Macroeconomic index-driven market trend prediction system

Abstract

The invention relates to the technical field of market trend prediction, and discloses a market trend prediction system driven by macroeconomic indicators. An index acquisition module of the system dynamically acquires core economic indexes such as GDP growth rate, CPI, PMI and currency supply; the data preprocessing module is used for carrying out layered noise reduction processing on the multi-source heterogeneous data; the feature engineering module constructs a market sensitive feature set through spatio-temporal feature fusion; the prediction model building module is used for building a multi-layer nonlinear prediction model based on a deep belief network; the dynamic adjustment module adopts reinforcement learning to optimize a decision threshold value and combines a Markov chain to carry out state transition planning; and the feedback iteration module analyzes and predicts deviation through Bayesian filtering and realizes strategy updating. According to the method, deep learning and reinforcement learning technologies are creatively fused, the prediction precision is remarkably improved through a dynamic calibration mechanism, and the method can be widely applied to the macroeconomic analysis fields of financial investment, industrial planning and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of market trend prediction, in particular to a macroeconomic indicator driven market trend prediction system. BACKGROUND

[0002] Accurate understanding of market trends has an undeniable impact on enterprise decision-making, investment layout and economic regulation. With the increasing global economic linkage, market fluctuations are influenced by multiple macroeconomic factors, and traditional prediction methods gradually show obvious limitations.

[0003] Existing market trend prediction relies on a few core economic indicators, such as single gross domestic product growth rate or money supply, ignoring the relevance and dynamic influence between different indicators, resulting in a one-sided perspective. Although some methods attempt to introduce multiple indicators, the data collection mechanism is fixed and cannot be adjusted in real time according to changes in the economic environment, making it difficult to cover new data dimensions brought by sudden economic events or policy adjustments, resulting in insufficient timeliness and comprehensiveness of input data.

[0004] In the data processing link, the existing technology often uses simple cleaning or unified formatting for processing of multi-source heterogeneous data, lacking a hierarchical noise reduction mechanism for economic data characteristics. Economic data contains both short-term fluctuations and long-term trend signals, and rough data processing can easily lead to noise interference with valid information or misdeletion of key signals, affecting the basis for subsequent analysis.

[0005] In terms of feature engineering, traditional methods focus on feature extraction in a single time dimension, ignoring spatial correlation between different regions and different economic sectors, as well as dynamic evolution rules in time series. This makes it difficult for the extracted features to accurately reflect the sensitive response of the market to macroeconomic changes, and the feature dimension and representativeness are insufficient, which cannot provide strong support for the prediction model.

[0006] After the construction of the prediction model, most systems maintain a static structure and are difficult to adapt to the dynamic changes of the market environment. The market is often affected by policy adjustments, unexpected events, etc., showing nonlinear and sudden characteristics, and fixed model parameters and structure cannot capture these dynamic rules, leading to an enlarged deviation between the predicted sequence and the actual trend. At the same time, the model decision threshold lacks a dynamic optimization mechanism, making it difficult to adjust the judgment standard according to the market state, further reducing the prediction adaptability.

[0007] Existing systems often lack effective feedback iteration mechanisms, and the deviation between the prediction results and the actual trend is difficult to be captured and used for model optimization in time. Even if the deviation is found, it is difficult to trace back to the feature extraction, model construction and other links for targeted adjustment, resulting in the system maintaining the initial performance for a long time and being unable to continuously evolve with the market evolution. SUMMARY

[0008] The present application aims to provide a macroeconomic indicator-driven market trend prediction system to solve the problems raised in the background art.

[0009] To achieve the above-mentioned purpose, the present application provides a macroeconomic indicator-driven market trend prediction system, which comprises:

[0010] An indicator collection module collects macroeconomic operation data and dynamically adapts the data collection mechanism, including gross domestic product growth rate, consumer price index, purchasing manager index and money supply;

[0011] A data preprocessing module is used to receive multi-source heterogeneous data, perform hierarchical noise reduction processing on the data, and obtain a standardized data set;

[0012] A feature engineering module is used to construct a market sensitive feature set through spatio-temporal feature fusion based on the standardized data set, to obtain a high-dimensional feature vector;

[0013] A prediction model construction module constructs a multi-layer nonlinear prediction model through a deep belief network based on the high-dimensional feature vector and market historical rules, and generates a trend prediction sequence;

[0014] A dynamic adjustment module optimizes the model decision threshold through a reinforcement learning algorithm after receiving the trend prediction sequence, and plans state transition for the prediction process in combination with a Markov chain, to realize dynamic calibration of prediction accuracy;

[0015] A feedback iteration module tracks the deviation degree of the prediction result from the actual market trend through Bayesian filtering, and feeds back the deviation analysis result to the feature engineering module, the prediction model construction module and the dynamic adjustment module, to update the prediction strategy.

