A stock price trend prediction method based on multi-order dynamic graph fusion

Abstract

This invention discloses a stock price trend prediction method based on multi-level dynamic graph fusion, relating to the fields of financial technology and artificial intelligence data analysis. The method mainly comprises three parts: multi-channel temporal feature extraction, collaborative modeling of macro and micro spatial structures, and multi-level feature fusion prediction. The steps include: First, constructing a multi-channel technical indicator sequence based on historical stock trading data, and using a multi-channel attention pooling GRU network to extract differentiated temporal features in parallel; Second, constructing a stock association hypergraph based on industry and concept labels, introducing graph information loss (GIL) as a feedback signal to drive an adaptive hyperedge reconstruction mechanism, dynamically adjusting hyperedge weights to filter noise and capture macro market hotspots; Simultaneously, constructing a simple graph within the hyperedge and combining contrastive learning constraints to semantically align stocks with similar technical patterns to enhance the discriminative power of micro local features; Finally, integrating macro hypergraph features and micro simple graph features through a cross-graph fusion module, concatenating them with temporal features, and inputting them into the prediction layer to complete the prediction of the stock price trend at the next moment. This method overcomes the limitations of existing static graph models in dynamically capturing high-order stock correlations and fine-grained technical pattern resonances, providing a new approach for financial spatiotemporal data mining.

Description

Technical Field

[0001] This invention relates to the field of financial technology and artificial intelligence data analysis, and more specifically to a method for predicting stock price trends based on multi-level dynamic graph fusion. Background Technology

[0002] The stock market is a typical complex dynamic system. Its price fluctuations are not only influenced by the historical trends of individual stocks, but also by higher-order cross-stock correlations such as industry sectors and thematic concepts. To capture and model these complex market patterns, early research primarily relied on statistical methods (such as ARIMA) or deep learning-based recurrent neural networks (such as LSTM and GRU) for time series analysis. While these methods can capture nonlinear long-term and short-term dependencies, they typically treat each stock as an independent entity, ignoring the pervasive spatial dependencies between stocks. To address this issue, graph neural networks (GNNs) and their variants (such as GCNs and GATs) were introduced to characterize binary synergistic effects between stocks. Subsequently, hypergraph neural networks (HGCNs), capable of expressing complex higher-order dependencies such as "one-to-many" or "many-to-many," have also become a research hotspot, utilizing hyperedges to model industry or thematic relationships.

[0003] However, existing related technical solutions still have significant limitations in practical applications. First, most existing hypergraph methods rely on predefined static structures or simple dynamic update mechanisms, making the models unstable when faced with rapid shifts in market hotspots and frequent changes in concept themes, failing to respond promptly to market dynamics. Second, existing models often treat hyperedges as fully connected sets, easily introducing redundant information and noise, and lack fine-grained modeling of the resonance of local technical patterns within hyperedges, making it difficult to effectively align macro-level sector information with micro-level individual stock technical patterns. Furthermore, in traditional models, temporal information is often simply spliced ​​into the spatial structure. This shallow spatiotemporal fusion method prevents temporal embedding from continuously influencing decisions during cross-graph fusion, limiting the model's ability to capture the linkage between macro and micro levels. Therefore, a new technical solution is urgently needed that can jointly model temporal dynamics and higher-order correlations between stocks, and possess structural adaptability, to address these problems. Summary of the Invention

[0004] In view of this, the present invention provides a stock price trend prediction method based on multi-level dynamic graph fusion to address the problems of existing stock price prediction models, such as lagging in capturing market hotspot rotation, lack of characterization of micro-technical pattern resonance, and shallow fusion of spatiotemporal features. This is of great significance for improving the accuracy of financial time-series data mining and assisting quantitative investment decision-making.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a stock price trend prediction method based on multi-level dynamic graph fusion, the specific steps of which include the following:

[0006] Data Input and Feature Construction Module: Acquire historical trading data of the target stock market and external industry and concept tag data; construct a multi-channel stock feature sequence containing multiple technical indicators, and construct an industry concept correlation tensor reflecting the relationship between the stock and its sector;

[0007] Multi-channel temporal coding module: The multi-channel stock feature sequence is input into the multi-channel attention pooling GRU network, and trend, volatility and momentum features are extracted through parallel channels respectively. The channel attention mechanism is used to adaptively fuse multi-source signals to generate a temporal embedded representation of the stock.

[0008] Macro Hypergraph Adaptive Reconstruction Module: Based on the industry concept association tensor, an initial stock association hypergraph is constructed. Graph Information Loss (GIL) is introduced as a feedback signal to evaluate the magnitude of spatial structure correction to temporal features. The hyperedge weights are dynamically adjusted by calculating hyperedge sensitivity, thereby reconstructing a dynamic hypergraph structure that can reflect current market hotspots.

