Intelligent auxiliary system and method for security transaction and liquidation based on large model

Abstract

The invention discloses an intelligent auxiliary system and method for security transaction and liquidation based on a large model, and belongs to the technical field of data processing systems or methods specially suitable for finance, and the method comprises the steps: receiving and preprocessing the natural language description and real-time constraint state information of the transaction intention of a user; extracting financial semantic features to generate an initial transaction semantic vector; generating a compliance transaction instruction through a limited decoder perceived in a legal form; constructing a joint intention tensor, and synchronously generating fund transfer and security transfer instructions; and combining the supervision rule anchoring reasoning chain and the gradient coordination mechanism optimization model to generate a transaction clearing instruction packet. According to the method, the problems of instruction non-compliance, coupon payment atomicity deficiency and the like caused by probability decoding in the prior art are solved, it is ensured that the instruction conforms to supervision and business specifications, atomicity execution and audit tracing under intelligent decision making are achieved, the business risk is reduced, and the financial supervision base line requirement is met.

Description

Technical Field

[0001] This invention relates to the field of data processing systems or methods specifically applicable to finance, and more particularly to intelligent auxiliary systems and methods for securities trading and clearing based on large models. Background Technology

[0002] Securities trading and clearing are the core cornerstones of modern financial markets. Trading involves accurately capturing investor intent, reaching contracts, and generating standardized trading instructions; while clearing is responsible for the transfer of asset ownership, final settlement of funds, and management of settlement risks. Although separated in time and space, they are logically complementary, jointly ensuring the liquidity of the global financial system. With the development of financial technology, securities business is undergoing a transformation from traditional electronic processes to highly intelligent ones. Large language models, with their potential in semantic understanding, multi-source information integration, and complex logical reasoning, are being widely introduced into the securities business chain. Through deep extraction of massive amounts of unstructured data such as research reports, regulatory announcements, and real-time news, intelligent auxiliary systems can help institutions capture signals and simulate strategies in complex markets, automating business processes to improve execution efficiency and reduce operational error rates.

[0003] However, existing intelligent auxiliary systems for securities trading and clearing based on large models exhibit significant technical limitations in practical deployments. Existing systems often treat trading decision-making logic and backend clearing execution logic as independent processes, resulting in a loosely coupled relationship between the two. When generating trading instructions, large models primarily rely on probabilistic output based on semantic rationality, making it difficult to perceive and embed the rigid constraints of the clearing process in real time, such as dynamic changes in cash positions, settlement windows for specific securities, or stringent formal requirements for message standards. This disconnect at the mechanism level leads to automated instructions being frequently rejected by the central clearing system after entering the clearing process due to minor formatting inconsistencies or logical violations. This technical deficiency not only increases the probability of settlement defaults but also forces the system to rely heavily on manual verification when handling complex transactions, severely limiting the reliability and security of intelligent methods in core financial business scenarios. Summary of the Invention

[0004] This invention overcomes the shortcomings of the prior art and provides an intelligent auxiliary system and method for securities trading and clearing based on a large model.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: an intelligent auxiliary method for securities trading and clearing based on a large model, comprising: S1. Receive the user's input of a natural language description of the transaction intent, synchronously obtain the corresponding real-time constraint status information, and perform preprocessing; wherein, the real-time constraint status information includes: market data, account cash position status, and deliverable securities holding information; S2. Based on the large model, extract financial semantic feature representations from the natural language description of transaction intent and real-time constraint state information to generate an initial transaction semantic vector; S3. Based on the initial transaction semantic vector, generate compliant transaction instructions through a legally perceptible restricted decoder; S4. Based on the initial transaction semantic vector, construct a joint intent tensor in the feature representation layer to generate fund transfer instructions and securities transfer instructions; S5. By combining compliant trading instructions, fund transfer instructions, and securities transfer instructions, a regulatory rule anchoring inference chain is constructed. Through a gradient coordination mechanism, the parameter updates of the large model are jointly optimized to generate a transaction settlement instruction package.

[0006] In a preferred embodiment of the present invention, in step S3, the legal form-aware restricted decoder is embedded with a deterministic finite automaton compiled from the rules of securities trading legal forms. When decoding each field of the generated instruction, the candidate tokens are filtered by the legality mask through the deterministic finite automaton to ensure that the structure and content of the generated instruction naturally conform to the formal requirements of the regulator and the exchange.

[0007] In a preferred embodiment of the present invention, step S4 includes: S41. Deconstruct the core features of the initial transaction semantic vector and extract the core sub-vectors related to funds and securities settlement; S42. Map the core sub-vectors to the semantic intent dimension, constraint state dimension, and atomic rule dimension to construct a three-dimensional joint intent tensor, and encode the atomicity requirement of payment against payment as a constraint matrix and embed it into the joint intent tensor. S43. Based on the joint intent tensor with embedded atomicity constraints, synchronously fill the fund transfer instruction template and the securities transfer instruction template to generate the fund transfer instruction and the securities transfer instruction with logical coupling.

[0008] In a preferred embodiment of the present invention, the preprocessing process in step S1 includes: performing dynamic standardization processing on the numerical real-time constraint status information to eliminate dimensional differences and retain the relative strength characteristics of the data relative to recent historical fluctuations; performing pre-liquidation logic calculation, by calculating the virtual available position after deducting frozen funds and estimated funds for pending orders from the total account balance, and the sellable position after deducting pledged and restricted quantities, embedding the pre-verification of liquidation rules into the data input stage.

[0009] In a preferred embodiment of the present invention, the gradient coordination mechanism in step S5 includes: calculating the gradients of the semantic loss of the transaction instruction, the atomicity loss of the clearing instruction, and the compliance loss of the regulatory rule; dynamically detecting and modulating the conflict between different gradient directions; and ensuring the collaborative optimization of model parameters under multi-objective constraints through adaptive weighting.

[0010] In a preferred embodiment of the present invention, the regulatory rule anchoring inference chain in step S5 transforms securities regulatory regulations, exchange business rules and brokerage risk control rules into structured inference entries and embeds them into the instruction verification and model feature activation process to achieve traceability and mandatory compliance constraints on the instruction generation logic.

[0011] In a preferred embodiment of the present invention, step S2 includes: constructing a gating fusion mechanism through a gating linear unit, injecting the hard constraint indicators obtained from the pre-clearing logic into the semantic feature stream, and dynamically suppressing the transmission of the original transaction intent features according to the severity of the real-time constraints, thereby realizing the front-end semantic risk control logic that prioritizes constraints over intent.

[0012] In a preferred embodiment of the present invention, the extraction of financial semantic features in step S2 includes: introducing a knowledge graph-based semantic enhancement mechanism, utilizing financial professional knowledge parameters stored within a large model to perform semantic calibration and vector representation enhancement on professional terms, securities codes, and business logic in the transaction intent.

