Housing policy intelligent decision-making auxiliary method and system based on large language model

Abstract

The invention provides a housing policy intelligent decision auxiliary method and system based on a large language model, and belongs to the technical field of housing policy decision support, and the method comprises the steps: constructing a market theory hypothesis based on a housing filtering mechanism; inputting proxy attributes and market variable data; based on an intelligent agent modeling technology, a simulation model for simulating housing decision behaviors of different income groups is constructed and operated, and market dynamics are deduced under various policy scenarios; key indexes such as transaction activeness, housing quality distribution and group structure in the simulation process are extracted and visualized; and finally, inputting the structured indexes into a large language model, and generating an intelligent decision-making auxiliary text containing scene analysis, result induction and policy suggestion. According to the method, the dynamic deduction ability of computational simulation and the semantic understanding ability of artificial intelligence are fused, refined and intelligent evaluation of the housing policy effect can be realized, and the scientificity and the foresight of housing policy design are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of housing policy decision support technology, and in particular to a method and system for intelligent decision support for housing policy based on a large language model. Background Technology

[0002] Housing is both a livelihood issue and a development issue. With the deepening of new urbanization, the liquidity of the housing market is directly related to the high-quality development of cities. Housing policy making is a complex decision-making process involving multiple stakeholders. Existing housing policy making and evaluation technologies often face the following problems: In terms of data, most existing analyses are based on low-frequency data such as population censuses, housing surveys, and housing price indices, exploring the impact of different policies on overall housing prices, investment, and other macroeconomic indicators, which are difficult to reflect rapidly changing market dynamics; in terms of models, they mostly rely on static statistical methods such as regression models and econometric models, making it difficult to capture the decision-making behavior of diverse and heterogeneous stakeholders and their nonlinear impact on market liquidity; in terms of policy interaction, current evaluations are mostly focused on single policies, unable to provide policymakers with dynamic evaluation and quantitative comparison of different policy combinations in complex scenarios; in terms of decision support, existing tools are mainly based on macroeconomic statistical analysis and empirical rules, lacking interactive and predictable intelligent assistance capabilities. Therefore, there is an urgent need to introduce a systematic solution based on intelligent agent simulation and large language models for housing policy simulation and intelligent decision support. Summary of the Invention

[0003] This invention aims to solve at least one of the technical problems existing in the prior art, and proposes a method and system for intelligent decision support for housing policy based on a large language model.

[0004] In a first aspect, embodiments of the present invention provide a housing policy intelligent decision support method based on a large language model, comprising:

[0005] S1. Constructing theoretical hypotheses for the housing market based on housing filtering mechanisms;

[0006] S2. Input agent attribute data and housing market variable data as initialization conditions for the agent model;

[0007] S3. Based on the theoretical assumptions and input data, construct and run an intelligent agent model to simulate housing transaction behavior decisions and market evolution of heterogeneous agents under different policy scenarios;

[0008] S4. Extract and visualize the micro-behavioral features and macro-market indicators during the operation of the intelligent agent model;

[0009] S5. Input the macroeconomic market indicators into the large language model in a structured manner to generate intelligent decision-making assistance text that includes policy scenario analysis, result summarization and policy recommendations.

[0010] Furthermore, the housing market theory assumptions based on the housing filtering mechanism in S1 include:

[0011] H1. Heterogeneity of market participants: The housing market is composed of three heterogeneous agents: high-income, middle-income, and low-income, who differ in their housing preferences, affordability, credit availability, and risk tolerance.

[0012] H2, the dual constraint assumption of housing transactions: the housing transaction behavior of the agent is simultaneously affected by financial constraints and market constraints;

[0013] H3. The dynamic evolution hypothesis of housing filtering mechanism: market liquidity affects the housing replacement chain between different income groups;

[0014] H4. Government Policy Intervention Hypothesis: The government intervenes in the market through tax, fiscal, financial, and rental policy tools to adjust the efficiency of the filtering mechanism and housing equity.

[0015] Furthermore, the agent attribute data in S2 includes the agent's income group, initial housing status, and housing quality; the housing market variable data includes demand-side factors such as the price-to-income ratio, income growth rate, loan interest rate, down payment ratio, and housing subsidies, as well as supply-side factors such as market liquidity, second-hand housing price-to-income ratio, second-hand housing transaction tax, and existing housing / household ratio.

