Quantitative financial data classification evaluation method and system based on artificial intelligence

Abstract

The invention discloses a quantitative financial data classification evaluation method and system based on artificial intelligence, and belongs to the field of financial science and technology, and the method comprises the steps: collecting financial data according to a pre-configured financial data source, and carrying out the preprocessing of the financial data; performing data standardization processing on the financial original data to obtain financial standard data, and constructing a financial knowledge graph based on the financial standard data; performing feature extraction on the financial knowledge graph to obtain financial feature data; a financial time prediction model based on the CEEMDAN-LSTM-ATTE is constructed; inputting the financial feature data into a financial time prediction model for training to obtain a prediction result; and evaluating the prediction result according to a preset scoring grade. According to the quantitative financial data classification evaluation method and system based on artificial intelligence provided by the invention, the future risk trend can be predicted according to the historical data and the current market condition, corresponding countermeasures are formulated in advance, and meanwhile, the prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of financial technology, and in particular to an artificial intelligence-based quantitative financial data classification and evaluation method and system. Background Art

[0002] Financial data refers to market data, company data, industry indices, and pricing data related to the financial industry. All data related to the financial industry can be incorporated into financial market big data systems, providing a reference for market analysis. There are many academic definitions, but a more general example is that the price data of a particular stock over a period of time, arranged in a predetermined chronological order, is considered financial time series data.

[0003] With existing technologies, improving financial data analysis systems requires continuously acquiring and analyzing data from specific financial websites to make the financial data more structured and contextualized. Once integrated, this data should be promptly stored in the target server's database and made available to users. Currently, with the continuous development of internet technology, financial data analysis systems are becoming increasingly technical. The innovation and integration of information systems enable financial institutions to process large amounts of data and information in their daily operations, enabling better financial management. In the course of financial transactions, accurate financial data provides users with a foundation for future financial decisions, charts the course for future financial careers, and plays a crucial role in the development of the financial industry.

[0004] Therefore, how to improve the efficiency of financial data analysis technology and make its operation easier plays an important role in improving the financial data analysis system. Summary of the Invention

[0005] The purpose of the present invention is to provide a quantitative financial data classification and evaluation method and system based on artificial intelligence to solve the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention provides an artificial intelligence-based quantitative financial data classification system, comprising:

[0007] A data collection and preprocessing module, used to collect financial data from pre-configured financial data sources and perform preprocessing;

[0008] A graph construction module is used to standardize financial data, obtain financial standard data, and construct a financial knowledge graph based on the financial standard data;

[0009] Feature extraction module, used to extract features from the financial knowledge graph, obtain financial feature data, and save it to the database;

[0010] Model building module, building a financial time prediction model based on CEEMDAN-LSTM-ATTE;

[0011] The prediction module inputs financial feature data into the financial time prediction model for training to obtain prediction results.

[0012] Preferably, the prediction module is further used to extract financial sample data from a preset sample database, and extract sample index items based on the financial sample data;

[0013] Calculate the sample scoring results of the sample indicator items according to the preset scoring rules;

[0014] The financial sample data is used as the input of the preset training model, and the sample risk score result is used as the output of the preset training model;

[0015] The preset training model is trained according to the input results and output results to obtain a financial time prediction model.

[0016] Preferably, the graph construction module is further used to clean the financial data to obtain financial filtered data;

[0017] Perform data adaptive filling on financial filtered data to obtain financial completed data;

[0018] Perform knowledge extraction on financial supplementary data to obtain financial standard data.

[0019] Preferably, the graph construction module is also used to extract triples from financial standard data according to a preset financial knowledge graph framework, and embed the triples into the preset financial knowledge graph framework to obtain a financial knowledge graph.

[0020] Preferably, the prediction module is further used to input the financial feature data into the financial time prediction model for scoring processing to obtain an evaluation result.

