Bond recommendation method and system based on machine learning and deep neural network

Abstract

The embodiment of the invention provides a bond recommendation method and system based on machine learning and a deep neural network. The method comprises the following steps: acquiring multi-source feature data related to bond transaction; and performing dynamic feature screening processing on the multi-source feature data, obtaining a key influence factor set by fusing correlation analysis and historical trend weighting, and generating a correlation coefficient prediction result based on correlation analysis. Inputting the key influence factor set into a multi-model collaborative prediction framework, and respectively generating a first prediction result, a second prediction result, a third prediction result and a fourth prediction result through a state space model, a long-short-term memory network, a gradient lifting model and a moving average convergence divergence model, and inputting the correlation coefficient prediction result into an AdaBoost classifier for training to obtain a bond price rise and fall prediction model, and outputting a bond recommendation result based on the bond price rise and fall prediction model, thereby improving the accuracy and efficiency of bond recommendation.

Description

Technical Field

[0001] The present application relates to the field of financial technology, and more specifically, to a bond recommendation method and system based on machine learning and deep neural networks. Background Art

[0002] In the financial sector, bond prices are a crucial investment tool, and predicting their price trends plays a crucial role in guiding investment decisions. Currently, financial institutions often use traditional machine learning methods to predict time series trends, such as bond prices, to aid trading decisions. However, traditional machine learning models have relatively simple network structures, and their data-fitting capabilities are limited when faced with the massive, high-dimensional data generated by financial markets. They struggle to capture the complex nonlinear relationships and underlying patterns within the data, resulting in poor prediction results.

[0003] Therefore, how to improve the accuracy and efficiency of bond recommendations is an urgent problem that needs to be solved. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a bond recommendation method and system based on machine learning and deep neural networks to improve the accuracy and efficiency of bond recommendation suggestions.

[0005] In a first aspect, the present application provides a bond recommendation method based on machine learning and deep neural networks, comprising: Acquiring multi-source feature data related to bond transactions, wherein the multi-source feature data includes bond price data and related market index data; Performing dynamic feature screening processing on the multi-source feature data, obtaining a set of key influencing factors by fusing correlation analysis with historical trend weighting, and generating a correlation coefficient prediction result based on the correlation analysis; The set of key influencing factors is input into a multi-model collaborative prediction framework, and a first prediction result, a second prediction result, a third prediction result, and a fourth prediction result are generated respectively through a state space model, a long short-term memory network, a gradient boosting model, and a moving average convergence divergence model. The multi-model collaborative prediction framework is retrained after the market close each day based on the multi-source feature data updated on that day; Inputting the correlation coefficient prediction result, the first prediction result, the second prediction result, the third prediction result, and the fourth prediction result into an AdaBoost classifier for training, learning the weighted influence of each prediction result through a cross-entropy loss function, and obtaining a bond price increase and decrease prediction model, wherein the bond price increase and decrease prediction model uses the difference between the closing price and the opening price as the increase and decrease standard; Output bond recommendation results based on the bond price increase and decrease prediction model.

[0006] Optionally, performing the dynamic feature screening process on the multi-source feature data, obtaining the key influencing factor set by fusing the correlation analysis with the historical trend weighting, and generating the correlation coefficient prediction result based on the correlation analysis includes: Calculating the correlation coefficients between the bond price data and the associated market index data to establish a characteristic correlation matrix; Performing the historical trend weighting processing on the feature correlation matrix, adjusting the weight proportions of correlations in different periods by a time decay function, and obtaining a weighted feature correlation matrix; Performing feature importance sorting based on the weighted feature correlation matrix, and extracting leading features from the result of the feature importance sorting as candidate influencing factors; Conduct dynamic adaptability verification on the candidate influencing factors and evaluate the stability indicators of each feature in the candidate influencing factors in different market cycles through a rolling window verification method; Filtering the key influencing factor set according to the product result of the stability index and the correlation weight, wherein the key influencing factor set includes interest rate characteristics, index characteristics, and trading volume characteristics; Based on the signs and numerical values ​​of the correlation coefficients between each of the features in the feature correlation matrix and the bond price data, a correlation coefficient prediction result representing the direction and intensity of the rise and fall is generated.

[0007] Optionally, inputting the key influencing factor set into the multi-model collaborative prediction framework, and generating the first prediction result, the second prediction result, the third prediction result, and the fourth prediction result respectively through the state space model, the long short-term memory network, the gradient boosting model, and the moving average convergence divergence model, includes: Reconstructing the key influencing factor set into a time series to form an input feature sequence with a fixed time window length; Inputting the input feature sequence into the state space model, capturing the periodic fluctuation characteristics of the feature sequence through a structured state space conversion mechanism, and outputting the first prediction result; Inputting the input feature sequence into the long short-term memory network, processing the long-term dependency of the feature sequence through a gated recurrent unit, and outputting the second prediction result; Inputting the input feature sequence into the gradient boosting model, and generating the third prediction result by superimposing the prediction results of multiple decision trees; The input feature sequence is input into the moving average convergence divergence model, and the fourth prediction result is generated by calculating the deviation value and the deviation average of the price fluctuation difference sequence.

[0008] Optionally, inputting the input feature sequence into the state space model, capturing the periodic fluctuation characteristics of the feature sequence through the structured state space conversion mechanism, and outputting the first prediction result includes: Performing dimension mapping on the input feature sequence through a fully connected network to generate an initial state vector with a uniform hidden dimension; Constructing a state transfer equation based on a linear combination of the current input feature and the hidden state at the previous moment, and dynamically evolving the initial state vector through the state transfer equation to generate a new hidden state; Discretize the new hidden state by matrix exponential operation and integral approximation calculation to generate a discretized state transfer matrix and an input weight matrix; Performing a multi-step update on the new hidden state based on the discretized state transfer matrix and the input weight matrix, dynamically adjusting the retention weights of different feature components in the hidden state, and generating a multi-step updated hidden state; The hidden state after multiple steps of updating is input into the output mapping layer, and the original prediction value of the state space model is generated by a combination of linear transformation and nonlinear activation function; The first prediction result is determined according to the original prediction value.

[0009] Optionally, inputting the input feature sequence into the long short-term memory network, processing the long-term dependency of the feature sequence through the gated recurrent unit, and outputting the second prediction result includes: Convert the input feature sequence into a tensor of preset time length and feature dimension, where the preset time length corresponds to the time span of continuous feature data, and the feature dimension corresponds to the number of features at each time node; Inputting the tensors of the preset time length and feature dimension into the respective network layers of the long short-term memory network in sequence, and performing feature extraction and dimension conversion processing on the tensors of the preset time length and feature dimension through the respective network layers; Inputting the feature data processed by each network layer into a linear function activation layer, and performing a linear transformation on the feature data through the linear function activation layer to generate a regression result, wherein the numerical value of the regression result corresponds to the long-term dependency analysis result of the input feature sequence; The second prediction result is generated according to the regression result.

[0010] Optionally, inputting the input feature sequence into the gradient boosting model and generating the third prediction result by superimposing prediction results of multiple decision trees includes: Converting the input feature sequence into a sample feature set in a preset sample format, wherein the preset sample format corresponds to the feature input dimension required by the gradient boosting model, and the sample feature set includes numerical information of each feature in the input feature sequence and correlation relationships between features; Inputting the sample feature set in the preset sample format into the gradient boosting model in sequence, and performing feature screening, decision tree construction, and prediction result superposition processing on the sample feature set through the gradient boosting model; The processed superposition prediction results are input into the output layer of the gradient boosting model, and the superposition prediction results are comprehensively calculated through linear integration processing to generate an overall prediction result, wherein the numerical value of the overall prediction result is positively correlated with the comprehensive influence strength of each feature in the sample feature set; The third prediction result is generated according to the total prediction result.

[0011] Optionally, inputting the input feature sequence into the moving average convergence divergence model and generating the fourth prediction result by calculating the deviation value and the mean deviation value of the price fluctuation difference sequence includes: extracting bond price data from the input feature sequence; Calculating a short-term weighted moving average and a long-term weighted moving average based on the bond price data, wherein a calculation period of the short-term weighted moving average is shorter than a calculation period of the long-term weighted moving average; Difference between the short-term weighted moving average and the long-term weighted moving average to generate a deviation value sequence; Calculating a moving average value for the deviation value sequence to generate a deviation mean value sequence; Subtracting the deviation value sequence from the deviation mean value sequence to generate a moving average convergence divergence indicator sequence; Performing trend analysis on the moving average convergence divergence indicator sequence, generating an upward trend signal when the moving average convergence divergence indicator sequence turns from negative to positive, and generating a downward trend signal when the moving average convergence divergence indicator sequence turns from positive to negative; The fourth prediction result is determined based on the upward trend signal or the downward trend signal.

