Stock index prediction method and system based on domestic large model multi-dimensional emotion features and ensemble learning

Abstract

This invention relates to the field of data information processing technology, specifically to a stock index prediction method and system based on multidimensional sentiment features and ensemble learning using a domestically developed large-scale model. The method includes: acquiring and preprocessing stock index trading data and financial news text; extracting and calibrating news sentiment scores using a domestically developed large-scale language model; constructing a five-dimensional sentiment feature based on the calibrated scores, including mean, standard deviation, trend, positive percentage, and negative percentage; simultaneously calculating traditional multidimensional technical indicators such as momentum, volatility, and trading volume; fusing the sentiment features and technical indicator features to form a fused feature set; and finally, training multiple heterogeneous machine learning models using time-series cross-validation, dynamically assigning weights based on their performance, performing weighted ensemble learning, and outputting a probability prediction of the stock index's direction of rise or fall. This invention achieves effective fusion of textual sentiment and market data, improves prediction accuracy and model robustness through ensemble learning, and is suitable for quantitative investment decision support.

Description

Technical Field

[0001] This invention relates to the field of data information processing technology, specifically to a stock index prediction method and system based on multidimensional sentiment features and ensemble learning of a domestically developed large model. Background Technology

[0002] Predicting stock market trends has long been a crucial challenge in the financial sector. Traditional technical analysis methods primarily rely on structured data such as price and volume to construct technical indicators, including the Relative Strength Index (RSI) and Moving Average Convergence Divergence (MACD). However, these indicators only reflect historical market trading patterns and fail to capture changes in investor sentiment triggered by unstructured textual information such as financial news and policy announcements. Numerous empirical studies have demonstrated that investor sentiment is a key driver of short-term market fluctuations.

[0003] Regarding the utilization of financial text information, existing technologies mainly have the following shortcomings: Sentiment analysis methods have limited accuracy: early methods based on sentiment dictionaries struggled to understand complex contextual semantics and implicit sentiment expressions, resulting in low accuracy.

[0004] There is a systematic distribution bias between models: the sentiment scores of general-purpose large language models (such as GPT) and financial domain-specific models (such as FinBERT) for the same text have significant differences in distribution, mean, and variance, which makes their output scores not directly comparable and difficult to integrate.

[0005] The emotional feature expression dimension is limited: Existing technologies usually only extract a single "average emotional score" as a feature, ignoring key dimensions such as the volatility and trend of emotions and the relative proportion of positive and negative information, resulting in limited information content of the extracted features.

[0006] Insufficient robustness of prediction models: On financial time series data with limited sample size and low signal-to-noise ratio, using a single complex deep learning model (such as LSTM) is prone to overfitting and has limited generalization ability. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a stock index prediction method and system based on multidimensional sentiment features of a domestically developed large model and ensemble learning. This method effectively integrates textual sentiment with market data and improves prediction accuracy and model robustness through ensemble learning, making it suitable for quantitative investment decision support.

[0008] This invention is achieved through the following technical solution: This paper presents a stock index prediction method based on multidimensional sentiment features and ensemble learning from a domestically developed large-scale model, comprising the following steps: S1. Multi-source data acquisition and preprocessing: Acquire daily trading data of stock indices and financial news text data within a specified time range, and perform preprocessing operations such as cleaning, normalization, and text merging by trading day.

[0009] S2. Sentiment Score Calculation and Calibration: Using structured cue words, a domestic large language model (such as DeepSeek) is guided to perform sentiment inference on daily news corpora to obtain an initial sentiment score; then, based on the sentiment score distribution of a financial domain-specific model (such as FinBERT), the initial score is quantile-calibrated to obtain a calibrated sentiment score.

[0010] Quantile calibration is a calibration process that aligns the distribution of initial sentiment scores with the score distribution of the target financial domain-specific sentiment analysis model, while preserving the heterogeneity information of the output scores of the large language model. This includes distribution stretching, mean shifting, extreme value correction, and low quantile optimization.

