Financial text sentiment feature optimization method and system based on multi-supervised signal LoRA fine-tuning

Abstract

The present application relates to the technical field of data information processing, in particular to a financial text sentiment feature optimization method and system based on multi-supervision signal LoRA fine-tuning. The method comprises: obtaining financial text and transaction data; generating at least three groups of sentiment scores based on a large language model, including original scores, first fine-tuned scores based on financial special model label fine-tuning, and second fine-tuned scores based on future yield label fine-tuning; calibrating each version of the scores and constructing multi-dimensional emotion features; fusing each version of the emotion features with technical index features to form multiple candidate feature sets; using time series cross-validation and multiple heterogeneous evaluation models to systematically evaluate and statistically compare the candidate feature sets, and selecting the optimal emotion feature source according to the performance; and deploying a prediction model based on the optimization result. The present application realizes empirical comparison of fine-tuning strategies and objective optimization of feature sources, can avoid invalid fine-tuning overhead, and has a standardized engineering implementation process.

Description

Technical Field

[0001] This invention relates to the field of data information processing technology, specifically to a method and system for selecting the best sentiment features of financial texts based on LoRA fine-tuning of multi-supervised signals. Background Technology

[0002] Automated sentiment analysis of unstructured texts such as financial news and research reports using large language models has become a key technology for quantitative investment and risk warning. However, when adapting general-purpose large language models to specialized financial sentiment analysis tasks, existing technical solutions still face a series of systemic challenges and shortcomings: 1. Domain semantic understanding bias leads to insufficient adaptation: The pre-training corpus of general-purpose large language models is mainly composed of general texts, which lacks a deep understanding of the terminology, expression habits, and implicit risk semantics specific to the financial field. For example, the model may have difficulty accurately distinguishing between the drastically different warning levels of "expected loss" and "performance below expectations" in the financial market, resulting in limitations on the accuracy and domain specialization of sentiment judgment.

[0003] 2. Lack of Empirical Guidance for Efficient Parameter Fine-Tuning Strategies: Efficient parameter fine-tuning techniques, such as LoRA, provide a feasible path for domain adaptation of large models under limited computing power. However, in the specific task of financial sentiment analysis, there is significant technical uncertainty regarding which supervisory signal to choose to guide fine-tuning. Practitioners face a dilemma: should they use the output of existing mature financial models (such as FinBERT) as labels for knowledge distillation-style fine-tuning, or use the actual future market returns as supervisory signals for prediction-oriented fine-tuning? There is a lack of systematic technical comparison and selection criteria regarding the impact of different fine-tuning strategies on the performance of the final downstream task (such as market forecasting).

[0004] 3. Subjective and Fixed Selection of Sentiment Feature Sources: When multiple sentiment feature acquisition methods exist simultaneously, including zero-sample outputs of general large models, multiple versions fine-tuned with different supervision signals, and results from domain-specific models, existing technologies typically rely on developers' experience to make singular and static selections. This subjective selection mode lacks a systematic optimization mechanism based on multi-source comparison and empirical evaluation, which may result in the selected sentiment features not being the optimal solution for the current task and failing to adapt to changes in the relative performance of various feature sources under different market environments.

[0005] 4. Lack of Standardized Implementation for Fine-tuning: Successfully applying LoRA technology to financial sentiment analysis involves adapting a series of hyperparameters and engineering details, such as rank selection, learning rate setting, training round control, and batch processing strategies. Currently, there is a lack of validated and standardized engineering implementation plans for this specific task, resulting in poor reproducibility between different teams or projects, low efficiency in technology implementation, and hindering the standardization development of related technologies.

[0006] In summary, existing technologies have clear shortcomings in terms of deep adaptation of financial sentiment analysis models, scientific selection of fine-tuning strategies, objective optimization of feature sources, and standardization of engineering practices. This invention aims to address these problems specifically by providing a complete technical solution. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for selecting the best sentiment features in financial texts based on LoRA fine-tuning with multiple supervision signals. This method fills the technical gap in financial sentiment analysis by lacking a systematic comparison and selection of the effects of different supervision signals for LoRA fine-tuning. It also solves the problems of relying on subjective experience in the selection of sentiment feature sources and lacking an objective multi-source comparison and evaluation mechanism. Furthermore, it provides a standardized and reproducible LoRA fine-tuning engineering implementation plan for financial sentiment analysis.

