Precious metal transaction system based on multi-source data fusion and intelligent matching engine

Abstract

This invention discloses a precious metals trading system based on multi-source data fusion and an intelligent matching engine, belonging to the field of precious metals trading technology. It solves the technical problems of existing trading systems, such as difficulty in integrating alternative data like on-chain behavior, difficulty in quantifying counterparty default and market disturbance risks during matching, and lack of ability to identify hidden collaborative relationships. The system includes: a multi-source data acquisition and fusion module to generate full-dimensional profile stream data of participants; a dynamic heterogeneous graph construction and update module to calculate the convergence degree of behavior between nodes based on on-chain behavior patterns and use it as graph edge weights; a counterparty intelligent scoring module that uses the order initiator as an anchor point, runs a heterogeneous graph neural network, and reasons along three types of meta-paths: supply chain transmission, related event impact, and hidden collaboration, outputting a risk-adjusted comprehensive matching score; and an intelligent matching and order routing module that sorts and filters according to the score and generates matching decisions, and after a transaction, sends back an updated graph structure to form a closed loop.

Description

Technical Field

[0001] This invention belongs to the field of precious metals trading technology, specifically relating to a precious metals trading system based on multi-source data fusion and an intelligent matching engine. Background Technology

[0002] In over-the-counter precious metals trading and large-scale order matching scenarios, order matching generally adopts an anonymous matching mechanism that prioritizes price and time. Under this matching mechanism, information such as the credit status, position preferences, position in the industry chain, and historical behavior patterns of the counterparty are scattered in different systems and are manually judged by traders before the transaction is completed. This is not only inefficient, but also has a significant lag in risk identification.

[0003] While existing technologies have attempted to incorporate alternative data into trading decisions, they suffer from the following shortcomings: First, it is difficult to semantically integrate unstructured alternative data such as on-chain address behavior and related event texts with structured data such as public holdings and market data. Differences in entity naming, time granularity, and relationship expression among different data sources lead to contradictions after data aggregation, making it unsuitable for direct automatic matching. Second, it is difficult to quantitatively assess the default probability of candidate counterparties and the potential related market disturbances after a transaction at the moment of matching, resulting in a disconnect between risk control and matching. Third, there is a lack of means to identify hidden collaborative relationships among participants, especially for groups of participants with similar on-chain behavior patterns but no obvious correlation in public data; existing systems cannot perceive such risks during the matching process. These deficiencies have long constrained the matching quality of existing precious metal trading systems to a single price dimension, preventing them from achieving comprehensive optimization after risk adjustment.

[0004] Therefore, a precious metals trading system is needed that can semantically fuse multi-source alternative data and quantify counterparty risk and identify hidden collaborative relationships during the matching process. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art. To this end, this invention proposes a precious metals trading system based on multi-source data fusion and intelligent matching engine. This system addresses the technical problems of existing precious metals trading systems, such as the difficulty in semantically fusing alternative data like on-chain address behavior with market data, the difficulty in quantitatively assessing the default probability of candidate counterparties and the associated market disturbances after the transaction at the moment of order matching, and the lack of ability to identify hidden collaborative relationships among participants, resulting in matching decisions relying solely on the price dimension and lagging risk control.

[0006] To address the aforementioned problems, this invention provides a precious metals trading system based on multi-source data fusion and an intelligent matching engine, comprising the following modules: Multi-source data acquisition and fusion module: used to collect multi-source alternative data related to precious metals market participants in real time, and to perform semantic alignment and cleaning of the multi-source alternative data with real-time market data to generate a full-dimensional profile stream of participants; Dynamic heterogeneous graph construction and update module: used to receive the full-dimensional profile stream data of the participants, dynamically construct and update the heterogeneous graph of precious metal market participants with precious metal market participants as nodes and the relationship between participants as edges, and calculate the behavioral convergence degree between nodes based on the on-chain address transaction behavior pattern in multi-source alternative data, and use the calculated behavioral convergence degree as the dynamic weight of the corresponding edge in the heterogeneous graph of precious metal market participants. The counterparty intelligent scoring module is used to run a pre-trained heterogeneous graph neural network model on the heterogeneous graph of the precious metals market participants when an order to be matched is received, with the order initiator node as the anchor point. Through node-level and semantic-level two-layer attention aggregation, it infers the potential default probability of each candidate counterparty node and the related market disturbance behavior that may be triggered after the order is completed along different meta-paths, and finally outputs the risk-adjusted comprehensive matching score of each candidate counterparty. Intelligent matching and order routing module: used to sort and filter candidate counterparties based on the risk-adjusted comprehensive matching score, and automatically generate matching decisions or routing suggestions.

[0007] Preferably, the multi-source data acquisition and fusion module includes: Multi-source alternative data includes publicly available holdings reports, industry chain upstream and downstream relationship data, related event text data, and on-chain address transaction behavior pattern data that has been anonymized; Semantic alignment and cleaning include entity identity alignment, timeline alignment, and semantic relationship alignment, specifically: The entity identity alignment adopts entity disambiguation and coreference resolution technology based on knowledge graph embedding to map different names of the same market participant in different data sources to the same globally unique identifier. The timeline alignment is used to maintain a unified event timeline, normalize the time granularity of data sources with different update frequencies, and adopt a strategy that combines forward padding and time-effect decay weighting for low-frequency data. When a major event related to a participant is detected, an instant refresh of the full-dimensional profile of the participant is immediately triggered. The semantic relationship alignment uses a relationship fusion inference engine to normalize and remove redundancy from multiple relationships between the same pair of participants in different data sources. When it is detected that the frequency of fund transactions between on-chain addresses of participants within a preset observation period is greater than a preset frequency threshold, and the statistical significance of the direction of fund transactions being opposite to the direction of changes in their public holdings is greater than a preset significance threshold, the transaction relationship between the two is strengthened into a suspected wash trading risk relationship, thereby generating structured full-dimensional profile stream data of participants.

[0008] Preferably, the entity disambiguation and coreference resolution technology calculates the similarity of name strings, the consistency of registration location, and the matching degree of holding behavior patterns to output entity alignment confidence. When the entity alignment confidence is greater than a first preset threshold, entities are automatically merged. When the second preset threshold is less than or equal to the entity alignment confidence and less than or equal to the first preset threshold, they are marked as pending manual confirmation.

[0009] Preferably, in the dynamic heterogeneous graph construction and update module, calculating the behavioral convergence degree between nodes based on on-chain address transaction behavior patterns in multi-source alternative data includes: On-chain behavioral feature vectors are extracted from the full-dimensional profile data of each participant. These on-chain behavioral feature vectors are multi-dimensional time-series tensors, with each dimension corresponding to a behavioral indicator in the most recent time series. Statistical values ​​within a time window; After aligning and normalizing the on-chain behavioral feature vectors of any two participant nodes in the same time window, the Pearson correlation coefficient and dynamic time warping distance are calculated for each feature dimension in both the short and long time windows. After weighted fusion, the behavioral similarity scores of the dimension are obtained. Finally, the behavioral similarity scores of all dimensions are combined to obtain the comprehensive behavioral convergence between the two participant nodes. The permutation test method is used to perform a statistical significance test on the overall behavioral convergence. When the actual observed overall behavioral convergence is greater than the preset quantile threshold of the random distribution, the convergence relationship is determined to be significant. A behavioral convergence relationship edge is created or updated in the heterogeneous graph of precious metal market participants, and the value of the overall behavioral convergence is used as the dynamic weight of the behavioral convergence relationship edge.

[0010] Preferably, in the intelligent scoring module for counterparties, a pre-trained heterogeneous graph neural network model is run on the heterogeneous graph of the precious metals market participants. First, anchor point perception initialization is performed, and the node corresponding to the order initiator in the heterogeneous graph is determined as the anchor node. Nodes with the opposite direction to the current order and with trading qualifications are selected to form a candidate counterparty set. The order attributes are encoded into an order attribute vector as additional input and injected into the heterogeneous graph neural network model. When performing node-level attention aggregation in the two-layer attention aggregation of node-level and semantic-level, for each candidate adversary node, its neighbor nodes are divided into transaction relationship neighbor groups, lending relationship neighbor groups, supply chain neighbor groups, co-occurrence of related events neighbor groups, and behavior convergence neighbor groups according to edge type. Within each relationship group, the attention weight of the neighbor nodes to the candidate adversary node is calculated. The attention weight is used to perform weighted aggregation of the neighbor features within each relationship group to obtain the neighborhood aggregation representation from the perspective of each relationship type. In the semantic-level attention aggregation of the node-level and semantic-level two-layer attention aggregation, the neighborhood aggregation representation under each relation type is regarded as a different semantic view. The attention weight is calculated for each semantic view and then weighted and summed to obtain the fusion graph embedding representation of the candidate opponent node.