[0013] In a preferred embodiment of the present invention, the generation of the transaction clearing instruction package in step S5 includes: integrating transaction instructions, fund instructions and securities instructions that have passed the end-to-end compliance verification, and attaching a traceability information module containing regulatory rule entries, gradient adjustment records and digital signatures to generate a standardized data package that conforms to the message standards of China Securities Depository and Clearing Corporation Limited.

[0014] This invention provides an intelligent auxiliary system for securities trading and clearing based on a large model, comprising: The data synchronization and preprocessing module is used to receive and integrate user trading intentions and real-time market status, perform pre-clearing logic calculations, and construct a multimodal integrated input tensor. The semantic understanding and feature extraction module is used to parse financial semantics through a large model encoder and use a gating fusion mechanism to implement risk control logic that prioritizes constraints over intentions, generating an initial transaction semantic vector. The compliant transaction instruction generation module is used to transform semantic vectors into compliant transaction instructions that meet the requirements of legal form using a constrained decoder based on deterministic finite automata constraints. The joint clearing instruction generation module is used to construct a joint intent tensor and embed atomicity constraint encoding, and synchronously and in parallel generate fund transfer instructions and securities transfer instructions that meet the atomicity requirements of delivery versus payment. The regulatory anchoring and joint optimization module is used to construct a regulatory rule anchoring inference chain, optimize model parameters through a multi-objective gradient coordination mechanism, and package and output a transaction clearing instruction package with audit traceability information.

[0015] This invention addresses the shortcomings of the prior art and has the following beneficial effects: This invention introduces a deterministic finite automaton generated by settlement rules into the decoding layer, constructing a legally form-aware restricted decoder. This decoder performs a legality mask filtering at the moment of token generation, ensuring that the generated transaction instructions naturally comply with the formal requirements of securities regulations, central counterparty rules, and exchange message standards. This achieves a unity of instruction formal validity and semantic rationality. Compared to existing technologies that generally equate semantic logical rationality with formal validity and lack underlying formal grammatical constraints, this invention effectively solves the problem that decision outputs cannot meet legal requirements for liquidation.

[0016] This invention designs a dual-channel joint action coding technology to construct a joint intent tensor that integrates transaction intent, real-time constraints, and atomic rules. This maps the atomicity requirement of payment against securities to joint constraints in a high-dimensional vector space, enabling the synchronous generation and logical coupling of fund transfer instructions and securities transfer instructions. This ensures the matching degree and synchronization of the two instructions. Compared with existing technologies that discretize the atomicity requirement of payment against securities into a sequence of instructions and have logical decoupling, this invention effectively avoids asymmetric deviations in settlement status caused by asynchronous communication or system fluctuations.

[0017] This invention introduces pre-clearing logic to calculate virtual available positions during the data preprocessing stage. It simultaneously acquires and standardizes real-time constraint information such as market conditions, cash positions, and deliverable positions, anchoring the intelligent decision-making benchmark within real trading constraints. This solves the problem of illusory suggestions caused by information gaps in large models. Compared with existing technologies that only perform simple data cleaning and ignore pre-verification of clearing rules, this invention ensures that the features of the input model implicitly contain hard business constraints, improves the executability of instruction generation, further reduces the risk of settlement default, and meets the special business needs of irrevocable securities clearing instructions.

[0018] This invention embeds multi-dimensional regulatory rules into the entire process of instruction verification and model optimization by constructing a regulatory rule anchoring inference chain and gradient coordination mechanism. It dynamically coordinates the update directions of three types of loss gradients: semantic, atomic, and compliance, to achieve traceable verification of instruction compliance and multi-objective collaborative optimization of the model. Compared with existing technologies, which cannot achieve deep embedding of regulatory rules and are prone to objective cancellation in model optimization, this invention ensures that instructions comply with regulatory requirements throughout the entire process, achieves atomic execution and audit traceability under intelligent decision-making, and meets the bottom-line requirements of financial regulation for the accuracy of instructions and the atomicity of operations. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a flowchart of a preferred embodiment of the present invention; Figure 2 This is a schematic diagram of the constraint gating fusion principle of a preferred embodiment of the present invention; Figure 3 This is a diagram of a restricted decoding mechanism according to a preferred embodiment of the present invention. Detailed Implementation

[0020] 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.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0022] Application Overview: This invention relates to a large-scale model-driven intelligent auxiliary system for securities trading and clearing, focusing on a real-time alignment mechanism between trading instruction generation and cross-entity clearing status. The essential characteristic of securities business lies in the fact that while the conclusion of a transaction (contract formation) and the completion of clearing (transfer of ownership) are legally separate, they must be strongly consistent in operation, and once the clearing process begins, instructions are irrevocable. The deficiency in existing technologies lies in the fact that systems generally equate the rationality of semantic logic with formal validity, ignoring the rigid formal requirements of clearing instructions under securities regulations and central counterparty rules. Due to the lack of underlying constraints on the formal syntax of clearing instructions, the decision output cannot meet the legal requirements for clearability, undermining the legal certainty of the intelligent decision-making chain and increasing the uncertainty risk in business execution.

[0023] The inability of existing technologies to achieve strong consistency in transaction and settlement execution stems from the probabilistic decoding mechanism of large-scale models. Character generation methods based on probability distributions cannot guarantee that the output strictly adheres to financial industry standards in terms of topological structure. Furthermore, the atomicity requirement of delivery versus payment (DVP) is discretized into a sequence of sequentially generated instructions in existing system architectures. Due to the logical decoupling between fund and security instructions, the model cannot guarantee the joint consistency of multiple instructions within the vector space. This lack of atomicity caused by probabilistic reasoning easily leads to asymmetric deviations in settlement states when facing asynchronous communication or system fluctuations. This failure of conventional systems makes it difficult to meet the minimum requirements of financial regulators for instruction accuracy and operational atomicity when handling high-value or cross-entity transactions.

[0024] This invention resolves the contradiction between probabilistic reasoning and rigid execution at the methodological level by introducing a restricted decoding method based on legal form awareness and a joint intent encoding mechanism. At the decoding layer, this invention introduces a deterministic finite automaton generated by settlement rules, performing a legality mask filtering at the moment of token generation to ensure that instructions naturally conform to formal requirements. Simultaneously, a dual-channel joint action encoding technology is designed to map the atomicity requirement of payment against securities to a joint constraint in a high-dimensional vector space, ensuring logical coupling between funds and securities instructions through synchronous projection of the intent tensor. Combined with a reasoning chain anchored to regulatory rules, this invention achieves atomic execution and audit traceability under intelligent decision-making.

