method provided by an embodiment of the present invention. As shown in Figure 2, the method includes the following steps:

S201. Obtain a financial data sample set D to be detected, the financial data sample set D to be detected includes a data attribute subset and a data subset corresponding to each data attribute in the data attribute subset;

Specifically, the financial data sample set D to be tested in the embodiment of the present invention is preprocessed and can be directly used for the original training set of the training model. The so-called preprocessing refers to supplementing, correcting and normalizing the missing and abnormal data. The value type of the data in each data subset is numeric.

S202. Perform M random sampling on the financial data sample set D to be tested according to the preset number of random sampling M, extract M data each time, and determine the sample set obtained by the kth random sampling as the training subset D _k , where k=1,2,3...M;

Specifically, the random sampling in the embodiment of the present invention is iterative random sampling with replacement, aiming to determine M different training subsets D _k from the sample set D, so as to prepare for the subsequent training of multiple decision trees .

S203. Using a preset noise reduction autoencoder model to learn a decision tree T _{k on the training subset D k ;}

Specifically, the so-called preset noise reduction autoencoder model is a neural network model containing a hidden layer, the leftmost layer of the neural network is called the input layer, the rightmost layer is called the output layer, and all nodes in the middle are composed of One layer is called the hidden layer, and this hidden layer is considered to encode the spatial layer. The so-called noise reduction refers to randomly adding noise to the financial training data in the training subset _Dk of the input layer during the learning stage, so that the learned encoder has strong robustness, thereby enhancing the generalization ability of the model. In practical applications, the number of hidden units in the hidden layer can be determined according to the number of data attributes included in the data attribute subset, that is, the number of hidden units is equal to the number of data attributes.

Each training subset Dk is input through the input layer, and the preset noise reduction _autoencoder model is used to encode the training subset to obtain the coding space of each training subset _Dk , so as to learn the decision in each coding space tree T _k , finally M decision trees can be obtained.

S204. Predict the financial status category of each data subset in the financial data sample set D to be detected according to the decision tree T _k , and the financial status category includes good financial status and abnormal financial status.

Specifically, in each coding space, use the corresponding decision tree model _Tk to vote on the financial situation, accumulate these votes and use the majority voting method to obtain the prediction result.

The multi-decision tree financial alarm method provided by the embodiment of the present invention obtains a plurality of different training subsets by randomly sampling the acquired financial data sample set to be detected, and then encodes a plurality of different training subsets to obtain multiple Different decision trees, and finally predict the subset of data to be detected by combining multiple decision trees. Therefore, the present invention can improve the accuracy and generalization ability of financial status prediction.

On the basis of the foregoing embodiments, step 203 in the method specifically includes:

Step 2031, using a preset noise reduction autoencoder model

Encode the training subset D _k in the l-layer neural network to obtain a new training subset Where W ^{(k, 1)} is the weight of the l-th layer neural network, and b ^{(k, l)} is the bias item of the l-th layer neural network.

Step 2032, in the new training subset The decision tree model T _k is obtained through learning above.

Specifically, this process can be regarded as encoding the "input data" D _k , so that the encoded data vector The typical characteristics of the input financial data can be preserved, so that the original financial data can be recovered more conveniently.

On the basis of the above-mentioned embodiments, the weight W ^{(k, 1)} and the bias term b ^{(k, l)} in step 2031 in the method are respectively according to the following formula:

Determined; where, λ is the weight decay parameter, x is the data subset under the specific data attribute, is a subset of data under a specific data attribute after adding noise, represents the output of the output layer, and represent the partial derivatives of the cost function J(W ^(k,l-1) ,b ^(k,l-1) ) with respect to W ^(k,l-1) and b ^(k,l-1) respectively.

Specifically, the initial value of the weight W={W ^(k,l) |k=1,2...M;l=1,2,3;} can be determined by making the input value equal to the output value. In practical application, and The calculation process is as follows:

Where δ ^(l) is the residual of layer l, specifically:

z ^(l+1) = W ^(k,l) a ^(l) +b ^(k,l) ,l=1

On the basis of the above-mentioned embodiments, step 2032 in the method specifically includes:

Step 20321, initialize the generated node node as the root node;

Step 20322, if If the financial data samples belong to a specific financial status category or the number of financial data samples is less than a preset threshold, the node is marked as a leaf node of the specific financial status category;

Step 20323, if different financial data samples belong to a specific financial status category and the number of financial data samples is not less than the preset threshold, then use the objective function

Among them, D _t is the data set corresponding to the current node t, D _tj is the data set of the child node j of the current node t; Ent(D) is the information entropy of D, reflecting the purity of the data D, the smaller the value, The higher the purity of D is; Gain represents the increase value of information entropy (information gain) before and after division, and the larger the value, the greater the "purity improvement" obtained by division. Select the test condition with the largest Gain as the division condition.