[0011] Preferably, in the counterparty intelligent scoring module, the potential default probability of each candidate counterparty node and the related market disturbance behavior that may be triggered after the order is completed are inferred along different meta-paths, including: Reasoning is conducted along the risk transmission path of the supply chain, the risk path of related event impacts, and the risk path of hidden collaborative risks. The supply chain risk transmission path sequentially passes through the order initiator node, the candidate counterparty node, the miner node, the refiner node, and the downstream bank node. The associated event impact risk metapath passes sequentially through candidate counterparty nodes, negative event nodes, and associated entity nodes; The hidden collaborative risk element path passes through candidate opponent nodes, high-risk participant nodes, and other participant nodes in sequence. For each meta-path, the feature vectors and edge weights in the full-dimensional profile stream data of the participants corresponding to each node on the path are sampled from the path instance and attention pooled to obtain the corresponding path vector.

[0012] Preferably, the counterparty intelligent scoring module outputs a risk-adjusted comprehensive matching score for each candidate counterparty, including: The path vectors corresponding to the supply chain transmission risk meta-path, the related event impact risk meta-path, and the hidden collaborative risk meta-path are fused through a gating mechanism to obtain a path fusion vector. The path fusion vector is then spliced ​​with the fusion graph embedding representation and input into the default probability prediction head and the related market disturbance prediction head, respectively. The default probability prediction header outputs the default probability of the candidate counterparty under the current order. ; The associated market disturbance prediction head outputs the expected market disturbance index exerted by candidate counterparties on the relevant precious metals within a preset time window after the transaction. The expected market disturbance index is obtained by comprehensively calculating the predicted abnormal fluctuation amplitude, the probability of order book imbalance, and the consistency of price shock direction. Simultaneously, based on the influence centrality measure of the candidate counterparties in the heterogeneous graph of precious metals market participants and the influence coefficient of their historical behavior on the market, the strategy value score of the candidate counterparties is calculated. ; The risk-adjusted composite matching score is calculated as follows: ,in, , and These are the weighting coefficients.

[0013] Preferably, the risk-adjusted comprehensive matching score is used to rank and filter candidate counterparties, including: A hard screening process is conducted on candidate counterparties, filtering out those with a comprehensive matching score ≤ the system's preset minimum threshold, those on the order initiator's prohibited trading list, and those whose tradable amount does not meet the minimum transaction unit. For the candidate competitors who pass the hard screening, a joint ranking function based on the comprehensive score and the quoted price is used for ranking. The joint ranking function is as follows: ,in, The normalized, risk-adjusted composite matching score. The price is the normalized price. This is a risk-price balancing factor, set by the order initiator before the transaction.

[0014] Preferably, the intelligent matching and order routing module includes: When the candidate counterpart with the highest overall matching score cannot fully absorb the total number of orders, the remaining order quantity is re-sorted in the remaining candidate set, the second best candidate counterpart is selected and its available sub-order quantity is allocated, and the process is recursively executed until all orders are allocated or the candidate set is exhausted. The transaction routing instructions are automatically generated based on the matching results of each sub-order. Specifically, for the part that can be directly executed on the electronic platform, the sub-order is automatically routed to the corresponding transaction channel through the FIX protocol; for the off-exchange transaction part, an electronic inquiry form is generated and pushed to the counterparty's terminal through a preset encrypted communication channel. After each transaction is completed, the transaction result is automatically sent back to the dynamic heterogeneous graph construction and update module in the form of a transaction feedback data packet. Based on this, the dynamic heterogeneous graph construction and update module adds a new transaction relationship edge or enhances the weight of an existing transaction relationship edge in the heterogeneous graph of precious metal market participants.

[0015] The beneficial effects of this invention are: This invention employs a three-layer processing approach—entity identity alignment, timeline alignment, and semantic relationship alignment—to integrate publicly available holdings reports, upstream and downstream relationships within the industry chain, related event texts, and desensitized on-chain address transaction behavior patterns with real-time market data into a comprehensive participant profile stream. Specifically, semantic relationship alignment strengthens the ordinary transaction relationship between participants into a suspected wash trading risk relationship when the frequency of fund transfers between participants' on-chain addresses exceeds a preset frequency threshold, and the statistical significance of this fund transfer direction being opposite to the direction of their publicly available holdings changes exceeds a preset significance threshold. This solves the problem of heterogeneous data not being directly aggregated and used, enabling the downstream graph construction module to obtain semantically clear and temporally consistent input data. This invention constructs an intelligent scoring module for counterparties, using the initiator node of the order to be matched as the anchor point. The order attributes are encoded as additional inputs and injected into a heterogeneous graph neural network. Node-level and semantic-level dual-layer attention aggregation is performed to obtain a fusion graph embedding. The module then infers along three meta-paths: supply chain transmission risk, related event impact risk, and hidden collaborative risk. The path vectors and the fusion graph embedding are concatenated and input into the default probability prediction head and the related market disturbance prediction head, respectively. Simultaneously, the strategy value score is calculated by combining influence centrality and historical behavior influence coefficients. The risk-adjusted comprehensive matching score is output according to a weighted formula. This expands the matching criteria from a single price quote to a three-dimensional comprehensive score that includes default risk, market disturbance risk, and strategy value. Moreover, the score is dynamically adjusted according to changes in order attributes, achieving order-level risk control upfront. This invention constructs an intelligent scoring module for counterparties, using the initiator node of the order to be matched as the anchor point. The order attributes are encoded as additional inputs and injected into a heterogeneous graph neural network. Node-level and semantic-level dual-layer attention aggregation is performed to obtain a fusion graph embedding. The module then infers along three meta-paths: supply chain transmission risk, related event impact risk, and hidden collaborative risk. The path vectors and the fusion graph embedding are concatenated and input into the default probability prediction head and the related market disturbance prediction head, respectively. Simultaneously, the strategy value score is calculated by combining influence centrality and historical behavior influence coefficients. The risk-adjusted comprehensive matching score is output according to a weighted formula. This expands the matching criteria from a single price quote to a three-dimensional comprehensive score that includes default risk, market disturbance risk, and strategy value. Moreover, the score is dynamically adjusted according to changes in order attributes, achieving order-level risk control upfront. This invention employs a rigid screening mechanism, followed by a risk-price balance factor to jointly rank the comprehensive score and the quoted price. When the highest-scoring counterparty cannot absorb all orders, the system recursively splits the order and generates routing instructions. After each transaction, the transaction result is sent back to the heterogeneous graph construction module to add or enhance transaction relationship edges. This ensures that each match not only achieves the optimal match at the moment but also feeds back into the graph structure, making subsequent scoring more accurate and forming a continuously self-optimizing closed loop. Attached Figure Description

[0016] Figure 1This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 As shown, this invention is a precious metals trading system based on multi-source data fusion and an intelligent matching engine, comprising the following modules: Multi-source data acquisition and fusion module: used to collect multi-source alternative data related to precious metals market participants in real time, and to perform semantic alignment and cleaning of the multi-source alternative data with real-time market data to generate a full-dimensional profile stream of participants; Dynamic heterogeneous graph construction and update module: used to receive the full-dimensional profile stream data of the participants, dynamically construct and update the heterogeneous graph of precious metal market participants with precious metal market participants as nodes and the relationship between participants as edges, and calculate the behavioral convergence degree between nodes based on the on-chain address transaction behavior pattern in multi-source alternative data, and use the calculated behavioral convergence degree as the dynamic weight of the corresponding edge in the heterogeneous graph of precious metal market participants. The counterparty intelligent scoring module is used to run a pre-trained heterogeneous graph neural network model on the heterogeneous graph of the precious metals market participants when an order to be matched is received, with the order initiator node as the anchor point. Through node-level and semantic-level two-layer attention aggregation, it infers the potential default probability of each candidate counterparty node and the related market disturbance behavior that may be triggered after the order is completed along different meta-paths, and finally outputs the risk-adjusted comprehensive matching score of each candidate counterparty. Intelligent matching and order routing module: used to sort and filter candidate counterparties based on the risk-adjusted comprehensive matching score, and automatically generate matching decisions or routing suggestions.

