Methods and systems for dynamic risk control of AI robots for financial transactions

By constructing a twin graph of the trading risk status and a dynamic risk envelope for the financial trading AI robot, the problem of poor adaptability of risk boundaries is solved, and the accuracy of dynamic prevention and control and risk intervention throughout the entire cycle is improved.

CN123071073APending Publication Date: 2026-11-10BEIJING YONGFEI CHANGKONG TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611163892.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-03
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies in financial trading AI robots have poor adaptability to risk boundaries, lack a dynamic prevention and control mechanism throughout the entire cycle, and are difficult to adapt to market fluctuations and changes in trading status, resulting in misjudgments, omissions, and delayed responses. They are unable to effectively combine predictive uncertainty to dynamically determine risk boundaries.

Method used

By acquiring multi-source operational data from financial trading AI robots, a trading status event stream is generated, a trading risk status twin graph is constructed, the dynamic risk context and envelope are determined, the risk trajectory of trading actions is deduced, and safe action projection and risk intervention are carried out under the constraints of the dynamic risk envelope, forming a closed-loop prevention and control process.

Benefits of technology

It improves the adaptability of risk boundaries, enhances the ability to prevent and control risks throughout the entire cycle, and improves the accuracy of risk intervention and the security of transaction execution.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123071073A_ABST
    Figure CN123071073A_ABST
Patent Text Reader

Abstract

This application provides a method and system for dynamic risk control of AI robots in financial transactions. The method includes: acquiring multi-source data, performing time-series correction and reliability assessment to form a transaction state event stream; constructing a transaction risk state twin graph from the event stream, representing the interaction relationships and forming a dynamic risk context; generating a risk budget, propagation constraints, and a dynamic risk envelope by combining the dynamic risk context, account risk boundaries, and interaction relationships; acquiring candidate actions, constructing risk scenarios, and deducing states to form candidate risk trajectories; fusing risk trajectories with prediction uncertainty to form risk prediction results, comparing them with the risk envelope to determine the feasible region; performing safe action projection on candidate actions exceeding the feasible region to form target transaction actions; executing the target action, comparing the two types of trajectories, propagating intervention according to deviation, and updating the state twin graph and risk envelope. This application can improve the adaptability of risk boundaries, enhance the full-cycle control capability, and improve the accuracy of risk intervention.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the fields of financial technology and artificial intelligence risk control technology, and in particular to a method and system for dynamic risk prevention and control of AI robots for financial transactions. Background Technology

[0002] As artificial intelligence technology is increasingly applied in quantitative trading and automated order execution, financial trading AI robots are gradually taking on tasks such as market analysis, strategy decision-making, and trade execution. To mitigate financial losses caused by market volatility, strategy distortion, or abnormal execution, risk control systems are typically configured to verify the risk of trading orders, account positions, and fund status.

[0003] Existing technologies mostly employ fixed thresholds, static rule bases, or preset risk scoring models to determine transaction amounts, position ratios, price deviations, and loss levels, with a focus on risk verification after the transaction is completed. While some solutions introduce machine learning models to identify anomalies, the model parameters and risk boundaries are usually fixed in the long term, resulting in limited interaction between the risk control system and the trading strategy and order execution modules, making it difficult to adjust synchronously according to real-time market conditions.

[0004] The aforementioned methods are ill-suited to the non-stationary environment of market volatility, liquidity, and constantly changing trading hours. They are prone to misjudgment, omissions, and delayed responses during extreme market conditions, high-frequency anomalies, or sudden changes in trading status. Furthermore, existing technologies lack a full-cycle control mechanism covering pre-trade prediction, in-trade monitoring, and post-trade correction. They cannot analyze the risk propagation and cumulative effects between different trading states, nor can they dynamically determine risk boundaries based on predictive uncertainty. For over-limit transactions, they typically only allow or block them, lacking the ability to correct actions while maintaining strategic intent. Summary of the Invention

[0005] In view of this, the embodiments of this application provide a method and system for dynamic risk prevention and control of AI robots for financial transactions, in order to solve the problems of poor risk boundary adaptability, insufficient dynamic prevention and control throughout the entire cycle, and lack of adaptive calibration in risk intervention in the existing technology.

[0006] A first aspect of this application provides a method for dynamic risk control of a financial trading AI robot, comprising: acquiring multi-source operational data of the financial trading AI robot, generating a trading state event stream after time-series correction and credibility assessment; constructing a trading risk state twin based on the trading state event stream to represent the interaction between market state, strategy state, trading action, and account state, and generating a dynamic risk context; determining risk budget and propagation constraints based on the dynamic risk context, account risk boundary, and interaction, and generating a dynamic risk envelope; acquiring candidate trading actions, constructing a risk deduction scenario matching the dynamic risk context, deducing state evolution based on the trading risk state twin, and generating candidate risk trajectories; generating risk prediction results based on candidate risk trajectories and prediction uncertainties, and comparing them with the dynamic risk envelope to determine the feasible domain of trading actions; when a candidate trading action exceeds the feasible domain of trading actions, projecting a safe action under the constraints of the dynamic risk envelope to generate a target trading action; executing the target trading action, performing deviation analysis between the trading execution trajectory and the candidate risk trajectory, implementing risk intervention based on deviation propagation, and updating the trading risk state twin and dynamic risk envelope based on trading feedback.

[0007] A second aspect of this application provides a dynamic risk control system for financial trading AI robots, comprising: an acquisition module for acquiring multi-source operational data of the financial trading AI robot, generating a trading state event stream after time-series correction and credibility assessment; a construction module for constructing a trading risk state twin graph representing the interaction between market state, strategy state, trading actions, and account state based on the trading state event stream, and generating a dynamic risk context; a determination module for determining risk budget and propagation constraints based on the dynamic risk context, account risk boundaries, and interaction relationships, and generating a dynamic risk envelope; and a deduction module for acquiring candidate trading actions and constructing a dynamic risk envelope. The system employs a risk scenario modeling approach, matching the risk context and extrapolating state evolution based on a transaction risk state twin graph to generate candidate risk trajectories. A comparison module generates risk prediction results based on candidate risk trajectories and prediction uncertainties, comparing them with the dynamic risk envelope to determine the feasible domain of the transaction action. A generation module projects a safe action under the constraints of the dynamic risk envelope when a candidate transaction action exceeds the feasible domain, generating a target transaction action. An update module executes the target transaction action, performs deviation analysis between the transaction execution trajectory and the candidate risk trajectory, implements risk intervention based on deviation propagation, and updates the transaction risk state twin graph and dynamic risk envelope based on transaction feedback.

[0008] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: By acquiring multi-source operational data from a financial trading AI robot, and generating a trading state event stream after time-series correction and credibility assessment, this application constructs a trading risk state twin graph representing the interaction between market state, strategy state, trading actions, and account state based on the trading state event stream, and generates a dynamic risk context. Based on the dynamic risk context, account risk boundaries, and interaction relationships, risk budgets and propagation constraints are determined, generating a dynamic risk envelope. Candidate trading actions are acquired, and risk simulation scenarios matching the dynamic risk context are constructed. State evolution is deduced based on the trading risk state twin graph, generating candidate risk trajectories. Risk prediction results are generated based on the candidate risk trajectories and prediction uncertainties, and compared with the dynamic risk envelope to determine the feasible domain of trading actions. When a candidate trading action exceeds the feasible domain, a safe action projection is performed under the constraints of the dynamic risk envelope to generate a target trading action. The target trading action is executed, and deviation analysis is performed between the trading execution trajectory and the candidate risk trajectory. Risk intervention is implemented based on deviation propagation, and the trading risk state twin graph and dynamic risk envelope are updated based on trading feedback. This application can improve the adaptability of risk boundaries, enhance full-cycle prevention and control capabilities, and improve the accuracy of risk intervention. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, 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 of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating the dynamic risk control method for AI robots in financial transactions provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of the AI ​​robot risk dynamic prevention and control system for financial transactions provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0011] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0012] Existing risk control technologies for financial trading AI robots typically employ fixed thresholds, static rule bases, or pre-set risk scoring models to verify transaction amounts, position ratios, price deviations, and loss levels. While some existing solutions introduce machine learning models to identify abnormal transactions, the data interaction between the risk model and trading strategies, order execution, and account status remains relatively fragmented. Risk assessment is mostly concentrated on compliance checks after the trading instruction is generated or the transaction is completed, and a dynamic control mechanism that runs through trading decisions, action execution, and result feedback has not yet been formed.

[0013] Because financial markets are inherently non-stationary, market volatility, liquidity levels, trading session characteristics, and account status are constantly changing, making it difficult for fixed risk thresholds to accurately reflect the actual risks under different market conditions. Furthermore, existing technologies typically assess market risk, strategy risk, and execution risk independently, making it difficult to identify the risk propagation and amplification processes between different trading states, and failing to fully consider the uncertainty of risk prediction results. In scenarios such as extreme market conditions, high-frequency fluctuations, order delays, or strategy output deviations, risk misjudgments, omissions, and delayed interventions are prone to occur; for trading actions exceeding risk limits, the only options are usually to allow or block them, making it difficult to perform safe corrections while maintaining the strategy's intent.

