An investment advisor evaluation method and system based on fixed-time-series pairing simulation

By using fixed-time input and a unified field format for paired simulation trajectory recording, the problem of unstable trajectory reproduction and comparison in conversational investment advisor evaluation is solved, thus achieving a more accurate evaluation of advisor strategies.

CN123115822APending Publication Date: 2026-11-17WUHAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611578087.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-13
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing conversational investment advisor evaluation methods are difficult to reliably reproduce and directly compare the trajectories of different advisor strategies in long-term simulations, and the inconsistent trajectories recording formats make it difficult to trace the sources of differences.

Method used

By adopting a fixed time-series input and paired simulation trajectory method, the first and second simulation instances are established by acquiring the same historical market time-series data and investor initialization, generating and recording trajectories with a unified field format, and calculating paired differential evaluation data.

Benefits of technology

It improves the repeatability of comparisons between different consultant strategies, reduces the impact of market and initialization differences, and enhances the independence and auditability of the simulation process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115822A_ABST
    Figure CN123115822A_ABST
Patent Text Reader

Abstract

This invention provides an investment advisor evaluation method and system based on fixed-time-series paired simulation. The method acquires fixed historical market time-series data, investor initialization data, the advisor strategy to be evaluated, a reference advisor strategy, and simulator configuration data. Under identical initialization, simulator configuration, and common random process control, two simulation instances are established. The same market observation data is input step-by-step, and advisor messages are generated by the two advisor strategies respectively. The investor simulator generates execution actions accordingly and updates the portfolio account status based on the same market feedback after the actions are determined. A first trajectory and a second trajectory are formed according to unified fields, preset trajectory indicators are calculated for each, and pairwise differences are performed to output evaluation data. This invention can reduce the interference of external input, initialization, and random sampling differences on strategy comparison, and improve the repeatability and traceability of long-term session strategy evaluation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence evaluation technology, and in particular to an investment advisor evaluation method and system based on fixed-time-series pairing simulation. Background Technology

[0002] With the development of large language models and conversational intelligent agents, conversational investment advisors can generate risk statements, market information, portfolio-related explanations, or behavioral suggestions based on market observations, investor-visible information, portfolio account status, and dialogue records. Advisor messages may influence investors' subsequent actions and change the subsequent portfolio account status through actual market feedback. Therefore, the evaluation of conversational investment advisors involves not only a single response but also a long-term interaction trajectory formed over multiple time steps.

[0003] Current evaluations of conversational investment advisors typically focus on a single round of responses, assessing results based on accuracy, compliance, completeness, or personalization. Another approach evaluates subsequent behavior based on existing trading logs or backtesting results. While these methods can reflect the content of responses or past outcomes, real-world logs usually only contain one track that has already occurred, making it difficult to obtain corresponding tracks for the same investor using different advisor strategies under the same external market conditions.

[0004] In long-term simulations, if different advisory strategies use different market time series inputs, investor initializations, simulator configurations, or random sampling states, the differences between the two sets of output trajectories simultaneously include differences in advisory strategies as well as differences in external inputs, initializations, and sampling. When subsequent actions change subsequent holdings, cash, and portfolio net asset value, these differences will continue to propagate to later time steps, making it difficult to stably reproduce and directly compare the simulation results between different advisory strategies.

[0005] Furthermore, if the simulation process only retains the final results, or if different runs use different record fields, the subsequent evaluation module will find it difficult to locate the source of trajectory differences by indexing the same investor and the same time. Therefore, there is a need for a conversational investment advisory evaluation method and system that can fix external time-series inputs and initialization conditions, generate paired simulation trajectories with corresponding field structures, and perform pairwise difference calculations based on trajectory records. Summary of the Invention

