A machine learning driven investment strategy configuration method

Abstract

The application discloses a kind of machine learning driven investment strategy configuration methods, it is related to investment strategy configuration technical field, including: the preference parameter in re-optimization mode router is regarded as hidden state variable, user feedback is mapped as noisy observation sequence, and state space model parameter set is constructed;The posterior mean and posterior variance are estimated by Kalman recursion;The hysteresis bandwidth is configured as the monotone non-decreasing function of posterior variance, and the uplink and downlink switching threshold is output in combination with current re-optimization mode and direction polarity;When posterior mean is continuously out of bounds in confirmation window and posterior variance is lower than preset upper limit, mode switching permission signal is output, otherwise output remains signal;When switching is allowed, activation mode and direction polarity are reversed, and investment strategy configuration instruction is output;The application effectively suppresses the frequent oscillation of strategy mode caused by the non-stationarity of user feedback by Kalman filtering, adaptive hysteresis band and double criteria, and improves the robustness and reliability of investment strategy configuration.

Description

Technical Field

[0001] This invention relates to the field of investment strategy allocation technology, and in particular to a machine learning-driven investment strategy allocation method. Background Technology

[0002] In the field of investment strategy allocation, adaptive adjustment mechanisms have become a key technological direction for improving portfolio management efficiency in order to cope with dynamic changes in the market environment. In existing technologies, machine learning systems are widely used to learn and track the personalized preferences of investment decision-makers. A common approach is to record users' confirmation or fine-tuning behavior regarding system recommendations and update preference parameters in the strategy routing module in real time, thereby gradually aligning the investment strategy generation process with the user's subjective intentions. Based on this idea, existing technologies have modeled the routing decisions of two re-optimization modes—partial repair and overall replanning—as online learning problems, using user interaction feedback as a supervisory signal to continuously correct the values ​​of preference parameters. However, in actual human-machine collaborative investment scenarios, user decision-making behavior is often not entirely rational or consistent. It may be affected by various factors such as emotional fluctuations, cognitive load, and external market noise, leading to significant randomness and inconsistencies in the confirmation or fine-tuning operations made by users at different times for the same or highly similar market disturbance scenarios. This non-stationary feedback signal makes the preference parameter estimation process based on online gradient descent or Bayesian updates unable to converge, resulting in the preference parameters continuously oscillating around the decision boundaries of the two re-optimization modes. Once preference parameters oscillate, the strategy routing system will frequently switch between local repair and overall replanning, causing significant fluctuations in investment plans and repeated adjustments to execution instructions. Existing online learning algorithms generally assume that user behavior follows a certain stationary distribution or at least has a consistent trend, lacking robustness to irrational fluctuations in real human-computer interaction. This makes it difficult to guarantee the stable convergence of preference parameters, thus limiting the reliability and practicality of adaptive investment strategies in actual deployment. Summary of the Invention

[0003] The purpose of this invention is to provide a machine learning-driven investment strategy configuration method to solve the problem in the prior art where the randomness and inconsistency of user behavior make it difficult for existing machine learning systems to converge preference parameters, causing strategy routing to oscillate between local repair and overall replanning.

[0004] To achieve the above objectives, this invention provides a machine learning-driven investment strategy configuration method, comprising the following steps: Step 100: Define the real-valued scalar variable representing the relative tendency between the local repair re-optimization mode and the overall replanning re-optimization mode in the re-optimization mode router as the preference parameter, and treat the preference parameter as a hidden state variable; map the confirmation operation and fine-tuning operation of the investment strategy scheme as a discrete-time scalar observation sequence with observation noise; model the time evolution of the preference parameter as a first-order random walk driven by zero-mean process noise; output the adaptive observation noise variance through a pre-trained lightweight feedforward neural network; finally output the state-space model parameter set containing the process noise variance, the adaptive observation noise variance, and the initial posterior estimate. Step 200: The prior prediction of the preference parameters, the calculation of the observation residuals, the calculation of the covariance of the observation residuals, the calculation of the Kalman gain, and the posterior update are completed in sequence, and the posterior mean estimate and posterior variance estimate of the preference parameters at the current time are output. Step 300: Configure the hysteresis band width as a monotonically non-decreasing function of the posterior variance estimate; combine the reoptimization mode currently activated by the reoptimization mode router and the directional polarity of the hysteresis band, and output the uplink handover threshold and downlink handover threshold at the current moment. Step 400: When the posterior mean estimate remains validly out of bounds within the preset confirmation window length, and the posterior variance estimate is always lower than the preset upper limit of variance within the confirmation window length, output a mode switching enable signal; otherwise, output a mode hold signal. Step 500: Upon receiving a mode switching permission signal, switch the activation mode of the re-optimization mode router from the current mode to the opposite mode and reverse the hysteresis polarity; upon receiving a mode hold signal, keep the activation mode and directional polarity unchanged; write the updated activation mode identifier and directional polarity back to the re-optimization mode router, and output the activation mode identifier as an investment strategy configuration instruction to control the external generation of the corresponding investment strategy scheme.

