An intelligent decision-making method and system based on reliability perception fusion and constraint scene deduction

Abstract

The application provides an intelligent decision-making method and system based on reliability perception fusion and constraint scene deduction. Through a history scene Bayesian prior and real-time logical consistency double weighting mechanism, the application automatically reduces the weight of contradictory and low credibility signals, and suppresses the excessive confidence of large models. The application discards the full connection multi-round debate architecture, adopts the single structured JSON output of analysts, and can adapt to the real-time trading needs of securities companies. The application does not need model fine-tuning and retraining, but only relies on daily text gradient and weekly knowledge distillation to update the market style adaptation prompt word; each analyst server locally retains the original market data, financial report data and news data, and only transmits the desensitized confidence and reasoning abstract structured signal across nodes. The original sensitive financial data does not leave the local node, and there is no need to additionally deploy privacy computing hardware, which meets the data isolation and information security supervision specifications of the financial industry.

Description

Technical Field

[0001] This invention relates to the fields of computer data processing and artificial intelligence financial technology, and more specifically, to an intelligent decision-making method and system based on reliability perception fusion and constraint scenario deduction. Background Technology

[0002] Stock prediction systems identify market trends by analyzing historical price sequences, technical indicators, news texts, and correlations between stocks, thereby providing investors with trading decision support and ensuring the steady appreciation of capital in a complex financial environment. Multi-agent collaboration, a recently emerging paradigm of large language model applications, simulates analysts in different roles (such as market analysts, fundamental analysts, and news analysts) reasoning and debating based on textual information from different channels (such as historical price sequences, technical indicators, and news texts) to generate comprehensive predictive text conclusions. It aims to integrate information from different channels and simulate the collaborative process of human expert teams to improve the accuracy and interpretability of predictions.

[0003] With the rapid development of large language model technology, combining multi-agent frameworks with graph neural networks has become a new research hotspot. Such methods can utilize graph structures to model dynamic dependencies between stocks (such as industry spillover relationships and supply chain transmission relationships), while leveraging the text understanding capabilities of large language models to process unstructured news and financial statement texts, thus providing a more comprehensive picture of market conditions. Meanwhile, to address the overconfidence and context window limitations that large language models are prone to in financial decision-making, multi-agent debate mechanisms have been introduced into the prediction process. These mechanisms allow agents to mutually verify each other's viewpoints through multiple rounds of natural language interaction, reducing the bias of a single model and improving the depth of reasoning.

[0004] However, in real financial markets, technical indicators, fundamental data, news sentiment, and the relationship between stocks are often filled with irrelevant or misleading information. Existing systems typically process this information by directly merging or simply splicing it together with equal weights, forcing the model to assign equal trust to all information. The model cannot dynamically resolve contradictions between pieces of information, nor can it suppress low-quality input. As a result, it is prone to making overconfident and erroneous judgments, believing that all information is true and valid. While multi-agent fully connected debate can improve reasoning quality, each agent needs to read and respond to messages from all other agents. The communication overhead for a single decision often exceeds 100,000 tokens and has a latency of tens of seconds, severely restricting the deployment in real-time trading scenarios. Furthermore, existing methods largely rely on short-term market volatility and risk appetite for decision-making, lacking explicit constraints on long-term risks (such as conditional value of risk), making them susceptible to significant losses during periods of high volatility. More importantly, the non-stationarity and concept drift (market patterns changing over time) of financial markets require models to continuously adapt to new market styles. Existing systems require the collection of large amounts of data and model fine-tuning or retraining, costing tens of GPU hours and exhibiting slow response times. Therefore, how to reliably integrate data from different information sources without retraining model parameters, reduce the resources and time consumed in exchanging information between agents, strictly limit potential investment losses (i.e., downside risk), and keep up with changes in market style at a low cost are pressing technical problems in the field of stock prediction based on large language models. Therefore, we propose an intelligent decision-making method and system based on reliability-aware fusion and constraint scenario inference. Summary of the Invention

[0005] In view of the deficiencies existing in the above-mentioned background art, the technical problem to be solved by the present invention includes: First, existing multi-agent stock prediction systems based on large language models struggle to determine the reliability of information from multiple sources when integrating information from various sources, and they also fail to flexibly reconcile conflicting information. As a result, the system treats all signals of different natures—technical indicators, fundamentals, news, relationship momentum, etc.—with the same level of trust. This amplifies noisy and contradictory information, leading the system to make overconfident and erroneous judgments. This invention aims to solve this problem by automatically adjusting the weight of each piece of information based on historically similar market scenarios and the logical consistency between current pieces of information. This minimizes the impact of less reliable information and prevents contradictory information from interfering with each other.

