An Ensemble Multi-Model Approach for Stock Return Forecasting Integrating Emotion-Driven Feature Extraction and Time Series Prediction Based on Variational Autoencoders (VAE) and Temporal Fusion Transformers (TFT)

บทคัดย่อ

This article aimed to study: 1) the impact of investor sentiment on stock returns in the automotive industry, and 2) the development and validation of the Adaptive Sentiment-Enhanced Temporal Regression Model (ASTRM), an ensemble model integrating Variational Autoencoders (VAE), Temporal Fusion Transformers (TFT), and Ordinary Least Squares (OLS) regression. This quantitative research utilizes historical secondary data from 269 companies in the Chinese automotive industry. Data was collected from the China Stock Market & Accounting Research (CSMAR) database, comprising historical stock prices and a sentiment index derived from social media posts from May 23, 2019, to September 21, 2022. The statistical analysis involved: 1) using VAE for feature extraction and data denoising, 2) employing TFT with LSTM and multi-head attention mechanisms for time-series forecasting, and 3) applying OLS regression to optimize the final predictions. Model performance was evaluated using metrics such as accuracy and adjusted R-squared. The findings indicate that the ASTRM model achieved a prediction accuracy of 91.12%, significantly outperforming the traditional LSTM model (70.63%) and a model without denoising (77.48%). Furthermore, investor sentiment was found to have a statistically significant influence on stock returns (p-values < 0.001). The ASTRM also demonstrated faster convergence and a lower final loss value, enhancing overall predictive performance.