Financial agent system based on big data and construction method
By fusing CNN-LSTM-Attention, XGBoost-GRU, and BERT-Transformer models, the problems of data integration and decision lag in the financial management system are solved, and efficient processing and accurate decision-making of multi-source heterogeneous data are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF TECH
- Filing Date
- 2025-12-08
- Publication Date
- 2026-04-17
AI Technical Summary
Existing financial management systems struggle to effectively integrate unstructured data and lack adaptive multi-model fusion mechanisms, resulting in insufficient data processing capabilities, delayed decision-making, and poor interpretability.
The system employs a CNN-LSTM-Attention model for financial risk warning, an XGBoost-GRU model for financial forecasting, and a BERT-Transformer model for intelligent accounting, achieving deep integration of multi-source heterogeneous data and various neural network models to adapt to the characteristics of financial scenarios.
It improved the accuracy of financial risk early warning, reduced forecasting errors, enhanced the accuracy of special business accounting, and reduced human interpretation bias.
Smart Images
Figure CN121883186A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and financial information technology, specifically to a financial intelligent agent system based on big data and its construction method. Background Technology
[0002] As enterprises become increasingly information-based, financial management systems face challenges such as insufficient data processing capabilities, delayed decision-making, passive risk identification, and a lack of personalization. Traditional systems primarily rely on structured data and preset rules, making it difficult to effectively integrate unstructured data (such as contract texts and market sentiment). Furthermore, the analysis process heavily depends on manual intervention, leading to inefficiency and decision-making biases.
[0003] While some existing solutions attempt to incorporate big data or AI technologies (such as OCR for invoice recognition and simple algorithms for revenue prediction), they mostly remain at the level of single-function optimization, failing to form a closed loop of "data collection - intelligent analysis - business implementation - decision feedback." More critically, existing solutions fail to address the core issue of deep integration between AI algorithms and financial calculation logic, particularly how to collaboratively utilize different types of data (numerical, time-series, text) and models with different characteristics (such as decision trees, recurrent neural networks, and Transformers) to make more accurate and interpretable financial decisions. These solutions typically employ a simple concatenation or parallelization of models followed by rigid concatenation, lacking an adaptive fusion mechanism capable of dynamically evaluating the contribution of each model, resulting in system performance bottlenecks and poor interpretability.
[0004] Therefore, there is an urgent need for an intelligent system that can deeply integrate multi-source heterogeneous data and various neural network models, and can adapt to the characteristics of financial scenarios. Thus, a new solution to the above problems needs to be proposed. Summary of the Invention
[0005] The purpose of this invention is to provide a financial intelligent agent system and its construction method based on big data, so as to solve the technical problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for building a financial intelligent agent system based on big data, comprising at least the following steps:
[0007] S1: A financial risk early warning engine is built by using a CNN-LSTM-Attention fusion model, which simultaneously processes the local features, temporal correlations and key factor weights of financial data. The local features are high-impact risk factors, and the temporal correlations are risk transmission trends, thus solving the problem that a single model cannot cover multi-dimensional risk features.
[0008] S2: A financial forecasting engine is built by using the XGBoost-GRU fusion model. Combining XGBoost's ability to process nonlinear financial features with GRU's advantages in modeling short time series data, the problem of poor nonlinear fitting and time lag in traditional forecasting models is solved.
[0009] S3: Build an intelligent accounting engine through the BERT-Transformer fusion model. Use BERT to understand unstructured accounting text and combine it with the rule mapping capability of Transformer to solve the problems of difficult special business processing and poor rule adaptation in traditional accounting models.
[0010] Furthermore, S1 includes at least the following steps:
[0011] S1.1: Local feature extraction using CNN;
[0012] Input the preprocessed financial feature matrix ;
[0013] in For the sample size, For feature dimensions;
[0014] Through convolution kernel Extracting local high-impact features, among which The size of the convolution window is set to 3-6 based on the financial data period, corresponding to a risk period of 3-6 months.