[0006] The purpose of this invention is to provide an investment advisor evaluation method and system for fixed-time-series paired simulation, which solves the problems of difficulty in comparing trajectories of different advisor strategies in long-term simulations due to inconsistencies in external time-series inputs, investor initialization, simulator configuration, or random sampling states, as well as the difficulty in tracing the source of differences due to inconsistent trajectory recording formats.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A conversational investment advisor evaluation method based on fixed-time input and paired simulation trajectories includes: Acquire fixed historical market time series data, investor initialization data, advisor strategies to be evaluated, reference advisor strategies, and preset simulator configuration data. The fixed historical market time series data includes market observation data and market feedback data, and the investor initialization data includes the initial portfolio account status and initial dialogue records. A first simulation instance and a second simulation instance are established based on the same investor initialization data and simulator configuration data. At each time step, the same market observation data corresponding to the time step is input into the first simulation instance and the second simulation instance respectively. The advisor strategy to be evaluated and the reference advisor strategy generate advisor messages based on the unified observable input in their respective simulation instances. The investor simulator generates execution actions based on market observation data, portfolio account status, dialogue records, and corresponding advisor messages at the current time step. It also updates the portfolio account status of the first and second simulation instances based on the same market feedback data at the corresponding time step. Simultaneously, it records time stamps, market observation data, advisor messages, execution actions, and portfolio account status in a unified field format to form a first trajectory and a second trajectory. The first trajectory and the second trajectory use the same fixed historical market time series data, investor initialization data, simulator parameters, action space, and the same sampling control for a common random process, but the advisor strategies are different. Preset fields are read from the first trajectory and the second trajectory, preset trajectory indicators are calculated respectively, and the first trajectory indicator value and the second trajectory indicator value corresponding to the initial data of the same investor are differentially divided in pairs to generate and output the differential evaluation data.

[0008] In one alternative implementation, the unified observable input includes market observation data at the current time step, visible fields in investor initialization data, the current portfolio account status, and the current conversation history; the advisor strategy to be evaluated and the reference advisor strategy receive the unified observable input through an input interface with the same field structure.

[0009] In one optional implementation, the investor simulator employs a multi-agent structure, including a risk signal processing sub-agent, a return signal processing sub-agent, an account constraint processing sub-agent, and an action aggregation sub-agent. The risk signal processing sub-agent is used to process data related to drawdown, volatility, portfolio concentration, and risk level. The return signal processing sub-agent is used to process data related to asset returns, price or net asset value trends, and relative performance. The account constraint processing sub-agent is used to process data related to available cash, portfolio ratio, trading constraints, and investor initialization constraints. The action aggregation sub-agent is used to aggregate the structured outputs of each processing sub-agent and generate execution actions.

[0010] In one optional implementation, the risk signal processing sub-agent, the profit signal processing sub-agent, and the account constraint processing sub-agent each output action score data corresponding to a preset action code, and the action score data is represented by a preset numerical field; the action aggregation sub-agent aggregates the action score data of each processing sub-agent corresponding to the same action code according to a preset aggregation weight, and determines the action to be executed based on the aggregation result, wherein the preset action code includes at least one or more of the following: buy, sell, or hold.

[0011] In one optional implementation, both the first trajectory and the second trajectory include step records for multiple time steps. Each step record includes at least an investor initialization identifier, a time identifier, market observation data, advisor messages, executed actions, and combined account status. The step records are stored in a uniform field format and are read according to the investor initialization identifier and the time identifier.

[0012] In one optional implementation, for the shared random processes in the first simulation instance and the second simulation instance, the same random seed, pre-generated random sequence, or other equivalent sampling control method is adopted; when the two advisory strategies produce different execution actions, the combined account state of the first simulation instance and the second simulation instance evolves according to their respective execution actions, but reads the same market feedback data at the same time step.

[0013] In one optional implementation, the preset trajectory indicators include at least one of the following: cumulative return indicators, annualized return indicators, maximum drawdown indicators, volatility indicators, or risk-adjusted return indicators calculated based on portfolio net asset value or asset return data; buy count, sell count, trading frequency, or holding duration indicators calculated based on execution action records; and net cash flow indicators or cash retention indicators calculated based on account cash flow records; the paired differential evaluation data is obtained by subtracting the first trajectory indicator value and the second trajectory indicator value corresponding to the initial data of the same investor according to the preset differential direction.

[0014] In one alternative implementation, compliance checks are performed on the advisor messages after the first and second trajectories are generated; the compliance checks generate compliance markers based on preset compliance rules, text classification models, or preset evaluation models, and the compliance markers are output as post-hoc diagnostic data and do not participate in the generation of the execution actions and the combined account status updates.