[0019] In one embodiment of the present invention, the multi-source data acquisition and fusion module includes: Multi-source alternative data includes publicly available holdings reports, industry chain upstream and downstream relationship data, related event text data, and on-chain address transaction behavior pattern data that has been anonymized; Semantic alignment and cleaning include entity identity alignment, timeline alignment, and semantic relationship alignment, specifically: The entity identity alignment adopts entity disambiguation and coreference resolution technology based on knowledge graph embedding to map different names of the same market participant in different data sources to the same globally unique identifier. The timeline alignment is used to maintain a unified event timeline, normalize the time granularity of data sources with different update frequencies, and adopt a strategy that combines forward padding and time-effect decay weighting for low-frequency data. When a major event related to a participant is detected, an instant refresh of the full-dimensional profile of the participant is immediately triggered. The semantic relationship alignment uses a relationship fusion inference engine to normalize and remove redundancy from multiple relationships between the same pair of participants in different data sources. When it is detected that the frequency of fund transactions between on-chain addresses of participants within a preset observation period is greater than a preset frequency threshold, and the statistical significance of the direction of fund transactions being opposite to the direction of changes in their public holdings is greater than a preset significance threshold, the transaction relationship between the two is strengthened into a suspected wash trading risk relationship, thereby generating structured full-dimensional profile stream data of participants.

[0020] Specifically, alternative data from multiple sources is acquired in parallel through four acquisition channels: The first channel is a public position report acquisition channel, which connects to the weekly position reports regularly published by exchanges and regulatory agencies to obtain the varieties, directions, quantities, and weekly changes of each participant in the precious metals futures and spot markets. The obtained raw reports are mostly in PDF or HTML format. The channel has a built-in text parser to convert the unstructured report content into structured four-tuple data, and attaches two timestamps: the report release date and the data effective date. The second channel is a supply chain upstream and downstream relationship acquisition channel, which connects to the interface of compliant supply chain data service providers, and is supplemented by text extraction from mining company annual reports and refiner announcements to obtain the supply and demand relationships between participants. The first channel, based on relationships, subordinate relationships, and long-term procurement agreements, employs a remotely monitored relationship extraction model to automatically identify relationship triples from unstructured text, in the form of "subject entity—relationship type—object entity." The second channel is a related event text collection channel, which captures publicly available event information related to precious metals market participants, links each event text to its entities, and associates the names of institutions mentioned in the text with the unique identifiers of participants within the system, outputting "participant—event—association strength" triples. The third channel is a de-identified on-chain address transaction behavior pattern collection channel, which obtains behavioral statistics at the address cluster level through a compliant blockchain data analysis platform interface, rather than the raw transaction records of a single address. The obtained content includes: the average daily amount transferred from address clusters to exchange hot wallets, the average daily amount transferred out of exchange hot wallets, the number and time interval of large transfers, the proportion of transfer frequency during periods of low liquidity, and the frequency of interaction with stablecoin issuer addresses. All of the above data are statistical feature values ​​after clustering and de-identification processing, and do not include raw transaction records traceable to specific individuals. Because the same participant is referred to inconsistently across different data sources—for example, using registration codes in position reports, abbreviations in event texts, and address cluster tags in on-chain data—it is necessary to unify entity identities first. This embodiment employs an entity disambiguation and coreference resolution model based on knowledge graph embedding. The model is a pre-trained dual-tower encoder. The input consists of multi-dimensional attribute vectors (name string vector, registration location code vector, and position behavior pattern vector for the past three months) for two entities from two different data sources. The output is the confidence score of the two entities pointing to the same real subject. During model training, positive samples are known pairs of the same entity across data sources (e.g., manually labeled historical data), and negative samples are randomly paired pairs of different entities. The loss function is a contrastive loss, specifically: ,in, For real labels (1 indicates the same entity, 0 indicates different entities). Alignment confidence score for model output; For time-sampling alignment, the update frequencies of different data sources vary significantly: market data is at the millisecond level, on-chain behavior data at the minute level, related event texts are triggered irregularly, public position reports are at the weekly level, and industry chain data is at the monthly level. This module maintains a unified event timeline. For high-frequency data such as market data and on-chain behavior, it is directly mapped to the timeline according to its original timestamp. For low-frequency data such as position reports and industry chain data, forward padding is used between two formal updates, that is, the most recent valid value is used, and a time decay factor is added to the value. ,in, The initial weights when the data is first released. The decay rate constant is This represents the time interval between the current moment and the data release time; the time-decrease factor will be passed to the downstream graph construction module along with the data to adjust the weight contribution of the corresponding edge; Based on this, an event-triggered refresh mechanism is set up: the system continuously monitors the associated event flow and on-chain behavior flow of each participant. When it detects that a large transfer exceeding the preset amount occurs in the on-chain address cluster of a participant, or when the participant appears in negative event text, the system immediately pulls the latest data of the participant from the relevant data source before the normal update cycle of low-frequency data arrives, overwrites the forward padding value, and resets the time decay factor to the initial value. The preset amount threshold is determined based on the 95th quantile of the on-chain transfer amount distribution of the product over the past 90 trading days. Based on the alignment of entities and time, this module normalizes and removes redundancy from multiple relationships between the same pair of participants. It deploys a relationship fusion inference engine, whose input is a set of relationship triples from different data sources, and whose output is the normalized relationship type and strength. This inference engine has a built-in set of conflict resolution rules: when two data sources provide different descriptions of the same relationship, a weighted decision is made based on the credibility of the data sources; when two relationship types semantically contain each other, the more specific relationship type is retained and an index pointing to the more general type is added; specifically, the relationship fusion inference engine performs cross-validation of wash trading risk, specifically: for any pair of participants A and B, the inference engine extracts the frequency of fund transactions between the associated address clusters of A and B within a preset observation period from on-chain behavioral data, and extracts the sequence of position change directions of A and B within the same period from public position data. When the frequency of fund transactions exceeds a preset threshold... When the statistical significance of the opposite direction of fund flows and open position changes exceeds a preset significance threshold, the inference engine rewrites the edge between A and B, which was originally marked as a normal trading relationship, as a suspected wash trading risk relationship. The preset frequency threshold is obtained by statistically analyzing the distribution of the frequency of on-chain fund flows between all known participants of this precious metal product within the same observation period, and taking the 90th quantile of this distribution as the preset frequency threshold. The statistical significance is calculated by constructing a binary sequence of fund flow direction sequence and open position change direction sequence (1 for consistent direction and 0 for opposite direction), and using the chi-square test to determine whether the event of "fund flow direction and open position direction are opposite" is independent of random distribution, and taking the p-value as the statistical significance index. In this embodiment, the preset significance threshold is 0.05, that is, when the p-value is less than 0.05, the null hypothesis that "opposite direction is purely random" is rejected, and the opposite relationship is considered to be statistically significant. After the above three alignment operations, the multi-source data acquisition and fusion module outputs structured full-dimensional participant profile stream data. Each data entry includes: timestamp, unique participant identifier, data source identifier, dimension type, feature vector, time decay factor, and relationship strength value. This data is pushed to the dynamic heterogeneous graph construction and update module in real time via a message queue.

[0021] In one embodiment of the present invention, the entity disambiguation and coreference resolution technology calculates the similarity of name strings, the consistency of registration location, and the matching degree of holding behavior patterns to output entity alignment confidence. When the entity alignment confidence is greater than a first preset threshold, entities are automatically merged. When the second preset threshold is less than or equal to the entity alignment confidence and less than or equal to the first preset threshold, they are marked as pending manual confirmation.