[0016] Preferably, macroeconomic operation data is collected and the data collection mechanism is dynamically adapted, including gross domestic product growth rate, consumer price index, purchasing manager index and money supply, including:

[0017] According to the economic cycle fluctuation characteristics and the correlation analysis of indicators, the time window, frequency and priority ranking of collection are determined;

[0018] According to the time window, frequency and priority ranking of collection, macroeconomic indicator data is collected, and the indicator data is processed in space-time alignment to obtain a standardized data set, the macroeconomic indicator data including gross domestic product growth rate, consumer price index, purchasing manager index and money supply.

[0019] Preferably, multi-source heterogeneous data is received, and the data is processed in hierarchical noise reduction to obtain a standardized data set, including:

[0020] The original collected data is divided into structured data and unstructured text data;

[0021] The noise reduction parameters and data cleaning rules are initialized; according to the difference of data types, the random fluctuations in the structured data are filtered by the moving average method, and the invalid information in the unstructured text is removed by semantic similarity matching;

[0022] According to the data integrity check result, the missing values are filled by interpolation method, and the abnormal data are identified and corrected by outlier detection algorithm, so as to obtain consistent data;

[0023] According to the dimension characteristics of macroeconomic indicators, the data is mapped to the preset interval by the maximum and minimum normalization method, so as to obtain the standardized data set.

[0024] Preferably, according to the standardized data set, market sensitive feature set is constructed by spatio-temporal feature fusion to obtain high-dimensional feature vector, including:

[0025] According to the standardized data set, the target dimension and association rule of feature construction are determined;

[0026] Initialize the feature pool, generate a set of basic statistical features and time series features as the initial feature candidate set;

[0027] The features in the candidate set are extended in time dimension and associated in cross-section dimension, the influence weight of the features on market trend is calculated by attention mechanism, and the key features are selected according to the weight value;

[0028] The selected key features are nonlinearly transformed and dimensionally combined to generate market sensitive feature set containing time domain, frequency domain and association features, so as to obtain high-dimensional feature vector.

[0029] Preferably, according to the high-dimensional feature vector and market historical law, a multi-layer nonlinear prediction model is constructed by deep belief network, and a trend prediction sequence is generated, including:

[0030] Obtain the high-dimensional feature vector and the corresponding market historical trend label;

[0031] Initialize the parameters of deep belief network;

[0032] According to the market historical law, the deep belief network architecture including input layer, hidden layer and output layer is constructed, the high-dimensional feature vector is input into the network for layer-by-layer pre-training, and the network weight is optimized by contrast divergence algorithm;

[0033] On the basis of pre-training, the network is fine-tuned by back propagation algorithm to minimize the mean square error between the predicted trend and the actual trend, so as to obtain the optimized multi-layer nonlinear prediction model;

[0034] Input the real-time high-dimensional feature vector into the optimized prediction model to generate a market trend prediction sequence for a future preset time period.

[0035] Preferably, after receiving the trend prediction sequence, the model decision threshold is optimized through a reinforcement learning algorithm, and a state transition plan is made for the prediction process in combination with a Markov chain to achieve dynamic calibration of prediction accuracy, including:

[0036] Receive the trend prediction sequence and preset the evaluation index and threshold range of prediction accuracy;

[0037] According to the preset evaluation index and the current prediction sequence, the prediction income under different thresholds is calculated through the Q-learning algorithm in reinforcement learning, and the optimal decision threshold is selected;

[0038] Under the constraint of the optimal threshold, the state transition probability of the market state is described through the Markov chain model, and the prediction sequence is multi-step state deduction to generate a dynamic adjustment strategy;

[0039] According to the dynamic adjustment strategy, the output result of the prediction model is calibrated in real time to achieve dynamic optimization of prediction accuracy.

[0040] Preferably, according to the preset evaluation index and the current prediction sequence, the prediction income under different thresholds is calculated through the Q-learning algorithm in reinforcement learning, and the optimal decision threshold is selected, including:

[0041] Set relevant parameters, including exploration rate, learning rate and discount factor;

[0042] According to the deviation of the current prediction sequence and the actual market trend, a reward function is constructed, and the reward value is positively correlated with the prediction accuracy;

[0043] The state-action value function is updated through the Q-learning algorithm, and in each iteration step, the threshold adjustment action is selected according to the current state, and the Q value is updated according to the reward function;

[0044] When the Q value converges or reaches the maximum number of iterations, the learning process is stopped, and the threshold that makes the Q value maximum is selected as the optimal decision threshold.

[0045] Preferably, the deviation degree of the prediction result and the actual market trend is tracked through Bayesian filtering, and the deviation analysis result is fed back to the feature engineering module, the prediction model construction module and the dynamic adjustment module to update the prediction strategy, including:

[0046] Collect the actual market trend data within the prediction period;

[0047] The deviation degree index is calculated by fusing the prediction result and the actual data through Bayesian filtering, including mean absolute error and trend direction accuracy rate;

[0048] The bias degree index is decomposed into feature contribution degree, model structure bias and threshold setting error, and a bias analysis report is generated;

[0049] The bias analysis report is fed back to the feature engineering module, the prediction model construction module and the dynamic adjustment module, triggering feature weight updating, model parameter optimization or threshold recalibration, to update the prediction strategy.