[0005] Furthermore, in step 100, the ratio of process noise variance to adaptive observation noise variance is limited to an interval. Inside, among which Greater than 0 and less than 1 Greater than 1 and less than the preset upper convergence bound; when the ratio obtained by adaptive estimation based on the observation residuals of the most recent preset number of observations exceeds the interval When the ratio is projected to the nearest endpoint of the interval, it is then substituted into the Kalman recursive estimate.

[0006] Furthermore, in step 100, the input of the lightweight feedforward neural network is the observation residual sequence of the most recent H time steps and the posterior variance sequence of the most recent H time steps. The network structure is a two-layer fully connected layer, and the final output is an adaptive observation noise variance. The training method of the lightweight feedforward neural network is: in the offline stage, it is pre-trained using historical user feedback data with the goal of maximizing the negative log-likelihood of the observation sequence.

[0007] Furthermore, step 200 also includes robustness processing for outlier observations: normalized observation residuals are calculated during each Kalman recursive estimation; when the absolute value of the normalized observation residuals exceeds a preset threshold of n times the standard deviation for k consecutive times, the posterior update is switched to a robust update mode based on Huber loss, where the robust update mode linearly weights the observation residuals exceeding the n-times-standard-deviation threshold; when the absolute value of the normalized observation residuals exceeds a preset threshold of n times the standard deviation for k consecutive times, the posterior update is switched to a robust update mode based on Huber loss, where the robust update mode linearly weights the observation residuals exceeding the n-times-standard-deviation threshold; when the absolute value of the normalized observation residuals exceeds a preset threshold of n times the standard deviation threshold for k consecutive times, the posterior update is switched to a robust update mode based on Huber loss, where ... The normalized observation residuals are restored to the preset value. When the error is below the standard deviation threshold, the posterior update mode is restored to the standard Kalman update; the formula for calculating the normalized observation residuals is: ,in Let be the normalized observation residual at time t. for Observe the residuals at all times. for The residual covariance is constantly observed, and k, k', n, and n' are all preset positive parameters.

[0008] Further, in step 200, the calculation process of the Kalman gain includes numerical condition number monitoring: before calculating the Kalman gain, the observation residual covariance is calculated first, and underflow detection is performed on the observation residual covariance; when the observation residual covariance is lower than the preset underflow threshold, the observation residual covariance is clamped upward to the preset underflow threshold before the reciprocal is calculated; when the Kalman gain calculated based on the result of the reciprocal exceeds the preset upper bound of the gain, the Kalman gain is clamped to the preset upper bound of the gain; the clamped Kalman gain is used for posterior update.

[0009] Furthermore, in step 300, the hysteresis band width is configured such that the monotonically non-decreasing function of the posterior variance estimation adopts a piecewise linear function, as follows: ; in, This represents the hysteresis band width function. and These represent the lower and upper bounds of the hysteresis band width, respectively. and These represent the first variance threshold and the second variance threshold, respectively. express Posterior variance estimation at time 1. , .

[0010] Furthermore, in step 300, the method for outputting the uplink handover threshold and downlink handover threshold at the current moment, combining the currently activated reoptimization mode of the reoptimization mode router and the directional polarity of the hysteresis band, is as follows: When the current activation mode is local repair and re-optimization mode, the out-of-bounds threshold for moving towards the overall replanning and re-optimization mode is taken as follows: ; When the current activation mode is global replanning and reoptimization mode, the out-of-bounds threshold for moving towards local repair and reoptimization mode is taken as follows: ; in Let be the downlink handover threshold at time t. Let be the uplink switching threshold at time t. Let t be the directional polarity of the hysteresis band. , for The hysteresis band width of the posterior variance estimate at time step [time]. The threshold difference.

