Financial time series data multi-scale feature analysis and prediction method and system

Abstract

The invention discloses a financial time series data multi-scale feature analysis and prediction method and system, and belongs to the technical field of financial data processing and deep learning, and the method comprises the steps: obtaining and preprocessing multi-source financial time series data; performing multi-scale decomposition by adopting discrete wavelet transform and empirical mode decomposition to generate a trend component, a periodic component and a noise component; respectively extracting time dependence characteristics of each component through a time sequence Transform coding module; multi-scale features are subjected to adaptive weighted fusion through a multi-scale attention fusion module; outputting a predicted value and a confidence interval through a probability prediction module; and the closed-loop optimization module adjusts decomposition parameters and attention weights according to prediction error feedback, so that different frequency components of financial time series data can be effectively separated, a multilevel time dependency relationship is captured, and adaptive feature fusion and online optimization are realized.

Description

Multi-scale Feature Analysis and Prediction Methods and Systems for Financial Time Series Data Technical Field

[0001] This invention relates to the fields of financial data processing and deep learning technology, specifically to a method and system for multi-scale feature analysis and prediction of financial time series data. Background Technology

[0002] Financial market time-series data is characterized by high noise, non-stationarity, and complex multi-scale cyclical features, making it difficult for traditional forecasting methods to effectively capture long-term dependencies and multi-scale evolutionary patterns. With the rapid development of deep learning technology, neural network-based financial time-series forecasting methods have been widely used, but existing methods still have shortcomings in handling the multi-scale characteristics of financial data.

[0003] Chinese invention patent application CN120316474A discloses a method for predicting financial time series data. This method obtains financial data as input data, uses a data augmentation module to enhance the input data to obtain time series data, and constructs a core network containing feature cores and time cores. The feature cores are used to learn the relationship between time series data and time series, and the time cores are used to learn the relationship between before and after time series. The core network is trained based on financial data to obtain an initial model. The parameters of the initial model are adjusted based on the amount of input data and temperature parameters to obtain a financial time series data prediction model. While this method enhances the model's information extraction capability to some extent through gating and skip connections, it suffers from the following technical problems: First, both the feature core and time core are built on fully connected layers, making it difficult to effectively capture the evolutionary patterns of different frequency components in financial time series data. Its processing capability for financial data with multi-level time dependencies, including minute, hourly, and daily timeframes, is limited. Second, this method does not perform multi-scale decomposition of the original financial time series, failing to separate and model trend, periodic, and noise components separately, leading to mutual interference between different frequency components and affecting prediction accuracy. Third, this method lacks an adaptive multi-scale feature fusion mechanism, making it unable to dynamically adjust the importance weights of features at different time scales based on data characteristics. Fourth, the model parameter adjustments are based solely on data volume and temperature parameters, lacking a closed-loop optimization mechanism based on prediction error feedback, making it difficult to achieve continuous optimization of decomposition parameters and fusion weights.

[0004] In recent years, wavelet transform and empirical mode decomposition (EMD), as effective multi-scale signal decomposition methods, have been applied to the field of financial time series analysis. Wavelet transform can decompose signals into different frequency levels, capturing local and global features; EMD can adaptively decompose nonlinear and non-stationary signals into multiple intrinsic mode functions. Meanwhile, the Transformer architecture, with its powerful long-range dependency modeling capabilities, has demonstrated excellent performance in time series prediction. However, current technologies have not yet effectively integrated wavelet decomposition, EMD, and the time series Transformer architecture to form a complete solution for multi-scale financial time series modeling.

[0005] Therefore, there is an urgent need for a financial time series prediction method and system that can perform multi-scale decomposition of financial time series data, capture the time dependence of different frequency components, adaptively fuse multi-scale features, and continuously optimize through a closed-loop mechanism. Summary of the Invention

[0006] To address the technical problems in existing technologies, such as insufficient multi-scale feature extraction of financial time series data, difficulty in capturing long-term dependencies, lack of adaptive feature fusion mechanisms, and closed-loop optimization capabilities, this invention provides a method and system for multi-scale feature analysis and prediction of financial time series data.

[0007] The method for multi-scale feature analysis and prediction of financial time series data provided by this invention includes: a multi-source data acquisition step, which acquires multi-source financial time series data including high-frequency market data, daily trading data, macroeconomic indicator data, market sentiment indicator data, and event calendar data; preprocessing the multi-source financial time series data by time alignment and missing value imputation to generate standardized financial time series data; a multi-scale decomposition step, in which the standardized financial time series data is input into a multi-scale decomposition module, which uses discrete wavelet transform to perform multi-level decomposition of the standardized financial time series data to obtain wavelet coefficients at different frequency levels, and simultaneously uses empirical mode decomposition to adaptively decompose the standardized financial time series data to obtain multiple intrinsic mode functions; based on the frequency characteristics of wavelet coefficients and intrinsic mode functions, frequency band matching and component fusion are performed to generate trend components, periodic components, and noise components; and a time series feature encoding step, in which the trend components, periodic components, and noise components are respectively input into a time series Transformer encoding module, which adds cross-scale position encoding information to each component and uses a multi-level self-attention machine. The system captures the temporal dependencies within each component and outputs trend feature vectors, periodic feature vectors, and noise feature vectors respectively. In the multi-scale feature fusion step, these vectors are input into a multi-scale attention fusion module. This module calculates adaptive weights for features at different time scales using a cross-scale attention mechanism and then weights and fuses them to generate a multi-scale fused feature vector. In the probabilistic prediction step, the multi-scale fused feature vector is input into a probability prediction module, which generates predicted distribution parameters for future multi-step time series based on the vector, outputting predicted values ​​and corresponding confidence intervals. In the closed-loop optimization step, the predicted values ​​are compared with actual observations to calculate the prediction error. This error is input into a closed-loop optimization module, which generates decomposition parameter update amounts and attention weight update amounts based on the prediction error. The decomposition parameter update amounts are fed back to the multi-scale decomposition module to adjust the decomposition parameters, and the attention weight update amounts are fed back to the multi-scale attention fusion module to update the adaptive weights.

[0008] Preferably, the discrete wavelet transform has 3 to 6 decomposition layers, and the empirical mode decomposition generates 4 to 8 intrinsic mode functions.