[0006] Second, in the existing "fully connected multi-agent debate" architecture, each agent has to read and respond to messages from all other agents, resulting in extremely high communication overhead and decision latency. A single decision often consumes over 100,000 tokens, with latency reaching tens of seconds, which is far from meeting the needs of real-time transactions. This invention aims to solve this problem by reducing repetitive and unnecessary communication between agents, lowering inference costs, and improving response speed.

[0007] Third, existing decision-making methods often neglect the quantitative control of long-term risks such as downside risk and conditional risk value. They often adjust risk appetite based solely on textual descriptions of short-term market fluctuations, without conducting stress tests for extreme market conditions. As a result, significant losses are easily incurred once the market experiences sharp fluctuations. This invention aims to solve this problem by strictly controlling potential losses while striving for maximum returns, truly achieving a balance between risk and return.

[0008] Fourth, existing systems struggle to adapt to the "concept drift" and non-stationary characteristics of financial markets in a low-cost and efficient manner, meaning that market rules and styles are constantly changing. Once the market style changes, it's necessary to collect large amounts of data again and fine-tune or retrain the model, which is both time-consuming and computationally expensive. This invention aims to solve this problem by not updating the model weights at all, i.e., not changing the internal parameters of the model. Instead, it relies solely on a natural language feedback mechanism to continuously optimize the model's inference logic, thereby avoiding performance degradation due to changes in market style and significantly reducing the cost of model updates.

[0009] To achieve the aforementioned objectives, this invention provides the following technical solution: an intelligent decision-making method based on reliability-aware fusion and constraint scenario deduction, applied to a distributed multi-server financial forecasting framework, wherein the framework comprises a central fusion server and multiple analyst servers, and the method includes the following steps: S1. Construct a distributed prediction framework for the observable Markov decision process in the graph enhancement part. Deploy role-differentiated large language model agents on each analyst server to analyze technical indicators, fundamentals, news semantics, and stock correlations, and output structured heterogeneous signals. S2. The heterogeneous signals output by each analyst server are aggregated to the central fusion server. The reliability-aware fusion module retrieves historical similar scenarios to generate Bayesian prior weights, performs logical consistency scoring on each signal, and calculates adaptive weights to resolve conflicts and obtain a unified state representation. S3. Input the state representation after conflict resolution into the constraint scenario review module, and generate three candidate strategies in parallel: wait and see, light position trial, and heavy position aggressive. For each strategy, perform three counterfactual scenarios of downtrend, oscillation, and uptrend, and calculate the profit and loss asymmetry ratio. Combined with the conditional value at risk constraint, perform risk pruning to select the optimal trading action. S4. The central fusion server compares the predicted action returns with the actual market returns, and uses a reflective strategy to modify and generate natural language text gradient updates for each agent's prompts. Every week, knowledge distillation is performed to extract general strategy principles, which are then solidified into the underlying prompts of each analyst agent, enabling the model to adapt and evolve while keeping the model weights frozen throughout the process.

[0010] As a preferred technical solution of the present invention, in step S1, the analyst server includes a market analyst server, a fundamental analyst server, a news and industry analyst server, and a momentum overflow analyst server; each server is equipped with an independently customized prompt word large language model, and only outputs desensitized JSON structured signals, and the original price, financial report, and news data are not transmitted across servers.

[0011] As a preferred technical solution of the present invention, the reliability perception fusion module in step S2 uses the Chroma vector database to store historical scene vectors, and uses cosine similarity to retrieve the top K similar historical scenes to calculate Bayesian prior weights; the arbitration big language model completes the logical consistency scoring from three dimensions: causal integrity, evidence sufficiency, and macro background fit, and obtains the final weight of each signal by weighted normalization.

[0012] As a preferred technical solution of the present invention, in step S3, the constraint scenario review module uses Monte Carlo tree search to complete multi-branch counterfactual deduction; the high-risk language model sets the CVaR downside risk threshold as a hard constraint, performs feasibility mask filtering on candidate strategies, and selects the transaction action corresponding to the maximum profit-loss asymmetry ratio after retaining only the risk compliance branch.

