Multi-dimensional fund flow dynamic prediction analysis method

Abstract

The invention provides a multi-dimensional fund flow dynamic prediction analysis method, and relates to the technical field of data processing, and the method comprises the steps: carrying out the time label standardization and delay correction processing of original data, and aligning the asynchronous data of different time granularities; carrying out self-adaptive calculation on the fluctuation rate, the periodic change rate and the cross correlation degree of the data of each dimension, and extracting an influence factor; calculating a time sequence dependency coefficient and interaction intensity among the dimensions, establishing a time sequence coupling mapping matrix, and generating capital flow associated topological data; when any dimension of data fluctuates abnormally, calculating the conduction intensity of the data along the topological path, calculating and correcting the time attenuation coefficient of the conduction intensity in the topological path, and dynamically adjusting the topological structure; detecting a high-sensitivity path, identifying a risk conduction source, tracing and positioning an abnormal chain, and generating capital flow risk trend prediction data; according to the invention, the autonomy and accuracy of the fund flow dynamic prediction analysis are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a multi-dimensional dynamic prediction and analysis method for cash flow. Background Technology

[0002] Currently, existing technologies primarily predict and analyze cash flows using time-series statistical models (such as ARIMA, VAR, or GARCH models) and some machine learning algorithms (such as random forests and support vector regression). These methods typically rely on single-dimensional historical transaction data as the main input, such as account transaction records, market trading volume, or asset price fluctuations. The prediction process often uses historical data within a fixed window as a basis, fitting time-series trends to predict future cash flow changes. However, when dealing with the multi-source and multi-dimensional coupling relationships of cash flows in complex economic systems, these methods often lack dynamic modeling capabilities and cannot fully reflect the nonlinear correlations between multiple factors such as macroeconomic indicators, industry capital flows, public opinion fluctuations, and policy changes. Consequently, the prediction results are insufficiently accurate in scenarios involving multiple interacting factors.

[0003] In practical applications, such as when commercial banks monitor the liquidity risk of corporate accounts, existing technologies often rely solely on historical account balances and transaction flows for short-term predictions. When sudden market events (such as adjustments to industry credit policies or disruptions in upstream supply chain funding) cause abnormal fluctuations in the cash flow of a particular type of enterprise, traditional models cannot dynamically capture the temporal correlation of this cross-dimensional information. Specifically, the bank's predictions may still show that the enterprise's cash flow is stable, but in reality, due to the combined effects of asynchronous factors such as upstream settlement delays and downstream payment lags, the enterprise's cash flow may have already entered a high-risk state. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-dimensional dynamic prediction and analysis method for cash flow, aiming to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A multi-dimensional dynamic prediction and analysis method for cash flow, the method comprising: Obtain historical transaction data, industry economic indicators, policy event information, supply chain settlement records, and market sentiment data of the target account group as raw data; The original data is time-stamped and delayed, and asynchronous data with different time granularities are aligned to generate time series matrix data. Based on time series matrix data, the volatility, periodic change rate and cross correlation of each dimension of data are adaptively calculated to extract the influencing factors that can reflect the dynamic characteristics of multiple dimensions and generate multi-dimensional influencing factor data. Based on multi-dimensional influencing factor data, the temporal dependence coefficient and interaction strength between each dimension are calculated, a temporal coupling mapping matrix is ​​established, and capital flow correlation topology data is generated. Based on the capital flow correlation topology data, when any dimension of data shows abnormal fluctuations, its transmission strength along the topology path is calculated, and the time lag effect is evaluated based on the response delay between nodes to obtain the time decay coefficient of the transmission strength in the corrected topology path, so as to dynamically adjust the topology structure and generate dynamic propagation data. Based on dynamic propagation data, highly sensitive paths are detected and risk transmission sources are identified. Abnormal chains are traced and located, and fund flow risk trend prediction data is generated. This data is then used to output dynamic prediction results and risk warning signals for the future fund flows of the target account group.

[0006] The above-described solution of the present invention has at least the following beneficial effects: First, by constructing a multi-dimensional input system that includes historical transaction data, industry economic indicators, policy event information, supply chain settlement records, and market sentiment data, the limitations of existing technologies that rely on single historical data are overcome. Through time-label standardization and delay correction, data at different time granularities are synchronized, enabling a unified expression of the multi-source dynamic characteristics of capital flows at the macro, industry, and micro levels.

[0007] Secondly, by adaptively calculating the volatility, periodicity, and cross-correlation of time-series matrix data, this method extracts influencing factors that reflect multidimensional dynamic characteristics, capturing the cyclical patterns and cross-dimensional linkages of capital flows. This method achieves dynamic updates at the feature layer, automatically adjusting the calculation focus as data changes, and more sensitively reflects the impact of external economic events on capital flows.

[0008] Furthermore, by establishing a temporal coupling mapping matrix and generating a capital flow correlation topology, a structured expression of the dependency relationship between macroeconomics, industry trends, and corporate behavior is realized, thereby quantifying the transmission relationship of multiple factors in a complex economic system and providing a reliable foundation for subsequent risk identification.

[0009] Furthermore, when abnormal fluctuations occur in any dimension, the propagation strength can be calculated along the topological path, and the time lag effect can be evaluated by combining the node response delay. The path strength can be corrected for time decay, so that the propagation structure can be dynamically updated, accurately reflecting the diffusion process of abnormal fluctuations in the network and avoiding the lagging judgment of static models.

[0010] Finally, by conducting risk detection and source tracing analysis on dynamic propagation data, highly sensitive paths are identified and risk transmission sources are located, generating prediction results for cash flow risk trends. In practical applications, such as banks monitoring the liquidity of enterprise groups, this invention can identify potential risks and output early warning signals in the early stages of supply chain settlement delays or policy adjustments, providing timely basis for financial decision-making. Attached Figure Description

[0011] Figure 1 This is a flowchart of the multi-dimensional dynamic prediction and analysis method for cash flow provided in an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown, embodiments of the present invention propose a multi-dimensional dynamic prediction and analysis method for cash flow, the method comprising: Obtain historical transaction data, industry economic indicators, policy event information, supply chain settlement records, and market sentiment data of the target account group as raw data; The original data is time-stamped and delayed, and asynchronous data with different time granularities are aligned to generate time series matrix data. Based on time series matrix data, the volatility, periodic change rate and cross correlation of each dimension of data are adaptively calculated to extract the influencing factors that can reflect the dynamic characteristics of multiple dimensions and generate multi-dimensional influencing factor data. Based on multi-dimensional influencing factor data, the temporal dependence coefficient and interaction strength between each dimension are calculated, a temporal coupling mapping matrix is ​​established, and capital flow correlation topology data is generated. Based on the capital flow correlation topology data, when any dimension of data shows abnormal fluctuations, its transmission strength along the topology path is calculated, and the time lag effect is evaluated based on the response delay between nodes to obtain the time decay coefficient of the transmission strength in the corrected topology path, so as to dynamically adjust the topology structure and generate dynamic propagation data. Based on dynamic propagation data, highly sensitive paths are detected and risk transmission sources are identified. Abnormal chains are traced and located, and fund flow risk trend prediction data is generated. This data is then used to output dynamic prediction results and risk warning signals for the future fund flows of the target account group.

[0014] In this embodiment of the invention, by performing dynamic predictive analysis on multi-dimensional capital flow data, a comprehensive judgment and trend prediction of the capital flow status from different sources is achieved. The method first extracts raw data from the target account group's historical transaction data, industry economic indicators, policy event information, supply chain settlement records, and market sentiment data. Through time-label standardization and delay correction, asynchronous data at different time granularities are unified in the time dimension, thereby obtaining time series matrix data that can be used for time series analysis. This step ensures the temporal consistency between multi-dimensional data, providing a foundation for subsequent dynamic feature extraction.

[0015] After data alignment, the volatility, periodicity, and cross-correlation of each dimension of the time series matrix are adaptively calculated to extract a set of influencing factors that reflect the changing characteristics among multiple dimensions. This process can identify the dominant changing trends and interactions of different dimensions in the time series, thereby generating multi-dimensional influencing factor data and providing a quantitative basis for subsequent time series relationship modeling.

