Utilizing Text Mining for Labeling Training Models from Futures Corpus in Generative AI

Abstract

For highly time-constrained, very short-term investors, reading and extracting valuable information from financial news poses significant challenges. The wide range of topics covered in these news articles further compounds the difficulties for investors. The diverse content adds complexity and uncertainty to the text, making it arduous for very short-term investors to swiftly and accurately extract valuable insights. Variations between authors, media sources, and cultural backgrounds also introduce additional complexities. Hence, performing a bull–bear semantic analysis of financial news using text mining technologies can alleviate the volume, time, and energy pressures on very short-term investors, while enhancing the efficiency and accuracy of their investment decisions. This study proposes labeling bull–bear words using a futures corpus detection method that extracts valuable information from financial news, allowing investors to quickly understand market trends. Generative AI models are trained to provide real-time bull–bear advice, aiding investors in adapting to market changes and devising effective trading strategies. Experimental results show the effectiveness of various models, with random forest and SVMs achieving an impressive 80% accuracy rate. MLP and deep learning models also perform well. By leveraging these models, the study reduces the time spent reading financial articles, enabling faster decision making and increasing the likelihood of investment success. Future research can explore the application of this method in other domains and enhance model design for improved predictive capabilities and practicality.

1. Introduction

Currently, the internet is flooded with fake news and futures transactions designed to mislead retail investors. These deceptive practices enable large investors to manipulate news or exploit public information, leading to widespread incorrect judgments among most investors. Generative AI has the potential to train models capable of analyzing and discerning news texts, automatically identifying the distinguishing features of fake news [1]. The model has the ability to learn patterns in the text, identify emotional biases, and recognize the distinguishing factors that separate fake news from genuine information. This knowledge can assist investors in discerning between real news and potential disinformation, enabling them to make more informed decisions [2]. Utilizing a model trained by generative AI, investors can gain valuable insights into the prediction and potential impact of news regarding bull–bear trends. The characteristic of index futures is the ability to profit from both “bull” and “bear” directions. Unlike stock price investments, where one generally profits only when stock prices rise, index futures operate differently. When you correctly predict the direction of the index, even if it moves downward, which represents a “bear” trend, you can still earn profits. This is distinct from stocks, where profits usually depend on a “bull” trend. It represents a different concept altogether. This guidance empowers investors to make more accurate investment decisions, resulting in improved outcomes [3]. Nevertheless, the labeling process involved in the preprocessing of these models is currently fraught with challenges [2]. The performance and accuracy of generative AI models are inherently tied to the quality and reliability of the training data. If the training data contains errors, biases, or inaccurate information, it can lead to misleading results generated by the models [4]. In the field of finance and economics, labels often adhere to general rules, but in more specific areas, such as assessing index futures, the duality of bull and bear positions complicates the labeling process. The overall direction of news may not be immediately apparent, and it may require comprehensive analysis across multiple sources. Additionally, there is currently a shortage of labels that capture the impact of news broadcasts that induce investor panic [5].

The primary focus of this research lies in addressing the challenges faced by current researchers, particularly in regards to the labeling of bull–bear news using semantic analysis for index futures. The ultimate goal is to establish a comprehensive futures corpus that can serve as an effective training resource for generative AI models. The article highlights the specific challenges faced by current researchers in labeling bull–bear news for semantic analysis in the context of index futures [6]. By identifying these challenges, the research aims to contribute to overcoming them and advancing the field [7]. The research aims to establish a comprehensive futures corpus that can serve as a valuable resource for training generative AI models. This focus on creating a specialized corpus for index futures distinguishes this research from previous studies [8,9] that may have focused on more general language or topic domains. Hence, this literature review has identified two primary research questions in this field. Firstly, what are the essential input factors for training, i.e., what are the key features on which machine learning relies for accurate labeling? Secondly, does the process of labeling contribute positively to the detection of articles using generative AI, thereby enhancing the accuracy rate and facilitating the creation of a profitable training model? To address the aforementioned research questions, this study will commence by conducting a comprehensive literature review to ascertain whether existing research methodologies have successfully answered these questions or if there are any limitations encountered by previous researchers. Subsequently, this study will propose innovative research methods aimed at empirical investigation through a research model, with the objective of validating the proposed hypotheses. Drawing from the empirical findings, the study will endeavor to provide explanations, discuss potential discoveries, and ultimately draw conclusions.