[0013] As a preferred technical solution of the present invention, the reflective strategy correction mechanism described in step S4 generates a text gradient containing error location, logic correction, and operational lessons learned at the close of trading each day, and adds it to the corresponding analyst agent's prompt word memo; weekly knowledge distillation filters out noise reflection caused by daily fluctuations, and outputs standardized long-term strategic memos broadcast to all analyst servers.

[0014] As a preferred technical solution of the present invention, each server adopts a star topology HTTP communication, the analyst server outputs a structured JSON signal only once, and the total token consumption for a single decision is controlled at 3000-5000, which reduces the communication overhead by more than 95% compared with full connection multi-round debate.

[0015] An intelligent decision-making system based on reliability perception fusion and constraint scenario inference, used to execute the intelligent decision-making method, the system comprising: Multiple analyst servers, each configured with an independent large language model intelligent agent, respectively correspond to market technical analysis, fundamental financial analysis, news industry sentiment analysis, and stock momentum spillover analysis, used to locally parse raw financial data and output heterogeneous structured signals; A central fusion server communicates with each analyst server via an internal network API, and includes: The reliability-aware fusion module is used to retrieve similar historical scenarios, calculate adaptive signal weights, and resolve signal conflicts among multiple agents. The constraint scenario review module is used for multi-branch counterfactual scenario simulation, risk constraint pruning, and outputting the optimal compliant transaction action. The adaptive evolution module includes a reflection and review sub-unit and a weekly knowledge distillation sub-unit, which are used to generate text gradients, extract general investment strategies, and update prompt words for each analyst agent. The vector database, message queue, and local log database store historical scenario vectors, server scheduling instructions, and daily decision and review logs, respectively.

[0016] As a preferred technical solution of the present invention, the reliability perception fusion module has a built-in arbitration big language model, is configured with special prompt words for weight calculation, inputs the current heterogeneous signal and the retrieved historical scene, and outputs the normalized multi-agent signal weight.

[0017] As a preferred technical solution of the present invention, the constraint scenario review module includes a review big language model and a risk big language model; the review big language model generates three types of candidate position strategies and completes multi-scenario return drawdown simulation; the risk big language model uses the upper limit of conditional value of risk as a constraint to intercept and filter strategies.

[0018] As a preferred technical solution of the present invention, the adaptive evolution module freezes the network weights of all large language models throughout the process, and only modifies the prompt words of each analyst server system to achieve adaptive market concept drift.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: First, it significantly improves prediction accuracy and risk control capabilities. Through a multi-perspective LLM signal encoder with differentiated roles (four analysts independently deployed: market, fundamentals, news, and momentum spillover), heterogeneous signals originally mixed in a large model are decoupled into four independent inference channels. Combined with Bayesian context gating and logical consistency scoring in the reliability-aware fusion mechanism, the system can dynamically suppress signals with low confidence and logical contradictions. Experiments show that on stocks with obvious downward trends, such as 000063.SZ, this invention increases the annualized return from 2.17% using the traditional LLM method (TradingAgents) to 16.21%, the Sharpe ratio from 0.37 to 1.55, and the maximum drawdown is controlled within 12.35%. Compared to the negative returns and high volatility of deep learning models (LSTM), both prediction accuracy and risk-adjusted returns achieve a qualitative leap.

[0020] Secondly, it significantly reduces model update costs, saving computing power and time. Existing technologies (such as TradingAgents or deep reinforcement learning models) require data re-collection, labeling, training, or fine-tuning of LLM parameters when facing market concept drift. A single retraining session typically requires tens of GPU hours or even days. This invention employs a reflective policy correction mechanism and weekly knowledge distillation. Model adaptation can be achieved simply by modifying the "battle memo" and strategic principles in the system prompts of each analyst server. No backpropagation or gradient updates are required throughout the process, and the model weights are completely frozen. Calculations show that the weekly distillation operation only requires calling LLM inference about a few dozen times, consuming less than 1% of the computing power of traditional retraining, and can be completed on a regular CPU server, significantly saving on computing power investment and electricity consumption.

[0021] Third, it reduces multi-agent communication overhead and improves system efficiency. Existing LLM multi-agent systems (such as TradingAgents) typically require all agents to engage in multiple rounds of fully connected debate, generating a large amount of natural language text in each round. This results in token consumption for a single decision often exceeding 100,000, with response latency reaching tens of seconds. This invention adopts a star topology of analyst server → central fusion server. Each analyst only needs to output a structured JSON signal once (typically <500 tokens). The central server synthesizes the signals and then performs one arbitration LLM and one deliberation LLM call. The total token consumption for a single decision is approximately 3,000 to 5,000, reducing communication data volume by more than 95% and decision latency from minutes to seconds. Simultaneously, distributed deployment allows four analysts to perform parallel inference on different physical servers, further shortening end-to-end response time.