[0050] Preferably, it also includes an index early warning module, which is used to generate macroeconomic index abnormal fluctuation early warning signals according to the trend prediction sequence and the preset risk threshold, including:

[0051] Obtain key node data in the trend prediction sequence and the preset risk threshold interval;

[0052] The fluctuation amplitude of the index prediction value is monitored by the statistical process control method, and when the prediction value exceeds the threshold interval, the early warning mechanism is triggered;

[0053] According to the early warning level classification, text early warning information and visual warning signs are generated and pushed to the system terminal.

[0054] Preferably, it also includes a model explanation module, which is used to analyze the key influencing factors of the prediction result and generate an explainability report, including:

[0055] Obtain the trend prediction sequence output by the prediction model and the corresponding high-dimensional feature vector;

[0056] Calculate the contribution degree of each macroeconomic index to the prediction result through SHAP value, and generate feature importance ranking;

[0057] Combine market historical data to build an influence path diagram to show how key indicators affect the final prediction result through feature correlation;

[0058] Integrate the contribution degree ranking and the influence path diagram into a natural language report, and mark the data source and calculation method.

[0059] Compared with the prior art, the present application has the following advantages:

[0060] The system collects macroeconomic operation data through the index collection module and realizes dynamic adaptation of the data collection mechanism, which can flexibly adjust the data source and collection frequency according to the changes of the economic environment, covering gross domestic product growth rate, consumer price index, purchasing manager index, money supply and other types of key indicators, breaking the limitations of traditional single index or fixed collection mode, making the obtained data more in line with the actual operation of the market scene, providing a more comprehensive and timely information basis for subsequent analysis.

[0061] The data preprocessing module performs hierarchical noise reduction processing on multi-source heterogeneous data, can distinguish different types of noise and effective signals in the data, reduces the interference of noise on key information through hierarchical processing, preserves the long-term trend and short-term fluctuation characteristics contained in the data, improves the purity and representativeness of the standardized data set, avoids information distortion caused by rough data processing, and provides higher quality input data for the feature engineering module.

[0062] The feature engineering module constructs a market sensitive feature set through spatio-temporal feature fusion, incorporates the dynamic changes in the time dimension and the associated effects in the space dimension into the feature extraction process, not only enriches the dimension of the features, but also enhances the relevance of the features to market trends, and the generated high-dimensional feature vector can more accurately reflect the influence path and degree of macroeconomic changes on the market, making the features more suitable for prediction needs and providing more valuable input for the prediction model construction.

[0063] The prediction model construction module constructs a multi-layer nonlinear prediction model based on a deep belief network, which can capture the complex nonlinear relationship between macroeconomic indicators and market trends. Compared with traditional linear models or shallow neural networks, it is more suitable for describing the complex dynamic rules of the market affected by multiple factors, and the generated trend prediction sequence can more accurately reflect the potential evolution direction of the market, improving the fit of the prediction sequence to the actual market trend.

[0064] The dynamic adjustment module optimizes the model decision threshold through reinforcement learning algorithm, and plans the state transition of the prediction process combined with Markov chain, so that the model can adjust the decision criteria according to the real-time state of the market, and plan the prediction path under different market states, realizing the dynamic calibration of prediction accuracy, avoiding the problem of insufficient adaptability of static models in market mutations, and making the prediction process better follow the market changes, maintaining the timeliness and accuracy of the prediction.

[0065] The feedback iteration module tracks the deviation degree of the prediction results and the actual market trend by means of Bayesian filtering, and feeds back the deviation analysis results to the feature engineering module, the prediction model construction module and the dynamic adjustment module, forming a complete closed-loop optimization mechanism. This mechanism can timely find the shortcomings of each link of the system, adjust the feature extraction method, model structure parameters and dynamic adjustment strategy, and promote the continuous optimization of the whole system with the evolution of the market, so that the system can continuously adapt to new market characteristics in the long-term operation, and maintain stable prediction performance.

[0066] The modules work together to form a complete and dynamic prediction system from data collection, processing, feature extraction, model construction to dynamic adjustment and feedback optimization, which can effectively cope with the complexity, dynamics and uncertainty of the market, making the market trend prediction more in line with actual needs, and providing more valuable reference for related economic activities. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 The working principle diagram of the macroeconomic indicator driven market trend prediction system described in the present application;

[0068] Figure 2 The working flowchart of the indicator collection module;

[0069] Figure 3 The working principle diagram of the feature engineering module;

[0070] Figure 4 The flowchart of the Q-learning algorithm optimizing the decision threshold. DETAILED DESCRIPTION

[0071] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0072] Please refer to Figure 1 The present application provides a macroeconomic indicator driven market trend prediction system, which comprises an indicator collection module, a data preprocessing module, a feature engineering module, a prediction model construction module, a dynamic adjustment module and a feedback iteration module, and each module is sequentially and cooperatively operated. The specific implementation is as follows:

[0073] The indicator collection module is responsible for collecting macroeconomic operation data, covering gross domestic product growth rate, consumer price index, purchasing manager index and money supply, and dynamically adapting the data collection mechanism according to economic characteristics.