[0022] Specifically, name string similarity is determined by first performing text normalization on the names or registration codes of any two entities to be aligned. This includes removing whitespace and punctuation, unifying capitalization, and converting full-width characters to half-width characters. After normalization, the following three metrics are calculated: Edit distance similarity: ,in, The Levenstein edit distance between the two strings. and The lengths of the two strings are given; character n-gram overlap: taking n=3, the two strings are divided into sets of character triples, and the Jaccard coefficient is calculated: ,in, and These are sets of character triplets for the two strings, respectively; Abbreviation matching score: if every character of one string appears in the other string in the correct order (e.g., "GS" appears in "GoldmanSachs"), the score is 1; otherwise, it is 0. Edit distance similarity Character n-gram overlap Matching degree with abbreviations We obtain the name string similarity by weighting and summing: The weights mentioned above are based on the grid search results of the labeled validation set; The consistency score of the registration location is calculated by comparing the registration location codes of two entities. If the two entities are registered in the same country, the consistency score is 1; if the registered countries are different but they belong to the same preset regional group (regional groups are divided according to the geographical proximity of major precious metal trading centers, such as "North America Group", "Western Europe Group" and "East Asia Group"), the consistency score is 0.5; otherwise, the consistency score is 0. For on-chain address clusters, the distribution vector of the region to which the main interaction addresses belong is used instead of the registration location code, and the consistency score is calculated based on the cosine similarity of the two distribution vectors, denoted as . By taking the publicly available weekly change values ​​of two entities' positions over the past 12 weeks, a matching degree of position behavior patterns is constructed. Specifically, the two sequences are first standardized using Z-scores to obtain the sequence. and Then calculate the Pearson correlation coefficient between the two. ,in, For covariance, and These are the standard deviations of the two sequences; simultaneously, the sign rate of the first-order difference sequence is calculated, which is obtained by comparing the direction of position changes (increase, decrease, or no change) week by week, and dividing the number of weeks with the same change by the total number of weeks. The matching degree of holding behavior patterns is the weighted sum of the two: ,in, To truncate the negative correlation coefficient to 0, since two entities with opposite holding directions are likely not the same entity, the weight values ​​are also derived from the grid search of the labeled validation set; Overall confidence level Among them, the matching degree of holding behavior pattern has the highest weight because in the precious metals market, holding behavior has the "fingerprint" characteristics of participants, and its distinguishability is significantly higher than that of name and registration location information. The above weight values ​​are obtained by fitting logistic regression on the labeled validation set: with the three sub-scores as features and "whether they are the same entity" as the label, the logistic regression model is trained, and the normalized regression coefficients are the weights of each sub-item. The first and second preset thresholds are calculated by taking an independent labeled validation set and calculating the overall confidence level for each sample pair. ,according to The values ​​are sorted from highest to lowest, and precision and recall are calculated point by point. The first preset threshold is the minimum value corresponding to the first precision reaching 0.95. The second preset threshold is the minimum value corresponding to the recall rate first reaching 0.90. value.

[0023] In one embodiment of the present invention, the dynamic heterogeneous graph construction and update module calculates the behavioral convergence degree between nodes based on on-chain address transaction behavior patterns in multi-source alternative data, including: On-chain behavioral feature vectors are extracted from the full-dimensional profile data of each participant. These on-chain behavioral feature vectors are multi-dimensional time-series tensors, with each dimension corresponding to a behavioral indicator in the most recent time series. Statistical values ​​within a time window; After aligning and normalizing the on-chain behavioral feature vectors of any two participant nodes in the same time window, the Pearson correlation coefficient and dynamic time warping distance are calculated for each feature dimension in both the short and long time windows. After weighted fusion, the behavioral similarity scores of the dimension are obtained. Finally, the behavioral similarity scores of all dimensions are combined to obtain the comprehensive behavioral convergence between the two participant nodes. The permutation test method is used to perform a statistical significance test on the overall behavioral convergence. When the actual observed overall behavioral convergence is greater than the preset quantile threshold of the random distribution, the convergence relationship is determined to be significant. A behavioral convergence relationship edge is created or updated in the heterogeneous graph of precious metal market participants, and the value of the overall behavioral convergence is used as the dynamic weight of the behavioral convergence relationship edge.

[0024] Specifically, on-chain behavioral feature vectors are extracted from the full-dimensional profile stream data of each participant node. ,for 3D temporal tensor, Five indicators are selected, each dimension corresponding to a statistical sequence of a behavioral indicator within the most recent N time windows. In this embodiment, the time window length is set to 24 hours, and N is 30, that is, the daily statistical sequence of the most recent 30 days is extracted; if a participant has no on-chain activity record in some time windows, the corresponding position is filled with 0; the five behavioral indicators are defined as follows: Indicator 1 ( ): Net inflow of funds into exchange hot wallets: Defined as the total amount transferred out of the participant's associated address cluster to known exchange hot wallet addresses within each time window, minus the total amount transferred in from exchange hot wallet addresses; Indicator 2 ( ): Frequency of large transaction events: defined as the number of transactions within each time window where the amount of a single transaction by the associated address cluster of the participant exceeds the 95th quantile of the on-chain transaction amount distribution of the past 90 trading days for that product; Indicator 3 ( ): Percentage of transfers during periods of low liquidity: Defined as the number of transfers made by the associated address cluster of the participant between 0:00 and 6:00 UTC within each time window, divided by the total number of transfers made by the cluster within that time window; Indicator 4 ( Stablecoin exchange activity: Defined as the total number of interactions between the participant's associated address cluster and the addresses of major stablecoin issuers and stablecoin liquidity pools of major decentralized exchanges within each time window; Indicator 5 ( Concentration of internal transfers within an address cluster: Defined as the total amount transferred by the participant's associated address cluster to other addresses within its own cluster within each time window, divided by the sum of the total amount transferred outwards and the total amount transferred inwards within that time window; thus, the on-chain behavior feature vector of each participant. Given a 5×30 matrix, where the first... line, number Column elements Indicates the first behavioral indicators in the first Statistical values ​​within a time window; For any two participant nodes and Take the feature vector of its chain behavior. and First, perform time window alignment: using the union of the two timestamps as a reference, fill missing positions with 0 to ensure... and A one-to-one correspondence on the timeline; then for each behavioral indicator... Z-score standardization was performed on the 30-day series data. After standardization, each behavioral indicator was analyzed. The similarity was calculated for both short-term and long-term windows. The short-term window covered the most recent 7 days (t=24 to t=30), and the long-term window covered the entire 30 days (t=1 to t=30). Specifically, the Pearson correlation coefficients of the standardized two sequences were calculated for both the short-term and long-term windows. ,in, and These are the two sequences within a short time window. Standardized values ​​for each time window and The summation is the mean within the corresponding window, and the summation range is all time points within the short-time window; for the long-time window... The same formula is used to calculate the distance at all 30 time points. The dynamic time warping distance is calculated using the existing DTW algorithm, which allows for non-linear scaling and alignment of sequences on the time axis to capture the morphological similarity of behavioral patterns, rather than simply comparing them point-by-point at the same time. The original DTW distance is calculated separately for short-term and long-term windows. and After normalization, we get: ,in, It is a very small positive number, with a value of 10. -6 , The closer to 0, the more similar the temporal patterns of the two sequences; for each behavioral indicator The Pearson correlation coefficient and dynamic time regularization distance over both the short-term and long-term windows are weighted and fused to obtain the behavioral similarity score for this dimension. ,in, ... and The weighting coefficients are obtained through the analytic hierarchy process (AHP). In this embodiment... , , and The values ​​are 0.30, 0.25, 0.25 and 0.20 respectively; All The participant's behavioral similarity scores across all dimensions are weighted and summed to obtain the participant's score. and Overall behavioral convergence among them: ,in, For the first The weights of each behavioral indicator are determined by training a gradient boosting tree model using labeled historical collaborative transaction events as supervision signals. This model uses five dimensions... The input feature is the value, and the binary label is "whether the two participants engage in coordinated trading in the same direction within a subsequent preset time period". After training, the importance score of each feature is output, and the importance score is normalized as the result. The labeled historical collaborative trading events refer to instances in historical data where two participants made large, similar-valued opening positions in the same precious metal commodity on the same trading day or adjacent trading days, and these actions were subsequently confirmed not to be driven by publicly available market information. This labeling process is completed manually in conjunction with subsequent market trend backtesting to form a training set. To ensure that the convergence edges added to the heterogeneous graph have a genuine risk meaning, this embodiment uses a permutation test to determine the statistical significance of this value. The specific operation of the permutation test is as follows: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] The on-chain behavior feature vector remains unchanged, and the participants... The on-chain behavior feature vector is randomly shuffled M times on the time axis. In this embodiment, M is 1000. After each shuffling, the overall behavioral convergence between the two is recalculated, resulting in 1000 random values, forming a random distribution. A preset quantile threshold for this random distribution is taken: in this embodiment, the upper 95th quantile of the random distribution is taken, which is the 50th value after sorting the 1000 random values ​​from largest to smallest. When the actual observed overall behavioral convergence... If the value is greater than the upper 95th quantile of the random distribution, the convergence relationship is considered significant at the 0.05 significance level. Subsequently, a convergence edge is created or updated in the heterogeneous graph of precious metal market participants, with the edge weight set to... The current value; if the edge already exists, it is updated using the exponential moving average: ,in, The edge weights before the update. For the updated edge weights, The value is set to 0.9; when the actual observed value does not exceed this quantile threshold, no behavioral convergence edge is created; if the edge already exists, its weight is attenuated. When the weight after decay is less than 0.1, the edge is deleted. Through the above significance test, only participant pairs whose behavioral pattern convergence is statistically significantly different from the random level will obtain behavioral convergence relationship edges in the heterogeneous graph, thus ensuring that the risk signal carried by the edge is credible.