[0022] Fourth, it achieves data privacy protection and avoids the leakage of raw financial data. In traditional centralized forecasting methods, all companies' financial reports, transaction records, and news data must be aggregated on a single server, posing data leakage and compliance risks. This invention adopts a distributed LLM architecture, where each analyst server only stores and infers local financial data. The servers only transmit anonymized textual signals (such as confidence levels and inference summaries) and system prompt updates; the original price series, financial details, and full-text news articles remain on their respective servers. This design meets the data isolation and privacy compliance requirements of financial institutions and eliminates the need for additional, expensive privacy-preserving computing hardware.

[0023] Experiments on multiple real-world A-share datasets demonstrate that this invention outperforms existing methods in terms of annualized return, Sharpe ratio, and maximum drawdown, while reducing communication overhead between multiple agents and improving the system's robustness and interpretability. Furthermore, the distributed LLM architecture employed in this invention allows each analyst server to perform independent inference using only local financial data, and uploads textual signals rather than raw data to a central server for fusion, achieving global collaborative prediction while strictly protecting data privacy. All LLM model weights remain unchanged, and domain adaptation is achieved only through dynamic updates of prompts, significantly reducing deployment and maintenance costs. Attached image description: Figure 1 The framework flowchart provided for this invention; Figure 2 The overall flowchart provided for this invention; Figure 3 This is an overall example diagram provided for the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are specific implementations of the present invention and are not limited to all embodiments.

[0025] Therefore, the following detailed description of embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely illustrates some embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] It should be noted that, in the absence of conflict, the embodiments and features and technical solutions in the embodiments of the present invention can be combined with each other. It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0027] Example 1: System Hardware Deployment: A local area network distributed system was built, deploying 4 analyst servers and 1 central fusion server. All devices ran Ubuntu 22.04, loaded the DeepSeek-7B large language model, synchronized the servers with NTP time, and automatically started the prediction process at 9:00 AM daily. Analyst server hardware configuration: Intel Xeon Gold 6248 CPU, RTX 4090 GPU, 64GB RAM, 1TB NVMe SSD; 4 servers correspond to four types of analysis roles: market, fundamentals, news industry, and momentum overflow, with local independent storage corresponding to the raw financial data.

[0028] Centralized converged server hardware configuration: dual Intel Xeon Gold 6248 CPUs, dual RTX 4090 GPUs, 128GB of memory, 2TB NVMe SSD; internally deployed 5 LLM inference instances for arbitration, deliberation, risk, review, and distillation, equipped with Chroma vector database, Redis message queue, and SQLite log database.

[0029] Servers transmit JSON-formatted de-identified signals via HTTPS protocol, and Redis message queues enable daily forecast task synchronization scheduling.

[0030] Reliability Awareness Fusion The JSON-formatted text output by each analyst server is then aggregated to the central fusion server. The central server maintains a ChromaDB analyst vector database, initially empty. After each prediction round, the system converts each analyst's output JSON text into vector form using ChromaDB's default all-MiniLM-L6-v2 encoder and stores it in the database, continuously enriching its content. Subsequently, the central server uses the current JSON text output by each analyst as a query, retrieving the K most similar historical scenes *st* (K = 3) from the database using cosine similarity. Based on the historical fusion weights of these scenes, it calculates Bayesian prior weights and outputs a preliminary message weight *wprior* based on historical experience (episodic memory). Where st is the current state, st-k is the historical state, sim(·) is the cosine similarity measure in the latent embedding space, and τ is the temperature parameter that controls the sharpness of the historical focus.

[0031] Then, the arbitration LLM on the central server performs a logical consistency score on the JSON-formatted text output by each analyst's server, scoring it from three dimensions: causal completeness, sufficiency of evidence, and relevance to the current macroeconomic context. The prompt for Arbitration LLM is as follows: {You are a weight calculator. Based on the following input, output the weights of 5 analysts (total 1.0).}

[0032] enter: Current report: {JSON formatted text output by each analyst's server st} Historical reference (for inspiration only, not mandatory): {Historical scenes retrieved from vector databases using st-k} Note: Even if history is similar, if the current logic is inconsistent or the signals are weak, its weight must be reduced. Focus on the present.