[0025] Exemplary method: like Figure 1 As shown, the intelligent auxiliary method for securities trading and clearing based on a large model includes: S1. Receive the user's input of a natural language description of the transaction intent, synchronously obtain the corresponding real-time constraint status information, and perform preprocessing; wherein, the real-time constraint status information includes: market data, account cash position status, and deliverable securities holding information; S2. Based on the large model, extract financial semantic feature representations from the natural language description of transaction intent and real-time constraint state information to generate an initial transaction semantic vector; S3. Based on the initial transaction semantic vector, generate compliant transaction instructions through a legally perceptible restricted decoder; S4. Based on the initial transaction semantic vector, construct a joint intent tensor in the feature representation layer to generate fund transfer instructions and securities transfer instructions; S5. By combining compliant trading instructions, fund transfer instructions, and securities transfer instructions, a regulatory rule anchoring inference chain is constructed. Through a gradient coordination mechanism, the parameter updates of the large model are jointly optimized to generate a transaction settlement instruction package.

[0026] In the intelligent assistance method for securities trading and clearing based on large models, step S1 serves as the data access and initial perception layer for the entire process, providing high-quality, structured, and timely input data for subsequent semantic understanding, instruction generation, and joint optimization.

[0027] Step S1 first receives the user's input of a natural language description of trading intent. This description refers to unstructured text input by investors or traders through a spoken, written, or natural language dialogue interface. Its core elements include semantic elements such as security type, buy / sell direction, price range, and holding intentions. It serves as the raw input for the large model to understand user needs. This invention chooses natural language as the input method because it possesses extremely high expressive flexibility and information density, capable of capturing the complex decision-making logic and implicit semantics generated by traders in the market environment. To ensure the model can accurately identify these intents, the system uses a pre-built financial domain-specific dictionary in the background to perform lexical segmentation on the input natural language.

[0028] While receiving the intent description, step S1 synchronously retrieves the real-time constraint status information directly related to this intent description from the external market data source via the internal data bus.

[0029] Real-time constraint status information is a collection of multi-dimensional vectors, mainly including market data, account cash position status, and deliverable securities holding information.

[0030] Market data refers to real-time buy and sell quotes, latest transaction prices, trading volumes, order book depths, and calculated volatility indicators for target securities and related derivatives across multiple exchanges. Its data sources include real-time data streams from securities exchanges and aggregated data interfaces from financial information providers.

[0031] Account cash position status refers to the real-time available cash balance, frozen funds, funds in transit, and credit limit of a user's designated trading account within the settlement banking system, reflecting the current cash availability for payment.

[0032] Deliverable securities holding information refers to details such as the quantity, cost, holding status, and whether the securities are pledged or subject to lock-up periods that are currently held in the same account or related accounts and meet the requirements of the exchange and China Central Depository & Clearing Co., Ltd. for sale or delivery.

[0033] By synchronously acquiring this information, this invention brings the benchmark of intelligent decision-making back from pure semantic logic to the constraints of real transactions, solving the problem of illusory suggestions caused by information gaps in large models in the prior art.

[0034] After completing the synchronous acquisition of data, step S1 performs data preprocessing on the acquired data. The heterogeneous raw data with different frequencies and formats are transformed into numerical, normalized feature vectors with uniform dimensions, so that they can be input into the subsequent large model.

[0035] For the natural language description of transaction intent, a pre-trained word segmenter is used for processing.

[0036] Specifically, a BERT word segmenter, incrementally pre-trained for financial corpora, is used to segment the input sentence into a sequence of sub-words. This text sequence is then converted into a token ID sequence, with necessary padding or truncation to form a text embedding sequence, adapting to the input length limitations of large models. Simultaneously, key financial entities are identified and labeled, such as security codes, trading directions, quantities, price types, and constraints. The resulting word segmentation sequence, along with its corresponding word embedding vectors, forms the basic representation of the text. Meanwhile, real-time constraint state information is vectorized and encoded.

[0037] For numerical real-time constraint state information, due to the significant differences in the dimensions of different indicators, directly inputting them into a large model can lead to weight imbalance during gradient updates. Therefore, a dynamic Z-score normalization method is used to process market data and capital data. Let the characteristic value at a certain time t be... By calculating the mean value within a window W over a past period and standard deviation This is then converted into a standardized score. The specific standardized calculation formula is as follows: ; This is a very small constant used to prevent the denominator from being zero. This approach eliminates the influence of dimensions while preserving the relative strength of the data with respect to recent historical fluctuations, enabling the model to detect abnormal price fluctuations or abnormal changes in funds.

[0038] For discrete state information, such as securities codes and account types, one-hot encoding or specific financial entity embedding vectors are used for representation.

[0039] Specifically, in the preprocessing of funding positions and open positions information, step S1 introduces a pre-liquidation logic calculation method. Traditional preprocessing is merely a simple data cleaning process, while this invention calculates the virtual available positions at this stage. The formula is expressed as: ; For virtual available funds, Total account balance For funds that have been frozen, The estimated total amount of funds required for all unexecuted buy orders i in the current order pool P. Similarly, the formula for calculating securities holdings is: ;in, For positions that can be sold, Total holdings For the amount pledged, This represents the quantity that cannot be sold on the same day (for the T+1 market). This calculation method embeds the pre-verification of settlement rules into the data preprocessing stage, ensuring that the features input to the large model already implicitly contain hard business constraints.

[0040] Finally, step S1 performs multimodal concatenation of the cleaned text sequence and the standardized real-time constraint state vector to construct a comprehensive input tensor containing complete contextual information.

[0041] Step S2 takes the multimodal integrated input tensor output from step S1 and uses it as a semantic encoding and feature extraction step. It utilizes the powerful deep representation capabilities of the large language model to transform the discrete symbolic input into a continuous, dense, and high-dimensional semantic space representation rich in financial business logic.

[0042] The large model in step S2 specifically refers to a Transformer-based deep neural network model that has undergone full pre-training or instruction-based fine-tuning using financial domain-specific corpora, including research reports, transaction logs, regulatory rules, and clearing rules. Specifically, the pre-training process employs a dual-task training strategy of masked language modeling and sentence relationship prediction. The masked language modeling task masks financial terminology, forcing the model to learn the semantic relationships between terms; the sentence relationship prediction task constructs matching sample pairs of trading intentions and clearing constraints, allowing the model to learn the logical correspondence between intention and constraint, thereby enabling the large model to possess semantic understanding and constraint perception capabilities in financial scenarios.

[0043] Financial semantic feature representation refers to mapping discrete word signals and continuous market state values ​​onto a distributed representation in the same high-dimensional continuous vector space. In this space, the geometric distance of features represents the semantic or policy similarity.