[0008] This invention is achieved through the following technical solution: A method for selecting the best sentiment features in financial texts based on LoRA fine-tuning using multi-supervised signals is provided, comprising the following steps: S1. Multi-source data acquisition and preprocessing: Acquire stock index trading data and financial news text data within a specified time range, and perform cleaning, normalization, and text merging by trading day.

[0009] S2. Multi-version sentiment score generation: Based on a large language model, at least three sets of sentiment scores are generated: the original sentiment score (S0) obtained through zero-shot inference using structured cue words; the first fine-tuned sentiment score (S1) obtained by fine-tuning with LoRA using a financial-specific model (such as FinBERT) sentiment score as the first supervisory signal; and the second fine-tuned sentiment score (S2) obtained by fine-tuning with LoRA using the actual market returns over the next 5 trading days as the second supervisory signal. Wherein: the first fine-tuned sentiment score (S1) is the sentiment score sequence output by the large language model for news text after fine-tuning with the first supervisory signal; the second fine-tuned sentiment score (S2) is the sentiment score sequence output by the large language model for news text after fine-tuning with the second supervisory signal.

[0010] S3. Sentiment Score Calibration and Feature Construction: Quantile calibration was performed on the three sentiment scores (S0, S1, and S2) to eliminate systematic distribution bias between different versions, resulting in the calibrated score S0. ’ S1 ’ S2 ’ Based on the time series of calibrated scores for each group, five dimensions of emotion features—mean, standard deviation, trend, positive proportion, and negative proportion—are calculated using a sliding window, forming corresponding five-dimensional emotion feature sets F0, F1, and F2. F0 corresponds to the original emotion score S0, representing the original emotion score obtained based on zero-shot inference, and the five-dimensional emotion feature set constructed after calibration. F1 corresponds to the first fine-tuned emotion score S1, representing the emotion score obtained through LoRA fine-tuning using FinBERT emotion scores as a supervisory signal, and the five-dimensional emotion feature set constructed after calibration. F2 corresponds to the second fine-tuned emotion score S2, representing the emotion score obtained through LoRA fine-tuning using the actual increase in the index over the next 5 days as a supervisory signal, and the five-dimensional emotion feature set constructed after calibration.

[0011] S4. Multi-feature set generation: Calculate multi-dimensional technical indicator feature set F based on market data. tech ; F tech Concatenate with F0, F1, and F2 respectively to form at least three candidate feature sets (feature sets A, B, and C).

[0012] S5. Multi-model Evaluation and Optimization: Multiple heterogeneous machine learning models (e.g., Random Forest, LightGBM, XGBoost) are selected as evaluation models. A time-series cross-validation framework is used to train and evaluate the performance of each model on candidate feature sets A, B, and C, with AUC as the core performance indicator. The average performance of each feature set across all evaluation models is compared. The optimal sentiment feature source is automatically selected based on a preset optimization rule (e.g., the model with the highest average AUC is prioritized; if the difference is not statistically significant, the version with the lowest engineering cost is selected). Statistical significance testing can be used as a supplement. The optimization rule is: the candidate feature set with the best average AUC across all evaluation models is selected as the optimal sentiment feature source; if the AUC differences among multiple candidate feature sets are not statistically significant, the version with the lowest engineering implementation cost is selected.

[0013] S6. Predictive Deployment: Based on the optimal sentiment feature source selected through selection, it is compared with F tech The features are fused to form the final feature set. The best-performing evaluation model is selected for final training to obtain the deployment model, which is used to predict new data.

[0014] S7: Rolling optimization update, repeating steps S1~S5 according to a preset cycle, and dynamically updating the selection of the optimal sentiment feature source using newly accumulated data.

[0015] Accordingly, the present invention also provides a system for implementing the above method, comprising: a data preprocessing module, a multi-version sentiment score generation module, a score calibration and feature construction module, a multi-feature set generation module, a multi-model evaluation and optimization module, and a prediction deployment module.

