Stock Price Forecasting Model Based on ARIMA-TCN

In response to the limitations of traditional stock prediction methods such as ARIMA in stock price forecasting, a hybrid ARIMA-TCN model has been proposed. The model aims to leverage the strengths of ARIMA in processing linear data and TCN in capturing nonlinear characteristics and long-term dependencies to enhance the accuracy and stability of stock price predictions. Using data from Baidu and Coca-Cola stocks from January 4, 2016, to December 30, 2022, and January 4, 2021, to December 29, 2023, as samples, the model first uses ARIMA and TCN with optimized hyperparameters through Bayesian optimization for forecasting, then determines the weights of the two models’ predictions through Bayes optimization, ultimately obtaining a weighted average final prediction result. Empirical research has found that the hybrid model performs better on small datasets than on large ones, and compared to benchmark models such as ARIMA, TCN, and LSTM, the hybrid model shows a significant decrease in RMSE values on the aforementioned datasets. The experimental results verify that the proposed stock price prediction model not only improves accuracy and stability but also increases the model’s flexibility. However, the process of Bayesian optimization for TCN model hyperparameters and weight determination makes the computational cost of the hybrid model high on large-scale datasets.