[0033] Calculation logic: 1. Assess signal strength (logic, signal strength).

[0034] 2. Handling conflict (technology vs. messaging, combining historical trends and dominant factors).

[0035] 3. Normalization.

[0036] Output: Strict JSON, no explanation, no code blocks.

[0037] Then, based on the output of the arbitration LLM... The final message weight wt is calculated from the previously calculated initial message weight wprior. Finally, the final message weight wt is concatenated with the original reports from each analyst to form a new text, resulting in the message state representation Ct after conflict resolution. .

[0038] Constraint Scenario Deliberation Subsequently, the Ct state text representation is input into the constraint scenario deliberation module, and the system retrieves similar scenario memories from the ChromaDB audit vector database on the central server based on the current state, injecting them into the prompt words. At the start of the deliberation, the system does not rush to draw conclusions. Instead, the deliberation LLM generates three significantly different candidate strategy branches in parallel: a wait-and-see approach (b_{wait}), a small-scale trial-and-error approach (b_{probe}), and a large-scale aggressive buy or sell approach (b_{agg}), thus covering a complete decision-making process from conservative to aggressive. Next, the deliberation LLM enters the core counterfactual deduction phase, which applies three distinct future market scenarios to each candidate branch: a downside risk scenario (such as a sudden liquidity shock or negative news), a sideways consolidation scenario (repeatedly consuming time costs), and an upside breakout scenario (momentum continuation or positive news drive), thereby simulating the specific performance and resilience of each branch under extreme favorable and unfavorable environments. Based on the simulated data, an asymmetric ratio is calculated for each branch. This ratio is calculated by dividing the expected potential return in the upside scenario by the worst-case drawdown in the downside scenario. This quantifies the profit-loss ratio of each strategy. A higher ratio indicates that the branch has "great potential to make money and limited room for loss".

[0039] in For upside potential gains, the denominator is the worst-case drawdown.

[0040] The prompt for reviewing LLM is as follows: {You are the core of an MCTS-driven strategy game. Task: Conduct scenario-response simulations for uncertainty, evaluate the robustness and profit-loss asymmetry of different position strategies, and provide recommendations on "intended position".}

[0041] Input: Current environment and analyst judgment (router weight is considered prior attention; low-weighted analyst opinions are treated with caution, but consensus is not mandatory. Looking for the strongest single signal: if there is a buy recommendation with an analyst confidence level ≥ 0.7, an attack can be initiated): {Message status represents Ct} Recent Practical Review Notes (To avoid repeating mistakes, not to replicate history): {Weekly Knowledge Distillation} Historical similarity case calibration: {Audit vector database search content} Thinking Process (MCTS, integrated into the final output): 1. Construct three mutually exclusive branches: A. Heavy position (>70%), B. Trial and error (20%-50%), C. Empty position / very light position.

[0042] 2. Stress test each branch: downside risk (stop-loss clarity), volatility wear, and upside resilience. Calculate asymmetry [maximum loss vs. expected return].

[0043] 3. Decision-making: Based on historical calibration, reject branches with uncontrollable consequences and select the branch where "the consequences of a wrong judgment are controllable and the benefits of a right judgment are maximized".

[0044] Output: Operation: [Buy / Wait] Decision and thought process: Briefly describe the three branch simulations, explain why the other two were abandoned, and elaborate on the selected option. After calculation, the JSON text output by the review LLM is input into the risk LLM. The risk LLM reviews the three branches output by the review LLM and performs pruning to select the final action.

[0045] Specifically, risk LLM is based on risk constraints. (Conditional Value at Risk (CVaR) cap, default 5%) and the requirement for consistency between the review LLM text and analyst logic to generate a feasibility mask: Then select the branch that satisfies the mask and has the largest asymmetric ratio. Key phrase for Risk LLM: {You are the Chief Risk Actuary. Core mission: Ensure every penny lost is worthwhile—allow for logical trial and error, and prevent mindless gambling.}

[0046] enter: Current environment and analyst assessment: {Message status representation Ct} Recent Practical Review: {Weekly Knowledge Distillation} Reviewing LLM: {Reviewing LLM Output} Historical similarity case calibration: {Risk vector database search content} Thinking process: Factual constraints: All risks / opportunities must be based on input data or generally accepted logic, and fabrication is strictly prohibited; if the manager's plan is sound, simply write "no obvious omissions".