[0011] Furthermore, when the directional polarity of the hysteresis band is reversed, the initial threshold difference of the re-optimization mode before the reversal is multiplied by a preset asymmetric amplification coefficient to obtain the amplified threshold difference. The asymmetric amplification coefficient is greater than one and not greater than the upper limit of the preset asymmetric amplification coefficient. Under the amplified state, the amplified threshold difference is regressed periodically to the threshold difference corresponding to the hysteresis band width calculated by the piecewise linear function according to the preset attenuation factor.

[0012] Further, in step 400, the preset confirmation window length is determined according to the following adaptive rule: using the posterior variance estimates of the most recent M time steps as samples, a fixed-order sliding quantile is calculated; when the sliding quantile is higher than a preset upper quantile threshold, the preset confirmation window length is increased by a preset growth factor; when the sliding quantile is lower than a preset lower quantile threshold, the preset confirmation window length is reduced by a preset contraction factor; and the preset confirmation window length is limited to the interval formed by the preset minimum length and the preset maximum length.

[0013] Further, in step 400, the preset upper limit of variance is determined according to the following adaptive rule: a fixed-order moving quantile is calculated using the posterior variance estimates of the most recent M time steps as samples, and the product of the moving quantile and the preset safety factor is used as the preset upper limit of variance, wherein the preset safety factor is greater than zero and less than one; and the preset upper limit of variance is restricted to the interval formed by the preset lower bound of variance and the preset upper bound of variance.

[0014] Therefore, the above-mentioned machine learning-driven investment strategy allocation method of the present invention has the following beneficial effects: 1. By modeling user feedback as noisy observations and using Kalman recursion estimation, uncertainty can be explicitly estimated using posterior variance, compared to existing techniques that directly update preference parameters using bare gradients. Single anomalous feedback no longer directly drives decision-making, fundamentally reducing ineffective oscillations near the decision boundary. 2. By driving the hysteresis band width with a posteriori variance, the switching threshold for periods with high feedback noise and high estimation uncertainty is automatically raised, avoiding aggressive mode switching during periods of drastic fluctuations in user behavior and ensuring the stability of policy routing. 3. By asymmetrically amplifying the directional polarity reversal and initial threshold difference after mode switching, the phenomenon of immediately switching back is eliminated, leaving sufficient time window for observing the true preference trend and eliminating repeated switching back and forth. 4. By confirming the dual adaptive criteria of window and upper limit of variance, the switching conditions are automatically adjusted according to the time-varying characteristics of the feedback environment, avoiding the failure of static threshold in long-term non-stationary operation, and taking into account both sensitivity and robustness. 5. A robust update model based on Huber loss is introduced, which performs linear weighting on the residuals of continuous outlier observations. At the same time, numerical condition number monitoring and underflow / upper bound clamping are implemented to effectively prevent single-point contamination or numerical divergence from propagating to the posterior estimate, thus ensuring the stable operation of the filter in a real human-computer interaction environment.

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is a flowchart of a machine learning-driven investment strategy configuration method according to the present invention. Detailed Implementation

[0017] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely illustrates selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0018] Please see Figure 1 A machine learning-driven investment strategy allocation method includes the following steps: Step 100: Construct the state-space model parameter set; The preference parameter is defined as a real-valued scalar variable representing the relative tendency between the local repair and re-optimization mode and the global reprogramming and re-optimization mode, denoted as . Its value range is reduced to an interval. ,in Approaching The characteristics tend towards a local repair and re-optimization mode. Approaching The representation tends towards a holistic replanning and optimization model; user feedback observations are denoted as... Its value also lies within the interval The value is obtained by mapping the user feedback type and the normalized value of the user feedback amplitude. This mapping is completed inside the re-optimization mode router and is not visible to the outside.

