A Hybrid Deep Learning Model Based on INFO-TCN- iTransformer and Its Effectiveness in Stock Price Prediction

Abstract Amid profound financial market transformations and rising uncertainty, accurately forecasting stock price volatility is essential for risk management and the high-quality development of capital markets. Stock price series often exhibit high noise, strong non-stationarity, and multi-scale fluctuation patterns, challenging traditional time series or single deep learning models to capture both local volatility and long-term dependencies. This study proposes an INFO-TCN-iTransformer framework that integrates Temporal Convolutional Networks (TCN) for extracting short- and medium-term local volatility features and an improved Transformer (iTransformer) for modeling long-term dependencies across multivariate time series, enabling effective fusion of local and global information. A Vector-weighted Average Optimization (INFO) algorithm is employed to globally optimize key hyperparameters, improving training stability and predictive accuracy. Empirical results on Chinese stock data demonstrate that the proposed model significantly outperforms baseline methods in prediction accuracy, stability, and generalization. The findings indicate that multi-scale feature learning combined with intelligent hyperparameter optimization provides an effective approach for stock price forecasting and supports quantitative investment and financial risk management.