Time series data prediction task based on Bayesian optimization

Time series data prediction has important application value in the fields of finance, meteorology, and industrial control. However, traditional prediction models often use fixed hyperparameters or manual adjustment parameters, and model selection consumes a lot of labor costs and time costs, which limits the performance of the model. Aiming at the problems of high computational cost and slow convergence speed of existing grid search and random search methods, this paper proposes a CNN-LSTM-Attention prediction framework based on Bayesian optimization. Through the Gaussian process surrogate model, the consumption of labor costs is effectively reduced. The experiment compares the optimization effects of mainstream models such as LSTM and Transformer on the stock price datasets. The results show that compared with the traditional optimization method, the RMSE and MAPE indexes of the proposed method are reduced by at least 0.041 and 0.00011 respectively, and the convergence speed is increased by 2-3 times, which confirms the effectiveness and necessity of the algorithm optimization.