The remainder of this paper is organized as follows: the next section provides a review of the related literature. Section 3 details research methods regarding a labeling training model based on futures corpus. Section 4 and Section 5 present the experimental process, the results, and a discussion. Finally, the last section summarizes our contributions and provides suggestions for future research.

2. Literature Review

There has been a notable emphasis on generative algorithms in recent research endeavors [10,11,12,13]. Nevertheless, in more specialized and niche fields, it is often necessary to provide a specific direction or concise summary of the text [14]. In the context of index futures, particularly in the case of relatively short-term contexts, making instantaneous judgments becomes challenging. This difficulty arises due to the limited availability of labeled articles in this specific field, resulting in insufficient or incorrect learning materials. When learning is based on inaccurate or inadequate sources, even achieving a high rate of machine learning prediction correctness may lead to erroneous outcomes due to overfitting, which is not beneficial to the majority of investors. For example, Gurrib et al., in their study published in 2022 [15], concentrated on commodity futures and did not undertake a labeling process specifically tailored for index futures. Similarly, Dai et al., in their study published in 2022 [16], and Wang et al., in their study published in 2020 [17], did not specifically address the labeling process for index futures. In these recent studies, addressing the first research question regarding the input factors for training and labeling, providing a comprehensive explanation, especially for index futures with short-term fluctuations, is challenging. Due to the dynamic nature of index futures and the varied influencing factors at play, it becomes difficult to identify a specific set of input factors that consistently apply across different scenarios. The uniqueness and complexity of short-term fluctuations in index futures result in significant variations in the influencing factors [18,19].

Indeed, prior to commencing the training of a generative AI model, a substantial amount of preparatory work needs to be undertaken in the text domain. This includes tasks such as data collection, where text data is gathered to serve as training material for the model [19,20]. In order to collect the necessary text data for training, various methods can be employed, such as utilizing web crawlers or executing targeted database queries. Once the data is collected, the next crucial step is data cleaning and preprocessing. This involves removing special characters, punctuation marks, numbers, and other unnecessary elements from the collected data. Additionally, handling missing or erroneous data is also part of the preprocessing phase to ensure that the data is of high quality and suitable for training the generative AI model [21]. Tokenization or sentence segmentation involves splitting the text into sequences of words or sentences. This process can be achieved by utilizing simple methods, such as whitespace or punctuation-based tokenization. To create a vocabulary, one must build it based on the word segmentation results, encompassing all the words or phrases present in the training data. Data encoding is a crucial step in converting text data into a numerical representation that can be effectively processed by the model.

In order to train text data, it is necessary to build a suitable model and select an appropriate architecture. There are several options to consider, such as recurrent neural networks (RNNs), convolutional neural networks (CNNs), or transformer models, among others. Each architecture offers unique advantages and is applicable in different scenarios. The choice of model architecture should be based on the specific characteristics of the text data and the objectives of the task at hand [22]. Once the selected model has been established, the next step is to train it using the training data. This process involves feeding the data into the model for forward propagation, calculating a loss function to measure the model’s performance, and utilizing the backpropagation algorithm to update the model’s parameters. Additionally, parameter adjustment is essential, in which model parameters are fine-tuned, based on the training process results and their impact on performance.

This adjustment may involve tweaking hyper parameters like learning rate and regularization to optimize the model’s performance [23]. Determining whether the annotation process is beneficial for article detection by generative AI, leading to improved accuracy and the development of a profitable training model, requires validation and testing. Trained models need to be evaluated using separate validation datasets to assess potential overfitting and overall performance. This process is complex and essential to answer the second research question. However, existing studies lack focus on index futures, particularly in the context of extremely short-term bull–bear texts, which makes it challenging to generate a fully labeled training set specifically for this domain [24]. To address the need for a comprehensive semantic library or corpus for subsequent generative AI training, it is essential to either compile an existing semantic library or create a new corpus specifically tailored for the desired training objectives. This involves gathering relevant and diverse text data from reliable sources, ensuring the inclusion of various linguistic patterns and domain-specific knowledge. By curating a robust semantic library or creating a new corpus, researchers can provide valuable resources for effectively training generative AI models.