[0014] To address the aforementioned technical issues, this application first acquires multi-source operational data from a financial trading AI robot. After time-series correction and credibility assessment, a trading state event stream is generated. Based on this event stream, a trading risk state twin graph representing the interactions between market state, strategy state, trading actions, and account state is constructed, forming a dynamic risk context. Subsequently, risk budgets and propagation constraints are determined based on the dynamic risk context, account risk boundaries, and state interactions, generating a dynamic risk envelope that changes with market and account state. For candidate trading actions output by the AI ​​robot, a risk deduction scenario matching the current dynamic risk context is constructed. The trading risk state twin graph is used to deduce the state evolution caused by the trading actions, generating candidate risk trajectories.

[0015] Furthermore, this application generates risk prediction results with credible boundaries based on candidate risk trajectories and prediction uncertainties, and compares them with the dynamic risk envelope to determine the feasible region of the trading action. When a candidate trading action exceeds the feasible region, a safe action projection is executed under the constraints of the dynamic risk envelope to generate a target trading action that has a small deviation from the original trading intention and meets the risk constraints. During the execution of the target trading action, a deviation analysis is performed between the actual trading execution trajectory and the candidate risk trajectory, the deviation propagation is tracked along the trading risk state twin graph, and risk intervention is implemented based on the impact of the deviation. Simultaneously, the trading risk state twin graph and the dynamic risk envelope are updated using trading feedback, thereby forming a closed-loop prevention and control process that connects pre-trade risk deduction, in-trade dynamic intervention, and post-trade feedback updates.

[0016] Through the above technical solutions, this application can dynamically adjust the risk boundary according to market conditions, strategy operation status, and account capacity, thereby improving the adaptability of the risk boundary to non-stationary trading environments; it can identify the risk propagation and superposition effects between trading states, enhancing the dynamic prevention and control capabilities throughout the trading cycle; and it can use predictive uncertainty and safe action projection to make credible judgments and within-constraint corrections on candidate trading actions, thereby improving the accuracy of risk intervention and the security of trading execution.

[0017] The technical solution of this application will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0018] Figure 1 This is a flowchart illustrating the dynamic risk control method for AI robots used in financial transactions provided in this application. Figure 1 As shown, the method may specifically include: S101: Acquire multi-source operational data of the financial trading AI robot, and generate a transaction status event stream after time-series correction and credibility assessment; S102, Based on the transaction state event flow, construct a transaction risk state twin graph that represents the interaction between market state, strategy state, transaction action and account state, and generate dynamic risk context; S103, determine the risk budget and propagation constraints based on the dynamic risk context, account risk boundaries and interaction relationships, and generate a dynamic risk envelope; S104: Obtain candidate transaction actions, construct risk simulation scenarios that match the dynamic risk context, simulate state evolution based on the transaction risk state twin graph, and generate candidate risk trajectories; S105: Generate risk prediction results based on candidate risk trajectories and prediction uncertainties, and compare them with dynamic risk envelopes to determine the feasible domain of trading actions; S106, When the candidate transaction action exceeds the feasible domain of the transaction action, a safe action projection is performed under the dynamic risk envelope constraint to generate the target transaction action; S107, execute the target transaction action, perform deviation analysis between the transaction execution trajectory and the candidate risk trajectory, implement risk intervention based on deviation propagation, and update the transaction risk status twin and dynamic risk envelope based on transaction feedback.

[0019] In some embodiments, acquiring multi-source operational data of a financial trading AI robot, and generating a trading status event stream after time-series correction and credibility assessment includes: Acquire multi-source operational data reflecting the trading environment, robot operation, and account status, and establish data associations based on data sources and trading objects; Based on a unified transaction clock and event causal sequence, the time offset of multi-source running data is identified and corrected to generate time-series aligned data. Data credibility is generated based on data source quality, cross-source state consistency, and transmission timeliness. State conflicts of the same transaction object are resolved based on data credibility and event causal order. The time-series aligned data after conflict resolution is converted into transaction status events with associated transaction objects, occurrence times, status changes, and data credibility, and then aggregated to generate a transaction status event stream.

[0020] Specifically, multi-source operational data is used to simultaneously characterize the external trading environment, internal decision-making status, and account status of the financial trading AI robot. Trading environment data can be collected from market data access nodes, exchange interfaces, or liquidity monitoring nodes, and includes at least price sequences, order book depth, trading activity, and trading session status. Robot operational data can be output from the strategy engine, order manager, and execution interface, and includes at least strategy version, signal strength, target position, candidate trading actions, and instruction execution status. Account status data can be provided by the funds and position management nodes, and includes at least available funds, position changes, margin requirements, and risk limits. The system assigns source identifiers to each data source and establishes data associations using financial asset identifiers, account identifiers, strategy task identifiers, and order identifiers, enabling market changes, strategy outputs, order actions, and account responses related to the same trading task to form a continuous data chain.

[0021] During time-series calibration, the high-precision clock of the trading server is first selected as the unified trading clock. The offset between the local time of each data acquisition node and the unified trading clock is recorded. Then, the dynamic clock deviation is estimated using periodic synchronization messages, interface round-trip latency, and confirmed trade reports. For an event chain where market conditions change first, strategies are calculated later, orders are sent, and trade reports are finally returned, sequential constraints are established according to the causal order of the events. If the local time of a certain order sending record is earlier than the time of the strategy signal that triggered the order, the corresponding timestamp is corrected according to the node clock deviation and link transmission latency. If multiple events cannot be directly sorted within the allowable error range, the relative order is determined using event numbers, parent-child task relationships, and state transition conditions. After calibration, data of different frequencies are mapped to a unified time axis, retaining the original time, calibration time, and calibration magnitude to form time-series aligned data.

[0022] Data reliability is determined by source quality, cross-source consistency, and transmission timeliness. Source quality is generated based on the node's historical packet loss rate, anomaly rate, verification pass rate, and interface stability. Cross-source consistency is obtained by comparing the state values, change directions, and related events of the same transaction object in different data sources to see if they match. Transmission timeliness is determined by the deviation of data arrival delay from the allowable latency of the business. The system normalizes and weights these indicators to obtain a data reliability between 0 and 1. When a state conflict occurs, states that do not meet the preconditions are first excluded according to the causal order of events, and then the reliability of the conflicting data is compared. When the reliability difference reaches the conflict judgment threshold, the current state is updated using highly reliable data. When the difference does not reach the threshold, the candidate state is retained and confirmed with a delay based on subsequent transaction reports, fund changes, or position changes, to avoid a single abnormal data directly changing the transaction state.

[0023] Furthermore, after conflict resolution, the system encapsulates each time-aligned data entry into a transaction status event. Each transaction status event is associated with at least the transaction object, event type, occurrence time, state change amount, source identifier, and data reliability, and the event inheritance relationship is recorded through the preceding event identifier. For fine-grained market data that occurs consecutively and affects the same object, status change events can be generated according to a preset aggregation window; for key events such as strategy switching, order submission, trade confirmation, or risk limit changes, they are recorded independently. Subsequently, all transaction status events are aggregated according to the corrected occurrence time and causal order of events, forming a continuously appending transaction status event stream for use in subsequent construction of transaction risk status twin graphs and generation of dynamic risk contexts.

[0024] For example, an AI robot receives a signal of a rapid increase in the price of a target asset at 10:15 AM and generates a buy order. The market data node records the time as 10:15:00:120, the strategy node records the signal time as 10:15:00:105, and the order node records the sending time as 10:15:00:098. The system detects that the order was sent before the strategy signal, which does not conform to the causal sequence of events. It is found that the order node's local clock is 35 milliseconds fast. After correction, the order sending time is 10:15:00:133, and the event sequence is restored to market change, strategy signal, order sending. Subsequently, the order manager shows that the order has been executed, but the trading interface has not yet returned a completion report. Since the current data reliability of the order manager is 0.72 and the historical reliability of the trading interface is 0.94, the system does not confirm the execution status for now; once the account holdings increase and the trading interface sends the completion report, the order status will be confirmed as executed, and the corresponding event reliability will be updated to 0.96.

[0025] The above methods can transform heterogeneous, inconsistent, and potentially conflicting transaction data into a continuous, time-series-based, and credibility-marked transaction status event stream. This reduces the interference of erroneous time relationships and low-quality data on risk assessment and provides a reliable data foundation for subsequent risk propagation analysis, transaction action simulation, and dynamic intervention.

[0026] In some embodiments, a transaction risk state twin graph representing the interaction between market state, strategy state, trading actions, and account state is constructed based on the transaction state event flow, and a dynamic risk context is generated, including: Based on the transaction object identifier and the causal sequence of events, the transaction status event flow is mapped to the corresponding market status, strategy status, transaction action and account status status nodes; Based on the temporal response and state change correlation between each state node, the direction of action, intensity of action, response delay and relationship credibility are determined, action edges between state nodes are established, and a transaction risk state twin graph is generated. Based on the sliding time window, update state nodes and action edges to identify market state transitions and strategy execution offsets; By integrating updated market conditions, strategy status, account status, and relationship credibility, a dynamic risk context is generated.

