Enterprise inter-bank fund prediction method based on multi-dimensional feature fusion and computer equipment

Abstract

The invention relates to the technical field of financial science and technology and artificial intelligence, and discloses an enterprise inter-bank fund prediction method based on multi-dimensional feature fusion and computer equipment, and the method comprises the steps: firstly integrating enterprise data, and screening high-value target customer groups based on transaction characteristics; extracting a multi-dimensional feature set covering a time sequence trend, a transaction network topology and an enterprise background; then, a Panel mode joint prediction model based on a TimeMixers architecture is constructed, and cross-enterprise feature transfer learning is realized by sharing bottom layer LSTM-CNN joint encoder parameters; and finally, training the model by using a staged strategy, fusing the output of the classification and regression model through a dynamic weighting algorithm, and generating a final predicted value. According to the method, the problems of insufficient generalization ability and single feature caused by the fact that traditional single-enterprise modeling is limited by data sparsity are effectively solved, and the prediction precision and model maintenance efficiency of inter-bank fund flow of enterprises are improved by mining industry general modes and multi-source heterogeneous features.

Description

Technical Field

[0001] This invention relates to the fields of financial technology and artificial intelligence, specifically to a method and computer equipment for predicting cross-bank funding for enterprises based on multi-dimensional feature fusion. Background Technology

[0002] Currently, with the in-depth development of fintech and the advancement of corporate-level large-scale capital circulation strategies, commercial banks are increasingly demanding more refined management of corporate clients' capital flows. This is especially true for large and medium-sized manufacturing enterprises and other real economy entities, whose capital turnover chains are long, frequent, and involve complex interbank interactions. Accurately predicting the inflow and outflow of interbank funds within a given timeframe has become a core element for banks to enhance their internal capital absorption capacity, strengthen liability quality management, and formulate precise strategies for stabilizing and increasing deposits.

[0003] For the aforementioned application scenarios of liquidity forecasting, existing technologies typically employ statistical time series analysis or basic machine learning algorithms. In practice, the system often treats a single enterprise as an independent research object, utilizing algorithms such as Autoregressive Moving Average (ARIMA) or Long Short-Term Memory (LSTM) networks, primarily fitting and extrapolating based on the linear trend of the enterprise's historical transaction amounts; or employing ensemble learning methods such as XGBoost, combining the enterprise's static financial indicators with historical cash flow statistics to construct a univariate model to achieve regression analysis of future cash status or simple classification of price fluctuations.

[0004] However, the aforementioned existing technologies have limitations when dealing with a massive number of enterprise customers with complex transaction behaviors. On the one hand, the approach of modeling each customer independently leads to a linear increase in the number of models with the user base, resulting in redundancy in computing and storage resources. Furthermore, the inability to achieve cross-entity parameter sharing and feature transfer makes it difficult to capture industry-wide common cash flow patterns (such as supply chain payment cycles) when dealing with single enterprises with sparse transaction data, resulting in weak model generalization ability. On the other hand, existing methods are often limited to single-dimensional time-series feature mining, neglecting key behavioral patterns such as the enterprise's topology and node centrality in the cash interaction network, making it difficult to effectively identify potential risk characteristics that exist at the edge of the transaction network despite having cash flow. In addition, there is a lack of a dynamic fusion mechanism for qualitative cash levels and quantitative specific amounts. The single output mode cannot ensure both the accuracy of risk warning and the precision of numerical prediction, failing to meet the actual needs of refined cash management.

[0005] Therefore, this invention provides a method and computer device for predicting cross-industry funds for enterprises based on multi-dimensional feature fusion, in order to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and computer equipment for predicting cross-industry funds for enterprises based on multi-dimensional feature fusion. This solves the problems of weak model generalization ability, lack of transaction network structure information, and low prediction accuracy in data-sparse scenarios caused by independent modeling of a single entity and single feature dimension in traditional enterprise fund prediction.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for predicting cross-industry funding for enterprises based on multi-dimensional feature fusion, comprising the following steps: S1. Data collection and preprocessing: Integrate enterprise customer account transaction data and enterprise basic information data to establish a unified data view, explore, clean and standardize the format of the raw data, and generate the initial modeling dataset. S2. Customer group segmentation and screening: Receive the initial modeling dataset, divide customers into different groups according to the enterprise's transaction characteristic indicators, screen out high stable value customer groups and potential value customer groups as target modeling objects, and eliminate no-value customers and low-value customer groups. S3. Construct a multi-dimensional feature engineering: For the target modeling object, extract a multi-dimensional feature set covering time-series features, transaction behavior features, and enterprise background features. The multi-dimensional feature set covers three-dimensional information of time-series patterns, transaction modes, and enterprise attributes. S4. Construct a Panel-based joint prediction model: Construct a three-dimensional hybrid neural network based on the TimeMixers architecture. The three-dimensional hybrid neural network integrates the time-series features, the transaction behavior features, and the enterprise background features. It uses the Panel model to jointly model multiple customer data. In the Panel model, the sample data of all the target modeling objects share the parameters of the underlying neural network joint encoder to achieve feature transfer learning. S5. Model Training and Dynamic Fusion Prediction: The Panel model joint prediction model is optimized using a phased training strategy. During prediction execution, the capital level classification model and the capital amount regression model are run in parallel. The weighted fusion algorithm is used to dynamically calculate and adjust the weights of the output results of the two models based on the confidence level of the capital level classification model and the validation set error index of the capital amount regression model. The predicted value of cross-bank capital changes for the enterprise in the future time period is obtained through weighted calculation.

[0008] By adopting the above technical solution, this invention effectively addresses the limitations of single-dimensional data in fund prediction. By integrating time series, transaction behavior, and corporate background characteristics, it comprehensively depicts the patterns of corporate fund flows. In particular, the use of Panel-based joint modeling allows the model to share underlying feature encoder parameters across multiple corporate entities, effectively solving the training difficulties caused by the sparsity of sample data from a single enterprise. Transfer learning is then used to capture common industry features. Simultaneously, the combination of a three-dimensional hybrid neural network with a TimeMixers architecture enables parallel processing of heterogeneous feature data. In the prediction phase, a dynamic weighted fusion algorithm adaptively combines the qualitative judgment of the classification model with the quantitative calculation of the regression model. This ensures accurate prediction of transaction probability while improving the accuracy of specific amount predictions and enhancing the robustness of the prediction results in real-world business scenarios.