[0015]
[0016] in The convolution outputs a feature map. As a bias term, the ReLU function addresses nonlinear relationships in financial features;
[0017] Then, the features are compressed using a max pooling layer to retain key risk factors:
[0018]
[0019] in, To determine the pooled window size, The feature vector after pooling;
[0020] S1.2: LSTM temporal correlation modeling;
[0021] The pooled feature vector Inputting data into an LSTM layer allows for the learning of time-series transmission relationships of risk factors. The LSTM cell state update formula is as follows:
[0022]
[0023] in: To prevent the retention of historical risk information, a forgetting gate is established. This is the input gate, which controls the input weights of the current feature; It is an output gate that controls the output of cell state; is the hidden layer dimension, and is the temporal feature output of LSTM; For the accumulation of elements; This is the weight matrix; For bias terms; For the first The feature vector after pooling at time step has a dimension of ; This represents the hidden state of the LSTM at the previous time step.
[0024] S1.3: Attention weight allocation;
[0025] A multi-head attention mechanism is introduced to quantify the influence weights of different risk factors, as shown in the following formula:
[0026] Calculate the similarity between the query, key, and value, where the query is the output of an LSTM. The Key is a CNN feature. The Value is the original risk characteristic. ;
[0027]
[0028] in: The dimension of the key should be set to avoid gradient vanishing due to excessively high dimensionality.
[0029] Attention weights are obtained by softmax normalization:
[0030]
[0031] in, For the first The weights of each risk factor, and ;
[0032] The weighted summation yields the final risk characteristics:
[0033]
[0034] in, This represents the original risk feature vector;
[0035] S1.4: Risk level output;
[0036] Will Input: Fully connected layer; Output: Risk level.
[0037]
[0038] in, To output the weight matrix; To output the bias term, the level with the highest probability is taken as the final result, and then... The causes of high-weight risks are identified.
[0039] Furthermore, S2 includes at least the following steps:
[0040] S2.1: XGBoost nonlinear feature extraction;
[0041] S2.2: GRU time series prediction;
[0042] S2.3: Fusion prediction output.
[0043] Furthermore, S2.1 includes at least the following steps:
[0044] Input prediction feature set Build via XGBoost Decision trees, outputting non-linear feature mapping values :
[0045]
[0046] in, For the first The leaf node region of a tree; This represents the weight of the node. For indicator functions;
[0047] To adjust the feature contribution, dynamic weighting coefficients are introduced to suit different enterprise business attributes.
[0048]
[0049] in, For the first XGBoost output of each feature; For the first The business weights of each feature, and .
[0050] Furthermore, S2.2 includes at least the following steps:
[0051] Will By time series Reorganized into timing input ;
[0052] Input GRU layer learning time sequence correlation:
[0053] GRU update gate and reset gate formulas:
[0054]
[0055] in; To update the gate, retrieve the value. ; To reset the door, retrieve the value. ; This is the weight matrix; This is a bias term.
[0056] Furthermore, S2.3 includes at least the following steps:
[0057] Combining the nonlinear characteristics of XGBoost with the temporal output of GRU, the final predicted value is obtained through a linear fusion formula:
[0058]
[0059] in, For fusion weights; This is the error correction term.
[0060] Furthermore, S3 includes at least the following steps:
[0061] S3.1: BERT Semantic Feature Extraction
[0062] Input accounting text For the segmented vocabulary, word vectors are generated using a pre-trained BERT model. Then, the text semantic vector is obtained through pooling. :
[0063]
[0064] in, This is the global semantic vector of the text. It is the first The word vectors of each word have a dimension of . ;
[0065] S3.2: Transformer Rule Mapping
[0066] Construct an "accounting standards rule base" to transform accounting rules into structured rule vectors. For rule feature dimensions;
[0067] Semantic vectors are constructed using the self-attention mechanism of the Transformer encoder. With regular vectors The mapping relationship is shown in the following formula:
[0068] Calculate the attention weights between the semantic vector and each rule vector:
[0069]
[0070] The weighted summation yields the rule-fitting vector:
[0071]
[0072] in, This represents the combination of accounting standards and rules that best suits the current accounting scenario;
[0073] S3.3: Integrated Accounting Output
[0074] semantic vectors Adaptation vector with rules After concatenation, the fully connected layer generates a specific accounting scheme (such as "goodwill impairment amount and amortization method"), as shown in the following formula:
[0075]
[0076] in, This represents a vector concatenation operation; To calculate the weight matrix; For bias terms; For structured accounting results (such as "Goodwill impairment amount: 5 million yuan, amortization period: 10 years"), the matching rule basis is also output.
