This paper proposes a two-stage self-supervised pretraining modeling method for stock price sequence prediction in financial markets. The method is designed to address challenges such as limited labeled data, complex structural patterns, and non-stationary temporal features. The framework consists of two phases: pretraining and fine-tuning. In the pretraining phase, two self-supervised tasks are constructed. One captures long-term trends, while the other models short-term fluctuations. In the fine-tuning phase, the learned representations are used for regression prediction to improve the model's ability to fit future price movements. In the encoder design, the method integrates multi-layer temporal sequence modeling units. This enables multi-granularity semantic extraction and structure-aware representation learning. For the experimental part, a dataset is built based on Tesla's historical stock data from 2010 to 2024. The model is systematically evaluated under different time windows, hidden dimensions, sampling frequencies, and perturbation settings. The experimental results show that the proposed method outperforms existing baseline models across multiple metrics. It effectively captures temporal dependencies while maintaining strong prediction stability and robustness. This study validates the effectiveness of the two-stage architecture in financial time series modeling. It also demonstrates the practical potential of self-supervised learning in low-supervision financial prediction tasks.

This paper proposes an anomaly detection algorithm based on large language models guided by knowledge graphs. The method combines the deep semantic understanding capability of large language models with the structural modeling advantages of knowledge graphs, forming a semantic-structural co-driven fraud detection framework. Specifically, the model first encodes transaction texts using a pretrained language model to generate semantic representations. Meanwhile, a financial knowledge graph is constructed based on entities such as transaction accounts, devices, and IP addresses, from which structural embeddings are obtained via graph neural networks. Next, a gated fusion mechanism is introduced to adaptively integrate the semantic and structural vectors. The fused representation is then fed into a classifier to produce fraud prediction results. In multiple comparative experiments, the proposed model outperforms single-modality approaches across metrics including accuracy, precision, recall, and F1-score, validating its effectiveness in complex financial scenarios. In addition, this study conducts multi-task adaptation experiments and anomaly ratio variation tests to evaluate the model's stability and robustness. The results show that the proposed method maintains strong performance across different environments, demonstrating its reliability and practical value.

This paper proposes a stock price prediction method based on a graph neural network architecture. The method is designed to address key characteristics of the stock market, including high nonlinearity, multivariable dependencies, and dynamically changing structural relationships. It constructs a dynamic stock graph to represent the evolving relationship network among individual stocks over time. A temporal-aware graph neural network module is designed to jointly model node features through structural propagation and temporal dependence. Specifically, the model incorporates multi-source heterogeneous information to build the dynamic graph structure. This enables explicit representation of the time-varying linkages between stocks within the graph. Graph convolution is then applied to extract structural features at each time step. A temporal module is used to model the evolution of these features over time. To validate the effectiveness of the method, the model is compared with existing graph-based and time-series models across multiple evaluation metrics. Ablation studies, robustness tests, and performance assessments under different market conditions are conducted to comprehensively analyze the model's behavior in various scenarios. Experimental results show that the proposed method achieves low prediction error while maintaining strong stability and generalization ability. It significantly improves the accuracy of modeling asset price trends in financial markets. This study provides a unified solution for structural and dynamic aspects of the stock prediction problem and extends the application scope of graph neural networks in financial time series analysis.

This paper proposes a stock price prediction model based on a combination method of recurrent neural networks (RNN) and attention mechanisms. The goal is to improve the forecasting accuracy of financial time series data. By applying the attention mechanism to the RNN architecture, the model can learn dynamic importance weights for different time steps. This allows better capture of dominant features in stock price volatility. The study utilizes real Apple Inc.'s stock prices between January 2024 and January 2025 to build models and validate the models. Normalizing and implementing train-test strategies were used to preprocess data in a bid to enhance stability within the models. Experiments utilize mean squared error (MSE), mean absolute error (MAE), and the coefficient of determination (R2) as evaluation metrics for the model. Comparative evaluation was performed against a set of mainstream popular models. Results show that the constructed RNN+Attention model outperforms traditional RNN, LSTM, MLP, and Transformer models both in prediction accuracy and fitting performance. In addition, visualizations demonstrate a strong correspondence between predicted values and actual values. The curve of training loss shows a clear downward trend, further supporting the effectiveness and convergence of the model. This method provides an efficient and realistic technical solution to stock price prediction using deep learning.

This paper proposes a user risk classification model that combines dynamic graph neural networks with a self-supervised pretraining mechanism. The model is designed to address the dual challenges of dynamic user behavior and label scarcity in financial scenarios. First, a time-series graph is constructed to represent user transaction behaviors across different time slices. A dynamic graph convolution module is then applied to extract evolutionary features. Next, a node-level contrastive self-supervised task is designed. By constructing positive and negative node pairs, the model adaptively optimizes the representation space. This enhances the discriminative power of node embeddings during the unsupervised stage. Finally, the pretrained node representations are fed into a multi-layer perceptron for the downstream risk classification task. An empirical analysis is conducted on a publicly available credit card fraud detection dataset. The results show that the proposed model outperforms existing dynamic graph models in terms of accuracy, recall, and F1-score. It demonstrates a stronger ability in identifying high-risk user behavior, validating the effectiveness and robustness of the approach.

This paper proposes a deep learning model for systemic risk identification and prediction in financial markets, integrating multi-scale convolutional structures, attention mechanisms, and residual connections. The model is designed to capture dynamic features at different temporal resolutions, enhance focus on key risk factors, and maintain training stability in complex financial data environments. Using the US Systemic Risk Dataset released by the Federal Reserve, the model transforms the task into a binary classification problem based on crisis-related indicators. All time series data are standardized, interpolated, and structured through sliding windows to preserve temporal continuity. Experimental comparisons show that the proposed model outperforms traditional methods such as MLP, CNN, Transformer, and BERT across accuracy, precision, and recall. Further analysis reveals the effectiveness of multi-scale and attention mechanisms in modeling complex dependencies and detecting early risk signals. Sensitivity experiments on convolutional kernel size demonstrate that integrating cross-feature information significantly improves performance, with 3x3 and 5x3 kernels showing optimal results. The training process exhibits fast convergence and low loss stability, confirming the model's robustness and efficiency. These findings validate the technical advantages of the proposed architecture in addressing the challenges of systemic risk modeling in high-dimensional and heterogeneous financial time series.