[0074] After receiving multi-source heterogeneous data, the data preprocessing module obtains a standardized data set through hierarchical noise reduction processing.

[0075] The feature engineering module constructs a market sensitive feature set based on the standardized data set, generates a high-dimensional feature vector through spatio-temporal feature fusion.

[0076] The prediction model construction module uses the high-dimensional feature vector and market historical rules to construct a multi-layer nonlinear prediction model through a deep belief network, and outputs a trend prediction sequence.

[0077] After receiving the trend prediction sequence, the dynamic adjustment module optimizes the model decision threshold with the help of reinforcement learning algorithm, and combines the state transition of Markov chain planning prediction process to realize dynamic calibration of prediction accuracy.

[0078] The feedback iteration module tracks the deviation degree of the prediction result from the actual market trend through Bayesian filtering, and feeds back the deviation analysis result to the feature engineering module, the prediction model construction module and the dynamic adjustment module to update the prediction strategy.

[0079] Embodiment 1: refer to Figure 2 During the operation of the index collection module, economic cycle fluctuation feature analysis and index correlation analysis are first carried out. The economic cycle fluctuation feature analysis needs to cover the duration, fluctuation amplitude and typical features of different cycle stages, such as identifying the specific performance of stages such as economic expansion, contraction, recovery, etc.; the index correlation analysis determines the correlation degree of different indicators in the process of influencing market trends by calculating the correlation coefficients between each macroeconomic indicator. Based on these analysis results, the time window, frequency and priority ranking of collection are determined. The setting of the time window needs to match the stage of the economic cycle, when the economy is in a rapid change stage, the time window can be set to a shorter cycle, and when the economy is in a relatively stable stage, the time window can be appropriately extended. The collection frequency is adjusted according to the time window, when the time window is shorter, the collection frequency is correspondingly increased to capture more detailed changes; when the time window is longer, the collection frequency can be appropriately reduced. The priority ranking is determined according to the close degree of the index and the market trend, and the index with high correlation degree is given higher priority in the collection process to ensure that key indicator data is acquired preferentially in the case of limited data collection resources.

[0080] After determining the time window, frequency and priority ranking, the index collection module starts to collect macroeconomic indicator data, including gross domestic product growth rate, consumer price index, purchasing manager index and money supply. The gross domestic product growth rate reflects the overall economic growth trend, the consumer price index reflects the price level change, the purchasing manager index indicates the degree of manufacturing and overall economic prosperity, and the money supply reflects the market liquidity. During the collection process, these index data need to be processed for time and space alignment. In terms of time alignment, data collected at different frequencies are unified to the same time dimension, such as coordinating monthly data with quarterly data to ensure that all index data are consistent in time marking. In terms of space alignment, for the same index data involving different regions or different statistical standards, standardization processing is carried out to eliminate data deviation caused by regional differences or different statistical standards.

[0081] After receiving the collected data and undergoing spatiotemporal alignment, the data preprocessing module first divides the raw data into structured data and unstructured text data. Structured data includes specific values ​​of various indicators, statistical dates, and other data with fixed formats; unstructured text data includes textual information such as economic policy documents, market analysis reports, and news articles. Subsequently, noise reduction parameters and data cleaning rules are initialized. The noise reduction parameters are set according to the data's fluctuation characteristics, while the data cleaning rules specify the standards and methods for data selection.

[0082] For structured data, a moving average method is used to filter out random fluctuations. The moving average method calculates the average value of data within a certain time window, smoothing out short-term, irregular fluctuations while preserving the overall trend of the data. For unstructured text data, semantic similarity matching is used to remove invalid information. Semantic similarity matching calculates the semantic correlation between words and sentences in the text to identify and remove content unrelated to macroeconomic indicators and market trends, such as removing advertising information and irrelevant entertainment news.

[0083] After noise reduction, data integrity verification is performed. This involves checking for missing data at each time point and ensuring the data values ​​are within reasonable ranges to determine the completeness and anomalies. For missing values, interpolation is used to fill in the gaps, calculating reasonable values ​​for the missing locations based on valid data before and after the missing data, following specific mathematical logic. Outlier detection algorithms are used to identify outliers. These algorithms analyze data distribution characteristics, find data points that deviate significantly from the overall distribution, and correct them based on the characteristics of surrounding data to achieve data consistency.

[0084] Based on the dimensional characteristics of macroeconomic indicators, standardized data is performed to ensure consistency. Different macroeconomic indicators have different dimensions; for example, GDP growth rate is expressed as a percentage, while money supply is expressed as a specific amount. These dimensional differences can affect the accuracy of subsequent processing. Max-min normalization maps all data to a preset interval, typically set to [0,1]. Max-min normalization converts the original data into dimensionless standardized values ​​by calculating the relative positions of the data with the maximum and minimum values ​​of the indicator, thus obtaining standardized numbers.