[0009] Preferably, in step S1, the collected data fields include customer number, transaction date, transaction time, borrower / lender identifier, transaction amount, fund usage field, basic enterprise characteristics, enterprise managed asset size, and information on financial products held by the enterprise; the cleaning of the original data includes: filling non-critical fields with missing values ​​using default values ​​or average values, removing records with missing values ​​in critical identifier fields, and filtering logically erroneous data with negative transaction amounts.

[0010] By adopting the above technical solutions, the integrity and logical correctness of the input model data are ensured, the interference of noisy data on feature extraction is reduced, and a data foundation is laid for subsequent high-precision modeling.

[0011] Preferably, in step S2, the transaction characteristic indicators include transaction stickiness indicators, transaction activity indicators, and transaction stability indicators; the transaction stickiness indicator is determined based on the number of months with transaction records within the statistical period; the transaction activity indicator is determined based on the average daily transaction frequency of customers; and the transaction stability indicator is determined based on the statistical variance of the historical transaction amount of customers.

[0012] By adopting the above technical solutions, a quantitative customer value assessment system has been established, which can accurately depict customer profiles from three dimensions: transaction frequency, continuity, and volatility, providing an objective basis for tiered screening.

[0013] Preferably, step S2 further includes: defining customers whose number of transaction months is less than a preset first threshold and whose transaction status is unstable as the low-value customers; defining customers whose number of transaction months is between the first threshold and a preset second threshold and whose transaction status is unstable as the low-value customer group; defining customers whose number of transaction months is between a preset third threshold and a preset fourth threshold and whose transaction status is stable as the high-stability-value customer group; defining customers whose number of transaction months is between the second threshold and the fourth threshold and whose transaction records have been continuously present in the most recent preset month as the potential-value customer group; and retaining the data of the high-stability-value customer group and the potential-value customer group as the input data of the target modeling object.

[0014] By adopting the above technical solutions, we can accurately remove noisy samples that are sparse, interrupted, or extremely unstable, and focus on the target group with modeling value. This avoids the negative impact of low-quality samples on the convergence of model parameters and improves the efficiency of model training and the feasibility of prediction.

[0015] Preferably, in step S3, the extraction of the time series features includes: using the Long Short-Term Memory Network sliding window technique to process historical fund flow sequences, setting the sliding window to move gradually on the time axis to capture continuous time slice data, analyzing the fund change sequence within the window through the gating mechanism of the Long Short-Term Memory Network unit, and calculating and extracting dynamic indicators of fund inflow slope and fund outflow slope.

[0016] By adopting the above technical solution, the long-term dependence of capital flows can be effectively captured by utilizing the memory capability of LSTM for long-sequence data. At the same time, the rate and trend of capital changes are quantified by the slope index, which enhances the model's sensitivity to short-term fluctuation trends.

[0017] Preferably, in step S3, the extraction of transaction behavior features includes: constructing a transaction time distribution matrix, which is a two-dimensional matrix defined by time periods and transaction frequencies; constructing a capital flow network graph, where nodes represent transaction accounts and edges represent capital flows; using a convolutional neural network to extract features from the transaction time distribution matrix and the capital flow network graph, capturing local pattern features and node centrality topological structure features through convolution operations using convolution kernels; the extraction of enterprise background features includes: obtaining discrete classification data of the enterprise's position in the industrial chain and industry attributes, inputting the discrete classification data into the embedding layer, and converting the category features into low-dimensional dense real number vectors.

[0018] By adopting the above technical solutions, the transaction behavior of enterprises is transformed into image-based or graph-structured data. The powerful spatial feature extraction capability of convolutional neural networks is used to mine the implicit transaction patterns and network topology. At the same time, the discrete static attributes are mapped to the continuous vector space through the embedding layer, realizing the effective alignment and fusion of static and dynamic features.

[0019] Preferably, in step S4, the TimeMixers architecture includes a joint encoder of a long short-term memory network and a convolutional neural network; the joint encoder contains parallel processing branches, which are respectively used to process the time series data of fund flows generated by the sliding window, the transaction time distribution matrix and the fund flow network graph data, and the embedding vector of enterprise background features; the parameter sharing mechanism in the Panel mode is configured as follows: the weight parameters and bias parameters in the joint encoder are shared among all input customer samples, and parameters adapted to the individual differences of different customers are retained only in the fully connected layer.

[0020] By adopting the above technical solution, a multi-branch parallel feature encoding structure was constructed, which can simultaneously process heterogeneous data from multiple sources, including temporal, spatial, and static attributes. In particular, in Panel mode, cross-enterprise knowledge transfer was achieved by sharing the parameters of the underlying LSTM-CNN joint encoder, reducing the number of model parameters, preventing overfitting, and retaining fully connected layers to handle individual differences, thus balancing the model's generalization ability and its adaptability to specific tasks.

[0021] Preferably, in step S5, the phased training strategy includes: a pre-training phase: inputting the full enterprise customer dataset into the model and updating all shared layer parameters in the Panel pattern joint prediction model; and a fine-tuning phase: freezing the shared layer parameters, updating only the fully connected layer parameters of the Panel pattern joint prediction model, and setting the learning rate of the fine-tuning phase to be lower than the learning rate of the pre-training phase.

[0022] By adopting the above technical solution, we first learn the general characteristics of capital flow through the full amount of data, and then fine-tune it to adapt to the parameters of specific layers. This not only makes full use of the massive data to mine common patterns, but also avoids destroying the learned general characteristics during the fine-tuning stage, effectively preventing the forgetting problem.