3. Labeling Methods for Generative AI

Based on the research methods and the issues discussed in recent literature, this study employs text processing techniques aligned with generative AI to collect text data and perform various preprocessing steps for the domain of extremely short-term trading in index futures. The methodology involves creating and labeling a bull–bear training set, as depicted in Figure 1. Initially, a widely recognized and extensive financial website is utilized as a web crawler to gather relevant articles. When employing deep learning for natural language processing, the articles are transformed, converting each word into a vector representation through techniques such as word embedding.

In Figure 1, pertinent word vectors can be employed to convert words into vectors. TF–IDF is commonly used in text retrieval and information retrieval tasks to measure the similarity between documents by calculating the weights of words. BERT, on the other hand, is widely applied in various natural language processing tasks, including text classification, named entity recognition, and question answering. Its ability to understand contextual information enables it to excel in handling complex semantic tasks. However, in this article, the focus is on labeling the text representing the bull and bear stock futures market, based on its binary nature and the characteristics of short texts. The objective is to quickly annotate the data to respond to very short-term trading. Therefore, BERT may not be the most suitable choice for this specific research area [25]. These innovative models utilize TF–IDF and Word2Vec techniques to enhance their effectiveness in analyzing and processing the data in this field. One widely used approach is Word2Vec, which is a neural network-based language model that represents words as high-dimensional vectors. In the context of Chinese financial news, Word2Vec can be utilized to train word vectors by considering multiple words or even word sequences as a single unit. This approach, known as a word vector model using multiple words or word sequences as a unit, enhances the representation power of the word vectors and captures more contextual information. Chinese words are often structurally complex, comprising multiple characters or morphemes. Consequently, Chinese text processing commonly involves word segmentation to divide the text into individual words or phrases. In the Word2Vec model, the segmented words serve as training units to derive vector representations for each word. These word vectors effectively capture semantic information and express contextual relationships between Chinese words. Subsequently, the vector representations of all words are concatenated to create a matrix of dimensions (N, d), where N represents the number of words in the text, and d represents the dimension of the word vector. This matrix can be perceived as a sequence of length N. Moving forward the sequence is subjected to a 1D convolutional layer. The convolution operation is applied to the sequence, allowing flexibility in defining the size of each convolution kernel.

The convolution operation treats neighboring words as local text segments, enabling the extraction of features from these segments. It can be visualized as a sliding window moving across the sequence, extracting features within the window at each step. Subsequently, a pooling layer is applied to reduce the extracted features from each convolution kernel to a single numerical value. Common pooling techniques, such as max pooling or average pooling, can be utilized for this purpose. This step aids in compressing the features extracted by each convolution kernel, resulting in a reduction in the number of model parameters. Finally, all the compressed features are concatenated and passed through a fully connected layer for classification. In this specific case, a three-layer fully connected network can be employed for classification. The number of neurons in the first layer can be set to match the feature dimension outputted by the pooling layer, while the number of neurons in the second layer can be determined by the number of categories for text classification. The “softmax” activation function is commonly used for classification tasks. In summary, the proposed deep learning text classification model utilizes convolution operations to extract sequence features and employs pooling layers for feature compression. This allows the text to be represented as a fixed-dimensional vector suitable for subsequent classification tasks. The model design can be adapted for various text classification applications and can be further optimized by adjusting the hyper parameters, such as the size, number, and pooling methods of the convolution kernels. Additionally, part-of-speech tagging, specifically Chinese part-of-speech tagging, involves assigning each word in a Chinese sentence its corresponding part of speech, such as noun, verb, adjective, etc. For unknown words, their part of speech can be inferred from the dictionary entry for that word.