[0085] Example 2: See Figure 3After receiving the standardized dataset output by the data preprocessing module, the feature engineering module first determines the target dimension and the association rule of feature construction according to the overall characteristics of the standardized dataset. The target dimension covers the time characteristics of the data and the association characteristics between different indicators, and the association rule explicitly shows the interaction mode between different features. Initialize the feature pool and generate a set of basic statistical features and time series features as the initial feature candidate set. The basic statistical features include mean, variance, range, and median of the data sequence, which can reflect the trend and dispersion degree of the data set; the time series features include autocorrelation coefficient, partial autocorrelation coefficient, and maximum and minimum values in the sliding window, which can reflect the time-varying law of the data.

[0086] The features in the candidate set are expanded in the time dimension and associated in the cross-section dimension. The time dimension expansion is achieved by calculating the feature values at different time lags, such as calculating the difference or ratio between the current value and the value of the previous month or quarter of a certain indicator, to capture the continuity and change rate of the feature in time. The cross-section dimension association is achieved by analyzing the correlation between different macroeconomic indicator features, such as calculating the correlation between the consumer price index feature and the money supply feature, to mine the collaborative change relationship between different indicators. Calculate the influence weight of the feature on the market trend through the attention mechanism, and the attention mechanism allocates weight values according to the correlation strength of the feature with the market trend change in historical data. The closer the correlation between the feature and the market trend change, the higher the weight value. According to the weight value, filter the key features, and set a weight threshold in the filtering process, retain the features with weight values higher than the threshold, and eliminate the features with weight values lower than the threshold.

[0087] Nonlinear transformation and dimension combination are performed on the filtered key features. Nonlinear transformation includes squaring, square root, logarithmic transformation, exponential transformation, etc. of the features, to capture the nonlinear relationship between the features and the market trend. Dimension combination is achieved by performing arithmetic operations, cross products, etc. on different features to generate new combined features, such as multiplying the purchasing manager index feature and the money supply feature to get a new feature reflecting the combined effect of manufacturing industry sentiment and market liquidity. Through the above processing, a market sensitive feature set containing time domain features, frequency domain features, and association features is generated, the time domain features reflect the dynamic change of the features with time, the frequency domain features reflect the periodic fluctuation characteristics of the features, and the association features reflect the mutual influence between different features, finally forming a high-dimensional feature vector.

[0088] The prediction model construction module obtains a high-dimensional feature vector and a corresponding market historical trend label, which is determined based on historical market data, including historical market price trend, trading volume trend, etc. The parameters of the deep belief network are initialized, including the number of network layers, the number of neurons in each layer, the type of activation function, the initial weight value, the bias term, etc. According to the market historical law, a deep belief network architecture containing an input layer, a hidden layer and an output layer is constructed. The number of neurons in the input layer is consistent with the dimension of the high-dimensional feature vector, which is used to receive the high-dimensional feature vector; the hidden layer contains multiple layers, and the number of neurons in each layer is set according to the complexity of the feature, which transforms and abstracts the input feature through a nonlinear activation function, and gradually extracts higher-level feature rules; the number of neurons in the output layer is set according to the prediction target, which is used to output the market trend prediction result.

[0089] The high-dimensional feature vector is input into the network for layer-by-layer pre-training. The pre-training process is performed layer by layer in the order from the input layer to the hidden layer, and the output of the previous layer is used as the input of the current layer during the training of each layer. The network weights are optimized by the contrastive divergence algorithm, which constructs a Gibbs sampling chain, calculates the difference between the sample data and the reconstructed data, and adjusts the network weights according to the difference to reduce the reconstruction error. On the basis of pre-training, the network is fine-tuned by the backpropagation algorithm, which starts from the output layer, calculates the mean square error between the predicted trend and the actual trend, and propagates the error back to each layer according to the original path. According to the error size, adjust the weights and bias terms of each layer to minimize the mean square error, and get the optimized multi-layer nonlinear prediction model. The real-time high-dimensional feature vector is input into the optimized prediction model, and the model outputs the market trend prediction sequence of the future preset time period through the nonlinear transformation of each layer. The preset time period can be set according to the actual application requirements.

[0090] Embodiment 3: refer to Figure 4 After the dynamic adjustment module receives the trend prediction sequence, the evaluation index of the preset prediction accuracy and the threshold range are received. The evaluation index includes the deviation amplitude of the predicted value and the actual value, the degree of agreement of the trend direction, etc. The threshold range is set according to the historical fluctuation of the market and the prediction demand, covering the allowed error interval under different confidence levels.

[0091] According to the preset evaluation index and the current prediction sequence, the prediction revenue under different thresholds is calculated by the Q-learning algorithm in reinforcement learning to select the optimal decision threshold. The exploration rate, learning rate and discount factor are set, the exploration rate is used to control the probability of trying new threshold in the decision-making process, the learning rate is used to adjust the amplitude of Q value update in each iteration, and the discount factor is used to weigh the proportion of current revenue and future potential revenue. The reward function is constructed according to the deviation of the current prediction sequence and the actual market trend, and the reward value is positively correlated with the prediction accuracy. When the deviation of the prediction result and the actual trend is within the preset range, the reward value is positive; when the deviation exceeds the preset range, the reward value is negative.