[0015] In one optional implementation, the fixed historical market time series data is further segmented according to a preset time interval or a preset market status label, and corresponding step records are read from the first trajectory and the second trajectory respectively according to the time identifier. The preset trajectory index and the paired differential evaluation data are recalculated in each segment to generate segmented evaluation results.

[0016] On the other hand, the present invention provides a conversational investment advisory evaluation system based on fixed-time input and paired simulation trajectories, comprising: The data access module is used to acquire fixed historical market time-series data, investor initialization data, advisor strategies to be evaluated, reference advisor strategies, and simulator configuration data, and to perform time index alignment and field standardization on the fixed historical market time-series data; the simulator configuration and calibration module is used to determine and fix the simulator configuration data based on preset configuration or statistical data of anonymous aggregated behavior of real investors. The simulation execution module is used to generate execution actions at each time step based on market observation data, portfolio account status, dialogue records and advisor messages, and update portfolio account status based on market feedback data. The paired trajectory generation module is used to run the advisor strategy to be evaluated and the reference advisor strategy under the same fixed historical market time series data, investor initialization data, simulator parameters, action space and the same sampling control for the common stochastic process, and generate the first trajectory and the second trajectory according to a unified field format. The indicator calculation and diagnosis module is used to read preset fields from the first trajectory and the second trajectory, calculate preset trajectory indicators, pairwise differential evaluation data, and optional posterior compliance diagnosis data; The results output module is used to output one or more of the following: paired differential evaluation data, aggregated statistics, segmented evaluation results, or posterior compliance diagnostic data.

[0017] Compared with the prior art, the present invention has the following beneficial effects: First, by fixing historical market time-series data, investor initialization data, simulator parameters, action space, and common stochastic process control, simulation differences caused by non-advisory strategy factors are reduced, improving the repeatability of comparisons between different advisory strategies. Second, by recording two long-term trajectories in a unified field format, advisor messages, executed actions, and combined account status at corresponding time steps can be directly aligned and tracked. Third, by calculating paired differential indicators for two trajectories under the same initialization conditions, the impact of market intervals and initialization differences on absolute indicators is reduced. Fourth, by separating the trajectory generation process from post-compliance diagnostics, the independence and auditability of the simulation process are improved. Fifth, by uniformly calibrating simulator parameters and fixing their configuration during the evaluation phase, the consistency of the simulation environment used by different advisory strategies is improved. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the conversational investment advisor assessment method provided by the present invention; Figure 2 This is a schematic diagram of a selectable multi-agent investor simulation structure provided by the present invention; Figure 3 This is a schematic diagram of the paired simulation trajectory evaluation process provided by the present invention; Figure 4 This is a schematic diagram of the structure of the conversational investment advisor assessment system provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] It should be noted that, in the embodiments of the present invention, "advisor strategy" refers to a computer program or model strategy that generates advisor messages based on a unified observable input, which can be implemented by a large language model, a rule system, a retrieval-enhanced generation system, a financial knowledge base system, or a combination of the above systems; "investor simulator" refers to a computing module that generates execution actions and drives the updating of the portfolio account based on market observation data, portfolio account status, dialogue records, and advisor messages; "fixed historical market time series data" refers to a historical market data sequence organized according to a unified time index and maintained in consistency across different advisor strategy evaluations; "paired simulation trajectory" refers to a first trajectory and a second trajectory generated separately by changing only the advisor strategy under common external time series input and common initialization conditions.

[0022] Example 1 like Figure 1 As shown, the investment advisor evaluation method based on fixed time-series pairing simulation provided in this embodiment of the invention includes steps 101 to 104.

[0023] Step 101: Obtain fixed historical market time series data, investor initialization data, advisor strategies to be evaluated, reference advisor strategies, and simulator configuration data.

[0024] The fixed historical market time-series data can be organized by trading day or other preset time granularity. Data records for each time step must distinguish between market observation data and market feedback data. Market observation data may include one or more of the following: time stamp, market index, asset stamp, fund or asset net value, asset return, risk indicator, or trend indicator; market feedback data includes actual returns, net value changes, or price changes used to update the portfolio account status after the end of that time step. Two simulation instances read market observation data and market feedback data for the same time step according to the same time index.