[0011] S3. Construction of five-dimensional emotion features: Based on the time series of calibrated emotion scores, a sliding window (e.g., 5 trading days) is set to calculate the emotion features of five dimensions: emotion mean, emotion standard deviation, emotion trend, positive emotion percentage, and negative emotion percentage.

[0012] S4. Construction of Technical Indicator Features: Based on preprocessed market data, calculate multi-dimensional technical indicators such as momentum, volatility, volume, and time.

[0013] S5. Feature Fusion: The sentiment feature set generated in step S3 is concatenated with the technical indicator feature set generated in step S4 to form a fused feature set. For trading days without valid news data, the sentiment features can be filled in using the forward imputation method, that is, the sentiment feature values ​​of the most recent valid trading day are used. To further prevent information from becoming outdated, a decay coefficient λ (0 < λ ≤ 1, for example, λ = 0.9) can be introduced, and the imputation values ​​are multiplied by λ raised to the power of n (n is the number of missing days).

[0014] S6. Multi-model ensemble learning and prediction based on time-series cross-validation: Select at least three heterogeneous machine learning models (such as Random Forest, LightGBM, and XGBoost) as base learners. Use time-series cross-validation to divide the training and validation sets, strictly preventing future information leakage. In each validation fold, train each base learner using the training set data, and use the AUC value on the validation set as a performance metric, allocating ensemble weights according to their performance. For the trading day to be predicted, each base learner independently outputs the probability of an increase. The system calculates the final ensemble prediction probability using a weighted average formula and outputs the direction signal accordingly.

[0015] Accordingly, the present invention also provides a system for implementing the above method, comprising: a data preprocessing module, an sentiment reasoning and calibration module, a multidimensional feature construction module, an integrated learning prediction module, and a result output module.

[0016] The data preprocessing module is used to acquire and clean stock index trading data and financial news text data; The sentiment reasoning and feature construction module is used to calculate multidimensional sentiment features based on news text; The technical indicator calculation module is used to calculate multi-dimensional technical indicators based on transaction data. The feature fusion module is used to concatenate emotional features and technical indicator features into a fused feature set; The ensemble learning prediction module is used to manage the training, validation, weight allocation, and weighted ensemble prediction of multiple base learners. The results output module is used to output the final prediction results of the rise or fall direction and related information.

[0017] The beneficial effects of this invention are: I. Comprehensive Expression of Emotional Characteristics, Providing Independent Incremental Information: The five-dimensional emotional characteristic system constructed in this invention comprehensively depicts market sentiment from five complementary perspectives: level, fluctuation, trend, positive density, and negative density, solving the problem of information loss from single features. Feature importance analysis shows that features such as the sentiment mean rank highly in importance, surpassing most traditional technical indicators, proving the value of the independent incremental information they provide.

[0018] II. Significantly Improved Prediction Accuracy: Empirical results show that the prediction model incorporating the multidimensional emotion features of this invention has a significantly better performance evaluation index (AUC) than the benchmark model that only uses technical indicators.

[0019] III. Strong System Robustness: Employing a multi-model weighted ensemble strategy, the system integrates base learners with diverse structures and learning strategies, effectively mitigating the risk of single models failing in specific market environments. Experiments show that the ensemble model outperforms individual base learners in stability under different market conditions (bull and bear markets). On the DeepSeek feature set, the ensemble AUC is further improved compared to the optimal single model (random forest).

[0020] IV. Excellent Engineering Practicality and Interpretability: The system has a built-in caching and scheduling mechanism, which reduces repeated calls to the large language model API and improves efficiency. The sentiment judgment output by the large language model is combined with the feature importance analysis of the tree model to form a traceable decision-making logic that is traceable throughout the entire chain of "input text - sentiment score - multi-dimensional features - integrated prediction", which meets the requirements of financial risk control and compliance audit. Detailed Implementation

[0021] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.