[0012] Optionally, the step of inputting the correlation coefficient prediction result, the first prediction result, the second prediction result, the third prediction result, and the fourth prediction result into the AdaBoost classifier for training, and learning the weighted influence of each prediction result through the cross entropy loss function to obtain the bond price rise and fall prediction model includes: Perform feature splicing on the correlation coefficient prediction result, the first prediction result, the second prediction result, the third prediction result, and the fourth prediction result to form an input feature vector of the AdaBoost classifier; Initializing multiple weak classifiers, each of the weak classifiers corresponds to a weight parameter of a prediction result; The weight parameters of each weak classifier are updated by iterative training, and the voting weight is dynamically adjusted according to the classification error. The smaller the classification error, the larger the corresponding weight parameter of the prediction result. Calculating the loss value of the integrated prediction result and the true rise and fall label based on the difference between the closing price and the opening price through the cross entropy loss function, and back-propagating to optimize the weight parameters of each weak classifier; The weight updating and loss optimization steps are repeated until the loss value converges to a preset threshold, thereby generating the bond price rise and fall prediction model.

[0013] Optionally, outputting the final bond recommendation result based on the bond price fluctuation prediction model includes: Inputting the multi-source feature data updated in real time into the bond price increase and decrease prediction model to obtain an increase and decrease prediction result based on the difference between the closing price and the opening price on the next trading day; Calculate the accuracy index of the rise and fall prediction results in three consecutive increasing time windows, where the accuracy index is the ratio of the number of correctly predicted days to the total number of days in each time window; Performing a weighted fusion of the accuracy index and the confidence of the prediction result to generate a comprehensive recommendation score; sorting the bonds according to the comprehensive recommendation scores and selecting the top-ranked bonds to form a recommendation list; The recommendation list is associated with the corresponding set of key influencing factors and the prediction results of each model and stored to generate the bond recommendation result.

[0014] In a second aspect, the present application provides a bond recommendation system based on machine learning and deep neural networks, wherein the bond recommendation system based on machine learning and deep neural networks includes a machine-readable storage medium and a processor, wherein the machine-readable storage medium stores machine-executable instructions. When the processor executes the machine-executable instructions, the bond recommendation system based on machine learning and deep neural networks implements the aforementioned bond recommendation method based on machine learning and deep neural networks.

[0015] The bond recommendation method and system based on machine learning and deep neural networks provided in this application first obtain multi-source feature data of bond transactions, then dynamically screen these data to obtain key influencing factors and correlation coefficient prediction results, then input the key influencing factors into a multi-model collaborative prediction framework to generate multiple prediction results, and then input various prediction results into the AdaBoost classifier to train a bond price rise and fall prediction model, and finally output the bond recommendation results based on the model. The comprehensive acquisition of multi-source data ensures the richness of information, dynamic feature screening can accurately focus on key influencing factors, multi-model collaborative prediction combined with daily retraining can improve the timeliness and diversity of predictions, and the AdaBoost classifier can optimize the comprehensive prediction effect by learning the weights of each prediction result, thereby improving the accuracy and efficiency of bond recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0017] Figure 1 A flowchart of a bond recommendation method based on machine learning and deep neural networks provided in an embodiment of the present application; Figure 2 A structural diagram of a bond recommendation system based on machine learning and deep neural networks provided in an embodiment of the present application.

[0018] The above drawings illustrate specific embodiments of the present application, which will be described in more detail below. These drawings and the textual description are not intended to limit the scope of the present application in any way, but rather to illustrate the concepts of the present application to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0019] To help those skilled in the art better understand the present invention, the following will provide a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. It is apparent that the described embodiments are only a portion of the embodiments of the present invention, not all of them. All other embodiments derived by those skilled in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0020] The terms "first," "second," and so on, in the specification and claims of the present invention and the accompanying drawings are used to distinguish between different items, not to describe a specific order. Furthermore, the terms "including," "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, apparatus, product, or end comprising a series of steps or elements is not limited to the listed steps or elements but may optionally include steps or elements not listed therein, or may optionally include other steps or elements inherent to such process, method, product, or end.

[0021] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute a separate or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0022] Figure 1 This is a flow chart of a bond recommendation method based on machine learning and deep neural networks provided in an embodiment of the present application. It should be understood that in other embodiments, the order of some steps in the bond recommendation method based on machine learning and deep neural networks of this embodiment can be shared with each other according to actual needs, or some steps can be omitted or maintained. Figure 1 As shown, the method may include the following steps: Step S110: Acquire multi-source feature data related to bond transactions, wherein the multi-source feature data includes bond price data and associated market index data.

[0023] In this embodiment, acquiring multi-source feature data related to bond trading is the foundation of the entire bond recommendation method. The bond price data and associated market index data contained in this multi-source feature data can reflect the bond market from different perspectives. Specifically, the required data is obtained after the market closes each day. For example, this data may include the daily opening price, closing price, high price, low price, trading volume, and the daily closing rates for 1-year Treasury bonds, 5-year Treasury bonds, 10-year Treasury bonds, Chicago Mercantile Exchange 10-year Treasury bonds, ChiNext Index, CSI 300 Index, CSI 500 Index, Nasdaq Index, RMB / USD exchange rate, interbank overnight interest rate, and interbank seven-day interest rate. This data records the price fluctuation range and trading activity of bonds over the past trading days. For example, the change from opening price to closing price can indicate whether the bond price rose or fell that day; the high and low prices can reflect the price fluctuation range; and trading volume data reflects market interest in the bond. Regarding interest rate data, Treasury bond interest rates of different maturities have different impacts on bond prices. Short-term Treasury bond interest rates likely reflect current market liquidity and short-term funding costs, while long-term Treasury bond interest rates are more closely tied to long-term macroeconomic expectations. Stock index data encompasses the comprehensive performance of different market sectors, such as large-cap and small-cap indices. The rise and fall of these indices influences investor capital flows, which in turn impacts the bond market. Exchange rate data reflects the exchange rate between different currencies. For the bond market, which involves cross-border investment, exchange rate fluctuations can affect the actual returns of bonds.

[0024] Next, we need to time-align historical trading data with associated market index data. Since these data may come from different sources and have different timestamps, we need to ensure they have consistent timestamps. For example, bond trading data may be recorded on a daily basis, while some associated market index data may be recorded on an hourly or even minutely basis. In this case, they need to be aligned to the same time granularity for subsequent analysis.

[0025] Afterwards, a data integrity check is performed. During the actual data collection process, data may be missing due to various reasons, such as data source failures and data transmission errors. To ensure data integrity, missing data is handled using the previous value filling method. That is, if data at a certain moment is missing, it is filled with data from the previous moment. For example, if the interest rate data for a certain maturity of a treasury bond on a certain day is missing, the interest rate data for that maturity of the treasury bond from the previous day is used to fill it in.

[0026] Subsequently, a preset wavelet basis function is used to perform multi-scale decomposition on the padded data. Wavelet basis functions have excellent time-frequency localization characteristics and can decompose data into components of different frequencies. Through this decomposition, the noise components and valid information in the data can be separated. Then, soft threshold processing is used to filter out the noise components, retaining only the valid information in the high-frequency fluctuation components. High-frequency fluctuation components often contain sudden market changes and short-term trends, which are of great significance for bond price prediction. For example, a dB10 wavelet basis can be selected to perform a three-layer wavelet decomposition on the data of each dimension, and a soft threshold function can be used to remove noise in each data dimension. Traditional filtering methods will over-suppress high-frequency signals, which may contain key information about short-term market fluctuations and important events. Over-suppression of high-frequency signals will cause the denoised sequence to lose a large amount of valuable information, making it impossible for the model to accurately capture subtle changes in the market, thereby affecting the accuracy of the prediction. The multi-scale wavelet decomposition technology of this embodiment can effectively solve this problem.

[0027] Finally, the denoised data is normalized. Because the data range and magnitude of different features can vary significantly—for example, bond price data can range from tens to hundreds of yuan, while trading volume data can be in the tens or even hundreds of thousands—this discrepancy can affect subsequent model training. Normalization ensures that each feature dimension conforms to a standard normal distribution, eliminating the influence of dimensionality and enabling the model to treat each feature more fairly. Furthermore, after the daily market close, daily trading data and related market index data are collected to update multi-source feature data to ensure timeliness and accuracy, providing the latest information for subsequent analysis and forecasting.

[0028] Step S120: performing dynamic feature screening processing on the multi-source feature data, obtaining a set of key influencing factors by fusing correlation analysis with historical trend weighting, and generating a correlation coefficient prediction result based on the correlation analysis.

[0029] After acquiring multi-source feature data, dynamic feature screening is required to identify key factors that significantly impact bond prices. Traditional feature screening methods for bond price prediction often employ simple correlation analysis, simply calculating the correlation coefficient between bond prices and various features (such as interest rates, indices, and trading volume). This approach lacks consideration of historical trends and constitutes a static feature screening model. It struggles to adapt to dynamic market changes and cannot promptly adjust feature selection based on new market conditions. Consequently, the selected features may not be the factors that currently have a key impact on bond prices, significantly reducing the accuracy of risk identification and the reliability of predictions. This embodiment, however, employs a dynamic feature screening mechanism. First, the correlation coefficients between bond price data and related market index data are calculated to establish a feature correlation matrix. The correlation coefficient measures the strength of the linear relationship between two variables, and its value ranges from -1 to 1. When the correlation coefficient is positive and close to 1, it means that there is a strong positive correlation between the two variables, that is, an increase in one variable will lead to an increase in the other variable; when the correlation coefficient is negative and close to -1, it means that there is a strong negative correlation between the two variables, that is, an increase in one variable will lead to a decrease in the other variable; when the correlation coefficient is close to 0, it means that there is almost no linear relationship between the two variables.

