Financial risk prediction method based on multi-objective ensemble learning algorithm

Abstract

The invention provides a financial risk prediction method based on a multi-objective ensemble learning algorithm, and aims to solve the problems of multi-risk type comprehensive evaluation, data complexity and nonlinear relation processing, dynamic change adaptability and the like in risk prediction in the financial field. According to the method, a financial time sequence is divided into a plurality of time blocks through a space-time block data preprocessing technology, and a heterogeneous model library is independently trained, so that the adaptability and prediction precision of a model are improved. Meanwhile, an NSGA-II multi-objective optimization algorithm is combined with Gibbs distribution to dynamically distribute weights, the prediction precision and the engineering practicability are balanced, and dynamic optimization of economic indexes and engineering indexes is achieved. In addition, a time decay and mixed attention mechanism is designed, multi-model prediction results are fused, market dynamic changes are further captured, and the accuracy and real-time performance of risk prediction are improved. According to the method, the precision and the real-time response capability of financial risk prediction are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of financial risk prediction, and more specifically to a financial risk prediction method based on a multi-objective ensemble learning algorithm. Background Technology

[0002] In the financial field, risk prediction has always been a crucial research area. With the continuous development and increasing complexity of financial markets, traditional risk prediction methods are struggling to meet the growing demands. Although existing technologies have advanced risk prediction to some extent, they still have many shortcomings that limit their widespread application in financial engineering practice.

[0003] First, current research largely focuses on optimizing single models, such as those based on graph neural networks and multimodal models. However, financial markets exhibit significant nonlinear and time-varying characteristics, making it difficult for single models to dynamically adapt to such volatile market conditions. For example, the "human-machine collaboration" framework proposed by Cao et al. relies on manual rule switching algorithms, which cannot achieve automated responses when market structures change abruptly; while the credit scoring model by Li et al. requires manual screening of weak signals, leading to feature engineering redundancy issues when the model migrates across markets, resulting in a significant decrease in generalization performance.

[0004] Secondly, existing models exhibit significant contradictions in multi-objective optimization. Financial engineering practice requires simultaneously satisfying both economic indicators (such as forecast accuracy) and engineering indicators (such as the speed of response to risk peaks), but existing models often prioritize a single indicator as their core optimization objective. For example, while Wang et al.'s TAGOL model improves the predicted R-squared to 0.89, it suffers from a 30-minute delay in capturing peak values ​​of extreme market fluctuations, failing to meet real-time risk control requirements. Lu et al.'s alternative data model over-relies on the AUC indicator, neglecting the stringent requirements of financial regulators for tail risk warnings.

[0005] Furthermore, the time series modeling mechanisms of existing models lack adaptability. Many models are directly ported from the field of natural language processing and fail to be optimized for the characteristics of financial time series. For example, Yang et al.'s multimodal model failed to distinguish between high-frequency noise and key event windows, resulting in risk signals being overwhelmed by noise; while Li et al.'s credit scoring model uses a standard Long Short-Term Memory (LSTM) network, which is insufficient in capturing long-range dependencies on quarterly financial data and has a high lag error rate.

[0006] Meanwhile, existing models suffer from insufficient feature interpretability, leading to increased regulatory compliance risks. Financial regulation places stringent requirements on the transparency and interpretability of model decision-making processes. However, existing models often lack sufficient feature interpretability, making it difficult for regulators to assess the model's risks and compliance.

[0007] Finally, existing models face a dual-indicator conflict in engineering applications. It is difficult to simultaneously consider traditional statistical indicators (such as R-squared) and financial engineering indicators (such as the accuracy of risk peak prediction), limiting the model's practicality in extreme markets. Furthermore, existing models suffer from response delays when dealing with sudden events, failing to capture dynamic market changes in a timely manner, resulting in delayed risk warnings.