[0023] Preferably, in step S5, the dynamic weighted fusion algorithm includes: obtaining the maximum probability value output by the fund level classification model after processing by the Softmax function as the classification confidence; obtaining the mean absolute error of the fund amount regression model on the validation set as the regression accuracy index; calculating dynamic weight coefficients based on the exponential function value of the classification confidence and the exponential function value related to the regression accuracy index, such that the dynamic weight coefficients increase when the classification confidence increases and decrease when the mean absolute error decreases; and performing a weighted summation of the predicted probability or mapping level value output by the fund level classification model and the predicted specific amount value output by the fund amount regression model according to the dynamic weight coefficients to generate the final cross-bank fund change prediction result.

[0024] By adopting the above technical solution, an adaptive weighting mechanism based on model performance was constructed. This mechanism utilizes the nonlinear characteristics of the exponential function to map the confidence level of the classification model and the error index of the regression model into dynamic weights. When the confidence level of the classification model is high, its weight is increased to leverage its qualitative judgment advantage; when the error of the regression model is small, i.e., its accuracy is high, its weight is increased to leverage its quantitative calculation advantage. This fusion method is more flexible and accurate than single-model or fixed-weight fusion, and can better cope with complex financial forecasting scenarios.

[0025] In a second aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the cross-bank funding prediction method for enterprises based on multi-dimensional feature fusion as described in the first aspect.

[0026] By adopting the above technical solution and executing the prediction method through hardware devices, the automated and intelligent prediction of interbank fund flows of enterprises is realized. It can handle large-scale concurrent data and provide efficient computing power support and decision-making basis for the fund liquidity management of banks and enterprises.

[0027] This invention provides a method and computer equipment for predicting cross-industry funding for enterprises based on multi-dimensional feature fusion. It has the following beneficial effects: 1. This invention constructs a panel pattern joint prediction model based on the TimeMixers architecture, sharing the parameters of the underlying neural network joint encoder across multiple enterprise samples, thus achieving cross-enterprise feature transfer learning. This technique effectively overcomes the problems of poor generalization ability and high maintenance costs caused by data sparsity in traditional single-enterprise independent modeling. It can capture common industry cash flow patterns such as supply chain payment cycles. While compressing model storage space, it only requires fine-tuning the fully connected layers to quickly adapt to different enterprises, improving the model's training efficiency and industry applicability.

[0028] 2. This invention employs a multi-dimensional feature fusion strategy combining LSTM and CNN, breaking the limitations of traditional methods that rely solely on single temporal features or static attributes. By capturing the long-term dependence and dynamic slope of capital flows using LSTM, and combining it with CNN to analyze the local patterns and node centrality topological features of the transaction time distribution matrix and capital flow network graph, and fusing the enterprise background vector processed by the embedding layer, the model can deeply mine the nonlinear correlations between temporal patterns, transaction behavior patterns, and enterprise attributes, thereby accurately identifying complex capital patterns such as high turnover rates but network marginalization, and improving prediction accuracy.

[0029] 3. This invention introduces a confidence-based dynamic weighted fusion algorithm, which runs classification and regression models in parallel and adaptively adjusts their output weights based on classification confidence and regression validation errors. This mechanism can dynamically optimize quantitative predictions of specific amounts while ensuring accurate qualitative judgments of fund change levels. It effectively balances the accuracy requirements of qualitative risk warning and quantitative fund management, enhancing the robustness and interpretability of prediction results in complex business scenarios. Attached Figure Description

[0030] Figure 1 This is a flowchart of the enterprise cross-industry fund prediction method based on multi-dimensional feature fusion of the present invention; Figure 2 This is a schematic diagram of the data acquisition and preprocessing process of the present invention; Figure 3 This is a schematic diagram of the customer group stratification screening process of the present invention; Figure 4 This is a schematic diagram of the multi-dimensional feature engineering construction process of the present invention; Figure 5 This is a schematic diagram illustrating the construction of the Panel pattern joint prediction model of the present invention; Figure 6 This is a schematic diagram of the phased training and dynamic weighted fusion prediction process of the present invention; Figure 7 This is a schematic diagram illustrating the percentage distribution of predicted value errors according to the present invention. Figure 8 This is a structural block diagram of the enterprise cross-industry capital prediction device based on multi-dimensional feature fusion according to the present invention; Figure 9 This is a schematic diagram of the hardware structure of the computer device of the present invention. Detailed Implementation

[0031] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] See attached document Figure 1 This invention provides a method for predicting cross-industry funds for enterprises based on multi-dimensional feature fusion. The method mainly includes five core steps: data collection and preprocessing, customer group stratification screening, multi-dimensional feature engineering construction, construction of panel pattern joint prediction model, and model training and dynamic fusion prediction.

[0033] The first step, S1, involves data collection and preprocessing. This step aims to integrate corporate customer account transaction data with basic corporate information to establish a unified data view. The collected data specifically includes customer ID, transaction date, transaction time, borrower / debtor identifier, transaction amount, fund usage field, basic corporate characteristics, corporate assets under management (AUM), and information on financial products held by the company. The system probes and cleans the raw data, handling missing and outlier values, and standardizes the data format to provide foundational input for subsequent feature extraction and model calculations.

[0034] Step S2 is then executed to segment and screen customer groups. This step groups customers based on the company's transaction characteristic indicators to determine the target objects suitable for modeling. The system calculates quantitative indicators for three dimensions: transaction stickiness, activity, and stability. The stickiness indicator is determined based on the number of months with transaction records within the statistical period; the activity indicator is determined based on the average daily transaction frequency of customers; and the stability indicator is determined based on the variance of customer transaction volume. Based on the calculation results of these three dimensions, customers are divided into different groups, and customers with high stable value and potential value are selected as target modeling objects. Customers with sparse, interrupted, or extremely unstable transaction data (i.e., those with no value or low value) are eliminated, thus ensuring the quality of the subsequent model training samples and the feasibility of predictions.