[0016] Based on multi-dimensional influencing factor data, the temporal dependency coefficients and interaction strengths between each dimension are further calculated to construct a temporal coupling mapping matrix, and capital flow correlation topology data is generated accordingly. This topology data can structurally describe the transmission path between macroeconomic indicators, industry capital flows, and corporate behavior, thereby establishing the correlation between capital flows and multiple factors.

[0017] After the topology is formed, when an abnormal fluctuation occurs in any dimension, the signal propagation strength is calculated along the topological path, and the time lag effect is assessed by combining the response delay between nodes. This allows for the correction of the time decay coefficient in the path, thus achieving dynamic adjustment of the topology. This process can reflect the interconnected changes between multi-dimensional data in real time, generating dynamic propagation data and providing a real-time feedback basis for anomaly propagation analysis.

[0018] After obtaining dynamic propagation data, the segments with high transmission intensity and continuous changes in the path are detected to identify highly sensitive paths and locate risk transmission sources. Through reverse tracing and diffusion analysis of abnormal chains, fund flow risk trend prediction data is generated. This data can be used to determine potential risk accumulation areas and their possible directions of influence during fund flows, and output dynamic prediction results and risk warning signals for the future fund flows of target account groups, thereby enabling early identification and response during the prediction phase.

[0019] For example, in a supply chain settlement system, if a company experiences temporary cash flow difficulties due to delayed payments from upstream suppliers, this method can identify abnormal paths in industry cash flow through multi-dimensional data alignment and dynamic propagation analysis. When this anomaly propagates to multiple company nodes within the same supply chain, it identifies risk chains with high propagation intensity and generates risk trend prediction results. This allows for early warning of potential cash flow risks in real-world scenarios, assisting financial institutions or corporate management in making liquidity decisions.

[0020] In a preferred embodiment of the present invention, the acquisition of historical transaction data, industry economic indicators, policy event information, supply chain settlement records, and market sentiment data of the target account group as raw data specifically includes: First, transaction flow information for the target enterprise group is extracted from the account databases of financial institutions, including basic records such as fund inflows, outflows, balance changes, and transaction frequency. Second, industry economic indicators, such as industry output, inventory changes, financing scale, and average settlement cycle, are obtained from publicly available industry statistics channels or third-party data interfaces. For policy event information, policy release dates, implementation areas, and content change tags related to the target industry can be extracted from policy announcement databases or news monitoring, and key events can be marked with timestamps. Supply chain settlement records include settlement cycles, accounts receivable and payable periods, and actual payment delays of upstream and downstream trading partners. Market sentiment data can be obtained through public opinion analysis, extracting sentiment bias values ​​or market confidence indices from sources such as news reports, social media, and market commentary. After unified coding and cleaning, the above data forms a structured raw dataset, providing multi-dimensional input sources for subsequent time series analysis.

[0021] In a preferred embodiment of the present invention, the original data is subjected to time-stamping standardization and delay correction processing. After aligning asynchronous data with different time granularities, time series matrix data is generated, specifically including: First, a unified time label is added to the raw records from different data sources, mapping policy event information, supply chain settlement records, and market sentiment data by date or week to align them with transaction data on the same timeline. Then, data with coarser time granularity (such as quarterly economic indicators) is extended using interpolation to align with daily or weekly data. For records with time lags, such as supply chain settlement information and policy effect data, a time offset correction method is used. This method calculates the average lag period by comparing the difference between the event occurrence time and the actual cash flow reaction time, and adjusts the data timeline accordingly. Finally, the aligned data from each dimension is integrated into a matrix in chronological order, with rows representing the time series and columns representing different data dimensions, forming a time series matrix that provides structured input for subsequent feature calculations.

[0022] In a preferred embodiment of the present invention, based on time series matrix data, the volatility, periodicity, and cross-correlation of each dimension of the data are adaptively calculated to extract influencing factors that can reflect multi-dimensional dynamic characteristics, generating multi-dimensional influencing factor data, including: Data smoothing and anomaly removal are performed on the time series matrix data to obtain time series feature data; Based on time series characteristic data, the magnitude of change in adjacent time periods is calculated within a time sliding window to determine volatility indicators for each dimension. Based on volatility indicators, pattern recognition is performed on the trend direction of each dimension within multiple period windows to generate periodic change rate data. By cross-matching periodic rate of change data with sequences of different time dimensions, the degree of response under time shift is analyzed to obtain cross-dimensional cross-correlation data. Volatility indicators, cyclical change rate data, and cross-correlation data are weighted and integrated according to their dimensional importance to generate multi-dimensional influencing factor data.

[0023] In this embodiment of the invention, through multi-level analysis of time series matrix data, core information reflecting dynamic changes can be extracted from the original capital flow data. First, data smoothing and anomaly removal effectively reduce the interference of short-term noise on trend judgment, making subsequent calculations more stable. Second, calculating the magnitude of change and determining the volatility index within a time sliding window quantifies the activity of capital flows in different time periods, providing a foundation for periodic feature identification. Subsequently, by performing pattern recognition on the trend directions of multiple periodic windows, periodic change rate data is generated, enabling the capture of the periodic fluctuation patterns of capital flows. Then, by cross-matching time series data of different dimensions, the response degree under time offset is calculated, enabling the discovery of potential linkages between dimensions. Finally, by weighted fusion of volatility, periodic change rate, and cross-correlation data, multi-dimensional influencing factor data is generated, ensuring that the results not only reflect single-dimensional trends but also demonstrate the dynamic coupling relationships between multiple dimensions, thus providing accurate input for capital flow topology modeling.

[0024] In a preferred embodiment of the present invention, pattern recognition is performed on the trend direction of each dimension within multiple periodic windows based on the volatility index to generate periodic rate of change data, specifically including: First, sliding period windows of different lengths, such as a week, a month, or a quarter, are selected in the time series matrix to capture short-term and long-term patterns of change. Within each window, the average rate of change and directional change of each dimension are calculated to determine whether the capital flow exhibits an upward, downward, or stable trend within that period. Then, the trend change patterns between consecutive time windows are analyzed, recording the number of trend reversals and their duration as periodic change characteristics. For example, when a dimension shows an upward trend over three consecutive periods, it is marked as a periodic positive change; if a short-term reversal occurs, it is marked as a volatility adjustment. By statistically analyzing the repetition and persistence of trend directions across different periods, periodic change rate data is generated to reflect the regular fluctuation characteristics of capital flows over time.

[0025] In a preferred embodiment of the present invention, the periodic rate of change data is cross-matched with sequences of different time dimensions to analyze the degree of response under time shift, thereby obtaining cross-dimensional cross-correlation data, specifically including: First, select periodic rate of change data from two or more different dimensions from the time series matrix, such as the rate of change in market sentiment and the rate of change in trading activity. Then, within a certain time offset range, progressively shift the change sequence of one dimension relative to the other dimension, and calculate the proportion or trend synchronization of the two dimensions with the same direction of change at each shift step. When the change directions of the two dimensions are consistent across multiple time periods, they are considered to have a strong positive correlation; if the change directions are opposite, they are considered to have a negative correlation. By comparing the changes in correlation under different offsets, the optimal response time difference between each dimension can be determined. Finally, this time difference and the corresponding correlation strength are recorded as cross-dimensional cross-correlation data, thus reflecting the mutual influence of different economic factors over time and providing a basis for subsequent multi-dimensional coupling modeling.

[0026] In a preferred embodiment of the present invention, based on multi-dimensional influencing factor data, the temporal dependency coefficients and interaction strengths between each dimension are calculated, a temporal coupling mapping matrix is ​​established, and capital flow correlation topology data is generated, including: Based on multi-dimensional influencing factor data, the differences in the direction and magnitude of change of each dimension between adjacent time points are calculated to obtain basic information on time-series changes. Based on the basic information of time series changes, calculate the response strength of each dimension to changes in other dimensions, and determine the time series dependency coefficient under time delay; The temporal dependency coefficients are aggregated according to the dimensional pairing method, and the duration and direction consistency of the interaction between each dimension are calculated to obtain the interaction strength information; Based on the interaction strength information and the temporal dependency coefficient, a temporal coupling mapping matrix is ​​established, and low-relevance paths are eliminated through a weight screening mechanism to generate capital flow correlation topology data.