[0047] Panoramic Deviation Scan: Check whether clear negative factors (downward) have been ignored or high-profit-loss ratio opportunities (upward); if relevant stock / sector data is available, analyze whether the "catch-up rise / follow-the-fall" qualitative assessment is reasonable.

[0048] Finally, the central server combines the outputs of the review LLM and the risk LLM to output the final transaction action.

[0049] Reflective strategy modification The central server calculates the predicted return based on the final transaction action and compares it with the actual return. The central server's review LLM generates a text gradient in natural language form.

[0050] Specifically, after each trading day, the central server obtains the actual market returns and inputs the decision-making inference chain along with the actual results into the review LLM. The review LLM generates text gradients in the format of natural language error analysis and correction suggestions, such as: "Error reason: Momentum overflow analysts ignored recent negative industry policies, leading to overestimation of valuation signals. Corrective measures: In the future, during periods of concentrated policy activity, increase the weight of news analysts by at least 0.2." These output reflection texts are then stored in corresponding vector databases so that other agents can access them when searching for relevant information, including errors made or successful experiences.

[0051] The strategy context of each analyst server is updated through cumulative injection operations, using the following formula: in This indicates that the compressed text gradient will be appended to the "memo" section of the corresponding analyst system prompt.

[0052] Review LLM prompts: { You are an objective review auditor. The core principle is to be factual and objective, and criticism for the sake of criticism is strictly prohibited. During the audit, you need to compare the following input information: current situation, role name, original plan and thought process, and actual outcome.

[0053] The task is to compare the thought process with the actual results and output one of three scenarios, only the corresponding tag content, without any explanation. The first scenario is that the thought process has a substantial error, i.e., a logical flaw leading to loss or missing out. In this case, output three tags: [LOOPHOLE] extract the specific erroneous statement from the original plan, [CORRECTION] provide the corrected logic, and [LESSON] extract a lesson to avoid. The second scenario is that the thought process is generally correct but has room for improvement, such as being in the right direction but lacking details, like insufficient profit or excessive panic. In this case, output [POINT] to point out the imperfections and [IMPROVEMENT] to provide specific improvement suggestions (such as position sizing or stop-loss adjustments). The third scenario is that the thought process is completely correct, i.e., accurate prediction and perfect execution. In this case, output [WINNING_POINT] to point out the key judgment that determined the outcome and [SUGGESTION] to extract a general operational suggestion from the successful case. Please strictly follow this format and only output the tag content corresponding to the selected scenario. Example 2: Complete daily decision-making process: Multiple analysts generate heterogeneous signals in parallel: At 9:00, the central server issues a prediction task, four analysts perform parallel inference, and return a JSON signal within 12 seconds. (1) Market Technical Analyst: Read the daily line, MACD and RSI indicators for the past 20 days, and output {"Decision":"Buy","Confidence":0.82,"Analysis Report":"Price rises above the 20-day moving average, MACD crosses with increasing volume, short-term bullish"}; (2) Fundamental Analyst: Read quarterly financial report indicators (debt-to-equity ratio 58%, PE 12 times, ROE 14%), and output {"Decision":"Buy","Confidence Level":0.42,"Analysis Report":"Valuation is lower than the industry average, and financial indicators are healthy"}; (3) News industry analyst: Read the positive news about 5G overseas orders and output {"Decision":"Buy","Confidence":0.20,"Sentiment Temperature":0.6,"Analysis Report":"Orders improve long-term profit expectations"}; (4) Momentum Overflow Analyst: Compare the returns of individual stocks with those of China Unicom (600050). The correlation is 0.72 > 0.6 threshold. Output {"Decision":"Buy","Confidence":0.82,"Linkage Status":"Strong","Spillover Effect":"Catch-up"}.

[0054] Central server reliability awareness fusion (1) Vector retrieval: The four signals are encoded into vectors and retrieved from the database. The top three historical scenes with the highest similarity are selected, and the mean of the Bayesian prior weights is calculated: Market 0.350, Fundamentals 0.298, News 0.199, Momentum 0.153; (2) Arbitration LLM logic score: Market 0.90, Fundamentals 0.85, News 0.70, Momentum 0.60; (3) Normalized final weights: Market 0.385, Fundamentals 0.314, News 0.180, Momentum 0.121; (4) Combine the weights and the original signals to generate the comprehensive market state Ct text.