[0025] Within the same time step, market observation data is provided before advisor messages and actions are generated; market feedback data is used to update the portfolio account status after the actions are determined, and is not used as observable input when the advisor strategy to be evaluated, the reference advisor strategy, or the investor simulator generates actions at that time step. This time-series isolation avoids using market feedback information that has not yet occurred at the current time step to generate current advisor messages or actions.

[0026] The investor initialization data may include one or more of the following: risk level, financial literacy level, asset size, initial holdings, initial cash, and initial conversation records. This data can be generated from a preset configuration or based on anonymous aggregated statistics of real users or user profile clustering results. The same investor initialization data is copied as the common initial input for both the first and second simulation instances.

[0027] The advisor strategy to be evaluated refers to the conversational investment advisor strategy that needs to be compared, and the reference advisor strategy refers to the reference strategy used to form the corresponding trajectory. The reference advisor strategy can be a strategy that does not provide advisor information, a strategy that only provides neutral market information, a strategy that provides risk warnings according to fixed rules, or other preset comparison strategies. Different strategies receive current market observation data with the same field structure, visible fields in investor initialization data, combined account status, and dialogue records through a unified input interface, and output advisor messages.

[0028] The simulator configuration data is obtained through the following methods: acquiring anonymous aggregated behavior statistics of real investors; determining calibration targets based on the anonymous aggregated behavior statistics, wherein the calibration targets include one or more of the following: portfolio risk structure, buy response data under market uptrend conditions, and sell response data under market downtrend conditions; adjusting simulator parameters and calculating the difference between the simulated aggregated behavior statistics and the calibration targets; fixing simulator parameters that meet preset conditions or preset optimization targets, and using the simulator parameters for subsequent trajectory generation of different advisory strategies.

[0029] Step 102: Invoke the investor simulator to generate execution actions based on the advisor message and update the portfolio account status.

[0030] The investor simulator receives current market observation data, current portfolio account status, current dialogue records, corresponding advisor messages, and simulator configuration data at each time step, and outputs the actions to be executed from the preset action space. These actions can include one or more of buying, selling, and holding; in implementations requiring multiple rounds of consultation, the action space can also include consultation actions. Based on the executed actions and the market feedback data from the current time step, the simulator updates holdings, cash, portfolio net asset value, and other preset account fields, and uses the updated portfolio account status as input for the next time step.

[0031] like Figure 2 As shown, in one optional embodiment, the investor simulator adopts a multi-agent structure, including a risk signal processing sub-agent, a return signal processing sub-agent, an account constraint processing sub-agent, and an action aggregation sub-agent. The risk signal processing sub-agent reads preset fields such as maximum drawdown, volatility, portfolio concentration, and risk level from current market observation data and portfolio account status, and outputs first action score data corresponding to each preset action code; the return signal processing sub-agent reads preset fields such as asset return, net asset value change trend, and relative performance, and outputs second action score data corresponding to each preset action code; the account constraint processing sub-agent reads preset fields such as available cash, position ratio, trading constraints, and investor initialization constraints, and outputs third action score data corresponding to each preset action code.

[0032] The action scoring data is stored using a unified field structure, with each preset action code corresponding to a numerical score. The score range is pre-fixed within the same evaluation task, for example, using a numerical value between 0 and 1 or a preset integer level. The action aggregation sub-agent aggregates the first, second, and third action score data corresponding to the same action code according to preset aggregation weights to obtain the aggregate score for each action code, and determines the action code that meets the preset selection rules as the action to be executed. Thus, the input fields, output fields, and action aggregation relationships of each processing sub-agent are all represented using a preset data structure.

[0033] When the preset action space includes consultation actions and the consultation action is determined as an action to be executed by the action aggregation sub-agent, the system can select a consultation statement corresponding to the current market observation data and portfolio account status from the preset consultation template set, or the preset text generation module can generate a consultation statement based on the same fields and write it into the current dialogue record; the advisor strategy to be evaluated or the reference advisor strategy generates an advisor response based on the updated dialogue record. This process does not require each processing sub-agent to output a consultation action separately.

