Hierarchical and asynchronous dynamic hypergraph network for stock trend prediction

Orthogonal factor-based biclustering algorithm (BCBOF) for high-dimensional data and its application in stock trend prediction

As a novel tool for predicting stock trends, hypergraphs are used to effectively represent high-order relationships among stocks, capturing symmetric dependencies inherent in market interactions. However, the instability of hyperedges limits their ability to capture dynamic stock changes, and existing methods neglect the influence of time decay on feature importance. To address these challenges, a hybrid hypergraph–dynamic graph attention network based on temporal decay attention and conditional aggregation for stock trend prediction, namely HDGAN, is developed. Specifically, we utilize dynamic graphs to capture the dynamic relationships among stocks, which mitigates the instability of the hyperedge structure in dynamic markets. A temporal decay attention mechanism is designed to identify important feature points in the evolution of stock prices, and then a conditional aggregation method is proposed to aggregate information from different pathways. Extensive experiments on A-share, NASDAQ, and NYSE datasets demonstrate HDGAN outperforms other state-of-the-art methods in stock trend prediction and investment return.

Unifying relational and market dynamics to enhance stock trend prediction

Stock trend prediction aims to classify trading decisions into buy, hold, and sell signals with the goal of ensuring traders maximize profits. With the application of deep learning models and their efficiency across multiple tasks, there is an increasing research direction into their applications in stock trend prediction. However, most of these models are not trained to understand complex market structures and are sometimes oversimplified, making them not very useful in real-world trading performance. This study proposes a meta-adaptive multi-horizon forecasting model that integrates historical price data with domain-specific structural features to improve market trend prediction across 1, 3, 7, and 10-day horizons. The model utilizes fair value gaps (FVG) derived via a deterministic three-bar imbalance rule that captures liquidity voids, inefficiency zones, and institutional trading points. These features are also augmented by order block attributes and rolling-window hurst exponents that capture fractal persistence and market transitions in a principled form. The proposed model employs a variable selection network for dynamic soft-gating over the input set, and a multi-scale projection module that transforms each sequence using learned pooled-up sampling before fusing them through trainable attention weights. This fused representation is then passed through the feature-wise linear modulation block, which is adjusted based on a global context vector that enables horizon-specific adaptability. The temporal dependencies are modeled with a gated attention unit encoder and a gated residual decoder that classifies the buy, hold, and sell labels across different future horizons. The model was evaluated on eight indices, namely BSE, DJIA, FTSE, NASDAQ, NIFTY, N225, S&P500, and SSE, using time series cross-validation with <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i> = 3, 5, and 7 fold settings. The results clearly indicate that the proposed model outperforms existing benchmark models in terms of average accuracy, precision, recall, and MCC over all evaluated indices, with values of 94.97%, 95.33%, 95.07%, and 0.9250, respectively. A real-time trading simulation assesses the practical impact of the window size, considering 5, 10, and 15 days, with subsequent statistical validation through a paired t-test and Wilcoxon signed-rank test against benchmarks. Conclusively, this study provides valuable insights and demonstrates how both short and long-term traders can effectively leverage the proposed model using historical price data, market cues, and smart money concepts to accurately classify stock price movement.

This paper investigates the application of long short-term memory (LSTM), one-dimensional convolutional neural networks (1D CNN), and logistic regression (LR), for predicting stock trends based on fundamental analysis. This research emphasises a company's financial statements and its intrinsic value for stock price trend forecasting. Using a dataset of 269 data points from publicly traded companies across various sectors from 2019 to 2023, we employ key financial ratios and the discounted cash flow (DCF) model for two tasks: annual stock price difference (ASPD) and difference between current stock price and intrinsic value (DCSPIV). Assessing the likelihood of profitability from relationship between financial data and price action, and the current discrepancy between 'true value' and market price, respectively. Our results demonstrate that LR models outperform CNN and LSTM models, achieving an average test accuracy of 74.66% for ASPD and 72.85% for DCSPIV, highlighting the benefits for portfolio managers in their decision-making processes.

Stock trend prediction is a complex and crucial task due to the dynamic and nonlinear nature of stock price movements. Traditional models struggle to capture the non-stationary and volatile characteristics of financial time series. To address this challenge, we propose the Hybrid Relational Transformer (HRformer), which specifically decomposes time series into multiple components, enabling more accurate modeling of both short-term and long-term dependencies in stock data. The HRformer mainly comprises three key modules: the Multi-Component Decomposition Layer, the Component-wise Temporal Encoder (CTE), and the Inter-Stock Correlation Attention (ISCA). Our approach first employs the Multi-Component Decomposition Layer to decompose the stock sequence into trend, cyclic, and volatility components, each of which is independently modeled by the CTE to capture distinct temporal dynamics. These component representations are then adaptively integrated through the Adaptive Multi-Component Integration (AMCI) mechanism, which dynamically fuses their information. The fused output is subsequently refined by the ISCA module to incorporate inter-stock correlations, leading to more accurate and robust predictions. Extensive experiments on the NASDAQ100 and CSI300 datasets demonstrate that HRformer consistently outperforms state-of-the-art methods, e.g., achieving about 0.83% higher Accuracy and 1.78% higher F1-score than TDformer on NASDAQ100, with Sharpe Ratios of 1.5354 on NASDAQ100 and 0.5398 on CSI300, especially in volatile market conditions. Backtesting results validate its practical utility in real-world trading scenarios, showing its potential to enhance investment decisions and portfolio performance.

