Stock price prediction with attentive temporal convolution-based generative adversarial network

Abstract

Stock price prediction presents significant challenges owing to the highly volatile and nonlinear nature of financial markets, which are influenced by various factors including macroeconomic conditions, policy changes, and market sentiment. Traditional prediction models such as ARIMA and classic linear regression models are often inadequate for capturing the complex dynamics of stock prices. The advent of deep learning has led to substantial improvements in prediction accuracy, with various recurrent neural networks widely employed for representation learning from stock sequences. However, recurrent networks such as LSTM and GRU may exhibit susceptibility to overfitting the training data, leading to suboptimal performance in real-world predictions due to the inherent noise and volatility of stock market data. Recent research has demonstrated that temporal convolutional networks (TCN) exhibit impressive capabilities in stock price prediction. A TCN can achieve extensive sequence memory by utilizing dilated convolutions, enabling it to capture long-term dependencies in time-series data, as well as causal convolution, ensuring that the model does not utilize future information when predicting future values, which is particularly crucial for time-series prediction. Nevertheless, stock market data typically contain substantial noise to which TCNs may be overly sensitive, thereby affecting the accuracy of the predictions. To address this issue, we propose a novel stock price prediction method based on the Generative Adversarial Networks (GANs) framework, utilizing an Attentive Temporal Convolutional Network (ATCN) as the generator, termed Attentive Temporal Convolution-based Generative Adversarial Network (ATCGAN). This approach employs a GAN framework to generate stock price data using an attentive temporal convolutional network as a generator, whereas a CNN-based discriminator evaluates the authenticity of the data. Adversarial training facilitates the model’s learning of the complex distribution of stock price data. Within the GAN framework, the TCN effectively captures long-term dependencies, combined with an attention mechanism for generating representative feature combinations, thereby enhancing prediction accuracy. Compared to the traditional ARIMA forecasting method, ACTGAN achieved a 78.29% reduction in Mean Absolute Error (MAE). Furthermore, when compared to the deep learning method GRU, ACTGAN reduced the Mean Absolute Error (MAE) by 51.01%. The experimental results demonstrate that the proposed GAN-based approach significantly outperforms the traditional methods and deep learning techniques.