[0025] In one embodiment of the present invention, in the intelligent scoring module for counterparties, a pre-trained heterogeneous graph neural network model is run on the heterogeneous graph of the precious metals market participants. First, anchor point perception initialization is performed, and the node corresponding to the order initiator in the heterogeneous graph is determined as the anchor node. Nodes with the opposite direction to the current order and with trading qualifications are selected to form a candidate counterparty set. The order attributes are encoded into an order attribute vector as additional input and injected into the heterogeneous graph neural network model. When performing node-level attention aggregation in the two-layer attention aggregation of node-level and semantic-level, for each candidate adversary node, its neighbor nodes are divided into transaction relationship neighbor groups, lending relationship neighbor groups, supply chain neighbor groups, co-occurrence of related events neighbor groups, and behavior convergence neighbor groups according to edge type. Within each relationship group, the attention weight of the neighbor nodes to the candidate adversary node is calculated. The attention weight is used to perform weighted aggregation of the neighbor features within each relationship group to obtain the neighborhood aggregation representation from the perspective of each relationship type. In the semantic-level attention aggregation of the node-level and semantic-level two-layer attention aggregation, the neighborhood aggregation representation under each relation type is regarded as a different semantic view. The attention weight is calculated for each semantic view and then weighted and summed to obtain the fusion graph embedding representation of the candidate opponent node.

[0026] Specifically, upon receiving an order to be matched, the anchor node is located in the heterogeneous graph using the globally unique identifier of the order initiator. If the node does not exist, it is first created by calling the dynamic heterogeneous graph construction and update module. Then, all nodes in the graph are traversed, and nodes that simultaneously meet the following conditions are selected to form a candidate counterparty set: the node type is a fund, miner, bank, or retailer; the cumulative change in holdings over the past 4 weeks is opposite to the direction of the current order; it is not on the restricted list and the tradable amount meets the minimum transaction unit; the order attributes are encoded into an order attribute vector, including the unique hot code of the commodity, the direction code, the quantity normalization value, and the price deviation, which is then mapped to an order injection vector through a single-layer fully connected network. , and the feature vector of the full-dimensional portrait of the node After concatenation, the initial hidden vectors of the nodes are obtained through a fully connected layer: ,in, For parameter matrices, As the bias vector, the initial representation of each node carries order condition information. For each candidate counterparty node, its neighbors are divided into transaction relationship groups, lending relationship groups, supply chain groups, co-occurrence groups of related events, and behavioral convergence groups according to edge type. Within each relationship group, the attention weights of neighboring nodes to candidate counterparty nodes are calculated. ,in, For relation type The corresponding trainable attention vector, Let be the attribute vector of the edge, containing the edge weight and the time decay factor; We obtain the relation type by weighting and summing the features of the neighbors within the group. Neighborhood aggregation vector from the perspective ; under the five relation types Treating them as five semantic views, calculate the weight of each view and merge them: , ,in, For semantic query vectors, For parameter matrices, For bias vectors, For graph embedding fusion, the node-level and semantic-level dual-layer attention follows the existing structure of heterogeneous graph attention networks. The difference in this invention is that order attribute vectors are injected into the initial features of nodes, and behavioral convergence relationship groups participate in aggregation as independent edge types. The model pre-training adopts multi-task joint learning, and the three output heads share the above attention layer parameters. The link prediction head outputs the probability of generating transaction relationships after concatenating the fused graph embeddings of two nodes and passing them through a fully connected layer. Its loss function is the binary cross-entropy: ,in, This represents the number of edge samples within the batch. For the first The label indicates whether a transaction has actually occurred on the edge within the next 7 days: 1 if a transaction has occurred, and 0 if no transaction has occurred. The probability value output by the link prediction header; the risk classification header is embedded in the node fusion graph and outputs the high-risk probability through a fully connected layer. Their loss functions are both binary cross-entropy. ,in, This represents the number of node samples within the batch. For the first Each node has a historical risk label, with 1 for high risk and 0 for low risk. The risk classification head outputs a high-risk probability; the perturbation regression head, after being embedded with the candidate counterparty fusion graph and concatenated with the order attribute vector, outputs the volatility prediction value 30 minutes after the transaction through a fully connected layer. Its loss function is the mean squared error. ,in, This refers to the number of historical order samples within the batch. For the first Volatility forecasts for each sample The true value of realized volatility 30 minutes after the trade is calculated from the annualized standard deviation of the minute-level closing price sequence for this instrument; the total loss is the weighted sum of the losses from the three sub-tasks: Among them, the weight values ​​are based on the fact that the market disturbance regression task is most directly related to the final scoring target, so the weight is the highest; link prediction and risk classification are auxiliary tasks, each accounting for 0.3; training uses the Adam optimizer with an initial learning rate of 0.001, which decays to 0.7 every 5 rounds, for a total of 30 rounds, with a batch size of 128, and the early stopping condition is that the total loss of the validation set does not decrease for 4 consecutive rounds; after pre-training, the semantic-level attention layer and output head are incrementally fine-tuned monthly using newly added transaction feedback data, and the node-level attention layer parameters are frozen; The order attributes include the type, direction, quantity, and price information of the order.

[0027] In one embodiment of the present invention, the counterparty intelligent scoring module infers the potential default probability of each candidate counterparty node and the related market disturbance behavior that may be triggered after the order is executed along different meta-paths, including: Reasoning is conducted along the risk transmission path of the supply chain, the risk path of related event impacts, and the risk path of hidden collaborative risks. The supply chain risk transmission path sequentially passes through the order initiator node, the candidate counterparty node, the miner node, the refiner node, and the downstream bank node. The associated event impact risk metapath passes sequentially through candidate counterparty nodes, negative event nodes, and associated entity nodes; The hidden collaborative risk element path passes through candidate opponent nodes, high-risk participant nodes, and other participant nodes in sequence. For each meta-path, the feature vectors and edge weights in the full-dimensional profile stream data of the participants corresponding to each node on the path are sampled from the path instance and attention pooled to obtain the corresponding path vector.

[0028] Specifically, the supply chain transmission risk meta-path sequentially passes through: order initiator node, candidate counterparty node, miner node, refiner node, and downstream bank node. Its semantics are: the order initiator has a transactional relationship with the candidate counterparty, the candidate counterparty lends to the miner, the miner supplies goods to the refiner, and the refiner supplies goods to the downstream bank. This path is used to capture the indirect impact of upstream supply chain risks transmitted along the industrial chain to the candidate counterparty. The associated event impact risk meta-path sequentially passes through: candidate counterparty node, negative event node, and associated entity node. Its semantics are: the candidate counterparty co-occurs with a negative event, and this negative event is also associated with another entity. This path is used to capture the impact risk of a negative event on the candidate counterparty after it spreads through associated entities. The hidden collaborative risk meta-path sequentially passes through: candidate counterparty node, high-risk participant node, and... Other participant nodes: This path is used to capture the risk of candidate adversaries indirectly exposing themselves to high-risk groups through hidden collaborative relationships; For each meta-path, path instance sampling is performed on the heterogeneous graph. The sampling method is as follows: Taking the current candidate adversary node as the second node of the path (in the meta-path of supply chain transmission risk, the first node is the order initiator node), the path is expanded hop by hop along the edge type defined by the meta-path. At each step, at most S neighbors that satisfy the edge type constraints are randomly selected to continue moving forward, where S is 5; After reaching the end of the path, a complete path instance is recorded; For each meta-path, P path instances are sampled for each candidate adversary node, where P is 10; If there are fewer than P path instances that can actually be generated, the actual number is used; If none can be generated, the meta-path is skipped, and the corresponding path vector is filled with a vector of all zeros; Each sampled path instance is converted into a path vector. Specifically, the full-dimensional feature vectors of all nodes on the path instance, and the edge attribute vectors of all edges (including edge weights and time-degradation factors), are extracted and concatenated in path order to obtain the original path vector. After Z-score normalization, the original path vector is mapped to a fixed-dimensional path vector through a single-layer fully connected network. ,in, This is the standardized original path vector. For parameter matrices, The bias vector is 64-dimensional. The original vector dimensions of each meta-path are different, but they are transformed by their respective... After mapping, it is unified to 64 dimensions; For multiple path instances sampled from the same candidate opponent under the same meta-path, attention pooling is performed to highlight path instances that are more valuable for risk assessment: ; ; ;in, For the first metapath The path vector of a path instance. It is a 64×64 parameter matrix. It is a 64-dimensional bias vector. This is a path query vector with a dimension of 64. For the first Attention weights for each path instance. This is the path vector aggregated under this meta-path. Thus, the supply chain transmission risk meta-path, the related event impact risk meta-path, and the hidden collaborative risk meta-path each yield a 64-dimensional path vector, denoted as follows: , and ; The parameters involved in meta-path reasoning include: path vector mapping matrix. Intrapath attention parameters , and These parameters are not trained separately, but are trained as part of the entire counterparty intelligent scoring module, in conjunction with the node-level attention layer, semantic-level attention layer, and downstream prediction head. During training, based on the pre-trained multi-task loss described in the previous embodiment, a path reasoning auxiliary loss is added. Specifically, for each path instance sampled from a meta-path, a path validity classification head is added. This classification head uses... Given the input, output the binary probability of whether the metapath significantly contributes to the final default prediction in the current order scenario, with the loss function being the binary cross-entropy: ,in, This represents the total number of metapath samples within the batch. For the first Labels of individual path samples The probability value output by the path validity classification header, and the label. During the training data construction phase, for each historical order sample, the output value of the default probability prediction header is calculated when the meta-path path vector is included, and the output value is calculated after removing the meta-path path vector. If the difference between the two exceeds a preset range, the label is 1, indicating that the meta-path makes a significant contribution to the prediction; otherwise, it is 0. In this embodiment, the preset range is 5%. The pre-training total loss is updated as follows: ,in, , and As defined in the previous embodiment, the training settings remain consistent with the previous embodiment: Adam optimizer, initial learning rate 0.001, decaying to 0.7 every 5 rounds, for a total of 30 rounds, batch size 128, early stopping condition is that the total loss of the verification set does not decrease for 4 consecutive rounds. After training, the path vectors of the three meta-paths are embedded and spliced ​​with the aforementioned fusion graph, and input to the downstream default probability prediction head and associated market disturbance prediction.