[0030] When calculating correlation coefficients, we can perform pairwise calculations for bond price data and each associated market index. For example, we can calculate the correlation coefficient between the closing price of a bond and the one-year Treasury bond rate, or the correlation coefficient between the closing price of a bond and the CSI 300 Index. Arranging these correlation coefficients into a matrix creates a characteristic correlation matrix. This matrix can intuitively demonstrate the strength of the correlation between bond prices and each associated market index.

[0031] Next, the feature correlation matrix can be weighted based on historical trends. Market conditions are constantly changing, and recent data often better reflects current market trends. Therefore, a time decay function is needed to adjust the weighting of correlations across different time periods. This function assigns different weights to correlations based on their timeliness, with more recent data receiving a greater weight. For example, an exponential decay function can be used, where the weight decreases exponentially over time. This results in a weighted feature correlation matrix that better reflects the impact of recent market factors on bond prices.

[0032] Step S121: Calculate the correlation coefficient between the bond price data and the associated market index data, and establish a feature correlation matrix.

[0033] In this step, correlation coefficients can be calculated using statistical correlation analysis. For each set of bond price data and associated market index data, their means are first calculated. For example, for the bond closing price data series {P1, P2, …, Pn} and the one-year Treasury bond interest rate data series {R1, R2, …, Rn}, their means μP and μR are calculated, respectively.

[0034] Then, the difference between each data point and the mean is calculated, namely (Pi-μP) and (Ri-μR).

[0035] Next, we calculate the sum of the products of these differences and divide it by the number of data points n and the product of the standard deviations of the bond closing price data and the 1-year Treasury bond rate data to get the correlation coefficient between them.

[0036] After calculating the correlation coefficients between all bond price data and the associated market indices, they are arranged in a matrix format according to a specific order to create a characteristic correlation matrix. Each row of this matrix represents a variable related to bond prices, and each column represents a variable related to the associated market index. The elements in the matrix represent the correlation coefficients between them.

[0037] Step S122: performing the historical trend weighting processing on the feature correlation matrix, adjusting the weight proportions of correlations in different periods by a time decay function, and obtaining a weighted feature correlation matrix.

[0038] The time decay function plays a key role in weighting historical trends. Assume the time decay function is f(t), where t represents the distance from the current time. For each element in the feature correlation matrix, multiply it by the corresponding time decay function value based on its corresponding time information. For example, if a correlation data item is recorded at a time t1 from the current time, the correlation data item is multiplied by f(t1).

[0039] This weighting operation is performed sequentially on correlation data at different times. After weighting, each correlation data point is assigned a different weight based on its timeliness, resulting in a weighted characteristic correlation matrix. This matrix more accurately reflects the impact of recent market factors on bond prices, making subsequent analysis more consistent with current market realities.

[0040] Step S123: performing feature importance sorting based on the weighted feature association matrix, and extracting the top features in the result of the feature importance sorting as candidate influencing factors.

[0041] Based on the weighted feature correlation matrix, the importance of each correlated market index feature can be ranked. This ranking is based on the absolute value of the correlation. A larger absolute value indicates a stronger correlation between the feature and the bond price, and a more significant impact on the bond price. For example, if the absolute value of the correlation coefficient between the Treasury bond interest rate of a certain maturity and the bond price is large in the weighted feature correlation matrix, then the Treasury bond interest rate feature of that maturity will be ranked higher in importance.

[0042] After the ranking is complete, the top features in the ranking are extracted as candidate influencing factors. These candidate influencing factors are preliminarily considered to have a significant impact on bond prices. The specific number of features extracted will be determined based on the actual situation and the requirements of the subsequent model. For example, if the subsequent model has a certain limit on the number of input features, a certain number of top-ranked features will be extracted as candidate influencing factors.

[0043] Step S124: Dynamically verify the adaptability of the candidate influencing factors, and evaluate the stability index of each feature in the candidate influencing factors in different market cycles through a rolling window verification method.

[0044] In this embodiment, dynamic adaptability verification of candidate influencing factors is performed to ensure that these factors consistently influence bond prices across different market cycles. Specifically, a rolling window verification method can be employed, dividing historical data into specific time windows. For example, the entire historical data can be divided into multiple time windows of equal length, with each window covering a certain number of trading days.

[0045] For each candidate influencing factor, its stability index is calculated within each time window. Stability indicators can take various forms, such as variance and standard deviation. For example, variance reflects the degree of data dispersion. A smaller variance indicates less fluctuation within the time window, and thus greater stability. By calculating the variance of each candidate influencing factor within each time window, a series of variance values ​​can be obtained.

[0046] Next, observe how these variances change across time windows. If the variance of a candidate influencing factor is relatively small and varies little across time windows, it indicates that the feature is relatively stable across different market cycles. Conversely, if the variance fluctuates significantly, it indicates that the feature is less stable. This way, the stability of each feature in the candidate influencing factor can be assessed across different market cycles.

[0047] Step S125: Filtering out the key influencing factor set according to the product result of the stability index and the correlation weight, wherein the key influencing factor set includes interest rate characteristics, index characteristics, and transaction volume characteristics.

[0048] After obtaining the stability index and correlation weight for each candidate impact factor, we multiply them together to create a composite score. This composite score takes into account both the stability of the feature and its correlation with bond prices. A composite score is calculated for each candidate impact factor.

[0049] Then, candidate influencing factors are screened based on the comprehensive scores. A threshold is set, and candidate influencing factors with comprehensive scores above the threshold are screened out to form a set of key influencing factors. Generally speaking, the set of key influencing factors will include interest rate characteristics, index characteristics, and trading volume characteristics. Interest rate characteristics, such as the interest rates of government bonds of different maturities, reflect the market's capital costs and macroeconomic expectations; index characteristics, such as various stock market indices, reflect the overall performance of the stock market and affect investors' capital allocation; trading volume characteristics reflect the trading activity of the bond market. Large trading volume often means that the market is paying close attention to the bond.

[0050] Step S126: Based on the signs and numerical values ​​of the correlation coefficients between each of the features in the feature correlation matrix and the bond price data, a correlation coefficient prediction result representing the direction and intensity of the rise and fall is generated.

[0051] Based on the sign and magnitude of the correlation coefficients between each feature in the feature correlation matrix and bond price data, a correlation coefficient prediction result can be generated. The sign of the correlation coefficient indicates the direction of change between the feature and the bond price. A positive correlation coefficient indicates a positive correlation between the feature and the bond price; that is, when the feature increases, the bond price is likely to increase. A negative correlation coefficient indicates a negative correlation between the feature and the bond price; that is, when the feature increases, the bond price is likely to decrease.

[0052] The magnitude of the correlation coefficient indicates the strength of the correlation. A larger value indicates a stronger correlation, and the more significant the feature's impact on bond prices. Based on the sign and magnitude of these correlation coefficients, each associated market index feature is analyzed. For example, if a stock index feature is positively correlated with bond prices and has a large correlation coefficient, then if the stock index is predicted to rise, bond prices are likely to rise as well, and the magnitude of the rise can be roughly estimated based on the correlation coefficient. By combining the analysis results of all features, a correlation coefficient prediction is generated, indicating the direction and strength of the price increase or decrease.

[0053] Step S130: Input the key influencing factor set into the multi-model collaborative prediction framework, and generate the first prediction result, the second prediction result, the third prediction result, and the fourth prediction result respectively through the state space model, the long short-term memory network, the gradient boosting model, and the moving average convergence divergence model.

[0054] The multi-model collaborative prediction framework is retrained after the daily close based on the multi-source feature data updated on that day.

[0055] In this embodiment, a set of key influencing factors is input into a multi-model collaborative prediction framework, and multiple different models are used to predict bond prices from different angles, which can improve the accuracy and reliability of the prediction. Most previous prediction methods rely on a single model or a single indicator for prediction, and do not fully integrate the advantages of multiple prediction methods. These different prediction results each have unique advantages, but they have not been organically combined to train a special rise and fall prediction classifier, resulting in the inability to fully utilize multiple information sources to improve the accuracy of the prediction. The multi-model collaborative prediction framework of this embodiment includes a state space model, a long short-term memory network, a gradient boosting model, and a moving average convergence divergence model. Each model has its own unique advantages and applicable scenarios.