[0022] The large language models used in this invention include, but are not limited to, DeepSeek and Wenxin Yiyan. Among them, DeepSeek, a domestically developed large language model, has advantages in Chinese financial text understanding and data security.

[0023] Example 1: 5-day forecast of the CSI 300 Index This embodiment uses the prediction of the rise and fall direction of the CSI 300 Index over the next 5 trading days as an example to explain in detail the implementation process of the present invention.

[0024] 1. Data Acquisition and Preprocessing By using financial data interfaces, we can obtain daily trading data (opening price, closing price, highest price, lowest price, trading volume, and turnover) of the CSI 300 Index from January 2016 to December 2025, as well as authoritative financial news text data for the same period.

[0025] News text is cleaned by removing HTML tags, URL links, special characters, and advertising text. Multiple news articles from the same trading day are merged into a single continuous text by natural date, forming a daily news corpus. For trading days with missing news, subsequent sentiment features are imputed.

[0026] 2. Calculation and calibration of affective scores To ensure the standardization, comparability, and interpretability of the model output, this invention designs a refined prompt template. This template adheres to four principles: domain adaptability, full sentiment spectrum coverage, output standardization, and interpretability. It encompasses a complete sentiment gradient from extremely negative to extremely positive, and is illustrated using typical financial scenarios (such as profit forecasting, policy announcements, and market volatility) as examples. The model is explicitly required to output a sentiment score within the range [-1, 1] with two decimal places and a concise judgment basis, in the format: "Score: [Value], Basis: [One Sentence]", avoiding redundant output and improving the interpretability and cross-sample consistency of the results.

[0027] The prompt words integrate financial terminology to correct semantic discrepancies between general and financial contexts, guiding the model to accurately identify implicit emotional logic in financial texts, rather than relying solely on word matching.

[0028] Specifically, examples of six levels of emotional intensity are included: Example 1 (Strongly Positive): Example scenario: such as "the company releases better-than-expected financial results", "major technological breakthrough", etc.; Example score: 1.0~0.8; Example 2 (Moderately Positive): Example scenarios: such as "steady performance growth" or "policy adjustments in line with expectations"; example score: 0.6~0.8; Example 3 (Neutral): Example scenario: such as "no major market news", "data meets general market expectations", etc.; Example score: 0.10~0.50; Example 4 (Moderately Negative): Example scenarios: such as "performance falling short of expectations", "slight tightening of industry regulatory policies", etc.; Example score: -0.60 to -0.10; Example 5 (Strongly Negative): Example scenarios: such as "the company is under investigation for serious violations" or "the core product has serious problems"; Example score: -0.80 to -0.60; Example 6 (Extremely Negative): Example scenarios: such as "the company faces bankruptcy risk" or "a systemic financial risk event breaks out"; Example score: -0.90 to -1.00.

[0029] By setting the model role as a financial sentiment analysis expert, the template clearly defines the model role, output format, and numerical range. It also guides the model to accurately identify implicit sentiment and correct general semantic biases by covering six typical financial scenario examples ranging from "strongly positive" to "extremely negative" (e.g., a company releases better-than-expected financial reports, performance falls short of expectations, systemic risk events occur, etc.).

[0030] The DeepSeek large language model is called via API. To reduce output randomness, the model parameters are set to: temperature=0.1, top-p=0.9, and max tokens=50. For efficient batch processing, the system implements an MD5 hash-based caching mechanism to avoid duplicate calculations of the same news content. Structured prompts containing role settings, score range constraints [-1, 1], and format requirements are used to perform batch sentiment inference on the daily news corpus, obtaining initial sentiment scores.

[0031] The system uses regular expressions to precisely extract the numerical value after "Score:" and constrains it to the range [-1, 1]. Simultaneously, it extracts the text after "Basis:", which is generated in real-time by the model based on prompts, the input news content, and its internal reasoning logic. This text provides a human-readable justification for the sentiment score and is used for subsequent interpretability analysis. For unparseable abnormal outputs, the system automatically assigns a neutral score of 0.0 to ensure the robustness of the process. Finally, the system yields the original DeepSeek sentiment score sequence and the corresponding judgment basis text.