[0009] Preferably, the number of encoder layers in the temporal Transformer encoding module is 4 to 8, and the number of attention heads in the multi-head self-attention mechanism is 4 to 16.

[0010] Preferably, the prediction step size is 1 to 30 time units, and the confidence level of the confidence interval is 90% to 99%.

[0011] This invention also provides a multi-scale feature analysis and prediction system for financial time series data, comprising: a multi-source data acquisition module for acquiring multi-source financial time series data and preprocessing it to generate standardized financial time series data; a multi-scale decomposition module for performing multi-scale decomposition on the standardized financial time series data using discrete wavelet transform and empirical mode decomposition to generate trend components, periodic components, and noise components; a time series Transformer encoding module for performing feature encoding on the trend components, periodic components, and noise components respectively, and outputting trend feature vectors, periodic feature vectors, and noise feature vectors; a multi-scale attention fusion module for calculating adaptive weights through a cross-scale attention mechanism and performing weighted fusion to generate a multi-scale fused feature vector; a probability prediction module for outputting predicted values ​​and confidence intervals based on the multi-scale fused feature vector; and a closed-loop optimization module for adjusting the decomposition parameters and attention weights according to the prediction error feedback.

[0012] The beneficial effects of this invention are as follows: First, by integrating discrete wavelet transform and empirical mode decomposition into a multi-scale decomposition method, this invention can effectively separate trend, periodic, and noise components in financial time series data, making the modeling of different frequency components independent of each other, avoiding mutual interference, and improving the accuracy and robustness of feature extraction. Second, by employing a temporal Transformer architecture to encode components at each scale separately, this invention can effectively capture multi-level time dependencies at the minute, hour, and day levels in financial time series data, overcoming the limitations of traditional fully connected networks in long-range dependency modeling. Third, the multi-scale attention fusion module designed in this invention can adaptively learn the importance weights of features at different time scales according to data characteristics, achieving dynamic fusion at the feature level and improving the adaptability of the prediction model. Fourth, the closed-loop optimization mechanism established in this invention can continuously adjust the decomposition parameters and attention weights based on prediction error feedback, achieving online learning and continuous optimization of the model, and improving the stability of prediction accuracy. Fifth, the probability prediction results output by this invention include predicted values ​​and confidence intervals, providing a quantitative risk reference for financial decision-making. Experimental results show that the prediction error of this invention on standard financial datasets is reduced by more than 35% compared to the traditional ARIMA method, and the prediction direction accuracy reaches more than 58%. Attached Figure Description

[0013] Figure 1 is a flowchart of the multi-scale feature analysis and prediction method for financial time series data of the present invention.

[0014] Figure 2 is an architecture diagram of the financial time series data multi-scale feature analysis and prediction system of the present invention. Detailed Implementation

[0015] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The method and system for multi-scale feature analysis and prediction of financial time series data provided by the present invention aim to solve the technical problems of high noise in financial time series data, complex multi-scale periodic characteristics, and the difficulty of traditional prediction methods in capturing long-term dependencies.

[0016] As shown in Figure 1, the financial time series data multi-scale feature analysis and prediction method of the present invention includes six core steps: multi-source data acquisition step, multi-scale decomposition step, time series feature encoding step, multi-scale feature fusion step, probability prediction step, and closed-loop optimization step. Each step forms a deeply coupled closed-loop collaborative relationship.

[0017] Step S1, multi-source data acquisition steps.

[0018] The multi-source data acquisition step is the initial stage of the entire method, responsible for collecting financial time-series data from multiple data sources and performing preprocessing. In one embodiment of the present invention, the multi-source financial time-series data includes five types of data: high-frequency market data, daily trading data, macroeconomic indicator data, market sentiment indicator data, and event calendar data.

[0019] High-frequency market data includes minute-level opening price, highest price, lowest price, closing price, and trading volume data, with a sampling frequency of 1 minute, covering the entire trading session of the trading day. In this embodiment, the high-frequency market data is obtained from the real-time market data interface of the stock exchange, and the data format is a structured record containing a timestamp and five price and volume fields. Daily trading data includes daily opening and closing prices, highest and lowest prices, trading volume, turnover, turnover rate, and other indicators, extracted from historical market data databases. Macroeconomic indicator data includes indicators such as GDP growth rate, consumer price index, purchasing managers' index, and money supply, with a data frequency of monthly or quarterly, aligned to the daily frequency using interpolation methods. Market sentiment indicator data includes market sentiment scores calculated based on news text and social media data, obtained by performing sentiment analysis on financial-related texts using a natural language processing model. Event calendar data includes discrete event markers such as earnings release dates, dividend ex-rights dates, and major policy release dates, represented using one-hot encoding.

[0020] Time alignment is a crucial step in multi-source data preprocessing. Because different data sources have varying sampling frequencies and time bases, all data needs to be aligned to a unified time axis. In this embodiment, minute-level timestamps are used as the base time axis. High-frequency market data is directly mapped to the base time point, daily data is expanded to each minute time point through repeated padding, and monthly and quarterly data are mapped to minute-level time points through linear interpolation or forward padding.

[0021] Missing value imputation addresses potential data gaps that may occur during data acquisition. For missing values ​​in high-frequency market data, time-weighted linear interpolation is used for imputation; for missing values ​​in daily data and macroeconomic indicator data, forward imputation is used to replace them with the nearest valid value. Furthermore, outliers are identified and corrected using a standard deviation-based threshold detection method.

[0022] Standardization transforms data with different dimensions and numerical ranges into a unified numerical space. In this embodiment, the Z-score standardization method is used to standardize numerical features. For financial time-series data, the standardized parameters are calculated based on statistics within a sliding window, with the window length set to the data from the past 60 trading days to accommodate the non-stationary nature of financial data. The standardized data has a mean of 0 and a standard deviation of 1, facilitating the training of subsequent neural network models.

[0023] After preprocessing, the generated standardized financial time series data has dimensions of (N, T, F), where N is the number of samples, T is the time step length, and F is the number of features. In this embodiment, the time step length T is set to 240, corresponding to 4 hours of minute-level data; the number of features F is determined based on the number of data sources included, including 5 price-volume features, 8 daily frequency indicator features, 6 macroeconomic features, 1 sentiment rating feature, and several event labeling features.