[0009] Microscopic Hyperedge In-line Contrast Learning Module: Construct a simple graph inside each hyperedge of the hypergraph, aggregate neighbor information using graph convolutional networks, and combine contrastive learning loss function to constrain the node representation inside the hyperedge, forcing stocks with similar technical patterns to align in the feature space and generating microscopic spatial features;

[0010] Spatiotemporal deep fusion and prediction module: Through the cross-graph fusion layer, the macroscopic hypergraph features that have been dynamically reconstructed and the microscopic simple graph features that have been optimized through comparative learning are nonlinearly integrated to form a unified spatial representation. This representation is then concatenated with the temporal embedding representation and input into the prediction layer to calculate the probability of stock price change at the next moment.

[0011] 2. Further, the stock price trend prediction method based on multi-level dynamic graph fusion is characterized in that the temporal attention pooling calculation formula in the multi-channel temporal coding module is: , in, For the first Channel at time The hidden state, For learnable parameters, This is channel-level time embedding.

[0012] 3. Further, the stock price trend prediction method based on multi-order dynamic graph fusion is characterized in that the formula for calculating the graph information loss (GIL) in the macro hypergraph adaptive reconstruction module is: ,in, The original temporal feature distribution of the nodes. The feature distribution after propagation via the hypergraph. Represents relative entropy.

[0013] 4. Further, the stock price trend prediction method based on multi-order dynamic graph fusion is characterized in that the update basis of the hyperedge weight in the macro hypergraph adaptive reconstruction module is the hyperedge sensitivity, and its calculation formula is as follows: in, The total loss of the model, For super-edge The gating factor. When Reduce weights to suppress noise when Weights are added to strengthen effective connections.

[0014] 5. Further, the stock price trend prediction method based on multi-order dynamic graph fusion is characterized in that the loss function formula used by the microscopic hyperedge in-circle contrast learning module is: in, These are positive sample pairs with similar technical forms within the super-edge. This is the temperature coefficient.

[0015] 6. Further, the stock price trend prediction method based on multi-level dynamic graph fusion is characterized in that the prediction formula of the spatiotemporal deep fusion and prediction module is: ,in, For the final timing embedding, The unified space representation after merging is defined by ||, which represents the splicing operation.

[0016] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a stock price trend prediction method based on multi-level dynamic graph fusion. The beneficial effects of the present invention are as follows:

[0017] 1. This invention addresses the problem of existing hypergraph models having static structures and being unable to adapt to rapid shifts in market trends by proposing an adaptive hyperedge reconstruction mechanism driven by graph information loss. This mechanism transforms "information distortion" into feedback signals, dynamically filtering and reconstructing the optimal sector topology, effectively filtering structural noise in the market, and significantly improving the model's response speed to sudden hot topics and sector rotations.

[0018] 2. While handling high-order correlations, this invention innovatively constructs a "simple graph within a hyperedge + comparative learning" framework. Compared to traditional methods that only focus on macro-industry correlations, this invention achieves deep alignment between macro-sector linkages and micro-individual stock technical resonances by establishing fine-grained technical form consistency constraints within the hyperedge, thereby enhancing the model's discriminative ability in local dynamics.

[0019] 3. This invention designs a multi-channel spatiotemporal collaborative coding and cross-graph fusion mechanism, breaking the limitation of traditional models that only perform shallow splicing of temporal information. By extracting trend, fluctuation, and momentum features in parallel through multi-channel attention pooling, and deeply fusing unified spatial representations and temporal embeddings before the final prediction layer, it ensures the continuous influence of temporal evolution laws in the decision-making process, thereby effectively controlling drawdown risk while ensuring high returns. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the process of a stock price trend prediction method based on multi-level dynamic graph fusion. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] This invention discloses a stock price trend prediction method based on multi-level dynamic graph fusion, the specific steps of which include the following:

[0024] Step 1: Data Input and Feature Construction. Obtain historical trading data of the target stock market and external industry and concept tag data. For each stock... Construct a multi-channel stock feature sequence that includes multiple technical indicators. This embodiment selects the original closing price and four derived technical indicators calculated from it: Moving Average (MA), Relative Strength Index (RSI), Moving Average Convergence Divergence (MACD), and Bollinger Band Width (BBW). These five technical indicators form a single-day feature vector. Simultaneously, based on the GICS industry classification standard and the Wikidata knowledge graph, an industry concept correlation tensor reflecting the relationship between the stock's sector affiliation is constructed. and For example, when a stock is on a trading day Belongs to a certain industry When the condition is met, the corresponding element of the tensor is set to 1; otherwise, it is set to 0.

[0025] Step Two: Multi-channel Temporal Encoding. The constructed multi-channel stock feature sequence is input into a multi-channel attention-pooling GRU network. First, based on semantics, the features are divided into three channels: trend, volatility, and momentum. These are then fed into independent GRU layers for parallel encoding to obtain the hidden state sequence. Subsequently, a temporal attention pooling (TAP) mechanism is introduced to weightedly pool the hidden states in the time series, calculated using the following formula: , .