[0016] Furthermore, the agent model in S3 sets a home purchase decision function and a home sale decision function for each type of agent;

[0017] The home purchase decision function is based on the standardized price-to-income ratio, income growth rate, loan interest rate, down payment ratio, and home purchase subsidy, combined with the sensitivity coefficients of different income groups to calculate the probability of home purchase;

[0018] The house-selling decision function is based on standardized market liquidity, the price-to-income ratio of second-hand houses, second-hand house transaction tax, and the housing stock / household ratio, combined with sensitivity coefficients for different income groups to calculate the probability of selling a house.

[0019] Furthermore, the home purchase decision function is as follows:

[0020] ;

[0021] in, For the home purchase decision function, These are the current housing market variables: price-to-income ratio, income growth rate, loan interest rate, down payment ratio, and housing subsidies. , , , , These are weighted parameters representing the differences in the sensitivity of home purchase decisions among different income groups, with lower weighted parameters for higher income groups.

[0022] Furthermore, the house-selling decision function is as follows:

[0023] ;

[0024] in, For the decision function of selling a house, These are the current housing market variables: market liquidity, the ratio of existing home prices to income, existing home transaction taxes, and the ratio of existing homes to households. , , , These are the weighting parameters representing the differences in the sensitivity of home-selling decisions among different income groups.

[0025] Furthermore, the macro market indicators extracted in S4 include: new home transaction volume, second-hand home transaction volume, rental market transaction volume, average housing quality, proportion of low-quality housing, number of homeowners and renters in different income groups, and number of inter-group exchange behaviors.

[0026] Furthermore, S4 also includes generating dynamic visualization charts based on extracted indicators, including market transaction volume curves, housing quality trend charts, population rental and purchase structure stacking charts, and a visualization interface that dynamically displays changes in agent status through periodic grids.

[0027] Furthermore, S5 specifically includes:

[0028] S51. Configure system prompts and user prompts. The system prompts are used to set the role and task framework of the large language model. The user prompts are used to structure and fill in the macro market indicators obtained in step S4.

[0029] S52. Call the large language model interface, input the configured prompt words, and generate a policy analysis summary for the simulated scenario;

[0030] S53. Output the generated summary after verifying its format and content rationality.

[0031] In a first aspect, embodiments of the present invention provide a housing policy intelligent decision support system based on a large language model, comprising:

[0032] The theoretical hypothesis building module is used to construct theoretical hypotheses about the housing market based on housing filtering mechanisms;

[0033] The data input module is used to input agent attribute data and housing market variable data;

[0034] The intelligent agent modeling and simulation module is used to construct an intelligent agent model based on the theoretical assumptions and input data, run simulations, and output market evolution results.

[0035] The feature extraction and visualization module is used to extract macro market indicators from simulation results and visualize them.

[0036] The intelligent decision support module is used to input the macro market indicators into the large language model and generate intelligent decision support text.

[0037] The technical effects of the intelligent decision-making support method for housing policy based on large language model disclosed in this invention are as follows: This method, by integrating ABM and LLM, achieves dynamic, refined simulation and intelligent interpretation of the effects of housing policies. Its technical effects are: 1) Through heterogeneous agent modeling, it can more realistically simulate market micro-behavior, identify the differentiated impact of policies on different groups, and improve the targeting of policy design (corresponding to S1, S3); 2) Multi-scenario simulation and visualization functions provide an intuitive experimental platform for the quantitative comparison of policy combinations, supporting iterative optimization (corresponding to S3, S4); 3) Utilizing LLM to automatically generate analysis reports transforms complex simulation data into easily understandable decision references, improving the system's intelligence level and decision-making support efficiency (corresponding to S5).

[0038] The technical advantages of the intelligent decision-making support system for housing policy based on a large language model disclosed in this invention are as follows: This system integrates theory, models, data, computation, and interactive interface into one system (corresponding to claim 10), providing a complete and user-friendly policy simulation experimental platform. Its technical advantages are: 1) Integration: Integrating the dispersed modeling, simulation, analysis, and decision-making steps into a single software environment lowers the barrier to entry and improves research efficiency; 2) Interactivity: Users can adjust parameters in real time through the interface, observe market evolution animations, and obtain intelligent analysis, forming a rapid feedback loop of "parameter adjustment - simulation operation - result interpretation"; 3) Scalability: Based on Python and modular design, it facilitates the addition of new agent types, behavioral rules, or the integration of more advanced AI models, making the system highly adaptable. Attached Figure Description

[0039] Figure 1 A flowchart illustrating a housing policy intelligent decision support method based on a large language model, provided as an embodiment of the present invention;

[0040] Figure 2 This is a diagram of the main interface of the housing policy simulation system provided in an embodiment of the present invention.

[0041] Figure 3 This is a schematic diagram illustrating the configuration of the system prompt in an embodiment of the present invention.