[0027] Specifically, the trading risk state twin graph is used to organize market changes, strategy decisions, trade execution, and account responses scattered throughout the trading state event stream into a graph structure with causal direction and dynamic parameters. The system first merges trading state events according to the trading object identifier, grouping events corresponding to the same financial asset, account, strategy task, and order into a related event set, and then mapping them to different state nodes based on event type. Market state nodes record price trends, volatility, liquidity, and market imbalances; strategy state nodes record strategy patterns, signal strength, target position size, and operational confidence; trading action nodes record action direction, order size, execution progress, and transaction status; and account state nodes record capital occupation, open interest, margin, and risk limits. Each node uses a state vector to store its current value, rate of change, update time, and event confidence, and establishes candidate interaction relationships between nodes through the causal sequence of events.

[0028] Action edges are used to describe the driving process of one state change on another. The system extracts the state sequences of candidate node pairs within a preset historical window and compares the temporal relationship between the changes of the preceding node and the responses of the following node. When the strategy signal strengthens after a market price jump, and this phenomenon repeats in multiple trading cycles, an action edge can be established from the market state node to the strategy state node; when trading actions cause changes in positions and capital occupation, an action edge can be established from the trading action node to the account state node.

[0029] The direction of action is determined by stable sequential response relationships, the intensity of action can be calculated based on the magnitude of the change in the subsequent state caused by a unit change in the preceding state, and the response delay can be determined based on the time difference when the two types of state changes reach their maximum correlation. Relationship credibility is generated by comprehensively considering the number of samples, state correlation stability, event credibility, and reverse verification results. For candidate relationships that appear only in a small number of samples or whose causal order is unstable, the relationship credibility is reduced or they are temporarily omitted from the graph structure to reduce the propagation of errors caused by accidental correlations.

[0030] Furthermore, after generating the initial graph structure, the system continuously receives new trading state events according to a sliding time window and updates the node state and edge parameters. The sliding time window can be set to a continuous number of minutes or several trading cycles based on the trading frequency, with adjacent windows retaining some overlapping data to maintain the continuity of state changes. When updating a node, events within the window are weighted and aggregated according to their occurrence time and confidence level to obtain the latest node state; when updating an edge, recent response samples are given higher weights, and smoothing constraints are set in conjunction with historical parameters to avoid sudden changes in the intensity of the effect or the response delay caused by a single anomaly. If the market volatility, liquidity structure, or price change pattern exceeds the migration threshold relative to the previous window, it is identified as a market state migration; if the strategy signal distribution, action output frequency, or target position deviates from the historical baseline of the corresponding strategy version, it is identified as a strategy operation deviation, and the direction and duration of the deviation are recorded.

[0031] Dynamic risk context describes the overall risk conditions at the current trading moment. The system extracts the current market state, strategy state, and account state from the updated graph structure and selects local subgraphs with reachable propagation paths to candidate trading actions. Subsequently, node states are corrected according to data reliability, and impact edges are attenuated according to relational reliability. Context features are generated by combining the degree of market state migration, strategy execution offset, and remaining account capacity. The dynamic risk context may include the current risk stage, main risk sources, key propagation paths, risk amplification nodes, and state reliability, and is associated with the trading object identifier and generation time for use in dynamic risk envelope generation and candidate trading action deduction.

[0032] For example, a trading robot executes a short-term buying strategy for a target asset. Within a sliding window, the asset price increases from 0.2% to 0.8% per minute, the available trading volume decreases by 35%, the strategy signal strength rises from 0.62 to 0.91, and large buy orders are continuously executed, increasing the account's open interest exposure from 48% to 71% of the set limit. The system maps price increases and market contraction to market status nodes, signal enhancement to strategy status nodes, buy orders to trading action nodes, and increased open interest exposure to account status nodes.

[0033] Historical response data shows that the trading deviation widens approximately 1.5 seconds after the order book contraction. A 10% increase in buying activity leads to an average 7% increase in open interest exposure, thus updating the strength and latency of the corresponding action edge. Because the current volatility structure significantly deviates from the recent baseline, the system identifies a market state shift; simultaneously, the strategy output size consistently exceeds the normal range for the same version, identifying a strategy deviation. After integrating the states and relationship credibility of each node, a dynamic risk context is generated, characterized primarily by decreased liquidity, aggressive strategies, and narrowing account capacity.

[0034] In this way, transaction state events can be transformed into an updatable and traceable risk impact network, accurately representing the dynamic relationship between the market, strategy, action and account, and timely identifying market state transitions and strategy operation deviations, providing a structured and reliable state basis for risk propagation constraints, dynamic risk envelopes and transaction action deduction.

[0035] In some embodiments, risk budgets and propagation constraints are determined based on dynamic risk context, account risk boundaries, and interaction relationships, and a dynamic risk envelope is generated, including: Based on the dynamic risk context, market disturbances, strategy deviations, and account carrying status are determined, and a total risk budget is generated by combining the account risk boundary. Identify risk propagation paths along the relationships in the transaction risk status twin graph, and determine the propagation contribution and coupling amplification degree based on the strength of the effect, response delay, and relationship credibility; The total risk budget is allocated based on the propagation contribution, and the corresponding risk sub-budget is adjusted according to the degree of coupling amplification to generate propagation constraints; By combining risk sub-budgets, propagation constraints, and uncertainty safety margins, the boundary of state changes that can be formed by candidate transaction actions is determined, and a dynamic risk envelope is generated.

[0036] Specifically, the dynamic risk envelope describes the range of market response, strategy response, and account status changes allowed to occur under the current trading environment and account carrying capacity. The system first extracts the degree of market disturbance, strategy deviation, and account carrying capacity from the dynamic risk context. The degree of market disturbance is generated based on the deviation of price fluctuations, liquidity contraction, order book imbalances, and trading shocks from the recent baseline; the degree of strategy deviation is generated based on the strategy signal strength, action frequency, target position size, and the difference between the actual output and the strategy version baseline; and the account carrying capacity is determined based on available funds, open positions, margin requirements, and remaining risk limits.

[0037] After unifying the risk direction and normalizing the intervals for various statuses, the current risk occupancy is formed. This is then combined with the account's preset maximum loss limit, position limit, capital occupation limit, and liquidity tolerance boundary to determine the account's risk boundary. The system uses the available space between the account's risk boundary and the current risk occupancy as the basic risk capacity, and reduces the basic risk capacity according to market disturbances and strategy deviations to generate the total risk budget available for allocation in the current trading cycle.

[0038] Risk propagation paths can be searched starting from candidate trading action nodes in the trading risk state twin graph, extending along edges with effective directions and sufficient relational credibility to strategy state, market state, and account state nodes. For each propagation path, the system calculates the state change propagation ratio based on the strength of each edge in the path, determines whether the risk will be fully manifested within the target trading period based on response latency, and uses relational credibility to attenuate the propagation ratio. The propagation contribution can be jointly determined by the path propagation ratio, the sensitivity of the target state, and the degree of risk manifestation.

[0039] In some examples, if multiple propagation paths converge to the same account state node, or if the same preceding state simultaneously amplifies multiple subsequent risk states, the degree of coupling amplification is calculated based on the degree of path overlap, the consistency of change direction, and the degree of temporal overlap of effects. Paths with short response delays, high effect strength, and high relationship credibility are assigned a larger propagation contribution; paths with unstable relationships or risks that cannot be fully manifested in the current period are assigned a reduced propagation contribution.

[0040] Furthermore, after completing the propagation analysis, the system allocates the total risk budget according to the proportion of each risk path's propagation contribution to the total propagation contribution, forming risk sub-budgets corresponding to price shocks, trading deviations, open positions, and capital occupation. The risk sub-budgets are not fixed; when a path exhibits significant coupling amplification, the available risk sub-budget for that path is compressed according to the degree of coupling amplification, and the released amount is reserved as a buffer budget rather than transferred to high-risk paths. Propagation constraints may include the maximum allowed state increment for a single path, the upper limit of joint changes for multiple related nodes, the minimum delay before risk reaches key account nodes, and mutual exclusion conditions between high-risk paths. This prevents candidate trading actions from exceeding the overall account capacity after being superimposed by multiple paths, even if each individual risk constraint is met.

[0041] When generating the dynamic risk envelope, the system uses each risk sub-budget as its capacity basis, defines the direction of state changes and joint relationships through propagation constraints, and sets uncertainty safety margins based on model prediction errors, data reliability, and relationship reliability. When prediction errors are large, data reliability is low, or the interaction relationships are unstable, the safety margin is increased and the state change boundaries are tightened; when the risk state is stable and historical prediction coverage is good, the boundaries can be appropriately relaxed within the account risk boundaries. The system separately determines the allowed price impact boundaries, transaction deviation boundaries, position change boundaries, and capital occupation boundaries for candidate trading actions, and forms a multi-dimensional envelope structure through propagation path association. The dynamic risk envelope updates with the sliding time window, automatically shrinking when market states shift, strategy deviations intensify, or the remaining account capacity decreases, and gradually releasing when the risk state recovers.

[0042] For example, an account has a maximum allowed daily risk loss of 1 million yuan, of which 580,000 yuan is currently used, leaving a basic risk capacity of 420,000 yuan. The target asset enters a state of high volatility and decreased liquidity, with a market disturbance reduction factor of 0.75. The strategy output scale is higher than the baseline, and the strategy deviation reduction factor is 0.8, thus determining a total risk budget of 252,000 yuan. Propagation analysis shows that the path from candidate buy action to trading impact, to transaction deviation, and then to position exposure contributes 55%, while the path from liquidity contraction to capital occupation contributes 45%, corresponding to initial risk sub-budgets of 138,600 yuan and 113,400 yuan, respectively. Since both paths increase margin usage, the coupling amplification degree reaches 0.2. The system further compresses the relevant risk sub-budgets and increases the safety margin, ultimately limiting the increase in position caused by buy action to no more than 4% of the account's net assets, the increase in capital occupation to no more than 12% of available funds, and the expected transaction deviation to no more than 0.35% of the current midpoint, thus forming the dynamic risk envelope at the current moment.