[0092] The state-action value function is updated iteratively by the Q-learning algorithm. In each iteration step, the current state is determined based on the size of the prediction error and the consistency of the trend direction, and the action includes increasing the decision threshold, reducing the decision threshold or keeping the decision threshold unchanged. After selecting the corresponding action according to the current state, the instantaneous reward is calculated combined with the reward function, and the Q value of the current state-action pair is updated according to the Q value of the next state, and the update formula is:

[0093]

[0094] wherein, represents the value of executing action in state ; represents the learning rate; represents the instantaneous reward obtained after executing action ; represents the discount factor; represents the new state transferred after executing action ; represents all possible actions that can be executed in new state ; represents the maximum Q value of all possible actions in new state .

[0095] In the iteration process, the Q value is continuously updated until the Q value converges or the preset maximum number of iterations is reached. When the stopping condition is met, the threshold that makes the Q value maximum is selected from all possible decision thresholds as the optimal decision threshold.

[0096] Under the constraint of optimal threshold, the transition probability of market state is described by Markov chain model. Market state is divided into different categories according to the value range of macroeconomic indicators and market trend characteristics, such as prosperity, stability, recession, etc. The transition probability between different states, i.e. the possibility of transferring from the current state to the next state, is calculated based on historical data to form a state transition matrix. The matrix is used to perform multi-step state deduction on the prediction sequence. According to the current market state and transition probability, the market state and corresponding probability distribution that may occur at future time nodes are predicted, and a dynamic adjustment strategy is generated.

[0097] According to the dynamic adjustment strategy, the output results of the prediction model are calibrated in real time. When the deduction result shows that the probability of significant change in future market state is high, the prediction sequence is modified in combination with the optimal decision threshold, so that the modified prediction result is more consistent with the state transition rule. Through the above process, dynamic calibration of prediction accuracy is realized, so that the prediction result can better adapt to market changes.

[0098] In the running of the feedback iteration module, first, the actual market trend data in the prediction period are collected, which cover multiple dimensions, such as index fluctuation of stock market, yield change of bond market, price trend of commodity, industry prosperity index and consumer confidence index, etc. These data come from multiple channels such as public information of stock exchange, statistical report of industry association, analysis result of market research institution, etc., to ensure that the data can fully reflect the actual market situation.

[0099] After the collection is completed, the prediction results are fused with the actual data by Bayesian filtering. Bayesian filtering calculates the posterior probability by combining prior probability with observation data, and constantly updates the cognition of the deviation between prediction results and actual data. In this process, the deviation index is calculated, among which the mean absolute error is the average of the absolute difference between the predicted value and the actual value, which intuitively presents the deviation degree in numerical level; the trend direction accuracy rate is the proportion of the number of times that the predicted trend direction is consistent with the actual trend direction in the total number of prediction times, which reflects the accuracy of the judgment on the overall trend of the market.

[0100] The bias degree index is decomposed into feature contribution degree, model structure bias, and threshold setting error. The calculation of feature contribution degree is based on the weight of each feature generated by the feature engineering module in the prediction process, which analyzes the proportion of the influence of different features on bias, for example, a feature with a high weight in prediction but weak correlation with market trends will contribute more to bias. Model structure bias starts with the network architecture of the prediction model construction module, including the number of neurons in each layer, the selection of activation function, the number of training iterations, and other factors. Unreasonable configuration of these factors will cause the model to deviate from the fitting of market rules. Threshold setting error is related to the decision threshold in the dynamic adjustment module. A threshold that is too high or too low will cause the model to deviate from the response to market changes. By comparing the prediction bias under different thresholds, the influence of threshold setting on overall bias is determined.

[0101] According to the decomposition results, a bias analysis report is generated, which details the specific numerical distribution of feature contribution degree, which parameters in the model structure cause bias, and the deviation range of threshold setting from the optimal value.

[0102] The bias analysis report is fed back to the feature engineering module, the prediction model construction module, and the dynamic adjustment module. After receiving the report, the feature engineering module adjusts the weight calculation method of feature selection according to the feature contribution degree. For features with abnormal contribution degree, re-evaluate their correlation with market trends, adjust their proportion in the feature set, or optimize feature representation through new nonlinear transformation methods. The prediction model construction module adjusts the architecture of the deep belief network according to the analysis results of the model structure bias, such as increasing or decreasing the number of hidden layers, adjusting the number of neurons in each layer, replacing the type of activation function, or modifying the parameters in the pre-training and fine-tuning process, and re-training the model to correct the structural bias. The dynamic adjustment module recalculates the range of decision threshold based on the threshold setting error information, and iterates again through reinforcement learning algorithm to optimize the Q value update process, so that the decision threshold is more consistent with the actual market fluctuations. Through such feedback and adjustment, the prediction strategy is continuously updated, and the system can continuously adapt to the dynamic changes of the market.