[0042] Figure 4 This is a schematic diagram illustrating the configuration of user prompts in an embodiment of the present invention.

[0043] Figure 5 This is a schematic diagram of policies generated by LLMs from multiple perspectives in an embodiment of the present invention. Detailed Implementation

[0044] To enable those skilled in the art to better understand the technical solutions of the present invention, exemplary embodiments of the present invention are described below in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0045] A housing policy intelligent decision support method based on a large language model, referencing Figures 1 to 5 As shown, the specific steps include:

[0046] Step S1: Construct theoretical hypotheses for the housing market based on housing filtering mechanisms.

[0047] By combining housing filtering mechanisms with typical housing market experience, a set of structural and behavioral core assumptions are set as the basis for agent modeling in urban housing markets.

[0048] A1. Heterogeneity Assumption of Market Participants (H1): The housing market consists of three heterogeneous agents: high-income groups, middle-income groups, and low-income groups. These three groups differ in their housing preferences, affordability, credit availability, and risk tolerance. High-income groups tend to upgrade their homes; middle-income groups primarily purchase their first homes; and low-income groups rely on government housing subsidies and rental market supply.

[0049] A2. The Dual Constraints Assumption of Housing Transactions (H2): The housing transaction behavior of agents is simultaneously affected by financial constraints (income level, loan interest rate, down payment ratio) and market constraints (market liquidity, second-hand housing transaction tax, housing stock, etc.). The dual constraints jointly determine whether the agent can complete the purchase, exchange, or rental decision.

[0050] A3. The dynamic evolution hypothesis of the housing filtering mechanism (H3): When market liquidity is high, high-income groups purchase new homes, and their original homes become affordable for low- and middle-income groups, forming a virtuous cycle of filtering. When market liquidity is insufficient or transaction taxes are high, high-income groups cannot easily upgrade their homes, and low- and middle-income groups find it difficult to transition from renting to buying, increasing pressure on the rental market.

[0051] A4. Government Policy Intervention Hypothesis (H4): The government intervenes in the housing market through various policy tools (tax policy, fiscal subsidies, financial policy, rental policy, etc.). Policy intervention can adjust housing filtering efficiency and affect housing accessibility and market fairness for high-income groups.

[0052] These assumptions provide the logical foundation for subsequent mathematical models and agent behavior rules, enabling simulation models to more realistically reflect the complex interactions among multiple stakeholders in the real market.

[0053] Step S2: Input agent attribute data and housing market variable data.

[0054] This module provides the ABM housing filtering simulation system with two core input data levels: agent attribute data and housing market variable data, which together constitute the initial conditions and behavioral driving basis for the operation of the ABM model.

[0055] B1. Agent Attribute Setting: Housing market participants are defined as high-income, middle-income, and low-income agents. Referring to the "China Income Distribution Annual Report 2023" and the actual distribution characteristics of income stratification groups in typical urban housing markets, the population probability is allocated to 20%, 50%, and 30% respectively. In the initial housing status setting, all high-income groups own their homes, 80% of middle-income groups own homes, and 60% of low-income groups own homes; the remainder enter the rental market. Each Household Agent has the following attributes: unique identifier (uid), income group (group), homeownership status (has_house), tenant status (is_renter), housing quality (house_quality) (or rental_quality if tenant), and whether it is a newly purchased home (is_new_home).

[0056] This invention uses a continuous housing quality index q∈[0.5, 5] to characterize the physical and comprehensive quality of a single housing unit, where a higher value indicates better quality. Housing quality decreases over time, reflecting the physical depreciation process of the house, and directly affecting the agent's willingness to upgrade their housing. For agents who own their own housing, housing quality decays at a fixed depreciation rate delta = 0.04 in each decision period, with a maximum quality of 5.0 for newly built housing. Housing quality is initialized based on income level differences: the quality of owner-occupied housing for high-income groups is concentrated in the range [4,5], for middle-income groups it is [2.5,4], and for low-income groups it is [0.5,3]. The quality of rental housing is randomly set within the corresponding range, providing a structural basis for subsequent housing filtering, upgrade replacement, and rental-purchase migration in the model. During the simulation, high-income groups mainly upgrade their housing by purchasing new homes, releasing existing housing stock; middle- and low-income groups mainly upgrade their housing by accepting filtered housing, while those who fail to purchase homes enter the rental market. All agents make decisions on purchasing, selling, exchanging, and upgrading their housing based on the set behavioral rules in each time period.