[0027] In this embodiment of the invention, by analyzing multi-dimensional influencing factor data, a correlation structure of capital flows at the macro, industry, and enterprise levels can be established. First, the differences in the direction and magnitude of changes between adjacent time points are calculated to obtain basic information on temporal changes, thereby reflecting the synchronization and offset relationships between different dimensions. Second, the response strength of each dimension to changes in other dimensions is calculated to determine the temporal dependency coefficient under time delay conditions, quantifying the mutual influence between different dimensions. Further, by summarizing these dependency coefficients and analyzing directional consistency, interaction strength information is obtained, reflecting the strength and duration of transmission between different economic factors. Finally, by establishing a temporal coupling mapping matrix and performing weight filtering, noisy or weakly correlated connections can be removed, retaining only stable and causally significant transmission paths. This method transforms capital flow relationships from discrete indicators into a visualized topological structure, providing a structured data foundation for anomaly propagation detection and dynamic risk identification.

[0028] In a preferred embodiment of the present invention, based on multi-dimensional influencing factor data, the differences in the direction and magnitude of change of each dimension between adjacent time points are calculated to obtain basic information on time-series changes, specifically including: First, time series data for each dimension are extracted from multi-dimensional influencing factor data, such as transaction activity, industry prosperity index, policy sensitivity, and supply chain settlement rate. Then, the changes in each dimension between adjacent time points are calculated to determine whether the change direction is upward, downward, or unchanged, and the magnitude of the change is recorded. To avoid interference from short-term fluctuations in trend judgment, the magnitude of the changes is smoothed, for example, using moving averages or weighted smoothing, to better reflect the overall trend. Next, the directional and magnitude change results for each dimension are combined to form a time-series change feature set. For dimensions that change continuously in the same direction, the duration of the change is recorded; for dimensions that frequently reverse, they are marked as high-volatility features. Finally, the above results are compiled into a time-series change basic information dataset, providing input for subsequent calculations of dependencies and interaction strengths.

[0029] In a preferred embodiment of the present invention, based on the basic information of time-series changes, the response intensity of each dimension to changes in other dimensions is calculated, and the time-series dependency coefficient under time delay is determined, specifically including: First, pairwise dimension matching is performed on the basic information of time-series changes to analyze whether a change in one dimension will cause a delayed response in another. For example, when the industry prosperity index rises, does the enterprise transaction activity show an upward trend after several time units? By statistically analyzing the consistency of the direction of change and the similarity of the magnitude across multiple time windows, the degree of influence of one dimension on another can be determined. Second, in the presence of time delay, the response strength at different time offsets is calculated, and the duration and stability of the response are compared to determine the optimal time delay length. The shorter the time delay and the higher the response strength, the stronger the dependency between the two dimensions. Finally, the dependency strength between different dimensions is normalized to obtain a value between the lowest and highest dependency levels, and this value is used as the time-series dependency coefficient to describe the dynamic response relationship between the two dimensions.

[0030] In a preferred embodiment of the present invention, the temporal dependency coefficients are aggregated according to dimensional pairing, and the duration and direction consistency of the interaction between each dimension are calculated to obtain the interaction strength information, specifically including: First, based on the temporal dependency coefficients between each dimension, their changing trends over different time periods are statistically analyzed to determine whether the dependency relationship is short-term or long-term stable. If one dimension consistently exerts a positive or negative influence on another dimension, the interaction lasts for a relatively long time. Second, the directional consistency of the dependency relationship is compared, i.e., whether the direction of the dependency relationship remains consistent across multiple time windows. High directional consistency indicates that the interaction between the two dimensions is stable. Next, the duration and directional consistency are combined, and the interaction strength is determined through a weighted evaluation method. Dimensions with high interaction strength represent factor combinations that significantly impact changes in capital flows. The final output interaction strength information can be used to construct a topological mapping matrix, making the transmission path of capital flows temporally reasonable and hierarchical.

[0031] In a preferred embodiment of the present invention, based on the capital flow-related topology data, when any dimension of the data shows abnormal fluctuations, its transmission strength along the topology path is calculated, and the time lag effect is evaluated based on the response delay between nodes to obtain the time decay coefficient of the transmission strength in the corrected topology path. The topology structure is then dynamically adjusted to generate dynamic propagation data, including: Based on the topological data of fund flow association, monitor the real-time change rate of the corresponding dimension of each node to form node status monitoring information; Based on node status monitoring information, when the change of any node exceeds the preset fluctuation threshold, anomaly propagation analysis is triggered to extract the adjacent paths of the abnormal node and form path candidate information. Based on the path candidate information, combined with the change magnitude of the abnormal nodes, the path weights, and the response magnitude of the target nodes, the transmission intensity of the abnormal signal on each path is calculated. Based on the path candidate information, the time difference of the state changes of adjacent nodes on the path is compared to obtain the response delay, and the time decay coefficient is determined according to the ratio of the response delay to the preset time window. Based on the time decay coefficient, the propagation intensity of each node is corrected for time decay to form path propagation analysis information; For paths whose propagation intensity exceeds a preset first propagation intensity threshold in the path propagation analysis information, perform topology updates, adjust path weights and directions, and form an updated topology structure. Dynamic propagation data is generated based on the changing trends between paths in the updated topology.

[0032] In this embodiment of the invention, dynamic response analysis of the fund flow propagation process is achieved through real-time monitoring and adjustment of the fund flow-related topology data. First, node status monitoring continuously tracks the real-time change rate of each dimension, ensuring that abnormal fluctuations can be quickly identified. When a node's change exceeds a threshold, its adjacent paths are extracted to form a candidate set, providing a data range for abnormal propagation analysis. Next, by combining the abnormal node's change amplitude, path weights, and target node's response amplitude, the propagation intensity of the abnormal signal in the path is calculated, quantifying the degree of fund anomaly diffusion. Subsequently, by calculating the response delay between adjacent nodes and determining the time decay coefficient, the propagation intensity is time-corrected, making the propagation analysis more consistent with actual delay characteristics. Finally, topology updates are performed on high-intensity paths, adjusting weights and directions, thereby making the network structure adaptive in the time dimension. This process achieves dynamic maintenance of the fund flow topology, enabling timely reflection of interactive changes between multi-dimensional factors and providing an accurate propagation basis for subsequent risk identification.

[0033] In a preferred embodiment of the present invention, based on path candidate information, and combining the change magnitude of abnormal nodes, path weights, and the response magnitude of target nodes, the propagation intensity of abnormal signals on each path is calculated, specifically including: First, anomalous nodes are identified in the topology, i.e., nodes whose rate of change in capital flow exceeds a preset threshold, and the upstream and downstream paths connected to these nodes are determined. Then, the magnitude of change at the anomalous nodes is extracted to measure the initial impact of the anomalous event. Next, path weights are obtained from the path candidate information; these weights reflect the strength of historical capital transmission relationships between nodes. Simultaneously, the response magnitude of the target node is extracted from its monitoring data, representing the degree of reaction of downstream nodes after being affected by upstream changes. The magnitude of change at the anomalous node, the path weight, and the response magnitude of the target node are weighted and fused, with paths having higher weights having a greater impact in the propagation calculation. By comparing the fusion results on different paths, the transmission capability of the anomalous signal on different paths can be quantified. The final output transmission strength data reflects the propagation range and impact of the anomalous signal in the capital flow-related topology, providing fundamental data support for subsequent time-lag analysis and risk identification.

[0034] In a preferred embodiment of the present invention, the propagation intensity of each node is corrected for time decay based on the time decay coefficient to form path propagation analysis information, specifically including: First, the correlation between the path propagation strength data and time decay coefficients obtained in the previous steps is extracted to ensure that the strength data of each path matches its response delay. Second, for paths with long response delays, their propagation strength is gradually reduced according to the time decay coefficient to reflect the characteristic that capital transmission in actual economic activities gradually weakens over time. For example, when the capital response between nodes on a path lags significantly, the effective influence of that path is reduced proportionally to time. For paths with short delays and synchronous changes, a higher strength value is maintained. Subsequently, the strength data of each path after decay correction is re-normalized to ensure that the strength values ​​of different paths are on the same order of magnitude for horizontal comparison. Finally, the corrected propagation strength results are output as path propagation analysis information to determine which paths still maintain high propagation efficiency under the influence of time factors, thus providing a decision-making basis for the dynamic updating of the topology.