[0029] In one embodiment of the present invention, the counterparty intelligent scoring module outputs a risk-adjusted comprehensive matching score for each candidate counterparty, including: The path vectors corresponding to the supply chain transmission risk meta-path, the related event impact risk meta-path, and the hidden collaborative risk meta-path are fused through a gating mechanism to obtain a path fusion vector. The path fusion vector is then spliced ​​with the fusion graph embedding representation and input into the default probability prediction head and the related market disturbance prediction head, respectively. The default probability prediction header outputs the default probability of the candidate counterparty under the current order. ; The associated market disturbance prediction head outputs the expected market disturbance index exerted by candidate counterparties on the relevant precious metals within a preset time window after the transaction. The expected market disturbance index is obtained by comprehensively calculating the predicted abnormal fluctuation amplitude, the probability of order book imbalance, and the consistency of price shock direction. Simultaneously, based on the influence centrality measure of the candidate counterparties in the heterogeneous graph of precious metals market participants and the influence coefficient of their historical behavior on the market, the strategy value score of the candidate counterparties is calculated. ; The risk-adjusted composite matching score is calculated as follows: ,in, , and These are the weighting coefficients.

[0030] Specifically, the path vectors of the three meta-paths have been obtained. , and A gating mechanism is used to integrate the three: , ,in, It is a 64×64 parameter matrix. It is a 64-dimensional bias vector. For the first The gate vector corresponding to the element path, This indicates element-wise multiplication. This is the merged path fusion vector; the path fusion vector is... Embedded with the aforementioned fusion graph The vectors are concatenated to obtain the final score input vector. , It is a 128-dimensional vector, containing both direct neighbor aggregation information and meta-path indirect risk information; the default probability prediction head is a two-layer fully connected network, with input... Output the probability of default. : , ,in, It is a parameter matrix of 128×64. It is a 64-dimensional bias vector. It is a parameter vector of 64×1. For scalar bias, The value of t ranges from (0, 1), representing the estimated default probability of the candidate counterparty in the current order scenario. The training label construction method for the default probability prediction head is as follows: counterparties that were successfully matched in the historical order samples and subsequently defaulted or were forced to liquidate are taken as positive samples (label 1), and the remaining normal performance samples are taken as negative samples (label 0). The loss function is binary cross-entropy, and its form is the same as described above. Consistent, denoted as ; The related market disturbance prediction head is also a two-layer fully connected network, but its output has three components, corresponding to the abnormal volatility amplitude, the probability of order book imbalance, and the consistency of price shock direction: , , , ,in, It is a parameter matrix of 128×64. It is a 64-dimensional bias vector. , and It is a 64-dimensional parameter vector. , and For scalar bias, To predict the abnormal price fluctuation range of this commodity within a preset time window after the transaction, by... The function guarantees a positive output. To predict the probability of order book imbalance, by The function is mapped to (0, 1); To ensure consistency in the predicted price shock direction, the tanh function maps the data to [-1, 1]. Positive values ​​indicate that the shock direction aligns with the order direction, while negative values ​​indicate the opposite. The expected market disturbance index... The above three components are combined and calculated as follows: ,in, This is the daily average of the realized volatility of this instrument over the past 90 trading days, used to normalize abnormal volatility into relative strength. The function truncates the final result to the interval [0, 1]. The closer the value is to 1, the stronger the expected market disturbance after the candidate counterparty's transaction. The training labels for the related market disturbance prediction head are constructed as follows: Realized volatility, order book imbalance duration, and price deviation direction of the instrument within a preset time window (30 minutes in this example) after historical order transactions are executed, and then converted in reverse according to the above formula. The true value is used as the regression label, and the head loss function is the mean squared error, the form of which is the aforementioned Consistent, denoted as ; Strategy Value Score This score reflects the positive impact of the candidate counterparty's subsequent market behavior on the order initiator after a transaction. It is calculated by weighting two objective indicators and does not involve subjective human assignment; influence centrality measure. The weighted out-degree centrality of candidate counterparty nodes in a heterogeneous graph of precious metal market participants is calculated as follows: The number of transaction relationship edges and co-occurrence edges of related events originating from the node is counted. The sum of the edge weights of these edges is used as the original centrality value. Then, a min-max normalization is performed on all nodes in the graph, mapping them to the interval [0, 1]. The higher the value, the wider the node's connections with other participants in the market, and the more likely its trading behavior is to influence other participants to follow suit; historical behavior influence coefficient Take all completed transaction records of the candidate counterparty within the past 12 months. For each record, calculate the consistency rate between the deviation direction of the price of the commodity relative to the pre-execution benchmark price within 30 minutes after the transaction and the direction of the order. If an order is a buy order and the price is higher than the pre-execution benchmark price within 30 minutes after the transaction, the consistency rate of that record is 1; otherwise, it is 0. The arithmetic mean of the consistency rates of all records is the result. The value range is [0, 1]; The closer the score is to 1, the more stable the participant's historical trading behavior is in guiding the market price direction; strategy value score The weights were determined through correlation analysis on the labeled validation set: using "whether a favorable follow-up effect occurred after matching" in historical orders as the label, the weights were calculated respectively. and When used alone, the discriminative power (AUC value) of the label is weighted according to the proportion of AUC value, resulting in an approximate ratio of 0.6 to 0.4; The risk-adjusted composite matching score is calculated as follows: ,in, , and The weighting coefficient is dynamically generated by the system based on the risk preference parameters set by the order initiator before the transaction. Specifically, the order initiator selects one of three preset risk preference levels when placing the order: conservative, balanced, and aggressive. The conservative level corresponds to... , and Balanced type correspondence , and Aggressive type correspondence , and ; The value ranges from [0, 1]. The closer it is to 1, the higher the overall matching value of the candidate counterparty in the current order scenario. All candidate counterparties... After the calculation is completed, the data is sorted from high to low and then passed to the intelligent matching and order routing module. In this embodiment, the parameters involved include: gating fusion parameters. , Default probability prediction header parameters , , and Related market disturbance forecast head parameters , , , , , , and These parameters are jointly trained with the aforementioned node-level attention layer, semantic-level attention layer, and meta-path reasoning part, and are not optimized separately; among them, the total training loss is further expanded based on the aforementioned embodiments: ,in, The cross-entropy loss of the default probability prediction head is used for binary classification. To correlate the mean squared error loss of the market disturbance prediction head, both are weighted at 0.5, reflecting their direct correlation with the final scoring target. The training settings are the same as in the previous example: Adam optimizer, initial learning rate 0.001, decaying to 0.7 every 5 rounds, for a total of 30 rounds, batch size 128, early stopping condition is that the total loss of the validation set does not decrease for 4 consecutive rounds. After training, the model is deployed online, and the prediction head parameters are incrementally fine-tuned monthly using newly added transaction feedback data.