[0043] The above approach allows for the dynamic configuration of risk capacity based on real-time market disturbances, changes in strategy operation, and account capacity. It also incorporates the propagation and coupling effects between states into the boundary generation process, avoiding the problem that fixed thresholds cannot reflect the superposition of risks. This provides a constraint basis that can be updated with the transaction status for candidate transaction action deduction, feasible domain determination, and safe action projection.

[0044] In some embodiments, candidate transaction actions are obtained, a risk simulation scenario matching the dynamic risk context is constructed, the state evolution is simulated based on the transaction risk state twin graph, and candidate risk trajectories are generated, including: Acquire candidate trading actions output by the financial trading AI robot, associate them with corresponding strategy objectives and trading status, and generate trading intentions to be deduced; Extract state disturbance factors that affect transaction execution based on dynamic risk context, and generate risk simulation scenarios according to the direction of disturbance, magnitude of change and coupling relationship; The transaction intentions to be simulated are loaded into each risk simulation scenario. Based on the direction of action, intensity of action and response delay in the transaction risk state twin, the state changes caused by the transaction actions are propagated in time intervals. By aggregating state changes, risk propagation paths, and the proximity of risk boundaries across different time periods, candidate risk trajectories are generated.

[0045] Specifically, candidate trading actions represent the trading control content that the financial trading AI robot intends to execute at the current decision-making moment. The system reads the action direction, target quantity, order price, execution time limit, and order splitting method from the strategy engine or order generation module, and associates them with the strategy task that triggers the action, target profit, target position, and current trading status. The current trading status may include the latest market status, strategy operation stage, account holdings, and available funds. Subsequently, the candidate trading actions, strategy objectives, and trading status are combined into a trading intention to be deduced, recording action parameters, allowable adjustment range, execution priority, and strategy objective maintenance conditions. This allows subsequent risk deduction not only to determine whether the action generates risk, but also to identify the correspondence between the risk source and the trading intention.

[0046] The risk simulation scenario is used to model the possible execution results of candidate trading actions under different state disturbances. The system filters state disturbance factors that have an impact path on candidate trading actions from the dynamic risk context, including price changes, increased volatility, liquidity contraction, execution delays, strategy signal offsets, and changes in account capacity. For each disturbance factor, the direction of the disturbance is determined based on the recent direction of state change, the magnitude of the change is determined based on historical percentile intervals, the current rate of change, and the degree of state migration, and the coupling relationship is determined based on the correlation path in the trading risk state twin graph. The system can generate a baseline scenario, a forward disturbance scenario, and a stress disturbance scenario. The baseline scenario maintains the current state, the forward disturbance scenario amplifies the current main trend, and the stress disturbance scenario combines multiple high-risk disturbances. Mutually exclusive conditions are set for disturbances with opposite directions that cannot occur simultaneously, and joint change conditions are set for disturbances that can reinforce each other, thus avoiding unreasonable scenarios generated by simple enumeration.

[0047] After loading the proposed trading intentions into each risk scenario, the system updates the trading risk status twin graph according to the preset simulation step size. In each time period, candidate trading actions are first written into the trading action node, and then propagated along the direction of influence to subsequent market status, strategy status, and account status nodes. The propagation increment can be determined based on the change in the status of the previous node and the intensity of its influence, and is written into the corresponding time period after the response delay; when the relationship reliability is low, the propagation increment is attenuated.

[0048] In some examples, if multiple propagation paths converge at the same state node during the same period, they are superimposed or canceled out according to the direction of change and coupling relationship. After the node state is updated, it continues to propagate to the next level node until the extrapolation time domain, the risk propagation termination condition, or the state boundary of a key account is reached. For incomplete transaction actions, subsequent action nodes can also be dynamically updated based on the remaining order volume and market liquidity to reflect the impact of segmented execution and unexecuted withdrawals on state evolution.

[0049] Candidate risk trajectories are generated according to the risk simulation scenarios. The system records node state values, state changes, trigger paths, propagation delays, and relationship reliability at each simulation period, and calculates the distance between the corresponding state and the dynamic risk envelope boundary. The degree of boundary proximity can be determined by the proportion of state change to the allowable change space; a proportion close to 1 indicates that the risk state is approaching the envelope boundary, while a proportion exceeding 1 indicates a boundary breach. The system connects the state changes in consecutive time periods in chronological order and marks the risk peak time, main propagation path, first approach to the boundary time, and boundary breach node, forming candidate risk trajectories that reflect the direction, speed, and accumulation degree of risk evolution.

[0050] For example, an AI robot plans to buy 80,000 units of a target asset using a limit order, with the strategy aiming to increase its position from 32% to 40% within 5 minutes. The current dynamic risk context indicates rising prices, decreasing market depth, a strong strategy signal, and limited remaining account funds. The system generates a baseline scenario, a forward disturbance scenario with continued price increases and a 20% decrease in liquidity, and a stress disturbance scenario with a 1.2% price jump, a 40% decrease in liquidity, and a delayed execution. In the stress disturbance scenario, the buy action initially causes a surge in trading volume, followed by a widening of the trading deviation approximately 2 seconds later, subsequently increasing holding costs and capital requirements. By the 3rd minute, capital requirements reach 92% of the allowable limit, and by the 4th minute, position exposure exceeds the corresponding risk sub-budget. The system marks the path from the trading action to the trading deviation, then to capital requirements and position exposure, as the primary risk propagation path. Candidate risk trajectories are formed by aggregating state changes and boundary proximity at various time points, providing input for subsequent risk prediction and trading action feasibility domain determination.

[0051] By using the above methods, risk scenarios with direction, magnitude, and coupling relationships can be constructed in combination with the current dynamic risk context before the actual execution of candidate transaction actions. The risk propagation process can be continuously deduced using a transaction risk state twin diagram, thereby improving the completeness and foresight of transaction risk evolution identification and providing a reliable basis for dynamic risk boundary comparison and safety action correction.

[0052] In some embodiments, risk prediction results are generated based on candidate risk trajectories and prediction uncertainties, and compared with dynamic risk envelopes to determine the feasible domain of trading actions, including: Extract the risk state peak, cumulative change and propagation intensity from candidate risk trajectories to generate trajectory risk characterization; Adaptive calibration of trajectory risk representation is performed based on historical extrapolation deviations matched with dynamic risk context to determine prediction intervals and corresponding prediction uncertainties. By integrating trajectory risk characterization, prediction interval, and prediction uncertainty, a risk prediction result containing risk prediction value and confidence boundary is generated. The credible boundary of the risk prediction result is compared with the state change boundary in the dynamic risk envelope item by item, and the set of trading actions that simultaneously satisfy each state change boundary is determined as the feasible region of the trading action.

[0053] Specifically, trajectory risk representation is used to compress the time-varying node states under different risk simulation scenarios into comparable risk features. For each candidate risk trajectory, the system divides risk dimensions according to market state, strategy state, trading action, and account state, and extracts state peaks, cumulative changes, and propagation intensity within the simulation time domain. The state peak can be the maximum state value after unifying risk directions or the maximum boundary occupancy ratio, used to represent the instantaneous risk ceiling; the cumulative change can be the weighted sum of risk increments in each time period according to duration, used to reflect continuous exposure and risk accumulation; the propagation intensity can be comprehensively generated based on the influence strength, coupling amplification degree, and relationship credibility in the main propagation path, used to describe the diffusion ability of candidate trading actions on subsequent states. The system can also record the peak occurrence time, the first approach to the risk boundary time, and the duration of high risk, and perform normalization on each feature to form a trajectory risk representation corresponding to the candidate trading action.

[0054] Since risk projection relies on state twins and perturbation scenarios, discrepancies may exist between the trajectory risk representation and the actual execution results. Therefore, the system uses historical projection deviations for adaptive calibration. Historical projection deviations are obtained by comparing historical candidate risk trajectories with their corresponding transaction execution trajectories on a time-by-time basis, including peak deviation, cumulative deviation, propagation delay deviation, and boundary occupancy deviation. Based on the market volatility phase, liquidity status, strategy execution offset, and account capacity level in the current dynamic risk context, the system selects calibration samples from historical samples that meet the matching criteria for state similarity, and configures sample weights according to state similarity, sample recentity, and historical data reliability. If a state transition occurs in the current market, the weight of the transitioned samples is increased, while the weight of samples in the stable phase is decreased to prevent early data from weakening the calibration effectiveness.