[0035] In a preferred embodiment of the present invention, topology updates are performed on paths whose propagation intensity in the path propagation analysis information exceeds a preset first propagation intensity threshold. This involves adjusting the path weights and directions to form an updated topology, specifically including: First, based on path propagation analysis information, the propagation strength value of each path is compared with a preset first propagation strength threshold. Paths with strengths above the threshold are marked as active paths, and those below the threshold are marked as decaying paths. Then, the weights of active paths are increased to give them greater influence in subsequent propagation analyses; for decaying paths, their weights are appropriately reduced or the connection is removed after multiple occurrences of low intensity, to simplify the network structure. Second, the directional consistency of paths is analyzed. When changes in upstream nodes continuously affect downstream nodes, but the downstream nodes' reverse response is weak, the path direction remains unchanged; if a bidirectional response is detected, a bidirectional path identifier is established, forming a feedback structure to capture cyclical capital flows. Finally, the updated path weights and directional information are re-aggregated to generate new topology data, enabling the network structure to dynamically reflect the real-time relationships between various dimensions, providing a structural foundation for continuous capital flow propagation monitoring.

[0036] In a preferred embodiment of the present invention, the preset fluctuation threshold is used to determine whether the changes in the node's capital flow meet the conditions for triggering abnormal propagation analysis.

[0037] Specifically, this threshold can be set based on the distribution of historical capital flow volatility, for example, by taking the average volatility of similar nodes over a recent period plus a certain number of standard deviations. When the real-time rate of change of a node exceeds this threshold, it is determined that the node has abnormal volatility, thereby triggering subsequent path extraction and propagation calculation steps.

[0038] The technical purpose of this design is to distinguish between normal economic fluctuations and abnormal event shocks, prevent the model from frequently reporting false alarms due to minor changes, and ensure high sensitivity in capturing sudden risks.

[0039] In a preferred embodiment of the present invention, the preset first transmission strength threshold is used to filter important capital flow transmission paths in the path propagation analysis information.

[0040] This threshold is set based on the path strength distribution in model training or historical samples, and is generally taken as a value above the average level. When the path propagation strength exceeds this threshold, the path is considered to have a real impact on fund propagation and should be included in the topology update process.

[0041] Its function is to eliminate low-intensity, highly random noise paths, thereby improving the computational efficiency and economic interpretability of topology updates.

[0042] In a preferred embodiment of the present invention, based on dynamic propagation data, highly sensitive paths are detected and risk transmission sources are identified, abnormal chains are traced and located, and fund flow risk trend prediction data is generated, including: Based on dynamic propagation data, paths with conduction strength higher than a preset second conduction strength threshold are selected to form a set of highly sensitive paths; Based on the temporal continuity and directional consistency in the set of highly sensitive paths, identify chain nodes that continuously trigger anomalies and form risk chain information; Perform reverse tracing on the risk chain information to identify the earliest source node where the anomaly occurred and generate risk transmission source location information; Based on the location information of the risk transmission source, analyze the rate and scope of the spread of funds from the source node to the downstream node to form risk diffusion information; By fitting risk diffusion information with historical time series trends, data predicting future cash flow risk trends can be generated.

[0043] In this embodiment of the invention, the identification of abnormal capital flow chains and the tracing of risk transmission paths are achieved through the analysis of dynamic propagation data. First, paths with transmission intensity exceeding a threshold are screened, allowing for rapid focusing on high-impact areas within complex topologies and reducing ineffective computation. Second, by identifying the temporal continuity and directional consistency of highly sensitive paths, continuous chains of abnormal capital flow propagation can be discovered, revealing the actual transmission process of risk across multiple dimensions. Furthermore, by tracing these chains in reverse, the earliest source node exhibiting anomalies is identified, thus pinpointing the starting point of the risk and facilitating intervention or verification of the source event. Subsequently, analysis of the diffusion rate and impact range of the source node quantitatively describes the transmission speed and attenuation characteristics of risk between different nodes. Finally, fitting the risk diffusion information with historical time series trends allows for prediction of the future direction and intensity of capital flow risk changes. This process achieves a closed loop from risk detection and path identification to trend prediction, quantifying the evolutionary patterns of capital flow risk and providing an operational basis for predictive decision-making.

[0044] In a preferred embodiment of the present invention, based on the risk transmission source location information, the rate and scope of the spread of funds from the source node to the downstream node are analyzed to form risk diffusion information, specifically including: First, the specific location of the source node and its adjacent downstream nodes are determined from the risk transmission source location information. Then, the response time difference and amplitude changes of the source node's capital flow to downstream nodes over a continuous time period are monitored to calculate the diffusion rate. The diffusion rate reflects how quickly the risk signal propagates in the topology: if the capital changes of downstream nodes are synchronized with the source node in a short time, the diffusion rate of that path is fast; conversely, it is a lagging diffusion path. Next, by comparing the response levels of multiple downstream nodes, the propagation range of the risk signal is determined, i.e., the number of nodes affected by the source node and their distribution hierarchy are identified. Finally, the diffusion rate and impact range data are combined to generate risk diffusion information, which describes the dynamic process of risk transmission from the source node to downstream nodes. This information helps management identify potential risk concentration areas and diffusion directions, providing data support for subsequent risk control and liquidity intervention.

[0045] In a preferred embodiment of the present invention, a preset second conduction strength threshold is used to identify highly sensitive paths in dynamically propagated data.

[0046] This threshold is typically set at a higher quantile of the first threshold, such as the top 10% or 20% intensity range. Paths above this threshold indicate that the risk is spreading rapidly and has a far-reaching impact within that channel.

[0047] With this setting, the model can focus on key risk channels and accurately identify systemic risk chains without having to traverse redundant paths throughout the entire network.

[0048] In a preferred embodiment of the present invention, volatility indicators, cyclical change rate data, and cross-correlation data are weighted and fused according to dimensional importance to generate multi-dimensional influencing factor data, including: Based on volatility indicators, periodic change rate data, and cross-correlation data from various dimensions, initial fusion input data reflecting time series characteristics from different dimensions are extracted. Sensitivity assessment is performed on the initial fusion input data, and the initial weighting coefficients are determined based on the fluctuation activity of each dimension in the historical time series and the current intensity of change, thereby generating weighted input data; Based on the weighted input data, the weight coefficients are dynamically adjusted. When the volatility or cycle change pattern of any dimension increases, its corresponding weight is increased, and dynamic weight adjustment data is generated. Based on the dynamically weighted data, redundancy detection and correction are performed on the cross-correlation data to reduce the redundant contributions between highly correlated dimensions and generate redundant correction data. The redundant correction data is synchronized and merged in time to generate multi-dimensional influence factor vector data. Normalization and smoothing processes are performed on the multi-dimensional impact factor vector data to eliminate the differences in feature scales between dimensions and generate multi-dimensional impact factor data.

[0049] In this embodiment of the invention, a dynamic weighting and redundancy correction mechanism is introduced during the generation of multi-dimensional influence factors, making the construction of influence factors more adaptive and accurate. First, initial input data is formed by extracting the volatility, periodicity, and cross-correlation of each dimension, ensuring the preservation of various dynamic features. Second, initial weight coefficients are determined through sensitivity assessment, making the weight allocation more consistent with the actual influence of each dimension in historical fluctuations and current changes. Subsequently, by dynamically adjusting the weights, when the change pattern of a certain dimension intensifies, its weight automatically increases, enabling flexible response capabilities when data fluctuations are significant. Simultaneously, redundancy detection and correction effectively suppress the duplicate contributions of highly correlated dimensions, maintaining the independence of the fused data. After time synchronization and normalization processing, the generated multi-dimensional influence factor data not only maintains feature stability but also dynamically reflects the temporal change trends of each dimension, thus providing more representative data input for subsequent topological mapping.