[0055] Constraint scenario simulation and risk pruning (1) The LLM was reviewed and three branches were generated: wait and see, small trial with 10% position, and aggressive buying with 50% position; (2) Three types of stress tests: downward plunge (maximum drawdown -5%), sideways consolidation, and upward positive (returns of 3% to +8%). (3) Calculate the profit / loss ratio: trial branch profit / loss = 3 / 1, aggressive branch = 8 / 6 ≈ 1.33; (4) Risk LLM verification CVaR drawdown constraint: both branches meet the 10% maximum drawdown threshold, and the 10% small trial with the higher profit-loss ratio is selected as the final trading instruction.

[0056] Post-market review and analysis The actual market data is obtained at the close of trading that day. The LLM is reviewed and compared with the prediction logic and the actual trend. A text gradient is generated: "The trading volume shrank in the afternoon, and the market analyst's optimistic prediction was biased; since the complete market data in the morning was not obtained, the market analyst's weight was reduced by 0.1". This reflection vector is stored in the database and added to the market analyst's prompt memo. The model weights are not modified throughout the process.

[0057] Weekly knowledge distillation: Every Friday after the market closes, the distillation process begins, summarizing all reflection texts from the past five days. The LLM filter removes noise from occasional daily fluctuations and extracts general strategy memos: ① Increase the weight of news analysts by 0.15 on policy announcement days; ② Reduce the weight of momentum overflow signals when market volatility is below 15%. Broadcast these two strategies to all analyst servers, solidify them into the role's permanent prompts, and complete a week of adaptive iteration. On the last trading day of each week, the central server performs weekly knowledge distillation, inputting all reflective strategy adjustments from the past 7 days into the distillation LLM, which is then processed by a non-linear distillation operator. After extracting common strategic principles and removing high-frequency noise, we obtain: Specifically, the distillation LLM reviews all reflective strategy adjustments made over the past 7 days, outputs a refined memo JSON text, and then broadcasts it to the agent LLMs on all central servers. This memo is then merged into the persistent part of their system prompts to reduce recent errors and achieve adaptive model evolution. Ultimately, the agent LLMs on each central server, through continuous memorization and optimization of similar scenarios, achieve accurate, robust, and interpretable stock prediction services.

[0058] Example 3: Backtesting Results Verification: We selected four A-share stocks from January 1, 2024 to June 1, 2024 for simulated backtesting, with an initial capital of 1 million yuan, a small trial position of 10%, and an aggressive position of 50%. The proposed solution achieved an annualized return of 15% to 33% in backtesting, with a Sharpe ratio of 1.25 to 3.79, and the maximum drawdown was controlled within 12.97%. Compared with the traditional TradingAgents multi-agent solution, which has an annualized return of 2.17%, a Sharpe ratio of 0.37, and high drawdown, the solution significantly improves both return and risk control. The average time for a single decision inference is 6.8 seconds, the token consumption per round is approximately 4200, and communication overhead is reduced by more than 95%.

[0059] Those skilled in the art can replace different base models such as Llama-3 and Qwen as needed, and adjust parameters such as relevance threshold, maximum drawdown constraint, and number of retrievals K, to achieve the core function of this invention without creative effort.

[0060] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely exemplary descriptions of the invention as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. An intelligent decision-making method based on reliability perception fusion and constraint scenario inference, characterized in that, An application is made to a distributed multi-server financial forecasting framework, the framework comprising a central fusion server and multiple analyst servers, the method comprising the following steps: S1. Construct a distributed prediction framework for the observable Markov decision process in the graph enhancement part. Deploy role-differentiated large language model agents on each analyst server to analyze technical indicators, fundamentals, news semantics, and stock correlations, and output structured heterogeneous signals. S2. The heterogeneous signals output by each analyst server are aggregated to the central fusion server. The reliability-aware fusion module retrieves historical similar scenarios to generate Bayesian prior weights, performs logical consistency scoring on each signal, and calculates adaptive weights to resolve conflicts and obtain a unified state representation. S3. Input the state representation after conflict resolution into the constraint scenario review module, and generate three candidate strategies in parallel: wait and see, light position trial, and heavy position aggressive. For each strategy, perform three counterfactual scenarios of downtrend, oscillation, and uptrend, and calculate the profit and loss asymmetry ratio. Combined with the conditional value at risk constraint, perform risk pruning to select the optimal trading action. S4. The central fusion server compares the predicted action returns with the actual market returns, and uses a reflective strategy to modify and generate natural language text gradient updates for each agent's prompts. Every week, knowledge distillation is performed to extract general strategy principles, which are then solidified into the underlying prompts of each analyst agent, enabling the model to adapt and evolve while keeping the model weights frozen throughout the process.