[0019] Preference parameters As a hidden state variable, each user feedback observation As a discrete-time observation with observational noise, the temporal evolution of the preference parameter is modeled as a first-order random walk, i.e., the process model is... ,in With a mean of zero and a variance of Process noise; observation model is ,in The mean is zero and the variance is The observation noise, and and They are independent of each other; the output includes process noise variance. Observation noise variance Estimation of initial and posterior mean of preference parameters Compared with the initial posterior variance estimate The state-space model parameter set. The meanings of each symbol are as follows: express The true value of the preference parameters at time 1. express User feedback observations at any given time express Process noise samples at time points, express Observational noise samples at any given time Represents the process noise variance. Represents the variance of observation noise. This represents the posterior mean estimate at the initial time. This represents the posterior variance estimate at the initial time. In this embodiment... , , Modeling preference parameters as random walks with process noise can explain their slow drift over long time scales, and modeling user feedback as scalar observations with observation noise can explicitly absorb random disturbances in human-computer interaction, thus enabling the use of an optimal linear estimation framework to explicitly quantify the uncertainty of the estimate in subsequent steps.

[0020] To make the posterior variance estimated by Kalman recursion more consistent with the true uncertainty and more accurately capture the non-stationary random characteristics of user feedback, the observation noise variance is dynamically adjusted through a machine learning model. Therefore, the observation noise variance is fixed. Replace with time-varying adaptive observation noise variance The adaptive observation noise variance is output in real time by a lightweight feedforward neural network. The input features of this lightweight feedforward neural network are the observation residual sequence and the posterior variance sequence of the most recent H time steps. The network structure is a two-layer fully connected layer with 16 hidden neurons and the activation function is ReLU. The final output is the adaptive observation noise variance. This lightweight feedforward neural network is pre-trained offline using historical user feedback data with the goal of maximizing the negative log-likelihood of the observation sequence.

[0021] To ensure that the Kalman gain remains within the stable convergence region, further constraints can be imposed. and ratio In the range Inside, among which Greater than 0 and less than 1 Greater than 1 and less than the preset upper convergence bound, in this embodiment Take 0.5, Set the value to 50. When the ratio obtained by adaptive estimation based on the observation residuals of the most recent preset number of observations exceeds the interval, project it to the nearest endpoint of the interval and then substitute it into the Kalman recursion to avoid two numerical ill-conditions: the ratio being too small, which leads to posterior mean tracking noise, and the ratio being too large, which leads to posterior mean failing to respond to true preference drift.

[0022] The process noise variance, adaptive observation noise variance, and initial posterior estimates (posterior mean estimate and posterior variance estimate at the initial time) are combined into a state-space model parameter set.

[0023] Step 200: For each arriving user feedback observation, perform Kalman recursive estimation based on the state-space model parameter set; The process sequentially completes prior prediction of preference parameters, calculation of observation residuals, calculation of observation residual covariance, calculation of Kalman gain, and posterior update to obtain the posterior mean estimate of preference parameters at the current time step. With posterior variance estimation The posterior mean estimate and the posterior variance estimate are then input into step 300.

[0024] The representative formula for the Kalman recursive estimation used is as follows: ; ; ; ; in, express Time based and prior mean estimates from previous observations, express Time based and the prior variance of previous observations, express Observe the residuals at all times. express Observe the residual covariance at all times. express Moment Kalman gain, express The posterior mean estimate at time t. express Posterior variance estimation at time step 1.

[0025] To prevent numerical divergence, numerical condition number monitoring is implemented during the Kalman gain calculation process: the condition number is calculated before calculating the Kalman gain. and perform underflow detection on it; when Below the preset underflow threshold At that time, clamp it upwards until Then find the reciprocal; when calculated according to the above formula... Exceeding the preset upper limit of gain At that time, Clamp to The clamped Kalman gain is used for subsequent posterior updates. In this embodiment... , .

[0026] To address sudden anomaly feedback, a robust treatment for anomaly observations is further introduced in step 200: defining normalized observation residuals. ;when continuous Second greater than The threshold value is 10 times the standard deviation (in this embodiment, it is taken as 10 times the standard deviation threshold). , When this step occurs, the subsequent update will be switched to a robust update mode based on Huber loss. The robust update mode... Exceeding The portion exceeding the standard deviation threshold is weighted linearly rather than quadratically, thereby suppressing the excessive pull of a single outlier on the posterior mean; when continuous Next (in this embodiment) ) Restore to Multiple standard deviation threshold (in this embodiment) When the following occurs, this step reverts the posterior update mode to the standard Kalman update. By introducing robustness measures for outlier observations, a single aggressive or contradictory feedback will not immediately cause a jump in the posterior mean, thus preventing noisy feedback from penetrating into the downstream mode-switching criterion.

