Abstract:Spatial-temporal graph neural networks (STGNNs) have achieved significant success in various time series forecasting tasks. However, due to the lack of explicit and fixed spatial relationships in stock prediction tasks, many STGNNs fail to perform effectively in this domain. While some STGNNs learn spatial relationships from time series, they often lack comprehensiveness. Research indicates that modeling time series using feature changes as tokens reveals entirely different information compared to using time steps as tokens. To more comprehensively extract dynamic spatial information from stock data, we propose a Double-Path Adaptive-correlation Spatial-Temporal Inverted Transformer (DPA-STIFormer). DPA-STIFormer models each node via continuous changes in features as tokens and introduces a Double Direction Self-adaptation Fusion mechanism. This mechanism decomposes node encoding into temporal and feature representations, simultaneously extracting different spatial correlations from a double path approach, and proposes a Double-path gating mechanism to fuse these two types of correlation information. Experiments conducted on four stock market datasets demonstrate state-of-the-art results, validating the model's superior capability in uncovering latent temporal-correlation patterns.

What problem does this paper attempt to address?

The problems that this paper attempts to solve are as follows: Existing spatio - temporal graph neural networks (STGNNs) and Transformer - based time - series prediction models perform poorly in stock market prediction, mainly due to the lack of effective modeling of the dynamic spatial relationships between nodes in the stock market. Specifically: 1. **Modeling time - series independently**: Many methods only focus on the time - series of a single node and ignore the interdependencies between different time - series. However, in the stock market, the entire market operates as a whole, and the time - series of different stocks have significant interactions at any time. 2. **Limitations of feature and time - step modeling**: For Transformer - based methods, using time - steps as tokens to model time - series often leads to poor performance because the characteristics of time - series are usually reflected in the changes of features rather than the information content of each time - step itself. 3. **Insufficient predefined spatial relationships**: Most spatio - temporal graph neural networks require predefined adjacency relationships (such as spatial distances in climate data or road connections in traffic networks), but these relationships do not clearly exist in the stock market. Although some ST - Transformers do not require predefined correlation relationships, they still learn the correlations between nodes from the perspective of time - steps, and experiments show that the correlations learned in this way are limited. To solve these problems, the paper proposes a new model - **Double - Path Adaptive - correlation Spatial - Temporal Inverted Transformer (DPA - STIFormer)** - aiming to more comprehensively extract the dynamic spatial information in stock data. This model improves existing methods in the following ways: - **Double - path adaptive - correlation mechanism**: Introduces a double - path adaptive - correlation block (DPABlock) to model the correlations between nodes from the feature and time dimensions respectively, and fuses these two types of correlation information through a gating mechanism. - **Inverted time - series modeling**: By regarding the feature changes of time - series as tokens instead of regarding time - steps as tokens, it can better capture the dynamic characteristics of time - series. - **Importance - weight mechanism**: Introduces an importance - weight mechanism to ensure that the influence degree of different features on nodes can be considered when calculating attention. Through these improvements, DPA - STIFormer can achieve state - of - the - art results on four stock market datasets, verifying its superior ability in revealing potential spatio - temporal correlation patterns.

Double-Path Adaptive-correlation Spatial-Temporal Inverted Transformer for Stock Time Series Forecasting

TCGPN: Temporal-Correlation Graph Pre-trained Network for Stock Forecasting

ST-Trader: A Spatial-Temporal Deep Neural Network for Modeling Stock Market Movement

HATR-I: Hierarchical Adaptive Temporal Relational Interaction for Stock Trend Prediction

Hierarchical Adaptive Temporal-Relational Modeling for Stock Trend Prediction

Spatial-Temporal Convolutional Transformer Network for Multivariate Time Series Forecasting

Temporal and Heterogeneous Graph Neural Network for Financial Time Series Prediction

DyGraphformer: Transformer combining dynamic spatio-temporal graph network for multivariate time series forecasting

Spatiotemporal Transformer for Stock Movement Prediction

Spatial correlation learning based on graph neural network for medium-term wind power forecasting

Stock Market Forecasting Based on Spatiotemporal Deep Learning

Multi-Channel Temporal Graph Convolutional Network for Stock Return Prediction

Deep transition network with gating mechanism for multivariate time series forecasting

MASTER: Market-Guided Stock Transformer for Stock Price Forecasting

Temporal-Relational hypergraph tri-Attention networks for stock trend prediction

Stock Price Prediction Using Temporal Graph Model with Value Chain Data

Spectral Temporal Graph Neural Network for Multivariate Time-series Forecasting

Price graphs: Utilizing the structural information of financial time series for stock prediction

Pre-training Enhanced Spatial-temporal Graph Neural Network for Multivariate Time Series Forecasting

Deep Convolutional Transformer Network for Stock Movement Prediction

Spatiotemporal Adaptive Neural Network for Long-term Forecasting of Financial Time Series