[0008] To address the aforementioned issues, the applicant proposes a financial risk prediction method based on a multi-objective ensemble learning algorithm. Summary of the Invention

[0009] The purpose of this invention is to provide a financial risk prediction method based on a multi-objective ensemble learning algorithm to solve the problems in the prior art.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a financial risk prediction method based on a multi-objective ensemble learning algorithm, comprising the following steps:

[0011] Step 1: Data preprocessing: Input the original financial time series data, dynamically divide the time window using the sliding window method, perform feature engineering, including standardization, missing value imputation and non-linear feature generation, and generate a block dataset;

[0012] Step 2: Model Training: Call the pre-trained model library for each time block. The pre-trained model library contains deep learning models, traditional machine learning models, and linear models. Perform distributed training and save the model parameters and performance metrics.

[0013] Step 3: Multi-objective weight calculation: The NSGA-II algorithm is called to generate the Pareto front, which is based on the trade-off between economic and engineering indicators. The Gibbs distribution weights are calculated and the model performance ranking is output.

[0014] Step 4: Attention Fusion and Prediction: Calculate the time utility decay factor, which is based on an exponentially decaying utility weight correction model. Call the hybrid attention network, which includes time-aware attention and performance-weighted attention, and output the final prediction result.

[0015] Step 5: Model Update and Iteration: Retrain the model based on the new time block data, dynamically adjust the Pareto front, and achieve online learning capabilities.

[0016] Optionally, the multi-objective ensemble learning algorithm includes:

[0017] Multi-objective Pareto front modeling: Define the model performance evaluation index as economic index and engineering index, and solve the Pareto front set through the NSGA-II multi-objective optimization algorithm;

[0018] Gibbs distribution weight function: The probability density function of Gibbs distribution is constructed based on the Pareto front curve level, and the parameters are dynamically tuned through Bayesian optimization to achieve a dynamic balance between economy and engineering.

[0019] Multi-objective optimization strategy: Employ a reinforcement learning framework to model the model training process as a Markov decision process and dynamically adjust the allocation ratio of training samples.

[0020] Optionally, the spatiotemporal block training includes:

[0021] Time window division rules: The financial time series is divided into continuous non-overlapping time spans, each time span contains k consecutive time points, and the length of the time window is dynamically adjusted according to the data frequency;

[0022] Block-based independent training mechanism: Independently train a heterogeneous model library consisting of seven benchmark models for each time span;

[0023] Dynamic window adapter: Automatically adjusts the time window length based on the data sampling frequency, supporting multi-granularity analysis.

[0024] Optionally, the multi-model dynamic collaboration includes:

[0025] Heterogeneous model library: contains deep learning models, traditional machine learning models and linear models, and ensures that the model parameters of each time block do not interfere with each other through data isolators;

[0026] Distributed training: Performs distributed training based on the Spark framework, saving model parameters and performance metrics.

[0027] Optionally, the attention mechanism weighted fusion includes:

[0028] Time decay factor calculation: Establish a utility weight correction model based on exponential decay to dynamically adjust the influence of historical time blocks;

[0029] Hybrid attention mechanism integration: Design a two-stage attention network, namely time-aware attention and performance-weighted attention, and dynamically adjust the model weights through time difference;

[0030] Gated normalization function: The output layer uses a gated normalization function to ensure prediction stability.

[0031] Optionally, the following steps from the summary of implementation methods may also be included:

[0032] Pareto Front Visualization Tool: Generates 3D scatter plots to illustrate the trade-offs between economic indicators, engineering indicators, and computational latency;

[0033] Adaptive weight allocator: Automatically adjusts Gibbs distribution parameters based on real-time market conditions;

[0034] Multi-objective optimization controller: coordinates the NSGA-II algorithm and the reinforcement learning reward function to ensure continuous updates of the Pareto front;

[0035] Model training sandbox: Creates an independent computing environment for each time span, using Docker container technology to achieve resource isolation;

[0036] Feature engineering pipeline: Standardization, missing value imputation, and non-linear feature generation are performed on each time block;

[0037] Model initialization: Initialize the heterogeneous model library during the spatiotemporal block preprocessing stage;

[0038] Model update and iteration: Retrain the model based on new time block data, dynamically adjust the Pareto front, and achieve online learning capabilities;

[0039] Exponential decay calculation engine: supports custom decay rates and dynamically generates time utility weights;

[0040] Hybrid Attention Fusion Unit: Processes predictions from different models in parallel and performs weighted fusion using Gibbs distribution weights;

[0041] Configurable attention parameter interface: Allows users to adjust the time window weight coefficient and the model performance weight coefficient.