[0024] Step S2, multi-scale decomposition step.

[0025] The multi-scale decomposition step receives standardized financial time-series data and uses a combination of discrete wavelet transform and empirical mode decomposition to decompose the original time series into three frequency-level components: trend component, periodic component, and noise component. This step is one of the core innovations of this invention, as it fully extracts the multi-scale features of financial time-series data by integrating two complementary decomposition methods.

[0026] Discrete wavelet transform is used to perform multi-level decomposition of standardized financial time series data. In this embodiment, the Daubechies-4 wavelet is used as the mother wavelet function. This wavelet has good time-frequency localization characteristics and is suitable for handling abrupt changes and fluctuations in financial time series data. The number of decomposition levels is set to 5, and each level produces an approximate coefficient sequence and a detail coefficient sequence. Let the original time series be... , No. Approximate coefficients after layer decomposition and detail coefficient Calculated using the following adaptive wavelet decomposition algorithm:

[0027] ,

[0028] ,

[0029] in, For the first The first layer of decomposition Approximation coefficients, For the first The first layer of decomposition A detailed coefficient, For the first The approximate coefficient sequence of the layer, These are the coefficients of the low-pass filter. These are the coefficients of the high-pass filter. For the original sequence index, This is the coefficient index after downsampling. For the first level decomposition, That is, the original time series .

[0030] Empirical Mode Decomposition (EMD) simultaneously performs adaptive decomposition on standardized financial time-series data, yielding multiple intrinsic mode functions (EMFs). Through an iterative sieving process, EMD decomposes the signal into a series of EMFs arranged from high to low frequency, along with a residual trend term. In this embodiment, EMD generates six EMFs, denoted as... and a residual term The extraction process of intrinsic mode functions follows a standard screening algorithm, with the stopping condition set to a standard deviation of less than 0.05 between two consecutive screening results.

[0031] Frequency band matching and component fusion are the core steps of the multi-scale trend-period-noise separation algorithm proposed in this invention. This algorithm fuses wavelet coefficients and intrinsic mode functions (IMFs) into components at three frequency levels based on their frequency characteristics. Specifically, it first calculates the center frequency of each wavelet detail coefficient sequence and IMF, and then divides the components according to a preset frequency threshold. Let the first preset frequency threshold be... The second preset frequency threshold is ,in The ingredient classification rules are as follows:

[0032] Trend Components From frequency lower The components include the highest-level approximation coefficients of wavelet decomposition. and the residual terms of empirical mode decomposition The fusion method is a weighted average:

[0033] ,

[0034] in, For trend components at time The value, The approximation coefficients of the 5th layer are reconstructed by upsampling at time 10. The value, For the empirical mode decomposition residual term at time... The value, and The adaptive weighting coefficients are determined based on the correlation between the two components and the original signal, satisfying the following conditions: .

[0035] Periodic components From frequency between and The components consist of, including the intermediate layer detail coefficients of wavelet decomposition. , Mid-frequency eigenmode functions of empirical mode decomposition , The fusion method is as follows:

[0036] ,

[0037] in, For the periodic component at time t The value, For the first After upsampling and reconstruction, the layer detail coefficients are at time [time value missing]. The value, For the first The intrinsic mode functions at time... The value, and The adaptive weighting coefficients are determined based on the energy proportion of each component.

[0038] noise components From frequency higher The components include the low-level detail coefficients of wavelet decomposition. , High-frequency intrinsic mode functions of empirical mode decomposition , The fusion method is as follows:

[0039] ,

[0040] in, For the noise component at time... The value, and These are adaptive weighting coefficients.

[0041] In this embodiment, the first preset frequency threshold Set to 0.02Hz, corresponding to a period of approximately 50 seconds; second preset frequency threshold. The frequency is set to 0.1Hz, corresponding to a period of approximately 10 seconds. These thresholds are determined based on the typical fluctuation cycles of financial markets and can be adjusted according to specific application scenarios.

[0042] Step S3, temporal feature encoding step.

[0043] The temporal feature encoding step receives the trend component, periodic component, and noise component output from the multi-scale decomposition step, and extracts features from each component using the temporal Transformer encoding module to capture the temporal dependencies within each component. The temporal Transformer encoding module is designed to fully consider the characteristics of financial time series data, employing cross-scale positional encoding and a multi-layer self-attention mechanism.

[0044] The temporal Transformer encoding module comprises four components: an input embedding layer, a cross-scale positional encoding layer, a multi-layer encoder stack, and an output projection layer. The input embedding layer is responsible for mapping the temporal components of the input to a high-dimensional feature space. In this embodiment, the embedding dimension is set to 256, implemented through a one-dimensional convolution operation with a kernel size of 3 and a stride of 1, while maintaining the temporal dimension unchanged.

[0045] The cross-scale positional coding layer is an innovative design proposed in this invention, used to add positional codes containing scale information to temporal components at different scales. Traditional Transformer positional coding only contains absolute positional information, while the cross-scale positional coding of this invention includes both absolute position and scale identifier information. The cross-scale positional coding algorithm is defined as follows:

[0046] ,

[0047] ,

[0048] in, For position In dimensions The location code value at that location, For position In dimensions The location code value at that location, For model dimensions, For the scale phase offset, the trend component... For periodic components For noise components By using different phase shifts, the encoder is able to distinguish inputs from different scales.

[0049] The multi-layer encoder stack is the core component of the temporal Transformer encoding module, consisting of multiple encoder layers with the same structure connected in series. In this embodiment, the number of encoder layers is set to 6. Each encoder layer contains two sub-layers: a multi-head self-attention sub-layer and a feedforward neural network sub-layer, with each sub-layer followed by layer normalization and residual connections.