2. The intelligent decision-making method based on reliability perception fusion and constraint scenario inference according to claim 1, characterized in that, In step S1, the analyst server includes a market analyst server, a fundamental analyst server, a news and industry analyst server, and a momentum overflow analyst server; each server is equipped with an independently customized large language model for prompt words, and only outputs desensitized JSON structured signals, and the original price, financial report, and news data are not transmitted across servers.

3. The intelligent decision-making method based on reliability perception fusion and constraint scenario inference according to claim 1, characterized in that, The reliability perception fusion module described in step S2 uses the Chroma vector database to store historical scene vectors and calculates Bayesian prior weights by retrieving the top K similar historical scenes using cosine similarity. The arbitration big language model scores logical consistency from three dimensions: causal integrity, evidentiary sufficiency, and macro-contextual fit. Weighted normalization is then used to obtain the final weight of each signal.

4. The intelligent decision-making method based on reliability perception fusion and constraint scenario inference according to claim 1, characterized in that, In step S3, the constraint scenario review module uses Monte Carlo tree search to complete multi-branch counterfactual deduction; the high-risk language model sets the CVaR downside risk threshold as a hard constraint, performs feasibility mask filtering on candidate strategies, and selects the transaction action corresponding to the maximum profit-loss asymmetry ratio after retaining only the risk compliance branch.

5. The intelligent decision-making method based on reliability perception fusion and constraint scenario inference according to claim 1, characterized in that, The reflective strategy correction mechanism described in step S4 generates a text gradient containing error location, logic correction, and operational lessons learned at the daily closing, and appends it to the corresponding analyst agent's prompt word memo; weekly knowledge distillation filters out noise reflection caused by daily fluctuations, and outputs standardized long-term strategic memos broadcast to all analyst servers.

6. The intelligent decision-making method based on reliability perception fusion and constraint scenario inference according to claim 1, characterized in that, Each server uses a star topology HTTP communication, and the analyst server outputs a structured JSON signal only once. The total token consumption for a single decision is controlled between 3,000 and 5,000, which reduces the communication overhead by more than 95% compared to full-connection multi-round debate.

7. An intelligent decision-making system based on reliability perception fusion and constraint scenario inference, characterized in that, The system is used to execute the intelligent decision-making method according to any one of claims 1 to 6, the system comprising: Multiple analyst servers, each configured with an independent large language model intelligent agent, respectively correspond to market technical analysis, fundamental financial analysis, news industry sentiment analysis, and stock momentum spillover analysis, used to locally parse raw financial data and output heterogeneous structured signals; A central fusion server communicates with each analyst server via an internal network API, and includes: The reliability-aware fusion module is used to retrieve similar historical scenarios, calculate adaptive signal weights, and resolve signal conflicts among multiple agents. The constraint scenario review module is used for multi-branch counterfactual scenario simulation, risk constraint pruning, and outputting the optimal compliant transaction action. The adaptive evolution module includes a reflection and review sub-unit and a weekly knowledge distillation sub-unit, which are used to generate text gradients, extract general investment strategies, and update prompt words for each analyst agent. The vector database, message queue, and local log database store historical scenario vectors, server scheduling instructions, and daily decision and review logs, respectively.

8. The intelligent decision-making system based on reliability perception fusion and constraint scenario inference according to claim 7, characterized in that, The reliability perception fusion module has a built-in arbitration big language model, is configured with special prompt words for weight calculation, takes the current heterogeneous signal and the retrieved historical scene as input, and outputs the normalized multi-agent signal weights.

9. The intelligent decision-making system based on reliability perception fusion and constraint scenario inference according to claim 7, characterized in that, The constraint scenario review module includes a review big language model and a risk big language model; the review big language model generates three types of candidate position strategies and completes multi-scenario return and drawdown simulations; the risk big language model uses the upper limit of conditional value at risk as a constraint to intercept and filter strategies.

10. The intelligent decision-making system based on reliability perception fusion and constraint scenario inference according to claim 7, characterized in that, The adaptive evolution module freezes the network weights of all large language models throughout the process, and only modifies the prompt words of each analyst server system to achieve adaptive market concept drift.

Images