[0016] The data preprocessing module is used to acquire and clean stock trading data and financial news text; A multi-version sentiment score generation module is used to perform zero-shot inference and LoRA fine-tuning with at least two different supervision signals to generate multiple sets of sentiment scores; The score calibration and feature construction module is used to calibrate the sentiment scores of each version and construct a five-dimensional sentiment feature set and a technical indicator feature set based on the calibrated scores. The multi-feature set generation module is used to merge the sentiment feature sets of various versions with the technical indicator feature sets to form multiple alternative feature sets; The multi-model evaluation and selection module is used to manage the time-series cross-validation evaluation process of multiple evaluation models and to make selection decisions based on the evaluation results. The prediction deployment module is used to train and deploy the final prediction model based on the best results.

[0017] The beneficial effects of this invention are: I. Filling the technological gap in selecting the best fine-tuning strategy: For the first time, this invention systematically compares the effects of different fine-tuning strategies, such as "knowledge distillation" and "prediction-oriented" approaches, within the same framework. It selects the optimal solution through empirical data rather than empirical judgment, thus resolving long-standing technological uncertainties. Experiments show that different fine-tuned versions exhibit significant differences in sentiment distribution (e.g., after FinBERT-based fine-tuning, the negative sentiment percentage reaches 93.6%, while the original sentiment score is only 0.2%). The optimal selection mechanism of this invention can effectively identify the best approach.

[0018] Second, avoid unnecessary computational overhead and improve efficiency: The optimization process may find that the sentiment feature performance of the original zero-shot inference is better than or equal to that of the fine-tuned version. In this case, the present invention can help users avoid investing in expensive and unnecessary fine-tuning training costs, and directly adopt the zero-shot solution, which significantly saves computational resources and engineering time.

[0019] Third, it provides a standardized engineering implementation plan: it gives suggested value ranges and typical configurations for key parameters such as LoRA rank, learning rate, and number of training rounds, forming a set of reproducible and transferable engineering specifications for fine-tuning LoRA in financial sentiment.

[0020] Fourth, the selection process is rigorous, interpretable, and dynamically adaptable: The introduction of time-series cross-validation and statistical testing as decision-making criteria makes the feature selection process transparent and auditable, meeting financial compliance requirements. Furthermore, by supporting "rolling selection" (periodic reassessment), it can dynamically track changes in market style and promptly switch to the current optimal feature source, thereby maintaining the long-term performance of the predictive model. 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] Example 1:

[0023] This embodiment uses the selection of the optimal sentiment feature source for predicting the rise and fall of the CSI 300 Index in the next 5 days as an example to illustrate the implementation process of the present invention.

[0024] S1. Data Acquisition and Preprocessing: Acquire daily price data for the CSI 300 Index from 2016 to 2025, along with relevant financial news during the same period. Clean the news text and merge it into a daily corpus, then normalize the price data.

[0025] The news text is cleaned by removing HTML tags, URL links, and special characters; standardizing the encoding format; and merging multiple news articles from the same trading day into a single text. After cleaning and merging, data on trading days with effective news coverage is obtained. For individual trading days without news coverage, subsequent sentiment scores are filled using a forward imputation method based on a decay coefficient. Specifically, for trading days lacking sentiment features, the corresponding feature values ​​of each sentiment feature are taken from the most recent effective trading day and multiplied by the decay coefficient λ raised to the power of n, where λ = 0.2 and n is the number of days between the current trading day and the most recent effective trading day. This operation aims to prevent the influence of outdated information from being excessively amplified.

[0026] Minimum-maximum normalization is applied to the market data, mapping each indicator to the [0, 1] interval.

[0027] S2, Multiple Versions of Sentiment Score Generation: Raw Sentiment Score (S0): Using the DeepSeek API with a temperature of 0, the raw sentiment score for each news item is obtained through structured prompts. The prompt template includes a role setting (financial sentiment analysis expert), a score range (-1 to 1, where -1 represents extreme negativity and 1 represents strong positivity), and output format requirements.

[0028] 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 strongly 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.