[0027] Step 300: Based on the posterior variance estimation, adaptively calculate the hysteresis band width, and combine the reoptimization mode currently activated by the reoptimization mode router and the directional polarity of the hysteresis band to output the uplink handover threshold and downlink handover threshold at the current moment. The hysteresis band width is configured as a monotonically non-decreasing function of the posterior variance estimate, specifically using a piecewise linear function, as follows: ; The meanings of each symbol are as follows: This represents the hysteresis band width function. and These represent the lower and upper bounds of the hysteresis band width, respectively. and These represent the first variance threshold and the second variance threshold, respectively. , In this embodiment, , , , .

[0028] This step combines the currently active reoptimization mode of the reoptimization mode router with the directional polarity of the hysteresis band. Determine the uplink handover threshold at the current moment. Downlink handover threshold When the current activation mode is local repair and re-optimization mode, the out-of-bounds threshold for moving towards the overall replanning and re-optimization mode is taken as follows: When the current activation mode is the overall replanning and reoptimization mode, the out-of-bounds threshold for moving towards the local repair and reoptimization mode is taken as follows: ; The threshold difference ensures that a valid out-of-bounds condition is met only when the posterior mean of the preference parameter moves toward the currently inactive re-optimization mode beyond the corresponding out-of-bounds threshold.

[0029] Furthermore, after each effective switch occurs, this step checks the directional polarity. Invert immediately; multiply the initial threshold difference in the reoptimization mode before inversion by a preset asymmetric amplification factor. To obtain the amplification threshold difference, where , The upper limit of the preset asymmetric magnification factor (in this embodiment) is set. , The amplification threshold difference is maintained for a preset number of decay cycles during the amplification state. (In this embodiment) ), and then according to the preset attenuation factor (In this embodiment) By periodically regressing to the threshold difference corresponding to the hysteresis band width calculated by the piecewise linear function, the short-term back-cut oscillations after the switch are significantly suppressed. This method ensures that the threshold difference required for the back-cut direction after a switch is strictly greater than the threshold difference required before the switch, providing a window for observing true preferences and avoiding continuous back-and-forth switching caused by noise feedback.

[0030] Step 400: Perform mode switching determination based on posterior mean estimation and uplink switching threshold and downlink switching threshold; Only when Within the preset confirmation window length The boundary crossing remains valid continuously (i.e., the boundary crossing condition towards the currently inactive mode continues to be met), and within this confirmation window... Always below the preset upper limit of variance When the output mode is active, a mode switching enable signal is output; otherwise, a mode hold signal is output.

[0031] To adapt to the time-varying characteristics of the feedback environment, the confirmation window length was further adjusted. Implement adaptive: with the most recent Each time step (in this embodiment) The posterior variance estimate of the sample is used to calculate the moving quantile of a fixed order. (In this embodiment) );when Higher than the preset upper quantile threshold At that time, According to the preset growth factor (In this embodiment) ) Magnify; when Below the preset lower quantile threshold At that time, According to the preset shrinkage factor (In this embodiment) Shrink; and Limited to the range Inside (in this embodiment) , At the same time, the upper limit of variance. Implement adaptive: with the most recent The posterior variance estimate at each time step is used to calculate the moving quantile of the same order for the sample, and the moving quantile is then compared with a preset safety factor. (In this embodiment) The product of ) is used as ; and will Limited to the range Internally. Using this method, when the system enters a period of high noise, Automatic elevation will simultaneously lengthen the confirmation window and lower the upper limit of variance, making the switching conditions more stringent; when the system enters a stable period, the confirmation window length will tighten and the upper limit of variance will be widened, and the system will restore a higher switching response sensitivity.

[0032] Step 500: Update the re-optimization mode router according to the signal output by the mode switching determination, and output the activation mode identifier as an investment strategy configuration instruction to control the external generation of the corresponding investment strategy scheme.