[0034] Before using an investor simulator to evaluate advisory strategies, the simulator configuration can be calibrated based on anonymized aggregated behavioral statistics of real investors. Calibration targets can include one or more of the following: portfolio risk structure, buy response under market uptrends, and sell response under market downtrends. The system generates candidate simulator parameters through a preset parameter search or optimization process and calculates the difference between the simulated aggregated behavioral statistics and the calibration targets to determine the simulator configuration that meets the preset conditions. After calibration, the determined simulator parameters are fixed for subsequent trajectory generation of different advisory strategies.

[0035] Step 103: Generate the first trajectory and the second trajectory under the same fixed historical market time series data and investor initialization conditions.

[0036] like Figure 3 As shown, for each investor's initialization data, the first and second simulation instances share the same fixed historical market time-series data, the same investor initialization data, the same simulator parameters, the same action space, and the same sampling control for a common stochastic process. The difference lies in that the first simulation instance calls the strategy of the advisor to be evaluated, while the second simulation instance calls the strategy of the reference advisor. Both simulations proceed according to the same time index.

[0037] Each trajectory consists of multiple time-step records. Each step record includes at least the investor initialization identifier, time identifier, market observation data, advisor messages, executed actions, and portfolio account status; in implementations with multiple rounds of consultation, consultation statements and updated dialogue records can also be recorded. Different trajectories are stored using a unified field structure for subsequent retrieval based on the investor initialization identifier and time identifier.

[0038] In one implementation, the first and second simulation instances use the same random seed, pre-generated random sequence, or other equivalent sampling control methods for the shareable stochastic process. Since the two advisory strategies may produce different actions, the subsequent combined account states of the two trajectories can evolve separately, but the same external market feedback data is still read at the same time step, thus preserving the long-term path differences generated by the advisory strategies.

[0039] The common random process refers to random elements in two simulation instances that have the same meaning and can share sampling control. If the investor simulator is a deterministic computation process, then no corresponding random sampling control is required; if the investor simulator includes a random sampling module, then the same random seed, pre-generated random sequence, or equivalent control method is used for the shareable random elements. Random processes that cannot be shared by the two advisory strategies themselves are not a necessary component of the common random process.

[0040] Step 104: Calculate the pairwise difference index of the paired simulation trajectory and output the evaluation data.

[0041] For each preset trajectory indicator, the system reads the corresponding fields from the first trajectory and the second trajectory according to the same investor initialization identifier, calculates the first trajectory indicator value and the second trajectory indicator value, and then calculates the difference between the two according to the preset difference direction to obtain the pairwise difference evaluation data corresponding to the investor initialization. For multiple investor initializations, the pairwise difference evaluation data can be averaged or grouped and aggregated according to preset weights, or statistical data such as mean, median, quantiles or confidence intervals can be output.

[0042] In one implementation, the preset trajectory indicators may include one or more of the following: cumulative return, annualized return, maximum drawdown, volatility, risk-adjusted return, number of buys, number of sells, trading frequency, holding duration, net cash change, or cash retention. Return and drawdown indicators are calculated separately from the combined net asset value or asset return fields in the first and second trajectories; action indicators are calculated separately from the executed action records in the two trajectories; and cash flow indicators are calculated separately from the account cash flow records in the two trajectories. All of the above indicators have clearly defined trajectory field sources.

[0043] In an optional implementation, a compliance check is performed after the trajectory is generated. The system processes the advisor's message according to preset compliance rules, a text classification model, or a preset evaluation model and outputs a compliance flag. The compliance check result is stored and output separately as post-hoc diagnostic data and is not used as input for the investor simulator to generate and execute actions or update the portfolio account status.

[0044] In an optional implementation, the system can also perform market interval segmentation analysis. Specifically, fixed historical market time series data can be divided into multiple intervals according to preset date boundaries, market status labels, or other preset conditions. Corresponding step records are extracted from the first and second trajectories based on time identifiers. Preset trajectory indicators and paired differential evaluation data are calculated in each interval to generate segmented evaluation results for different time intervals.

[0045] Example 2 This embodiment provides a set of exemplary parameters for illustrating the process of generating paired simulation trajectories and calculating paired differences. These parameters are used to illustrate specific implementations of the invention and do not constitute a limitation on the range of market data, trading rules, or indicator values.