[0029] 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.

[0030] 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~ -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: -1.00 to -0.90.

[0031] 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.).

[0032] Inference was performed using a locally deployed DeepSeek large language model. To reduce output randomness, the model parameters were set as follows: temperature = 0.1, top-p = 0.9, and max tokens = 50. The model output was standardized and post-processed: the numerical value after "score:" was precisely extracted using regular expressions and constrained to the range [-1, 1]. Simultaneously, the text after "basis:" was extracted. This text, generated in real-time by the model based on prompts, the input news content, and its internal reasoning logic, represents the human-readable reasoning behind the sentiment score and is used for subsequent interpretability analysis. For unparseable abnormal outputs, the system automatically assigned a neutral score of 0.0 to ensure the robustness of the process. Finally, the original DeepSeek sentiment score sequence and the corresponding judgment basis text were obtained.

[0033] First fine-tuning of sentiment score (S1): Using FinBERT's sentiment scores (positive probability - negative probability) for the same news as the supervision signal, the DeepSeek base model is fine-tuned using LoRA.

[0034] The second fine-tuning of the sentiment score (S2) uses the actual increase in the index over the next 5 days as a monitoring signal and fine-tunes the base model using the same LoRA configuration.

[0035] During parameter tuning, this invention conducted grid search experiments on the hyperparameters for LoRA fine-tuning. The experiments show that when the LoRA rank is between 16 and 64, the learning rate is between 1e-5 and 5e-5, the number of training epochs is between 2 and 5, and the batch size is between 4 and 16, the model performance is stable and reaches a relatively good level. The parameter combination used in this embodiment (rank r=32, learning rate 2e-5, number of training epochs 3, batch size 8) is a typical configuration within this range.

[0036] The base model for both fine-tuning tasks is DeepSeek-LLM-7B-Chat. LoRA parameter configuration: rank r=32, learning rate 2e-5, number of training epochs 3, batch size 8, maximum sequence length 512, optimizer AdamW.

[0037] S3, Score Calibration and Feature Construction: Quantile calibration was performed on S0, S1, and S2 respectively. The fractional distributions were then mapped to a standard normal distribution using the Quantile Transformer (standard quantile transformation method), yielding S0 and S2 respectively. ’ S1 ’ S2 ’The calibration parameters are fitted only to the data from each fold of the training set and applied to the corresponding validation set. This ensures that the calibration process does not introduce future information and prevents data leakage. Using a 5-day sliding window, five-dimensional sentiment features are calculated based on each calibrated score, forming F0, F1, and F2. Simultaneously, 29-dimensional technical indicator features F1 are calculated. tech .

[0038] The multidimensional sentiment feature set includes the sentiment mean, sentiment standard deviation, sentiment trend, positive sentiment percentage, and negative sentiment percentage calculated based on a sliding window. Specifically: the sentiment mean is the arithmetic mean of the calibrated scores within the window; the sentiment standard deviation is the sample standard deviation of the calibrated scores within the window; the sentiment trend is the difference between the calibrated scores of the last trading day and the first trading day of the window, reflecting the direction of sentiment change; the positive sentiment percentage is the ratio of days with a calibrated score > 0.1 within the window to the window length; and the negative sentiment percentage is the ratio of days with a calibrated score < -0.1 within the window to the window length.

[0039] S4. Multi-feature set generation: Construct three 34-dimensional candidate feature sets: Feature set A (F tech +F0), feature set B (F tech +F1), feature set C(F tech +F2). Simultaneously construct a structure containing only F. tech The benchmark set and the set containing FinBERT features and the set containing F tech A reference set of features is used as a comparison.

[0040] S5. Multi-model evaluation and selection: Five heterogeneous models—Random Forest, LightGBM, XGBoost, SVM, and MLP—were selected as evaluators, and 11-fold time-series cross-validation was performed. The average AUC (AUC, the area under the ROC curve, is used to measure the overall ranking ability and the ability to distinguish between positive and negative samples) of each feature set on each model was recorded.