[0031] In one embodiment of the present invention, the risk-adjusted comprehensive matching score is used to rank and filter candidate counterparties, including: A hard screening process is conducted on candidate counterparties, filtering out those with a comprehensive matching score ≤ the system's preset minimum threshold, those on the order initiator's prohibited trading list, and those whose tradable amount does not meet the minimum transaction unit. For the candidate competitors who pass the hard screening, a joint ranking function based on the comprehensive score and the quoted price is used for ranking. The joint ranking function is as follows: ,in, The normalized, risk-adjusted composite matching score. The price is the normalized price. This is a risk-price balancing factor, set by the order initiator before the transaction.

[0032] Specifically, for each node in the candidate counterparty set, the following three conditions are checked in sequence, and the node is removed if any condition is not met: Condition 1: The overall matching score is higher than the system's preset minimum threshold. The system's preset minimum threshold is obtained by taking the 10th percentile of the overall matching score distribution of the actual counterparties in all completed matched orders in the past 90 trading days. This threshold is updated on a rolling basis as the system runs and is recalculated every 30 trading days; Condition 2: The candidate counterparty is not on the order initiator's prohibited trading list. The prohibited trading list is maintained by the order initiator before placing the order and stored in the list index table in the system. Matching is performed using a globally unique identifier as the index; Condition 3: The tradable amount of the candidate counterparty is not lower than the minimum trading unit of the current order variety. The tradable amount is the tradable amount value recorded in the most recent full-dimensional profile update of the candidate counterparty. The minimum trading unit is determined by the exchange or over-the-counter market rules and is read by the system from the variety configuration table; For the candidate competitors who passed the hard screening, their overall matching score and offer price were normalized respectively. Among them, the overall matching score... Since the value is already in the range [0, 1], we directly use the original value as the normalization result: Price normalization needs to be considered in conjunction with the current order direction. For sell orders, the higher the bid price offered by the candidate counterparty, the more advantageous it is for the order initiator. The normalization formula is: For buy orders, the lower the selling price offered by the candidate counterparty, the more advantageous it is for the order initiator. The normalization formula is: ,in, The buy offer from the candidate counterparty. The selling price of the candidate counterparty. and These represent the maximum and minimum prices quoted by all candidate competitors who passed the hard screening process. Take 10 -6 Used to prevent division by zero, after normalization The value ranges from [0, 1], and the closer it is to 1, the more favorable the quote is to the order initiator. in, This is a risk-price balancing factor, set by the order initiator before the transaction, with a value range of [0, 1]. At that time, the ranking was based entirely on the risk-adjusted composite matching score; At that time, it degenerates into the traditional price-price priority ranking; When taking the median value, the overall score and the quoted price are weighted proportionally; the order initiator sets this via a slider on the order placement interface. The system default value is 0.6; for all candidate opponents, the value is... The values ​​are sorted from high to low to obtain the final candidate sequence; after sorting, the system proceeds to the order matching and routing execution stage in sequence.

[0033] In one embodiment of the present invention, the intelligent matching and order routing module includes: When the candidate counterpart with the highest overall matching score cannot fully absorb the total number of orders, the remaining order quantity is re-sorted in the remaining candidate set, the second best candidate counterpart is selected and its available sub-order quantity is allocated, and the process is recursively executed until all orders are allocated or the candidate set is exhausted. The transaction routing instructions are automatically generated based on the matching results of each sub-order. Specifically, for the part that can be directly executed on the electronic platform, the sub-order is automatically routed to the corresponding transaction channel through the FIX protocol; for the off-exchange transaction part, an electronic inquiry form is generated and pushed to the counterparty's terminal through a preset encrypted communication channel. After each transaction is completed, the transaction result is automatically sent back to the dynamic heterogeneous graph construction and update module in the form of a transaction feedback data packet. Based on this, the dynamic heterogeneous graph construction and update module adds a new transaction relationship edge or enhances the weight of an existing transaction relationship edge in the heterogeneous graph of precious metal market participants.

[0034] Specifically, let the current total number of orders be... The candidate sequence after joint sorting is The system from Starting with the top-ranked candidate, read its available trading quota. ;like Not less than If the candidate competitor accepts the entire order, the sub-order quantity will be... The recursion ends if... Less than Then The corresponding quantity is allocated to the candidate counterparty as the first sub-order, and the remaining quantity... Subsequently, the assigned candidate opponents were removed from... Remove from the middle, and re-perform joint sorting on the remaining candidate set. The value remains unchanged, and the normalization required and (Recalculate within the remaining candidate set), select a new top-ranked candidate opponent, and repeat the above allocation process; if the remaining candidate set is empty but... If the value is still greater than zero, the unexecuted portion will be returned to the order initiator as a pending order, or transferred to the off-exchange brokerage channel according to the unexecuted processing strategy preset by the order initiator. For each allocated sub-order, a corresponding routing instruction is generated based on the type of trading venue where the counterparty is located. For sub-orders that can be directly executed on the electronic platform, the system encapsulates the sub-order parameters (product code, direction, quantity, limit price, counterparty identifier) ​​into a standard FIX protocol message and sends it through the established FIX session between the system and the corresponding trading channel. After sending, the system listens for execution report messages, and marks the sub-order status as executed upon receiving confirmation of the transaction. For sub-orders in the over-the-counter (OTC) trading portion, the system generates an electronic inquiry order. The inquiry order includes: product, direction, quantity, price, counterparty identifier, and inquiry validity period. The inquiry order is pushed to the counterparty's terminal through the system's preset encrypted communication channel (using the TLS 1.3 protocol). After the counterparty confirms the transaction within the inquiry validity period, the transaction report is returned to the system through the same channel, and the sub-order status is marked as confirmed. After each sub-order is matched, the system automatically generates a transaction feedback data packet containing the following fields: transaction timestamp, globally unique identifier of the order initiator, globally unique identifier of the counterparty, commodity code, transaction direction, transaction quantity, and transaction price. This data packet is pushed in real time to the dynamic heterogeneous graph construction and update module via an internal message queue. Upon receiving the packet, the module locates the order initiator node and the counterparty node in the heterogeneous graph of precious metals market participants and checks whether a transaction relationship edge already exists between them. If not, a new transaction relationship edge is created, with the edge weight initialized as the normalized product of the transaction quantity and transaction amount. The normalization method is to divide the transaction amount by the average daily transaction amount of the commodity over the past 30 trading days, truncated to [0.1, 10] as the initial weight. If the edge already exists, the edge weight is updated according to the following formula: ,in, To update the weights of the edges in the transaction relationship before the update. This is the normalized weight value corresponding to this transaction. The updated weights ensure that the transaction relationship edges in the heterogeneous graph always reflect the activity level of the actual transaction relationships between participants in the most recent 30 trading days, providing a continuously updated graph structure foundation for the subsequent scoring module.

[0035] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A precious metals trading system based on multi-source data fusion and intelligent matching engine, characterized in that, Includes the following modules: Multi-source data acquisition and fusion module: used to collect multi-source alternative data related to precious metals market participants in real time, and to perform semantic alignment and cleaning of the multi-source alternative data with real-time market data to generate a full-dimensional profile stream of participants; Dynamic heterogeneous graph construction and update module: used to receive the full-dimensional profile stream data of the participants, dynamically construct and update the heterogeneous graph of precious metal market participants with precious metal market participants as nodes and the relationship between participants as edges, and calculate the behavioral convergence degree between nodes based on the on-chain address transaction behavior pattern in multi-source alternative data, and use the calculated behavioral convergence degree as the dynamic weight of the corresponding edge in the heterogeneous graph of precious metal market participants. The counterparty intelligent scoring module is used to run a pre-trained heterogeneous graph neural network model on the heterogeneous graph of the precious metals market participants when an order to be matched is received, with the order initiator node as the anchor point. Through node-level and semantic-level two-layer attention aggregation, it infers the potential default probability of each candidate counterparty node and the related market disturbance behavior that may be triggered after the order is completed along different meta-paths, and finally outputs the risk-adjusted comprehensive matching score of each candidate counterparty. Intelligent matching and order routing module: used to sort and filter candidate counterparties based on the risk-adjusted comprehensive matching score, and automatically generate matching decisions or routing suggestions.