[0055] For the weighted calibration samples, the system statistically analyzes the distribution of positive and negative biases in each risk dimension. Positive bias indicates that the actual risk is higher than the projected risk, and is used to expand the upper boundary of the risk prediction value; negative bias indicates that the actual risk is lower than the projected risk, and is used to form the lower boundary. The system selects the bias quantiles according to a preset coverage level and superimposes them onto the trajectory risk representation to obtain the prediction interval for each risk dimension. Prediction uncertainty is generated based on the prediction interval width, the number of effective samples, the degree of bias dispersion, and the degree of state matching. When there are insufficient samples, large bias fluctuations, or a low degree of state matching, the prediction uncertainty is increased; when the sample coverage is sufficient and historical predictions are stable, the prediction uncertainty is reduced. The system can also continuously check the coverage ratio of the prediction interval to the actual state. When the actual coverage ratio is lower than the target ratio, the error boundary is widened; when it is consistently higher than the target ratio, the error boundary is gradually narrowed.

[0056] The risk prediction result consists of the central predicted value in the trajectory risk representation, the calibrated prediction interval, and the prediction uncertainty. The system uses the trajectory risk representation as the central predicted value and expands the prediction interval based on the prediction uncertainty to form a credible boundary. For states where larger values ​​indicate higher risk, a credible upper boundary is used for comparison; for states where smaller values, such as liquidity margin and available funds, indicate higher risk, a credible lower boundary is used for comparison. Each risk prediction result is associated with candidate trading actions, risk scenario projection, prediction time domain, and credibility, in order to trace the risk dimensions and scenarios that make the trading action infeasible.

[0057] When determining the feasible region of a trading action, the system compares the credible boundary with each of the position change boundary, capital occupation boundary, trading deviation boundary, and risk loss boundary in the dynamic risk envelope. Only when a candidate trading action satisfies all state change boundaries under all valid risk scenarios and the joint propagation constraint is not breached is it included in the feasible region. For continuously adjustable action parameters, action combinations with different trading volumes, execution rhythms, and price tolerance ranges can be generated near the candidate action to perform risk simulation and credible boundary comparison, thus forming a parameter region composed of multiple feasible actions. If an action only satisfies the boundary under the baseline scenario but exceeds the limit under the stress scenario, it is not included in the feasible region.

[0058] For example, the predicted center of exposure for a candidate buy action is 3.6% of the account's net assets. Under similar historical conditions, the 90th percentile positive deviation of the actual value from the projected value is 0.7%, and the credible upper bound after uncertainty expansion is 4.5%. The current dynamic risk envelope limits the increase in holdings to no more than 4%, therefore this action does not meet the holding boundary. The system adjusts the buy quantity sequentially to 90%, 80%, and 70% of the original action and re-projects. The credible upper bound for holding exposure corresponding to the 80% action is 3.8%, the credible upper bound for capital occupation is 10.6%, and the credible upper bound for transaction deviation is 0.31%, all of which do not exceed their respective boundaries, and the joint propagation constraint is met. Therefore, action combinations with 80% or lower risk are included in the feasible domain of trading actions.

[0059] The above method can convert candidate risk trajectories into risk prediction results with credible boundaries that have been calibrated by historical biases. It can also determine the feasible domain of trading actions under the condition of considering prediction errors and multidimensional risk constraints, reduce misjudgments caused by single-point prediction biases, and provide a clear and calculable constraint space for subsequent safe action projections.

[0060] In some embodiments, the trajectory risk representation is adaptively calibrated based on historical extrapolation deviations matched with the dynamic risk context to determine the prediction interval and the corresponding prediction uncertainty, including: Based on the similarity between the historical dynamic risk context and the current dynamic risk context, the historical extrapolation deviation is hierarchically matched to generate a state calibration sample set. Identify state transitions in dynamic risk contexts and adjust sample weights in the state calibration sample set based on the transition time between the sample's corresponding state and the current state. Based on the weighted historical extrapolation deviation, time-series conformal calibration is performed to determine the quantile error boundary corresponding to the trajectory risk characterization and generate the prediction interval. The quantile error boundary is corrected based on the historical coverage deviation of the prediction interval, and the prediction uncertainty is generated based on the correction magnitude and interval width.

[0061] Specifically, historical projection deviation is used to characterize the difference between the risk projection results and the actual transaction execution results. Historical deviation records can be generated according to the trading object, strategy version, market stage, and account status. Each record must include at least the historical dynamic risk context, trajectory risk representation, actual risk status, projection deviation for each risk dimension, and the time of occurrence. The system first performs unified feature processing on the historical and current dynamic risk contexts, converting market volatility, liquidity, price trends, strategy bias, account exposure, and data credibility into consistent state features, and then assigns matching weights based on the degree of influence of each feature on risk prediction.

[0062] The weighted distance between the historical state and the current state is then calculated, dividing the historical deviation records into a high-similarity layer, a neighboring similarity layer, and a supplementary sample layer. The high-similarity layer provides the primary calibration basis, the neighboring similarity layer increases sample coverage, and the supplementary sample layer participates in calibration only when there are insufficient effective samples, limiting its influence with a lower weight, thereby generating a state calibration sample set.

[0063] State transitions reflect the process of a market or strategy's operating state shifting from one stable mode to another. The system identifies transition events such as sudden increases in volatility, sharp drops in liquidity, price reversals, or changes in strategy output patterns by comparing the state center, rate of change, and relational structure within a continuous sliding window, and determines the timing of these transitions. For each sample in the state calibration sample set, the system calculates the relationship between the sample's state and the current state in the stages before and after the transition, as well as the time interval between the transitions. If the sample and the current state are in the same transition stage and the time interval is short, the sample weight is increased; if the sample is in the stage before the transition while the current state has already entered the stage after the transition, the sample weight is decreased based on the time interval and state difference; if historical samples have followed the same transition path as the current one, the weight is appropriately increased. Weights can also be adjusted in conjunction with sample freshness, data reliability, and the completeness of trade execution to prevent outdated samples or incomplete feedback from interfering with the current calibration.

[0064] Time-series conformal calibration is used to correct prediction errors without altering the temporal order of candidate risk trajectories or the direction of risk change. The system arranges the weighted historical projection deviations according to risk dimensions and prediction periods, obtaining sequences of peak deviation, cumulative deviation, propagation delay deviation, and boundary occupancy deviation. For each prediction period, weighted quantile errors are calculated based on the target coverage level, and anomalous jumps in error boundaries are eliminated through monotonic constraints between adjacent periods, ensuring that the upper error boundary corresponding to the continuous rise in risk does not shrink without justification. For samples where the peak occurrence time is offset, time axis alignment is first performed based on the propagation delay deviation, and then the corresponding quantile errors are calculated. The quantile error boundaries are superimposed on the central predicted value of the trajectory risk representation to form the upper and lower prediction boundaries of each risk dimension as they change over time, thereby generating the prediction interval.

[0065] To verify the reliability of the predicted interval, the system uses a rolling historical window to statistically analyze the proportion of actual risk states falling within the predicted interval and compares this proportion with the target coverage rate to obtain the historical coverage deviation. When the actual coverage rate is lower than the target coverage rate, it indicates that the interval is too narrow, and the quantile error boundary is widened according to the coverage gap and recent out-of-bounds deviation. When the actual coverage rate is consistently higher than the target coverage rate and the interval width is too large, the quantile error boundary is gradually narrowed while maintaining a minimum safety margin. The correction magnitude can be set with an upper limit for single-cycle variation to prevent calibration parameters from fluctuating drastically due to short-term anomalies. The prediction uncertainty is generated by combining the corrected interval width, coverage correction magnitude, number of effective samples, and state matching degree. The wider the interval, the larger the correction magnitude, the fewer the samples, or the lower the matching degree, the higher the prediction uncertainty.

[0066] For example, the target asset has transitioned from a low-volatility phase to a high-volatility and liquidity-contracting phase. The system filters 120 deviation samples from historical records, of which 38 are in the same high-volatility phase, 52 are in the adjacent migration phase, and the remaining samples belong to the low-similarity supplementary layer. The current state migration has occurred 8 minutes ago. The system increases the weight of recent samples after the migration and decreases the weight of samples before the migration. After time-series conformal calibration, the predicted value of the position exposure center corresponding to the candidate buy action is 3.4%, with an error boundary of 0.6% at the 90th percentile, forming an upper boundary of 4.0%. Rolling validation reveals that the historical coverage rate under similar states is only 84%, lower than the 90% target coverage rate. Therefore, the upper error boundary is revised to 0.8%, the predicted upper boundary is adjusted to 4.2%, and the prediction uncertainty is increased from 0.18 to 0.27. This result is used for subsequent comparison with the dynamic risk envelope to avoid underestimating risk due to an excessively high proportion of low-volatility samples before the migration.

[0067] By employing the above methods, historical projection biases can be utilized in a differentiated manner according to the current state characteristics and migration stage. Furthermore, by continuously correcting the prediction interval through time-series conformity preservation and coverage feedback, the adaptability and stability of the prediction credibility boundary under non-stationary trading environments can be improved, providing a more reliable basis for determining the feasible domain of trading actions based on uncertainty.

[0068] In some embodiments, when a candidate transaction action exceeds the feasible region of a transaction action, a safe action projection is performed under dynamic risk envelope constraints to generate a target transaction action, including: Based on the risk prediction results, identify the state change boundaries that candidate trading actions exceed, and determine the parameters of the actions to be corrected and the corresponding correction direction. The projected target is to minimize the strategic intent deviation between candidate trading actions and modified trading actions. The projected model is constructed with dynamic risk envelope and account risk boundary as constraints. The parameters of the action to be corrected are solved collaboratively to generate corrected transaction actions that fall within the feasible region of the transaction action; The risk simulation is re-executed for the corrective trading action, and the corrective trading action that satisfies the dynamic risk envelope is identified as the target trading action.