[0041] The optimization results (as shown in Table 1 below) show that, on the random forest model, feature set A (original sentiment score) has the highest AUC (0.6386), outperforming feature set B (first fine-tuned sentiment score), feature set C (second fine-tuned sentiment score, 0.633), and the reference set FINBERT (0.6289). Compared with historical experimental results, the performance of the original sentiment score (DeepSeek) has reversed from slightly inferior to FinBERT (original AUC 0.6332 vs. 0.6339) to superior (new AUC 0.6386 vs. 0.6289), and its performance improvement relative to the baseline set (NONE) without sentiment features (△AUC=0.0394) has significantly improved statistical significance (p-value of paired t-test improved from 0.347 to 0.081, close to the significance boundary of 0.05). According to the optimization rules, the sentiment feature of the original sentiment score is selected as the optimal source.

[0042] Table 1 Optimal Selection Results

[0043] To verify the statistical significance of the performance differences, a paired t-test was performed. The test results showed that: DeepSeek (raw sentiment score) and FinBERT (reference): t=0.768, p=0.460. The p-value is much greater than 0.05, indicating that although DeepSeek has a higher mean AUC, the performance difference between the two is not statistically significant.

[0044] DeepSeek (raw sentiment score) and NONE (benchmark): t=1.941, p=0.081. The p-value is close to the significance level of 0.05, indicating that the performance improvement brought by the sentiment features of the raw sentiment score is marginally significant and its economic significance is clear.

[0045] According to the selection rule, feature set A (original sentiment score) has the highest average AUC. Although the difference between it and feature set C (second fine-tuned sentiment score) is not statistically significant (p>0.05), considering that feature set A does not require any fine-tuning training and has the lowest engineering implementation cost, the system automatically selects the original sentiment score sentiment feature as the optimal sentiment feature source under the comprehensive trade-off between performance and cost.

[0046] This optimal decision has also been consistently validated on other evaluation models such as LightGBM and XGBoost: the original sentiment score or the second fine-tuned sentiment score is generally better than the first fine-tuned sentiment score, and the original sentiment score achieves the highest or second-highest performance on most models.

[0047] S6. Predictive Deployment: Based on the optimal selection results, the original sentiment score and sentiment features are combined with F... techThe random forest model is trained and deployed on all historical data.

[0048] S7: Rolling optimization update, repeating steps S1~S5 according to a preset cycle, and dynamically updating the selection of the optimal sentiment feature source using newly accumulated data.

[0049] For newly input financial news text and market data, obtain the raw sentiment score according to the above process, construct a five-dimensional sentiment feature, integrate it with technical indicators, input it into the deployment model, and output the prediction of rise and fall in the next 5 days.

[0050] Example 2:

[0051] This embodiment, based on Embodiment 1, demonstrates the scalability and robustness of the method and provides an in-depth analysis of the fine-tuning results.

[0052] (1) Add fine-tuning task: Add a third fine-tuning task, which uses abnormal market returns as a binary monitoring signal for fine-tuning, generates S3 and includes it in the comparison, further enriching the comparison dimensions of the fine-tuning strategy.

[0053] (2) Parameter Comparison: A combination comparison experiment was conducted on the hyperparameters of LoRA fine-tuning. Specifically, the LoRA rank was tested with three configurations: r=16, 32, and 64; the learning rate with three configurations: 1e-5, 2e-5, and 5e-5; the number of training epochs with three configurations: 2, 3, and 5; and the batch size with three configurations: 4, 8, and 16. Experimental results show that within the above parameter range, the model performance is generally stable, with the combination of rank r=32, learning rate 2e-5, number of training epochs 3, and batch size 8 achieving the best balance between performance and efficiency. Performance and model efficiency were considered comprehensively when selecting the optimal combination.

[0054] (3) Rolling selection mechanism: The selection cycle is set to once per quarter. At the end of each quarter, the selection process of S1 to S5 is re-executed using the latest accumulated data to dynamically update the optimal sentiment feature source in order to cope with changes in market style.

[0055] (4) Robustness verification of calibration methods: To verify the dependence of the selection framework on the selection of the front-end calibration method, two different quantile calibration methods were used to process the sentiment scores of each version, and the selection process was run independently for comparison: Method 1 (Standard Method): The standard quantile transformation method is used to map the fractional distribution to the standard normal distribution.

