Financial management improvement method based on artificial intelligence

Abstract

The invention discloses a financial management improvement method based on artificial intelligence, and the method comprises the following analysis steps: S1, carrying out the collection of the income and expenditure data of a financial system, and carrying out the preprocessing of the data; s2, performing classification processing on the income and expenditure data through clustering analysis; s3, constructing a differential autoregressive moving average (ARIMA) model, extracting a linear feature relationship, and performing income and expenditure prediction; s4, constructing a long short-term memory neural network LSTM model, extracting a nonlinear feature relationship and performing income and expenditure prediction; and S5, constructing a combined model based on the weight ratio, and fusing the feature relationships of the two single prediction models to obtain a final prediction result. S6, performing result verification on the test set data by using the obtained combined model, and predicting income and expenditure of the next month; according to the method, on the basis of processing and clustering analysis of the income and expenditure data, the models are established respectively to extract the linear and nonlinear characteristic relation of the income and expenditure data, fusion prediction is carried out, the financial budget precision is improved, and prediction errors are reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of financial management, and particularly relates to a financial management improvement method based on artificial intelligence. BACKGROUND

[0002] In enterprise financial management, the accuracy of financial budget and the control level of cash flow are key factors affecting the operation efficiency and risk control of enterprises. Traditional financial budget and cash flow management often rely on historical data and manual experience, and there are problems such as low data processing efficiency, inaccurate prediction, lack of dynamic adjustment, etc. Therefore, it is very important to study intelligent financial management methods to improve the accuracy of financial budget, strengthen dynamic prediction of cash flow, reduce human errors, improve data efficiency, optimize fund allocation and reduce financial risks.

[0003] Therefore, in order to solve the above problems, it is necessary to develop a financial management improvement method based on artificial intelligence. SUMMARY

[0004] The purpose of the application is to overcome the shortcomings of the prior art and provide a financial management improvement method based on artificial intelligence. Based on the processing and clustering analysis of income and expenditure data, models are established to extract the linear and nonlinear characteristic relationships of the income and expenditure data, and fusion prediction is performed to improve the accuracy of financial budget and reduce prediction errors.

[0005] The purpose of the application is achieved by a financial management improvement method based on artificial intelligence, comprising the following analysis steps:

[0006] S1, respectively collecting the income and expenditure data of the financial system, and preprocessing the data, including data cleaning, deduplication and completion;

[0007] S2, respectively clustering the income and expenditure data based on the Gaussian mixture model clustering algorithm, classifying the income and expenditure data, and determining the characteristics of different types of data;

[0008] S3, constructing a difference autoregressive moving average ARIMA model to extract the linear characteristic relationship of the income and expenditure data and make income and expenditure prediction;

[0009] S4, optimizing the hyperparameters of the long short-term memory neural network LSTM by improving the particle swarm algorithm, constructing an optimized long short-term memory neural network LSTM model, extracting the nonlinear characteristic relationship of the income and expenditure data, and making income and expenditure prediction;

[0010] S5. Based on the prediction results of the two single prediction models, ARIMA and LSTM, establish an optimization objective function, solve for the weights of the two single prediction models, construct a combined model based on the weight ratio, fuse the feature relationships of the two single prediction models, and obtain the final prediction result.

[0011] S6. Validate the obtained combined model against the test set data and predict the income and expenditure for the next month.

[0012] Furthermore, in step S1, data cleaning is used to remove invalid, erroneous, or duplicate data records; data deduplication is used to ensure the uniqueness of data and avoid duplicate data interfering with the analysis results; data completion is used to complete missing data to ensure data integrity, specifically by using the mean method or interpolation method.

[0013] Furthermore, in step S2, a Gaussian mixture model clustering algorithm is used to classify the income and expenditure data. Therefore, for a certain data sample set... Assuming the sample follows a Gaussian mixture distribution, the Gaussian mixture distribution contains a total of A Gaussian distribution model, specifically represented as: , This represents the Gaussian probability density function. , Represents random variables dimensionality 3D random variables , Indicates the first The weight coefficients of each component model , Indicates the first The mean vector of each component model Indicates the first The covariance matrix of each component model.