[0057] B2. Housing Market Variable Setting: The dynamic evolution of the housing market essentially stems from the gradual decision-making behavior of micro-agents under both supply and demand conditions. On the one hand, the market must be able to effectively release and provide available housing resources (i.e., housing supply capacity); on the other hand, different income groups must possess the purchasing power and willingness to complete transactions (i.e., effective demand from housing agents). Accordingly, this system sets housing market variables according to the supply and demand logic, as follows:

[0058] B21. Supply-Side Variable Settings: Supply-side variables determine the housing supply release capacity and availability within the housing filter chain. On the supply side, Market Liquidity (ML) measures the transaction efficiency of the secondary housing and rental markets, determining whether upstream housing supply can be released smoothly; its value ranges from 0% to 100%. The Resale Price-to-Income Ratio (RPIR) reflects the affordability of filtered housing for low- and middle-income groups, ranging from 1.0 to 10.0; excessively high values ​​can hinder the filter chain's capacity. The Secondary Housing Transaction Tax (ST) represents the marginal cost of upgrading housing; excessively high ST values ​​can suppress transaction activity, and are set at 0%-10% in the model. Finally, the Housing Stock-to-Household Ratio (HSR) reflects the sufficiency of market supply, ranging from 0.1 to 5.0; higher values ​​indicate ample housing supply, which is beneficial for the smooth operation of the filter chain.

[0059] B22. Demand-Side Variable Setting: Demand-side variables determine whether an agent is capable of completing a transaction and successfully entering the filtering chain. On the demand side, the price-to-income ratio (PIR) is the most crucial threshold variable, measuring the ratio of overall housing prices to household income; its value ranges from 5 to 40, and a value exceeding 30 often signifies a severe affordability crisis. Income growth (IG) determines changes in household purchasing power, and in the model, it is set at -5% to 10%, covering scenarios from economic contraction to high growth. The loan rate (LR) and down payment ratio (DPR) together constitute financial constraints, set at 3%–8% and 10%–50% respectively; rising interest rates or excessively high down payments significantly inhibit the willingness to purchase a home. The government housing subsidy (GS) simulates government fiscal intervention, with a range of 0%–20%. The higher the subsidy, the easier it is for low- and middle-income groups to cross the threshold and enter the market.

[0060] Table 1. Setting of key variables in the model

[0061]

[0062] Step S3: Build and run the intelligent agent model to perform multi-scenario simulation.

[0063] This module is the core of the system. Using agents as the basic unit, it drives the reproduction and simulation of the housing market under different policy environments by setting behavioral rules and multi-scenario parameters. The specific steps are as follows:

[0064] C1. Setting Behavioral Rules: At the beginning of each time step, this step first standardizes the input variables, i.e., the ratio of any market variable to the preset benchmark value, and then calculates the trigger probability of various decision-making behaviors of the agent.

[0065] C11. Home Purchase Decision: For agents without existing housing, their willingness to purchase a home is determined by their sensitivity to the price-to-income ratio, loan interest rates, income growth, and down payment ratio. The home purchase decision function is as follows:

[0066] ;

[0067] in, For the home purchase decision function, These are the current housing market variables: price-to-income ratio, income growth rate, loan interest rate, down payment ratio, and housing subsidies. , , , , These are the weighted parameters representing the differences in the sensitivity of home purchase decisions among different income groups, with lower weighted parameters for higher income groups. Higher income groups are less sensitive to the price-to-income ratio and credit constraints in their home purchase decisions; lower income groups show higher weighted parameters for down payment thresholds and home purchase subsidies. See Table 2 for reference.

[0068] Table 2 Sensitivity Coefficient for Home Purchase Decisions

[0069]

[0070] The decision function for selling a house is as follows:

[0071] ;

[0072] in, For the decision function of selling a house, These are the current housing market variables: market liquidity, the ratio of existing home prices to income, existing home transaction taxes, and the ratio of existing homes to households. , , , These are the weighting parameters for the differences in the sensitivity of home-selling decisions among different income groups. See Table 3.

[0073] Table 3 Sensitivity coefficient for selling a house

[0074]

[0075] C13. Home Swap and Upgrade Decisions: High-income families prioritize home swapping when housing quality declines and new housing supply is available. The quality of newly purchased homes is concentrated in the high-end range, and the vacated medium-to-high-quality existing homes flow into the secondary market. Low- and middle-income families, on the other hand, are more likely to upgrade by selecting homes from the secondary market that meet the minimum quality threshold, and their original homes are then released to even lower-income families or new families. To characterize the heterogeneous behavior of high-income and low- and middle-income families during housing quality changes, this invention constructs two types of home swapping decision mechanisms: spontaneous home swapping behavior (Swap) based on housing quality in high-income families and improvement-oriented upgrade behavior (Upgrade) based on random triggering in low- and middle-income families.