[0046] Historical data of a single risky asset across five consecutive time steps are selected as fixed historical market time-series data. For ease of demonstration, the market feedback returns for the five time steps are set to 1%, -2%, 3%, 1%, and -1%, respectively. Investor initialization data includes initial cash of 100,000 yuan, initial portfolio value of 0 yuan, a risk level of "medium," and an empty initial dialogue record. The first simulation instance is connected to the advisor strategy A to be evaluated, and the second simulation instance is connected to the reference advisor strategy B; both instances use the same simulator parameters, action space, and common stochastic process control. In practical applications, market feedback data for each time step is directly read from the selected fixed historical market time-series data.

[0047] For ease of explanation, the preset action space in this embodiment is buy, sell, and hold; the buy action purchases the risky asset at 20% of the available cash at the time of the action, the sell action sells at 20% of the market value of the risky asset holdings at that time, and the hold action does not adjust the account holdings and does not incur transaction fees. The above transaction ratios and fee settings are only used for calculation explanation in this embodiment.

[0048] At each time step, both simulation instances first receive the same market observation data, but do not receive market feedback data that has not yet been used for account updates at that time step. After the corresponding advisor strategy generates advisor messages, the investor simulator generates execution actions respectively; after the execution actions are determined, the holding market value and portfolio net asset value are updated according to the same market feedback return rate at that time step. In this embodiment, the execution actions for the first trajectory are buy, hold, buy, sell, and hold in sequence; the execution actions for the second trajectory are buy, buy, sell, hold, and hold in sequence. The portfolio net asset value at each time step is shown in the table below.

[0049]

[0050] In this embodiment, the cumulative return R = (terminated portfolio NAV - initial portfolio NAV) / initial portfolio NAV × 100%; the maximum drawdown MDD is the maximum value of the ratio of the difference between the historical peak portfolio NAV and the current portfolio NAV to the historical peak portfolio NAV within the evaluation interval. For indicators where a larger value is better, the pairwise difference is calculated by subtracting the second trajectory indicator value from the first trajectory indicator value; for indicators where a smaller value is better, the pairwise difference is set in the opposite direction so that a positive value indicates the improvement of the advisor's strategy relative to the reference advisor's strategy.

[0051] According to the aforementioned trajectory records and calculation rules, the net asset value of the first trajectory's terminated portfolio was 100,866.93 yuan, with a cumulative return of 0.8669% and a maximum drawdown of 0.4032%; the net asset value of the second trajectory's terminated portfolio was 100,324.50 yuan, with a cumulative return of 0.3245% and a maximum drawdown of 0.7226%. For the cumulative return, calculated by subtracting the indicator value of the second trajectory from the first trajectory's indicator value, the paired difference is +0.5424 percentage points; for the maximum drawdown, calculated by subtracting the indicator value of the first trajectory from the second trajectory's indicator value, the paired difference is +0.3194 percentage points. The system outputs the above difference results, the indicator calculation direction, and the identification information of the two trajectories together.

[0052] When the evaluation object includes initial data from multiple investors, the above-mentioned paired simulation process can be repeated for each initial data, and the mean, median, quantile or confidence interval can be calculated for the paired difference results corresponding to each initialization, thereby forming aggregated evaluation data of the advisor strategy to be evaluated relative to the reference advisor strategy.

[0053] Example 3 like Figure 4 As shown in the figure, this embodiment provides an investment advisor evaluation system based on fixed-time-series paired simulation. The system includes a data access module, a simulator configuration and calibration module, a simulation execution module, a paired trajectory generation module, an indicator calculation and diagnosis module, and a result output module.

[0054] The data access module is used to import fixed historical market time series data, investor initialization data, advisor strategies to be evaluated, reference advisor strategies, and simulator configuration data. It also performs time index alignment, field cleaning, and field standardization on the fixed historical market time series data, and outputs data records in a unified format.

[0055] The simulator configuration and calibration module is used to read the preset simulator configuration, or search for simulator parameters based on the statistical data of anonymous aggregated behavior of real investors and the preset calibration target; after determining the simulator parameters, the relevant parameters are fixed and provided to the simulation execution module and the paired trajectory generation module.

[0056] The simulation execution module is used to read market observation data, current portfolio account status, current dialogue records and advisor messages at each time step, call the investor simulator to generate execution actions, and update the portfolio account status according to the market feedback data of the current time step. When a multi-agent structure is adopted, the simulation execution module further includes a risk signal processing sub-agent, a return signal processing sub-agent, an account constraint processing sub-agent, and an action aggregation sub-agent, and its inputs and outputs adopt the aforementioned unified structured fields.