[0050] The multi-head self-attention sublayer captures the dependencies between different positions in a sequence through multiple parallel attention heads. In this embodiment, the number of attention heads is set to 8, and each head has a dimension of 32. The attention calculation process is as follows: Let the embedding representation of the input sequence be... First, the query matrix is ​​obtained through linear transformation. Key matrix Sum matrix :

[0051] ,

[0052] in, The weight matrix is ​​a learnable matrix. Let's define the dimension of a single attention head. Then, compute the scaled dot product attention:

[0053] ,

[0054] Multi-head attention concatenates the outputs of multiple attention heads and then performs a linear transformation:

[0055] ,

[0056] in, For the first The output of each attention head, To focus on the number of heads, This is for outputting the projection matrix.

[0057] The feedforward neural network sublayer performs a non-linear transformation on the features at each location, containing two linear layers and a GELU activation function. In this embodiment, the middle dimension of the feedforward network is set to 1024. The computation process of the feedforward network is as follows:

[0058] ,

[0059] in, and This is the weight matrix. and For bias vectors, This represents the intermediate dimension of the feedforward network.

[0060] The output projection layer projects the sequence features output by the encoder onto a feature vector of fixed dimensions. In this embodiment, global average pooling is used to aggregate the temporal dimension, and then a linear layer is used to project it onto a 256-dimensional feature vector.

[0061] The trend component, periodic component, and noise component are processed by independent temporal Transformer encoding modules. The three encoding modules share the same network structure but have independent parameters, and each outputs a trend feature vector. Periodic eigenvectors and noise feature vector Each feature vector has a dimension of 256.

[0062] Step S4, multi-scale feature fusion step.

[0063] The multi-scale feature fusion step receives the trend feature vector, periodic feature vector, and noise feature vector output by the temporal feature encoding step, and performs adaptive weighted fusion through the multi-scale attention fusion module to generate a multi-scale fused feature vector.

[0064] The core of the multi-scale attention fusion module is an adaptive multi-scale attention weight calculation algorithm, which can dynamically adjust the importance weights of features at different scales based on the characteristics of the input data. Unlike simple concatenation or averaging fusion methods, the fusion method of this invention can learn the correlation between scales and assign weights accordingly.

[0065] The computation process of the cross-scale attention mechanism is as follows: First, the three feature vectors are stacked to form a scale feature matrix. ,in The feature dimension is used. Then, a query matrix is ​​generated through a linear transformation. Key matrix Sum matrix :

[0066]

[0067] in, This is a learnable weight matrix. Cross-scale attention scores are calculated using scaled dot products:

[0068] ,

[0069] in, This is the inter-scale correlation score matrix. The adaptive weights are obtained by softmax normalizing the correlation scores:

[0070] ,

[0071] in, Each row represents the weight allocation for each scale during feature fusion. The fused features are calculated using a weighted sum:

[0072] ,

[0073] To further enhance the fusion effect, this invention adds a gating mechanism after weighted summation. The gating coefficient is calculated in the following way:

[0074] ,

[0075] in, It is the sigmoid activation function. and Here are the gating parameters. The final multi-scale fused feature vector is:

[0076] ,

[0077] in, This represents element-wise multiplication. The gating mechanism enables the model to dynamically balance between adaptive weighted fusion and simple average fusion.

[0078] In this embodiment, multi-scale fusion feature vectors The dimension is 256, the same as the input feature vector dimension. The multi-scale attention fusion module also includes a residual connection and layer normalization operations to improve training stability.

[0079] Step S5, probability prediction step.

[0080] The probabilistic prediction step receives a multi-scale fused feature vector, generates prediction distribution parameters for future multi-step time series through the probabilistic prediction module, and outputs the predicted values ​​and corresponding confidence intervals. Unlike traditional point prediction methods, this invention employs a probabilistic prediction framework, which can quantify the uncertainty of the prediction.

[0081] The probability prediction module comprises three parts: a feature transformation layer, a distribution parameter generation layer, and a sampling output layer. The feature transformation layer performs a nonlinear transformation on the multi-scale fused feature vector, expanding the expressive power of the features. In this embodiment, the feature transformation layer consists of two fully connected layers with an intermediate dimension of 512 and the activation function being GELU.

[0082] The distribution parameter generation layer is responsible for outputting the mean and variance parameters of the predicted distribution. This invention assumes the predicted distribution follows a Gaussian distribution; therefore, it is necessary to generate the mean. and variance Two parameters. The probability prediction distribution generation algorithm is defined as follows:

[0083] ,

[0084] ,

[0085] in, The output of the feature transformation layer, This is the weight matrix. This is the bias vector. The variance parameter is generated using the logarithmic space to ensure non-negativity.

[0086] For multi-step prediction, the distribution parameter generation layer outputs the mean and variance parameters for each prediction step. In this embodiment, the prediction step size is set to 10, meaning the values ​​are predicted for the next 10 time units. Parameter generation for multi-step prediction is achieved through an autoregressive approach, with each prediction step based on the result of the previous step.

[0087] The sampling output layer generates predicted values ​​and confidence intervals based on the mean and variance parameters. The predicted value is taken from the mean parameter:

[0088] ,

[0089] The confidence interval is calculated based on the quantiles of the Gaussian distribution. For the confidence level... Confidence interval:

[0090] ,

[0091] in, For a moment The predicted value, For a moment The confidence interval, For standard normal distribution Upper quantile. In this embodiment, the confidence level is set to 95%, corresponding to... .

[0092] The probability prediction module is trained using the negative log-likelihood loss function:

[0093] ,

[0094] in, For negative log-likelihood loss, To predict the step size, For a moment The true value, For a moment The predicted mean, For a moment The prediction variance is calculated. This loss function simultaneously optimizes the accuracy of the predicted mean and the reasonableness of the prediction variance.

[0095] Step S6, closed-loop optimization step.

[0096] The closed-loop optimization step is another core innovation of this invention. By establishing a feedback path from prediction error to decomposition parameters and attention weights, the model can achieve online learning and continuous optimization. The closed-loop optimization module receives the comparison results between predicted values ​​and actual observations, generates update amounts for decomposition parameters and attention weights, and feeds them back to the multi-scale decomposition module and the multi-scale attention fusion module, respectively.

[0097] The prediction error is calculated using a combination of mean square error and direction error:

[0098] ,

[0099] in, The first term represents the mean square error, and the second term represents the direction error. For balance coefficient, This is the indicator function. The introduction of the direction error term makes the model more focused on the accuracy of the predicted direction, which is particularly important for financial trading decisions.