[0032] Using the sentiment score output by the FinBERT financial model concurrently as the calibration benchmark, the initial score of DeepSeek was processed using the quantile calibration method to obtain the "calibrated sentiment score". After calibration, the consistency of the distribution of the two scores was significantly improved.

[0033] FinBERT, a pre-trained model in the Chinese financial domain, was selected as the calibration distribution benchmark. The positive (p_pos), negative (p_neg), and neutral (p_neu) probabilities output by FinBERT were converted into continuous sentiment scores ranging from [-1, 1] using the formula s_FB = p_pos - p_neg, making them consistent with the dimensions of DeepSeek scores.

[0034] The calibration process is performed in sequence through the following four steps, with all parameters determined through time-series cross-validation: (1) Distribution stretching: Align the standard deviation of the original DeepSeek score distribution with the standard deviation of the FinBERT score. Specifically, stretch it to 60% to 100% of the FinBERT standard deviation to adjust the difference in dispersion between the two.

[0035] (2) Mean shift: The mean of the stretched score is shifted by 50% to 100% towards the FinBERT mean to adapt to the relatively cautious nature of the overall sentiment distribution of financial texts.

[0036] (3) Extreme value correction: Based on the 3σ principle, outliers are shortened to the boundary, and extreme values ​​that exceed the mean ± 2 to 7 times the standard deviation are cut off to eliminate the interference of outliers.

[0037] (4) Low quantile optimization: The values ​​of the low quantile interval (such as the lowest 10th percentile) of the sentiment score distribution are moderately reduced by 0.5 to 5% to more accurately match the subtle distribution of negative sentiment in financial texts, while strictly protecting the scores in the neutral interval ([-0.1, 0.1]) from changing.

[0038] To prevent future information leaks and ensure the scientific rigor of the calibration, the entire calibration process was conducted within a time-series cross-validation framework. Specifically, the data was divided into multiple training and test sets in chronological order (e.g., using a 2- to 20-fold division). Within each fold, only the training set data from historical time periods was used to fit the parameters of the calibration steps described above, and then this was independently applied to future test set data. Finally, all test set results were concatenated in chronological order to obtain the final calibrated sentiment score (DeepSeek_calib).

[0039] As shown in Table 1, the mean (-0.118) and standard deviation (0.179) of the calibrated DeepSeek score are very close to those of FinBERT (mean -0.124, standard deviation 0.185), and the proportion of negative samples has increased significantly from 0.2% to 55.7%, effectively correcting the bias of the general model in judging the sentiment of financial texts. The Pearson correlation coefficient also increased from 0.21 before calibration to 0.52.

[0040] Table 1 Comparison of Emotional Score Distribution FinBERT 83(3.4%) 1368(56.3%) 979(40.3%) -0.124 0.185 DeepSeek Original 754(31.0%) 6(0.2%) 1670(68.7%) 0.087 0.152 DeepSeek after calibration 95(3.9%) 1352(55.7%) 983(40.4%) -0.118 0.179 3. Multidimensional feature construction 3.1 Calculation of Five-Dimensional Emotional Characteristics: Set the sliding window length W = 5 trading days. For each trading day T, calculate the calibrated sentiment score within the window [T-4, T]. Sentiment mean: the arithmetic mean of the scores within the window.

[0041] Sentiment standard deviation: the sample standard deviation of scores within the window.

[0042] Sentiment trend: The difference between the score on day T and the score on day T-4.

[0043] Positive sentiment percentage: The percentage of days with a score greater than the threshold of 0.1 out of the window length.

[0044] Negative sentiment percentage: The percentage of days with a score less than the threshold -0.1 out of the window length.