[0077] A financial intelligent agent system based on big data is constructed using a method for building a financial intelligent agent system based on big data.
[0078] It includes at least a financial risk early warning engine, a financial forecasting engine, and an intelligent accounting engine.
[0079] Compared with the prior art, the beneficial effects of the present invention are:
[0080] The model accuracy of this invention is significantly improved, and the accuracy of financial risk early warning is increased from that of traditional single models. It utilizes multi-dimensional feature processing via CNN-LSTM-Attention to cover static, dynamic, and textual risk factors.
[0081] The error rate of financial forecasting has been reduced by integrating nonlinear fitting and short-time series modeling capabilities through XGBoost-GRU, making it suitable for complex scenarios such as raw material price fluctuations and changes in customer traffic. The accuracy of special business accounting has been improved by solving the problem of "textual clause understanding + rule matching" through BERT-Transformer, avoiding human interpretation bias. Attached Figure Description
[0082] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0083] Figure 1 This is a schematic diagram of the system of the present invention;
[0084] Figure 2 This is a schematic diagram of the financial risk early warning engine of the present invention;
[0085] Figure 3 This is a schematic diagram of the financial forecasting engine of the present invention;
[0086] Figure 4 This is a schematic diagram of the intelligent accounting engine of the present invention. Detailed Implementation
[0087] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0088] This invention specifically relates to a financial intelligent agent system that integrates a multi-neural network collaborative model, big data processing, and financial calculation logic. It can be applied to core scenarios such as enterprise financial accounting, risk warning, budget forecasting, and tax planning, solving the problem of insufficient adaptability of a single AI model in complex financial scenarios.
[0089] Example 1:
[0090] Please see Figures 1-4 A method for building a big data-based financial intelligent agent system includes at least the following steps:
[0091] S1: A financial risk early warning engine is built by using a CNN-LSTM-Attention fusion model. It simultaneously processes the local features, temporal correlations, and key factor weights of financial data. The local features are high-impact risk factors, and the temporal correlations are risk transmission trends, thus solving the problem that a single model cannot cover multi-dimensional risk features.
[0092] S2: A financial forecasting engine is built by using the XGBoost-GRU fusion model. By combining XGBoost's ability to handle nonlinear financial characteristics (such as raw material price fluctuations) with GRU's modeling advantages for short time series data (such as monthly cash flow), the problem of poor nonlinear fitting and time lag in traditional forecasting models is solved.
[0093] S3: Build an intelligent accounting engine through the BERT-Transformer fusion model. Use BERT to understand unstructured accounting text (such as M&A contract terms) and combine it with Transformer's rule mapping capabilities to solve the problems of difficult special business processing and poor rule adaptation in traditional accounting models.
[0094] S1 includes at least the following steps:
[0095] S1.1: Local feature extraction using CNN;
[0096] Input the preprocessed financial feature matrix ;
[0097] in For the sample size, Features include accounts receivable turnover, customer credit scores, policy keyword vectors, etc.
[0098] Through convolution kernel Extracting local high-impact features, among which The size of the convolution window is set to 3-6 based on the financial data period, corresponding to a risk period of 3-6 months.
[0099]
[0100] in The convolution outputs a feature map. As a bias term, the ReLU function addresses non-linear relationships in financial characteristics (such as the non-linear risk of a sudden drop in credit score).