[0014] Furthermore, step S2 involves estimating the parameters of the Gaussian mixture model using the maximum likelihood method, specifically including the following steps:

[0015] ① Initialization ① Calculate the parameters of each Gaussian distribution model; ② Calculate the probability of generating each data point under each mixture component model, expressed as: In the formula: Indicates sample A certain data in, Representing sample data In the The probability density function under the component model; ③ The expectation-maximization algorithm is used to estimate each parameter, specifically expressed as: , , , and according to the estimated model parameter value, the parameter of the first component model is updated , , ; ④the model parameter updated in step ③ is brought into step ② for calculation, and whether the maximum likelihood function of the Gaussian mixture model reaches an extreme value is judged, which is expressed as: , if the extreme value is reached, the parameter value of each component model finally trained is output, and if the extreme value is not reached, steps ② and ③ are repeated; ⑤according to the final parameter, the probability value of each sample under each component model is calculated, and each sample is divided into corresponding models according to the probability value, and the clustering result is obtained.

[0016] Further, the step S3 of constructing the difference autoregressive moving average model specifically includes the following steps:

[0017] ①determine whether the time series of the income and expenditure data is a pure random sequence, specifically by white noise test Q statistic , if the data is a white noise sequence, it is a pure random sequence, and there is no correlation between the data, and the process is ended, if the data is not a white noise sequence, there is a linear feature for extraction, and the next step is performed;

[0018] ②the stationarity of the time series of the income and expenditure data is tested by ADF unit root test method, if it is a stationary sequence, the next step is performed, if it is a non-stationary sequence, difference method is used to stabilize the sequence, and after it is a stationary sequence, step ① is returned to continue the judgment according to the process;

[0019] ③analyze the autocorrelation function ACF and the partial autocorrelation function PACF of the model, determine the model parameters suitable for the current sequence for model identification, including difference parameter , autoregressive order and moving average order ;

[0020] ④the optimal order of the model is automatically identified by using grid search, the AIC value of each model is calculated, the parameter corresponding to the model with the minimum AIC value is selected to construct the ARIMA model, and the model order is determined;

[0021] ⑤the model fitting effect is judged by white noise test of the model residual, if it passes the residual test, the next step is performed, if it does not pass, steps ③ and ④ are returned to identify and determine the order of the model again, until it passes the residual test; when the residual sequence obeys the normal distribution and presents the white noise characteristics, it means that the model fitting is good enough, and the useful information is fully extracted;

[0022] ⑥the ARIMA model is constructed, and the time series sequence is predicted. ​

[0023] Further, the step S4 of optimizing the hyperparameters of the long short-term memory neural network LSTM by the improved particle swarm algorithm includes optimizing and solving the step length, learning rate, batch size and number of hidden layer neurons of the LSTM model.

[0024] Further, the step S4 of optimizing the parameters of the long short-term memory neural network LSTM by the improved particle swarm algorithm includes the following steps:

[0025] ① initializing the particle parameters, including the initial position of each particle in the particle swarm and velocity , learning factor and , inertia weight , number of particles, maximum number of iterations, solution range of parameters and particle dimension; ② taking the root mean square error MSE of the loss function of LSTM on the training set as the fitness function, calculating the fitness value of the particle; ③ updating the speed and position information of each particle, and updating the historical optimal position and global optimal position; ④ judging whether the iteration termination condition is met, if not, returning to step ② for continuous calculation, if yes, proceeding to the next step; ⑤ outputting the fitness value and optimal position information, i.e. the optimal parameters of LSTM obtained by solving, and constructing the LSTM model according to the optimal parameters for prediction.

[0026] Further, the fitness function in step ② is expressed as: , , respectively represent the predicted value and the true value; the speed updating formula of the particle in step ③ is expressed as: , wherein: , respectively represent the speed of the particle , , w represents the inertia weight, , , respectively represent the maximum and minimum inertia weight, , respectively represent the current and maximum number of iterations, represents the control factor, represents the learning factor, represents a random number between 0 and 1, represents the historical optimal position of the particle individual, represents the global optimal position of the group; the position updating formula is expressed as: , respectively represent the particle , The position of the user equipment.