[0076] Swap refers to the quality of housing occupied by high-income families. When it is below the set threshold (e.g.) ), under the condition that there is a supply of new homes in the market (such as If the agent sells their original home and enters the home purchase decision sequence, this action is counted as a high-income home-changing event in the model. Mathematically, it is defined as:

[0077] ;

[0078] in, This indicates that the agent belongs to the high-income group; Indicates the housing quality threshold; This refers to the supply of new homes in the market.

[0079] Upgrade refers to the act of upgrading or exchanging housing by low- and middle-income families while they still own their own homes, triggered by a predetermined random probability. The Bernoulli distribution, used to simulate whether a one-off random trial occurs, is a discrete probability distribution and, in this invention, represents whether the upgrade or replacement occurs during that period. When low- and middle-income families still own their own homes (i.e.,...) When, using probability parameters Trigger upgrade behavior:

[0080] ;

[0081] Where λ is the probability parameter for an upgrade (e.g., 0.2), indicating that an upgrade occurs with probability λ (Upgrade=1) and does not occur with probability 1-λ (Upgrade=0). If an upgrade occurs, then: the agent sells their original home and adds it to the secondary housing market supply pool; the agent re-enters the home purchase decision-making process, based on their income group. Weighting determines whether to choose a new or used home.

[0082] C14. Matching rules for the secondary housing market and rental market: When the agent successfully enters the purchase intention state (i.e., it has been triggered)... If a new home is unavailable, the system will guide the agent to try and find suitable properties in the secondary housing market. If the agent has no property and the purchase attempt fails, the system will automatically enter the rental market without further evaluation. The matching process is based on an "acceptable housing quality threshold." "Settings: Among them, For agent i, the income group The corresponding maximum acceptable property quality is determined by the following matching rules:

[0083] ;

[0084] Among them, if there are housing quality issues in the second-hand housing pool If a property is available, the agent will acquire it and complete the transaction in a first-in, first-out order; if no suitable second-hand property is available, the agent will enter the rental market, and the system will assign it rental housing quality based on income group.

[0085] C2. Multi-Scenario Simulation Setting: To evaluate the impact of different policy interventions on housing market transaction behavior, housing quality evolution paths, and changes in housing population structure, this system sets up three mutually exclusive policy scenarios based on the ABM model: a baseline scenario, a credit stimulus scenario, and a fiscal subsidy scenario. Specifically, the baseline scenario simulates the self-organizing state of the housing system under natural market evolution without external intervention. The credit stimulus scenario simulates loosening monetary policies, such as lowering loan interest rates and reducing down payment thresholds, to promote home purchase conversion. The fiscal subsidy scenario simulates direct fiscal support measures, such as increasing the proportion of home purchase subsidies and reducing transaction taxes and fees, to improve home purchase accessibility for low-income groups. Each scenario is mapped to a specific combination of parameters and has a differentiated impact on agent decision-making based on behavioral functions. It should be noted that the above three scenarios are only typical policy combinations selected in the simulation process. The system itself allows users to flexibly set and combine these scenarios within a wider range of policy tools through a parameter adjustment panel, thereby realizing scenario extrapolation and analysis under multiple policy paths.

[0086] Table 4 Policy parameter settings under different policy scenarios

[0087]

[0088] Step S4: Feature extraction and visualization.

[0089] This module is responsible for extracting key feature indicators from micro-level agent behavior during simulation and visually displaying the dynamic evolution of the housing market in a graphical manner. Its function consists of two steps:

[0090] D1. Housing Market Feature Extraction: After initializing the ABM model, different policy scenarios can be continuously implemented in the system for n time steps. At each time step, the system will summarize the results of home purchase, sale, exchange, and rental behaviors at the micro-agent layer into a set of statistical indicators, including various transaction activities (new home transaction volume, second-hand home transaction volume, rental market transaction volume), average housing quality and the proportion of low-quality housing, inter-group replacement behavior (number of upgrades and replacements for low- and middle-income groups and number of home exchanges for high-income groups), and dynamic characteristics of population property rights structure (number of homeowners and renters among high, middle, and low-income groups). The above data is recorded in the history container in the form of time series, ensuring that the simulation output of each step can be tracked, compared, and reproduced for subsequent intelligent analysis and policy interpretation using the large language model.