[0056] Before inputting the key influencing factors into the model, they need to be reconstructed into time series, forming input feature sequences with fixed time window lengths. Time series reconstruction organizes the previously discrete key influencing factor data into chronological order, allowing the model to better capture temporal dependencies within the data. For example, a fixed time window length can be selected to combine key influencing factor data over a continuous period of time into an input feature sequence. This way, each input feature sequence captures the changes in each key influencing factor within that time window.

[0057] Step S131: inputting the input feature sequence into the state space model, capturing the periodic fluctuation characteristics of the feature sequence through a structured state space conversion mechanism, and outputting the first prediction result.

[0058] When a sequence of input features is fed into a state-space model, the model processes this data through a structured state-space transformation mechanism. The core of the state-space model is the discretization and spatialization of the structured state-space model. The continuous state-space model describes the evolution of the state h(t) over time, which can be expressed as h'(t) = Ah(t) + Bx(t). The output equation maps the state to the output, which can be expressed as y(t) = Ch(t) + Dx(t). After discretization, we get h_t = A_barh_{t-1} + B_barx_t, where A_bar = e^(ΔA) and B_bar = (ΔA)^(-1)(e^(ΔA)-I)ΔB.

[0059] First, the input feature sequence is dimensionally mapped using a fully connected network to generate an initial state vector with a uniform hidden dimension. The fully connected network integrates and transforms the different dimensions of the input feature sequence so that all input data can be represented in a unified hidden dimension space. For example, the input feature sequence may contain multiple key influencing factors of different types, each with varying dimensions. These factors are mapped into a fixed-dimensional latent space using a fully connected network to generate the initial state vector.

[0060] Next, a state transition equation is constructed based on the linear combination of the current input features and the hidden state at the previous moment. This equation describes how the state transitions from one moment to the next, taking into account the current input features and the hidden state at the previous moment. Using this equation, the initial state vector is dynamically evolved to generate a new hidden state. This new hidden state contains more time series information, reflecting not only the current input features but also the previous state information.

[0061] The new hidden state is then discretized using matrix exponential operations and integral approximation. Since state-space models are typically continuous-time models, discretization is required in practical calculations. Matrix exponential operations and integral approximation convert the continuous state transition process into a discrete form, generating a discretized state transition matrix and an input weight matrix. The discretized state transition matrix describes the transition relationship between states in discrete time steps, while the input weight matrix represents the influence of input features on state transitions.

[0062] A multi-step update is performed on the new hidden state based on the discretized state transition matrix and the input weight matrix. At each time step, the hidden state is updated based on the discretized state transition matrix and the input weight matrix. During this update, the retained weights of different feature components in the hidden state are dynamically adjusted. For example, feature components that are more important for bond price prediction are given higher weights, while less important feature components are appropriately reduced in weight. After these multi-step updates, the updated hidden state is generated.

[0063] Finally, the hidden state, updated through multiple steps, is input into the output mapping layer. The output mapping layer converts the hidden state into the raw prediction value of the state-space model through a combination of linear transformations and nonlinear activation functions. Linear transformations allow for linear combinations of hidden states, while nonlinear activation functions introduce nonlinear factors, increasing the model's expressive power. Based on this raw prediction value, the first prediction result is determined. For example, if the raw prediction value is greater than a certain threshold, the bond price is predicted to rise; if the raw prediction value is less than a certain threshold, the bond price is predicted to fall.

[0064] For example, the standardized features x of the previous N days can be used as input data, that is, an N*16 tensor, which is input to a fully connected layer with 16 hidden layers. Then, it is input to a Mamba layer with 16 hidden layers and 3 Mamba network layers. The output is then sent to a fully connected layer with 16 hidden layers. It then passes through a Dropout layer with a probability of 0.2 and finally passes through a PReLU activation function. Output values ​​greater than 0 are considered to be rising, and values ​​less than 0 are considered to be falling.

[0065] The Adam optimizer is used for model training. Combining the strengths of AdaGrad and RMSProp, the Adam optimizer adaptively adjusts the learning rate of each parameter, enabling the model to converge to the optimal solution more quickly during training. The mean squared error (MSE) is used as the loss function. MSE measures the difference between the model's predicted values ​​and the true values. By minimizing MSE, the model's predicted results can be made closer to the true values.

[0066] For example, the early stopping mechanism can be set to a certain number of rounds during model training. For example, when the loss function does not decrease significantly within a certain number of rounds, the training is stopped to prevent the model from overfitting. The number of training rounds can also be set with an upper limit, for example, set to a certain value, to ensure that the model has enough training times but is not overtrained. The initial learning rate will be set according to the specific situation. The appropriate initial learning rate can enable the model to converge quickly in the early stages of training. During the training process, the parameters of the model are continuously adjusted so that the loss function gradually decreases until the early stopping condition is met or the preset number of training rounds is reached, at which point the state space model training is completed.

[0067] Step S1311: Dimension mapping is performed on the input feature sequence through a fully connected network to generate an initial state vector with a uniform hidden dimension.

[0068] When performing dimensional mapping through a fully connected network, the network consists of multiple neurons, each of which receives all elements in the input feature sequence and performs a weighted sum. Assume the dimension of the input feature sequence is m, and the hidden dimension of the fully connected network is n. Each neuron has m weights, one for each of the m elements in the input feature sequence. The neuron multiplies each element of the input feature sequence by the corresponding weight, then adds these products, adds a bias term, and finally performs a nonlinear transformation using an activation function (such as the ReLU function). In this way, after processing by the fully connected network, the input feature sequence is mapped from m dimensions to an n-dimensional latent space, generating an initial state vector with uniform latent dimensions. This initial state vector better represents the information in the input feature sequence and facilitates subsequent processing of the state-space model.

[0069] Step S1312: constructing a state transfer equation based on a linear combination of the current input feature and the hidden state at the previous moment, and dynamically evolving the initial state vector through the state transfer equation to generate a new hidden state.

[0070] When constructing the state transfer equation, the linear combination of the current input features and the hidden state at the previous moment can be considered. Assume that the current input feature vector is x(t) and the hidden state vector at the previous moment is h(t-1). The state transfer equation can be expressed as h(t)=A*h(t-1)+B*x(t), where A is the state transfer matrix and B is the input weight matrix. The state transfer matrix A describes the influence of the hidden state at the previous moment on the current hidden state, and the input weight matrix B describes the influence of the current input features on the current hidden state. Through this state transfer equation, the initial state vector is dynamically evolved. Starting from the initial state vector h(0), based on the current input feature vector x(1), the hidden state h(1) of the first time step is calculated. Then, using h(1) as the hidden state at the previous moment, combined with the input feature vector x(2) of the next time step, h(2) is calculated. And so on, the hidden state is continuously updated, and finally a new hidden state is generated. This new hidden state contains the information of the input feature sequence at multiple time steps and can better reflect the dynamic changes of the feature sequence.

[0071] Step S1313: Discretize the new hidden state through matrix exponential operation and integral approximation calculation to generate a discretized state transfer matrix and an input weight matrix.

[0072] Since state-space models are typically continuous-time models, they need to be discretized for computation on a computer. Matrix exponential operations and integral approximations play a key role in this discretization process. First, the continuous-time state transition equation is converted to discrete-time form through matrix exponential operations. Matrix exponential operations calculate the discrete-time state transition matrix based on the continuous-time state transition matrix and the time interval. Simultaneously, integral approximations are used to account for the influence of input features on the state, resulting in a discretized input weight matrix. For example, numerical integration methods (such as the Euler method or the Runge-Kutta method) are used to approximate the integral of the input features over a time interval to obtain the input weight matrix. After these operations, the new hidden state is discretized, resulting in the discretized state transition matrix and input weight matrix.

[0073] The discretized state transition matrix describes the state transition relationship between discrete time steps. It reflects how the hidden state at the previous moment affects the hidden state at the current moment. This matrix can be used to predict the changes in the hidden state at different time steps. The input weight matrix represents the influence of input features on state transitions. Different input features may have different contributions to state transitions. The input weight matrix quantifies these contributions, allowing the role of input features to be accurately accounted for when updating the state.

[0074] Step S1314: performing a multi-step update on the new hidden state based on the discretized state transfer matrix and the input weight matrix, dynamically adjusting the retention weights of different feature components in the hidden state, and generating a hidden state that has undergone multi-step updates.

[0075] After obtaining the discretized state transfer matrix and the input weight matrix, we can perform a multi-step update on the new hidden state. At each time step, the hidden state is updated based on the discretized state transfer matrix and the input weight matrix. Specifically, the hidden state at the previous moment is multiplied by the discretized state transfer matrix, and then the current input features are multiplied by the input weight matrix to obtain the current hidden state.

[0076] During the update process, the retained weights of different feature components in the hidden state are dynamically adjusted. This is because different feature components may have different importance for bond price prediction, and this importance may also change at different time steps. For example, during periods of high market volatility, certain feature components reflecting market sentiment may be more important for bond price prediction, while during periods of relative market stability, certain macroeconomic indicator feature components may be more critical. By dynamically adjusting the weights, the hidden state can more accurately reflect the information relevant to bond price prediction.