[0027] Further, the step S5 is based on the error square sum of the combination model to establish an optimization objective function, which is specifically represented as: , wherein: represents the length of the time series, represents the prediction error of the combination model at the time, , , respectively represents the prediction error of the ARIMA model and the LSTM model at the time, , respectively represents the weight of the ARIMA model and the LSTM model, and satisfies , the Lagrange multiplier method is used to solve , and , .

[0028] Due to the adoption of the above technical solutions, the application has the following beneficial effects:

[0029] (1) The data cleaning, deduplication and completion processing are performed on the income and expenditure data, and the Gaussian mixture model clustering algorithm is used to perform clustering analysis on the income and expenditure data, so that the income and expenditure data are classified and processed, the characteristics of different types of data are determined, and the data processing efficiency is improved.

[0030] (2) The linear feature relationship and the nonlinear feature relationship of the income and expenditure data are extracted by using the difference autoregressive moving average ARIMA model and the long short-term memory neural network LSTM model respectively, and the two kinds of feature relationships are fused based on the weight ratio, a combination model is constructed for prediction, the prediction advantages of the two kinds of models are fused, the prediction accuracy is effectively improved, and the prediction effect is improved. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is a flowchart of the application. DETAILED DESCRIPTION

[0032] The technical solutions of the application will be further specifically described below by examples in combination with the drawings.

[0033] As shown in Figure 1 , a financial management improvement method based on artificial intelligence includes the following analysis steps:

[0034] S1, the income and expenditure data of the financial system are collected respectively, and the data is preprocessed, including data cleaning, deduplication and completion.

[0035] Preferably, in step S1, data cleaning is used to remove invalid, erroneous, or duplicate data records; data deduplication is used to ensure the uniqueness of data and avoid duplicate data interfering with the analysis results; data completion is used to complete missing data to ensure data integrity, specifically by using the mean method or interpolation method.

[0036] S2. Cluster analysis of income and expenditure data is performed based on Gaussian mixture model clustering algorithm to classify the income and expenditure data and determine the characteristics of different types of data.

[0037] Preferably, in step S2, a Gaussian mixture model clustering algorithm is used to classify the income and expenditure data. Therefore, for a given data sample set... Assuming the sample follows a Gaussian mixture distribution, the Gaussian mixture distribution contains a total of A Gaussian distribution model, specifically expressed as:

[0038]

[0039] In the formula: This represents the Gaussian probability density function.

[0040]

[0041] , Represents random variables dimensionality 3D random variables , Indicates the first The weight coefficients of each component model , Indicates the first The mean vector of each component model Indicates the first The covariance matrix of each component model.

[0042] Preferably, step S2 involves estimating the parameters of the Gaussian mixture model using the maximum likelihood method, specifically including the following steps:

[0043] ① Initialization The parameters of a Gaussian distribution model.

[0044] ② Calculate the probability of each data point being generated under each mixture component model, expressed as:

[0045]

[0046] In the formula: Indicates sample A certain data in, Representing sample data In the the probability density function under the component model.

[0047] ③ The expectation maximization algorithm is used to estimate each parameter, which is specifically represented as:

[0048]

[0049]

[0050]

[0051] and according to the estimated model parameter value, the first component model parameter , , .

[0052] ④ The model parameters updated in step ③ are brought into step ② for calculation to determine whether the maximum likelihood function of the Gaussian mixture model has reached an extreme value, which is represented as:

[0053]

[0054] If the extreme value is reached, the parameter values of each component model obtained after final training are output, and if the extreme value is not reached, steps ② and ③ are repeated.

[0055] ⑤ According to the final parameters, the probability value of each sample under each component model is calculated, and each sample is divided into corresponding models according to the probability value to obtain the clustering result.

[0056] S3, construct a difference autoregressive moving average ARIMA model, extract the linear characteristic relationship of the income and expenditure data, and perform income and expenditure prediction.