[0033] Upon receiving a mode switching permission signal, this step switches the activation mode of the re-optimization mode router from the current mode to the opposite mode and writes the activation mode identifier to the re-optimization mode router; simultaneously, it adjusts the directional polarity according to step 300. Invert and magnify by the asymmetric factor. The initial threshold difference in the opposite direction is amplified so that the threshold difference required for immediate re-switching after the handover is strictly greater than the threshold difference required before the handover. Upon receiving a mode hold signal, this step maintains the active mode and directional polarity unchanged. The updated active mode identifier and directional polarity are written back to the re-optimized mode router for use in steps 300 and 400 at the next moment. After reading the updated active mode identifier, the external investment strategy generation module generates the corresponding scheme according to its internal rules and returns the scheme identifier.

[0034] In this embodiment, data interaction with the external investment strategy generation module is strictly limited to the following three types of standardized interface data: Current scheme identifier, which is a unique number assigned internally by the external investment strategy generation module to the current output scheme; User feedback type, which is an enumerated value including at least two values: confirmation and fine-tuning; User feedback amplitude normalized value, mapped to an interval. The real number.

[0035] It should be noted that the Kalman recursive estimation in step 200 can be replaced by any one of extended Kalman filtering, unscented Kalman filtering, or particle filtering: when using extended Kalman filtering or unscented Kalman filtering, the state transition and observation mapping can be generalized to a nonlinear form, and the posterior variance estimation is still output by the filter; when using particle filtering, the posterior variance estimation is replaced by the sample variance of the particle set or an uncertainty measure equivalent to the sample variance.

[0036] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A machine learning-driven investment strategy allocation method, characterized in that, Includes the following steps: Step 100: Define the real scalar variable representing the relative tendency between local repair re-optimization mode and global replanning re-optimization mode in the re-optimization mode router as preference parameter, and use the preference parameter as a hidden state variable; map the confirmation operation and fine-tuning operation of the investment strategy scheme as a discrete-time scalar observation sequence with observation noise; The temporal evolution of the preference parameters is modeled as a first-order random walk driven by zero-mean process noise; an adaptive observation noise variance is output through a pre-trained lightweight feedforward neural network; the final output includes the process noise variance, the adaptive observation noise variance, and the state-space model parameter set of the initial posterior estimate. Step 200: The prior prediction of the preference parameters, the calculation of the observation residuals, the calculation of the covariance of the observation residuals, the calculation of the Kalman gain, and the posterior update are completed in sequence, and the posterior mean estimate and posterior variance estimate of the preference parameters at the current time are output. Step 300: Configure the hysteresis band width as a monotonically non-decreasing function of the posterior variance estimate; combine the reoptimization mode currently activated by the reoptimization mode router and the directional polarity of the hysteresis band, and output the uplink handover threshold and downlink handover threshold at the current moment. Step 400: When the posterior mean estimate remains validly out of bounds within the preset confirmation window length, and the posterior variance estimate is always lower than the preset upper limit of variance within the confirmation window length, output a mode switching enable signal; otherwise, output a mode hold signal. Step 500: When a mode switching permission signal is received, the active mode of the reoptimization mode router is switched from the current mode to the opposite mode, and the hysteresis polarity is reversed; when a mode hold signal is received, the active mode and directional polarity remain unchanged. The updated activation mode identifier and directional polarity are written back to the re-optimized mode router, and the activation mode identifier is output as an investment strategy configuration instruction to control the external generation of the corresponding investment strategy scheme.

2. The machine learning-driven investment strategy allocation method according to claim 1, characterized in that, In step 100, the ratio of process noise variance to adaptive observation noise variance is limited to an interval. Inside, among which Greater than 0 and less than 1 Greater than 1 and less than the preset upper convergence bound; When the ratio obtained by adaptive estimation based on the observation residuals of the most recent preset number of observations exceeds the interval When the ratio is projected to the nearest endpoint of the interval, it is then substituted into the Kalman recursive estimate.

3. The machine learning-driven investment strategy allocation method according to claim 2, characterized in that, In step 100, the input of the lightweight feedforward neural network is the observation residual sequence of the most recent H time steps and the posterior variance sequence of the most recent H time steps. The network structure is a two-layer fully connected network, and the final output is the adaptive observation noise variance. The training method of the lightweight feedforward neural network is: in the offline stage, historical user feedback data is used to pre-train the network with the goal of maximizing the negative log-likelihood of the observation sequence.