Through the analysis of sentence structure, word semantics, and linguistic rules, part-of-speech tagging can be performed by assigning each word its appropriate part of speech. In addition, TF–IDF (term frequency–inverse document frequency) is a widely used text mining technique that can be applied to evaluate the importance of each word within the financial news corpus. It calculates the word frequency of each term within the document and then determines the inverse document frequency of each term using the IDF formula. By multiplying the TF and IDF values, the TF–IDF value of each word can be obtained. By calculating the TF–IDF values for both bull and bear words, the final classification of the article as either bull or bear can be determined. This process involves collecting a significant number of financial articles, annotating them accordingly, and assembling them into a training set for the subsequent steps.

The second step involves constructing a futures corpus, as depicted in Figure 2. Initially, financial articles need to be collected and compiled into a text dataset. These articles consist of news reports sourced from reputable financial websites. The relevance of these texts to bull and bear futures is verified through the reports provided by financial experts on the websites. Once the labeled training set is available, the entire text dataset is utilized as a corpus. Preprocessing techniques, such as synonym and antonym identification, are employed to establish the bull and bear terms. Additionally, various text exploration techniques are applied to the generated training set, including tasks like word segmentation, word frequency statistics, and TF–IDF calculations to determine the weights of the vocabulary. Through processing and analysis of the corpus, more valuable text information can be obtained.

Once the corpus is established, a portion of the text can be chosen as a test set for model evaluation and validation purposes. Furthermore, the training set can be retained within the corpus to facilitate the retrieval of suitable training data for subsequent model training. In summary, it is a common practice to first construct a training set and then build a corpus. This approach allows for a more targeted construction of the corpus, enhancing the efficiency and accuracy of text mining tasks. By ensuring a focused and relevant training set, the subsequent corpus can be designed with greater precision, leading to improved outcomes in the text analysis and mining processes. The third step involves making real-time predictions regarding articles that can influence actual bull–bear trends, as illustrated in Figure 3. This step focuses on capturing relevant real-time financial news, inputting it into the trained training set, and retrieving the appropriate training data from the established corpus based on the article type. Instant predictions can then be generated. Various algorithms can be employed for this purpose, with deep learning being a notable approach that can be considered.

Figure 4 illustrates the scenario without using the futures corpus as a baseline section, allowing it so serve as a control group for comparing state-of-the-art methods. This approach highlights the significance of utilizing the futures corpus in the real-time text mining of financial news. Without the construction of a futures corpus, one is often limited by surface-level keywords like “bull” or “bear”, as meaningful information may be concealed behind these keywords. The futures corpus generated through the iterative labeling process using machine learning serves as a crucial variable, significantly enhancing the contributions of this article. In comparison to the current research, the proposed method encompassing the futures corpus should improve the accuracy of bull or bear predictions.

Before proposing the experimental process, a research model is constructed, as shown in Figure 5. The left side represents the preprocessing step of text mining. Various calculations, including PoS tagging and TF–IDF, are performed to generate multiple empty words. These empty words serve as the training set, with some extracted from the corpus to save collection time. After feedback, the valid data is saved back to the corpus for subsequent extractions. To evaluate the effectiveness of the training models, the latest real-time articles from actual financial websites are obtained for prediction and judgment. These real-time data will be labeled to increase the effectiveness of the training set. The effectiveness of the model requires assessment through training and testing using various algorithms. To cover a wide range of possibilities, several popular algorithms, such as support vector machines (SVMs), multi-layer perceptron (MLP), random forest, and deep learning, will be used for efficiency verification. Figure 5 depicts the scenario in which the futures corpus is utilized, serving as a control group for comparison with state-of-the-art methods as the baseline section. This approach effectively highlights the significance of employing the futures corpus in the real-time text mining of financial news. Without the construction of a futures corpus, the analysis is often limited by surface-level keywords such as “bull” or “bear,” overlooking meaningful insights that lie beneath these keywords. The futures corpus generated through iterative labeling using machine learning plays a pivotal role, significantly enhancing the contribution of this article. In comparison to the current research, the proposed method, which incorporates the futures corpus, is expected to improve the accuracy of bull or bear predictions. The next section will compare and validate this comparative model against the method proposed in this article.