[0021] Preferably, the steps of the CEEMDAN-LSTM-ATTE financial time prediction model are as follows:

[0022] The CEEMDAN decomposition of the financial time series yields a series of intrinsic mode functions IMF1, IMF2, ..., IMF n and a residual sequence Res;

[0023] The subsequences are normalized and then used as input variables of the LSTM-ATTE model with different parameters for prediction, and a series of prediction values ​​F1, F2, ..., F are obtained. n , F n+1 ;

[0024] The predicted values ​​of each subsequence are then denormalized and superimposed to obtain the final prediction result.

[0025] A quantitative financial data classification and evaluation method based on artificial intelligence includes the following steps:

[0026] S1. Collect financial data and pre-process it according to pre-configured financial data sources;

[0027] S2. Perform data standardization on the original financial data to obtain financial standard data, and build a financial knowledge graph based on the financial standard data;

[0028] S3. Extract features from the financial knowledge graph to obtain financial feature data;

[0029] S4. Build a financial time prediction model based on CEEMDAN-LSTM-ATTE;

[0030] S5. Input the financial feature data into the financial time prediction model for training to obtain prediction results;

[0031] S6. Evaluate the prediction results according to the preset scoring levels.

[0032] Therefore, the present invention utilizes the aforementioned AI-based quantitative financial data classification and evaluation method and system to collect, process, and analyze various financial data in real time, identifying potential risk points and implementing appropriate preventative and countermeasure measures. By utilizing big data analysis technology, financial institutions can rapidly process and analyze massive amounts of data, identifying patterns and anomalies within the data and accurately assessing risk. Furthermore, big data analysis technology can also quantitatively assess risk, providing financial institutions with a more scientific basis for decision-making. By learning and simulating the human decision-making process, it can predict future risk trends based on historical data and current market conditions, and formulate appropriate countermeasures in advance.

[0033] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flow chart of a method for classifying and evaluating quantitative financial data based on artificial intelligence according to the present invention;

[0035] Figure 2 This is a schematic diagram of a combined prediction model based on CEEMDAN-LSTM-ATTE according to an embodiment of the present invention;

[0036] Figure 3 Schematic diagram of the implementation of the attention mechanism of an embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram of the original network according to an embodiment of the present invention;

[0038] Figure 5 A schematic diagram of an improved network with an attention mechanism added to an embodiment of the present invention;

[0039] Figure 6 This is a schematic diagram of the CEEMDAN decomposition results of the Shanghai Composite 50 Index according to an embodiment of the present invention;

[0040] Figure 7 This is a schematic diagram of the IMF component prediction results of CEEMDAN according to an embodiment of the present invention;

[0041] Figure 8 This is a schematic diagram of the prediction results of the Shanghai Composite 50 Index CEEMDAN-LSTM-ATTE according to an embodiment of the present invention;

[0042] Figure 9 Schematic diagram of linear regression analysis of the Shanghai Composite 50 Index according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort shall fall within the scope of protection of the present invention.

[0044] See also Figure 1 , a quantitative financial data classification and evaluation method based on artificial intelligence, comprising the following steps:

[0045] S1. Collect financial data and pre-process it according to pre-configured financial data sources;

[0046] S2. Perform data standardization on the original financial data to obtain financial standard data, and build a financial knowledge graph based on the financial standard data;

[0047] S3. Extract features from the financial knowledge graph to obtain financial feature data;

[0048] S4. Build a financial time prediction model based on CEEMDAN-LSTM-ATTE;

[0049] S5. Input the financial feature data into the financial time prediction model for training to obtain prediction results;

[0050] S6. Evaluate the prediction results according to the preset scoring levels.

[0051] An artificial intelligence-based quantitative financial data classification system, comprising:

[0052] A data collection and preprocessing module, used to collect financial data from pre-configured financial data sources and perform preprocessing;

[0053] A graph construction module is used to standardize financial data, obtain financial standard data, and construct a financial knowledge graph based on the financial standard data;

[0054] Feature extraction module, used to extract features from the financial knowledge graph, obtain financial feature data, and save it to the database;

[0055] Model building module, building a financial time prediction model based on CEEMDAN-LSTM-ATTE;

[0056] The prediction module inputs financial feature data into the financial time prediction model for training to obtain prediction results.