[0057] Preferably, the step S3 of constructing the difference autoregressive moving average model specifically includes the following steps:

[0058] ① Determine whether the time series of the income and expenditure data is a pure random sequence, specifically by white noise test Q statistic , if the data is a white noise sequence, it is a pure random sequence, and there is no correlation between the data, and the process is ended, if the data is not a white noise sequence, there is a linear characteristic for extraction, and the next step is performed.

[0059] Preferably, the white noise test Q statistic is specifically represented as: , wherein: represents the observation period of the sequence, represents the specified maximum delay period, represents the delay order, represents the delay the estimated value of the coefficient of the sample sub-item, , representing the sample autocovariance of the time series at lag , representing the i-th observation of the time series, representing the mean of the time series, representing the variance of the time series.

[0060] ② The stationarity of the time series of the income and expenditure data is tested by the ADF unit root test method. If it is a stationary series, the next step is performed. If it is a non-stationary series, the sequence is stabilized by using the difference method until it is a stationary sequence, and then step ① is returned to continue the judgment according to the process.

[0061] Preferably, the ADF unit root test method is specifically represented as: , wherein, 、 represent the values of the time series at 、 , represent the coefficient of the lag value in the regression of the lag value, represent the i-th difference value of , i.e. the difference value between and , represent the selected lag order, represent the lag difference coefficient of the regression of the lag difference, represent the residual error at .

[0062] ③ The autocorrelation function ACF and the partial autocorrelation function PACF of the model are analyzed to determine the model parameters suitable for the current sequence for model identification, including the difference parameter , the autoregressive order and the moving average order .

[0063] ④ The optimal order of the model is automatically identified by using the grid search, and the Akaike information criterion AIC value of each model is calculated, , wherein: represents the number of model parameters, represents the likelihood function, and the parameters corresponding to the model with the minimum AIC value are selected to construct the ARIMA model, so as to realize model order determination.

[0064] ​​​​⑤ Perform a white noise test on the residuals of the model to determine the model fit. If the residual test is passed, proceed to the next step. If it is not passed, return to steps ③ and ④ to re-identify and determine the order of the model until the residual test is passed. When the residual sequence follows a normal distribution and exhibits white noise characteristics, it indicates that the model fit is good enough and the useful information has been fully extracted.

[0065] ⑥ Once the ARIMA model is built, time series prediction is performed.

[0066] S4. The hyperparameters of the Long Short-Term Memory Neural Network (LSTM) are optimized by improving the particle swarm optimization algorithm. An optimized LSTM model is constructed to extract the nonlinear feature relationships of income and expenditure data and to predict income and expenditure.

[0067] Preferably, in step S4, the hyperparameters of the Long Short-Term Memory Neural Network (LSTM) are optimized by improving the particle swarm optimization algorithm, including optimizing the step size, learning rate, batch size, and number of hidden layer neurons of the LSTM model.

[0068] Preferably, step S4 optimizes the parameters of the Long Short-Term Memory (LSTM) neural network using an improved particle swarm optimization algorithm, specifically including the following steps:

[0069] ① Initialize particle parameters, including the initial position of each particle in the particle swarm. and speed It also includes learning factors and Inertia weight The number of particles, the maximum number of iterations, the solution range of parameters, and the particle dimension.

[0070] ② Using the root mean square error (MSE) of the LSTM loss function on the training set as the fitness function, calculate the fitness value of the particles; preferably, the fitness function in step ② is expressed as:

[0071]

[0072] , , These represent the predicted value and the actual value, respectively.

[0073] ③ Update the velocity and position information of each particle, and update the historical best position and the global best position; preferably, the particle velocity update formula in step ③ is expressed as:

[0074]

[0075] In the formula: , They represent particles respectively , speed, Indicates inertia weight, , These represent the maximum and minimum inertia weights, respectively. , These represent the current and maximum number of iterations, respectively. Indicates control factor. Represents the learning factor. Represents a random number between 0 and 1. This represents the historical best position of an individual particle. This represents the global optimal position of the group; the position update formula is expressed as:

[0076]

[0077] , They represent particles respectively , The location.