Robust stock trend prediction via volatility detection and hierarchical multi-relational hypergraph attention

Abstract: The increasing complexity and volatility of financial markets demand intelligent, technology-driven systems that enable accurate forecasting, risk management, and data-driven decision-making. This paper presents the Nifty50 Decision Tool, a Decision Support System (DSS) designed for stock trend prediction and portfolio analysis. The system integrates two roles—Users and Administrators—to streamline stock selection, forecasting, and decision recommendations. Investors can select stocks from the Nifty50 index (e.g., Reliance, Infosys), visualize technical indicators such as Moving Averages, RSI, MACD, Bollinger Bands, and VWAP, and receive short-term forecasts generated using ARIMA and LSTM models. Administrators leverage the backend to manage data acquisition, validate forecasts, and store user analysis history in SQLite for continuous improvement. The tool combines linear and non-linear forecasting approaches with technical analysis, providing users with actionable insights such as Buy, Hold, or Sell signals. Built with HTML, CSS, and JavaScript for the frontend, Flask APIs for backend, Streamlit for visualization, and SQLite for data storage, the system emphasizes lightweight deployment and scalability. Experimental evaluation demonstrates reliable short-term forecasting, effective visualization, and improved decision accuracy, making the Nifty50 Decision Tool a practical DSS for both academic study and real-world financial decision support. Keywords: Stock Market, Decision Support System, Nifty50, ARIMA, LSTM, Technical Indicators, Forecasting, Portfolio Analysis, Financial Analytics.

Leveraging cosine similarity for stock trend prediction with options chain data and piecewise linear encoding

Stock trend prediction has long been an important research direction in the financial field, and it is also an extremely challenging task. Currently, most studies focus on a single prediction model to find a better prediction scheme by comparing the effects of different algorithms.This paper proposes a stock trend prediction method based on a Blending ensemble learning approach, which combines 55 technical indicators such as Exponential Moving Averages (EMA) and Relative Strength Index (RSI). PCA dimensionality reduction is used to further simplify the data representation of the features after SOM dimensionality reduction. The method employs two high-performing machine learning models with distinct algorithmic characteristics as base learners and Logistic Regression as the meta-learner to construct an efficient ensemble prediction framework. Using Apple Inc.'s stock (AAPL) as the research subject, the study utilises the confusion matrix as the core performance evaluation metric. Experimental results demonstrate that optimised through hyperparameter tuning. Experimental results indicate that the Blending ensemble learning model, optimized through hyperparameter tuning, outperforms the single prediction models in terms of accuracy.

MA-MSTNet: mixed attention-based multi-scale temporal network for stock trend prediction

In stock market prediction, using news headlines to anticipate stock trends has become increasingly important. Analyzing sentiment from these headlines makes it possible to predict the stock price trends of the targeted company and profit from the resulting trades. This study examines the impact of company-related news headlines on stock price trends. The objectives of this study are as follows: First, we propose a multi-attention network that incorporates the strength of long short-term memory (LSTM) and bidirectional encoder representations from transformers (BERT) to model domain-specific and domain-general features in news headlines to predict the stock price trend of companies. Second, the proposed model can model and evaluate the effect of news on the stock price trend of different companies. Third, we construct the Bursa Malaysia news headline dataset and automatically align headlines with target companies and their stock price trend. This study proposes that the LSTM WITH ATTENTION +BERT model should use domain-specific and domain-general features to predict stock price trends using news headlines. The proposed model is compared to several convention models and deep learning models. The LSTM WITH ATTENTION +BERT model achieved an accuracy of 50.68%, showing notable improvements over other approaches. It surpassed the Decision Tree by 11.2%, Naïve Bayes by 20.13%, and Support Vector Machine by 5.12%. Compared to the CNN, LSTM, and BERT models, the proposed model is 4.27%, 2.91% and 1.64% higher, respectively, in terms of accuracy. These results highlight the strength of the proposed model.