[0057] The prediction module is further used to extract financial sample data from a preset sample database and extract sample indicator items based on the financial sample data;

[0058] Calculate the sample scoring results of the sample indicator items according to the preset scoring rules;

[0059] The financial sample data is used as the input of the preset training model, and the sample risk score result is used as the output of the preset training model;

[0060] The preset training model is trained according to the input results and output results to obtain a financial time prediction model.

[0061] The graph construction module is also used to clean financial data and obtain financial filtered data;

[0062] Perform data adaptive filling on financial filtered data to obtain financial completed data;

[0063] Perform knowledge extraction on financial supplementary data to obtain financial standard data.

[0064] The graph construction module is also used to extract triples from financial standard data according to the preset financial knowledge graph framework, and embed the triples into the preset financial knowledge graph framework to obtain a financial knowledge graph.

[0065] The prediction module is also used to input financial feature data into the financial time prediction model for scoring and obtaining evaluation results.

[0066] like Figure 2 , the steps of the CEEMDAN-LSTM-ATTE financial time prediction model are as follows:

[0067] The CEEMDAN decomposition of the financial time series yields a series of intrinsic mode functions IMF1, IMF2, ..., IMF n and a residual sequence Res;

[0068] The subsequences are normalized and then used as input variables of the LSTM-ATTE model with different parameters for prediction, and a series of prediction values ​​F1, F2, ..., F are obtained. n , F n+1 ;

[0069] The predicted values ​​of each subsequence are then denormalized and superimposed to obtain the final prediction result.

[0070] The attention mechanism originates from the selective attention mechanism of human vision. It scans the entire image, identifies areas of focus, and devotes more attention resources to obtaining more target information while ignoring other information. Therefore, the core of the attention mechanism in deep learning is to identify the correlation between the predicted target and input features. When neural network models process large amounts of input information, the attention mechanism can be used to selectively process key input information, thereby improving the efficiency of the network.

[0071] The calculation of the attention mechanism is the weighted sum of the input vector. According to Bahdanau’s definition, assuming that the hidden layer state at each time step is h t The calculation process of the model based on the attention mechanism is as follows:

[0072] (1) Calculate the given h j value and target state s t-1 Similarity, that is, the state h at each time t j Weight:

[0073] e tj =a(s t-1 ,h j );

[0074] (2) Normalized weight coefficient α tj :

[0075]

[0076] (3) For state h t Weighted average:

[0077]

[0078] Where T is the total number of time steps in the input sequence; a(·) is called the alignment model, which is a nonlinear function used to measure h j With st-1 The higher the matching degree, the higher the attention distribution; c t It is the "context" vector based on the attention mechanism. The implementation process of the attention mechanism is as follows Figure 3 shown.

[0079] As research on the attention mechanism deepens, it has spawned numerous variants. Among them, self-attention can extract relevant information from its own information, demonstrating promising results in reading comprehension, text inheritance, and text summarization. As the name suggests, self-attention captures dependencies within a time series, making it easier to capture long-range interdependencies. While the recurrent structures of LSTMs and RNNs can capture long-range dependencies, this requires information accumulation over several time steps, and the longer the time step, the less effective information can be captured. Therefore, an improved LSTM algorithm based on the attention mechanism is proposed.

[0080] In the improved algorithm, an attention mechanism is added after the LSTM layer to establish long-distance dependencies between input time series, to "dynamically" generate weights of different connections, and thus process variable-length information sequences. Figure 4 As shown, in the traditional LSTM network, the output h at the last moment is usually used t To calculate the C vector, which is the input vector of the next layer of neural network, this method loses the output information of the LSTM layer at other times. Figure 5 As shown, in the improved network with attention mechanism, all outputs h1 to h t , are input into the Attention layer, and the new output is obtained by weighted summation. vector, this improved network model is referred to as LSTM-ATTE.