[0050] In a preferred embodiment of the present invention, sensitivity evaluation is performed on the initial fused input data, and initial weighting coefficients are determined based on the fluctuation activity of each dimension in historical time series and the current intensity of change, thereby generating weighted input data, specifically including: First, volatility analysis is performed on the time series data for each dimension. By calculating the average amplitude of data changes over consecutive time periods, the activity level of that dimension in historical periods is determined. Higher volatility indicates a stronger dynamic influence of that dimension on changes in capital flows. Second, the rate and direction of change within the current period are extracted to measure the sensitivity of that dimension to short-term changes. Subsequently, historical volatility activity is weighted and combined with the current intensity of change, and its weight range is determined using empirical thresholds or training samples. Dimensions with higher weights represent a greater contribution to capital flow trends. Finally, the determined weight coefficients are applied to the initial fusion input data to weight the data of each dimension, forming weighted input data, which provides a basis for subsequent dynamic weight adjustments.

[0051] In a preferred embodiment of the present invention, the weighting coefficients are dynamically adjusted based on the weighted input data. When the volatility or cyclical change pattern of any dimension increases, its corresponding weight is increased, and dynamic weight adjustment data is output. Specifically, this includes: First, a continuous monitoring window is set on the time series of the weighted input data to detect the changing trends of volatility and cyclical patterns in each dimension in real time. When a continuous increase in the volatility of a certain dimension or an increasing trend in cyclical changes is detected, that dimension is marked as a high-dynamic dimension. Next, the weights are reallocated based on the marking results, appropriately increasing the weight values ​​of high-dynamic dimensions while decreasing the weights of low-dynamic dimensions to maintain overall weight balance. Subsequently, the adjusted weights are smoothed over the time dimension to maintain the continuity of the change process and avoid instability in the calculation results caused by abrupt weight changes. Finally, the dynamically adjusted weight data is output, reflecting the changes in the importance of different dimensions in the real-time economic environment, providing support for adaptive feature fusion.

[0052] In a preferred embodiment of the present invention, redundancy detection and correction are performed on the cross-correlation data based on the dynamic weight adjustment data to reduce redundant contributions between highly correlated dimensions and generate redundancy correction data, specifically including: First, based on dynamically weighted data, the dimension group with higher weights at the current moment is selected, and the correlation between each dimension is calculated. When the correlation between two or more dimensions exceeds a preset correlation threshold, information redundancy is identified. Then, the dominant dimension identification method is used to determine the dimension that contributes the most to the capital flow characteristics, and this dimension is retained as the dominant dimension; the feature contribution values ​​of secondary dimensions are reduced proportionally to weaken their repetitive influence. Next, the weights of all dimensions are renormalized to ensure that the total weight of the data remains consistent. This process effectively avoids the over-amplification of similar information during the fusion process, making the generated multi-dimensional data more reflective of independent characteristics. Finally, redundancy correction data is output, ensuring the reliability of multi-dimensional feature fusion.

[0053] In a preferred embodiment of the present invention, the redundancy correction data is time-synchronized and fused to generate multi-dimensional influence factor vector data, specifically including: First, the data from each dimension, after redundancy correction, are realigned according to time labels to ensure that different dimensions correspond to the same state of capital flow change at the same time point. Then, the aligned data is merged and summarized by calculating a weighted average or synthesizing based on trend direction within a time window, integrating the multi-dimensional data into a single vector structure. This fusion process comprehensively considers the current weights and trends of each dimension, ensuring that important dimensions are more fully reflected in the fusion result. Next, the fused vector is smoothed to eliminate local distortions caused by abnormal fluctuations. The final multi-dimensional influence factor vector data simultaneously reflects time trends, dimensional characteristics, and dynamic weight adjustment results, providing an accurate input basis for subsequent normalization and topology analysis.

[0054] In a preferred embodiment of the present invention, a temporal coupling mapping matrix is ​​established based on interaction strength information and temporal dependency coefficients, and low-relevance paths are eliminated through a weighted filtering mechanism to generate capital flow correlation topology data, including: Based on the interaction strength information and time-series dependency coefficient, the dependency coefficient and interaction strength are hierarchically mapped according to different attribute categories of macro indicators, industry capital trends and corporate behavior, generating multi-level mapping data; Based on the multi-level mapping data, the main transmission weight between the macro and industry levels, and the secondary transmission weight between the industry and enterprise levels are calculated to form a multi-level weight mapping matrix data. Based on the multi-layer weighted mapping matrix data, the temporal stability of each path is evaluated within a preset time window to generate path stability data. Based on path stability data, the path weights are dynamically adjusted to form a dynamic coupling mapping matrix. Weighted filtering is performed on the dynamic coupling mapping matrix data to remove low-relevance paths and optimize directional consistency, generating capital flow related topology data.

[0055] In this embodiment of the invention, by introducing a hierarchical mapping and dynamic filtering mechanism in the process of modeling the correlation between capital flows, the structural expression of the time-series coupling matrix is ​​made more consistent with the logic of economic hierarchy. First, based on the different characteristics of the macro, industry, and enterprise levels, the interaction strength and time-series dependency coefficient are mapped hierarchically, which can clearly distinguish the transmission effects between different levels. Second, by calculating the main transmission weight and the secondary transmission weight, the master-slave relationship of capital flows is established, revealing the guiding role of macro variables on downstream capital behavior. Furthermore, the stability of the path is evaluated within a preset time window, and the weights are dynamically adjusted according to their change characteristics to maintain the timeliness and accuracy of the topology. Finally, through weight filtering and direction optimization, low-correlation or unstable paths are removed, and dominant connections with high influence and consistent direction are retained. This process makes the hierarchical structure of the capital flow topology clearer and the transmission path more reliable, providing a stable basic topological framework for subsequent anomaly propagation calculations.

[0056] In a preferred embodiment of the present invention, based on the interaction strength information and the time-series dependency coefficient, and according to different attribute categories such as macroeconomic indicators, industry capital trends, and corporate behavior, a hierarchical mapping of the dependency coefficient and the interaction strength is performed to generate multi-level mapping data, specifically including: First, the time-series dependency coefficients and interaction strength information generated in the preceding steps are classified by attribute, categorizing them into different levels: macro level (e.g., monetary policy, interest rates, inflation rates), industry level (e.g., net capital flow in the industry, financing scale, accounts receivable period in the industrial chain), and enterprise level (e.g., transaction activity, cash flow stability). Second, within each level, the dependency coefficients between dimensions are matrixed, providing a structured expression of the interaction relationships between factors within the same level. Subsequently, the dependency information between different levels is mapped, ensuring that changes in the macro level correspond to fluctuations in the industry level, and changes in the industry level are further transmitted to the enterprise level. This mapping process, through setting the inter-level dependency ratios, ensures that the influence strength between different levels conforms to the actual economic transmission logic. The final multi-level mapping data output reflects the capital flow transmission framework from macro to micro, providing a structural foundation for subsequent transmission weight calculations.

[0057] In a preferred embodiment of the present invention, based on multi-level mapping data, the primary transmission weight between the macro and industry levels, and the secondary transmission weight between the industry and enterprise levels, are calculated to form multi-level weight mapping matrix data, specifically including: First, dependency data between the macro and industry levels is extracted to calculate the strength of the impact of macroeconomic factors on changes in industry funding. For example, when a decrease in policy interest rates leads to an increase in industry financing activity, the primary transmission weight of this path will rise. Second, the dependency relationship between the industry and enterprise levels is analyzed to identify the transmission effect of changes in industry funding on enterprise liquidity indicators, which is defined as the secondary transmission weight. Subsequently, a normalization method is used to ensure that the sum of the primary and secondary weights is within a comparable range, maintaining a relative balance between different levels. Finally, the transmission weights between each level are stored in matrix form, forming a multi-layer weight mapping matrix data, providing a quantitative structural basis for capital flow topology modeling.

[0058] In a preferred embodiment of the present invention, the temporal stability of each path is evaluated within a preset time window based on multi-layer weight mapping matrix data to generate path stability data, specifically including: First, the multi-layer weight mapping matrix is ​​updated within multiple consecutive time windows, each representing an independent analysis period. The stability of a path is determined by comparing the magnitude of weight changes for the same path in adjacent time windows. When the path weight changes are small and in a consistent direction, it indicates high stability over time; otherwise, it is considered a short-term anomalous transmission path. Second, the stability results for all paths are statistically analyzed, calculating the average stability level of each path over several past periods. Subsequently, weight smoothing is performed on paths with low stability to eliminate fluctuations caused by short-term disturbances. The final output path stability data accurately reflects the temporal reliability of transmission paths at each level, providing a basis for dynamic weight adjustment.