[0077] After multiple updates, a hidden state is generated. This hidden state combines the input feature information and state transition information of multiple time steps, and can more comprehensively capture the periodic fluctuation characteristics of the feature sequence.

[0078] Step S1315: The hidden state that has undergone multiple steps of updating is input into the output mapping layer, and the original prediction value of the state space model is generated through a combination of linear transformation and nonlinear activation function.

[0079] The hidden state, updated over multiple steps, is input to the output mapping layer. The output mapping layer converts the hidden state into the original predictions of the state-space model. First, a linear transformation is performed, which allows the hidden states to be linearly combined. For example, each component of the hidden state is multiplied by its corresponding weight, and then these products are summed to produce a linear combination. This linear combination provides a preliminary integration of the hidden state information.

[0080] Next, the linear combination result is processed using a nonlinear activation function. Nonlinear activation functions can introduce nonlinear factors and increase the model's expressiveness. Common nonlinear activation functions include the Sigmoid function and the ReLU function. Different nonlinear activation functions have different characteristics. Choosing the appropriate nonlinear activation function can help the model better fit the data.

[0081] Through the combination of linear transformation and nonlinear activation function, the original prediction value of the state space model is finally generated. This original prediction value is the preliminary prediction result of the bond price based on the hidden state.

[0082] Step S1316: Determine the first prediction result according to the original prediction value.

[0083] The first prediction is determined based on the raw prediction value generated by the state-space model. A threshold can be set and the raw prediction value compared to the threshold. If the raw prediction value is greater than the threshold, the bond price is predicted to increase; if the raw prediction value is less than the threshold, the bond price is predicted to decrease. For example, if the raw prediction value is greater than a pre-set positive threshold, the model predicts that the bond price will increase in the future. In this case, the first prediction is a bond price increase. Conversely, if the raw prediction value is less than a pre-set negative threshold, the first prediction is a bond price decrease.

[0084] Step S132: inputting the input feature sequence into the long short-term memory network, processing the long-term dependency of the feature sequence through the gated recurrent unit, and outputting the second prediction result.

[0085] The input feature sequence is fed into a Long Short-Term Memory (LSTM) network. LSTM networks are well-suited for processing sequence data with long-term dependencies and are particularly important for processing time series data from the bond market. The core of LSTM networks is the gated recurrent unit (GRU), which effectively handles long-term dependencies in feature sequences.

[0086] Step S1321: converting the input feature sequence into a tensor of preset time length and feature dimension, wherein the preset time length corresponds to the time span of continuous feature data, and the feature dimension corresponds to the number of features at each time node.

[0087] First, the input feature sequence is converted into a tensor of a preset time length and feature dimension. The preset time length determines the time span of the continuous feature data and can be set based on specific needs and data characteristics. For example, if you are interested in short-term bond price fluctuations, you can set a shorter preset time length; if you want to analyze long-term market trends, you can set a longer preset time length.

[0088] The feature dimension corresponds to the number of features at each time point and contains information about multiple key influencing factors. The input feature sequence is organized according to the preset time length and feature dimension to form a tensor. This tensor can be better processed in the long short-term memory network, allowing the network to simultaneously consider information from both the time and feature dimensions.

[0089] Step S1322: Input the tensors of the preset time length and feature dimension into each network layer of the long short-term memory network in sequence, and perform feature extraction and dimension conversion processing on the tensors of the preset time length and feature dimension through each network layer.

[0090] Tensors of preset time length and feature dimensions are sequentially fed into the various layers of a LSTM network. LSTM networks typically consist of multiple layers, each with its own specific function. During the input process, each layer extracts features and converts the dimensions of the tensor.

[0091] The network layer extracts features relevant to bond price prediction from the input tensor. Different layers may extract features at different levels, from low-level features to higher-level features. Furthermore, the network layer transforms the tensor's dimensions, allowing the features to be expressed in different dimensional spaces for better processing by subsequent layers.

[0092] Step S1323: Input the feature data processed by the various network layers into the linear function activation layer, and perform linear transformation on the feature data through the linear function activation layer to generate a regression result.

[0093] The numerical value of the regression result corresponds to the long-term dependency analysis result of the input feature sequence.

[0094] The feature data processed by each network layer is fed into the linear function activation layer. The linear function activation layer performs a linear transformation on the feature data to generate a regression result. The linear transformation linearly combines the various components of the feature data to produce a regression value.

[0095] The magnitude of this regression result corresponds to the analysis of the long-term dependencies in the input feature sequence. The LSTM network processes the long-term dependencies in the feature sequence through gated recurrent units, and the regression result reflects the impact of these long-term dependencies on bond prices. For example, a large regression result indicates that the long-term dependencies in the input feature sequence indicate an upward trend in bond prices; conversely, a small regression result indicates a possible downward trend in bond prices.

[0096] Step S1324: Generate the second prediction result based on the regression result.

[0097] A second prediction is generated based on the regression results generated by the linear function activation layer. Similarly, a threshold can be set and the regression results compared to this threshold. If the regression result is greater than the threshold, the bond price is predicted to rise; if the regression result is less than the threshold, the bond price is predicted to fall. In this way, the regression results are converted into predictions of bond price increases and decreases, resulting in the second prediction output by the long short-term memory network.

[0098] Step S133: inputting the input feature sequence into the gradient boosting model, and generating the third prediction result by superimposing the prediction results of multiple decision trees.

[0099] The input feature sequence is fed into the gradient boosting model. The gradient boosting model is an ensemble learning model that generates a final prediction by stacking the predictions of multiple decision trees. A decision tree is a tree-based decision-making model, where each decision tree can analyze and predict the input feature sequence.

[0100] Step S1331: converting the input feature sequence into a sample feature set in a preset sample format, wherein the preset sample format corresponds to the feature input dimension required by the gradient boosting model, and the sample feature set includes the numerical information of each feature in the input feature sequence and the correlation relationship between features.

[0101] Convert the input feature sequence into a set of sample features in a predefined sample format. The predefined sample format is determined by the feature input dimensions required by the gradient boosting model. The gradient boosting model has certain requirements for the input data format, and the input feature sequence needs to be converted according to these requirements.

[0102] The sample feature set contains the numerical information for each feature in the input feature sequence and the inter-feature correlations. The numerical information for each feature represents the specific value of each key influencing factor in the input feature sequence, while the inter-feature correlations reflect the interactions and influences between these key influencing factors. For example, there may be a correlation between interest rate features and stock index features, and this correlation will be included in the sample feature set.

[0103] Step S1332: inputting the sample feature set in the preset sample format into the gradient boosting model in sequence, and performing feature screening, decision tree construction and prediction result superposition processing on the sample feature set through the gradient boosting model.

[0104] The sample feature set in a pre-set sample format is sequentially fed into the gradient boosting model. The gradient boosting model then filters the sample feature set to identify the most important features for bond price prediction. This feature filtering reduces the interference of unnecessary features on the model, improving its efficiency and accuracy.

[0105] Next, the gradient boosting model constructs multiple decision trees. Each decision tree analyzes and predicts the sample feature set based on the selected features. During the decision tree construction process, nodes are divided according to the feature values, forming a tree-like structure. Each decision tree outputs a prediction result.

[0106] Finally, the gradient boosting model stacks the predictions from multiple decision trees. This stacking can be done in a weighted manner, assigning different weights to each decision tree based on its performance and accuracy. These weighted predictions are then summed together to produce a composite prediction.

[0107] Step S1333: Input the processed superposition prediction results into the output layer of the gradient boosting model, and perform comprehensive calculation on the superposition prediction results through linear integration processing to generate a total prediction result. The numerical value of the total prediction result is positively correlated with the comprehensive influence intensity of each feature in the sample feature set.

[0108] The processed superimposed prediction results are fed into the output layer of the gradient boosting model. The output layer integrates the superimposed prediction results to generate a final prediction. Through linear integration, the components of the superimposed prediction results are linearly combined. For example, the prediction results of each decision tree are multiplied by their corresponding weights, and these products are then summed to obtain the final prediction result.

[0109] The magnitude of the total prediction result is positively correlated with the combined influence of each feature in the sample feature set. The combined influence of each feature in the sample feature set refers to the combined effect of each key influencing factor on the bond price. A large total prediction result indicates that the combined influence of each feature in the sample feature set is causing the bond price to rise; conversely, a small total prediction result indicates that the bond price may be declining.

[0110] Step S1334: Generate the third prediction result based on the total prediction result.

[0111] A third prediction is generated based on the total prediction generated by the gradient boosting model. A threshold is set and the total prediction is compared with the threshold. If the total prediction is greater than the threshold, the bond price is predicted to rise; if the total prediction is less than the threshold, the bond price is predicted to fall. In this way, the total prediction is converted into a prediction of whether the bond price will rise or fall, resulting in the third prediction output by the gradient boosting model.

[0112] Step S134: inputting the input feature sequence into the moving average convergence divergence model, and generating the fourth prediction result by calculating the deviation value and the deviation average of the price fluctuation difference sequence.