[0100] The closed-loop decomposition parameter optimization algorithm calculates the update amount of the decomposition parameters based on the prediction error. The decomposition parameters include the fusion weights from wavelet decomposition. , , and frequency threshold , The update amount is calculated based on the gradient information of the prediction error with respect to the decomposition parameters:

[0101] ,

[0102] in, To decompose the parameter update amount, Represents the decomposition parameter set, To decompose the learning rate of the parameters, The gradient of the prediction error with respect to the decomposition parameters is calculated using the backpropagation algorithm. Due to the differentiability of the decomposition operation, the gradient can be efficiently propagated to the decomposition parameters.

[0103] In this embodiment, the decomposition parameter learning rate The value is set to 0.001, employing a momentum gradient descent strategy, with the momentum coefficient set to 0.9. The update process also introduces parameter constraints to ensure that the fused weights remain non-negative and normalized, and that the frequency threshold remains within a reasonable range.

[0104] The calculation of attention weight updates is similarly based on the gradient of the prediction error with respect to the attention parameters:

[0105] ,

[0106] in, For attention weight update amount, This represents the parameters in the multi-scale attention fusion module. The learning rate is the attention parameter. In this embodiment, Set to 0.0005, slightly lower than the decomposition parameter learning rate, to maintain the stability of the attention mechanism.

[0107] The closed-loop optimization is performed using an online learning model, updating the parameters after each new batch of observation data is received. To prevent excessively drastic parameter updates from causing model instability, this invention introduces an update magnitude limiting mechanism:

[0108] ,

[0109] in, The maximum update magnitude is set to 0.01 in this embodiment. Furthermore, the closed-loop optimization also employs an exponential moving average strategy to smooth parameter updates:

[0110] ,

[0111] in, The coefficient for the exponential moving average is set to 0.99 in this embodiment.

[0112] The closed-loop optimization mechanism enables the method of this invention to adapt to the dynamic changes in financial markets. When market characteristics change, the decomposition parameters and attention weights can be automatically adjusted to adapt to the new data distribution without the need for manual intervention to retrain the model.

[0113] The model training method of this invention adopts an end-to-end approach, and the training process includes five stages: data preparation, model initialization, forward propagation, loss calculation, and backpropagation.

[0114] In the data preparation phase, the raw financial time-series data is first preprocessed, including missing value imputation, outlier correction, and standardization. Then, the data is divided into training, validation, and test sets in a ratio of 7:1.5:1.5. The training set is used for model parameter learning, the validation set for hyperparameter tuning and early stopping detection, and the test set for final performance evaluation. To avoid data leakage, the data is strictly divided chronologically, ensuring that the timestamps of the validation and test sets are later than those of the training set.

[0115] During model initialization, the parameters of each module are initialized. The linear layer parameters in the temporal Transformer encoding module use the Xavier uniform initialization method to ensure stable signal variance during forward and backward propagation. The query, key, and value projection matrices in the multi-head self-attention mechanism use orthogonal initialization, which helps accelerate model convergence. The fusion weight parameters in the multi-scale decomposition module are initialized with a uniform distribution, ensuring that the contributions of each component are equal in the initial state. The learning rate parameter in the closed-loop optimization module is initialized with a small value to avoid overly drastic parameter updates in the early stages of training.

[0116] During the forward propagation phase, the input data sequentially passes through the multi-source data acquisition module, the multi-scale decomposition module, the temporal Transformer encoding module, the multi-scale attention fusion module, and the probability prediction module to generate predicted distribution parameters. During forward propagation, the intermediate outputs of each module are saved for subsequent backpropagation calculations. To improve training efficiency, a batch processing method is adopted, with a batch size set to 64.

[0117] In the loss calculation phase, a composite loss function is used to comprehensively measure prediction performance. The composite loss function consists of three parts: negative log-likelihood loss measures the accuracy of probability prediction, mean squared error loss measures the bias in point prediction, and direction loss measures the correctness of the prediction direction. The expression for the composite loss function is:

[0118] ,

[0119] in, This is the composite loss value. For negative log-likelihood loss, For mean square error loss, For directional loss, and This is the balance coefficient. In this embodiment, Set to 0.5. Set it to 0.3.

[0120] During the backpropagation phase, the gradient of the loss function with respect to each parameter is calculated using an automatic differentiation mechanism, and then the model parameters are updated using the Adam optimizer. The hyperparameters of the Adam optimizer are set as follows: learning rate 0.001, first-order momentum coefficient 0.9, second-order momentum coefficient 0.999, and numerical stability term 1e-8. To prevent gradient explosion, a gradient pruning strategy is adopted, with a pruning threshold set to 1.0.

[0121] The model training employs an early stopping strategy to prevent overfitting. Training is stopped and the model parameters with the lowest validation set loss are saved when the validation set loss does not decrease for 10 consecutive training epochs. Furthermore, a learning rate decay strategy is used, multiplying the learning rate by 0.5 every 20 training epochs. The total training epoch limit is set to 200 epochs. In the experimental environment of this embodiment, the model typically converges within 60 to 80 training epochs.

[0122] To verify the effectiveness of the method of this invention, comparative experiments were conducted on standard financial datasets. The experimental datasets included minute-level market data for the CSI 300 Index, S&P 500 Index, and Bitcoin prices, spanning from January 2020 to December 2024. The experimental environment consisted of a server configured with NVIDIA A100 GPUs, and the deep learning framework used was PyTorch 2.0.

[0123] The evaluation metrics include four indicators: root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAS%), and directional accuracy. RMSE measures the deviation between the predicted and actual values; MAE measures the absolute magnitude of the prediction error; MAS%) measures the relative error; and directional accuracy measures the accuracy of the predicted direction of rise or fall. Lower RMSE and MAE indicate higher prediction accuracy, while higher directional accuracy indicates more accurate predictions of the direction of rise or fall.