[0026] Step 3: Adaptive Reconstruction of the Macro Hypergraph. An initial stock association hypergraph is constructed based on the industry concept association tensor, where each industry or concept is considered a hyperedge. To address the issue of static structures failing to adapt to market hotspot rotation, Graph Information Loss (GIL) is introduced as a feedback signal. GIL calculates the original temporal feature distribution of nodes. Feature distribution after hypergraph propagation The relative entropy (KL divergence) between them is used to assess the magnitude of the correction of temporal features by spatial structure: .

[0027] Based on this, the total loss of the calculation model is affected by the hyperedge gating factor. The partial derivatives are used as hyperedge sensitivity. And adjust the hyperedge weights dynamically accordingly. When the sensitivity When the connection is strong, the weight of the hyperedge is increased to strengthen effective connections; conversely, the weight is decreased to suppress noise. Through this mechanism, a dynamic hypergraph structure that reflects current market trends can be reconstructed.

[0028] Step 4: Microscopic Hyperedge In-line Contrast Learning. A simple graph is further constructed within each hyperedge of the macroscopic hypergraph to capture microscopic-level technological morphological resonances. First, the cosine similarity of technological features between node pairs within the hyperedge is calculated, and a sparse adjacency matrix is ​​constructed based on an adaptive threshold. Next, a graph convolutional network (GCN) is used to aggregate neighbor information. To enhance the discriminative power of node representations, a contrastive learning loss function is introduced. Node representation inside the constrained hyperedge: .

[0029] Step 5: Spatiotemporal Deep Fusion and Prediction. The dynamically reconstructed macroscopic hypergraph features are then fused through a cross-graph fusion layer. Compared with microscopic simple graph features optimized through comparative learning The data is spliced ​​together and integrated through nonlinear layers to form a unified spatial representation. Finally, it is combined with the aforementioned temporal embedding representation. The data is then concatenated, and the prediction layer calculates the probability of the stock price change at the next time step. If the predicted probability If the value is positive, it is considered an increase; otherwise, it is considered a decrease.

[0030] The present invention proposes a stock price trend prediction method based on multi-level dynamic graph fusion, which solves the limitations of existing models in balancing macro-sector and micro-stock information and adapting to rapid shifts in market hotspots. The method has a clear logic and is easy to implement.

[0031] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A stock price trend prediction method based on multi-level dynamic graph fusion, characterized in that, The specific steps include the following: Data Input and Initial Feature Construction Module: Acquire historical trading data of the target stock market and external industry and concept label data, construct a multi-channel stock feature sequence containing multiple technical indicators, and construct an industry concept correlation tensor reflecting the relationship between the stock's sector based on industry and concept classification; Multi-channel temporal coding module: The multi-channel stock feature sequence is input into the multi-channel attention pooling GRU network. Trend, volatility and momentum features are extracted through parallel channels, and time step information is aggregated using the temporal attention pooling mechanism. Finally, the multi-source signals are adaptively fused through the channel attention mechanism to generate the temporal embedding representation of the stock. Macro Hypergraph Adaptive Reconstruction Module: Based on the industry concept association tensor, an initial stock association hypergraph is constructed. Graph Information Loss (GIL) is introduced as a feedback signal to evaluate the magnitude of spatial structure correction to temporal features. The hyperedge weights are dynamically adjusted by calculating hyperedge sensitivity, thereby reconstructing a dynamic hypergraph structure that can reflect current market hotspots. Microscopic Hyperedge In-line Contrast Learning Module: Construct a simple graph inside each hyperedge of the hypergraph, aggregate neighbor information using graph convolutional networks, and combine contrastive learning loss function to constrain the node representation inside the hyperedge, forcing stocks with similar technical patterns to align in the feature space and generating microscopic spatial features; Spatiotemporal deep fusion and prediction module: Through the cross-graph fusion layer, the macroscopic hypergraph features that have been dynamically reconstructed and the microscopic simple graph features that have been optimized through comparative learning are nonlinearly integrated to form a unified spatial representation. This representation is then concatenated with the temporal embedding representation and input into the prediction layer to calculate the probability of stock price change at the next moment.

2. The stock price trend prediction method based on multi-level dynamic graph fusion according to claim 1, characterized in that, The formula for calculating the temporal attention pooling in the multi-channel temporal coding module is as follows: , .

3. The stock price trend prediction method based on multi-level dynamic graph fusion according to claim 1, characterized in that, The formula for calculating the graph information loss (GIL) in the macroscopic hypergraph adaptive reconstruction module is as follows: .

4. The stock price trend prediction method based on multi-level dynamic graph fusion according to claim 1, characterized in that, The update of hyperedge weights in the macroscopic hypergraph adaptive reconstruction module is based on hyperedge sensitivity, and its calculation formula is denoted as: .

5. The stock price trend prediction method based on multi-level dynamic graph fusion according to claim 1, characterized in that, The contrastive learning loss function formula used in the microscopic hyperedge interior contrastive learning module is as follows: .

6. The stock price trend prediction method based on multi-level dynamic graph fusion according to claim 1, characterized in that, The prediction formula of the spatiotemporal depth fusion and prediction module is: .

Images