[0042] Beneficial effects: Solves the problem of missing multi-algorithm collaboration mechanisms.

[0043] This invention employs a spatiotemporal block training paradigm, independently training multiple benchmark models (deep learning models, traditional machine learning models, and linear models) and then weightedly fusing them to achieve collaborative work among multiple algorithms. This mechanism overcomes the problem of insufficient generalization ability of a single model and improves the model's adaptability and prediction accuracy in complex financial scenarios.

[0044] Resolving contradictions in multi-objective optimization:

[0045] This invention employs the NSGA-II multi-objective optimization algorithm and the Gibbs distribution weight function to achieve a dynamic balance between economic indicators (such as prediction accuracy) and engineering indicators (such as risk peak response speed). This strategy solves the problem that existing models struggle to balance traditional statistical indicators with financial engineering indicators in engineering applications, thereby improving the model's practicality and engineering value.

[0046] Improve the adaptability of time series modeling mechanisms:

[0047] This invention designs a two-stage attention network (time-aware attention and performance-weighted attention), combined with a time utility decay factor, to improve the local sensitivity and long-range memory capability of time-series modeling. This design addresses the problem of insufficient adaptability of existing attention mechanisms to the non-stationary characteristics of high-frequency financial data, reduces prediction lag, and improves the timeliness of risk warning.

[0048] Enhancing feature interpretability and reducing regulatory compliance risks:

[0049] Although this invention does not directly mention specific improvements in feature interpretability, it indirectly enhances the model's interpretability through multi-dimensional risk quantification modeling and modular design. Financial institutions can more easily understand the model's risk assessment basis, thereby reducing regulatory compliance risks.

[0050] Resolving conflicting metrics in engineering applications:

[0051] By employing a multi-objective optimization strategy, this invention strikes a balance between prediction accuracy and risk peak response speed, ensuring that the model can simultaneously meet the requirements of both economic and engineering indicators in engineering applications.

[0052] Improve the speed of response to emergencies:

[0053] The real-time dynamic adjustment capability and adaptive model update mechanism of this invention enable the model to adapt more quickly to market changes, including the occurrence of unexpected events. By assigning higher weight to recent data and automatically triggering the model retraining process, the model can adjust its prediction results in a timely manner, reducing the lag in risk warnings. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the process of an embodiment of the present invention;

[0055] Figure 2 This is a flowchart of Embodiment 3 of the present invention. Detailed Implementation

[0056] The preferred embodiments of the present invention are described below with reference to the accompanying drawings to make the technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0057] This invention proposes a financial risk prediction method based on a multi-objective ensemble learning algorithm. This method constructs a new generation of intelligent risk control system through an innovative spatiotemporal block training paradigm, a multi-objective optimization mechanism based on evolutionary game theory, and a dynamic attention network incorporating temporal characteristics. This system can effectively improve the accuracy and real-time performance of financial risk prediction, providing financial institutions with more comprehensive and accurate risk prediction services.

[0058] The technical solution of the present invention includes the following modules and steps:

[0059] The spatiotemporal segmented data preprocessing module is the foundation of this invention. This module divides the financial time series into continuous, non-overlapping time spans, each containing k consecutive time points. The time window length is dynamically adjusted based on the data frequency. For example, for minute-level data, k = 60; for daily data, k = 21. A sliding window method is used to divide the time windows, retaining a certain number of overlapping time points between adjacent windows to maintain temporal continuity. For each time span, a heterogeneous model library consisting of seven benchmark models is independently trained, including deep learning models (such as LSTM, GRU), traditional machine learning models (such as RF, SVM, GBDT, LightGBM), and linear models (such as OLS, Ridge). A data isolator ensures that the model parameters of each time block do not interfere with each other, preventing information leakage. Simultaneously, each time block undergoes standardization, missing value imputation (based on time series interpolation), and nonlinear feature generation (such as volatility surface fitting).

