Paper Reading | NGAT: Long-Term Stock Prediction Using Graph Attention Networks

Paper link: arxiv.org/pdf/2507.0201... Code link: github.com/FreddieNIU/N... Graph representation learning methods have been widely applied in the financial domain, enhancing company representations through inter-firm relationships. However, existing approaches face three major challenges: (1) the advantages of relational information are constrained by downstream task design; (2) graph models specifically designed for stock prediction often suffer from poor generalizability due to high computational complexity; and (3) empirically constructed firm-relation graphs lack effective comparative evaluation across different graph structures. To address these issues, this paper proposes a long-term stock prediction task and develops NGAT—a node-level Graph Attention Network specifically tailored for firm-relation graphs—while experimentally demonstrating the limitations of existing graph comparison methods that rely solely on downstream task performance. Experimental results on two datasets consistently validate the effectiveness of both the proposed task and model. 1 Background Graph Neural Networks (GNNs) learn contextual embeddings reflecting graph structure via message-passing mechanisms that aggregate neighborhood nodes, establishing themselves as the mainstream approach in graph representation learning. This success has spurred their application in stock prediction—modeling firms as nodes and inter-firm relationships as edges to capture momentum spillover effects. Existing studies typically follow a paradigm of “constructing firm-relation graphs → applying GNNs to aggregate node embeddings → evaluating on downstream tasks”, yet face the following challenges: Limitations of short-term prediction: Most existing work focuses on next-day trend prediction (a one-day forecasting horizon), failing to fully exploit the GNN’s capacity to capture long-term momentum spillover (which may span multiple days); Insufficient model adaptability: General-purpose GNNs (e.g., GCN, GAT) assume homogeneous node feature distributions, whereas nodes (firms) in firm-relation graphs exhibit distinct behavioral patterns, rendering shared attention mechanisms inappropriate; Lack of graph construction evaluation: Firm-relation graphs are built empirically, with no effective method for comparing the quality of different graph structures based on downstream task performance. 2 Problem Definition This paper aims to address the following core problems: 1. Modeling long-term momentum spillover: Existing short-term predictions (e.g., next-day) cannot adequately leverage the lagged effect of inter-firm momentum spillover (potentially spanning multiple days) and neglect volatility—a key risk indicator; 2. Adaptability to firm-relation graphs: Shared attention mechanisms in general-purpose GNNs are unsuitable for firm-relation graphs where node (firm) feature distributions are highly heterogeneous; 3. Evaluation of graph construction methods: There is a lack of effective methods to compare graph structure quality based on downstream task performance. 3 Method 3.1 Task Definition: Long-Term Stock Prediction Return Mean Prediction: Predict whether the average return over the next T trading days exceeds that over the past T trading days. The label is defined as where is the log return of stock s on day d, and p_{d}^{s} denotes the adjusted closing price. Labels are categorized into four scenarios: surge (LP, N+, positive return in prior period followed by higher positive return), rebound (LN, N+, negative return in prior period followed by positive return), pullback (LP, N−, positive return in prior period followed by negative return), and crash (LN, N−, negative return in prior period followed by even lower return). Volatility Prediction: Predict the standard deviation of returns over the next T trading days, defined as 3.2 Model Architecture: NGAT

NP-DNN: A DNN-Based Stock Market Price Prediction Method Outperforming LLMs with 99.21% Accuracy

Original article: mp.weixin.qq.com/s/Twgh... Stock price prediction is an important yet challenging topic in economics due to the market’s inherent random fluctuations. Conventional statistical time-series forecasting methods often fail to effectively predict the probabilistic range of future stock prices. This paper proposes NP-DNN—a stock price prediction method that applies z-score normalization to standardize stock price data and employs a Multi-Layer Perceptron (MLP) to learn complex nonlinear relationships among stock market prices. Compared with other methods integrating Large Language Models (LLMs), NP-DNN achieves an accuracy of 99.21%. Abstract Stock market price prediction is a significant interdisciplinary research area relying on the intersection of finance, statistics, and economics. Conventional statistical time-series forecasting methods often fail to effectively predict the probabilistic range of future stock prices. This paper proposes NP-DNN to predict stock prices, using z-score normalization to standardize stock price data and an MLP to learn complex nonlinear relationships among stock market prices. Compared with other methods integrating Large Language Models (LLMs), NP-DNN achieves an accuracy of 99.21%. Introduction Stock price prediction is an important yet challenging topic in economics due to the market’s inherent random fluctuations. Prior to investment, data analysis is essential; large volumes of stock price data significantly influence investment decisions, and stock price movements are affected by factors such as supply and demand. Existing methods struggle to cope with complex market dynamics, whereas AI, machine learning (ML), and deep learning (DL) techniques can capture intricate relationships and yield more accurate predictions. This study predicts stock prices through the following components: · Neural Prophet captures historical patterns and handles irregularly spaced time series, thereby improving the quality of inputs fed into the Deep Neural Network (DNN); · Deep Neural Networks learn complex nonlinear relationships to achieve accurate price predictions; · Multi-Layer Perceptron (MLP) performs feature extraction to enhance prediction performance and provides higher-quality inputs for the DNN. Methodology This study employs NP-DNN to predict stock prices based on extracted features and hyperparameter optimization via Optuna. Using the Crunchbase dataset as input, preprocessing techniques—including z-score normalization and missing-value imputation—are applied to ensure model consistency. An MLP learns underlying patterns to generate feature representations. Figure 1 illustrates the stock price prediction process. Dataset The Crunchbase dataset is a publicly available database containing LinkedIn profile data accessible to researchers. The platform provides business and professional information for both public and private companies. Experiments were conducted on this dataset in September 2021. Data are categorized into three types: organizations, people, and investments. Additionally, the dataset includes samples of companies that have left LinkedIn profiles on Crunchbase. Preprocessing Preprocessing the Crunchbase data enhances the quality of raw numerical data and improves the accuracy and value of financial information. Z-score standardization is applied to normalize the data, facilitating the identification of financial statistical regularities and patterns, and supporting analyses such as stock market prediction. A two-stage approach addresses data irregularity: first, outliers are removed using the pandas library; second, temporal interpolation is performed via linear interpolation to estimate missing values between data points.

Pentair plc. (PNR) Stock Forecasts

Oct 2, 2026

Argus Research lowered its price target for Pentair plc. to $51.00, while the current stock price stands at $53.03.

Yahoo Finance Australia

2026-10-02

Pentair plc. (PNR) Stock Forecasts

2026-10-20

Barclays sets USD 136 target for PPG Industries stock

Pentair plc. (PNR) Stock Forecasts

FAQ