[0124] The comparison methods include the traditional ARIMA model, LSTM network, original Transformer model, and the core network method in CN120316474A. Experimental results show that the root mean square error (RMSE) of the proposed method on the CSI 300 index dataset is 0.0142, a reduction of 34.9% compared to ARIMA's 0.0218, a reduction of 20.2% compared to LSTM's 0.0178, a reduction of 13.9% compared to the original Transformer's 0.0165, and a reduction of 10.7% compared to CN120316474A's 0.0159. In terms of direction accuracy, the proposed method achieves 58.3%, an improvement of 6.2 percentage points compared to ARIMA's 52.1%. Similar performance improvements were also observed on the S&P 500 index and Bitcoin datasets.

[0125] Ablation experiments verified the contributions of each module. Removing the multi-scale decomposition module increased the root mean square error by 18.5%; removing the cross-scale location encoding increased the root mean square error by 7.2%; removing the multi-scale attention fusion module increased the root mean square error by 12.3%; and removing the closed-loop optimization module increased the root mean square error by 9.8%. These results indicate that each module of the present invention makes a significant contribution to the overall performance.

[0126] Furthermore, a specific analysis was conducted on the effectiveness of the closed-loop optimization module. After 30 days of online operation, compared with the model without closed-loop optimization, the model with closed-loop optimization further reduced its root mean square error on the test set by 5.3%, indicating that the closed-loop optimization mechanism can effectively adapt to dynamic changes in the market environment. The performance advantage of the closed-loop optimized model was particularly evident during periods of significant market volatility.

[0127] Performance analysis of different prediction step lengths shows that the method of this invention has advantages in both short-term and medium-term prediction. For 1-step prediction, the directional accuracy reaches 61.2%; for 5-step prediction, the directional accuracy is 58.7%; and for 10-step prediction, the directional accuracy is 56.1%. As the prediction step length increases, the prediction difficulty increases and the performance decreases somewhat, but it is still better than the comparative methods.

[0128] Stability analysis under different market environments shows that the method of this invention maintains stable predictive performance in bull markets, bear markets, and sideways markets. During bull markets, the root mean square error (RMSE) is 0.0138; during bear markets, it is 0.0151; and during sideways markets, it is 0.0144. The performance difference across the three market environments is controlled within 10%, indicating that the method of this invention has good robustness.

[0129] As shown in Figure 2, the financial time series data multi-scale feature analysis and prediction system of the present invention includes six core modules: multi-source data acquisition module 1, multi-scale decomposition module 2, time series Transformer encoding module 3, multi-scale attention fusion module 4, probability prediction module 5, and closed-loop optimization module 6. The data flow and feedback relationship between the modules are as follows.

[0130] The multi-source data acquisition module 1 is responsible for collecting and preprocessing financial time-series data from multiple data sources. This module includes a data interface unit, a time alignment unit, a missing value handling unit, and a standardization unit. The data interface unit connects to the stock exchange market data system, macroeconomic database, and sentiment analysis service through a standardized API interface to acquire high-frequency market data, daily trading data, macroeconomic indicator data, market sentiment indicator data, and event calendar data. In this embodiment, the data interface unit supports multiple data protocols, including the FIX protocol for acquiring exchange market data, a RESTful interface for acquiring macroeconomic data, and WebSocket for acquiring real-time sentiment indicators. The time alignment unit aligns data of different frequencies to a unified time base, using a strategy combining forward imputation and linear interpolation to handle the alignment problem of data of different frequencies. The missing value handling unit uses the interpolation and imputation strategies described in the method embodiment to handle missing data, while recording the location information of missing values ​​for subsequent quality monitoring. The standardization unit performs Z-score standardization on the data, with the standardization parameters calculated based on a sliding window to adapt to the non-stationary characteristics of financial data. The output of the multi-source data acquisition module 1 is standardized financial time series data. The data dimension is a three-dimensional tensor consisting of the number of samples multiplied by the time step multiplied by the number of features, and it is transmitted to the multi-scale decomposition module 2.

[0131] Multi-scale decomposition module 2 receives standardized financial time-series data and performs multi-scale decomposition processing. This module includes a wavelet transform unit, an empirical mode decomposition unit, and a component fusion unit. The wavelet transform unit uses a discrete wavelet transform algorithm to decompose the input data into multiple layers, outputting wavelet coefficients at different frequency levels. In this embodiment, the wavelet transform unit uses the Daubechies-4 wavelet as the default mother wavelet function, and the default number of decomposition layers is set to 5, but it can be dynamically adjusted according to the feedback from the closed-loop optimization module. The empirical mode decomposition unit uses a sieving algorithm to decompose the input data into multiple intrinsic mode functions. The sieving process uses cubic spline interpolation to construct upper and lower envelopes, and the stopping condition is set to the standard deviation of two adjacent sieving results being less than a preset threshold. The component fusion unit, according to the frequency band matching and component fusion algorithm described in the method embodiment, fuses the wavelet coefficients and intrinsic mode functions into trend components, periodic components, and noise components. The weight parameters of component fusion can be adjusted online by the closed-loop optimization module 6 to adapt to the optimal decomposition strategy under different market environments.

[0132] The temporal Transformer encoding module 3 receives the trend component, periodic component, and noise component respectively, and performs feature encoding. This module includes three parallel Transformer encoders, each with the same structure but independent parameters, specifically processing the temporal components at their respective scales. Each encoder includes four sub-units: an embedding unit, a positional encoding unit, an encoder stacking unit, and a projection unit. The embedding unit maps the input sequence to a high-dimensional feature space using one-dimensional convolution with a kernel size of 3 and an output dimension of 256. The positional encoding unit adds the cross-scale positional encoding described in the method embodiment, enabling the encoder to distinguish input information from different scales through different phase shifts. The encoder stacking unit contains six encoder layers, each including a multi-head self-attention sub-layer, a feedforward neural network sub-layer, two normalization sub-layers, and two residual connections. The multi-head self-attention sub-layer is configured with eight attention heads, each with a dimension of 32. The feedforward neural network sub-layer has an intermediate dimension of 1024, and the activation function is GELU. The projection unit aggregates the encoder output into a fixed-dimensional feature vector, uses global average pooling to aggregate the time dimension, and then projects it linearly to 256 dimensions. The output of the temporal Transformer encoding module 3 is three 256-dimensional vectors: a trend feature vector, a periodic feature vector, and a noise feature vector, which are then transmitted to the multi-scale attention fusion module 4.