[0060] The multi-objective dynamic weight allocation module is the core of the technical solution of this invention. This module defines the model performance evaluation index as the economic index R. 2 The Pareto front set is solved using the NSGA-II multi-objective optimization algorithm, along with the engineering metric A_peak (peak prediction accuracy). Based on the Pareto front curve levels, a Gibbs probability density function is constructed, and parameters are dynamically tuned using Bayesian optimization to achieve a dynamic balance between economic efficiency and engineering feasibility. A reinforcement learning framework is employed to model the model training process as a Markov decision process, dynamically adjusting the training sample allocation ratio based on current market volatility, model performance metrics, and computational resource utilization. This module also includes Pareto front visualization tools, an adaptive weight allocator, and a multi-objective optimization controller to support the implementation and monitoring of the multi-objective optimization process.

[0061] The temporal utility decay and attention fusion module is the key to the technical solution of this invention. This module establishes a utility weight correction model based on exponential decay, dynamically adjusting the influence of historical time blocks and assigning higher weights to recent data to improve the model's response speed to dynamic market changes. A two-stage attention network is designed, namely time-aware attention and performance-weighted attention, which are weighted and fused using Gibbs distribution weights. The output layer uses a gated normalization function to ensure prediction stability. This module also includes an exponential decay calculation engine, a hybrid attention fusion engine, and a configurable attention parameter interface to support the implementation and adjustment of the temporal utility decay and attention fusion process.

[0062] The technical approach of this invention includes the following steps: data segmentation and preprocessing, independent model training, multi-objective weight calculation, attention fusion and prediction, and model updating and iteration. In the data segmentation and preprocessing step, raw financial time series data is input, time windows are dynamically divided, feature engineering is performed, and segmented datasets are generated. In the independent model training step, a pre-trained model library is called for each time block to perform distributed training, and model parameters and performance metrics are saved. In the multi-objective weight calculation step, the NSGA-II algorithm is called to generate the Pareto front, Gibbs distribution weights are calculated, and the model performance ranking is output. In the attention fusion and prediction step, the time utility decay factor is calculated, a hybrid attention network is called, and the final prediction result is output. In the model updating and iteration step, the model is retrained based on the new time block data, the Pareto front is dynamically adjusted, and online learning capability is achieved.

[0063] The module collaborative working mechanism of this invention includes data flow interaction, control signal linkage, and resource allocation coordination. The output of the spatiotemporal segmentation module serves as the input of the multi-objective optimization module, and the weight parameters generated by the multi-objective optimization module are passed to the attention fusion module. The multi-objective optimization module dynamically adjusts the window length k of the spatiotemporal segmentation based on the reinforcement learning signal. The attention fusion module requests GPU / TPU resource allocation from the model training module according to computational latency requirements.

[0064] By implementing the technical solution of this invention, a new generation of intelligent risk control system can be constructed, effectively solving problems such as the lack of multi-algorithm collaboration mechanisms, contradictions in multi-objective optimization, and insufficient adaptability of time-series modeling mechanisms in existing financial risk prediction methods. This method provides financial institutions with more comprehensive, accurate, and real-time risk prediction services, possessing significant innovation and practicality, and is expected to have a profound impact on the field of financial risk prediction.

[0065] The technical solution of this invention utilizes a spatiotemporal block training paradigm to divide financial time series into continuous, non-overlapping time spans. Multiple benchmark models are independently trained for each time span, enabling collaborative work among multiple algorithms and improving the model's adaptability and prediction accuracy in complex financial scenarios. Simultaneously, a multi-objective optimization algorithm and weighting function are employed to achieve a dynamic balance between economic and engineering indicators, addressing the difficulty of simultaneously considering traditional statistical and financial engineering indicators in engineering applications. Furthermore, a two-stage attention network is designed, incorporating a time utility decay factor, to improve the local sensitivity and long-range memory capabilities of time series modeling, resolving the insufficient adaptability of existing attention mechanisms to the non-stationary characteristics of high-frequency financial data. All modules are loosely coupled through standardized interfaces.

[0066] Example 1: Stock Market Risk Prediction

[0067] In the stock market, investors and financial institutions face a variety of risks, including market risk, credit risk, and liquidity risk. These risks are intertwined and have a complex impact on portfolio value. To effectively manage these risks, an investment bank decided to adopt a financial risk prediction method based on a multi-objective ensemble learning algorithm.