In summary, the process of capturing and analyzing real-time financial news to make instant predictions on futures bull–bear trends involves several key steps. Firstly, relevant financial news must be captured from multiple sources using crawlers or similar methods, followed by the performance of a text analysis on the collected news. Secondly, a corpus must be established to store classified futures bull–bear financial news, serving as a training data source for the prediction model. Thirdly, the model must be trained using techniques like deep learning, enabling it to automatically analyze news and predict its impact on actual trends, such as bull and bear futures. Real-time financial news is then input into the trained model to obtain predictions on the actual bull or bear positions and their influence on factors such as futures prices. It is important to acknowledge the challenges and limitations involved in this process. The vast amount of news information requires quick and accurate analysis. Additionally, the factors influencing actual bull–bear trends, such as futures prices, are complex and require comprehensive consideration during correlation analysis. Therefore, careful model design and training are necessary to effectively address these challenges. This article distinguishes itself from previous studies through significant model innovations. Figure 1 presents the bull and bear words training set, Figure 2 illustrates the unique futures corpus, and Figure 3 showcases the impact of real-time forecasting. These models are original creations within this article, setting it apart from the existing literature. Consequently, the paper goes beyond being a mere flowchart, as it not only outlines the methods and experiments conducted, but also emphasizes the innovative models developed.

4. Evaluation

For the research study, we collected financial news articles from well-known websites, such as Juheng.com (https://news.cnyes.com/news/id/5159323, accessed on 1 July 2023) and Business Times (https://ctee.com.tw/news/futures/852930.html, accessed on 1 July 2023), from January 2023 to May 2023. A total of 50 news articles were collected during this period. These articles were subjected to text exploration and processing to extract relevant information. Based on the analysis, we established a set of 100 bull and bear words (https://drive.google.com/file/d/19SqTtMMYiG4yjxPALbCTb9ECRYebu7SM/view?usp=sharing, accessed on 1 July 2023) that are significant for futures trading. The established corpus has conducted article training to train the model. The training process involved utilizing various techniques such as word segmentation, part-of-speech tagging, and TF–IDF calculations to process the text data. These methods aimed to enhance the accuracy and effectiveness of the training model. It is important to note that this research study focused specifically on the domain of financial futures news, and the collected articles were obtained from a specific timeframe. This allowed us to gather a targeted dataset and ensure its relevance to the research objectives.

Table 1. The terms accessed from the futures corpus.

Bull			Bear
Positive	Support	Steady growth	Downtrend	Pressure	Intensify
Long position	Enhance	Strengthen	Decline	Lackluster	Weaken
Interest rate cut	Expansion	Rise	Price limit down	Bearish	Panic
Rise	Recovery	Low rebound	Fall below	Pessimistic	Bearish profits
Appreciation	Fill the gap	High growth potential	Sharp decline	Significant drop	Low demand
Profit	Pullback	Rebound	Heavy losses	Net selling	Interest rate hike
Trading volume surge	Return of funds	Buy excess	Downward trend	Capital outflow	Recession
Reach new highs	Rebound	Improve	Weakening	Decline	Lack of strength
Breakthrough	Inflow	Buying sentiment	Weakness	Bearish dominance	Unrelenting
Strong performance	Return flow	Strong demand	Sluggish trend	Foreign selling	Warning
Turn stronger	Rise	Supply shortage	Feeble trend	Short selling	Sharp drop
Robust	Buying pressure	Positive outlook	Slump	Continued decline	Sell-off
Uptrend	Long position	Active stock market	Bottom fishing	Bearish	Weakening trend
Growth	Breakthrough	Stability	Bearish view	Bear market	Short selling
Sharp increase	Increase	Sharp rise	Fall	Break below the moving average	Selling sentiment
Upward movement	Slight rise	Turn positive	Selling pressure	Decrease	Unstable
Bullish	Stability		Risk	Lower

presents the bull and bear terms obtained from the futures corpus. These terms, extracted from financial articles, have undergone preprocessing calculations, including TF–IDF, to justify their inclusion in the corpus. By accessing these terms directly from the corpus, the interpretation process is accelerated, resulting in a more comprehensive vocabulary and improved accuracy in judgments.