Step 20324, according to test conditions will Divide into two subsets and build two child nodes.

Step 20325, for each subset, repeat step 322 and step 323.

On the basis of the above-mentioned embodiments, step 4 in the method specifically includes:

Using the combined base classifier T ^*

Classify a data subset x under a specific data attribute, where δ(true)=1, δ(false)=0, T _k (x) is the prediction of the decision tree T _k on the data subset x under a specific data attribute.

On the basis of the above-mentioned embodiments, the subset of data attributes in this method includes: asset-liability ratio, return on net assets, return on assets, net profit rate, total asset turnover, accounts receivable turnover, current assets At least one data attribute among turnover rate, growth rate of main business, growth rate of total assets and growth rate of net profit.

Fig. 3 is a schematic flowchart of a multi-decision tree financial early warning method provided by another embodiment of the present invention. First, randomly sample the financial data sample D to be tested to obtain training subsets: D ₁ , D ₂ , D ₃ ... D _k ; secondly, obtain weight W ^(k,1) and bias item b ^{(k,l )} ; Then, the above training subsets are encoded by the preset noise reduction autoencoding model: D ₁ , D ₂ , D ₃ ... D _k to obtain a new training subset: After that, iteratively learn multiple decision trees on the new training subset: T ₁ , T ₂ , T ₃ ... T _k ; finally use the decision tree model to predict the financial status category of each data subset, and pass the majority vote The method combines these forecast results to obtain the final financial warning situation.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it can still be Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent replacements are made to some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims (6)

1. A multi-decision tree financial early warning method is characterized in that, comprising: Step 1. Obtain a financial data sample set D to be detected, the financial data sample set D to be detected includes data attribute subsets and data subsets corresponding to each data attribute in the data attribute subsets; Step 2. Perform M random sampling on the financial data sample set D to be tested according to the preset random sampling times M, extract M data each time, and determine the sample set obtained by the kth random sampling as the training subset D _k , where k=1,2,3...M; Step 3, using the preset noise reduction autoencoder model to learn the decision tree T _{k on the training subset D k ;} Step 4. According to the decision tree T _k , predict the financial status category of each data subset in the financial data sample set D to be detected, and the financial status category includes good financial status and abnormal financial status. 2. The method according to claim 1, wherein said step 3 specifically comprises: Step 31. Use the preset noise reduction autoencoder model Encode the training subset D _k in the l-layer neural network to obtain a new training subset Where W ^{(k, 1)} is the weight of the l-th layer neural network, and b ^{(k, l)} is the bias item of the l-th layer neural network. Step 32, in the new training subset The decision tree model T _k is obtained through learning above. 3. method according to claim 2, is characterized in that, weight W ^{(k, 1)} and bias item b ^(k, 1) in the described step 31 are according to following formula respectively: Determined; where, λ is the weight decay parameter, x is the data subset under the specific data attribute, is a subset of data under a specific data attribute after adding noise, represents the output of the output layer, and represent the partial derivatives of the cost function J(W ^(k,l-1) ,b ^(k,l-1) ) with respect to W ^(k,l-1) and b ^(k,l-1) respectively. 4. The method according to claim 2, wherein said step 32 specifically comprises: Step 321, initialize the generation node node as the root node; Step 322, if If the financial data samples belong to a specific financial status category or the number of financial data samples is less than a preset threshold, the node is marked as a leaf node of the specific financial status category; Step 323, if different financial data samples belong to a specific financial status category and the number of financial data samples is not less than the preset threshold, then use the objective function Search test conditions, where D _t is the data set corresponding to the current node t, D _tj is the data set of the child node j of the current node t, and Ent(D _tj ) is the information entropy of D _tj ; Step 324, according to test condition will Divide into two subsets and build two child nodes. Step 325, for each subset, repeat step 322 and step 323. 5. The method according to claim 1, wherein said step 4 specifically comprises: Using the combined base classifier T ^* Classify a data subset x under a specific data attribute, where δ(true)=1, δ(false)=0, T _k (x) is the prediction of the decision tree T _k on the data subset x under a specific data attribute. 6. The method according to any one of claims 1-5, wherein the data attribute subset includes: asset-liability ratio, return on net assets, return on assets, net profit rate, total asset turnover, should At least one data attribute among the turnover rate of accounts collection, turnover rate of current assets, growth rate of main business, growth rate of total assets and growth rate of net profit.