[0044] In the specific implementation process, the comprehensive input tensor output in step S1 Where L is the total sequence length and d is the dimension of the hidden layers of the model, which is directly input into the large model. The model iteratively interacts and aggregates information at each position in the sequence, including lexical units representing intent and state labels representing constraints, through its multi-layered self-attention mechanism and feedforward neural network. The self-attention mechanism allows any two elements in the sequence to establish a direct connection. The specific attention calculation formula is as follows: ;in, In other words, the query, key, and value are all obtained by projecting the input of this layer through different linear transformations. Features derived from transaction intent Feature representation derived from real-time constraint information, This is the dimension of the attention head, used for scaling to prevent the gradient from vanishing due to excessively large dot products. In a multi-head attention mechanism, multiple such attention heads are computed in parallel, and their outputs are concatenated and then subjected to a linear projection. Through multi-layer stacking, the model can construct a complex inference chain from local lexical dependencies to global cross-modal semantic associations.

[0045] Step S2 further optimizes the extracted features using financial semantic feature representation. Financial semantic feature representation refers to the distributional expression in the latent space that accurately characterizes the core elements and potential risk attributes of securities trading. Since general language models often struggle to distinguish the precise meaning of financial terms in specific contexts, step S2 introduces a semantic enhancement mechanism based on knowledge graphs. During feature extraction, the model activates internally stored financial knowledge parameters to calibrate the input features.

[0046] like Figure 2 As shown, for the hard constraint indicators such as the virtual available position pre-calculated in step S1, step S2 forcibly injects them into the main stream of semantic features through a gating fusion mechanism. Specifically, a gating linear unit is used to control the degree of correction of the constraint information to the text semantics, and the formula can be expressed as: ;in, For text features, As a constraint feature, The Sigmoid activation function is used. For the fused compliance semantic features, For gating projection parameters, For feature transformation parameters, For Hadama accumulation, To achieve layer normalization, the fused features are standardized. Using this formula, when constraints are extremely strict, the gating mechanism suppresses the transmission of the original intent features, enhancing the dominance of the constraint features, thereby realizing a risk control logic where constraints take precedence over intent at the semantic level.

[0047] After completing the deep feature extraction and fusion, step S2 generates the initial transaction semantic vector. The initial transaction semantic vector is not a single vector, but a tensor sequence with a temporal structure, which carries all the source information required to generate subsequent compliance instructions.

[0048] Step S3, building upon the initial transaction semantic vector output from Step S2, resolves the contradiction between the probabilistic decoding mechanism of the large model and the rigid requirements of securities trading instruction format. Through a legally perceptive, constrained decoder, it transforms high-dimensional semantic features into compliant trading instructions that conform to regulatory rules and industry standards. This provides a legally valid and business-executable foundational instruction carrier for subsequent clearing instruction generation. Securities trading instructions are the core link between investors' trading intentions and the exchange's execution system. They not only need to carry clear trading semantics but also strictly adhere to the formal requirements of securities regulations, central counterparty rules, and exchange messaging standards. Any formal flaws will result in direct rejection, leading to trading delays or failures. Step S3 addresses this core requirement by constructing a precise conversion path from semantics to compliant instructions.

[0049] A legal form-aware constrained decoder is a generative decoder that embeds legal form rules for securities trading and exchange message standards. It differs from the free probabilistic decoding mode of traditional large models. By imposing deterministic formal constraints during the decoding process, it ensures that every token and field generated complies with preset legal and business norms. Legal form awareness means that the decoder can accurately identify and embed the core formal requirements of securities trading instructions, including field format specifications, field value ranges, and logical constraints between fields. Constraint means that the decoding path is constrained by a deterministic finite automaton, filtering out token candidates that do not meet the formal requirements and forcing the decoder to complete instruction generation within the compliant path.

[0050] A compliant trading instruction refers to a structured trading instruction that possesses both semantic rationality and formal validity. It includes core components such as an instruction header, a target information module, a trading parameter module, a risk control verification module, and a signature authentication module. The instruction header must contain unique information such as the instruction number, initiation time, and account identifier. The target information module must clearly identify the security code, market sector, and target type. The trading parameter module must cover core trading elements such as trading direction, order quantity, order price, and order type. The risk control verification module must embed regulatory requirement fields such as risk level and compliance identifier. The signature authentication module must contain digital signature information to ensure the immutability of the instruction. This type of instruction can be directly connected to the exchange's trading execution system without manual intervention.

[0051] Before constructing a restricted decoder for legal form perception, it is first necessary to build a rule base for securities trading legal forms, which is the core basis for restricted decoding.

[0052] The construction of the rule base needs to integrate multi-dimensional formal constraints, including three major categories: regulatory requirements at the securities regulatory level, message standards at the exchange level, and risk control rules at the brokerage level.

[0053] Securities regulatory requirements cover investor suitability management, trading instrument qualification constraints, and compliance of order types; exchange-level message standards cover instruction field structure, field length, field data type, and field encoding format; brokerage-level risk control rules cover account position limits, sufficient capital constraints, and trading frequency limits. The rule base is constructed using a structured storage method, transforming each constraint rule into a field constraint entry. Each entry includes attributes such as field name, field type, value range, related field constraints, and violation penalty weight. To ensure the rule base's real-time performance and adaptability, the system synchronizes rule update announcements from regulatory agencies and exchanges via interfaces, automatically triggering iterative updates to the rule base.

[0054] After the rule base is built, it needs to be compiled into a deterministic finite automaton, which is the key technology for implementing constrained decoding. A deterministic finite automaton is a mathematical model with deterministic state transition capabilities. It can transform the formal constraints in the rule base into explicit state transition paths, forcing the decoder to generate tokens within compliant paths.

[0055] The construction of a deterministic finite automaton involves defining four core elements: the state set, the input character set, the state transition function, the initial state, and the final state. The state set corresponds one-to-one with the field structure of compliant trading instructions, including the initial state, instruction header filling state, target information filling state, trading parameter filling state, risk control verification filling state, signature authentication filling state, and instruction termination state; each state corresponds to the filling process of one field module of the instruction; the input character set refers to the set of all legal tokens that conform to the rule base constraints.

[0056] For example, the input character set for the target information filling state is all security code tokens that conform to the x-digit numeric format. The state transition function is the core of the deterministic finite automaton, defined as the mapping relationship between the current state and the input tokens. That is, when the decoder inputs a valid token in a certain state, the automaton will transition to the next corresponding filling state. If the input token does not conform to the rule base constraints, the automaton rejects the token, maintains the current state, and waits for a new valid token input. The starting state is the initial trigger state for instruction generation, and the terminating state is the state where all fields of the instruction have been filled and passed the verification. Only when the automaton transitions to the terminating state will the decoder stop generating tokens and output a complete and compliant trading instruction. The compilation formula of the deterministic finite automaton can be expressed as follows: ,in, For a set of states, For the input character set, This is the state transition function. This is the initial state. As a termination state, this formula transforms unstructured constraints in the rule base into structured mathematical models, enabling precise constraints on the decoding process.