[0057] The paired trajectory generation module is used to establish a first simulation instance and a second simulation instance based on the initialization data of the same investor. Under the same fixed historical market time series data, the same simulator parameters, the same action space, and the same sampling control for the common random process, the strategy of the advisor to be evaluated and the reference advisor strategy are run respectively, and the first trajectory and the second trajectory are recorded according to a unified field format.

[0058] The indicator calculation and diagnosis module is used to read preset fields from the first trajectory and the second trajectory according to the investor initialization identifier and the time identifier, calculate the preset trajectory indicators and the corresponding pairwise differential evaluation data respectively, and perform market interval segment calculation or post-compliance detection according to the configuration.

[0059] The output module is used to output one or more of the following: pairwise differential evaluation data, aggregated statistics, segmented evaluation results, or posterior compliance diagnostic data. The output results can be provided in the form of files, database records, interface-returned data, or visual reports.

[0060] Example 4 This embodiment provides an electronic device. The electronic device may include a processor, a memory, a communication interface, and a bus. The memory stores a computer program, and when the processor executes the computer program, it can implement the conversational investment advisory evaluation method described in any of the above embodiments.

[0061] Example 5 This embodiment provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, can implement the conversational investment advisory evaluation method described in any of the above embodiments.

[0062] Example 6 This invention also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, can implement the conversational investment advisory evaluation method described in any of the above embodiments.

[0063] It should be noted that the conversational investment advisor assessment results in this invention are used for offline simulation assessment and technical diagnostics, and do not directly constitute investment advice for real investors. In practical applications, the investor initialization fields, action space, simulator calibration targets, post-compliance rules, and indicator configurations can be adjusted according to specific asset types, market data sources, evaluation objectives, and compliance requirements. However, different advisor strategies to be compared should use a consistent fixed configuration in the same evaluation task.

[0064] For technical details not specifically described in this specification, those skilled in the art can implement them using conventional technical means adapted to the technical solutions of this invention.

[0065] It should be understood that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Those skilled in the art can adjust, substitute, or combine the above embodiments without departing from the concept of the present invention. The scope of protection of the present invention is determined by the appended claims.

Claims

1. An investment advisor evaluation method based on fixed-time-series paired simulation, characterized in that, include: Acquire fixed historical market time series data, investor initialization data, advisor strategies to be evaluated, reference advisor strategies, and preset simulator configuration data. The fixed historical market time series data includes market observation data and market feedback data, and the investor initialization data includes the initial portfolio account status and initial dialogue records. A first simulation instance and a second simulation instance are established based on the same investor initialization data and simulator configuration data. At each time step, the same market observation data corresponding to the time step is input into the first simulation instance and the second simulation instance respectively. The advisor strategy to be evaluated and the reference advisor strategy generate advisor messages based on the unified observable input in their respective simulation instances. The investor simulator generates execution actions based on market observation data, portfolio account status, dialogue records, and corresponding advisor messages at the current time step. After the execution actions are generated, the portfolio account status of the first and second simulation instances is updated based on the same market feedback data at the corresponding time step. At the same time, the time stamp, market observation data, advisor messages, execution actions, and portfolio account status are recorded in a unified field format to form a first trajectory and a second trajectory. The first trajectory and the second trajectory use the same fixed historical market time series data, investor initialization data, simulator parameters, action space, and the same sampling control for a common random process, but the advisor strategies are different. Preset fields are read from the first trajectory and the second trajectory, preset trajectory indicators are calculated respectively, and the first trajectory indicator value and the second trajectory indicator value corresponding to the initial data of the same investor are differentially divided in pairs to generate and output the differential evaluation data.

2. The method of claim 1, wherein, The unified observable input includes market observation data at the current time step, visible fields in investor initialization data, the current portfolio account status, and the current conversation record; the advisor strategy to be evaluated and the reference advisor strategy receive the unified observable input through an input interface with the same field structure; The market feedback data is not part of the unified observable input at the current time step.