[0059] In a preferred embodiment of the present invention, the path weights are dynamically adjusted based on path stability data to form dynamically coupled mapping matrix data, specifically including: First, path stability data is matched with the current multi-layer weight mapping matrix. When a path exhibits high stability over time, its weight is increased to enhance its influence in the capital flow transmission network; conversely, for paths with low stability or fluctuating directions, their weights are appropriately reduced to minimize their interference with the overall model. Second, to avoid topological instability caused by abrupt weight changes, a sliding weighting method is used for smooth adjustment, that is, the path weights are gradually updated based on the trend of multi-layer weight changes. Finally, the updated weights of all paths are reintegrated into a new dynamic coupling mapping matrix, enabling the topological structure to adapt to changes over time, thus better reflecting the dynamic transmission characteristics of capital flows under different economic environments.

[0060] In a preferred embodiment of the present invention, weight filtering is performed on the dynamic coupling mapping matrix data to eliminate low-relevance paths and optimize directional consistency, generating capital flow association topology data, specifically including: First, based on the weights in the dynamic coupling mapping matrix, all paths are sorted by relevance, and a threshold is set to filter out low-weight paths. This process reduces noise or weak dependencies, making the network structure more representative. Next, directional consistency checks are performed on the remaining paths. By comparing the changing trends of upstream and downstream nodes, it is determined whether the path direction conforms to actual economic logic. When a reverse path is detected, its direction is recalibrated or invalid connections are deleted. Then, normalization is performed on the filtered paths to maintain a balance in the topology weights within the overall structure. The final output of the capital flow correlation topology data clearly reflects the multi-level transmission relationships between macro, industry, and enterprise levels, providing a stable and interpretable network foundation for subsequent dynamic propagation analysis and risk identification.

[0061] In a preferred embodiment of the present invention, the transmission intensity of the abnormal signal on each path is calculated based on path candidate information, combined with the change amplitude of the abnormal node, the path weight, and the response amplitude of the target node, including: Based on the path candidate information, obtain the state change data of the starting node and the target node on the path to form the path state input data; Based on the path status input data, calculate the change amplitude of the starting node to reflect the initial energy of the abnormal signal, and generate change amplitude data; Based on the path topology, extract the corresponding path weights to represent the basic transmission relationships between nodes, and generate path weight data. Based on the rate of change of the state of the target node, determine the response magnitude of the target node and generate target response data; The change magnitude data, path weight data, and target response data are weighted and fused together to generate path transmission strength data.

[0062] In this embodiment of the invention, the propagation capability of abnormal signals is quantitatively characterized by calculating path transmission strength during anomaly propagation analysis. First, by obtaining state change data of the starting and target nodes from path candidate information, the input and output states of the abnormal event in the time dimension can be accurately reflected. Second, by calculating the change amplitude of the starting node and combining it with path weights and the response amplitude of the target node, the propagation capability of the abnormal signal on different paths can be comprehensively reflected. The multi-factor weighted fusion mechanism ensures that the signal strength considers both the impact scale of the anomaly source and the absorption capacity of the target node, giving the calculation results a hierarchical and directional quality. The final generated path transmission strength data can serve as an indicator of anomaly propagation efficiency for subsequent time-delay correction and topology updates. This process enables the identification of high-intensity propagation paths under abnormal conditions, thereby more accurately characterizing the propagation mechanism of financial anomalies in multidimensional networks and providing reliable basic data support for subsequent dynamic risk assessment.

[0063] In a preferred embodiment of the present invention, based on the path status input data, the change amplitude of the starting node is calculated to reflect the initial energy of the abnormal signal, and change amplitude data is generated, specifically including: First, the changes in capital flow at the starting node over consecutive time points are extracted from the path status input data. The actual intensity of capital fluctuations is determined by calculating the magnitude of these increases or decreases within the analysis period. Second, it is analyzed whether the change belongs to normal cyclical fluctuations or sudden abnormal fluctuations, based on historical average change rates or the normal fluctuation range of the industry. When the change magnitude exceeds a preset change magnitude threshold, the node is marked as an abnormal source node. Subsequently, different influence weights are assigned to different types of abnormal signals (e.g., excessively rapid inflow, excessively rapid outflow, or violent fluctuations) to distinguish the nature of abnormal capital flows. Finally, the direction of change, fluctuation magnitude, and weights are combined to form change magnitude data, which is used to measure the energy intensity of abnormal signals in the initial stage of propagation, providing quantitative input for subsequent path propagation calculations.

[0064] In a preferred embodiment of the present invention, path weights are extracted based on the path topology to represent the basic transmission relationships between nodes, and path weight data is generated, specifically including: First, the structural information of the current path is read from the capital flow correlation topology data to determine the connection strength between the starting node and the target node. This connection strength can be derived from indicators such as the transaction frequency, capital flow volume, or interdependence between the two nodes in historical data. Second, to avoid abnormally high path weights caused by a single abnormal transaction, the original connection strength is smoothed to reflect long-term stable capital relationships. Then, the weight values ​​are adjusted according to the direction of capital flow within the time window: if the path direction is consistent with the current abnormal signal propagation direction, the path weight is appropriately increased; if not, its influence is reduced. Finally, the adjusted weight results are output as path weight data, which is used for fusion calculation with the node change amplitude and the target node response amplitude to ensure that the path propagation strength assessment more accurately reflects the characteristics of actual economic activities.

[0065] In a preferred embodiment of the present invention, the response amplitude of the target node is determined based on the state change rate of the target node, and target response data is generated, specifically including: First, the fund flow data of the target node is monitored in real time during the anomaly propagation analysis period, recording its rate of change and direction. When significant fluctuations occur in upstream nodes, the proportion of fund flow changes in the target node in subsequent time periods is calculated to reflect its responsiveness. Second, to avoid misjudgments due to short-term fluctuations, the rate of change of the target node is smoothed using a moving average or trend fitting method, ensuring the results reflect its stable response level to upstream changes. Subsequently, the response intensity and direction are categorized: when the direction of change of the target node is consistent with that of the upstream node, it indicates positive propagation; if the direction is opposite, it indicates a reverse offsetting effect. Finally, the average response rate and directional characteristics of the target node in each time period are summarized to generate target response data, providing input parameters for subsequent multi-factor fusion calculations of anomaly signals.

[0066] In a preferred embodiment of the present invention, the change magnitude data, path weight data, and target response data are weighted and fused using multiple factors to generate path transmission strength data, specifically including: First, the three types of input data are standardized to ensure comparability on the same numerical scale. Then, fusion weights are assigned based on the importance of each data type; for example, amplitude data reflects the initial strength of the anomalous signal, path weight data reflects the accessibility of the transmission channel, and target response data represents the intensity of the response after the signal arrives. Next, the three indicators are fused using weighted summation or hierarchical weighting methods to ensure the final result considers signal source energy, path transmission capability, and target response sensitivity. Subsequently, to avoid excessive influence from a single extreme value, truncated averaging or quantile adjustment methods are used to smooth the results. The final output path transmission strength data reflects the comprehensive propagation effect of the anomalous signal along the path, providing a precise basis for subsequent time-delay correction and dynamic topology updates.

[0067] In a preferred embodiment of the present invention, based on path candidate information, the time difference between the state changes of adjacent nodes on the path is compared to obtain the response delay, and a time decay coefficient is determined based on the ratio of the response delay to a preset time window, including: Based on the path candidate information, extract the time points of state changes of adjacent nodes on the path to form node time series data; Based on the node time series data, calculate the time interval between state changes between adjacent nodes and generate node response delay data; Based on the node response delay data, compare it with the preset time window to determine the delay ratio and generate time delay ratio data; Based on the time delay ratio data, the degree of time attenuation of the transmitted signal along the path is evaluated, the time attenuation coefficient is calculated, and the time attenuation coefficient data is generated.