[0101] Then, the features are compressed using a max pooling layer to retain key risk factors:
[0102]
[0103] in, To determine the pooled window size, The feature vector after pooling;
[0104] S1.2: LSTM temporal correlation modeling;
[0105] The pooled feature vector Inputting into an LSTM layer, the system learns the time-series transmission relationships of risk factors (e.g., "credit score decline → loan default after 3 months"). The LSTM cell state update formula is as follows:
[0106]
[0107] in: To prevent data loss, control the retention of historical risk information (such as "whether cooperation records from 1 year ago affect current risks"). This is the input gate, which controls the input weights of the current feature; It is an output gate that controls the output of cell state; is the hidden layer dimension, and is the temporal feature output of LSTM; For the accumulation of elements; This is the weight matrix; For bias terms; For the first The feature vector after pooling at time step has a dimension of ; This represents the hidden state of the LSTM at the previous time step.
[0108] S1.3: Attention weight allocation;
[0109] A multi-head attention mechanism is introduced to quantify the impact weights of different risk factors (e.g., the weight of "related enterprise default" is higher than that of "industry average default rate"), as shown in the following formula:
[0110] Calculate the similarity between Query, Key, and Value, where Query is the output of LSTM. Key is the CNN feature. Value represents the original risk characteristic. ;
[0111]
[0112] in: The dimension of the key should be set to avoid gradient vanishing due to excessively high dimensionality.
[0113] Attention weights are obtained by softmax normalization:
[0114]
[0115] in, For the first The weights of each risk factor, and ;
[0116] The weighted summation yields the final risk characteristics:
[0117]
[0118] in, This represents the original risk feature vector;
[0119] S1.4: Risk level output;
[0120] Will Input to a fully connected layer, output risk level (low / medium / high), corresponding probability vector
[0121]
[0122] in, To output the weight matrix; To output the bias term, the level with the highest probability is taken as the final result, and then... High-weight risk factors (e.g.) (This is the main risk factor).
[0123] S2 includes at least the following steps:
[0124] S2.1: XGBoost nonlinear feature extraction;
[0125] S2.2: GRU time series prediction;
[0126] S2.3: Fusion prediction output.
[0127] S2.1 includes at least the following steps:
[0128] Input prediction feature set (Including historical revenue, raw material prices, customer traffic, etc.), built using XGBoost Decision trees, outputting non-linear feature mapping values :
[0129]
[0130] in, For the first The leaf node region of a tree; This represents the weight of the node. This is an indicator function (1 if the feature belongs to this node, 0 otherwise);
[0131] For different business attributes of enterprises, dynamic weighting coefficients are introduced (manufacturing) service industry Adjusting feature contribution:
[0132]
[0133] in, For the first XGBoost output of each feature; For the first The business weights of each feature, and .
[0134] S2.2 includes at least the following steps:
[0135] Will By time series (e.g., "data from the past 12 months") are reorganized into time-series inputs. ;
[0136] Input GRU layer learning time sequence correlation:
[0137] The formulas for the GRU update gate (which controls the retention of historical information) and reset gate (which controls the input of current information) are as follows:
[0138]
[0139] in; To update the gate, retrieve the value. ; To reset the door, retrieve the value. ; This is the weight matrix; This is a bias term.
[0140] S2.3 includes at least the following steps:
[0141] Combining the nonlinear characteristics of XGBoost with the temporal output of GRU, the final predicted value is obtained through a linear fusion formula:
[0142]
[0143] in, To incorporate weights (optimized through cross-validation, such as when predicting quarterly revenue) It emphasizes nonlinear characteristics; when predicting monthly cash flow. (Focusing on temporal correlation) This is the error correction term (dynamically adjusted based on historical prediction deviations).