[0103] In the runtime of the index early warning module, the key node data in the trend prediction sequence and the preset risk threshold interval are obtained. The key node data includes the trend turning points appearing in the prediction sequence, the extreme points of the index value, and the nodes with the same change direction for a plurality of continuous periods, etc. These nodes can reflect the time when the market trend may change significantly. The preset risk threshold interval is determined based on the historical fluctuation range of macroeconomic indicators, the economic policy control target, and the reasonable interval for market stable operation. The threshold interval of different indicators is set individually according to its own characteristics. For example, the risk threshold interval of the consumer price index is determined in combination with the price stability control target, and the risk threshold interval of the money supply is set in reference to the reasonable level of market liquidity.

[0104] The fluctuation amplitude of the predicted value of the index is monitored by the statistical process control method. The statistical process control method calculates the control limit of the predicted value and tracks the fluctuation of the predicted value within the control limit. When the predicted value exceeds the preset threshold interval, the early warning mechanism is triggered. After the early warning mechanism is started, the early warning level is classified according to the amplitude, duration, and number of indicators involved when the predicted value exceeds the threshold. Different levels correspond to different early warning response mechanisms. The generated text early warning information needs to clearly indicate the index name, current predicted value, threshold range, and change trend of the predicted value. The text expression is simple and clear, and ambiguity expression is avoided. The visual warning mark displays the position relationship between the index predicted value and the threshold interval in the form of a chart. For example, the upper and lower limits of the threshold are marked with different colored lines in the trend chart, and the predicted points exceeding the threshold are marked with special symbols, so that the early warning information is intuitive and easy to understand. The generated early warning information and mark are sent to the system terminal through the built-in push mechanism of the system. The terminal includes a computer client, a mobile device application, etc., to ensure that relevant personnel can receive the early warning content in time.

[0105] When the model explanation module runs, the trend prediction sequence output by the prediction model and the corresponding high-dimensional feature vector are obtained. The contribution of each macroeconomic indicator to the prediction result is calculated through the SHAP value. The calculation of the SHAP value is based on the feature transmission path inside the model and reflects the actual influence of each indicator in the prediction process. After the calculation is completed, each macroeconomic indicator is sorted according to the absolute value of the contribution degree. The sorting result can clearly show which indicators play a major role in the prediction and which indicators have a smaller impact. Combined with historical market data, an influence path diagram is constructed. The influence path diagram represents the correlation between indicators and the action path of indicators on the prediction result through nodes and lines. For example, it shows how the change of money supply affects the consumer price index and then affects the prediction result of market trend. The thickness of the lines in the path diagram can represent the degree of influence. The contribution degree sorting and the influence path diagram are integrated into a natural language report. The report needs to explain the calculation process of the SHAP value in detail, including the algorithm principle and parameter setting used, the meaning of each node and line in the influence path diagram, the specific institution of the data source and the data release time, the theoretical basis of the calculation method, and ensure that the report content is detailed and the logic is clear, so that the reader can understand the formation process of the prediction result and the specific influence of each indicator.

[0106] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0107] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A macroeconomic indicator driven market trend forecasting system, characterized by, Comprise: Index collection module, collect macroeconomic data, and dynamically adapt the data collection mechanism, including GDP growth rate, consumer price index, purchasing managers index and money supply; Data preprocessing module, for receiving multi-source heterogeneous data, processing the data in layers to get standardized data set; Feature engineering module, for constructing market sensitive feature set through spatio-temporal feature fusion according to standardized data set, to get high-dimensional feature vector; Prediction model construction module, according to high-dimensional feature vector and market historical law, through deep belief network to construct multi-layer nonlinear prediction model, and generate trend prediction sequence; Dynamic adjustment module, after receiving the trend prediction sequence, through reinforcement learning algorithm to optimize model decision threshold, and combine Markov chain to plan state transition of prediction process, to realize dynamic calibration of prediction accuracy; Feedback iteration module, through Bayesian filtering to track the deviation degree of prediction results and actual market trend, and feed back the deviation analysis results to feature engineering module, prediction model construction module and dynamic adjustment module, to update prediction strategy.

2. A macroeconomic indicator driven market trend forecasting system according to claim 1, wherein, Collect macroeconomic data, and dynamically adapt the data collection mechanism, including GDP growth rate, consumer price index, purchasing managers index and money supply, including: According to the characteristics of economic cycle fluctuation and index correlation analysis, determine the time window, frequency and priority order of collection; According to the time window, frequency and priority order of collection, collect macroeconomic index data, and process the index data in space-time alignment to get standardized data set, including GDP growth rate, consumer price index, purchasing managers index and money supply.

3. A macroeconomic indicator driven market trend forecasting system according to claim 1, wherein, Receive multi-source heterogeneous data, process the data in layers to get standardized data set, including: Divide the original collected data into structured data and unstructured text data; Initialize noise reduction parameters and data cleaning rules; According to the difference of data type, filter the random fluctuation in structured data by moving average method, and remove invalid information in unstructured text by semantic similarity matching; According to the data integrity check result, fill in the missing value by interpolation method, identify abnormal data by outlier detection algorithm and correct it, to get consistent data; According to the dimension characteristics of macroeconomic indicators, map the data to the preset interval by maximum and minimum normalization method, to get standardized data set.