[0091] D2. Market Trend Visualization Output: Based on the aforementioned time-series data, the system generates transaction volume curves for new homes, secondhand homes, and rental markets; upgrade and replacement curves for high-income and low-to-middle-income groups; trend charts of average housing quality and the proportion of low-quality housing; and a stacked chart of population rental and purchase structures. Furthermore, the system provides a 15×15 periodic grid built on CanvasGrid to dynamically visualize the position and status changes of each agent. Specifically, an agent's housing status is distinguished by shape (homeownership is marked as a circle, renting as a square); the agent's income group is distinguished by the color of the shape (high-income is marked as blue, middle-income as green, and low-income as red); and the agent's housing quality is distinguished by the area of ​​the shape (higher quality, larger shape area). New homes are highlighted with black circles. Based on the migration, transaction, and replacement behaviors during the ABM simulation process, updates are performed in real-time on a time-step basis. The interface supports single-step execution ("Step") and overall reset ("Reset") operations, thus intuitively presenting the micro-evolution process of the housing market.

[0092] Step S5: LLM Intelligent Assisted Decision Generation.

[0093] E1. Prompt Configuration: This step acts as an intermediary between the multi-agent simulation system and the large language model intelligent feedback system. Its core objective is to use the core dynamic indicators output from steps C and D (such as transaction activity, housing quality evolution, group replacement behavior, and population structure changes) as input features, and guide the large language model to automatically generate targeted policy recommendations from multiple perspectives through prompt engineering.

[0094] E11. Configure System Prompt: The system prompt defines the roles of the large language model (e.g., policymaker, regulator, researcher), the task logic framework, and the output format specifications, clarifying the standard format for policy scenario presentation, result summarization, and recommendation generation. Role-based design is necessary because the same simulation result may be interpreted differently depending on the analytical perspective: policymakers focus more on institutional design and tool innovation, regulators emphasize risk identification and regulatory mechanisms, while researchers tend to focus on model expansion and mechanism interpretation.

[0095] E12. Configure User Prompt: User prompt provides structured policy input parameters and model output data, covering core variables such as price-to-income ratio (PIR) and income growth rate (IG), as well as simulation result indicators such as new home transaction volume, housing quality evolution, and population distribution, serving as the basis for decision-making reasoning in the large language model.

[0096] E13. Iterative Testing and Evaluation: Through repeated testing of the output, the task instructions, constraints, and variable expressions in the prompts are gradually adjusted to ensure stable convergence. By checking whether all placeholders in the prompts are filled and whether the values ​​are within a reasonable range, inconsistencies in input are prevented. Finally, the final prompt content is determined when the generated result corresponding to the prompt reaches the minimum error.

[0097] E2. Intelligent generation of simulation summary: The system calls the large language model interface, takes the simulation input parameters and output indicators as prompt words, and generates a summary that includes scenario parameters, result summaries and policy recommendations.

[0098] E21, LLM Interface Configuration: By calling the GPT-4o interface provided by OpenAI, the system prompts and user prompts configured in step E1 are passed as input messages to the large language model. The interface parameters must explicitly specify key elements such as the model version being called (e.g., "gpt-4o"), sampling temperature (e.g., 0.2, used to control the stability and diversity of the generated results), and maximum output length (e.g., 3000 tokens, used to ensure the completeness of the summarized content).

[0099] E22. Structured Input Data Transmission: During the call process, the core simulation results obtained in steps C and D (new home transaction volume, second-hand home transaction volume, rental transaction volume, average housing quality, proportion of low-quality housing, group purchase and rental distribution, etc.) are structured and filled with placeholders in the user prompts to form a standardized input data package, ensuring that the large language model can accurately identify the simulation scene and parameter background.

[0100] E23. Output Content Generation and Standardization: After receiving input, the large language model automatically generates a simulation summary from multiple perspectives based on the role identities and task logic framework set by the prompt words. The output results are presented in a unified format of three parts: "policy scenario parameters - model result summary - policy recommendations," avoiding simple data restatement and emphasizing trend identification, mechanism logic, and the generation of targeted recommendations.

[0101] E24. Result Verification and Feedback Storage: The generated simulation summary must undergo format compliance checks (e.g., whether it contains the three parts, whether the number of suggested items meets the requirements) and content rationality checks (e.g., whether there are missing data or logical contradictions), and be stored together with historical summary results. The simulation summary generated by the large language model is only for auxiliary reference; the final decision must be interpreted, revised, and adopted by policymakers or research teams.

[0102] This embodiment employs a double-blind experimental design, inviting 30 domain experts to evaluate policy recommendations generated by the LLM (Learning by Managers) and those generated by human experts across multiple dimensions. The evaluation dimensions included relevance, logicality, heuristics, feasibility, and preference, using a 5-point Likert scale (1 = strongly disagree, 5 = strongly agree). To assess the consistency between the LLM and expert ratings, weighted Kappa coefficients and paired-samples t-tests were used to analyze the differences between the two sets of ratings. The results showed no significant difference between the LLM and human expert ratings on the relevance and logicality dimensions (p > 0.05), with weighted Kappa coefficients of 0.635 and 0.654, respectively, indicating "high consistency." In terms of practical usability: on the feasibility, preference, and heuristic dimensions, the LLM ratings were significantly lower than those of human experts (p < 0.05), and the weighted Kappa values ​​were low, ranging from 0.256 to 0.335, indicating poor consistency.