[0057] After compiling the deterministic finite automaton, the next step is to build the architecture of the legal form perception restricted decoder. This decoder is an improvement on the Transformer decoder architecture. In addition to the original self-attention layer and feedforward network layer, two core modules are added: a legal form perception layer and a DFA mask layer, to achieve deep integration of semantic features and formal constraints.

[0058] The initial transaction semantic vector output in step S2 is input into the self-attention layer of the decoder. The self-attention layer captures the semantic relationships of core elements such as the transaction target, transaction direction, and order price by calculating the attention weights between feature vectors, and generates semantically enhanced feature vectors.

[0059] The semantically enhanced feature vector is input to the legal form perception layer, which embeds a vector representation of a securities regulatory rule knowledge graph. The knowledge graph contains elements such as rule nodes, field nodes, and constraint relationship edges. By calculating the cosine similarity between the semantically enhanced feature vector and the knowledge graph node vector, rule constraints that match the current trading scenario are activated. For example, when the semantic feature vector contains information about a stock listed on the Science and Technology Innovation Board, the legal form perception layer will automatically activate the constraint vector of the investor suitability certification field and guide the decoder to generate that field.

[0060] like Figure 3 As shown, the feature vector enhanced by the legal form perception layer is input to the DFA masking layer, which is the core execution module of constrained decoding. The DFA masking layer reads the current state of the deterministic finite automaton in real time and performs masking filtering on the candidate token set generated by the feedforward network according to the state transition function, retaining only the legal tokens that meet the constraints of the current state. The masking formula can be expressed as follows: ;in, The original token candidate set generated by the feedforward network. The matrix is ​​a DFA mask, where elements corresponding to valid tokens have a value of 1, and elements corresponding to invalid tokens have a value of 0. This formula enables precise filtering of invalid tokens. Finally, the decoder generates the token for the current field based on the masked set of valid token candidates and drives the deterministic finite automaton to the next state. This process repeats the self-attention calculation, legal form awareness, DFA mask filtering, and token generation steps until the automaton reaches the termination state, generating a complete compliant transaction instruction.

[0061] After generating the initial compliant trading instruction, step S3 requires constructing a multi-layered compliance verification mechanism to ensure the formal validity and logical consistency of the instruction. The first layer is formal verification, which checks the completeness, format correctness, and length compliance of the instruction's fields based on field constraint entries in the rule base. The second layer is logical verification, which checks the logical relationships between fields. The third layer is regulatory compliance verification, which checks whether the instruction complies with specific regulatory requirements based on regulatory rules activated by the legal form perception layer.

[0062] Step S4 receives the initial transaction semantic vector generated in step S2, which serves as the key execution layer for transforming transaction intent into physical settlement actions. A joint intent tensor is constructed in the feature representation layer, and based on this tensor, fund transfer instructions and securities transfer instructions that meet the atomicity requirements of payment against securities are generated synchronously and in parallel.

[0063] The joint intent tensor is a high-dimensional structured tensor that integrates semantic features of transaction intent, real-time constraint state features, and atomic rules of securities delivery and payment. Its dimensions cover semantic intent dimension, constraint state dimension, and atomic rule dimension. The features of each dimension are interconnected and mutually constrained, ensuring that the generation basis of funds and securities orders has strong mathematical consistency.

[0064] A fund transfer instruction is a standardized instruction that specifies the account to which funds are transferred, the amount transferred, the purpose of the funds, the time limit for arrival of funds, and the rules for deducting handling fees. It is the legal basis for fund settlement and its format must strictly follow the fund settlement message standard.

[0065] A securities transfer instruction is a standardized instruction that specifies the underlying security code, the quantity to be transferred, the holding custody account, the settlement seat, and the restricted sale status. It is the core certificate for the transfer of securities ownership and is directly related to the legal effect of the settlement result.

[0066] Before constructing the joint intent tensor, it is necessary to first deconstruct the core features of the initial transaction semantic vector. The initial transaction semantic vector is a tensor sequence with a temporal structure and a dimension of [missing information]. ; For timing step size, As a feature dimension, it covers core information such as the trading target, trading volume, trading price, capital constraints, and securities constraints.

[0067] The deconstruction process employs a financial business-oriented attention pooling mechanism, extracting specialized sub-vectors based on the core needs of fund and securities order generation, specifically including transaction target sub-vectors. Transaction quantity subvector Transaction price subvector Available capital subvector Deliverable securities sub-vector and delivery time limit subvector The formula for calculating attention pooling is: ,in, Each core sub-vector is a dedicated query vector whose parameters are pre-trained using financial business rules, enabling it to accurately focus on feature information related to instruction generation. These are the key and value vectors of the initial transaction semantic vector, respectively. This attention pooling operation filters out redundant information from the initial transaction semantic vector, extracts core features directly related to the generation of clearing instructions, and provides a structured feature foundation for constructing the joint intent tensor.

[0068] After completing the core feature deconstruction, the next step is the dimensional expansion and feature mapping of the joint intent tensor, which is the core construction content of step S4. The dimensional expansion of the joint intent tensor needs to cover the semantic intent dimension. Constraint state dimension and atomic rule dimension There are three core dimensions, each corresponding to different feature attributes. The semantic intent dimension includes subjective transaction demands such as the transaction target, transaction quantity, and transaction direction, which is the source basis for instruction generation; the constraint state dimension includes objective constraint features such as available funds balance, number of deliverable securities, and amount of frozen funds, which are rigid prerequisites for the execution of instructions; the atomic rule dimension includes business rule features such as securities delivery versus payment matching rules, settlement time limit synchronization rules, and tax deduction rules, which are the core principles to ensure the atomicity of instructions.

[0069] The formula for calculating dimensional expansion is as follows: ,in, This is a tensor product operation used to achieve deep fusion of core subvectors and dimension embedding matrices. These are three dedicated embedding matrices, the parameters of which are pre-trained using a financial business rules dataset. They are capable of mapping core sub-vectors to a high-dimensional space with business semantics. This is a tensor concatenation operation used to fuse three-dimensional feature tensors into a single three-dimensional joint intent tensor, with the final dimension being... * * Through this dimensional expansion and feature mapping, the originally scattered core sub-vectors are integrated into a structured tensor whole. Each tensor element carries triple information: transaction intent, real-time constraints, and atomic rules, achieving strong correlation at the feature level.