[0056] Method 2 (Heterogeneity Preservation Method): This method employs a calibration approach designed to align distributions while preserving the original differences in the models. The method primarily involves the following steps: First, within a time-series cross-validation framework, to prevent future information leakage; second, using the sentiment score distribution of the Financial Institutional Sentiment Analysis (FinBERT) model as the target distribution, the initial scores are sequentially subjected to: (1) distribution stretching, adjusting its dispersion to align with the target distribution; (2) mean shifting, translating its center position; (3) extreme value correction, handling outliers; and (4) low quantile optimization, finely adjusting the matching of negative intervals. The design goal is to make the calibrated scores comparable to the target distribution (e.g., the FinBERT score distribution) in descriptive statistics (such as mean and standard deviation), without excessively eliminating reasonable distributional differences (i.e., "heterogeneity") caused by fundamental model differences. After the above calibration, the sentiment scores of each version can be compared on the same benchmark, and their original heterogeneity characteristics are preserved to the greatest extent.

[0057] The heterogeneity-preserving quantile calibration method is designed to eliminate systematic distribution biases among sentiment scores of different models or versions (making their means, standard deviations, and other statistics comparable) while preserving reasonable distribution differences caused by the sentiment judgment logic of the source model (such as the large language model).

[0058] Experimental results show that after preprocessing the original sentiment score, the first fine-tuned sentiment score, and the second fine-tuned sentiment score using the two calibration methods described above, the proposed optimization framework consistently selects the original sentiment score (feature set A) as the optimal sentiment feature source. Furthermore, the performance ranking (AUC) relationship of each feature source remains consistent under both calibration methods. For example, on the random forest model, the AUC of the original sentiment score is approximately 0.639 when using heterogeneity-preserving calibration, and approximately 0.640 when using QuantileTransformer calibration. This demonstrates that the optimization framework of this invention is robust to the specific selection of the front-end calibration method, and its optimization conclusion is stable and independent of any particular calibration algorithm, thus broadening its applicability in diverse engineering environments.

[0059] (5) In-depth analysis of the fine-tuning results: Analysis of the sentiment score distribution generated after fine-tuning revealed that the sentiment distribution of "DeepSeek (fine-tuned version)" obtained by fine-tuning based on the return rate label exhibits the characteristics of "neutral to positive with very few negative ones" (e.g., 17.3% positive, 0.7% negative, and 82.0% neutral), which is completely different from the "highly negative" distribution (e.g., a negative proportion of 93.6%) after fine-tuning based on the FinBERT label. This intuitively reveals from the data level how different supervision signals fundamentally change the sentiment judgment logic of the model. Even if the sentiment distribution of the fine-tuned model changes significantly, under the evaluation framework of this invention, the performance of its corresponding feature set B and feature set C in the prediction task may still not be better than the original sentiment score A. This result profoundly confirms the beneficial effect of this invention in "avoiding unnecessary fine-tuning": not all efforts aimed at improving the model can be transformed into improvements in downstream task performance, and the optimal selection mechanism of this invention can effectively identify and avoid such ineffective investment.

[0060] Experimental results show that by increasing the number of fine-tuning tasks and periodically rolling the selection, this method can continuously track the optimal sentiment feature source and maintain good predictive performance in different market stages.

[0061] 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 method for selecting the best sentiment features of financial texts based on LoRA fine-tuning using multi-supervised signals, characterized in that: Includes the following steps: S1. Multi-source data acquisition and preprocessing: Acquire stock index trading data and financial news text data within a specified time range, and perform cleaning and standardization processing. S2. Multi-version sentiment score generation: Based on the large language model, at least three sets of sentiment scores are generated, including: the original sentiment score obtained through zero-shot inference, the first fine-tuned sentiment score obtained through LoRA fine-tuning of the first supervision signal, and the second fine-tuned sentiment score obtained through LoRA fine-tuning of the second supervision signal, which is different from the first supervision signal. S3. Sentiment Score Calibration and Feature Construction: The sentiment scores of each version are calibrated to eliminate distribution bias between versions; based on the calibrated sentiment scores of each version, a corresponding five-dimensional sentiment feature set is constructed. S4. Multi-feature set generation: The five-dimensional sentiment feature sets corresponding to each version are merged with the technical indicator feature sets constructed based on transaction data to form multiple candidate feature sets; S5. Multi-model evaluation and selection: Under the same time-series cross-validation framework, multiple heterogeneous evaluation models are used to evaluate the predictive performance of multiple candidate feature sets; based on the evaluation results, the optimal sentiment feature source is selected from multiple candidate feature sets according to the preset selection rules. S6. Prediction Deployment: Based on the optimal sentiment feature sources selected in S5, construct the final feature set and train the prediction model for subsequent financial prediction tasks.