3. The method of claim 1, wherein, The investor simulator adopts a multi-agent structure, including a risk signal processing sub-agent, a return signal processing sub-agent, an account constraint processing sub-agent, and an action aggregation sub-agent. The risk signal processing sub-agent is used to process data related to drawdown, volatility, portfolio concentration, and risk level. The return signal processing sub-agent is used to process data related to asset returns, price or net asset value trends, and relative performance. The account constraint processing sub-agent is used to process data related to available cash, portfolio ratio, trading constraints, and investor initialization constraints. The action aggregation sub-agent is used to aggregate the structured outputs of each processing sub-agent and generate execution actions.

4. The method of claim 3, wherein, The risk signal processing sub-agent, the profit signal processing sub-agent, and the account constraint processing sub-agent each output action score data corresponding to a preset action code, and the action score data is represented by a preset numerical field; The action aggregation sub-agent aggregates the action score data corresponding to the same action code of each processing sub-agent according to the preset aggregation weight, and determines the action to be executed based on the aggregation result. The preset action code includes at least one or more of the following: buy, sell, or hold.

5. The method of claim 1, wherein, Both the first trajectory and the second trajectory include step records for multiple time steps. Each step record includes at least an investor initialization identifier, a time identifier, market observation data, advisor messages, executed actions, and combined account status. The step records are stored in a uniform field format and are read according to the investor initialization identifier and the time identifier.

6. The method of claim 1, wherein, For the shared stochastic processes in the first simulation instance and the second simulation instance, the same random seed, pre-generated random sequence, or other equivalent sampling control methods are adopted; when the two advisory strategies produce different execution actions, the combined account state of the first simulation instance and the second simulation instance evolves according to their respective execution actions, but reads the same market feedback data at the same time step.

7. The method of claim 1, wherein, The preset trajectory indicators include at least one of the following: cumulative return indicators, annualized return indicators, maximum drawdown indicators, volatility indicators, or risk-adjusted return indicators calculated based on portfolio net asset value or asset return data; buy count, sell count, trading frequency, or holding duration indicators calculated based on execution action records; and net capital change indicators or capital retention indicators calculated based on account cash flow records; the paired differential evaluation data is obtained by subtracting the first trajectory indicator value and the second trajectory indicator value corresponding to the initial data of the same investor according to the preset differential direction.

8. The method of claim 1, wherein, It also includes performing compliance checks on the advisor messages after the first trajectory and the second trajectory are generated; the compliance checks generate compliance markers based on preset compliance rules, text classification models or preset evaluation models, and the compliance markers are output as post-hoc diagnostic data and do not participate in the generation of the execution actions and the combined account status update.

9. The method of claim 1, wherein, It also includes segmenting the fixed historical market time series data according to a preset time interval or a preset market status label, reading the corresponding step records from the first trajectory and the second trajectory according to the time identifier, recalculating the preset trajectory index and the paired differential evaluation data in each segment, and generating segmented evaluation results.

10. An investment advisor evaluation system based on fixed timing pairing simulation, characterized by, include: The data access module is used to acquire fixed historical market time-series data, investor initialization data, advisor strategies to be evaluated, reference advisor strategies, and simulator configuration data, and to perform time index alignment and field standardization on the fixed historical market time-series data; the simulator configuration and calibration module is used to determine and fix the simulator configuration data based on preset configuration or statistical data of anonymous aggregated behavior of real investors. The simulation execution module is used to generate execution actions at each time step based on market observation data, portfolio account status, dialogue records and advisor messages, and update portfolio account status based on market feedback data. The paired trajectory generation module is used to run the advisor strategy to be evaluated and the reference advisor strategy under the same fixed historical market time series data, investor initialization data, simulator parameters, action space and the same sampling control for the common stochastic process, and generate the first trajectory and the second trajectory according to a unified field format. The indicator calculation and diagnosis module is used to read preset fields from the first trajectory and the second trajectory, calculate preset trajectory indicators, pairwise differential evaluation data, and optional posterior compliance diagnosis data; The result output module is used to output one or more of the pairwise differential evaluation data, aggregated statistics, segmented evaluation results, or posterior compliance diagnostic data. The investment advisor evaluation system based on fixed-time-series pairing simulation is used to perform the steps in the investment advisor evaluation method based on fixed-time-series pairing simulation as described in any one of claims 1-9.