[0070] After dimensional expansion, embedded encoding of atomic constraints is required, a crucial step to ensure logical coupling between funds and securities instructions. The atomicity of delivery versus payment (DVP) requires strict matching between the fund transfer amount and the securities transfer market value, and complete synchronization between the fund arrival time and the securities transfer time. This invention transforms these business rules into a mathematical constraint matrix, embedding a joint intent tensor through tensor operations. First, the atomic constraint matrix is ​​constructed. This matrix is ​​a two-dimensional matrix, with dimensions equal to the sum of the capital feature dimension and the security feature dimension. When two features satisfy the delivery versus payment rule, the matching weight in the matrix is ​​1; otherwise, it is 0. This achieves a mathematical expression of the business rules. Then, the constraint matrix is ​​embedded into the joint intent tensor using tensor multiplication, calculated using the following formula: Through this operation, each element in the joint intent tensor is assigned an atomic constraint weight, ensuring that the relationship between the funding characteristics and the security characteristics strictly conforms to the delivery versus payment (DVP) business rules.

[0071] After constructing the joint intent tensor and embedding atomic constraints, the system proceeds to the synchronous generation of fund transfer instructions and securities transfer instructions. This invention employs a generation mechanism combining standard financial industry template filling with real-time atomicity verification to ensure the standardization and validity of the instructions. First, the system pre-sets fund transfer instruction templates and securities transfer instruction templates that comply with regulatory requirements. These templates include core fields such as instruction number, initiating account, receiving account, underlying security code, transfer amount, transfer quantity, settlement time limit, commission rate, and digital signature. The field format and length strictly adhere to exchange message standards.

[0072] Then, the joint intention tensor based on embedded atomicity constraints Extract the corresponding features and populate them into the template. The core field of the fund transfer instruction, the transfer amount, is calculated from the transaction quantity, transaction price, and tax rate. The calculation formula is as follows: ,in, For transaction fee rates, For the transaction stamp duty rate, this formula ensures that the allocated amount includes the transaction principal, handling fees, and taxes, covering all funding needs; the core field of the securities transfer instruction, the transfer quantity, is calculated by matching the transaction quantity with the number of deliverable securities, using the following formula: ,in, The formula specifies the number of deliverable securities. It ensures that the number of securities transferred does not exceed the account's deliverable holdings, thus avoiding delivery defaults due to insufficient holdings.

[0073] After the initial instructions are generated, the system also needs to perform real-time atomicity verification to check the matching degree between the fund transfer amount and the securities transfer market value, ensuring the atomicity consistency of the two instructions. If the matching degree meets the requirements, the standardized fund transfer instructions and securities transfer instructions are directly output; if the matching degree does not meet the requirements, the reverse adjustment mechanism is initiated, by adjusting the attention weights of the fund features and securities features in the joint intent tensor, re-extracting features to fill the template, until the matching degree meets the error threshold requirements.

[0074] The fund transfer instructions and securities transfer instructions generated in step S4 not only have a standardized format and accurate content, but also carry dual characteristic information of atomicity constraints and real-time constraints, providing high-quality instruction input for step S5 to construct the regulatory rule anchoring inference chain.

[0075] In the intelligent auxiliary method for securities trading and clearing based on a large model, step S5 takes over the compliant trading instructions output by step S3 and the fund transfer instructions and securities transfer instructions output by step S4. As a closed loop link for joint optimization and instruction generation in the whole process, it constructs a regulatory rule anchored inference chain and combines a gradient coordination mechanism to jointly optimize the parameters of the large model, and finally generates a trading and clearing instruction package with compliance, atomicity and traceability.

[0076] Securities trading and clearing involve multiple parties and multiple layers of regulatory rules. Once an instruction is generated and submitted, it is irrevocable. Step S5 is precisely based on the core requirement of strong consistency between trading and clearing, realizing the integration of decentralized instructions into standardized instruction packages and iterative optimization of the model.

[0077] The regulatory rule anchoring inference chain refers to the logical chain that transforms securities regulatory laws, exchange business rules, and brokerage risk control requirements into structured inference rules and embeds them into the instruction verification and model optimization process. This inference chain can bind the generation process of each instruction to the corresponding regulatory rules, enabling traceable verification of instruction compliance.

[0078] The gradient coordination mechanism refers to resolving and modulating conflicting gradients by dynamically detecting the consistency of the gradient directions of different loss functions. This ensures that when the model is simultaneously optimizing multiple objectives such as the compliance of trading orders and the atomicity of clearing orders, there will be no problem of chaotic parameter update directions or the cancellation of optimization objectives.

[0079] A transaction clearing instruction package refers to a standardized data package containing compliant transaction instructions, fund transfer instructions, securities transfer instructions, and instruction traceability information. This instruction package not only contains execution instructions that can directly connect to the exchange and clearing system, but also contains the regulatory rules basis and parameter optimization process records for each instruction, and has complete audit traceability capabilities.

[0080] In the specific implementation process, the first step is to construct a regulatory rule anchoring inference chain. This chain requires integrating multi-dimensional regulatory rules, including departmental regulations issued by securities regulatory authorities, trading and clearing business rules formulated by securities exchanges, and risk control rules internal to securities firms. These unstructured rule texts are transformed into structured inference rule entries, each containing four core elements: rule number, constrained object, constraint condition, and consequences of violation. For example, a rule entry for a fund transfer instruction might be rule number J001, constrained object: fund transfer amount, constraint condition: the transfer amount must not exceed the account's available funds balance, and consequences of violation: instruction rejection and triggering a risk control warning. These rule entries are then embedded into the instruction verification process. The compliant trading instructions generated in step S3 and the fund transfer and securities transfer instructions generated in step S4 are verified one by one. Instructions that pass verification are marked with a compliance flag, while those that fail trigger a reverse correction mechanism, backtracking to the corresponding step to adjust feature weights and regenerate the instruction. Simultaneously, the rule entries are bound to the model's feature extraction process, enabling the model to automatically activate the corresponding rule constraints when generating feature vectors, thus achieving anchoring of regulatory rules at the feature level.

[0081] After completing the construction of the regulatory rule anchoring inference chain, the next step is to jointly optimize the model parameters under the gradient coordination mechanism. Step S5's core is to jointly optimize the temporal contrastive loss and attention smoothing loss through the gradient coordination mechanism. Let the model's learnable shared parameters be θ, and the temporal contrastive loss be... Attention smoothing loss is Then the gradients of the two losses with respect to the parameter θ are respectively: The cosine similarity of two gradients is used to determine the direction conflict. If the cosine similarity is greater than 0, the gradient directions are consistent and can be directly updated. If the cosine similarity is less than or equal to 0, the gradient directions conflict and gradient projection needs to be performed.