[0078] ④ Determine if the iteration termination condition is met. If not, return to step ② to continue calculation. If it is met, proceed to the next step.

[0079] ⑤ Output the fitness value and the optimal location information, that is, the optimal parameters of the LSTM obtained by solving. Construct an LSTM model based on the optimal parameters and make predictions.

[0080] S5. Based on the prediction results of the two single prediction models, ARIMA and LSTM, establish an optimization objective function, solve for the weights of the two single prediction models, construct a combined model based on the weight ratio, and fuse the feature relationships of the two single prediction models to obtain the final prediction result.

[0081] Preferably, in step S5, the optimization objective function is established based on minimizing the sum of squared errors of the combined model, specifically expressed as follows: In the formula: Indicates the length of the time series. Indicating the combination model in Prediction error at time, , These represent the ARIMA model and the LSTM model respectively. Prediction error at time, Let the weights of the ARIMA model and the LSTM model be respectively, satisfying the following conditions: Using the Lagrange multiplier method to Solving the problem yields the following results: , .

[0082] S6, the obtained combination model is used to check the result of the test set data, and the income and expenditure of the next month are predicted.

[0083] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application, and are not limited thereto. Although the present application has been described in detail with reference to the above examples, those skilled in the art should understand that the specific embodiments of the present application can still be modified or replaced equivalently without departing from the spirit and scope of the present application, and any modification or equivalent replacement without departing from the spirit and scope of the present application should be covered in the scope of the claims of the present application.

Claims

1. An artificial intelligence-based financial management improvement method, characterized by: The method comprises the following analysis steps: S1, collecting and preprocessing the income and expenditure data of the financial system, including data cleaning, deduplication and completion; S2, clustering analysis of the income and expenditure data based on Gaussian mixture model clustering algorithm, classifying the income and expenditure data and determining the characteristics of different types of data; S3, constructing a difference autoregressive moving average ARIMA model to extract the linear characteristic relationship of the income and expenditure data and make income and expenditure prediction; S4, optimizing the hyperparameters of the long short-term memory neural network LSTM by improving the particle swarm algorithm, constructing an optimized long short-term memory neural network LSTM model, extracting the nonlinear characteristic relationship of the income and expenditure data, and making income and expenditure prediction; S5, establishing an optimization objective function according to the prediction results of the ARIMA model and the LSTM model, solving the weights of the two single prediction models, constructing a combination model based on the weight ratio, fusing the characteristic relationships of the two single prediction models, and obtaining the final prediction result; S6, verifying the test set data with the obtained combination model and predicting the income and expenditure in the next month.

2. The method of claim 1, wherein: In step S1, data cleaning is used to eliminate invalid, erroneous or duplicate data records; data deduplication is used to ensure the uniqueness of data and avoid interference of repeated data on the analysis result; data completion is used to complete the missing data to ensure the integrity of the data, and the mean method or interpolation method is used for filling.

3. The method of claim 1, wherein: The step S2 adopts a Gaussian mixture model clustering algorithm to classify the income and expenditure data. For a data sample set , assuming that the sample is subject to a mixed Gaussian distribution, the mixed Gaussian distribution has Gaussian distribution models, which are specifically represented as: , represents a Gaussian probability density function, , represents the dimension of a random variable , a random variable , represents the weight coefficient of the th component model, , represents the mean vector of the th component model, represents the covariance matrix of the th component model.

4. The method of claim 3, wherein: In step S2, the parameters of the mixture model are estimated using the maximum likelihood method, specifically including the following steps: ① Initialization ① Calculate the parameters of each Gaussian distribution model; ② Calculate the probability of generating each data point under each mixture component model, expressed as: In the formula: Indicates sample A certain data in, Representing sample data In the The probability density function under the component model; ③ The expectation-maximization algorithm is used to estimate each parameter, specifically expressed as: , , And based on the estimated model parameter values, update the first... Parameters of a component model , , ④ Substitute the updated model parameters from step ③ into step ② for calculation, and determine whether the maximum likelihood function of the Gaussian mixture model has reached an extreme value, expressed as: If an extreme value is obtained, output the parameter values ​​of each component model obtained from the final training. If no extreme value is obtained, repeat steps ② and ③. ⑤ Based on the final parameters, calculate the probability value of each sample under each component model, and classify each sample into the corresponding model according to the probability value to obtain the clustering result.