[0035] Next, step S3 is executed to construct multi-dimensional feature engineering. For the selected target modeling objects, the system extracts a multi-dimensional feature set covering time-series features, transaction behavior features, and enterprise background features. Time-series feature extraction uses a sliding window technique to process historical fund flows, capturing recent fund flow trends and dynamic indicators such as slope. Transaction behavior feature extraction involves constructing a transaction time distribution matrix or a fund flow network graph, and using a convolutional neural network to parse the matrix or graph structure, capturing local pattern features and topological features such as node centrality. Enterprise background feature extraction uses an embedding layer to vectorize discrete static information such as the enterprise's position in the industry chain and industry attributes. Finally, a feature set containing multiple highly discriminative derived features is constructed, covering three-dimensional information of time-series patterns, transaction patterns, and enterprise attributes.

[0036] Next, step S4 is executed to construct a Panel-pattern joint prediction model. This step builds a three-dimensional hybrid neural network based on the TimeMixers architecture and uses the Panel pattern to jointly model multi-customer data. The TimeMixers architecture integrates the aforementioned time-series features, transaction behavior features, and enterprise background features. In Panel mode, the sample data of all target enterprises share the underlying neural network joint encoder parameters, realizing feature transfer learning. This model structure includes a shared parameter layer and a task-specific layer. By sharing most of the model parameters among multiple entities, it compresses the model storage space and enables the model to learn common patterns of capital flows based on industry-standard data.

[0037] Finally, step S5 is executed to train the model and perform dynamic fusion prediction. This step utilizes a phased training strategy to optimize the model, including a pre-training phase that updates the shared layer parameters using the full dataset, and a fine-tuning phase that updates only the fully connected layer parameters for specific data. During prediction execution, the system simultaneously runs the capital level classification model and the capital amount regression model. Through a dynamic weighted fusion algorithm, the system dynamically calculates and adjusts the weights of the two model outputs based on the confidence level of the classification model and the validation set error index of the regression model. Finally, the weighted calculation yields the predicted value of interbank capital changes for the enterprise in the future time period. This prediction includes the specific predicted amount of interbank transfers in or out and the corresponding capital level range, thus achieving a quantitative prediction of the enterprise's interbank capital flows.

[0038] See attached document Figure 2 In one embodiment of the present invention, step S1, data acquisition and preprocessing, specifically includes the following process: The system first establishes a connection with the bank's corporate transaction behavior view database, extracting raw transaction data and related enterprise data based on a preset time window. The system then performs data integration, consolidating and aggregating the scattered transaction log data with the enterprise's static attribute data. The data fields collected at this stage specifically include customer ID, transaction date, transaction time, borrower / lender identifier, transaction amount, and fund usage field. Simultaneously, the system also collects basic enterprise characteristic data corresponding to the customer ID, enterprise managed asset size data, and data on the enterprise's financial product holdings. All of these data together constitute the initial modeling dataset.

[0039] The system performs data exploration on the initial modeling dataset, using statistical analysis methods to check the data's completeness and distribution patterns. The system scans the dataset for null values, duplicate records, and outliers that do not conform to business logic. For data quality issues discovered during exploration, the system performs data cleaning. For non-critical fields with missing values, the system fills in the missing values ​​using default values ​​or the mean; for critical identifier fields with missing values, the system removes the corresponding records. For anomalous data with obvious logical errors, such as negative transaction amounts, the system performs cleaning and filtering operations.

[0040] The system further standardizes the cleaned data before feature extraction. It converts the transaction date and time fields into a unified time-series format to support subsequent time-series feature extraction. The system standardizes the encoding of borrower / lender identifiers and fund usage fields to ensure consistent classification across data from different sources. It also standardizes the accuracy of numerical data such as transaction amounts and company asset management scale, removing format noise. After the above collection, exploration, cleaning, and standardization processes, the system generates a structured standard dataset, which serves as the foundational input data for subsequent customer group analysis and model building.

[0041] See attached document Figure 3 In one embodiment of the present invention, step S2, which involves segmenting and screening customer groups based on transaction characteristics, specifically includes the following process: The system first calculates three quantitative indicators based on the company's historical transaction data: transaction stickiness, transaction activity, and transaction stability. The transaction stickiness indicator is calculated by counting the number of months with transaction records within the most recent 20-month time window; a higher number of months indicates stronger transaction stickiness. The transaction activity indicator is calculated based on the average daily transaction frequency of customers within the statistical period, and customer activity is categorized into high or low levels based on the frequency value. The transaction stability indicator is calculated based on the statistical variance of customers' historical transaction amounts, measuring the magnitude of transaction fluctuations and classifying the company's transaction status into three levels: stable, relatively stable, and unstable.

[0042] Based on the calculation results of the three dimensions of transaction stickiness, transaction activity, and transaction stability, the system divides all pending enterprise customers into four different customer groups. The system defines the first customer group as Customer Group 0, characterized by fewer than 10 months of transactions, low transaction activity, and unstable transaction status, specifically no transaction records in the last 6 months, or fluctuating transaction amounts. The system defines the second customer group as Customer Group 1, characterized by between 10 and 15 months of transactions, low transaction activity, and unstable transaction status, specifically inconsistent monthly transaction amounts, with numerous months showing zero transaction amounts.

[0043] The system defines the third customer group as Customer Group 2. This group is characterized by having 18 to 20 months of transactions, high transaction activity, and stable transaction status. Specifically, transaction amounts are generally maintained at the same level, with very few months showing similar amounts. The system defines the fourth customer group as Customer Group 3. This group is characterized by having 15 to 20 months of transactions, with continuous transaction records in the last 3 months. It also exhibits high transaction activity and relatively stable transaction status, specifically with transaction amounts generally at the same level or remaining at the same level for the last 3 months, with only a few months showing different transaction amount levels.