[0045] 3.2 Calculation of technical indicators: Based on market data, we calculate momentum indicators (such as 1-day momentum, 2-day momentum, RSI), volatility indicators (such as Bollinger Band width), volume indicators (such as volume ratio), and time indicators.

[0046] 3.3 Feature Fusion: The five-dimensional sentiment characteristics and multi-dimensional technical indicators mentioned above are concatenated column by column to form a fused feature vector for each trading day.

[0047] 4. Model Training and Ensemble Prediction Random Forest, LightGBM, and XGBoost were selected as base learners. Five-fold time-series cross-validation was used for model training and evaluation. At each fold, historical data was used to train each model, and performance was evaluated on the subsequent validation set. Taking the detailed results of a single fold of the Random Forest model on the DeepSeek feature set as an example (as shown in Table 2), its AUC reached 0.6198 to 0.6735, demonstrating the rigor of the model evaluation process and the stability of the results.

[0048] Table 2. Examples of detailed results for single-fold random forest model (DeepSeek feature set) 1 0.5489 0.5500 0.5435 0.5467 0.6198 2 0.6213 0.6250 0.6024 0.6135 0.6688 3 0.6085 0.6136 0.5814 0.5971 0.6318 4 0.6468 0.6500 0.6190 0.6341 0.6735 5 0.5617 0.5652 0.5455 0.5552 0.5986 average 0.5974 0.6008 0.5784 0.5893 0.6385 Note: F1 score is the harmonic mean of precision and recall, and AUC is the area under the ROC curve, used to measure the overall ranking ability of the model and its comprehensive ability to distinguish between positive and negative samples.

[0049] As shown in Table 3 below, the ensemble weights W are allocated according to the AUC performance of each base learner on the validation set, in a manner that is positively correlated with their AUC values. m Specifically, the weights W for each model can be calculated using the AUC value itself, the square of the portion of the AUC exceeding the random guess baseline (0.5), or by exponential weighting. m For the date to be predicted, the three base learners independently output the predicted probability P of an increase. m The final ensemble prediction probability P ensemble According to formula P ensemble =∑(W m ×P m ) / ∑(W m ) Calculate. If P ensemble If the value is greater than 0.5, the index is predicted to rise in the next 5 days; otherwise, it is predicted to fall.

[0050] Table 3 shows the average AUC (5-fold average) for each feature set. LightGBM (Lightweight Gradient Boosting Machine Model) 0.5734 0.5904 0.5974 0.6004 XGBoost (Extreme Gradient Boosting Model) 0.5911 0.6051 0.6114 0.6139 SVM (Support Vector Machine) model 0.5700 0.5988 0.5970 0.5775 Random Forest 0.6016 0.6339 0.6385 0.6423 MLP (Multilayer Perceptron Model) 0.6028 0.5894 0.5857 0.5601 5. Analysis of Empirical Results On the test set, the predictive performance of the method in this embodiment was verified in multiple dimensions: Significantly improved prediction performance: As shown in Table 3, after adding the DeepSeek sentiment features of this invention to the random forest model, the AUC significantly improved from the baseline of 0.6016 to 0.6385, an increase of 3.69 percentage points. This improvement is not only reflected in the AUC, but also in the accuracy, precision, recall, and F1 score, as shown in Table 4, confirming the comprehensive gains of the sentiment features.

[0051] Table 4. Comparison of various metrics of the Random Forest model (50% average) None 0.5808 0.5835 0.5620 0.5725 FinBERT 0.5940 0.5962 0.5741 0.5850 DeepSeek 0.5974 0.6008 0.5784 0.5893 Both 0.5915 0.5943 0.5710 0.5824 Statistical and Economic Significance: Although the paired t-test (as shown in Table 5) indicates that the p-value (0.296) for the DeepSeek feature improvement over the benchmark did not reach the statistical significance level of 0.05 due to the limited sample size, the 3.69 percentage point increase in AUC has clear economic significance. In quantitative trading, such an improvement may correspond to a considerable enhancement in annualized returns.