[0082] If gradient directions conflict, smooth the loss gradient. Project to Orthogonal directions yield the gradient after conflict resolution. .

[0083] Then, a coordination coefficient λ is introduced to dynamically balance the contribution intensity of each gradient to the parameter update. The gradients of the translation task, the contrastive loss gradient, and the smoothing loss gradient after conflict resolution are fused to obtain the final total update gradient. ,in, For the translation task gradient.

[0084] After constructing the regulatory rule anchoring inference chain, the model parameter joint optimization phase under the gradient coordination mechanism is initiated. First, the set of loss functions for model optimization is determined, comprising three core loss functions: transaction instruction semantic loss, clearing instruction atomicity loss, and regulatory rule compliance loss. The transaction instruction semantic loss measures the semantic similarity between the generated transaction instruction and the user's original transaction intent; the clearing instruction atomicity loss measures the matching degree between fund transfer instructions and securities transfer instructions; and the regulatory rule compliance loss measures the degree of compliance between instructions and regulatory rules. The gradients of each loss function with respect to the model parameters are calculated: semantic loss gradient, atomicity loss gradient, and compliance loss gradient. Next, the consistency of these gradient directions is assessed by calculating the cosine similarity between gradients to measure directional differences. If the cosine similarity is positive, it indicates consistent gradient directions, allowing direct superposition for parameter updates; if the cosine similarity is negative, it indicates conflicting gradient directions, in which case the conflicting gradients are projected to directions orthogonal to the dominant gradient to eliminate conflicting components. Subsequently, an adaptive gradient modulation mechanism is introduced. By calculating the exponential moving average of each gradient norm, the contribution strength of each gradient component to parameter updates is dynamically determined, preventing a single loss function from dominating the model optimization direction. For example, when the market is in a state of high volatility, the contribution weight of the compliance loss gradient is increased to ensure that the orders strictly comply with the risk control rules; when the user's trading intention is clear and the constraints are lenient, the contribution weight of the semantic loss gradient is increased to prioritize meeting the user's needs.

[0085] After gradient coordination and parameter updates are completed, the final transaction clearing instruction package is generated. This package integrates compliant trading instructions, fund transfer instructions, and securities transfer instructions that have passed regulatory rule verification. It also includes an instruction traceability information module, which contains the corresponding regulatory rule entry for each instruction, gradient adjustment records during parameter optimization, instruction generation timestamps, digital signatures, and other information. The instruction package format strictly adheres to the message standards of the China Securities Depository and Clearing Corporation Limited (CSDC) to ensure direct interface with the exchange's trading execution and clearing systems. After generating the instruction package, the system performs a final end-to-end verification, checking the completeness of instruction fields, logical consistency between instructions, and compliance with regulatory rules. Instruction packages that pass verification are submitted to the execution module; those that fail trigger a manual review process.

[0086] Exemplary system: A large-scale model-based intelligent auxiliary system for securities trading and clearing includes the following modules: The data synchronization and preprocessing module receives and integrates user trading intentions and real-time market status to construct standardized multimodal input data. This includes: a user intention receiving unit, which receives user-inputted trading intention descriptions via a natural language interface and calls a financial-specific word segmenter for lexical processing; a real-time status acquisition unit, which synchronously acquires real-time market data, account balances, and deliverable securities holdings related to the current trading intention via an internal data bus and external data interface; a data standardization and feature encoding unit, which performs dynamic sliding window normalization on numerical status information, embeds entities into discrete information, and executes pre-clearing logic calculations to generate feature vectors with implicit business constraints; and a multimodal input construction unit, which concatenates and projects the processed text lexical sequence with the status feature vectors to generate a comprehensive input tensor containing complete context.

[0087] The semantic understanding and feature extraction module is used to deeply analyze financial semantics from the fused input data and generate high-dimensional feature representations that imply intent and constraints, including: The large-scale encoder unit employs a Transformer architecture model pre-trained on financial domain corpora, and performs deep interaction and feature fusion on the input tensor through a multi-layer self-attention mechanism; the financial knowledge enhancement unit embeds financial business knowledge graph parameters, and performs semantic calibration and enhancement on professional terms and business logic during feature extraction; the constraint gating fusion unit dynamically and controllably injects hard constraint features such as funds and positions into the main semantic feature stream through a gating linear unit mechanism, realizing the pre-positioning of risk control logic; and the semantic vector generation unit extracts compact and robust initial transaction semantic vectors from the complex state sequence output by the encoder through a task-aware attention pooling mechanism.

[0088] The compliant trading instruction generation module is used to transform high-level semantic vectors into executable trading instructions that strictly comply with regulatory and business requirements, including: The system comprises the following units: a legal form rule base unit, which stores and maintains structured constraint entries compiled from securities regulations, exchange message standards, and brokerage risk control rules; a deterministic finite automaton compilation unit, which compiles the formal constraints in the rule base into deterministic finite automata (DFAs) with explicit state transitions, serving as a rigid framework for the decoding path; a restricted decoder unit, based on an improved Transformer decoder architecture, integrating a DFA mask layer, which performs legality mask filtering on the vocabulary at each time step of instruction generation based on the current state of the DFA, ensuring that the output token is strictly compliant; and an instruction verification and optimization unit, which performs multi-level verification of the generated initial instructions in terms of form, logic, and regulatory compliance, and reorders candidate instructions using semantic alignment loss to output the optimal compliant trading instruction.

[0089] The joint clearing instruction generation module is used to synchronously generate fund transfer and securities transfer instructions with inherent atomicity from shared semantic features, ensuring the delivery versus payment principle, including: The intent feature deconstruction unit decouples specific sub-vectors such as transaction target, quantity, price, and constraints from the initial transaction semantic vector through business-oriented attention pooling. The joint intent tensor construction unit merges the core sub-vectors with the embedding matrices of three dimensions: semantic intent, constraint state, and atomic rules, through dimensional expansion and feature mapping to construct a three-dimensional joint intent tensor. The atomicity constraint encoding unit transforms the business rules of payment versus securities transactions into mathematical constraint matrices and embeds them into the joint intent tensor through tensor operations, forcing logical coupling between fund and security features at the feature level. The dual-channel instruction synchronous generation unit, based on the embedded constraint joint intent tensor, fills the standard instruction templates for fund transfer and securities transfer in parallel, and ensures strict matching of the two instructions in terms of amount and time limit through atomic real-time verification and reverse adjustment mechanisms.