[0081] The steps of CEEMDAN decomposition are as follows:

[0082] First, add a series of adaptive white noise to the original time series x(t):

[0083] x i (t)=x(t)+ω0ε i (t),i∈{1,…,I};

[0084] Among them, x i (t) is the time series of adding white noise for the i-th time, ω0 is the noise coefficient, ε i (t) is the white noise added for the i-th time, and I is the number of integrations, which is generally a very small number (between 10 and 20).

[0085] Decompose the new time series x using the EMD algorithm i (t), for the first IMF component decomposed Take the mean:

[0086]

[0087] Removing c1(t) from the original time series x(t) yields the first residual series:

[0088] r1(t)=x(t)-c1(t);

[0089] For r1(t)+ω1E1[ε i (t)] Continue EMD decomposition to obtain the second IMF component:

[0090]

[0091] Among them, E j (·) is the j-th IMF component obtained by EMD decomposition.

[0092] Repeat the following steps to calculate the remaining IMF components.

[0093] r k (t) = r k-1 (t)-c k (t),k=2,3,…,K,;

[0094]

[0095] Where K is the total number of modes.

[0096] When the residual sequence cannot be further decomposed, that is, the number of extreme points does not exceed 2, the algorithm ends. The residual sequence is:

[0097]

[0098] The original time series x(t) can be expressed as follows after CEEMDAN decomposition:

[0099]

[0100] The closing price series of the Shanghai Composite Index (SSE50) from January 2, 2004 to January 9, 2018, totaling 3408 trading days, was selected as the experimental object. CEEMDAN decomposition was performed with an integration order of I = 50 and white noise with a standard deviation of 0.2 added. The closing price series was decomposed into 8 intrinsic mode functions (IMF1 to IMF8) and a residual sequence Res using the Python language, arranged from high frequency to low frequency. The decomposition results are shown in Figure 2. Figure 6 shown.

[0101] IMF1 to IMF8 components have different frequency ranges. IMF1 is the highest-frequency component in the CEEMDAN decomposition and therefore captures the highest-frequency fluctuations in the closing price series of the SSE 50 Index. IMF8 is the last component in the CEEMDAN decomposition and also the lowest-frequency component, reflecting the lowest-frequency fluctuations in the closing price series of the SSE 50 Index. The residual sequence, Res, is the remainder of the original sequence minus the sum of all IMF components. It represents the overall trend of changes in the closing price of the SSE 50 Index. To better understand the variation patterns of the closing price series of the SSE 50 Index after CEEMDAN decomposition, further periodic analysis and research is conducted on each IMF component and the residual sequence. First, the period of each IMF component is defined by obtaining the number of maximum and minimum points of each IMF component:

[0102]

[0103] Wherein, N is the number of original data, N1 is the sum of maximum and minimum points, N=3408.

[0104] Secondly, define the variance share of each IMF component as:

[0105]

[0106] Among them, ξ is the variance of the original sequence, ξ1 is the variance of the IMF component, and A represents the proportion of the volatility of each IMF component to the overall volatility of the closing price of the SSE 50 Index.

[0107] We then calculated the mean and Pearson correlation coefficient for each component to observe the correlation between the IMF component and the original series. Finally, through a statistical description of each IMF component, we gained a deeper understanding of the fluctuation pattern of the closing price of the SSE 50 Index. The statistical results are shown in Table 1.

[0108] Table 1 Correlation statistics of each component of the Shanghai Composite 50 Index after CEEMDAN decomposition

[0109] IMF mean cycle Variance % Pearson correlation coefficient IMF1 0.44 1.48 0.07％ 0.0527 IMF2 0.75 3.11 0.08％ 0.0734 IMF3 -0.25 7.04 0.17％ 0.0473 IMF4 -2.48 16.38 0.51％ 0.0398 IMF5 -0.41 40.57 0.90％ 0.1392 IMF6 -53.0 126.22 20.21％ 0.4809 IMF7 -6.80 284.00 19.57％ 0.6162 IMF8 63.20 1136.00 33.89％ 0.6655 Res 1932.29 3408.00 6.88％ 0.3994