[0069] Specifically, the safety action projection is used to constrain the action parameters when candidate transaction actions do not meet the feasible domain of the transaction action. This ensures that the correction result retains the original strategy objective as much as possible while avoiding direct rejection of the transaction, which would interrupt strategy execution. The system first reads the confidence boundaries of each risk dimension in the risk prediction results and compares them item by item with the state change boundaries in the dynamic risk envelope, calculating the boundary exceedance amount and exceedance ratio. For states where increased values ​​lead to increased risk, the difference between the confidence upper boundary and the allowable upper limit is used as the exceedance amount; for states where decreased values ​​such as available funds and liquidity margins lead to increased risk, the difference between the allowable lower limit and the confidence lower boundary is used as the exceedance amount.

[0070] Subsequently, the main propagation paths corresponding to the out-of-limit state are traced back along the twin graph of the trading risk state to identify the action parameters that contribute significantly to the out-of-limit result. Action parameters to be corrected may include trading size, order price tolerance range, execution time limit, order splitting rhythm, and order cancellation conditions. The direction of correction is determined based on the marginal impact of parameter changes on the risk state. For example, when position exposure exceeds the limit, priority is given to reducing the trading size; when transaction deviation exceeds the limit, the single order size can be reduced or the execution time limit extended; when capital occupation exceeds the limit, the target position can be reduced or the execution order adjusted.

[0071] The safe action projection model uses candidate trading actions as the projection starting point and sets minimizing the strategy intent deviation as the optimization objective. The strategy intent deviation can be calculated by weighting factors such as target position completion rate, degree of adherence to the trading direction, expected execution timing deviation, and target profit loss. Trading direction, strategy task identifier, and necessary execution deadline can be set as immutable conditions, while trading size, order price, and order splitting rhythm can be set as adjustable parameters.

[0072] Model constraints include the boundaries of each state change in the dynamic risk envelope, account risk boundaries, risk propagation constraints, and trading rule constraints. For situations where multiple risk dimensions are jointly affected by the same action parameter, joint constraints are established to prevent the increase of other risks after a single risk adjustment. For situations where parameters are dependent on each other, such as when the number of order splits and execution time limits need to change synchronously after a decrease in trading size, parameter linkage is used to ensure the executable nature of the adjustment action.

[0073] During collaborative problem-solving, the system determines the constraint sensitivity of each action parameter to be corrected based on the aforementioned marginal impact, prioritizing the adjustment of parameters that can reduce risk more significantly with a smaller deviation from the intended strategy. Local candidate solutions can be generated first along the risk-reducing direction, and then iteratively updated using constraint search or sequential optimization. Each iteration predicts changes in position exposure, capital occupation, trading deviation, and expected loss caused by changes in action parameters, and checks whether the action has entered the feasible region of the trading action.

[0074] When multiple corrective measures meet the constraints, the one with the smallest deviation from the strategic intent and the largest margin of safety is selected. When no feasible solution exists to maintain the original execution period, non-core strategic conditions can be relaxed step by step, such as extending the execution time or reducing the target completion rate, but the account risk boundary must not be exceeded. If all allowed adjustment parameters fail to form a feasible solution, a protective action of stopping execution or transferring to manual review is generated to prevent high-risk actions from being issued.

[0075] Furthermore, upon receiving the corrected trading action, the system does not execute it directly. Instead, it reconstructs a risk scenario that matches the current dynamic risk context and updates the trading action nodes according to the corrected action parameters, re-deducing the state evolution. The system compares the newly generated risk prediction results with the dynamic risk envelope to confirm that the credible boundaries of each risk dimension are within the allowable range, the joint propagation constraints have not been breached, and the prediction uncertainty has not exceeded the allowable level. If the above conditions are met, the corrected trading action is determined as the target trading action; if there are still boundary exceedances, the parameters are adjusted based on the remaining exceedance amount and verified again until an actionable action is obtained or a protective action is triggered.

[0076] For example, an AI robot plans to buy 80,000 units of a target asset within 4 minutes. The credible upper bound for position exposure is 4.6%, and the credible upper bound for capital occupation is 13.5%, while the dynamic risk envelope is limited to 4% and 12%, respectively. Backtesting results show that the transaction size contributes the most to position exposure, while the single order size and execution time limit have a significant impact on capital occupation and execution deviation. The system keeps the trading direction and strategy task unchanged, and uses the target position completion loss, execution delay, and price deviation as the combined factors to constitute the strategy intent deviation. A collaborative solution yields a revised scheme: the transaction size is reduced to 68,000 units, the single order size is reduced by 25%, and the execution time limit is extended to 6 minutes. After re-analysis, the credible upper bound for position exposure decreases to 3.9%, the credible upper bound for capital occupation decreases to 11.4%, and the credible upper bound for execution deviation is 0.29%, all satisfying the dynamic risk envelope. Therefore, this revised scheme is selected as the target trading action.

[0077] By using the above method, candidate trading actions that exceed risk constraints can be projected into the feasible domain. While maintaining the trading direction and main strategy objectives, the action parameters can be adjusted in a coordinated manner to reduce strategy losses caused by simple interception, and at the same time improve the adaptability of the target trading action to the dynamic risk boundary and account capacity.

[0078] In some embodiments, the target trading action is executed, the trading execution trajectory is compared with the candidate risk trajectory to perform deviation analysis, risk intervention is implemented based on deviation propagation, and the trading risk status twin and dynamic risk envelope are updated based on trading feedback, including: Execute the target transaction action, collect the action status and status response according to the transaction sequence, and generate the transaction execution trajectory; The transaction execution trajectory is time-series aligned with the candidate risk trajectory, the state response deviation is calculated, and the deviation propagation path is traced along the transaction risk state twin graph to determine the source of the deviation, the scope of its impact, and the urgency of the risk. Determine the target risk intervention actions based on the risk urgency and remaining risk budget, execute the target risk intervention actions, and re-analyze the remaining transaction process; Based on the transaction feedback after intervention, the corresponding state nodes, interaction relationships and risk budget parameters are corrected. After verification through historical playback, the transaction risk state twin graph and dynamic risk envelope are updated.

[0079] Specifically, the transaction execution trajectory is used to continuously record the changes in the target trading action from its issuance to its completion, as well as the resulting market, strategy, and account responses. After the system executes the target trading action, it collects the action statuses such as order submission, exchange confirmation, partial execution, order cancellation and resubmission, and execution completion according to a unified trading clock. Simultaneously, it acquires status responses such as transaction price, order book depth, changes in open interest, capital occupation, and strategy signal adjustments. After time-series correction, various data are associated with order identifiers and trading object identifiers, and trajectory nodes are formed using a fixed sampling period or key event triggering method. Each trajectory node records the time of occurrence, the percentage of action completion, the actual state value, the amount of state change, and data reliability. Adjacent trajectory nodes are connected according to their execution sequence to generate a transaction execution trajectory that reflects the actual trading process.

[0080] During deviation analysis, the system uses the start time of the target transaction action as a benchmark, mapping the transaction execution trajectory and the corresponding candidate risk trajectory to a unified time axis. For trajectories with different sampling frequencies, alignment can be achieved using event anchors and piecewise interpolation. Event anchors include order submission, first trade, completion of predetermined percentage, and execution termination. The system calculates the numerical deviation, direction of change deviation, and response delay deviation between the actual state and the projected state for each time period, and converts the deviation into boundary occupancy increments based on the dynamic risk envelope.

[0081] Subsequently, starting from the state node exhibiting a significant deviation, the preceding action edge is traced backward along the trading risk state twin graph to determine the initial source of the deviation. Then, the potentially affected state nodes are searched backward along the direction of action to form the deviation propagation path. The source of the deviation can be attributed to changes in market disturbances, changes in strategy output, abnormal order execution, or changes in account status. The scope of impact is determined based on the number of reachable nodes, the propagation level, and the remaining execution time domain.

[0082] Risk urgency is used to characterize the likelihood and time pressure of deviations exceeding risk boundaries during the remaining trading process. The system generates risk urgency by comprehensively considering the current deviation magnitude, deviation growth rate, the strength of the propagation path, the expected time to reach key account nodes, and the reliability of the relationship. When the deviation is small and propagates slowly, monitoring can continue or the pace of subsequent orders can be reduced; when the deviation expands rapidly but there is still a remaining risk budget, the number of unexecuted orders can be reduced, the execution time limit extended, the order price adjusted, or new orders suspended; when the deviation is about to exceed the account risk boundary or the remaining risk budget is insufficient, unexecuted orders can be cancelled, hedging actions can be executed, or the trading task can be terminated. The system selects target risk intervention actions based on risk reduction capability, execution delay, and strategic intent loss, and writes the intervention actions into the trading action node, re-analyzing the remaining trading process. The corresponding intervention action is only executed if the re-analysis result meets the dynamic risk envelope; otherwise, it is upgraded to a higher protection level.

[0083] Transaction feedback includes the action status before and after intervention, actual state response, deviation changes, and final risk outcome. The system uses transaction feedback to correct deviations in state nodes and related action edges. If the actual transaction impact consistently exceeds the projected value, the strength of the action edge from the transaction action to the market state is increased; if the actual account response occurs later than predicted, the corresponding response delay is increased; if a propagation path repeatedly fails to produce the expected response, the reliability of the relationship is reduced.