[0133] The multi-scale attention fusion module 4 receives feature vectors from three scales and performs adaptive weighted fusion. This module comprises three sub-units: an attention calculation unit, a weight generation unit, and a fusion output unit. The attention calculation unit calculates the inter-scale correlation score according to the cross-scale attention mechanism described in the method embodiment. The three input feature vectors are first stacked to form a scale feature matrix, then a query, key, and value matrix is ​​generated through a learnable linear transformation, and finally, the inter-scale correlation is calculated through a scaled dot product. The weight generation unit generates adaptive weights based on the correlation score, using softmax normalization to ensure the weights sum to 1. The fusion output unit performs the weighted fusion operation, while introducing a gating mechanism to balance the contributions of adaptive fusion and average fusion. The multi-scale attention fusion module 4 also includes residual connections and layer normalization to improve training stability. The attention weight parameters of this module can be adjusted online by the closed-loop optimization module 6 to adapt to dynamic changes in data characteristics.

[0134] The probability prediction module 5 receives multi-scale fused feature vectors and generates prediction results. This module includes three sub-units: a feature transformation unit, a distribution parameter generation unit, and a sampling output unit. The feature transformation unit performs two layers of nonlinear transformation on the fused features, with an intermediate dimension of 512 and the activation function being GELU, to enhance the expressive power of the features. The distribution parameter generation unit contains two parallel linear layers, which output the mean parameter and variance parameter of the prediction distribution, respectively. The variance parameter is generated through logarithmic space to ensure non-negativity, and then converted to the actual variance value through an exponential function. The sampling output unit generates predicted values ​​and confidence intervals based on the distribution parameters. The predicted values ​​are taken from the mean parameter, and the confidence intervals are calculated based on the quantiles of the Gaussian distribution. For multi-step prediction, an autoregressive approach is used to generate the prediction distribution parameters for each step sequentially. The output of the probability prediction module 5 includes a sequence of predicted values ​​for the next multiple steps and a corresponding sequence of confidence intervals.

[0135] The closed-loop optimization module 6 receives predicted and actual observation values ​​and performs closed-loop optimization processing. This module includes three sub-units: an error calculation unit, a gradient calculation unit, and a parameter update unit. The error calculation unit calculates the prediction error, including a mean squared error term and a direction error term, comprehensively reflecting the numerical accuracy and directional accuracy of the prediction. The gradient calculation unit calculates the gradient of the prediction error with respect to the decomposition parameters and attention weights using the backpropagation algorithm, employing computational graph automatic differentiation technology for efficient gradient calculation. The parameter update unit generates parameter update amounts according to the momentum gradient descent strategy described in the method embodiment and applies update amplitude limits and exponential moving average smoothing to ensure the stability of parameter updates. The closed-loop optimization module 6 outputs the decomposition parameter update amounts and feeds them back to the multi-scale decomposition module 2 to adjust the fusion weights and frequency thresholds of the wavelet decomposition; it also outputs the attention weight update amounts and feeds them back to the multi-scale attention fusion module 4 to adjust the parameters of the cross-scale attention mechanism. The execution frequency of the closed-loop optimization is configurable, supporting two modes: updating immediately upon receiving a batch of data and batch updating after accumulating multiple batches of data.

[0136] In one deployment embodiment of the system of this invention, the system runs on a server configured with an 8-core CPU, 32GB of memory, and an NVIDIA RTX 4090 GPU. The system adopts a microservice architecture, with each module deployed as a Docker container, and data transmission between modules achieved through a message queue. The message queue uses Kafka technology, supporting high-throughput and low-latency data transmission, while providing data persistence and fault tolerance. The system supports two operating modes: real-time prediction mode and batch prediction mode. In real-time prediction mode, the system receives the latest market data every minute and outputs prediction results, with an end-to-end latency of less than 100 milliseconds, meeting the low-latency requirements of high-frequency trading scenarios. In batch prediction mode, the system supports backtesting analysis of historical data, with a daily processing capacity exceeding 1 million records, which can be used for strategy development and parameter optimization scenarios.

[0137] This invention also provides a visual interface and API for user interaction and system integration. The visual interface is implemented using Web technology and employs a responsive design to support both desktop and mobile access. Interface functions include: real-time prediction result display, showing predicted values ​​and confidence intervals in line graph format; historical prediction accuracy analysis, statistically analyzing prediction performance indicators for different time periods; visualization of multi-scale decomposition results, displaying time-series waveforms of trend components, periodic components, and noise components; feature importance analysis reports, showing the contribution of features at different scales to the prediction results; and model parameter monitoring, displaying real-time trends in decomposition parameters and attention weights. The API supports both RESTful and WebSocket protocols. The RESTful interface is used for batch data querying and configuration management, while the WebSocket interface is used for real-time data push. The API provides a robust authentication and access control mechanism, supporting concurrent access by multiple users.

[0138] The fault tolerance and recovery mechanism of this invention is designed as follows: Each module implements a health check interface, and the monitoring service periodically probes the module status and triggers alarms and automatic restarts when an anomaly occurs. Critical data is stored using a primary-backup dual-write strategy to ensure no data loss. Model parameters are periodically saved as checkpoints, supporting recovery from any checkpoint. The message queue is configured with a message retry mechanism, automatically re-delivering failed messages, with a maximum of 3 retries. System logs employ a distributed log collection scheme, supporting end-to-end tracing and problem localization.