2. The method for selecting the best sentiment features of financial text based on LoRA fine-tuning of multi-supervised signals according to claim 1, characterized in that: In step S2, the first supervisory signal is the sentiment score of the financial news text by the FinBERT financial sentiment analysis model, and the second supervisory signal is the actual return of the stock index over the next 5 trading days.

3. The method for selecting the best sentiment features of financial text based on LoRA fine-tuning of multi-supervised signals according to claim 2, characterized in that: In step S2, the LoRA fine-tuning tasks corresponding to the first and second supervision signals share the same base model, LoRA rank, learning rate, number of training rounds, and batch size configuration.

4. The method for selecting the best sentiment features of financial text based on LoRA fine-tuning of multi-supervised signals according to claim 3, characterized in that: The LoRA rank ranges from 16 to 64, the learning rate ranges from 1e-5 to 5e-5, the number of training epochs ranges from 2 to 5, and the batch size ranges from 4 to 16.

5. The method for selecting the best sentiment features of financial text based on LoRA fine-tuning of multi-supervised signals according to claim 1, characterized in that: In step S3, the calibration process is performed using either the standard quantile transformation method or the heterogeneity-preserving quantile calibration method, and the calibration aligns the distribution of sentiment scores in each version with the target distribution.

6. The method for selecting the best sentiment features of financial text based on LoRA fine-tuning of multi-supervised signals according to claim 1, characterized in that: In step S3, the five-dimensional emotion feature set includes the emotion mean, emotion standard deviation, emotion trend, positive emotion percentage, and negative emotion percentage calculated based on a sliding window.

7. The method for selecting the best sentiment features of financial text based on LoRA fine-tuning of multi-supervised signals according to claim 1, characterized in that: In step S5, the selection rule is as follows: select the candidate feature set with the best average AUC across all evaluation models as the optimal sentiment feature source; if the AUC differences among multiple candidate feature sets are not statistically significant, then select the version with the lowest engineering implementation cost.

8. The method for selecting the best sentiment features of financial text based on LoRA fine-tuning of multi-supervised signals according to claim 7, characterized in that: Step S5 also includes: performing a statistical significance test on the performance evaluation results and using the test results as a reference for the optimal decision.

9. The method for selecting the best sentiment features of financial text based on LoRA fine-tuning of multi-supervised signals according to claim 1, characterized in that, It also includes step S7: rolling optimization update, repeating steps S1 to S5 according to a preset cycle, and dynamically updating the selection of the optimal sentiment feature source using the newly accumulated data.

10. A financial text sentiment feature optimization system implementing the method as described in any one of claims 1 to 9, characterized in that, include: The data preprocessing module is used to acquire and clean stock trading data and financial news text; A multi-version sentiment score generation module is used to perform zero-shot inference and LoRA fine-tuning with at least two different supervision signals to generate multiple sets of sentiment scores; The score calibration and feature construction module is used to calibrate the sentiment scores of each version and construct a five-dimensional sentiment feature set and a technical indicator feature set based on the calibrated scores. The multi-feature set generation module is used to merge the sentiment feature sets of various versions with the technical indicator feature sets to form multiple alternative feature sets; The multi-model evaluation and selection module is used to manage the time-series cross-validation evaluation process of multiple evaluation models and to make selection decisions based on the evaluation results. The prediction deployment module is used to train and deploy the final prediction model based on the best results.

Images