In machine learning, various metrics are used to evaluate the performance of a model. Some commonly used metrics include accuracy, precision, and recall, which provide insights into different aspects of the model’s performance. Accuracy is the proportion of correct predictions among all predictions made by the model. It measures the overall prediction accuracy of the model. A higher accuracy indicates a stronger predictive ability of the model. Precision is the proportion of true positive predictions among all positive predictions made by the model. It measures the model’s ability to correctly predict positive examples. A high precision indicates that the model is accurate in identifying positive examples. Recall is the proportion of true positive predictions among all actual positive examples in the dataset. It measures the model’s ability to detect positive examples. A high recall indicates that the model can effectively capture positive examples. These metrics are important for evaluating the performance of a model, as they provide insights into its predictive accuracy, its ability to identify positive examples, and its ability to capture all relevant positive examples. Depending on the specific objectives and requirements of the research or application, different metrics may be given more importance.

To put it simply, the accuracy rate assesses the overall predictive capability of the model, while the precision rate and recall rate evaluate the model’s predictive capability and detection ability for positive examples, respectively. Typically, a trade-off exists between precision and recall, meaning that as we increase precision, recall tends to decrease, and vice versa. The practical applications select different metrics based on the specific problem at hand. For instance, in a medical diagnosis problem, our focus is on detecting actual patients, which emphasizes recall. On the other hand, in a credit card fraud problem, our priority is detecting fraud cases, placing greater importance on precision. When dealing with class imbalance, where there is a significant difference in the number of positive and negative examples, relying solely on accuracy may yield misleading outcomes. For instance, consider the prediction of tumor patients. If only 1% of the samples in the training set are actual tumor patients, a model that simply predicts all samples as non-tumor would still achieve 99% accuracy. However, such a model would not be useful in practice, as it fails to detect any actual tumor cases. This highlights the limitations of relying solely on accuracy when dealing with imbalanced datasets. On the other hand, when dealing with balanced classes, in which the number of positive and negative examples is equal, using accuracy can be a better choice. For instance, when predicting the bull–bear trend of the stock market, in which the probability of a bull or bear market is approximately 50:50, accuracy is a more suitable metric. It accurately reflects the model’s performance in predicting the correct bull–bear trend, taking into account both true positives and true negatives.

The evaluation of the research model’s performance is carried out by analyzing its accuracy and utilizing information from the confusion matrix. This information is obtained by calculating the rates of true positive (TP), true negative (TN), false positive (FP), and false negative (FN) results, based on the classifier’s observations. In this context, “positive” and “negative” pertain to the classifier’s predictions in relation to the observed data. Accuracy is determined by computing the ratio of accurate predictions to the total number of predictions made for bull and bear features. This calculation yields a measurement of how many correct predictions the model has made in relation to the entire set of predictions. This accuracy value can be expressed through the following equation:

$$\mathrm{Accuracy} \mathrm{Rate}=\frac{TP+TN}{TP+TN+FP+FN}$$

(1)

The selection of an appropriate evaluation metric depends on the class balance of the dataset. When there is a balanced distribution of positive and negative examples, the accuracy rate can better assess the model’s performance. However, when there is a significant class imbalance, relying solely on accuracy may yield misleading results. In such cases, alternative metrics, like precision, recall, or F1 score, should be considered. In the specific experiment, the random forest model achieved an accuracy rate of 80%, SVMs also achieved 80% accuracy, MLP achieved 76% accuracy, and deep learning achieved 66% accuracy (

Table 2. The accuracy of algorithms.

Algorithms	Random Forest	SVMs	MLP	Deep Learning
Accuracy	80%	80%	76%	66%

). These results provide an initial indication of the models’ performance, but further analysis and comparison of other metrics are necessary to make a comprehensive assessment.