[0139] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for multi-scale feature analysis and prediction of financial time series data, characterized in that, include: The multi-source data acquisition steps involve acquiring multi-source financial time-series data, including high-frequency market data, daily trading data, macroeconomic indicator data, market sentiment indicator data, and event calendar data. The multi-source financial time-series data is preprocessed with time alignment and missing value imputation to generate standardized financial time-series data. The multi-scale decomposition step involves inputting the standardized financial time-series data into a multi-scale decomposition module. The multi-scale decomposition module uses discrete wavelet transform to perform multi-level decomposition on the standardized financial time-series data to obtain wavelet coefficients at different frequency levels. Simultaneously, empirical mode decomposition is used to adaptively decompose the standardized financial time-series data to obtain multiple intrinsic mode functions. Based on the frequency characteristics of the wavelet coefficients and the intrinsic mode functions, frequency band matching and component fusion are performed to generate trend components, periodic components, and noise components. In the temporal feature encoding step, the trend component, the periodic component, and the noise component are respectively input into the temporal Transformer encoding module. The temporal Transformer encoding module adds cross-scale position encoding information to each component and captures the temporal dependency relationship within each component through a multi-layer self-attention mechanism, and outputs the trend feature vector, periodic feature vector, and noise feature vector respectively. The multi-scale feature fusion step involves inputting the trend feature vector, the periodic feature vector, and the noise feature vector into a multi-scale attention fusion module. The multi-scale attention fusion module calculates adaptive weights for features at different time scales through a cross-scale attention mechanism. Based on the adaptive weights, the trend feature vector, the periodic feature vector, and the noise feature vector are weighted and fused to generate a multi-scale fused feature vector. In the probability prediction step, the multi-scale fused feature vector is input into the probability prediction module, which generates prediction distribution parameters for future multi-step time series based on the multi-scale fused feature vector, and outputs the predicted value and the corresponding confidence interval.

2. The method for multi-scale feature analysis and prediction of financial time series data according to claim 1, characterized in that, In the multi-scale decomposition step, the discrete wavelet transform has 3 to 6 decomposition levels, the wavelet basis functions used are Daubechies wavelets or Symlet wavelets, and the empirical mode decomposition generates 4 to 8 eigenmode functions.

3. The method for multi-scale feature analysis and prediction of financial time series data according to claim 1, characterized in that, In the temporal feature encoding step, the number of encoder layers in the temporal Transformer encoding module is 4 to 8, the number of attention heads in the multi-head self-attention mechanism of each layer is 4 to 16, and the hidden layer dimension is 128 to 512.

4. The method for multi-scale feature analysis and prediction of financial time series data according to claim 1, characterized in that, In the probability prediction step, the prediction step size is 1 to 30 time units, and the confidence level of the confidence interval is 90% to 99%.

5. The method for multi-scale feature analysis and prediction of financial time series data according to claim 1, characterized in that, In the multi-scale decomposition step, the frequency band matching and component fusion include: dividing the frequency band according to the frequency range corresponding to the wavelet coefficients and the instantaneous frequency of the intrinsic mode function; assigning components with frequencies lower than a first preset frequency threshold to the trend component; assigning components with frequencies between the first preset frequency threshold and the second preset frequency threshold to the periodic component; and assigning components with frequencies higher than the second preset frequency threshold to the noise component.

6. The method for multi-scale feature analysis and prediction of financial time series data according to claim 1, characterized in that, In the multi-scale feature fusion step, the cross-scale attention mechanism calculates the adaptive weights of features at different time scales by: concatenating the trend feature vector, the periodic feature vector, and the noise feature vector to form a query matrix, a key matrix, and a value matrix; calculating the correlation score between scales through scaled dot product attention; and generating the adaptive weights based on the correlation score.

7. The method for multi-scale feature analysis and prediction of financial time series data according to claim 1, characterized in that, It also includes a closed-loop optimization step, which compares the predicted value with the actual observed value, calculates the prediction error, inputs the prediction error into the closed-loop optimization module, the closed-loop optimization module generates decomposition parameter update amount and attention weight update amount according to the prediction error, feeds back the decomposition parameter update amount to the multi-scale decomposition module to adjust the decomposition parameters, and feeds back the attention weight update amount to the multi-scale attention fusion module to update the adaptive weight. The calculation of the decomposition parameter update is based on the gradient information of the prediction error with respect to the decomposition parameters, and the calculation of the attention weight update is based on the gradient information of the prediction error with respect to the attention weight. The update process adopts the momentum gradient descent strategy.

8. The method for multi-scale feature analysis and prediction of financial time series data according to claim 1, characterized in that, The temporal Transformer encoding module includes an input embedding layer, a cross-scale position encoding layer, a multi-layer encoder stack, and an output projection layer. Each layer in the multi-layer encoder stack includes a multi-head self-attention sub-layer, a feedforward neural network sub-layer, and two layer normalization sub-layers.

9. The method for multi-scale feature analysis and prediction of financial time series data according to claim 1, characterized in that, The probability prediction module includes a feature transformation layer, a distribution parameter generation layer, and a sampling output layer. The feature transformation layer performs a nonlinear transformation on the multi-scale fused feature vector. The distribution parameter generation layer outputs the mean parameter and variance parameter of the prediction distribution. The sampling output layer generates the predicted value and the confidence interval based on the mean parameter and the variance parameter.

10. A multi-scale feature analysis and prediction system for financial time series data, used to implement the multi-scale feature analysis and prediction method for financial time series data as described in any one of claims 1-9, characterized in that, include: The multi-source data acquisition module is used to acquire multi-source financial time-series data, which includes high-frequency market data, daily trading data, macroeconomic indicator data, market sentiment indicator data, and event calendar data. The module performs time alignment and missing value filling preprocessing on the multi-source financial time-series data to generate standardized financial time-series data. The multi-scale decomposition module is used to receive the standardized financial time series data, perform multi-scale decomposition on the standardized financial time series data using discrete wavelet transform and empirical mode decomposition, and generate trend components, periodic components and noise components based on frequency band matching and component fusion. The temporal Transformer encoding module is used to receive the trend component, the periodic component, and the noise component respectively, add cross-scale position encoding information to each component, and output the trend feature vector, periodic feature vector, and noise feature vector respectively through a multi-layer self-attention mechanism; the multi-scale attention fusion module is used to receive the trend feature vector, the periodic feature vector, and the noise feature vector, calculate adaptive weights through a cross-scale attention mechanism, and perform weighted fusion based on the adaptive weights to generate a multi-scale fused feature vector; The probability prediction module is used to generate prediction distribution parameters for future multi-step time series based on the multi-scale fusion feature vector, and output the predicted value and confidence interval; the closed-loop optimization module is used to generate decomposition parameter update amount and attention weight update amount according to the prediction error between the predicted value and the actual observation value, and feed the decomposition parameter update amount back to the multi-scale decomposition module, and feed the attention weight update amount back to the multi-scale attention fusion module.

Images