[0113] The input feature sequence is fed into the moving average convergence divergence model. The moving average convergence divergence model is primarily used to analyze bond price trends. It generates forecasts by calculating the deviations and mean deviations of the price fluctuation difference sequence.

[0114] Step S1341: extracting bond price data from the input feature sequence.

[0115] Extract bond price data from the input feature sequence. The bond price data here can be the closing price, opening price, etc. Extracting bond price data is the basis for calculating the moving average convergence divergence model, and subsequent calculations will be based on this price data.

[0116] Step S1342: Calculate a short-term weighted moving average and a long-term weighted moving average based on the bond price data, wherein the calculation period of the short-term weighted moving average is shorter than the calculation period of the long-term weighted moving average.

[0117] The short-term weighted moving average and the long-term weighted moving average are calculated based on the extracted bond price data. The calculation period of the short-term weighted moving average and the long-term weighted moving average are different, and the calculation period of the short-term weighted moving average is shorter than that of the long-term weighted moving average.

[0118] When calculating a weighted moving average, each data point is weighted differently based on its timeliness. Data points closer to the current time are weighted more heavily. For example, for a short-term weighted moving average, bond price data from a recent period is selected, weighted differently based on the chronological order of these data points, and then the weighted average is calculated. For a long-term weighted moving average, a similar weighted average is performed using bond price data from a longer period.

[0119] Step S1343: Subtract the short-term weighted moving average from the long-term weighted moving average to generate a deviation value sequence.

[0120] The difference between the calculated short-term weighted moving average and the long-term weighted moving average is used to generate a deviation series. The deviation series reflects the difference between short-term and long-term bond price trends. A positive deviation indicates that the short-term bond price trend is higher than the long-term bond price trend; a negative deviation indicates that the short-term bond price trend is lower than the long-term bond price trend.

[0121] Step S1344: Calculate a moving average value for the deviation value sequence to generate a deviation mean value sequence.

[0122] Calculate a moving average on a sequence of deviation values ​​to generate a sequence of deviation means. Calculating a moving average can smooth the sequence of deviation values, reduce the impact of short-term fluctuations, and better reflect the long-term trend of the deviation values. For example, select a certain number of deviation values ​​and average them to obtain a deviation mean. Over time, a series of deviation means are calculated sequentially to form a sequence of deviation means.

[0123] Step S1345: Subtract the deviation value sequence from the deviation mean value sequence to generate a moving average convergence divergence indicator sequence.

[0124] Subtracting the deviation series from the mean deviation series generates the moving average convergence divergence indicator series. The moving average convergence divergence indicator series is the core output of the moving average convergence divergence model. It combines the difference between short-term and long-term bond price trends and the changing trend of this difference.

[0125] Step S1346: performing trend analysis on the moving average convergence divergence indicator sequence, generating an upward trend signal when the moving average convergence divergence indicator sequence changes from negative to positive, and generating a downward trend signal when the moving average convergence divergence indicator sequence changes from positive to negative.

[0126] Perform trend analysis on the Moving Average Convergence Divergence indicator series. When the Moving Average Convergence Divergence indicator series turns from negative to positive, it indicates that the short-term bond price trend has begun to exceed the long-term bond price trend, and the market may be on the rise. At this time, an uptrend signal is generated. Conversely, when the Moving Average Convergence Divergence indicator series turns from positive to negative, it indicates that the short-term bond price trend has begun to fall below the long-term bond price trend, and the market may be on the decline. At this time, a downtrend signal is generated.

[0127] Step S1347: Determine the fourth prediction result based on the upward trend signal or the downward trend signal.

[0128] The fourth prediction is determined based on the generated upward or downward trend signal. If an upward trend signal appears, meaning the Moving Average Convergence Divergence indicator sequence turns from negative to positive, this indicates that the short-term bond price trend is beginning to surpass the long-term bond price trend, potentially signaling a market uptrend. In this case, a bond price increase is predicted. If a downward trend signal appears, meaning the Moving Average Convergence Divergence indicator sequence turns from positive to negative, this indicates that the short-term bond price trend is beginning to fall below the long-term bond price trend, potentially signaling a market downturn. In this case, a bond price decrease is predicted. In this manner, a prediction of bond price increases or decreases is obtained based on the analysis results of the Moving Average Convergence Divergence model, representing the fourth prediction.

[0129] Step S140: Input the correlation coefficient prediction result, the first prediction result, the second prediction result, the third prediction result, and the fourth prediction result into the AdaBoost classifier for training, and learn the weight influence of each prediction result through the cross entropy loss function to obtain the bond price rise and fall prediction model.

[0130] The correlation coefficient prediction results, the first prediction results, the second prediction results, the third prediction results, and the fourth prediction results are input into an AdaBoost classifier for training. The AdaBoost classifier is an ensemble learning algorithm that iteratively trains multiple weak classifiers to gradually improve classification accuracy. Traditional bond price prediction often uses only a single model, which fails to fully integrate the advantages of multiple information sources. However, this embodiment uses the AdaBoost classifier to fuse the prediction results of multiple models, enabling a more comprehensive consideration of the impact of various factors on bond prices.

[0131] Step S141: performing feature concatenation on the correlation coefficient prediction result, the first prediction result, the second prediction result, the third prediction result, and the fourth prediction result to form an input feature vector of the AdaBoost classifier.

[0132] The correlation coefficient prediction results, the first prediction results, the second prediction results, the third prediction results, and the fourth prediction results are subjected to feature concatenation. Feature concatenation combines these prediction results from different sources in a certain order to form a high-dimensional input feature vector. For example, the components of the correlation coefficient prediction results are first arranged in sequence, and then the first prediction results, the second prediction results, the third prediction results, and the fourth prediction results are concatenated. This input feature vector contains the bond price prediction information from multiple models, providing rich data for training the AdaBoost classifier, enabling the classifier to judge bond price fluctuations from multiple perspectives.

[0133] Step S142: Initialize multiple weak classifiers, each of which corresponds to a weight parameter of a prediction result.

[0134] Initialize multiple weak classifiers, each with a corresponding weight parameter for a prediction result. Weak classifiers can use simple classification models, such as decision stumps. A decision stump is a simple decision tree consisting of a root node and two leaf nodes, offering a simple structure and fast computational speed. The weight parameter for each weak classifier represents the importance of the prediction result in the final classification. Initially, each weak classifier is assigned an initial weight parameter. This initial weight parameter can be set based on experience or simple statistical methods. For example, the initial weight parameters of all weak classifiers can be set to the same value.

[0135] Step S143: updating the weight parameters of each weak classifier by adopting an iterative training method, and dynamically adjusting the voting weight according to the classification error. The smaller the classification error, the larger the weight parameter corresponding to the prediction result.

[0136] An iterative training approach is used to update the weight parameters of each weak classifier. In each iteration, the AdaBoost classifier classifies the input feature vector based on the current weak classifier and calculates the classification error. The classification error refers to the difference between the weak classifier's classification result and the true label. The true label is determined by the difference between the closing price and the opening price. If the difference is positive, the true label is up; if the difference is negative, the true label is down.

[0137] The voting weight is dynamically adjusted based on the classification error. Predictions with smaller classification errors have larger corresponding weight parameters, meaning that these predictions have a higher voting weight in the final classification. For example, if a weak classifier has a small classification error, it indicates that the corresponding prediction result is more accurate, and its weight parameter will be increased; conversely, if the classification error is large, its weight parameter will be decreased. Specifically, in each iteration, an adjustment factor is calculated based on the classification error, and this adjustment factor is then used to update the weight parameters of the weak classifier. The calculation of the adjustment factor is generally related to the magnitude of the classification error. The smaller the classification error, the larger the adjustment factor, and the greater the increase in the weight parameter.

[0138] Step S144: Calculate the loss value of the integrated prediction result and the true rise and fall label based on the difference between the closing price and the opening price through the cross entropy loss function, and optimize the weight parameters of each weak classifier through back propagation.

[0139] The cross-entropy loss function is used to calculate the loss between the ensemble prediction and the true rise / fall label, which is based on the difference between the closing and opening prices. The ensemble prediction is a weighted combination of the classification results of multiple weak classifiers according to the weight parameters. The cross-entropy loss function measures the degree of difference between the ensemble prediction and the true rise / fall label.

[0140] When calculating the cross-entropy loss function, the probabilities of predicted increases and decreases are calculated based on the ensemble predictions, and then compared with the actual increase / decrease labels. For example, if the actual increase / decrease label is an increase, the cross-entropy loss value will be smaller if the ensemble predictions have a higher probability of an increase; conversely, if the probability of an increase is lower, the cross-entropy loss value will be larger.

[0141] After obtaining the loss value, the backpropagation algorithm is used to optimize the weight parameters of each weak classifier. The backpropagation algorithm is an optimization method based on gradient descent, which adjusts the weight parameters according to the gradient of the loss value. Specifically, the partial derivative of the loss value with respect to each weak classifier weight parameter is calculated, and then the weight parameters are updated based on the direction and magnitude of the partial derivative. If the partial derivative is positive, it means that increasing the weight parameter will increase the loss value, so the weight parameter is reduced; conversely, if the partial derivative is negative, the weight parameter is increased. By continuously iteratively updating the weight parameters, the loss value gradually decreases until it reaches a smaller value or meets the preset convergence conditions.