A comparison between the baseline and the proposed method, as shown in

Table 3. The comparison of the baseline and proposed methods.

Futures Corpus	Random Forest	SVMs	MLP	Deep Learning	Average
√	80%	80%	76%	66%	75.50%
×	74%	80%	66%	56%	69%

, reveals that the majority of algorithms exhibit significantly improved performance with the proposed method, incorporating the futures corpus. The baseline method achieves an average accuracy of only 69%, which is much lower compared to the proposed method’s accuracy of 75.5%.

5. Discussion

The results suggest that random forest and SVMs performed relatively well, while deep learning exhibited a poorer performance. There are several potential reasons for this disparity. One possible cause is the small sample size of the dataset. When the dataset is small, models can easily be overfitting, meaning that they become too specialized in capturing the idiosyncrasies of the training data and fail to generalize well to unseen data. This can lead to significant differences in performance between different models. To address this issue, one option is to increase the size of the training dataset. Collecting more data can help improve the model’s ability to learn patterns and make more accurate predictions. Alternatively, techniques such as cross-validation can be employed to mitigate overfitting and more effectively assess the model’s generalization performance. By addressing the small sample size issue and employing appropriate techniques for model evaluation, it is possible to obtain more reliable and meaningful results that accurately reflect the performance of different models.

The quality of the data and the selection of relevant features can significantly impact the performance of a model. Poor data quality, such as missing values or incorrect labels, can lead to inaccurate predictions. Similarly, inadequate feature selection may include irrelevant or redundant features, hindering the model’s ability to learn meaningful patterns. To address these issues, it is crucial to carefully evaluate the data source and its quality. Cleaning the dataset by handling missing values, outliers, or inconsistencies can greatly improve the model’s performance. Additionally, employing effective feature selection techniques, such as statistical analysis or domain knowledge, can help identify the most informative features for the prediction task. Moreover, the model’s hyper parameters, such as the maximum depth of decision trees or regularization coefficients in SVMs, play a vital role in performance. Optimal hyper parameter selection is crucial for achieving the best results. Techniques such as grid search or random search can be used to systematically explore different hyper parameter combinations and identify the ones that yield the highest performance, based on evaluation metrics. By addressing data quality issues, refining feature selection, and fine-tuning the model’s hyper parameters, the performance of the models can be enhanced, resulting in more accurate predictions.

While some algorithms, like SVMs, show comparable accuracy between the proposed method and the baseline, there are factors contributing to this observation. One reason could be the inherent nature of SVMs, which map data to a higher-dimensional space and seek an optimal hyperplane that maximizes the margin between different classes, resulting in good generalization capabilities. Another factor could be the limitation of the sample size, which may not adequately reflect the differences between the two methods. Nonetheless, overall, the proposed method consistently outperforms the baseline method for the majority of algorithms, underscoring its significant contribution in this study.

In summary, achieving optimal model performance requires careful consideration and the adjustment of multiple factors. The quality and quantity of data, feature selection, hyper parameters, and inherent assumptions of the model all contribute to its performance. By conducting a thorough analysis and making appropriate adjustments to these factors, the model can be optimized to achieve the best possible performance. This iterative process of evaluation and refinement is crucial in order to ensure accurate and reliable predictions.

6. Conclusions

Financial articles contain complex and diverse information that plays a crucial role in investment decision making. When compared to the baseline, the majority of algorithms show significant improvement using the proposed method, which incorporates the futures corpus. The baseline achieves an average accuracy of only 69%, whereas the proposed method achieves a much higher accuracy of 75.5%. The proposed bull–bear futures text detection method effectively extracts valuable insights from financial articles, enabling quick identification of market trends. Machine learning models, such as random forest and support vector machines (SVMs), achieve an impressive accuracy rate of 80%. By utilizing these models, investors can receive real-time bull–bear advice, adapt to market changes, and develop effective trading strategies. The method in this study greatly reduces reading time, facilitates faster decision making, and enhances the likelihood of investment success. In future research, this approach can be expanded to other domains, further improving model designs to enhance predictive capabilities and practicality.