[0068] In this embodiment of the invention, a time decay correction mechanism is established by analyzing the response delay between nodes in the path, enabling the abnormal propagation model of capital flow to dynamically adjust in the time dimension. First, by extracting the state change time points of adjacent nodes on the path and calculating the time interval, the propagation delay of the signal in the path can be accurately quantified. Second, by comparing the delay value with a preset time window to obtain the delay ratio, the relative lag of the path propagation can be determined. Further, by evaluating the degree of signal strength attenuation caused by the delay ratio, a time decay coefficient is calculated, achieving time correction of the propagation intensity. This correction process allows for a reasonable distinction of the propagation effects of different paths under time differences, avoiding errors caused by the synchronicity assumption. Finally, by inputting the corrected propagation intensity data into the topology update stage, the path weights and directions can be dynamically adjusted, allowing the topology structure to automatically optimize over time. This mechanism improves the model's adaptability to the timeliness of capital flow propagation, ensuring that the simulation results of abnormal propagation more closely match the actual capital flow response characteristics.

[0069] In a preferred embodiment of the present invention, the time attenuation of the transmitted signal along the path is evaluated based on the time delay ratio data, and a time attenuation coefficient is calculated to form time attenuation coefficient data, specifically including: First, the response delay ratios between nodes on each path are extracted from the time delay ratio data obtained in the previous steps, and the delay changes of the paths within different time windows are compared. Second, for paths with large response delays, it is determined that the transmitted signal of this path exhibits a significant attenuation trend over time, and the delay ratio is combined with the actual response amplitude to assess the degree of signal strength weakening during propagation. For example, when the time delay of a path exceeds the average lag period, its signal strength decreases proportionally to reflect the gradual weakening of transmitted energy. Subsequently, the attenuation degree of different paths is classified, and paths with lighter and heavier attenuation are marked separately, and the corresponding time attenuation coefficients are calculated. This coefficient represents the proportion of effective strength retained when the signal propagates on that path. Finally, the attenuation coefficients of all paths are summarized and structured to form time attenuation coefficient data, providing a basis for subsequent propagation strength correction and topology updates, enabling the model to dynamically reflect the actual attenuation characteristics of the capital flow signal in the time dimension.

[0070] In a preferred embodiment of the present invention, risk diffusion information is fitted with historical time series trends to generate future time period capital flow risk trend prediction data, including: Based on risk diffusion information, risk propagation rate data of the risk source node and its downstream nodes are extracted to generate risk propagation input data; Based on the risk propagation input data, the trend of capital flow changes at each node in the historical time series is extracted to generate historical trend data; By aligning the risk transmission input data with historical trend data in terms of time and standardizing the magnitude, comparable fitting data can be formed. Based on comparable fitting data, perform multi-period trend comparison within a preset time window, calculate the degree of deviation between short-term abnormal propagation trend and long-term stable trend, and generate trend deviation data. Based on trend deviation data, predict the direction and intensity of risk propagation in the future time period, and generate capital flow risk trend prediction data.

[0071] In this embodiment of the invention, dynamic prediction of future cash flow risk trends is achieved by fitting risk diffusion information with historical time series trends. First, by extracting risk propagation rate data from risk transmission sources and downstream nodes, a propagation path model of risk among multiple nodes can be established. Second, by aligning these propagation data with historical cash flow trends in terms of time and magnitude, the changing patterns within different time periods can be reflected on a unified scale. Furthermore, by performing trend comparisons within multi-period windows and calculating the degree of deviation between short-term anomalies and long-term stable trends, potential risk shift signals can be captured. Based on the trend deviation results, the future propagation direction and intensity of risk can be predicted, allowing for early identification of the potential scope and timing of cash flow anomalies. The resulting risk trend prediction data can provide decision-making support for management entities, enabling the formulation of fund allocation or risk control strategies. This process integrates risk identification and trend prediction, transforming dynamic cash flow analysis from passive monitoring to proactive prediction.

[0072] In a preferred embodiment of the present invention, based on comparable fitting data, a multi-period trend comparison is performed within a preset time window to calculate the degree of deviation between the short-term abnormal propagation trend and the long-term stable trend, generating trend deviation data, specifically including: First, the corresponding portions of the risk propagation input sequence and historical time series are extracted from comparable fitted data, and the two are time-aligned to allow direct comparison of cash flow change curves at the same time scale. Second, within a preset analysis time window (e.g., one week, one month, or one quarter), the direction, slope, and volatility of short-term risk propagation trends are compared with long-term stable trends. When the direction of change of the short-term trend is inconsistent with the long-term trend or the volatility deviates significantly, a trend deviation is identified in that segment. Subsequently, the duration and deviation magnitude of short-term abnormal trends are statistically analyzed to assess the degree of difference between them and the long-term trend. The greater the deviation, the stronger the impact of the risk signal on the overall stability of cash flow. Finally, these deviation magnitude and duration parameters are structured and output to generate trend deviation data for subsequent risk direction and intensity prediction. This process can quantitatively characterize the difference between abnormal trends and steady-state trends, providing quantitative indicators for predicting future changes in cash flow risk.

[0073] In a preferred embodiment of the present invention, based on trend deviation data, the direction and intensity of risk propagation in a future time period are predicted to generate cash flow risk trend prediction data, specifically including: First, the deviation direction, duration, and magnitude of the deviation data are analyzed to comprehensively assess the overall trend of current cash flow risk changes. When the deviation direction is opposite to the long-term trend, the risk is considered to be in a diffusion phase; if the deviation magnitude gradually decreases, it indicates that the risk is converging. Second, based on the subsequent change patterns of similar deviation scenarios in historical trends, the direction of risk propagation in the future time period is assessed, i.e., determining which funding channels or nodes the risk signal will spread to. Subsequently, by comparing the ratio of historical deviation magnitudes to current deviation magnitudes, the intensity trend of the risk signal in the future stage is estimated. Increased intensity indicates an expansion of potential risk, while decreased intensity indicates risk mitigation. Finally, the predicted propagation direction, intensity, and possible duration are integrated to generate cash flow risk trend prediction data, which is used to output dynamic risk warning information for the future time period. This process achieves a logical closed loop from trend deviation identification to future risk trend prediction, enabling the model to have a forward-looking analytical capability for abnormal cash flow evolution.

[0074] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-dimensional dynamic prediction and analysis method for cash flow, characterized in that, The method includes: Obtain historical transaction data, industry economic indicators, policy event information, supply chain settlement records, and market sentiment data of the target account group as raw data; The original data is time-stamped and delayed, and asynchronous data with different time granularities are aligned to generate time series matrix data. Based on time series matrix data, the volatility, periodic change rate and cross correlation of each dimension of data are adaptively calculated to extract the influencing factors that can reflect the dynamic characteristics of multiple dimensions and generate multi-dimensional influencing factor data. Based on multi-dimensional influencing factor data, the temporal dependence coefficient and interaction strength between each dimension are calculated, a temporal coupling mapping matrix is ​​established, and capital flow correlation topology data is generated. Based on the capital flow correlation topology data, when any dimension of data shows abnormal fluctuations, its transmission strength along the topology path is calculated, and the time lag effect is evaluated based on the response delay between nodes to obtain the time decay coefficient of the transmission strength in the corrected topology path, so as to dynamically adjust the topology structure and generate dynamic propagation data. Based on dynamic propagation data, highly sensitive paths are detected and risk transmission sources are identified. Abnormal chains are traced and located, and fund flow risk trend prediction data is generated. This data is then used to output dynamic prediction results and risk warning signals for the future fund flows of the target account group.

2. The multi-dimensional dynamic prediction and analysis method for capital flows according to claim 1, characterized in that, Based on time series matrix data, the volatility, periodicity, and cross-correlation of each dimension of the data are adaptively calculated to extract influencing factors that reflect multi-dimensional dynamic characteristics, generating multi-dimensional influencing factor data, including: Data smoothing and anomaly removal are performed on the time series matrix data to obtain time series feature data; Based on time series characteristic data, the magnitude of change in adjacent time periods is calculated within a time sliding window to determine volatility indicators for each dimension. Based on volatility indicators, pattern recognition is performed on the trend direction of each dimension within multiple period windows to generate periodic change rate data. By cross-matching periodic rate of change data with sequences of different time dimensions, the degree of response under time shift is analyzed to obtain cross-dimensional cross-correlation data. Volatility indicators, cyclical change rate data, and cross-correlation data are weighted and integrated according to their dimensional importance to generate multi-dimensional influencing factor data.