[0044] The system performs sample screening based on the classification results to determine the target customer groups for final modeling. Because the transaction data of customer groups 0 and 1 exhibit sparse, discontinuous, or random characteristics, failing to provide continuous and stable time-series information to support predictive model training, the system removes customer data belonging to customer groups 0 and 1 from the modeling dataset. The system identifies customer groups 2 and 3 as meeting the conditions for fund prediction modeling. Due to their high continuity and relatively stable transaction behavior characteristics, the system retains customer data belonging to customer groups 2 and 3, using them as valid sample inputs for subsequent feature engineering and model training.

[0045] See attached document Figure 4 In one embodiment of the present invention, step S3, multi-dimensional feature engineering construction, specifically includes the following process: The system first performs time-series feature extraction to capture the dynamic evolution of corporate cash flows. It utilizes a sliding window technique with a Long Short-Term Memory (LSTM) network to process historical transaction amount sequences. The system sets the sliding window's time span to the most recent 30 days, and this window moves progressively along the time axis to capture continuous time slices of data. Through the gating mechanism of LSTM units, the system analyzes the cash flow sequence within the window, calculating and extracting dynamic indicators of cash inflow and outflow slopes. These indicators quantify the rate and trend of cash growth or decline within a short period, reflecting the time-series dependence of cash flows.

[0046] The system then performs transaction behavior feature extraction to analyze the enterprise's transaction patterns and network structure. First, the system constructs a transaction time distribution matrix based on the time periods and corresponding transaction frequencies. This two-dimensional matrix, defined by the time periods and transaction frequencies, quantifies the periodic distribution and stability of the enterprise's transaction behavior. Simultaneously, the system constructs a capital flow network graph based on counterparties and lending relationships, where nodes represent transaction accounts and edges represent capital flows. Building upon this, the system employs a convolutional neural network to perform deep feature extraction on the aforementioned matrix and network graph data. The system uses 3x3 convolutional kernels to perform convolution operations on the input data, extracting local pattern features from the capital flow network graph through local receptive fields and capturing topological features, including node centrality. These features reflect the enterprise's core position in the capital network and the clustering of counterparties.

[0047] The system synchronously performs enterprise background feature encoding to process non-numerical static attribute data. The system acquires discrete classification data regarding the enterprise's position in the industry chain and its industry attributes. This discrete data is then input into a pre-defined embedding layer for processing. The embedding layer uses table lookup and mapping operations to transform high-dimensional, sparse categorical features into low-dimensional, dense real-valued vectors. Specifically, the system converts each static attribute into a 128-dimensional feature vector. This vectorized representation method preserves the implicit relationships between different industries and positions in the industry chain while aligning static discrete features with dynamic continuous features in the vector space.

[0048] Finally, the system performs feature set integration. The system concatenates and fuses the time-series dynamic indicators extracted by LSTM, the transaction behavior topology features extracted by CNN, and the enterprise background feature vectors generated by the embedding layer. The system then filters and standardizes the fused features, ultimately constructing a derived feature set containing 17 high-discrimination indicators. This feature set comprehensively covers information across three dimensions: time-series patterns of funds, transaction spatial patterns, and inherent enterprise attributes, providing a multi-perspective, high-density input data foundation for the subsequent construction of the hybrid neural network model.

[0049] See attached document Figure 5 In one embodiment of the present invention, step S4, constructing the Panel pattern joint prediction model, specifically includes the following process: The system constructs a hybrid neural network model based on the TimeMixers architecture, which is designed as a three-dimensional hybrid network structure capable of simultaneously processing and fusing time-series features, transaction behavior features, and corporate background features. The system configures an LSTM-CNN joint encoder as the core feature extraction component within the TimeMixers architecture. This joint encoder contains parallel processing branches. The Long Short-Term Memory (LSTM) network branch receives and processes the time-series fund flow data generated by the sliding window, extracting long-term dependencies and trends in the time dimension. The Convolutional Neural Network (CNN) branch receives and processes the transaction time distribution matrix and fund flow network graph data, extracting local transaction patterns and network topology in the spatial dimension through convolution operations. The fully connected layer branch receives the embedding vectors of corporate background features. The system concatenates the feature vectors output from these three branches in a fusion layer to generate a high-dimensional feature representation containing multi-dimensional information.

[0050] The system employs a Panel model, a multi-agent collaborative modeling approach, to train the aforementioned TimeMixers architecture model. In Panel mode, instead of building independent time-series models for each individual enterprise customer, the system treats all customer data that meets the modeling selection criteria as a single dataset. The system standardizes time-series and feature data from different enterprises before inputting them into the same neural network model for training. This modeling approach allows the model to transcend the limitations of a single agent and learn common patterns and characteristics of cash flow across the entire industry or customer group, such as the payment cycle of specific supply chain links or industry-wide cash flow fluctuations.

[0051] In the implementation of Panel mode, the system achieves cross-enterprise feature transfer learning through a shared parameter layer mechanism. During model training, the weight and bias parameters in the LSTM-CNN joint encoder are shared among all input customer samples, meaning that data from different customers uses the same set of network parameters when passing through the encoder. This mechanism achieves a model parameter reuse rate of over 85%. Compared to the traditional method of storing a separate set of model parameters for each customer, the parameter sharing mechanism breaks the linear growth relationship between the number of models and the number of customers, reduces the space occupied by model files on storage media, and lowers the GPU memory consumption during model training and inference. The system retains only a small number of specific parameters in the final fully connected layer or output layer to adapt to the individual differences of different customers, thereby constructing a joint prediction model with both generalization and personalized adaptation capabilities.

[0052] See attached document Figure 6 In one embodiment of the present invention, step S5, phased training and dynamic weighted fusion prediction, specifically includes the following process: The system employs a phased training strategy to optimize the parameters of the constructed Panel pattern joint prediction model. First, a pre-training phase is performed, where the system inputs the entire enterprise customer dataset, after cleaning and feature engineering, into the model. During this phase, the system fully updates the shared layer parameters (weights of the LSTM-CNN joint encoder) to capture common cross-enterprise cash flow features. Then, a fine-tuning phase begins, where the shared layer parameters are frozen and updated only the fully connected layer parameters. In this fine-tuning phase, the learning rate is set to one-tenth of the pre-training learning rate. This low-learning-rate fine-tuning strategy adapts the model to the specific data distribution of each enterprise, preventing catastrophic forgetting and maintaining the model's generalization ability on common features.