[0144] S3 includes at least the following steps:
[0145] S3.1: BERT Semantic Feature Extraction
[0146] Enter the accounting text (such as the goodwill impairment clause). For the segmented vocabulary, word vectors are generated using a pre-trained BERT model. ( (This is the base dimension of BERT), and then the text semantic vector is obtained through pooling. :
[0147]
[0148] in, As a global semantic vector for text, it can capture the semantic relationships of key accounting information such as "goodwill impairment ratio and amortization period"; It is the first The word vectors of each word have a dimension of . ;
[0149] S3.2: Transformer Rule Mapping
[0150] Construct an "accounting standards rule base" to transform accounting rules (such as "goodwill impairment testing should be combined with the present value of future cash flows") into structured rule vectors. For rule feature dimensions;
[0151] Semantic vectors are constructed using the self-attention mechanism of the Transformer encoder. With regular vectors The mapping relationship is shown in the following formula:
[0152] Calculate the attention weights between the semantic vector and each rule vector:
[0153]
[0154] The weighted summation yields the rule-fitting vector:
[0155]
[0156] in, This represents the combination of accounting standards and rules that best suits the current accounting scenario;
[0157] S3.3: Integrated Accounting Output
[0158] semantic vectors Adaptation vector with rules After concatenation, the fully connected layer generates a specific accounting scheme (such as "goodwill impairment amount and amortization method"), as shown in the following formula:
[0159]
[0160] in, This represents a vector concatenation operation; To calculate the weight matrix; For bias terms; For structured accounting results (such as "Goodwill impairment amount: 5 million yuan, amortization period: 10 years"), the matching rule basis is also output.
[0161] Example 2:
[0162] A big data-based financial intelligent agent system is constructed using the construction method of a big data-based financial intelligent agent system proposed in Example 1.
[0163] It includes at least a financial risk early warning engine, a financial forecasting engine, and an intelligent accounting engine.
[0164] Based on the specific embodiments described above, the following practical application examples are proposed:
[0165] Example: Application of the Financial Intelligent Agent System at ABC Manufacturing Co., Ltd.
[0166] Background Description
[0167] Company Overview: ABC Manufacturing Co., Ltd. is a mid-sized electronic component manufacturer whose main customers include home appliance companies and automobile manufacturers. The company's financial data includes monthly financial statements, supply chain information, customer credit records, and unstructured contract texts (such as merger and acquisition agreements).
[0168] Financial challenges:
[0169] Risk warning: Frequent customer payment delays, but traditional systems cannot identify the trend of credit risk transmission in advance.
[0170] Demand forecasting: The price fluctuations of raw materials (such as copper and plastics) lead to inaccurate revenue forecasts.
[0171] Accounting complexity: A recent acquisition of a subsidiary involved goodwill impairment testing, requiring the handling of complex contract terms and accounting standards.
[0172] Application of system building methods
[0173] The system adopts the construction method described in the document, with three engines handling different tasks respectively. The specific steps and outputs are as follows.
[0174] Application of financial risk early warning engine (CNN-LSTM-Attention fusion model);
[0175] Input data: Preprocessed financial feature matrix (12-month sample, 50-dimensional features, including accounts receivable turnover, customer credit score, industry policy keyword vectors, etc.)
[0176] Processing steps:
[0177] S1.1 CNN Local Feature Extraction:
[0178] kernel size (Corresponding to a 3-month risk period), extract local high-impact features (such as a customer's credit score suddenly dropping).
[0179] formula: Output feature map .
[0180] Max pooling (window size) Compress the features to obtain the pooled vector. .
[0181] S1.2 LSTM Temporal Correlation Modeling:
[0182] enter At the LSTM layer, learn about risk transmission (such as how a decline in credit score can lead to loan default within 3 months).
[0183] LSTM Cell State Update: Using the Forgetting Gate Control historical information, input gate Control the current feature weights. Output time-series features. (Hidden layer dimension 32).
[0184] S1.3 Attention weight allocation:
[0185] Calculate LSTM output CNN features The similarity is normalized to obtain the attention weights. .
[0186] For example, the weights show the weights of the "related company default" factor. "Industry average default rate" weight .
[0187] Weighted summation: .
[0188] S1.4 Risk Level Output:
[0189] Fully connected layer output probability: .