5. The method of claim 1, wherein: The step S3 of constructing the difference autoregressive moving average model specifically comprises the following steps: ① judging whether the time series of the revenue and expenditure data is a pure random sequence, specifically by judging through a white noise test Q statistic . If the data is a white noise sequence, it is a pure random sequence, and there is no correlation between the data. If the data is not a white noise sequence, there is a linear feature for extraction, and the next step is performed; ② performing an ADF unit root test on the stationarity of the time series of the revenue and expenditure data. If it is a stationary sequence, the next step is performed. If it is a non-stationary sequence, a difference method is used to stabilize the sequence until it is a stationary sequence, and then the step ① is returned to continue the judgment according to the flow; ③ analyzing the autocorrelation function ACF and the partial autocorrelation function PACF of the model to determine the model parameters suitable for the current sequence for model identification, including the difference parameter , the autoregressive order , and the moving average order ; ④ using a grid search to automatically identify the optimal order of the model. The AIC value of each model is calculated, and the parameters corresponding to the model with the minimum AIC value are used to construct an ARIMA model to realize model order determination; ⑤ performing a white noise test on the residual of the model to judge the fitting effect of the model. If the residual test is passed, the next step is performed. If the residual test is not passed, the model is re-identified and ordered in the steps ③ and ④ until the residual test is passed; when the residual sequence obeys a normal distribution and presents a white noise characteristic, it indicates that the model fitting is good enough, and useful information is fully extracted; ⑥ the ARIMA model is constructed, and the time series sequence is predicted.

6. The method of claim 1, wherein: In step S4, the hyperparameters of the long short-term memory neural network LSTM are optimized by improving the particle swarm algorithm, including optimizing and solving the step length, learning rate, batch size and number of hidden layer neurons of the LSTM model.

7. The method of claim 1, wherein: The step S4 is to optimize the parameters of the long short-term memory neural network LSTM by improving the particle swarm algorithm, and specifically includes the following steps: ①initializing particle parameters, including the initial position and speed of each particle in the particle swarm , a learning factor , , an inertia weight , the number of particles, the maximum number of iterations, the solution range of the parameters, and the particle dimension; ②taking the root mean square error MSE of the loss function of the LSTM on the training set as a fitness function, calculating the fitness value of the particle; ③updating the speed and position information of each particle, and updating the historical optimal position and the global optimal position; ④judging whether the iteration termination condition is met, if not, returning to step ② for continuous calculation, and if yes, proceeding to the next step; ⑤outputting the fitness value and the optimal position information, i.e. the optimal parameters of the LSTM obtained by solving, and constructing an LSTM model according to the optimal parameters for prediction.​ 8. The method of claim 7, wherein: The fitness function in step ② is expressed as: , , respectively represent the predicted value and the true value; the velocity updating formula of the particle in step ③ is expressed as: , wherein: , respectively represent the velocity of the particle , , represents the inertial weight, , respectively represent the maximum and minimum inertial weight, , respectively represent the current and maximum iteration number, represents the control factor, represents the learning factor, represents a random number between 0 and 1, represents the historical optimal position of the particle individual, represents the global optimal position of the group; the position updating formula is expressed as: , respectively represent the position of the particle , .

9. The method of claim 1, wherein: The step S5 is based on the combination model error square sum minimum optimization objective function, and is specifically represented as: , wherein: represents the time series length, represents the prediction error of the combination model at time, , respectively represents the prediction error of the ARIMA model and the LSTM model at time, respectively represents the weight of the ARIMA model and the LSTM model, and satisfies , the Lagrange multiplier method is used to solve , and , is obtained.

Images