[0084] The risk budget parameters are adjusted based on actual risk consumption and remaining capacity after intervention, but the magnitude of each adjustment is limited by the version update boundary. The system generates candidate update versions from the corrected results and uses historical trading data under similar market conditions for playback to compare the trajectory deviation, risk boundary coverage, and number of erroneous interventions before and after the update. If the verification results meet the stability conditions, the trading risk state twin graph and dynamic risk envelope are updated; otherwise, the current version is retained.

[0085] For example, the target trading action plan was to buy 68,000 units of asset within 6 minutes. By the second minute of execution, the actual order book depth had decreased by 28% compared to the projected value, and the trading deviation had widened from the predicted 0.22% to 0.41%, propagating along the path from trading deviation to capital occupation and then to account risk exposure. The system determined the source of the deviation to a sudden liquidity contraction, predicting that capital occupation would approach the upper limit of the dynamic risk envelope in 90 seconds, reaching a high level of risk urgency, with a remaining risk budget of only 46,000 yuan. The system chose to suspend new orders, reduce the remaining single order volume by 40%, and extend the execution time by 2 minutes. After re-projection showed that the credible upper boundary of capital occupation had fallen back to the allowable range, intervention was executed. After the transaction ended, the system increased the intensity of the liquidity contraction path by 0.12 and shortened the response latency by 0.4 seconds based on the actual response, and completed parameter updates after historical playback verification.

[0086] By employing the above methods, abnormal deviations between the actual state and the projected state can be identified in a timely manner during the transaction execution process. Interventions that match the urgency and remaining risk budget can be implemented along the risk propagation relationship. Furthermore, the risk model and risk boundaries can be continuously revised using the validated feedback, thereby improving the timeliness of risk response, the accuracy of intervention, and the adaptability of subsequent projections in transactions.

[0087] The following are system embodiments of this application, which can be used to execute the method embodiments of this application. For details not disclosed in the system embodiments of this application, please refer to the method embodiments of this application.

[0088] Figure 2 This is a schematic diagram of the structure of the dynamic risk control system for AI robots in financial transactions provided in this application embodiment. Figure 2 As shown, the system includes: The acquisition module 201 is used to acquire multi-source operational data of the financial trading AI robot, and generate a transaction status event stream after time-series correction and credibility assessment. Module 202 is used to construct a transaction risk state twin graph based on the transaction state event flow, representing the relationship between market state, strategy state, transaction action and account state, and to generate dynamic risk context; Module 203 is used to determine the risk budget and propagation constraints based on the dynamic risk context, account risk boundaries and interaction relationships, and to generate a dynamic risk envelope. The deduction module 204 is used to acquire candidate transaction actions, construct risk deduction scenarios that match the dynamic risk context, deduce state evolution based on the transaction risk state twin graph, and generate candidate risk trajectories. Comparison module 205 is used to generate risk prediction results based on candidate risk trajectories and prediction uncertainties, and compare them with dynamic risk envelopes to determine the feasible domain of transaction actions; The generation module 206 is used to project a safe action under dynamic risk envelope constraints to generate a target transaction action when the candidate transaction action exceeds the feasible domain of the transaction action. The update module 207 is used to execute the target transaction action, perform deviation analysis between the transaction execution trajectory and the candidate risk trajectory, implement risk intervention based on the deviation propagation, and update the transaction risk status twin map and dynamic risk envelope based on transaction feedback.

[0089] In some embodiments, Figure 2 The acquisition module 201 acquires multi-source operational data reflecting the trading environment, robot operation, and account status, and establishes data associations according to data sources and trading objects; it identifies and corrects the time offset of multi-source operational data based on a unified trading clock and event causal order, and generates time-series aligned data; it generates data credibility based on data source quality, cross-source state consistency, and transmission timeliness, and resolves state conflicts of the same trading object based on data credibility and event causal order; it converts the conflict-resolved time-series aligned data into trading state events that are associated with trading objects, occurrence times, state changes, and data credibility, and aggregates them to generate a trading state event stream.

[0090] In some embodiments, Figure 2 The construction module 202 maps the transaction state event flow to corresponding market state, strategy state, transaction action, and account state state nodes according to the transaction object identifier and the causal sequence of events. Based on the temporal response and state change correlation between each state node, it determines the direction of action, intensity of action, response delay, and relationship credibility, establishes action edges between state nodes, and generates a transaction risk state twin graph. Based on the sliding time window, it updates the state nodes and action edges, identifies market state migration and strategy operation offset, and integrates the updated market state, strategy state, account state, and relationship credibility to generate a dynamic risk context.

[0091] In some embodiments, Figure 2 The determination module 203 determines market disturbances, strategy deviations, and account carrying status based on the dynamic risk context, and generates a total risk budget by combining the account risk boundary; it identifies risk propagation paths along the interaction relationships in the transaction risk state twin graph, and determines the propagation contribution and coupling amplification degree based on the intensity of the effect, response delay, and relationship credibility; it allocates the total risk budget based on the propagation contribution, and corrects the corresponding risk sub-budget according to the coupling amplification degree to generate propagation constraints; it combines the risk sub-budget, propagation constraints, and uncertainty safety margin to determine the state change boundaries that can be formed by candidate transaction actions, and generates a dynamic risk envelope.

[0092] In some embodiments, Figure 2The deduction module 204 acquires candidate trading actions output by the financial trading AI robot, associates them with corresponding strategy objectives and trading states, and generates trading intentions to be deduced; it extracts state disturbance factors affecting trading execution based on dynamic risk context, and generates risk deduction scenarios according to the direction of disturbance, magnitude of change, and coupling relationship; it loads the trading intentions to be deduced into each risk deduction scenario, and propagates the state changes caused by the trading actions in time periods based on the direction of action, intensity of action, and response delay in the trading risk state twin graph; it gathers the state changes, risk propagation paths, and proximity of risk boundaries in each time period to generate candidate risk trajectories.

[0093] In some embodiments, Figure 2 The comparison module 205 extracts the risk state peak, cumulative change, and propagation intensity from the candidate risk trajectory to generate a trajectory risk representation; it adaptively calibrates the trajectory risk representation based on the historical extrapolation deviation matched with the dynamic risk context to determine the prediction interval and the corresponding prediction uncertainty; it integrates the trajectory risk representation, prediction interval, and prediction uncertainty to generate a risk prediction result containing the risk prediction value and the confidence boundary; it compares the confidence boundary of the risk prediction result with the state change boundary in the dynamic risk envelope item by item, and determines the set of trading actions that simultaneously satisfy each state change boundary as the feasible region of the trading action.

[0094] In some embodiments, Figure 2 The comparison module 205 performs hierarchical matching of historical extrapolation deviations based on the similarity between the historical dynamic risk context and the current dynamic risk context, generating a state calibration sample set; identifies the state transition of the dynamic risk context, and adjusts the sample weights in the state calibration sample set according to the transition time between the sample's corresponding state and the current state; performs time-series conformal calibration based on the weighted historical extrapolation deviations, determines the quantile error boundary corresponding to the trajectory risk characterization, and generates a prediction interval; corrects the quantile error boundary based on the historical coverage deviation of the prediction interval, and generates prediction uncertainty based on the correction magnitude and interval width.

[0095] In some embodiments, Figure 2 The generation module 206 identifies the state change boundaries that candidate trading actions exceed based on the risk prediction results, determines the parameters of the actions to be corrected and the corresponding correction direction; takes the minimum deviation of the strategy intent between the candidate trading action and the corrected trading action as the projection target, and uses the dynamic risk envelope and account risk boundary as constraints to construct a safe action projection model; collaboratively solves the parameters of the actions to be corrected to generate corrected trading actions that fall within the feasible domain of the trading action; re-executes risk deduction on the corrected trading actions, and determines the corrected trading actions that satisfy the dynamic risk envelope as the target trading actions.

[0096] In some embodiments, Figure 2The update module 207 executes the target transaction action, collects the action status and status response according to the transaction time sequence, and generates the transaction execution trajectory; it aligns the transaction execution trajectory with the candidate risk trajectory in time sequence, calculates the status response deviation, and traces the deviation propagation path along the transaction risk status twin graph to determine the source of deviation, scope of impact, and risk urgency; it determines the target risk intervention action based on the risk urgency and remaining risk budget, executes the target risk intervention action, and re-deduces the remaining transaction process; it corrects the corresponding status nodes, interaction relationships, and risk budget parameters based on the transaction feedback after intervention, and updates the transaction risk status twin graph and dynamic risk envelope after historical playback verification.

[0097] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0098] Figure 3 This is a schematic diagram of the electronic device 3 provided in an embodiment of this application. Figure 3 As shown, the electronic device 3 of this embodiment includes: a processor 301, a memory 302, and a computer program 303 stored in the memory 302 and executable on the processor 301. When the processor 301 executes the computer program 303, it implements the steps in the various method embodiments described above. Alternatively, when the processor 301 executes the computer program 303, it implements the functions of each module / unit in the various system embodiments described above.

[0099] Electronic device 3 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 3 may include, but is not limited to, processor 301 and memory 302. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or different components.