[0190] result: (High risk) The system marks the main risk cause as "default by related companies" (weight 0.6).
[0191] Real-world application: The system issued a warning to ABC Company two months in advance that its key client (an automaker) posed a payment risk and recommended tightening its credit policy. Subsequently, the client did indeed default on payments, thus mitigating the risk.
[0192] Financial forecasting engine application (XGBoost-GRU fusion model)
[0193] Input data: Predicted feature set (Historical revenue, raw material prices, customer traffic), time series (Data from the past 12 months).
[0194] Processing steps:
[0195] S2.1 XGBoost Nonlinear Feature Extraction:
[0196] Construct T=100 decision trees and output nonlinear mappings. .
[0197] Introducing dynamic business weights: Manufacturing characteristic weights , Adjusted output .
[0198] S2.2 GRU timing prediction:
[0199] Will Reassembling the timing input .
[0200] GRU layer learns short-sequence patterns: update gate And control historical information retention, reset door Control the current input. Output timing characteristics. .
[0201] S2.3 Fusion prediction output:
[0202] Linear fusion: .
[0203] Fusion weights (Focusing on non-linear characteristics, due to quarterly forecasting), error term Adjusted based on historical bias.
[0204] Practical Application: The system predicted next quarter's revenue to be 52 million yuan (actual value was 51 million yuan), with an error rate of only 2%, far lower than the traditional model (15%). The system identifies raw material price fluctuations as the main influencing factor, helping the company to make purchases in advance.
[0205] Application of intelligent accounting engine (BERT-Transformer fusion model)
[0206] Input data: Unstructured accounting text (Merger and acquisition contract terms, such as "goodwill impairment test must be based on cash flow over the next 5 years").
[0207] Processing steps:
[0208] S3.1 BERT semantic feature extraction:
[0209] BERT model generates word vectors Mean pooling yields the global semantic vector. .
[0210] S3.2 Transformer rule mapping:
[0211] The accounting standards rule base contains rule vectors (e.g., IFRS goodwill impairment rules).
[0212] Calculate attention weights: The weighted summation yields the rule adaptation vector. .
[0213] S3.3 Fusion Accounting Output:
[0214] Concatenate vectors: .
[0215] Practical Application: Outputs a structured accounting scheme: "Goodwill Impairment Amount: 5 million RMB, Amortization Period: 10 years", and matches the rule based on "IFRS 36 - Goodwill Impairment Test". The system automatically processes contract terms, reducing manual accounting errors and time (from 5 days to 1 hour).
[0216] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for constructing a financial intelligent agent system based on big data, characterized in that: At least the following steps are included: S1: A financial risk early warning engine is built using a CNN-LSTM-Attention fusion model, which simultaneously processes the local features, temporal correlations, and key factor weights of financial data. The local features are high-impact risk factors, and the temporal correlations are risk transmission trends. S2: A financial forecasting engine is built using the XGBoost-GRU fusion model, combining XGBoost's ability to process nonlinear financial features with GRU's advantages in modeling short-time-series data. S3: Build an intelligent accounting engine through the BERT-Transformer fusion model, use BERT to understand unstructured accounting text, and combine it with the rule mapping capabilities of Transformer.