[0053] During the model inference and prediction phase, the system establishes a dual-model fusion architecture. On one hand, the system utilizes gradient boosting decision tree models such as LightGBM to construct a classification model to predict whether a customer will experience cross-bank fund changes in the coming month, or to predict the level range of such changes. On the other hand, the system uses a regression model based on the TimeMixers architecture to predict the specific numerical values ​​of cross-bank fund changes. The system obtains the output probability distribution of the classification model and the predicted real values ​​of the regression model through parallel computing, providing input data for subsequent fusion calculations.

[0054] The system then applies a dynamic weighted fusion algorithm to integrate the classification and regression results to generate the final prediction output. The system first calculates dynamic weight coefficients based on the confidence level of the classification model and the historical validation error of the regression model. The system defines classification confidence as the maximum probability value output by the classification model after processing with the Softmax function. The system defines regression accuracy as the mean absolute error calculated by the regression model on the validation set. The system calculates the weight coefficients according to the following formula. : ; in, Represents the weighting coefficient; Represents the natural constant; Represents classification confidence. This represents the mean absolute error of the regression model on the validation set. This formula reflects the adaptive weight allocation mechanism of the model: when the classification confidence is high, The value increases; when the mean absolute error of the regression model is small, i.e., the prediction accuracy is high, the second term in the denominator increases, leading to... The value decreases, thus giving the regression model a higher weight.

[0055] Finally, the system performs a weighted sum of the classification model output and the regression model output based on the calculated weight coefficients to obtain the final prediction result. The output calculation formula is as follows: ; in, This represents the final fusion prediction result; This represents the predicted probability or the mapped rank value output by the classification model. This represents the predicted monetary value output by the regression model. Through this dynamic weighting mechanism, the system can automatically adjust the contribution of the two models based on the differences in performance of different samples in classification and regression tasks, outputting a predicted value that balances qualitative accuracy and quantitative precision.

[0056] See attached document Figure 7 In one specific application embodiment of the present invention, the method is applied to the fund management system for large and medium-sized manufacturing enterprise customers of a provincial branch of a commercial bank. This embodiment aims to verify the effectiveness of the enterprise interbank fund prediction method based on multi-dimensional feature fusion in actual business scenarios, specifically targeting relevant enterprise customers within the jurisdiction of the Jiangsu Provincial Branch, and modeling and predicting their interbank fund inflows and outflows for the next month.

[0057] The system first preprocesses the collected customer transaction data and segments customers based on their transaction characteristics. The system categorizes customers into four groups based on four dimensions: transaction stickiness, activity level, stability, and overall business value. Customer stickiness is determined by the number of months with transactions within the past 20 months; activity level is determined by the average daily transaction frequency; and stability is determined by the variance of transaction volume. The specific customer group segmentation criteria and characteristics are shown in Table 1 below. Table 1. Customer Group Segmentation Criteria

[0058] Based on the classification results, the system eliminates customer groups 0 and 1, which do not meet the modeling requirements, and selects customer groups 2 and 3 as the target modeling groups. For the selected target groups, the system uses a combined strategy of the LightGBM and TimeMixer models for training. The LightGBM model performs the classification task, predicting whether a customer will engage in cross-bank transactions and their corresponding fund level range within the next month. The TimeMixer model performs the regression task, predicting specific fund changes based on a panel pattern and a joint encoder of LSTM-CNN. The system introduces 17 derived features, including a transaction time distribution matrix and a fund flow network graph, and fuses the classification and regression results using a dynamic weighting algorithm.

[0059] The system uses the company's transaction data from the past three months as a validation set to evaluate the model's predictive performance on customer groups 2 and 3. Evaluation metrics include the true positive rate, true negative rate, and the accuracy of transaction tier predictions under different tier standards. The true positive rate refers to the ratio of predicted transactions that actually occur, while the true negative rate refers to the ratio of predicted no transactions that actually occur. Transaction tiers are divided into five levels: tens of thousands, millions, tens of millions, hundreds of millions, and billions, and two standards: four levels: millions, tens of millions, hundreds of millions, and billions. The specific classifications and tier prediction evaluation results are shown in Table 2 below. Table 2. Evaluation Table of Model Prediction Results

[0060] Based on the completed price level prediction, the system further verifies the accuracy of the prediction of the actual amount of interbank transactions. The system calculates the predicted value. Compared with actual value The percentage of error between them is calculated using the following formula: .

[0061] Figure 7 The chart shows the percentage distribution of predicted error for customer groups 2 and 3 across the two dimensions of interbank transfer-in and interbank transfer-out. The stacked bars of different colors in the chart represent different error ranges, including five intervals: less than 0.25, 0.25 to 0.5, 0.5 to 0.75, 0.75 to 1, and greater than 1.

[0062] Combination Figure 7 The data analysis shows that for the high-value customer group (customer group 2), the prediction error for interbank transfers in is the best, with over 50% of the samples having an error less than 0.25 (blue block). Combined with samples with errors between 0.25 and 0.5 (orange block), the percentage of prediction results with a cumulative error within 50% is approximately 75%. For interbank transfers out of customer group 2, approximately 40% of the samples have an error less than 0.25, and approximately 60% have a cumulative error within 50%. For the potential-value customer group (customer group 3), the prediction accuracy for interbank transfers in is slightly lower than that of customer group 2, with approximately 40% of the samples having an error less than 0.25 and approximately 60% having a cumulative error within 50%. However, for interbank transfers out of customer group 3, approximately 30% of the samples have an error less than 0.25, and nearly 50% have a cumulative error within 50%. Overall, for enterprise clients with high stability, the method described in this invention can control the error of most prediction results within the range of 25% to 50%, verifying the effectiveness of the model in specific monetary prediction tasks.