[0100] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0101] The memory 302 can be an internal storage unit of the electronic device 3, such as a hard disk or memory of the electronic device 3. The memory 302 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 3. The memory 302 can also include both internal and external storage units of the electronic device 3. The memory 302 is used to store computer programs and other programs and data required by the electronic device.

[0102] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0103] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a readable storage medium (e.g., a computer-readable storage medium). Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which may be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0104] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for dynamic risk control of AI robots for financial transactions, characterized in that, include: Acquire multi-source operational data of the financial trading AI robot, and generate a transaction status event stream after time-series correction and credibility assessment; Based on the transaction state event flow, a transaction risk state twin graph is constructed to represent the interaction between market state, strategy state, transaction action and account state, and a dynamic risk context is generated. Based on the dynamic risk context, account risk boundaries, and the interaction relationships, risk budgets and propagation constraints are determined, and a dynamic risk envelope is generated. Candidate transaction actions are acquired, a risk simulation scenario matching the dynamic risk context is constructed, the state evolution is simulated based on the transaction risk state twin graph, and candidate risk trajectories are generated. Based on the candidate risk trajectory and the predicted uncertainty, a risk prediction result is generated and compared with the dynamic risk envelope to determine the feasible domain of the transaction action. When the candidate transaction action exceeds the feasible domain of the transaction action, a safe action projection is performed under the dynamic risk envelope constraint to generate the target transaction action. The target transaction action is executed, and the deviation analysis between the transaction execution trajectory and the candidate risk trajectory is performed. Risk intervention is implemented based on the deviation propagation, and the transaction risk status twin graph and the dynamic risk envelope are updated based on the transaction feedback.

2. The method according to claim 1, characterized in that, The process of acquiring multi-source operational data from the financial trading AI robot, and generating a trading status event stream after time-series correction and credibility assessment, includes: Acquire multi-source operational data reflecting the trading environment, robot operation, and account status, and establish data associations based on data sources and trading objects; Based on a unified transaction clock and event causal sequence, the time offset of the multi-source running data is identified and corrected to generate time-series aligned data. Data credibility is generated based on data source quality, cross-source state consistency, and transmission timeliness. State conflicts of the same transaction object are resolved based on the data credibility and the causal order of the events. The time-series aligned data after conflict resolution is converted into transaction status events that relate to the transaction object, the time of occurrence, the state change, and the data credibility, and then aggregated to generate the transaction status event stream.

3. The method according to claim 1, characterized in that, The process involves constructing a transaction risk state twin graph based on the transaction state event flow, representing the interaction between market state, strategy state, trading actions, and account state, and generating a dynamic risk context, including: Based on the transaction object identifier and the causal sequence of events, the transaction status event stream is mapped to status nodes corresponding to market status, strategy status, transaction action, and account status. Based on the temporal response and state change correlation between each state node, the action direction, action intensity, response delay and relationship reliability are determined, action edges between state nodes are established, and the transaction risk state twin graph is generated. The state nodes and action edges are updated based on a sliding time window to identify market state transitions and strategy execution offsets. The dynamic risk context is generated by integrating the updated market status, strategy status, account status, and relationship credibility.

4. The method according to claim 1, characterized in that, The step of determining the risk budget and propagation constraints based on the dynamic risk context, account risk boundaries, and the interaction relationships, and generating a dynamic risk envelope, includes: Based on the dynamic risk context, market disturbances, strategy deviations, and account carrying status are determined, and a total risk budget is generated by combining the account risk boundary. Identify risk propagation paths along the interaction relationships in the transaction risk state twin graph, and determine propagation contributions and coupling amplification degrees based on interaction strength, response delay, and relationship credibility; The total risk budget is allocated based on the propagation contribution, and the corresponding risk sub-budget is modified according to the degree of coupling amplification to generate the propagation constraints; By combining the risk sub-budget, the propagation constraints, and the uncertainty safety margin, the boundary of state changes that can be formed by candidate transaction actions is determined, and the dynamic risk envelope is generated.

5. The method according to claim 1, characterized in that, The steps of acquiring candidate transaction actions, constructing a risk simulation scenario matching the dynamic risk context, and inferring state evolution based on the transaction risk state twin graph to generate candidate risk trajectories include: The candidate trading actions output by the financial trading AI robot are obtained, and the corresponding strategy objectives and trading states are associated to generate the trading intention to be deduced. Based on the dynamic risk context, extract the state disturbance factors that affect transaction execution, and generate risk simulation scenarios according to the direction of disturbance, the magnitude of change and the coupling relationship; The proposed transaction intention is loaded into each risk scenario, and the state changes caused by the transaction action are propagated in time intervals based on the direction of action, intensity of action, and response delay in the transaction risk state twin graph. The candidate risk trajectory is generated by aggregating the state changes, risk propagation paths, and the proximity of risk boundaries at different time periods.

6. The method according to claim 1, characterized in that, The step of generating risk prediction results based on the candidate risk trajectories and prediction uncertainties, and comparing them with the dynamic risk envelope to determine the feasible domain of the transaction action includes: The risk state peak, cumulative change and propagation intensity are extracted from the candidate risk trajectories to generate a trajectory risk characterization. The trajectory risk representation is adaptively calibrated based on the historical extrapolation deviation matched with the dynamic risk context to determine the prediction interval and the corresponding prediction uncertainty. By integrating the trajectory risk representation, the prediction interval, and the prediction uncertainty, a risk prediction result containing the risk prediction value and the confidence boundary is generated; The credible boundary of the risk prediction result is compared item by item with the state change boundary in the dynamic risk envelope, and the set of transaction actions that simultaneously satisfy each state change boundary is determined as the feasible region of the transaction action.

7. The method according to claim 6, characterized in that, The adaptive calibration of the trajectory risk representation based on historical extrapolation deviations matched with the dynamic risk context, to determine the prediction interval and corresponding prediction uncertainty, includes: Based on the similarity between the historical dynamic risk context and the current dynamic risk context, the historical extrapolation deviation is hierarchically matched to generate a state calibration sample set. Identify the state transitions of the dynamic risk context, and adjust the sample weights in the state calibration sample set according to the transition time interval between the sample's corresponding state and the current state; Based on the weighted historical extrapolation deviation, time-series conformal calibration is performed to determine the quantile error boundary corresponding to the trajectory risk characterization and generate the prediction interval. The quantile error boundary is corrected based on the historical coverage deviation of the prediction interval, and the prediction uncertainty is generated based on the correction magnitude and the interval width.

8. The method according to claim 1, characterized in that, When the candidate transaction action exceeds the feasible region of the transaction action, a safe action projection is performed under the dynamic risk envelope constraint to generate the target transaction action, including: Based on the risk prediction results, identify the state change boundaries that the candidate transaction actions exceed, and determine the action parameters to be corrected and the corresponding correction direction. The projection objective is to minimize the strategic intent deviation between the candidate transaction action and the corrected transaction action, and the dynamic risk envelope and account risk boundary are used as constraints to construct a safe action projection model. The parameters of the action to be corrected are solved collaboratively to generate corrected transaction actions that fall within the feasible region of the transaction action; The risk simulation is re-executed on the modified trading action, and the modified trading action that satisfies the dynamic risk envelope is determined as the target trading action.

9. The method according to claim 1, characterized in that, The execution of the target transaction action, including deviation analysis between the transaction execution trajectory and the candidate risk trajectory, risk intervention based on deviation propagation, and updating the transaction risk status twin and the dynamic risk envelope based on transaction feedback, includes: Execute the target transaction action, collect the action status and status response according to the transaction time sequence, and generate the transaction execution trajectory; The transaction execution trajectory is time-aligned with the candidate risk trajectory, the state response deviation is calculated, and the deviation propagation path is traced along the transaction risk state twin graph to determine the source of the deviation, the scope of its impact, and the urgency of the risk. Based on the risk urgency and remaining risk budget, determine the target risk intervention action, execute the target risk intervention action, and re-analyze the remaining transaction process; Based on the transaction feedback after intervention, the corresponding state nodes, interaction relationships, and risk budget parameters are corrected. After verification through historical playback, the transaction risk state twin graph and the dynamic risk envelope are updated.

10. A dynamic risk control system for AI robots in financial transactions, characterized in that, include: The acquisition module is used to acquire multi-source operational data of the financial trading AI robot, and generate a transaction status event stream after time-series correction and credibility assessment; The construction module is used to construct a transaction risk state twin graph representing the interaction between market state, strategy state, trading action and account state based on the transaction state event stream, and generate dynamic risk context; The determination module is used to determine the risk budget and propagation constraints based on the dynamic risk context, account risk boundary and the interaction relationship, and generate a dynamic risk envelope; The deduction module is used to acquire candidate transaction actions, construct risk deduction scenarios that match the dynamic risk context, deduce state evolution based on the transaction risk state twin graph, and generate candidate risk trajectories. The comparison module is used to generate risk prediction results based on the candidate risk trajectories and prediction uncertainties, and compare them with the dynamic risk envelope to determine the feasible domain of the transaction action; The generation module is used to generate a target transaction action by projecting a safe action under the dynamic risk envelope constraint when the candidate transaction action exceeds the feasible domain of the transaction action. The update module is used to execute the target transaction action, perform deviation analysis between the transaction execution trajectory and the candidate risk trajectory, implement risk intervention based on deviation propagation, and update the transaction risk status twin and the dynamic risk envelope based on transaction feedback.