[0052] Table 5. Results of paired t-tests (random forest model) DeepSeek vs None (None) 0.6385 0.6016 1.201 0.296 DeepSeek vs FinBERT 0.6385 0.6339 0.343 0.749 Table 6. Ranking of Feature Importance (Both Feature Sets, Random Forest) 1 mom2 (2-day momentum) 0.089 Momentum 2 mom1 (1-day momentum) 0.082 Momentum 3 bb_width (Boolean bandwidth) 0.074 Volatility 4 RSI_14 (Relative Strength Index) 0.068 Volatility 5 deepseek_mean (DeepSeek sentiment mean) 0.061 Emotional 6 finbert_mean (FinBERT sentiment mean) 0.058 Emotional 7 ma2_ratio (MA2 deviation rate) 0.055 Momentum 8 deepseek_trend (DeepSeek sentiment trend) 0.049 Emotional 9 ret_std (standard deviation of returns) 0.047 Rate of return 10 vol_ma5_ratio (volume ratio) 0.045 Trading volume 11 finbert_trend (FinBERT sentiment trend) 0.042 Emotional 12 price_position (price position) 0.040 Rate of return 13 deepseek_std (DeepSeek Sentiment Standard Deviation) 0.038 Emotional 14 ret_mean (mean return) 0.036 Rate of return 15 finbert_std (FinBERT sentiment standard deviation) 0.034 Emotional Feature importance is prominent: Feature importance analysis (as shown in Table 6) shows that in the fused feature set, the importance of deepseek_mean (sentiment mean) is 0.061, ranking 5th, surpassing most traditional technical indicators, which conclusively proves that the sentiment features extracted by this invention provide incremental prediction information independent of price and trading volume.

[0053] Table 7 Predictive performance of ensemble models (weighted ensemble, γ=1.0) None 0.5734 0.5911 0.6016 0.6123 +1.07% FinBERT 0.5904 0.6051 0.6339 0.6412 +0.73% DeepSeek 0.5974 0.6114 0.6385 0.6478 +0.93% Both 0.6004 0.6139 0.6423 0.6512 +0.89% Ensemble learning enhances robustness: After adopting weighted ensemble (as shown in Table 7), the AUC of the ensemble model reached 0.6478, which is 0.93 percentage points higher than the single optimal model (random forest), verifying the effectiveness of the ensemble strategy.

[0054] Cross-market effectiveness: As shown in Table 8, the method of the present invention can improve the prediction performance in both bull and bear markets, and the improvement is more significant in bull markets (4.77%), indicating that it has strong environmental adaptability and robustness.

[0055] Table 8. Prediction Performance under Different Market Conditions bull market 0.6235 0.6712 4.77% bear market 0.5847 0.6089 2.42% Example 2: Adaptability Validation for Different Prediction Windows This embodiment is basically the same as the implementation of embodiment 1, except that the prediction target is adjusted to the direction of rise or fall in the next 1 day, 3 days, 5 days and 10 days respectively.

[0056] Table 9 Performance improvements under different prediction windows 1st 0.5231 0.5378 1.47% 3rd 0.5684 0.5912 2.28% 5th 0.6016 0.6385 3.69% 10th 0.5942 0.6123 1.81% As shown in Table 9, the method of this invention can effectively improve forecasting performance for different forecasting windows, with the best improvement effect (3.69%) for the 5-day window, successfully capturing the mainstream and sustained cycle of the impact of financial news sentiment on the market. This provides flexible and reliable technical support for timing strategies for different investment horizons.

[0057] Of course, the above description is not limited to the examples above. Technical features not described in this invention can be implemented by or using existing technology, and will not be repeated here. The above embodiments are only used to illustrate the technical solutions of this invention and are not intended to limit this invention. This invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention do not depart from the spirit of this invention and should also fall within the scope of protection of the claims of this invention.