[0068] The implementation steps are as follows:

[0069] Data collection and preprocessing:

[0070] Collect historical stock price data, financial statements, macroeconomic indicators, and news and public opinion data.

[0071] The spatiotemporal block data preprocessing module is used to divide the data into continuous non-overlapping time windows, with each window containing a certain number of continuous time points (e.g., k=21 for daily data).

[0072] The data within each time window are standardized, missing values ​​are filled, and non-linear features are generated.

[0073] Model training:

[0074] A heterogeneous model library consisting of seven independently trained benchmark models, including LSTM, GRU, RF, SVM, GBDT, LightGBM, and OLS.

[0075] Using a multi-objective dynamic weight allocation module, the NSGA-II multi-objective optimization algorithm is employed to solve for economic indicators (such as R²). 2 The Pareto frontier set of ) and engineering indicators (such as peak prediction accuracy).

[0076] The Gibbs distribution probability density function is constructed based on the Pareto front curve level, and the weights of each model are dynamically adjusted.

[0077] Attention fusion and prediction:

[0078] Calculate the time utility decay factor to give more attention to recent data.

[0079] The prediction results of each model are integrated using a hybrid attention mechanism, and the model weights are dynamically adjusted through time-aware attention and performance-weighted attention.

[0080] The final prediction results are output, including the range of stock price fluctuations and the timing of risk peaks.

[0081] The implementation results are as follows:

[0082] This investment bank has successfully improved the accuracy of stock market risk prediction by implementing a financial risk prediction method based on a multi-objective ensemble learning algorithm.

[0083] The model can capture real-time market dynamics and significantly improves its ability to capture peak values ​​of extreme market fluctuations, thus meeting the needs of real-time risk control.

[0084] Investment banks can adjust their portfolios based on forecasts, reducing potential losses and enhancing market competitiveness.

[0085] Example 2: Bank Credit Risk Assessment

[0086] In banking lending, accurately assessing borrowers' credit risk is crucial for reducing non-performing loan rates and ensuring the quality of bank assets. One commercial bank has decided to adopt a financial risk prediction method based on a multi-objective ensemble learning algorithm to improve the accuracy and efficiency of credit risk assessment.

[0087] The implementation steps are as follows:

[0088] Data collection and preprocessing:

[0089] Collect structured data such as borrowers' personal information, credit records, income statements, and balance sheets.

[0090] Collect unstructured data such as news reports and social media information to assess borrowers’ social reputation and potential risks.

[0091] The spatiotemporal block data preprocessing module is used to divide the data into time windows suitable for credit risk assessment.

[0092] Model training:

[0093] Independent training is available for heterogeneous model libraries suitable for credit risk assessment, including deep learning models (such as LSTM and GRU) and traditional machine learning models (such as RF and SVM).

[0094] The multi-objective dynamic weight allocation module is used to optimize indicators such as prediction accuracy and risk peak response speed.

[0095] By using Pareto frontier visualization tools, we can analyze the performance of different models on various indicators and select the optimal model combination.

[0096] Attention fusion and prediction:

[0097] Calculate the time utility decay factor to ensure that recent data has a greater weight in credit risk assessment.

[0098] Integrating the predictions from various models using a hybrid attention mechanism improves the accuracy and robustness of the evaluation.

[0099] It outputs borrowers' credit scores and potential risk levels, providing a scientific basis for banks' credit decisions.

[0100] The implementation results are as follows:

[0101] This commercial bank has significantly improved the accuracy and efficiency of credit risk assessment by implementing a financial risk prediction method based on a multi-objective ensemble learning algorithm.

[0102] The model can comprehensively consider multiple aspects of borrower information, including structured and unstructured data, to provide a more comprehensive risk assessment.

[0103] Banks can formulate more reasonable credit policies based on forecast results, reduce non-performing loan ratios, and improve asset quality.