[0090] The regulatory anchoring and joint optimization module integrates and verifies all generated instructions, constructs a traceable inference chain, optimizes model parameters by coordinating multi-objective losses, and finally packages and outputs an executable liquidation instruction package, including: The regulatory rule anchoring inference chain unit transforms multi-source regulatory rules into structured inference items and binds them to the instruction verification and model feature activation process to achieve traceable verification of instruction compliance. The multi-objective gradient coordination unit dynamically calculates and analyzes the gradient directions of transaction semantic loss, settlement atomicity loss, and regulatory compliance loss. Through conflict projection and adaptive modulation mechanisms, it resolves optimization objective conflicts and achieves collaborative updates of model parameters. The transaction settlement instruction package assembly unit integrates compliant transaction instructions, fund transfer instructions, and securities transfer instructions that have passed full-link verification, adds a traceability information module containing rule basis and optimization records, and generates a final instruction package that conforms to industry message standards. The system feedback and iteration unit monitors the instruction execution results and feeds back execution feedback (such as success or rejection reasons) to the training process to continuously optimize the model parameters and rule base completeness in the above modules.

[0091] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. An intelligent auxiliary method for securities trading and clearing based on a large model, characterized in that, include: S1. Receive the user's input of a natural language description of the transaction intent, synchronously obtain the corresponding real-time constraint status information, and perform preprocessing; wherein, the real-time constraint status information includes: market data, account cash position status, and deliverable securities holding information; S2. Based on the large model, extract financial semantic feature representations from the natural language description of transaction intent and real-time constraint state information to generate an initial transaction semantic vector; S3. Based on the initial transaction semantic vector, generate compliant transaction instructions through a legally perceptible restricted decoder; S4. Based on the initial transaction semantic vector, construct a joint intent tensor in the feature representation layer to generate fund transfer instructions and securities transfer instructions; S5. By combining compliant trading instructions, fund transfer instructions, and securities transfer instructions, a regulatory rule anchoring inference chain is constructed. Through a gradient coordination mechanism, the parameter updates of the large model are jointly optimized to generate a transaction settlement instruction package.

2. The intelligent auxiliary method for securities trading and clearing based on a large model according to claim 1, characterized in that: In step S3, the legal form-aware restricted decoder is embedded with a deterministic finite automaton compiled from the rules of securities trading legal forms. When decoding each field of the generated instruction, the candidate tokens are filtered by the legality mask through the deterministic finite automaton to ensure that the structure and content of the generated instruction naturally conform to the formal requirements of the regulator and the exchange.

3. The intelligent auxiliary method for securities trading and clearing based on a large model according to claim 1, characterized in that: Step S4 includes: S41. Deconstruct the core features of the initial transaction semantic vector and extract the core sub-vectors related to funds and securities settlement; S42. Map the core sub-vectors to the semantic intent dimension, constraint state dimension, and atomic rule dimension to construct a three-dimensional joint intent tensor, and encode the atomicity requirement of payment against payment as a constraint matrix and embed it into the joint intent tensor. S43. Based on the joint intent tensor with embedded atomicity constraints, synchronously fill the fund transfer instruction template and the securities transfer instruction template to generate the fund transfer instruction and the securities transfer instruction with logical coupling.

4. The intelligent auxiliary method for securities trading and clearing based on a large model according to claim 1, characterized in that: The preprocessing process in step S1 includes: performing dynamic standardization on the numerical real-time constraint status information to eliminate dimensional differences and retain the relative strength of the data relative to recent historical fluctuations; performing pre-liquidation logic calculations to embed the pre-verification of liquidation rules into the data input stage by calculating the virtual available position after deducting frozen funds and estimated funds for pending orders from the total account balance, and the sellable position after deducting pledged and restricted quantities.

5. The intelligent auxiliary method for securities trading and clearing based on a large model according to claim 1, characterized in that: The gradient coordination mechanism in step S5 includes: calculating the gradients of the semantic loss of trading instructions, the atomicity loss of clearing instructions, and the compliance loss of regulatory rules; dynamically detecting and modulating conflicts between different gradient directions; and ensuring the collaborative optimization of model parameters under multi-objective constraints through adaptive weighting.

6. The intelligent auxiliary method for securities trading and clearing based on a large model according to claim 1, characterized in that: The regulatory rule anchoring inference chain in step S5 transforms securities regulatory regulations, exchange business rules, and brokerage risk control rules into structured inference items and embeds them into the instruction verification and model feature activation process, thereby achieving traceability and mandatory compliance constraints on the instruction generation logic.

7. The intelligent auxiliary method for securities trading and clearing based on a large model according to claim 1, characterized in that: Step S2 includes: constructing a gating fusion mechanism through gating linear units, injecting the hard constraint indicators obtained from the pre-clearing logic into the semantic feature stream, and dynamically suppressing the transmission of the original transaction intent features according to the severity of the real-time constraints, so as to realize the semantic risk control logic of constraints being superior to intents.

8. The intelligent auxiliary method for securities trading and clearing based on a large model according to claim 1, characterized in that: The extraction of financial semantic features in step S2 includes: introducing a knowledge graph-based semantic enhancement mechanism, utilizing financial professional knowledge parameters stored within the large model to perform semantic calibration and vector representation enhancement on professional terms, security codes, and business logic in the transaction intent.

9. The intelligent auxiliary method for securities trading and clearing based on a large model according to claim 1, characterized in that: The generation of the transaction clearing instruction package in step S5 includes: integrating transaction instructions, fund instructions and securities instructions that have passed the end-to-end compliance verification, and attaching a traceability information module containing regulatory rule entries, gradient adjustment records and digital signatures to generate a standardized data package that conforms to the message standards of China Securities Depository and Clearing Corporation Limited.

10. An intelligent auxiliary system for securities trading and clearing based on a large model, wherein the intelligent auxiliary method for securities trading and clearing based on a large model according to any one of claims 1-9 is characterized in that, include: The data synchronization and preprocessing module is used to receive and integrate user trading intentions and real-time market status, perform pre-clearing logic calculations, and construct a multimodal integrated input tensor. The semantic understanding and feature extraction module is used to parse financial semantics through a large model encoder and use a gating fusion mechanism to implement risk control logic that prioritizes constraints over intentions, generating an initial transaction semantic vector. The compliant transaction instruction generation module is used to transform semantic vectors into compliant transaction instructions that meet the requirements of legal form using a constrained decoder based on deterministic finite automata constraints. The joint clearing instruction generation module is used to construct a joint intent tensor and embed atomicity constraint encoding, and synchronously and in parallel generate fund transfer instructions and securities transfer instructions that meet the atomicity requirements of delivery versus payment. The regulatory anchoring and joint optimization module is used to construct a regulatory rule anchoring inference chain, optimize model parameters through a multi-objective gradient coordination mechanism, and package and output a transaction clearing instruction package with audit traceability information.

Images