Claims

1. A stock index prediction method based on multidimensional sentiment features and ensemble learning of a domestically developed large-scale model, characterized in that: Includes the following steps: S1. Multi-source data acquisition and preprocessing: Acquire daily trading data of stock indices and financial news text data within a specified time range, and clean and standardize the data. S2. Sentiment Inference and Score Acquisition: Sentiment inference is performed on the preprocessed news text based on a large language model to obtain a sentiment score; S3. Construction of multi-dimensional emotion features: Based on the time series of emotion scores obtained in S2, multiple dimensions of emotion features are calculated through a rolling time window to form an emotion feature set; S4. Construction of technical indicator features: Based on the daily trading data of the stock index obtained in S1, multi-dimensional technical indicators are calculated to form a set of technical indicator features; S5. Feature Fusion: The emotional feature set and the technical indicator feature set are combined to form a fused feature set; S6. Ensemble Learning Prediction: Based on the temporal cross-validation framework, multiple heterogeneous base learners are trained using a fusion feature set. Ensemble weights are assigned according to the validation set performance of each base learner, and the final prediction result of the stock index rise and fall direction is obtained through weighted ensemble.

2. The stock index prediction method based on multidimensional sentiment features and ensemble learning of a domestically developed large model as described in claim 1, characterized in that: In step S3, the emotional characteristics of multiple dimensions include: emotional mean, emotional standard deviation, emotional trend, percentage of positive emotions, and percentage of negative emotions.

3. The stock index prediction method based on multidimensional sentiment features and ensemble learning of a domestically developed large model as described in claim 1, characterized in that: In step S4, the multidimensional technical indicators include at least one of the following: momentum indicators, volatility indicators, and trading volume indicators.

4. The stock index prediction method based on multidimensional sentiment features and ensemble learning of a domestically developed large model as described in claim 1, characterized in that: Step S6 specifically includes: S6.1 Select at least three heterogeneous machine learning models as base learners; S6.

2. Use time series cross-validation to divide the training set and validation set, use the training set data to train each base learner, and evaluate the performance on the validation set, using the area under the curve (AUC) as the performance metric. S6.3 Assign ensemble weights to each base learner based on its AUC value on the validation set. The weights are positively correlated with the AUC value. S6.4 For the sample to be predicted, each base learner independently outputs the prediction probability, and the system performs a weighted average according to the ensemble weights to obtain the final ensemble prediction probability.

5. The stock index prediction method based on multidimensional sentiment features and ensemble learning of a domestically developed large model as described in claim 1, characterized in that: Step S2 specifically includes: S2.

1. Guide the large language model to perform sentiment inference through structured prompts and output the initial sentiment score; S2.

2. Based on the score distribution of a sentiment analysis model specifically for the financial field, the initial sentiment score is quantile-calibrated to obtain the calibrated sentiment score. S3 is calculated based on the calibrated emotion score.

6. The stock index prediction method based on multidimensional sentiment features and ensemble learning of a domestically developed large model as described in claim 5, characterized in that: Quantile calibration is a calibration process that aligns the distribution of initial sentiment scores with the score distribution of the target financial domain-specific sentiment analysis model, while preserving the heterogeneity information of the output scores of the large language model. This includes distribution stretching, mean shifting, extreme value correction, and low quantile optimization.

7. A system for implementing the stock index prediction method as described in any one of claims 1 to 6, characterized in that: include: The data preprocessing module is used to acquire and clean stock index trading data and financial news text data; The sentiment reasoning and feature construction module is used to calculate multidimensional sentiment features based on news text; The technical indicator calculation module is used to calculate multi-dimensional technical indicators based on transaction data. The feature fusion module is used to concatenate emotional features and technical indicator features into a fused feature set; The ensemble learning prediction module is used to manage the training, validation, weight allocation, and weighted ensemble prediction of multiple base learners. The results output module is used to output the final prediction results of the rise or fall direction and related information.