[0103] Table 5. Results of Consistency Test for Policy Feedback Quality Assessment Dimensions

[0104]

[0105] Note: Significance: , , NS indicates no significance. Kappa agreement strength is based on: <0 = poor, 0.00-0.20 = very low, 0.21-0.40 = low, 0.41-0.60 = moderate, 0.61-0.80 = high, and 0.81-1.00 = almost perfect.

[0106] Based on the same inventive concept, embodiments of the present invention also provide a housing policy intelligent decision support system based on a large language model, comprising:

[0107] The theoretical hypothesis building module is used to construct theoretical hypotheses about the housing market based on housing filtering mechanisms;

[0108] The data input module is used to input agent attribute data and housing market variable data;

[0109] The intelligent agent modeling and simulation module is used to construct an intelligent agent model based on the theoretical assumptions and input data, run simulations, and output market evolution results.

[0110] The feature extraction and visualization module is used to extract macro market indicators from simulation results and visualize them.

[0111] The intelligent decision support module is used to input the macro market indicators into the large language model and generate intelligent decision support text.

[0112] This system is implemented using Python. Through modular design, it integrates the core functions of housing filtering simulation, user interface interaction, and intelligent analysis into a single, functional software system. The overall structure consists of the following parts:

[0113] F1. Operating Environment and Dependencies: The system runs in a Python 3 environment. The main open-source libraries it depends on include: mesa (implementing the core logic of agent-based simulation), streamlit (building an interactive interface to implement parameter input, result display and visualization), numpy and matplotlib (used for calculation and trend plot drawing during the simulation process), and openai (used to call large language models to generate simulation summaries and policy recommendations).

[0114] F2. Functional Module Division: The system includes five main modules: ① Data Input and Parameter Setting Module: Key variables such as the house price-to-income ratio, income growth rate, and loan interest rate are set using sliders and selection boxes as the input basis for simulation; ② Agency Behavior and Market Evolution Module: Housing behavior models for different income groups are constructed, rules for buying, selling, renting, and exchanging houses are set, and their dynamic interactions in the market are simulated; ③ Housing Transaction Behavior Visualization Module: The housing market transaction situation at each step, changes in housing quality, and the evolution of population structure are displayed graphically; ④ Large Language Model Intelligent Summary Module: Policy analysis text is automatically generated based on the core indicators output by the simulation.

[0115] F3. Platform Integration and Usage: Users only need to install the required dependencies and enter "streamlitrun filename.py" in the command line to run the system locally. The interface provides a complete workflow including parameter configuration, result viewing, and summary generation. If the user does not provide the API key for the large language model, the system can also call the built-in static suggestions to ensure the integrity of basic functions.

[0116] Compared with the prior art, the advantages of this invention are:

[0117] 1. This housing market simulation system, by setting differentiated behavioral rules for high, middle, and low-income groups and combining them with multi-parameter configurable policy inputs (such as price-to-income ratio, loan interest rates, and subsidy ratios), can dynamically simulate the replacement, rental, and transaction behaviors among groups, realistically reflecting the evolution of housing filtering mechanisms. This behavior-oriented simulation approach helps identify filtering bottlenecks and housing equity issues, thereby improving the pertinence and scientific rigor of housing policy design.

[0118] 2. This system integrates Agent-Based Modeling (ABM) and Large Language Modeling (LLM) technologies to simultaneously simulate individual housing decision-making behavior and automatically generate policy recommendations, thereby constructing a feedback loop between micro-level behavior and macro-level policies. This approach overcomes the limitation of traditional models that can only perform one-way simulations, enabling simulation results to directly serve policy understanding and institutional optimization, thus improving the system's intelligence level and policy applicability.

[0119] 3. This system integrates core functions such as parameter configuration, behavioral simulation, result visualization, and intelligent decision-making. It can display market transaction activity, changes in housing quality, and the evolution of population structure in real time, and supports intelligent summarization in multi-role scenarios. This integrated design not only provides a visualized experimental environment for complex policy solutions, but also facilitates policy researchers, managers, and the public in understanding and evaluating the effects of housing policies, enhancing the system's promotional value in scientific research, teaching, and practice.