[0110] As can be seen, the periods, variance contributions, and Pearson coefficients of each IMF component increase with decreasing frequency. IMF1–IMF5 have shorter periods, less than 40 days, and a smaller variance contribution, representing high-frequency sequences. This suggests a minimal impact on the overall fluctuations of the SSE 50 Index closing price. IMF6–IMF8 have longer periods, larger variance contributions, and larger Pearson coefficients, representing low-frequency sequences, indicating a strong correlation with the original sequence. The Res component represents the overall trend of the original sequence, and its Pearson coefficient reaches 0.3994, indicating a significant impact on the original sequence. Therefore, the higher the frequency of the IMF component, the smaller its impact on the SSE 50 Index closing price series.

[0111] Based on the data characteristics of the IMF components, the LSTM-ATTE model parameters corresponding to each IMF component are determined and then predicted. The parameter settings are shown in Table 2, where Unit is the number of output neurons in the LSTM, L2_rate is the L2 regularization coefficient, and Ir_rate is the learning rate.

[0112] Table 2 Parameters of the CEEMDAN-LATM-ATTE model

[0113] IMF Unit L2_rate lr_rate IMF1 60 0.1 0.01 IMF2 40 0.0001 0.01 IMF3 40 0 0.01 IMF4 40 0 0.01 IMF5 40 0 0.01 IMF6 40 0.001 0.001 IMF7 40 0.0001 0.01 IMF8 40 0 0.001 Res 40 0 0.01

[0114] The predictions for each subsequence are as follows Figure 7 As shown in Table 3, the errors are large, and it can be seen that the prediction errors of high-frequency subsequences, especially IMF1 and IMF2, are large because they have strong volatility. As the IMF component tends to be stable, the fit between the predicted value and the true value becomes better and better.

[0115] Table 3 Prediction error of subsequences

[0116] IMF MAE RMSE MAPE IMF1 12.95 23.81 585.22 IMF2 4.03 6.15 129.84 IMF3 0.69 1.20 18.39 IMF4 0.35 0.56 2.30 IMF5 0.22 0.30 70.33 IMF6 0.97 1.24 10.79 IMF7 3.23 3.72 5.21 IMF8 1.91 2.73 0.75 Res 0.52 0.65 0.02

[0117] Finally, the final result of the Shanghai Composite 50 Index closing price test set prediction is obtained by adding the prediction results of each IMF component and the residual component Res with equal weights, as shown in the figure below: Figure 8 As shown in the figure, the CEEMDAN-LSTM-ATTE model fits the original data well. To better evaluate the two combined models, the prediction errors of the EMD-LSTM-ATTE and CEEMDAN-LSTM-ATTE models are compared in Table 4. The MAE error of the CEEMDAN-LSTM-ATTE model is 8.45% lower than that of the EMD-LSTM-ATTE model, indicating that the CEEMDAN algorithm decomposes the original time series more accurately than the EMD algorithm, improving overall prediction capabilities.

[0118] Table 4 Comparison of prediction errors between EMD-LSTM-ATTE and CEEMDAN-LSTM-ATTE models

[0119]

[0120]

[0121] In order to more accurately evaluate the performance of the model, the predicted value is, the original value is, assuming that the linear equation is Y = a·X + b, the determination coefficient R 2 It is used to test the degree of association between two variables. If a is close to 1, the deviation between the predicted value and the original value is smaller. The linear regression results are as follows Figure 9 The parameters are shown in Table 5. The slope a of the linear regression is close to 1, and the coefficient of determination R 2 is also close to 1, which shows that the model's predicted values ​​are very close to the original values.