2. The precious metals trading system based on multi-source data fusion and intelligent matching engine according to claim 1, characterized in that, The multi-source data acquisition and fusion module includes: Multi-source alternative data includes publicly available holdings reports, industry chain upstream and downstream relationship data, related event text data, and on-chain address transaction behavior pattern data that has been anonymized; Semantic alignment and cleaning include entity identity alignment, timeline alignment, and semantic relationship alignment, specifically: The entity identity alignment adopts entity disambiguation and coreference resolution technology based on knowledge graph embedding to map different names of the same market participant in different data sources to the same globally unique identifier. The timeline alignment is used to maintain a unified event timeline, normalize the time granularity of data sources with different update frequencies, and adopt a strategy that combines forward padding and time-effect decay weighting for low-frequency data. When a major event related to a participant is detected, an instant refresh of the full-dimensional profile of the participant is immediately triggered. The semantic relationship alignment uses a relationship fusion inference engine to normalize and remove redundancy from multiple relationships between the same pair of participants in different data sources. When it is detected that the frequency of fund transactions between on-chain addresses of participants within a preset observation period is greater than a preset frequency threshold, and the statistical significance of the direction of fund transactions being opposite to the direction of changes in their public holdings is greater than a preset significance threshold, the transaction relationship between the two is strengthened into a suspected wash trading risk relationship, thereby generating structured full-dimensional profile stream data of participants.

3. The precious metals trading system based on multi-source data fusion and intelligent matching engine according to claim 2, characterized in that, The entity disambiguation and coreference resolution technology calculates the similarity of name strings, the consistency of registration location, and the matching degree of holding behavior patterns to output entity alignment confidence. When the entity alignment confidence is greater than the first preset threshold, entities are automatically merged. When the second preset threshold is less than or equal to the entity alignment confidence and less than or equal to the first preset threshold, they are marked as pending manual confirmation.

4. The precious metals trading system based on multi-source data fusion and intelligent matching engine according to claim 1, characterized in that, In the dynamic heterogeneous graph construction and update module, the behavioral convergence degree between nodes is calculated based on the on-chain address transaction behavior patterns in multi-source alternative data, including: On-chain behavioral feature vectors are extracted from the full-dimensional profile data of each participant. These on-chain behavioral feature vectors are multi-dimensional time-series tensors, with each dimension corresponding to a behavioral indicator in the most recent time series. Statistical values ​​within a time window; After aligning and normalizing the on-chain behavioral feature vectors of any two participant nodes in the same time window, the Pearson correlation coefficient and dynamic time warping distance are calculated for each feature dimension in both the short and long time windows. After weighted fusion, the behavioral similarity scores of the dimension are obtained. Finally, the behavioral similarity scores of all dimensions are combined to obtain the comprehensive behavioral convergence between the two participant nodes. The permutation test method is used to perform a statistical significance test on the overall behavioral convergence. When the actual observed overall behavioral convergence is greater than the preset quantile threshold of the random distribution, the convergence relationship is determined to be significant. A behavioral convergence relationship edge is created or updated in the heterogeneous graph of precious metal market participants, and the value of the overall behavioral convergence is used as the dynamic weight of the behavioral convergence relationship edge.

5. The precious metals trading system based on multi-source data fusion and intelligent matching engine according to claim 1, characterized in that, In the intelligent scoring module for counterparties, a pre-trained heterogeneous graph neural network model is run on the heterogeneous graph of the precious metals market participants. First, anchor point perception initialization is performed, and the node corresponding to the order initiator in the heterogeneous graph is determined as the anchor point node. Nodes with the opposite direction to the current order and with trading qualifications are selected to form a candidate counterparty set. The order attributes are encoded into an order attribute vector as additional input and injected into the heterogeneous graph neural network model. When performing node-level attention aggregation in the two-layer attention aggregation of node-level and semantic-level, for each candidate adversary node, its neighbor nodes are divided into transaction relationship neighbor groups, lending relationship neighbor groups, supply chain neighbor groups, co-occurrence of related events neighbor groups, and behavior convergence neighbor groups according to edge type. Within each relationship group, the attention weight of the neighbor nodes to the candidate adversary node is calculated. The attention weight is used to perform weighted aggregation of the neighbor features within each relationship group to obtain the neighborhood aggregation representation from the perspective of each relationship type. In the semantic-level attention aggregation of the node-level and semantic-level two-layer attention aggregation, the neighborhood aggregation representation under each relation type is regarded as a different semantic view. The attention weight is calculated for each semantic view and then weighted and summed to obtain the fusion graph embedding representation of the candidate opponent node.

6. The precious metals trading system based on multi-source data fusion and intelligent matching engine according to claim 1, characterized in that, In the counterparty intelligent scoring module, the potential default probability of each candidate counterparty node and the related market disturbance behavior that may be triggered after the order is completed are inferred along different meta-paths, including: Reasoning is conducted along the risk transmission path of the supply chain, the risk path of related event impacts, and the risk path of hidden collaborative risks. The supply chain risk transmission path sequentially passes through the order initiator node, the candidate counterparty node, the miner node, the refiner node, and the downstream bank node. The associated event impact risk metapath passes sequentially through candidate counterparty nodes, negative event nodes, and associated entity nodes; The hidden collaborative risk element path passes through candidate opponent nodes, high-risk participant nodes, and other participant nodes in sequence. For each meta-path, the feature vectors and edge weights in the full-dimensional profile stream data of the participants corresponding to each node on the path are sampled from the path instance and attention pooled to obtain the corresponding path vector.

7. The precious metals trading system based on multi-source data fusion and intelligent matching engine according to claim 1, characterized in that, The counterparty intelligent scoring module outputs a risk-adjusted comprehensive matching score for each candidate counterparty, including: The path vectors corresponding to the supply chain transmission risk meta-path, the related event impact risk meta-path, and the hidden collaborative risk meta-path are fused through a gating mechanism to obtain a path fusion vector. The path fusion vector is then spliced ​​with the fusion graph embedding representation and input into the default probability prediction head and the related market disturbance prediction head, respectively. The default probability prediction header outputs the default probability of the candidate counterparty under the current order. ; The associated market disturbance prediction head outputs the expected market disturbance index exerted by candidate counterparties on the relevant precious metals within a preset time window after the transaction. The expected market disturbance index is obtained by comprehensively calculating the predicted abnormal fluctuation range, the probability of order book imbalance, and the consistency of price shock direction. Simultaneously, based on the influence centrality measure of the candidate counterparties in the heterogeneous graph of precious metals market participants and the influence coefficient of their historical behavior on the market, the strategy value score of the candidate counterparties is calculated. ; The risk-adjusted composite matching score is calculated as follows: ,in, , and These are the weighting coefficients.

8. The precious metals trading system based on multi-source data fusion and intelligent matching engine according to claim 7, characterized in that, The risk-adjusted comprehensive matching score is used to rank and filter candidate counterparties, including: A hard screening process is conducted on candidate counterparties, filtering out those with a comprehensive matching score ≤ the system's preset minimum threshold, those on the order initiator's prohibited trading list, and those whose tradable amount does not meet the minimum transaction unit. For the candidate competitors who pass the hard screening, a joint ranking function based on the comprehensive score and the quoted price is used for ranking. The joint ranking function is as follows: ,in, The normalized, risk-adjusted composite matching score. The price is the normalized price. This is a risk-price balancing factor, set by the order initiator before the transaction.

9. The precious metals trading system based on multi-source data fusion and intelligent matching engine according to claim 1, characterized in that, The intelligent matching and order routing module includes: When the candidate counterpart with the highest overall matching score cannot fully absorb the total number of orders, the remaining order quantity is re-sorted in the remaining candidate set, the second best candidate counterpart is selected and its available sub-order quantity is allocated, and the process is recursively executed until all orders are allocated or the candidate set is exhausted. The transaction routing instructions are automatically generated based on the matching results of each sub-order. Specifically, for the part that can be directly executed on the electronic platform, the sub-order is automatically routed to the corresponding transaction channel through the FIX protocol; for the off-exchange transaction part, an electronic inquiry form is generated and pushed to the counterparty's terminal through a preset encrypted communication channel. After each transaction is completed, the transaction result is automatically sent back to the dynamic heterogeneous graph construction and update module in the form of a transaction feedback data packet. Based on this, the dynamic heterogeneous graph construction and update module adds a new transaction relationship edge or enhances the weight of an existing transaction relationship edge in the heterogeneous graph of precious metal market participants.

Images