[0142] Step S145: Repeat the weight updating and loss optimization steps until the loss value converges to a preset threshold, thereby generating the bond price fluctuation prediction model.

[0143] The weight update and loss optimization steps are repeated. In each iteration, the ensemble predictions are recalculated based on the new weight parameters, the cross-entropy loss is recalculated, and the weight parameters are updated using the backpropagation algorithm. As the iterations progress, the loss value gradually decreases.

[0144] A preset threshold is set. When the loss converges to this threshold, the model has achieved satisfactory training results. At this point, iterations are stopped, and a bond price fluctuation prediction model is generated. This model integrates information from multiple prediction results and continuously optimizes the weight parameters of weak classifiers to more accurately predict bond price fluctuations.

[0145] Step S150: Outputting bond recommendation results based on the bond price fluctuation prediction model.

[0146] In this embodiment, after obtaining the bond price increase and decrease prediction model, the bond recommendation results can be output based on the model to provide investors with valuable investment advice.

[0147] Step S151: inputting the multi-source feature data updated in real time into the bond price rise and fall prediction model to obtain a rise and fall prediction result based on the difference between the closing price and the opening price on the next trading day.

[0148] The bond price fluctuation prediction model is fed with real-time, multi-source feature data. This data is updated daily after the market close and contains the latest bond prices and associated market indices. Based on this real-time data, the model predicts the difference between the closing and opening prices of the bond for the next trading day, generating a price increase or decrease prediction. For example, if the model outputs a positive difference, the bond price is predicted to rise on the next trading day; if the difference is negative, the bond price is predicted to fall.

[0149] Step S152: Calculate the accuracy index of the rise and fall prediction results in three consecutive increasing time windows, where the accuracy index is the ratio of the number of correctly predicted days to the total number of days in each time window.

[0150] Calculate the accuracy of the rise and fall prediction results within three consecutive increasing time windows. A time window is a period of time divided by a certain time span. Three consecutive increasing time windows means that the next time window contains the previous time window, and the time span is gradually increasing. For example, the first time window can be the most recent time period, the second time window extends back a certain number of days based on the first time window, and the third time window extends further based on the second time window.

[0151] For each time window, the number of days with correct predictions and the total number of days are counted. Correct predictions refer to the number of days for which the model's predicted price fluctuations matched the actual price fluctuations, while the total number of days refers to the number of trading days within that time window. The ratio of correct predictions to total days is used as the accuracy metric for that time window. By calculating the accuracy metrics for three different time windows, the model's forecast accuracy can be evaluated from different time scales. A higher accuracy rate within a shorter time window indicates that the model has strong predictive power for short-term market fluctuations; a higher accuracy rate within a longer time window indicates that the model has good long-term stability.

[0152] Step S153: performing weighted fusion on the accuracy index and the confidence of the prediction result to generate a comprehensive recommendation score.

[0153] A weighted fusion of the accuracy metric and the confidence level of the prediction results is performed. The confidence level of the prediction results is the degree of certainty of the model's own predictions. It can be calculated through internal mechanisms within the model, such as the probability value output by the model. If the model outputs a prediction of an increase, and the given probability of an increase is high, then the model has a high degree of confidence in the prediction result.

[0154] The accuracy metric and the confidence level of the prediction results are each multiplied by their corresponding weights, and then added together to obtain a composite recommendation score. The weighting can be adjusted based on the actual situation. If accuracy is more important, a higher weight can be assigned to it; if confidence in the model is more important, a higher weight can be assigned to confidence. For example, based on historical data analysis, the importance of accuracy and confidence levels for bond recommendations in different market environments can be determined, and the weighting can be set accordingly.

[0155] Step S154: Sort the bonds according to the comprehensive recommendation scores, and select the bonds with the highest ranking to form a recommendation list.

[0156] Rank bonds based on their overall recommendation scores. All analyzed bonds are ranked from highest to lowest based on their overall recommendation scores, with higher scores indicating a higher recommendation value. Top-ranked bonds are selected to form a recommended list. The recommended list contains bonds identified as having the greatest investment potential based on model predictions and comprehensive evaluation. These bonds demonstrate high accuracy and confidence in predicting price increases and decreases, allowing investors to make investment decisions based on this list.

[0157] Step S155: The recommendation list is associated with the corresponding set of key influencing factors and the prediction results of each model and stored to generate the bond recommendation result.

[0158] The recommendation list is stored in association with the corresponding set of key influencing factors and the prediction results of each model. The key influencing factor set includes information such as interest rate characteristics, index characteristics, and trading volume characteristics that have a significant impact on bond prices. The prediction results of each model include the first prediction result of the state-space model, the second prediction result of the long-short-term memory network, the third prediction result of the gradient boosting model, and the fourth prediction result of the moving average convergence divergence model. This associated storage allows investors to understand the prediction basis and relevant influencing factors for each recommended bond, generating a complete bond recommendation result. For example, investors can view the key influencing factors of a recommended bond to understand the factors that led to the bond's recommendation. At the same time, they can also view the prediction results of each model and compare the predictions of different models for the bond's price increase or decrease, thereby more comprehensively evaluating the bond's investment value.

[0159] The method provided in the embodiment of the present application first obtains multi-source feature data of bond transactions, then performs dynamic feature screening on these data to obtain prediction results of key influencing factors and correlation coefficients, then inputs the key influencing factors into a multi-model collaborative prediction framework to generate multiple prediction results, and then inputs various prediction results into the AdaBoost classifier to train a bond price rise and fall prediction model, and finally outputs bond recommendation results based on the model. The comprehensive acquisition of multi-source data ensures the richness of information, dynamic feature screening can accurately focus on key influencing factors, multi-model collaborative prediction combined with daily retraining can improve the timeliness and diversity of predictions, and the AdaBoost classifier can optimize the comprehensive prediction effect by learning the weights of each prediction result, thereby improving the accuracy and efficiency of bond recommendation results.

[0160] Figure 2 This is a structural diagram of a bond recommendation system 100 based on machine learning and deep neural network provided in an embodiment of the present application. Figure 2 As shown, the processor 120 can be used in the bond recommendation system 100 based on machine learning and deep neural networks, and is used to perform the functions of the present invention.

[0161] The bond recommendation system 100 based on machine learning and deep neural networks can be a general-purpose server or a special-purpose server, both of which can be used to implement the bond recommendation method based on machine learning and deep neural networks of the present invention. Although only one server is shown in this invention, for convenience, the functions described in this invention can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0162] For example, the bond recommendation system 100 based on machine learning and deep neural networks may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. For example, the bond recommendation system 100 based on machine learning and deep neural networks may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present invention may be implemented based on these program instructions. The bond recommendation system 100 based on machine learning and deep neural networks may also include an input / output (I / O) interface 150 between the computer and other input / output devices.

[0163] For ease of explanation, only one processor is described in the bond recommendation system 100 based on machine learning and deep neural networks. However, it should be noted that the bond recommendation system 100 based on machine learning and deep neural networks in the present invention may also include multiple processors, so the steps performed by one processor described in the present invention may also be performed jointly or individually by multiple processors. For example, if the processor of the bond recommendation system 100 based on machine learning and deep neural networks performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually by one processor. For example, the first processor performs step A and the second processor performs step B, or the first processor and the second processor perform steps A and B together.

[0164] An embodiment of the present invention discloses a computer-readable storage medium storing a computer program for electronic data exchange, wherein the computer program enables a computer to execute the steps of the bond recommendation method based on machine learning and deep neural network of the aforementioned embodiment.

[0165] An embodiment of the present invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to enable a computer to execute the steps of the bond recommendation method based on machine learning and deep neural network of the aforementioned embodiment.

[0166] The device embodiments described above are merely illustrative. Modules described as separate components may or may not be physically separate, and components shown as modules may or may not be physical modules, i.e., they may be located in one place or distributed across multiple network modules. Some or all of these modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0167] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus the necessary general hardware platform, or of course, by means of hardware. Based on this understanding, the above technical solution, in essence, or the portion that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, including a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM), or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium capable of storing or storing data.

[0168] Finally, it should be noted that what is disclosed above is only a preferred embodiment of the present invention, which is only used to illustrate the technical solution of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features therein can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A bond recommendation method based on machine learning and deep neural network, characterized in that: include: Acquiring multi-source feature data related to bond transactions, wherein the multi-source feature data includes bond price data and related market index data; Performing dynamic feature screening processing on the multi-source feature data, obtaining a set of key influencing factors by fusing correlation analysis with historical trend weighting, and generating a correlation coefficient prediction result based on the correlation analysis; The set of key influencing factors is input into a multi-model collaborative prediction framework, and a first prediction result, a second prediction result, a third prediction result, and a fourth prediction result are generated respectively through a state space model, a long short-term memory network, a gradient boosting model, and a moving average convergence divergence model. The multi-model collaborative prediction framework is retrained after the market close each day based on the multi-source feature data updated on that day; Inputting the correlation coefficient prediction result, the first prediction result, the second prediction result, the third prediction result, and the fourth prediction result into an AdaBoost classifier for training, learning the weighted influence of each prediction result through a cross-entropy loss function, and obtaining a bond price increase and decrease prediction model, wherein the bond price increase and decrease prediction model uses the difference between the closing price and the opening price as the increase and decrease standard; Output bond recommendation results based on the bond price increase and decrease prediction model.

