Research And Optimization Of Time Series Prediction Based On Neural Network

Time series analysis,prediction and mining are important means of data analysis and modeling.They are widely used in the fields of meteorology,finance and engineering and have important research value.Stocks,as one of the most typical time series data,a barometer of financial markets,have drawn investors' attention since they emerged.Because of the stock's own long-term development trend,cyclical changes,seasonal changes and irregular movements,it is more complicated than other statistics.The continued growth of highly volatile and irregular data has created an urgent need to develop more efficient analytical methods that can extract meaningful statistical conclusions.As a process of exploring valuable hidden information,data mining has played a significant role in the analysis of financial time series.Therefore,data mining provides investors with forward-looking and knowledge-based decisions to help investors succeed in making profits with a smaller investment risk.In this context,the thesis studies the financial time series prediction algorithm based on neural network in detail.The main research contents and contributions are as follows:First,an extremist learning machine-based Computational Efficient Functional Link Artificial Neural Network(ELM-CEFLANN)model is proposed.The model can effectively reflect the nonlinear relationship between complex financial time series data through CEFLANN,can achieve a high prediction accuracy,but also can reduce the training time due to excessive parameters through ELM.In addition,a comprehensive performance evaluation system that considers both theoretical and practical aspects is proposed to help assess the effectiveness of the model.Finally,using the CSI 300 stock index,the model is verified with outstanding performance in all aspects,compared with other traditional machine learning algorithms.Second,a Particle Swarm Optimization-Online Sequential Extreme Learning Machine(PSO-OS-ELM)is applied to CEFLANN.For stock data in the actual application has to join the new data at any time,OS-ELM which only needs to be trained by the current batch data is put forward.According to the low prediction accuracy of OS-ELM applied to CEFLANN,PSO is proposed for optimization.Finally,the CSI-300 data is used to prove that the model's stability test accuracy is obviously better than that before optimization,and the model generalization ability is limited.Thirdly,SFLA-ELM algorithm using random frog leaping algorithm is applied to CEFLANN.Aiming at the problem that ELM input layer weight and hidden layer offset are randomly selected during initialization,which lead to over-reliance on data set and the structure is too swollen,an SFLA-ELM algorithm is proposed,which mainly uses SFLA to select the input weight of ELM algorithm and the hidden layer offset value.In order to reduce the prediction error generated by the model when different data sets are used,and to improve the generalization ability of the model.Experiments show that the model is more generalized,but the prediction accuracy and return on investment are less than PSO-ELM.