2. The method for building a big data-based financial intelligent agent system according to claim 1, characterized in that: S1 includes at least the following steps: S1.1: Local feature extraction using CNN; Input the preprocessed financial feature matrix ; in For the sample size, For feature dimensions; Through convolution kernel Extracting local high-impact features, among which The size of the convolution window is set to 3-6 based on the financial data period, corresponding to a risk period of 3-6 months. in The convolution outputs a feature map. As a bias term, the ReLU function addresses nonlinear relationships in financial features; Then, the features are compressed using a max pooling layer to retain key risk factors: in, To determine the pooled window size, The feature vector after pooling; S1.2: LSTM temporal correlation modeling; The pooled feature vector Inputting data into an LSTM layer allows for the learning of time-series transmission relationships of risk factors. The LSTM cell state update formula is as follows: in: To prevent the retention of historical risk information, a forgetting gate is established. This is the input gate, which controls the input weights of the current feature; It is an output gate that controls the output of cell state; is the hidden layer dimension, and is the temporal feature output of LSTM; For the accumulation of elements; This is the weight matrix; For bias terms; For the first The feature vector after pooling at time step has a dimension of ; This represents the hidden state of the LSTM at the previous time step. S1.3: Attention weight allocation; A multi-head attention mechanism is introduced to quantify the influence weights of different risk factors, as shown in the following formula: Calculate the similarity between the query, key, and value, where the query is the output of an LSTM. The Key is a CNN feature. The Value is the original risk characteristic. ; in: The dimension of the key should be set to avoid gradient vanishing due to excessively high dimensionality. Attention weights are obtained by softmax normalization: in, For the first The weights of each risk factor, and ; The weighted summation yields the final risk characteristics: in, This represents the original risk feature vector; S1.4: Risk level output; Will Input: Fully connected layer; Output: Risk level. in, To output the weight matrix; To output the bias term, the level with the highest probability is taken as the final result, and then... The causes of high-weight risks are identified.
3. The method for building a big data-based financial intelligent agent system according to claim 1, characterized in that: S2 includes at least the following steps: S2.1: XGBoost nonlinear feature extraction; S2.2: GRU time series prediction; S2.3: Fusion prediction output.
4. The method for building a big data-based financial intelligent agent system according to claim 3, characterized in that: S2.1 includes at least the following steps: Input prediction feature set Build via XGBoost Decision trees, outputting non-linear feature mapping values : in, For the first The leaf node region of a tree; This represents the weight of the node. For indicator functions; To adjust the feature contribution, dynamic weighting coefficients are introduced to suit different enterprise business attributes. in, For the first XGBoost output of each feature; For the first The business weights of each feature, and .
5. The method for building a big data-based financial intelligent agent system according to claim 4, characterized in that: S2.2 includes at least the following steps: Will By time series Reorganized into timing input ; Input GRU layer learning time sequence correlation: GRU update gate and reset gate formulas: in; To update the gate, retrieve the value. ; To reset the door, retrieve the value. ; This is the weight matrix; This is a bias term.
6. The method for building a big data-based financial intelligent agent system according to claim 5, characterized in that: S2.3 includes at least the following steps: Combining the nonlinear characteristics of XGBoost with the temporal output of GRU, the final predicted value is obtained through a linear fusion formula: in, For fusion weights; This is the error correction term.
7. The method for building a big data-based financial intelligent agent system according to claim 1, characterized in that: The S3 includes at least the following steps: S3.1: BERT Semantic Feature Extraction Input accounting text For the segmented vocabulary, word vectors are generated using a pre-trained BERT model. Then, the text semantic vector is obtained through pooling. : in, This is the global semantic vector of the text. It is the first The word vectors of each word have a dimension of . ; S3.2: Transformer Rule Mapping Construct an "accounting standards rule base" to transform accounting rules into structured rule vectors. For rule feature dimensions; Semantic vectors are constructed using the self-attention mechanism of the Transformer encoder. With regular vectors The mapping relationship is shown in the following formula: Calculate the attention weights between the semantic vector and each rule vector: The weighted summation yields the rule-fitting vector: in, This represents the combination of accounting standards and rules that best suits the current accounting scenario; S3.3: Integrated Accounting Output semantic vectors Adaptation vector with rules The process involves concatenating the data and inputting the fully connected layer to generate a specific computational scheme, as shown in the following formula: in, This represents a vector concatenation operation; To calculate the weight matrix; For bias terms; The results are structured and the matching rules are also output.
8. A financial intelligent agent system based on big data, characterized in that: It is constructed using the method for building a big data-based financial intelligent agent system as described in any one of claims 1-7; The aforementioned big data-based financial intelligent agent system includes at least a financial risk early warning engine, a financial forecasting engine, and an intelligent accounting engine.