[0104] Example 3

[0105] In the complex environment of high-frequency trading in financial markets, dramatic fluctuations, like turbulent tides, constantly test the accuracy and stability of predictive models. Recently, a groundbreaking study introduced the physical concept of cascading effects into the field of financial forecasting, proposing the MOI-Transformer model based on a multi-objective attention mechanism. By capturing the chain reaction of anomalous shocks, it reconstructs the high-frequency volatility prediction paradigm in the spatiotemporal dimensions. This model innovatively constructs a five-layer architecture comprising a decomposition module, a history perception module, a pattern perception module, an MOI perception module, and a prediction module, forming a complete closed loop from data deconstruction to risk prediction.

[0106] In the data preprocessing stage, the decomposition module acts like a sophisticated spectrum analyzer, breaking down the original financial time series into historical records across multiple frequency bands. This decomposition strategy is not a simple time-domain segmentation, but rather, based on mathematical tools such as wavelet transform, it maps data containing different fluctuation characteristics to corresponding frequency domain spaces. High-frequency bands capture instantaneous market sentiment pulses, while low-frequency bands reflect macroeconomic cycle fluctuations. The data in each frequency band maintains its independent characteristics while forming an organic whole through frequency coupling mechanisms, providing multi-dimensional analytical material for subsequent modules.

[0107] The historical perception module employs an improved multi-head cross-attention mechanism to construct a spatiotemporal sample association network. This module overcomes the limitation of traditional attention mechanisms that only focus on the time dimension by designing a spatiotemporal joint encoder to simultaneously capture the temporal continuity and spatial similarity of historical samples. In its implementation, each attention head focuses on a specific time scale, quantifying the correlation of market states across different time periods by calculating the dot product similarity between the query vector and the key vector. This design enables the model to identify both daily-level periodic patterns and capture minute-level abnormal fluctuation patterns, forming a comprehensive historical memory bank.

[0108] The morphology perception module introduces a dual-layer gating mechanism to achieve refined extraction of morphological features. The outer gating acts as a data filter, filtering noise signals based on a volatility threshold; the inner gating acts as a feature selector, dynamically adjusting the weights of each frequency component through a self-attention mechanism. In the feature extraction stage, the module adopts an improved Transformer encoder structure, incorporating morphological constraint terms into the self-attention calculation, enabling the model to maintain sensitivity to local morphological features while paying attention to global dependencies. Experiments show that this dual-layer gating design improves the model's accuracy in recognizing typical candlestick patterns such as V-shaped reversals and step-like rises.

[0109] As a core innovation of the model, the MOI perception module constructs a multi-objective optimization framework for energy level transformation. This module breaks through the traditional single-task learning paradigm, decoupling reliability assessment and validity verification into two parallel optimization objectives. At the energy transformation level, the module designs an adaptive energy function that maps historical fluctuation energy to a feature representation of the current state. By introducing Pareto optimal frontier theory, the module searches for solutions in the parameter space that simultaneously optimize both objectives. This multi-objective optimization strategy effectively alleviates the trade-off between accuracy and robustness in traditional models.

[0110] The prediction module employs a hybrid LSTM-CNN architecture, forming a prediction engine that fuses spatiotemporal features. The LSTM network is responsible for capturing long-term dependencies in time series data; its cell states act as memory carriers, conveying the evolutionary trajectory of market states. The CNN network extracts local pattern features through spatial scanning of convolutional kernels. The outputs of both are fused through a feature concatenation layer to form composite features that combine global trend judgment and local pattern recognition. Finally, fully connected layers map these features to specific risk indicators, achieving end-to-end prediction from data to decision.

[0111] In the empirical evaluation phase, the research team constructed a multi-asset test set covering stocks, futures, and foreign exchange. Comparative experiments showed that the MOI-Transformer outperformed the benchmark model in key indicators such as volatility prediction and extreme risk warning. Particularly during the global financial market turmoil of 2025, the model issued a risk warning 30 minutes in advance, and its predicted curve closely matched the actual volatility trajectory, validating the effectiveness of cascading effect modeling. This model not only provides a new analytical tool for high-frequency trading, but its multi-objective optimization and spatiotemporal joint modeling concepts also open up new paths for model innovation in the fintech field.