[0120] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A housing policy intelligent decision support method based on a large language model, characterized in that, include: S1. Constructing theoretical hypotheses for the housing market based on housing filtering mechanisms; S2. Input agent attribute data and housing market variable data as initialization conditions for the agent model; S3. Based on the theoretical assumptions and input data, construct and run an intelligent agent model to simulate housing transaction behavior decisions and market evolution of heterogeneous agents under different policy scenarios; S4. Extract and visualize the micro-behavioral features and macro-market indicators during the operation of the intelligent agent model; S5. Input the macroeconomic market indicators into the large language model in a structured manner to generate intelligent decision-making assistance text that includes policy scenario analysis, result summarization and policy recommendations.

2. The method according to claim 1, characterized in that, The housing market theory assumptions based on the housing filtering mechanism in S1 include: H1. Heterogeneity of market participants: The housing market is composed of three heterogeneous agents: high-income, middle-income, and low-income, who differ in their housing preferences, affordability, credit availability, and risk tolerance. H2, the dual constraint assumption of housing transactions: the housing transaction behavior of the agent is simultaneously affected by financial constraints and market constraints; H3. The dynamic evolution hypothesis of housing filtering mechanism: market liquidity affects the housing replacement chain between different income groups; H4. Government Policy Intervention Hypothesis: The government intervenes in the market through tax, fiscal, financial, and rental policy tools to adjust the efficiency of the filtering mechanism and housing equity.

3. The method according to claim 1, characterized in that, The agent attribute data in S2 includes the agent's income group, initial housing status, and housing quality; the housing market variable data includes demand-side factors such as the price-to-income ratio, income growth rate, loan interest rate, down payment ratio, and housing subsidies, as well as supply-side factors such as market liquidity, second-hand housing price-to-income ratio, second-hand housing transaction tax, and existing housing stock / household ratio.

4. The method according to claim 1, characterized in that, The agent model in S3 sets up a home purchase decision function and a home sale decision function for each type of agent; The home purchase decision function is based on the standardized price-to-income ratio, income growth rate, loan interest rate, down payment ratio, and home purchase subsidy, combined with the sensitivity coefficients of different income groups to calculate the probability of home purchase; The house-selling decision function is based on standardized market liquidity, the price-to-income ratio of second-hand houses, second-hand house transaction tax, and the housing stock / household ratio, combined with sensitivity coefficients for different income groups to calculate the probability of selling a house.

5. The method according to claim 4, characterized in that, The home purchase decision function is as follows: ； in, For the home purchase decision function, These are the current housing market variables: price-to-income ratio, income growth rate, loan interest rate, down payment ratio, and housing subsidies. , , , , These are weighted parameters representing the differences in the sensitivity of home purchase decisions among different income groups, with lower weighted parameters for higher income groups.

6. The method according to claim 4, characterized in that, The decision function for selling the house is as follows: ； in, For the decision function of selling a house, These are the current housing market variables: market liquidity, the ratio of existing home prices to income, existing home transaction taxes, and the ratio of existing homes to households. , , , These are the weighting parameters representing the differences in the sensitivity of home-selling decisions among different income groups.

7. The method according to claim 1, characterized in that, The macro market indicators extracted in S4 include: new home transaction volume, second-hand home transaction volume, rental market transaction volume, average housing quality, proportion of low-quality housing, number of homeowners and renters in different income groups, and number of inter-group exchange behaviors.

8. The method according to claim 1, characterized in that, The S4 also includes generating dynamic visualization charts based on the extracted indicators, including market transaction volume curves, housing quality trend charts, population rental and purchase structure stacking charts, and a visualization interface that dynamically displays changes in agent status through periodic grids.

9. The method according to claim 1, characterized in that, S5 specifically includes: S51. Configure system prompts and user prompts. The system prompts are used to set the role and task framework of the large language model. The user prompts are used to structure and fill in the macro market indicators obtained in step S4. S52. Call the large language model interface, input the configured prompt words, and generate a policy analysis summary for the simulated scenario; S53. Output the generated summary after verifying its format and content rationality.

10. A housing policy intelligent decision support system based on a large language model, characterized in that, include: The theoretical hypothesis building module is used to construct theoretical hypotheses about the housing market based on housing filtering mechanisms; The data input module is used to input agent attribute data and housing market variable data; The intelligent agent modeling and simulation module is used to construct an intelligent agent model based on the theoretical assumptions and input data, run simulations, and output market evolution results. The feature extraction and visualization module is used to extract macro market indicators from simulation results and visualize them. The intelligent decision support module is used to input the macro market indicators into the large language model and generate intelligent decision support text.

Images