3. The multi-dimensional dynamic prediction and analysis method for capital flows according to claim 1, characterized in that, Based on multi-dimensional influencing factor data, the temporal dependency coefficients and interaction strengths between each dimension are calculated, a temporal coupling mapping matrix is ​​established, and capital flow correlation topology data is generated, including: Based on multi-dimensional influencing factor data, the differences in the direction and magnitude of change of each dimension between adjacent time points are calculated to obtain basic information on time-series changes. Based on the basic information of time series changes, calculate the response strength of each dimension to changes in other dimensions, and determine the time series dependency coefficient under time delay; The temporal dependency coefficients are aggregated according to the dimensional pairing method, and the duration and direction consistency of the interaction between each dimension are calculated to obtain the interaction strength information; Based on the interaction strength information and the temporal dependency coefficient, a temporal coupling mapping matrix is ​​established, and low-relevance paths are eliminated through a weight screening mechanism to generate capital flow correlation topology data.

4. The multi-dimensional dynamic prediction and analysis method for capital flows according to claim 1, characterized in that, Based on the capital flow correlation topology data, when any dimension of data shows abnormal fluctuations, its transmission strength along the topology path is calculated, and the time lag effect is evaluated based on the response delay between nodes to obtain the time decay coefficient of the transmission strength in the corrected topology path. This allows for dynamic adjustment of the topology structure and generation of dynamic propagation data, including: Based on the topological data of fund flow association, monitor the real-time change rate of the corresponding dimension of each node to form node status monitoring information; Based on node status monitoring information, when the change of any node exceeds the preset fluctuation threshold, anomaly propagation analysis is triggered to extract the adjacent paths of the abnormal node and form path candidate information. Based on the path candidate information, combined with the change magnitude of the abnormal nodes, the path weights, and the response magnitude of the target nodes, the transmission intensity of the abnormal signal on each path is calculated. Based on the path candidate information, the time difference of the state changes of adjacent nodes on the path is compared to obtain the response delay, and the time decay coefficient is determined according to the ratio of the response delay to the preset time window. Based on the time decay coefficient, the propagation intensity of each node is corrected for time decay to form path propagation analysis information; For paths whose propagation intensity exceeds a preset first propagation intensity threshold in the path propagation analysis information, perform topology updates, adjust path weights and directions, and form an updated topology structure. Dynamic propagation data is generated based on the changing trends between paths in the updated topology.

5. The multi-dimensional dynamic prediction and analysis method for capital flows according to claim 1, characterized in that, Based on dynamic propagation data, highly sensitive paths are detected and risk transmission sources are identified. Abnormal chains are traced and located, generating financial flow risk trend prediction data, including: Based on dynamic propagation data, paths with conduction strength higher than a preset second conduction strength threshold are selected to form a set of highly sensitive paths; Based on the temporal continuity and directional consistency in the set of highly sensitive paths, identify chain nodes that continuously trigger anomalies and form risk chain information; Perform reverse tracing on the risk chain information to identify the earliest source node where the anomaly occurred and generate risk transmission source location information; Based on the location information of the risk transmission source, analyze the rate and scope of the spread of funds from the source node to the downstream node to form risk diffusion information; By fitting risk diffusion information with historical time series trends, data predicting future cash flow risk trends can be generated.

6. The multi-dimensional dynamic prediction and analysis method for capital flows according to claim 2, characterized in that, Volatility indicators, cyclical change rate data, and cross-correlation data are weighted and fused according to dimensional importance to generate multi-dimensional influencing factor data, including: Based on volatility indicators, periodic change rate data, and cross-correlation data from various dimensions, initial fusion input data reflecting time series characteristics from different dimensions are extracted. Sensitivity assessment is performed on the initial fusion input data, and the initial weighting coefficients are determined based on the fluctuation activity of each dimension in the historical time series and the current intensity of change, thereby generating weighted input data; Based on the weighted input data, the weight coefficients are dynamically adjusted. When the volatility or cycle change pattern of any dimension increases, its corresponding weight is increased, and dynamic weight adjustment data is generated. Based on the dynamically weighted data, redundancy detection and correction are performed on the cross-correlation data to reduce the redundant contributions between highly correlated dimensions and generate redundant correction data. The redundant correction data is synchronized and merged in time to generate multi-dimensional influence factor vector data. Normalization and smoothing processes are performed on the multi-dimensional impact factor vector data to eliminate the differences in feature scales between dimensions and generate multi-dimensional impact factor data.

7. The multi-dimensional dynamic prediction and analysis method for capital flows according to claim 3, characterized in that, Based on interaction strength information and temporal dependency coefficients, a temporal coupling mapping matrix is ​​established, and low-relevance paths are eliminated through a weighted screening mechanism to generate capital flow correlation topology data, including: Based on the interaction strength information and time-series dependency coefficient, the dependency coefficient and interaction strength are hierarchically mapped according to different attribute categories of macro indicators, industry capital trends and corporate behavior, generating multi-level mapping data; Based on the multi-level mapping data, the main transmission weight between the macro and industry levels, and the secondary transmission weight between the industry and enterprise levels are calculated to form a multi-level weight mapping matrix data. Based on the multi-layer weighted mapping matrix data, the temporal stability of each path is evaluated within a preset time window to generate path stability data. Based on path stability data, the path weights are dynamically adjusted to form a dynamic coupling mapping matrix. Weighted filtering is performed on the dynamic coupling mapping matrix data to remove low-relevance paths and optimize directional consistency, generating capital flow related topology data.

8. The multi-dimensional dynamic prediction and analysis method for capital flows according to claim 4, characterized in that, Based on the path candidate information, combined with the change magnitude of the abnormal nodes, path weights, and the response magnitude of the target nodes, the propagation intensity of the abnormal signal on each path is calculated, including: Based on the path candidate information, obtain the state change data of the starting node and the target node on the path to form the path state input data; Based on the path status input data, calculate the change amplitude of the starting node to reflect the initial energy of the abnormal signal, and generate change amplitude data; Based on the path topology, the corresponding path weights are extracted to represent the basic transmission relationships between nodes, and path weight data is generated. Based on the rate of change of the state of the target node, determine the response magnitude of the target node and generate target response data; The change magnitude data, path weight data, and target response data are weighted and fused together to generate path transmission strength data.

9. The multi-dimensional dynamic prediction and analysis method for capital flows according to claim 4, characterized in that, Based on the path candidate information, the time difference between the state changes of adjacent nodes on the path is compared to obtain the response delay. Then, a time decay coefficient is determined based on the ratio of the response delay to a preset time window, including: Based on the path candidate information, extract the time points of state changes of adjacent nodes on the path to form node time series data; Based on the node time series data, calculate the time interval between state changes between adjacent nodes and generate node response delay data; Based on the node response delay data, compare it with the preset time window to determine the delay ratio and generate time delay ratio data; Based on the time delay ratio data, the degree of time attenuation of the transmitted signal along the path is evaluated, the time attenuation coefficient is calculated, and the time attenuation coefficient data is generated.

10. The multi-dimensional dynamic prediction and analysis method for capital flows according to claim 5, characterized in that, By fitting risk diffusion information with historical time series trends, predictive data on future cash flow risk trends is generated, including: Based on risk diffusion information, risk propagation rate data of the risk source node and its downstream nodes are extracted to generate risk propagation input data; Based on the risk propagation input data, the trend of capital flow changes at each node in the historical time series is extracted to generate historical trend data; By aligning the risk transmission input data with historical trend data in terms of time and standardizing the magnitude, comparable fitting data can be formed. Based on comparable fitting data, perform multi-period trend comparison within a preset time window, calculate the degree of deviation between short-term abnormal propagation trend and long-term stable trend, and generate trend deviation data. Based on trend deviation data, predict the direction and intensity of risk propagation in the future time period, and generate capital flow risk trend prediction data.

Images