4. The machine learning-driven investment strategy allocation method according to claim 3, characterized in that, Step 200 also includes robustness processing for outlier observations: Normalized observation residuals are calculated during each Kalman recursive estimation; when the absolute value of the normalized observation residuals exceeds a preset threshold of n times the standard deviation for k consecutive times, the posterior update is switched to a robust update mode based on Huber loss. The robust update mode linearly weights the observation residuals exceeding the n-times-standard-deviation threshold; when the absolute value of the normalized observation residuals exceeds a preset threshold of n times the standard deviation threshold for k consecutive times... The normalized observation residuals are restored to the preset value. When the error is below the threshold of one standard deviation, the posterior update mode will be restored to the standard Kalman update. The formula for calculating the normalized observation residuals is: ,in Let be the normalized observation residual at time t. for Observe the residuals at all times. for The residual covariance is constantly observed, and k, k', n, and n' are all preset positive parameters.

5. The machine learning-driven investment strategy allocation method according to claim 4, characterized in that, In step 200, the calculation process of Kalman gain includes numerical condition number monitoring: before calculating Kalman gain, the observation residual covariance is calculated, and underflow detection is performed on the observation residual covariance; When the observation residual covariance is lower than the preset underflow threshold, the observation residual covariance is clamped upward to the preset underflow threshold before the reciprocal is calculated; when the Kalman gain calculated based on the result of the reciprocal exceeds the preset upper bound, the Kalman gain is clamped to the preset upper bound; the clamped Kalman gain is used for posterior update.

6. The machine learning-driven investment strategy allocation method according to claim 5, characterized in that, In step 300, the hysteresis band width is configured such that the monotonically non-decreasing function for the posterior variance estimation is a piecewise linear function, as follows: ； in, This represents the hysteresis band width function. and These represent the lower and upper bounds of the hysteresis band width, respectively. and These represent the first variance threshold and the second variance threshold, respectively. express Posterior variance estimation at time 1, , .

7. The machine learning-driven investment strategy allocation method according to claim 6, characterized in that, In step 300, the method for outputting the uplink handover threshold and downlink handover threshold at the current moment, combining the currently activated reoptimization mode of the reoptimization mode router and the hysteresis band direction polarity, is as follows: When the current activation mode is local repair and re-optimization mode, the out-of-bounds threshold for moving towards the overall replanning and re-optimization mode is taken as follows: ; When the current activation mode is global replanning and reoptimization mode, the out-of-bounds threshold for moving towards local repair and reoptimization mode is taken as follows: ; in Let be the downlink handover threshold at time t. Let be the uplink switching threshold at time t. Let t be the directional polarity of the hysteresis band. , for The hysteresis band width of the posterior variance estimate at time step [time]. The threshold difference.

8. The machine learning-driven investment strategy allocation method according to claim 7, characterized in that, When the directional polarity of the hysteresis band is reversed, the initial threshold difference of the re-optimization mode before the reversal is multiplied by the preset asymmetric amplification coefficient to obtain the amplified threshold difference. The asymmetric amplification coefficient is greater than one and not greater than the upper limit of the preset asymmetric amplification coefficient. Under the amplified state, the amplified threshold difference is regressed periodically to the threshold difference corresponding to the hysteresis band width calculated by the piecewise linear function according to the preset attenuation factor.

9. The machine learning-driven investment strategy allocation method according to claim 8, characterized in that, In step 400, the preset confirmation window length is determined according to the following adaptive rule: using the posterior variance estimates of the most recent M time steps as samples, a fixed-order sliding quantile is calculated; when the sliding quantile is higher than the preset upper quantile threshold, the preset confirmation window length is increased by a preset growth factor; when the sliding quantile is lower than the preset lower quantile threshold, the preset confirmation window length is reduced by a preset contraction factor. The default confirmation window length is limited to the range between the default minimum length and the default maximum length.

10. The machine learning-driven investment strategy allocation method according to claim 9, characterized in that, In step 400, the preset upper limit of variance is determined according to the following adaptive rule: the posterior variance estimates of the most recent M time steps are used as samples to calculate the moving quantile of a fixed order, and the product of the moving quantile and the preset safety factor is used as the preset upper limit of variance, wherein the preset safety factor is greater than zero and less than one; and the preset upper limit of variance is restricted to the interval formed by the preset lower limit of variance and the preset upper limit of variance.

Images