2. The bond recommendation method based on machine learning and deep neural network according to claim 1, characterized in that: The performing of the dynamic feature screening process on the multi-source feature data, obtaining the key influencing factor set by fusing the correlation analysis with the historical trend weighting, and generating the correlation coefficient prediction result based on the correlation analysis, includes: Calculating the correlation coefficients between the bond price data and the associated market index data to establish a characteristic correlation matrix; Performing the historical trend weighting processing on the feature correlation matrix, adjusting the weight proportions of correlations in different periods by a time decay function, and obtaining a weighted feature correlation matrix; Performing feature importance sorting based on the weighted feature correlation matrix, and extracting leading features from the result of the feature importance sorting as candidate influencing factors; Conduct dynamic adaptability verification on the candidate influencing factors and evaluate the stability indicators of each feature in the candidate influencing factors in different market cycles through a rolling window verification method; Filtering the key influencing factor set according to the product result of the stability index and the correlation weight, wherein the key influencing factor set includes interest rate characteristics, index characteristics, and trading volume characteristics; Based on the signs and numerical values ​​of the correlation coefficients between each of the features in the feature correlation matrix and the bond price data, a correlation coefficient prediction result representing the direction and intensity of the rise and fall is generated.

3. The bond recommendation method based on machine learning and deep neural network according to claim 1, characterized in that: The step of inputting the key influencing factor set into the multi-model collaborative prediction framework, and respectively generating the first prediction result, the second prediction result, the third prediction result, and the fourth prediction result through the state space model, the long short-term memory network, the gradient boosting model, and the moving average convergence divergence model includes: Reconstructing the key influencing factor set into a time series to form an input feature sequence with a fixed time window length; Inputting the input feature sequence into the state space model, capturing the periodic fluctuation characteristics of the feature sequence through a structured state space conversion mechanism, and outputting the first prediction result; Inputting the input feature sequence into the long short-term memory network, processing the long-term dependency of the feature sequence through a gated recurrent unit, and outputting the second prediction result; Inputting the input feature sequence into the gradient boosting model, and generating the third prediction result by superimposing the prediction results of multiple decision trees; The input feature sequence is input into the moving average convergence divergence model, and the fourth prediction result is generated by calculating the deviation value and the deviation average of the price fluctuation difference sequence.

4. The bond recommendation method based on machine learning and deep neural network according to claim 3, characterized in that: Inputting the input feature sequence into the state space model, capturing the periodic fluctuation characteristics of the feature sequence through the structured state space conversion mechanism, and outputting the first prediction result includes: Performing dimension mapping on the input feature sequence through a fully connected network to generate an initial state vector with a uniform hidden dimension; Constructing a state transfer equation based on a linear combination of the current input feature and the hidden state at the previous moment, and dynamically evolving the initial state vector through the state transfer equation to generate a new hidden state; Discretize the new hidden state by matrix exponential operation and integral approximation calculation to generate a discretized state transfer matrix and an input weight matrix; Performing a multi-step update on the new hidden state based on the discretized state transfer matrix and the input weight matrix, dynamically adjusting the retention weights of different feature components in the hidden state, and generating a multi-step updated hidden state; The hidden state after multiple steps of updating is input into the output mapping layer, and the original prediction value of the state space model is generated by a combination of linear transformation and nonlinear activation function; The first prediction result is determined according to the original prediction value.

5. The bond recommendation method based on machine learning and deep neural network according to claim 3, characterized in that: Inputting the input feature sequence into the long short-term memory network, processing the long-term dependency of the feature sequence through the gated recurrent unit, and outputting the second prediction result includes: Convert the input feature sequence into a tensor of preset time length and feature dimension, where the preset time length corresponds to the time span of continuous feature data, and the feature dimension corresponds to the number of features at each time node; Inputting the tensors of the preset time length and feature dimension into the respective network layers of the long short-term memory network in sequence, and performing feature extraction and dimension conversion processing on the tensors of the preset time length and feature dimension through the respective network layers; Inputting the feature data processed by each network layer into a linear function activation layer, and performing a linear transformation on the feature data through the linear function activation layer to generate a regression result, wherein the numerical value of the regression result corresponds to the long-term dependency analysis result of the input feature sequence; The second prediction result is generated according to the regression result.

6. The bond recommendation method based on machine learning and deep neural network according to claim 3, characterized in that: Inputting the input feature sequence into the gradient boosting model and generating the third prediction result by superimposing the prediction results of multiple decision trees includes: Converting the input feature sequence into a sample feature set in a preset sample format, wherein the preset sample format corresponds to the feature input dimension required by the gradient boosting model, and the sample feature set includes numerical information of each feature in the input feature sequence and correlation relationships between features; Inputting the sample feature set in the preset sample format into the gradient boosting model in sequence, and performing feature screening, decision tree construction, and prediction result superposition processing on the sample feature set through the gradient boosting model; The processed superposition prediction results are input into the output layer of the gradient boosting model, and the superposition prediction results are comprehensively calculated through linear integration processing to generate an overall prediction result, wherein the numerical value of the overall prediction result is positively correlated with the comprehensive influence strength of each feature in the sample feature set; The third prediction result is generated according to the total prediction result.

7. The bond recommendation method based on machine learning and deep neural network according to claim 3, characterized in that: Inputting the input feature sequence into the moving average convergence divergence model, and generating the fourth prediction result by calculating the deviation value and the deviation mean of the price fluctuation difference sequence, includes: extracting bond price data from the input feature sequence; Calculating a short-term weighted moving average and a long-term weighted moving average based on the bond price data, wherein a calculation period of the short-term weighted moving average is shorter than a calculation period of the long-term weighted moving average; Difference between the short-term weighted moving average and the long-term weighted moving average to generate a deviation value sequence; Calculating a moving average value for the deviation value sequence to generate a deviation mean value sequence; Subtracting the deviation value sequence from the deviation mean value sequence to generate a moving average convergence divergence indicator sequence; Performing trend analysis on the moving average convergence divergence indicator sequence, generating an upward trend signal when the moving average convergence divergence indicator sequence turns from negative to positive, and generating a downward trend signal when the moving average convergence divergence indicator sequence turns from positive to negative; The fourth prediction result is determined based on the upward trend signal or the downward trend signal.

8. The bond recommendation method based on machine learning and deep neural network according to claim 1, characterized in that: The step of inputting the correlation coefficient prediction result, the first prediction result, the second prediction result, the third prediction result, and the fourth prediction result into the AdaBoost classifier for training, and learning the weighted influence of each prediction result through the cross entropy loss function to obtain the bond price rise and fall prediction model includes: Perform feature splicing on the correlation coefficient prediction result, the first prediction result, the second prediction result, the third prediction result, and the fourth prediction result to form an input feature vector of the AdaBoost classifier; Initializing multiple weak classifiers, each of the weak classifiers corresponds to a weight parameter of a prediction result; The weight parameters of each weak classifier are updated by iterative training, and the voting weight is dynamically adjusted according to the classification error. The smaller the classification error, the larger the corresponding weight parameter of the prediction result. Calculating the loss value of the integrated prediction result and the true rise and fall label based on the difference between the closing price and the opening price through the cross entropy loss function, and back-propagating to optimize the weight parameters of each weak classifier; The weight updating and loss optimization steps are repeated until the loss value converges to a preset threshold, thereby generating the bond price rise and fall prediction model.

9. The bond recommendation method based on machine learning and deep neural network according to claim 1, characterized in that: Outputting the final bond recommendation result based on the bond price fluctuation prediction model includes: Inputting the multi-source feature data updated in real time into the bond price increase and decrease prediction model to obtain an increase and decrease prediction result based on the difference between the closing price and the opening price on the next trading day; Calculate the accuracy index of the rise and fall prediction results in three consecutive increasing time windows, where the accuracy index is the ratio of the number of correctly predicted days to the total number of days in each time window; Performing a weighted fusion of the accuracy index and the confidence of the prediction result to generate a comprehensive recommendation score; sorting the bonds according to the comprehensive recommendation scores and selecting the top-ranked bonds to form a recommendation list; The recommendation list is associated with the corresponding set of key influencing factors and the prediction results of each model and stored to generate the bond recommendation result.

10. A bond recommendation system based on machine learning and deep neural network, characterized in that: The invention comprises a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by a computer, the bond recommendation method based on machine learning and deep neural network described in any one of claims 1 to 9 is implemented.

Images