[0122] Table 5 Parameters in linear regression analysis

[0123] parameter CEEMDAN-LSTM-ATTE a 1.089 b -20.36 <![CDATA[R 2 ]]> 0.9942

[0124] Based on the same dataset, ARIMA, BP, RNN, LSTM, and LSTM-ATTECEEMDAN-LSTM models were selected for comparison. The prediction results are shown in Tables 6 and 7. The experimental results show that the CEEMDAN-LATM-ATTE model significantly outperforms the above models, achieving lower prediction errors. MAE was reduced by 77.86%, 71.11%, 63.95%, 49.63%, 48.11%, and 29.30%, respectively, while DS was increased to 76.17%. The combined model significantly outperformed the single model, demonstrating that EMD and CEEMDAN decomposition can improve the predictive performance of neural networks. The CEEMDAN-LATM-ATTE model performed particularly well, demonstrating the superior predictive capabilities of the combined model proposed in this example.

[0125] Table 6 Comparison of prediction indicators of different models

[0126]

[0127] Table 7. Five performance improvement percentages of CEEMDAN-LSTM-ATTE compared with other models

[0128]

[0129] Therefore, the present invention adopts the above-mentioned artificial intelligence-based quantitative financial data classification and evaluation method and system to improve the prediction accuracy.

[0130] 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 the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A quantitative financial data classification system based on artificial intelligence, characterized by: include: A data collection and preprocessing module, used to collect financial data from pre-configured financial data sources and perform preprocessing; A graph construction module is used to standardize financial data, obtain financial standard data, and construct a financial knowledge graph based on the financial standard data; Feature extraction module, used to extract features from the financial knowledge graph, obtain financial feature data, and save it to the database; Model building module, building a financial time prediction model based on CEEMDAN-LSTM-ATTE; The prediction module inputs financial feature data into the financial time prediction model for training to obtain prediction results.

2. The artificial intelligence-based quantitative financial data classification system according to claim 1, characterized in that: The prediction module is further used to extract financial sample data from a preset sample database and extract sample indicator items based on the financial sample data; Calculate the sample scoring results of the sample indicator items according to the preset scoring rules; The financial sample data is used as the input of the preset training model, and the sample risk score result is used as the output of the preset training model; The preset training model is trained according to the input results and output results to obtain a financial time prediction model.

3. The artificial intelligence-based quantitative financial data classification system according to claim 2, characterized in that: The graph construction module is further used to clean the financial data to obtain financial filtered data; Perform data adaptive filling on financial filtered data to obtain financial completed data; Perform knowledge extraction on financial supplementary data to obtain financial standard data.

4. The artificial intelligence-based quantitative financial data classification system according to claim 3, characterized in that: The graph construction module is also used to extract triples from financial standard data according to the preset financial knowledge graph framework, and embed the triples into the preset financial knowledge graph framework to obtain a financial knowledge graph.

5. The artificial intelligence-based quantitative financial data classification system according to claim 4, characterized in that: The prediction module is also used to input financial feature data into the financial time prediction model for scoring and obtaining evaluation results.

6. The artificial intelligence-based quantitative financial data classification system according to claim 1, characterized in that: The steps of the CEEMDAN-LSTM-ATTE financial time prediction model are as follows: The CEEMDAN decomposition of the financial time series yields a series of intrinsic mode functions IMF1, IMF2, ..., IMF n and a residual sequence Res; The subsequences are normalized and then used as input variables of the LSTM-ATTE model with different parameters for prediction, and a series of prediction values ​​F1, F2, ..., F are obtained. n , F n+1 ; The predicted values ​​of each subsequence are then denormalized and superimposed to obtain the final prediction result.

7. Apply the artificial intelligence-based quantitative financial data classification and evaluation method according to claims 1-6, characterized in that: The following steps are involved: S1. Collect financial data and pre-process it according to pre-configured financial data sources; S2. Perform data standardization on the original financial data to obtain financial standard data, and build a financial knowledge graph based on the financial standard data; S3. Extract features from the financial knowledge graph to obtain financial feature data; S4. Build a financial time prediction model based on CEEMDAN-LSTM-ATTE; S5. Input the financial feature data into the financial time prediction model for training to obtain prediction results; S6. Evaluate the prediction results according to the preset scoring levels.

Images