[0112] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0113] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A financial risk prediction method based on a multi-objective ensemble learning algorithm, characterized in that, Includes the following steps: Step 1: Data preprocessing: Input the original financial time series data, dynamically divide the time window using the sliding window method, perform feature engineering, including standardization, missing value imputation and non-linear feature generation, and generate a block dataset; Step 2: Model Training: Call the pre-trained model library for each time block. The pre-trained model library contains deep learning models, traditional machine learning models, and linear models. Perform distributed training and save the model parameters and performance metrics. Step 3: Multi-objective weight calculation: The NSGA-II algorithm is called to generate the Pareto front, which is based on the trade-off between economic and engineering indicators. The Gibbs distribution weights are calculated and the model performance ranking is output. Step 4: Attention Fusion and Prediction: Calculate the time utility decay factor, which is based on an exponentially decaying utility weight correction model. Call the hybrid attention network, which includes time-aware attention and performance-weighted attention, and output the final prediction result. Step 5: Model Update and Iteration: Retrain the model based on the new time block data, dynamically adjust the Pareto front, and achieve online learning capabilities.

2. The financial risk prediction method according to claim 1, characterized in that, The multi-objective ensemble learning algorithm includes: Multi-objective Pareto front modeling: Define the model performance evaluation index as economic index and engineering index, and solve the Pareto front set through the NSGA-II multi-objective optimization algorithm; Gibbs distribution weight function: The probability density function of Gibbs distribution is constructed based on the Pareto front curve level, and the parameters are dynamically tuned through Bayesian optimization to achieve a dynamic balance between economy and engineering. Multi-objective optimization strategy: Employ a reinforcement learning framework to model the model training process as a Markov decision process and dynamically adjust the allocation ratio of training samples.

3. The financial risk prediction method according to claim 1, characterized in that, The spatiotemporal block training includes: Time window division rules: The financial time series is divided into continuous non-overlapping time spans, each time span contains k consecutive time points, and the length of the time window is dynamically adjusted according to the data frequency; Block-based independent training mechanism: Independently train a heterogeneous model library consisting of seven benchmark models for each time span; Dynamic window adapter: Automatically adjusts the time window length based on the data sampling frequency, supporting multi-granularity analysis.

4. The financial risk prediction method according to claim 1, characterized in that, The multi-model dynamic collaboration includes: Heterogeneous model library: contains deep learning models, traditional machine learning models and linear models, and ensures that the model parameters of each time block do not interfere with each other through data isolators; Distributed training: Performs distributed training based on the Spark framework, saving model parameters and performance metrics.

5. The financial risk prediction method according to claim 1, characterized in that, The attention mechanism weighted fusion includes: Time decay factor calculation: Establish a utility weight correction model based on exponential decay to dynamically adjust the influence of historical time blocks; Hybrid attention mechanism integration: Design a two-stage attention network, namely time-aware attention and performance-weighted attention, and dynamically adjust the model weights through time difference; Gated normalization function: The output layer uses a gated normalization function to ensure prediction stability.

6. The financial risk prediction method according to claim 1, characterized in that, It also includes the following steps in the summary of implementation methods: Pareto Front Visualization Tool: Generates 3D scatter plots to illustrate the trade-offs between economic indicators, engineering indicators, and computational latency; Adaptive weight allocator: Automatically adjusts Gibbs distribution parameters based on real-time market conditions; Multi-objective optimization controller: coordinates the NSGA-II algorithm and the reinforcement learning reward function to ensure continuous updates of the Pareto front; Model training sandbox: Creates an independent computing environment for each time span, using Docker container technology to achieve resource isolation; Feature engineering pipeline: Standardization, missing value imputation, and non-linear feature generation are performed on each time block; Model initialization: Initialize the heterogeneous model library during the spatiotemporal block preprocessing stage; Model update and iteration: Retrain the model based on new time block data, dynamically adjust the Pareto front, and achieve online learning capabilities; Exponential decay calculation engine: supports custom decay rates and dynamically generates time utility weights; Hybrid Attention Fusion Unit: Processes predictions from different models in parallel and performs weighted fusion using Gibbs distribution weights; Configurable attention parameter interface: Allows users to adjust the time window weight coefficient and the model performance weight coefficient.

Images