[0063] See attached document Figure 8The present invention provides an enterprise cross-industry capital prediction device based on multi-dimensional feature fusion. The device includes a data acquisition module, a clustering and screening module, a feature construction module, a joint modeling module, and a fusion prediction module.

[0064] The data acquisition module is configured to connect to the bank's internal data storage system and obtain a view table of corporate transaction behavior. The data fields collected by this module include customer ID, transaction date, transaction time, borrower / debtor identifier, transaction amount, purpose of funds, basic corporate characteristics, size of assets under management, and information on financial products held by the company. The data acquisition module is also configured to clean the raw data, handle missing and outlier values, and transmit the processed, standardized data to the clustering and filtering module.

[0065] The segmentation and filtering module connects to the data acquisition module and is configured to receive standardized data and calculate enterprise transaction characteristic indicators. This module calculates transaction stickiness by counting the number of months with transaction records within the past twenty months, calculates activity based on average daily customer transaction frequency, and calculates stability based on the variance of customer transaction volume. Based on the calculated stickiness, activity, and stability indicators, the module categorizes customers into four groups: no-value customers, low-value customers, high-stability-value customers, and potential-value customers. This module is further configured to execute filtering logic, eliminating no-value and low-value customers, retaining only high-stability-value and potential-value customers as the target modeling group, and transmitting the target group data to the feature construction module.

[0066] The feature construction module is connected to the clustering and filtering module and is configured to extract multi-dimensional features from the target group data. This feature construction module includes a time-series feature extraction unit, configured to capture capital flow trends using a long short-term memory network sliding window technique and extract dynamic indicators of the slope of capital inflows and outflows over the past thirty days. It also includes a transaction behavior feature extraction unit, configured to construct a capital flow network graph and use a convolutional neural network to parse the graph and extract node centrality topological features. Finally, it includes a corporate background feature encoding unit, configured to vectorize the static information of a company's position in the industry chain and its industry attributes through an embedding layer. The feature construction module ultimately outputs a multi-dimensional derived feature set, including a transaction time distribution matrix, to the joint modeling module.

[0067] The joint modeling module is connected to the feature construction module and is configured to build and train the predictive model. This joint modeling module uses a panel model for multi-agent joint modeling, achieving multi-customer feature transfer learning through shared parameter layers. The joint modeling module is configured to build a hybrid neural network based on the TimeMixers architecture, which integrates time-series features, transaction behavior features, and corporate background features. The training process of this joint modeling module includes a pre-training phase and a fine-tuning phase. In the pre-training phase, a general model is trained using a corporate dataset, with the shared layers using a joint encoder of a long short-term memory network and a convolutional neural network. In the fine-tuning phase, the parameters of the shared layers are frozen, and only the parameters of the fully connected layers are updated. The joint modeling module simultaneously maintains a classification model for predicting fund levels and a regression model for predicting specific monetary values.

[0068] The fusion prediction module connects with the joint modeling module and is configured to generate the final interbank fund transfer forecast results. This fusion prediction module employs a dynamic weighted fusion algorithm, dynamically calculating weight coefficients based on the confidence level of the classification model and the mean absolute error of the regression model. The module reads the maximum probability value output by the classification model as the classification confidence level and the mean absolute error of the regression model on the validation set, calculating the respective weights of the classification and regression results through exponential operations. The fusion prediction module then weights and sums the probability output of the classification model and the numerical output of the regression model according to the calculated weights, outputting the predicted interbank transfer inflow and outflow amounts for the company in the coming month.

[0069] See attached document Figure 9 The present invention also provides a computer device, which includes a processor, a memory, a network interface, and a bus. The processor, memory, and network interface communicate and transmit data with each other through the bus.

[0070] The memory configuration is used to store computer programs and related data, including transaction time distribution matrices, eigenvectors, and model parameters generated during the processing of the various modules of the aforementioned device. The memory can be high-speed random access memory or non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0071] The network interface configuration is used to communicate with external devices or the bank's internal network, receive raw transaction data, and send prediction results to downstream business systems.

[0072] The processor is configured to invoke computer programs stored in memory to execute a cross-bank fund prediction method based on multi-dimensional feature fusion. Specifically, the processor executes the following steps: integrating multiple data fields from the intra-bank corporate transaction behavior view data table for data feature processing; segmenting customers based on corporate transaction stickiness, activity, and stability dimensions, and determining the modeled customer groups; performing feature engineering on the modeled customer groups to extract time-series features, transaction behavior features, and corporate background features; training the model based on the Panel pattern and TimeMixers architecture, including pre-training and fine-tuning stages; and using a dynamic weighted fusion algorithm to combine the outputs of the classification and regression models to calculate and output the prediction results for cross-bank fund transfers in and out. By executing the above instructions, the processor constructs a model for multiple customers within the same time series, reduces computational power consumption through parameter sharing, and improves prediction accuracy through multi-dimensional feature fusion.

[0073] An embodiment of the present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor of a computer device, it implements the various steps of the above-described method for predicting cross-bank funding based on multi-dimensional feature fusion. The storage medium may include various media capable of storing program code, such as read-only memory, random access memory, disk, or optical disk. When the computer program is executed, it can perform data acquisition and preprocessing, customer segmentation and screening based on transaction characteristics, construction of multi-dimensional features including a transaction time distribution matrix, joint modeling based on Panel pattern and TimeMixers architecture, and dual-model fusion prediction based on a dynamic weighting algorithm.