For a selected portfolio of large-cap blue-chip stocks in China A-share market, this study selects and quantifies three categories of textual information with comparatively notably low average daily volume: responses from Secretaries of the Boards of Listed Companies (RSB), Comments by Internet Influencers on Listed Companies (CII), and Official Press Releases of High-level Meetings of the Communist Party of China and Central Government (R-M-P&amp;G). Then, for every category of textual information, the predictive role of a Gaussian Kernel Learning-Support Vector Machine(GKL-SVM) model, employing the information independently, is explored and assessed on stock (price) trends. Based on a comparative analysis of A-share market transaction data for this investment portfolio over the past three years, the GKL-SVM model, individually utilizing any type of textual information, demonstrates some enhancement in predicting short-term trends of Daily Closing (DC) stock prices compared to the coin-toss benchmark. These improvements further enable the corresponding Intraday Decision-making Short-term (IDS) trading strategy to achieve positive average daily returns in a simulated trading environment. Furthermore, by innovatively adapting the Multi-Kernel Learning (MKL) applied in Support Vector Machine (SVM) models into a Multi-Gauss-Kernel Learning (MGKL) framework consisting of the three GKs employed, respectively, for the above three different categories of textual information, i.e. RSB, CII, and R-M-P&amp;G, the corresponding MGKL-SVM model is constructed for DC stock (price) trend prediction. This novel modeling approach integrating multiple categories of textual information not only significantly reduces the volume of textual data required and the computational complexity for intelligent analysis but also achieves a 10[Formula: see text] percentage point improvement in prediction accuracy compared to the three single-category (text-processing) GKL-SVM models (termed RSB, CII, and R-M-P&amp;G (category) GKL-SVM models). The simulated trading results further demonstrate that the IDS trading strategy generates superior average daily returns relative to its single-category GKL-SVM counterparts. Finally, applying the MGKL-SVM model to Weekly Closing (WC) stock (price) trend prediction achieves significantly higher accuracy than daily prediction, indicating its potential practical value for medium-to-long-term value investing.

MLSC: A Multi-label Stock Classifier for Multi-horizon Stock Trend Prediction

In response to the challenges of multimodal data integration, real-time information retrieval, model hallucination, and lack of interpretability in financial stock analysis, this paper proposes an innovative financial analysis framework—FSframe. It aims to address multiple challenges in stock analysis within the financial sector. The framework integrates various technological modules to provide comprehensive and efficient solutions for stock trend prediction and financial question answering tasks. First, FSframe optimizes large language models (LLMs), enhancing their adaptability to financial tasks, and incorporates prompt engineering to mitigate potential hallucination issues during the generation process, thereby improving the accuracy and reliability of the analysis. Secondly, the framework introduces Retrieval-Augmented Generation (RAG) technology, creating a dynamically updated financial knowledge base that enables the model to retrieve and integrate the latest market data, providing real-time external knowledge support for tasks. Furthermore, FSframe adopts a sparse attention mechanism, optimizing the processing efficiency of time-series data by filtering irrelevant information and focusing on key points, while also achieving efficient integration of time-series and textual data. Finally, through its modular design, FSframe organically combines the aforementioned advanced technologies, forming an innovative solution that blends multimodal data processing with real-time analysis, offering strong technical support for intelligent analysis in the financial sector. Validation on large-scale financial datasets (including historical stock prices, financial news, and market announcements) shows that FSframe significantly improves prediction accuracy and real-time responsiveness in stock trend forecasting and financial question answering tasks. Experimental results indicate that FSframe offers significant advantages in multimodal data integration, real-time performance, and interpretability, demonstrating excellent task adaptability and addressing the shortcomings of traditional methods. The FSframe framework not only provides an innovative solution for stock analysis in the financial sector but also opens new pathways for the development of intelligent financial technologies.

Wavelet Denoising and Double-Layer Feature Selection for Stock Trend Prediction

Leveraging multi-time-span sequences and feature correlations for improved stock trend prediction

Stock constitutes a crucial element of the financial market, and accurately forecasting stock trends remains a significant and unresolved issue. Nonetheless, the stock’s considerable complexity renders accurate prediction of stock trends more challenging. This paper proposes a novel multi-perspective approach that converts the time series prediction challenge into an image classification problem, referred to as the Multi-perspective Denoise Transformer (MPDTransformer). We initially multi-factor features into two-dimensional images employing a multi-perspective approach to more comprehensively explain the actual market conditions and enhance the model’s practicality and adaptability; secondly, we utilize a Convolutional Autoencoder (CAE) to extract features, which effectively eliminates noise and enhances data purity; finally, to comprehensively capture the temporal relationships within the data and gain a deeper understanding of the overall time series, we employ a Transformer for prediction. Experimental results demonstrate that our method outperforms other prevalent stock trend prediction techniques.

Abstract Due to the dynamic of stock markets, predicting stock price trends remains a massive challenge when utilizing machine learning models, especially in Bayesian models. Indeed, the distributions of input features extracted from stock datasets are seldom normal, making Gaussian density less discriminative. This paper introduces a novel predictive model, the Bayesian Classifier with Heuristic Likelihood Estimation (BC-HLE), in which heuristic likelihood estimation instead of the Gaussian density was utilized to release the normality assumption in the conventional naïve Bayes classifier. Our model leverages the concept of the p-value, which evaluates how close the testing value is to the expected value on a distribution. This approach yields a more accurate likelihood estimate without normality assumption. When tested on 55 stock datasets from the S&amp;P500 index, the proposed BC-HLE model outperformed conventional Gaussian classifiers and such machine learning models as support vector machines and multi-layer perceptron, regarding prediction accuracy. Additionally, its superiority was verified on the returns on investment when the stock trend prediction was applied to a simulated trading system. The experimental outcomes show that the proposed model is a reliable enhancement of Bayesian classifiers and can contribute to decision for the stock market investment.