4. The macroeconomic indicator driven market trend forecasting system of claim 1, wherein, According to standardized data set, construct market sensitive feature set through spatio-temporal feature fusion, to get high-dimensional feature vector, including: According to the standardized data set, determine the target dimension and association rule of feature construction; Initialize feature pool, generate a set of basic statistical features and time series features as initial feature candidate set; Expand the features in the candidate set in time dimension and correlate in cross section dimension, calculate the influence weight of the features on market trend by attention mechanism, and select key features according to the weight value; Nonlinear transformation and dimension combination are carried out on the selected key features to generate market sensitive feature set containing time domain, frequency domain and association features, to get high-dimensional feature vector.

5. A macroeconomic indicator driven market trend forecasting system according to claim 1, wherein, According to the high-dimensional feature vector and market historical law, a multi-layer nonlinear prediction model is constructed by a deep belief network, and a trend prediction sequence is generated, including: Obtain high-dimensional feature vectors and corresponding market historical trend labels; Initialize deep belief network parameters; According to the market historical law, a deep belief network architecture containing input layer, hidden layer and output layer is constructed, the high-dimensional feature vector is input into the network for layer-by-layer pre-training, and the network weight is optimized by contrast divergence algorithm; On the basis of pre-training, the network is fine-tuned by back propagation algorithm to minimize the mean square error of predicted trend and actual trend, so as to obtain the optimized multi-layer nonlinear prediction model; Input real-time high-dimensional feature vector into the optimized prediction model to generate market trend prediction sequence in future preset time period.

6. A macroeconomic indicator driven market trend forecasting system according to claim 1, wherein, After receiving the trend prediction sequence, the model decision threshold is optimized by reinforcement learning algorithm, and the state transition planning of prediction process is carried out combined with Markov chain to realize dynamic calibration of prediction accuracy, including: Receive the trend prediction sequence, and preset the evaluation index and threshold range of prediction accuracy; According to the preset evaluation index and the current prediction sequence, the prediction income under different threshold values is calculated by Q-learning algorithm in reinforcement learning, and the optimal decision threshold is selected; Under the constraint of the optimal threshold, the state evolution of the prediction sequence is carried out by Markov chain model to describe the transition probability of market state, and the dynamic adjustment strategy is generated; According to the dynamic adjustment strategy, the output result of the prediction model is calibrated in real time to realize the dynamic optimization of prediction accuracy.

7. A macroeconomic indicator driven market trend forecasting system according to claim 1, wherein, According to the preset evaluation index and the current prediction sequence, the prediction income under different threshold values is calculated by Q-learning algorithm in reinforcement learning, and the optimal decision threshold is selected, including: Set relevant parameters, including exploration rate, learning rate and discount factor; According to the deviation of current prediction sequence and actual market trend, a reward function is constructed, and the reward value is positively correlated with the prediction accuracy; The state-action value function is updated by Q-learning algorithm, in each iteration step, the threshold adjustment action is selected according to the current state, and the Q value is updated according to the reward function; When the Q value converges or reaches the maximum iteration times, stop the learning process, select the threshold with the maximum Q value as the optimal decision threshold.

8. A macroeconomic indicator driven market trend forecasting system according to claim 1, wherein, The deviation degree of prediction result and actual market trend is tracked by Bayesian filtering, and the deviation analysis result is fed back to the feature engineering module, prediction model construction module and dynamic adjustment module to update the prediction strategy, including: Collect the actual market trend data in the prediction period; Fuse the prediction result and the actual data by Bayesian filtering to calculate the deviation degree index, including mean absolute error and trend direction accuracy; Decompose the deviation degree index into feature contribution, model structure deviation and threshold setting error, and generate a deviation analysis report; The deviation analysis report is fed back to the feature engineering module, prediction model construction module and dynamic adjustment module to trigger feature weight update, model parameter optimization or threshold recalibration to update the prediction strategy.

9. The macroeconomic indicator driven market trend forecasting system of claim 1, wherein, The index early warning module is configured to generate a macroeconomic index abnormal fluctuation early warning signal according to the trend prediction sequence and a preset risk threshold, including: Obtaining key node data in the trend prediction sequence and a preset risk threshold interval; Monitoring the fluctuation amplitude of the index prediction value through a statistical process control method, and triggering an early warning mechanism when the prediction value exceeds the threshold interval; Generating text early warning information and visual warning signs according to early warning level classification, and pushing to a system terminal.

10. The macroeconomic indicator driven market trend forecasting system of claim 1, wherein, The model explanation module is configured to analyze key influencing factors of the prediction result and generate an explainability report, including: Obtaining a trend prediction sequence output by the prediction model and a corresponding high-dimensional feature vector; Calculating the contribution degree of each macroeconomic index to the prediction result through SHAP value, and generating feature importance ranking; Combining market historical data to construct an influence path diagram, and showing how key indicators affect the final prediction result through feature correlation; Integrating the contribution degree ranking and the influence path diagram into a natural language report, and marking the data source and calculation method.

Images