Claims

1. A cross-industry funding prediction method for enterprises based on multi-dimensional feature fusion, characterized in that: Includes the following steps: S1. Data collection and preprocessing: Integrate enterprise customer account transaction data and enterprise basic information data to establish a unified data view, explore, clean and standardize the format of the raw data, and generate the initial modeling dataset. S2. Customer group segmentation and screening: Receive the initial modeling dataset, divide customers into different groups according to the enterprise's transaction characteristic indicators, screen out high stable value customer groups and potential value customer groups as target modeling objects, and eliminate no-value customers and low-value customer groups. S3. Construct a multi-dimensional feature engineering: For the target modeling object, extract a multi-dimensional feature set covering time-series features, transaction behavior features, and enterprise background features. The multi-dimensional feature set covers three-dimensional information of time-series patterns, transaction modes, and enterprise attributes. S4. Construct a Panel-based joint prediction model: Construct a three-dimensional hybrid neural network based on the TimeMixers architecture. The three-dimensional hybrid neural network integrates the time-series features, the transaction behavior features, and the enterprise background features. It uses the Panel model to jointly model multiple customer data. In the Panel model, the sample data of all the target modeling objects share the parameters of the underlying neural network joint encoder to achieve feature transfer learning. S5. Model Training and Dynamic Fusion Prediction: The Panel model joint prediction model is optimized using a phased training strategy. During prediction execution, the capital level classification model and the capital amount regression model are run in parallel. The weighted fusion algorithm is used to dynamically calculate and adjust the weights of the output results of the two models based on the confidence level of the capital level classification model and the validation set error index of the capital amount regression model. The predicted value of cross-bank capital changes for the enterprise in the future time period is obtained through weighted calculation.

2. The enterprise cross-industry funding prediction method based on multi-dimensional feature fusion according to claim 1, characterized in that, In step S1, the data fields collected include customer number, transaction date, transaction time, borrower / lender identifier, transaction amount, fund usage field, basic enterprise characteristics, enterprise managed asset size, and information on financial products held by the enterprise. The cleaning of the original data includes: filling non-critical fields with missing values ​​using default values ​​or average values; removing records with missing values ​​in critical identifier fields; and filtering logically erroneous data with negative transaction amounts.

3. The enterprise cross-industry funding prediction method based on multi-dimensional feature fusion according to claim 1, characterized in that, In step S2, the transaction characteristic indicators include transaction stickiness indicators, transaction activity indicators, and transaction stability indicators. The transaction stickiness index is determined based on the number of months with transaction records within the statistical period; The transaction activity index is determined based on the average daily transaction frequency of customers; The transaction stability index is determined based on the statistical variance of the customer's historical transaction amount.

4. The enterprise cross-industry funding prediction method based on multi-dimensional feature fusion according to claim 3, characterized in that, Step S2 further includes: Customers whose number of transactions in a month is less than a preset first threshold and whose transaction status is unstable are defined as the valueless customers. Customers whose number of transactions in a given month falls between the first threshold and the preset second threshold and whose transaction status is unstable are defined as the low-value customer group. Customers whose number of transactions in a month is between a preset third threshold and a preset fourth threshold and whose transaction status is stable are defined as the high-stability-value customer group. Customers whose number of transaction months falls between the second threshold and the fourth threshold and who have a continuous transaction record in the most recent preset month are defined as the potential value customer group. The data of the high-value customer group and the potential-value customer group are retained as input data for the target modeling object.

5. The enterprise cross-industry funding prediction method based on multi-dimensional feature fusion according to claim 1, characterized in that, In step S3, the extraction of the temporal features includes: The Long Short-Term Memory Network (LSTM) sliding window technique is used to process historical fund flow sequences. The sliding window is set to move gradually on the time axis to capture continuous time slice data. The fund change sequence within the window is analyzed through the gating mechanism of the LTM unit, and dynamic indicators of fund inflow slope and fund outflow slope are calculated and extracted.

6. The enterprise cross-industry funding prediction method based on multi-dimensional feature fusion according to claim 5, characterized in that, In step S3, the extraction of the transaction behavior features includes: Construct a transaction time distribution matrix, which is a two-dimensional matrix defined by time periods and transaction frequencies; Construct a fund flow network graph, where nodes represent transaction accounts and edges represent fund flows. The convolutional neural network is used to extract features from the transaction time distribution matrix and the capital flow network graph. Convolutional operations are performed by convolution kernels to capture local pattern features and node centrality topological structure features. The extraction of enterprise background features includes: obtaining discrete classification data of the enterprise's position in the industrial chain and industry attributes, inputting the discrete classification data into the embedding layer, and converting the category features into low-dimensional dense real number vectors.

7. The enterprise cross-industry funding prediction method based on multi-dimensional feature fusion according to claim 1, characterized in that, In step S4, the TimeMixers architecture includes a joint encoder of a long short-term memory network and a convolutional neural network; The joint encoder includes parallel processing branches, which are respectively used to process the time series data of fund flows generated by the sliding window, the transaction time distribution matrix and the fund flow network graph data, as well as the embedding vector of enterprise background features. The parameter sharing mechanism in the Panel mode is configured such that the weight parameters and bias parameters in the joint encoder are shared among all input customer samples, and parameters adapted to the individual differences of different customers are retained only in the fully connected layer.

8. The enterprise cross-industry funding prediction method based on multi-dimensional feature fusion according to claim 1, characterized in that, In step S5, the phased training strategy includes: Pre-training phase: Input the full enterprise customer dataset into the model and update the shared layer parameters in the Panel pattern joint prediction model. Fine-tuning phase: Freeze the shared layer parameters, update only the fully connected layer parameters of the Panel mode joint prediction model, and set the learning rate of the fine-tuning phase to be lower than the learning rate of the pre-training phase.

9. The enterprise cross-industry funding prediction method based on multi-dimensional feature fusion according to claim 1, characterized in that, In step S5, the dynamic weighted fusion algorithm includes: The maximum probability value output by the fund level classification model after processing by the Softmax function is used as the classification confidence score. The mean absolute error of the fund amount regression model on the validation set is used as the regression accuracy index. The dynamic weight coefficient is calculated based on the exponential function value of the classification confidence and the exponential function value related to the regression accuracy index, so that the dynamic weight coefficient increases when the classification confidence increases and decreases when the mean absolute error decreases. The predicted probability or mapping level value output by the fund level classification model and the predicted specific amount value output by the fund amount regression model are weighted and summed according to the dynamic weight coefficient to generate the final cross-bank fund change prediction result.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the enterprise cross-bank funding prediction method based on multi-dimensional feature fusion as described in any one of claims 1 to 9.

Images