Research On Neural Network Optimization And Stock Index Prediction Based On Swarm Intelligence Algorithm

The stock price index,which measures and reflects the overall trend of the stock market,is the "barometer" of economic development.It can accurately reflect the basic trend of economic operation,and provide the basis for the government to regulate the market.It's also an important reference indicator for investors to judge the economic prospects,and provide guidance for investors to make reasonable investment decisions.Therefore,accurate stock index prediction has great significance in great theory and reality.This paper takes the modeling of stock index prediction and the empirical analysis of stock index prediction model as the research topic.Each stock index prediction model includes data preprocessing,feature extraction,feature selection and stock index prediction model based on neural network.Comprehensive learning particle swarm optimizer is an improved particle swarm optimizer,which can solve the premature convergence problem of particle swarm optimizer.Therefore,based on the research of comprehensive learning particle swarm optimizer,this paper proposes four kinds of improved comprehensive learning particle swarm optimizer,and applies it to the feature selection and neural network parameter optimization in three stock index prediction models,in order to obtain a better feature subset and a higher precision of stock index prediction model.The specific research content and innovation points of this paper are as follows:(1)Stock index prediction model ?: Firstly,based on stochastic perturbation operation and simulated annealing algorithm,a hybrid algorithm of integrated comprehensive learning particle swarm optimizer and simulated annealing algorithm is proposed.This method can expand the exploration area of individual particle optimal solution,and further improve the local search ability of the algorithm and the convergence precision of the algorithm in the later iteration of the algorithm.To solve the problem of slow convergence speed in the late iteration of comprehensive learning particle swarm optimizer.Through the tests of various standard functions,the hybrid algorithm proposed in this paper significantly improves the accuracy of standard functions.Secondly,in the stock index prediction model based on neural network,the comprehensive learning particle swarm optimizer is used to optimize the weights and thresholds of neural network.The neural network hyperparameters and the lengths of input features are optimized by a binary particle swarm optimizer.Finally,the optimal feature subset and the feature set composed of all the features obtained by the feature selection algorithm based on random forest were used as inputs for four neural network-based stock index prediction models.The results show that the index prediction model composed of the feature selection algorithm based on random forest and the improved comprehensive learning particle swarm optimizer has the highest accuracy.(2)Stock index prediction model ?: First,a feature selection algorithm based on multi-objective binary particle swarm optimizer and neural network is proposed.The multi-objective binary particle swarm optimizer is used for feature selection and neural network hyperparameter optimization simultaneously.The results show that compared with the binary particle swarm feature selection algorithm and the mutation binary particle swarm feature selection algorithm,the multi-objective optimization algorithm can effectively reduce the complexity of neural network and the number of optimal features.Secondly,based on the difference mutation and quasi-newton method,this paper proposes a hybrid algorithm composed of comprehensive learning particle swarm optimizer and differential evolution algorithm.This method can enhance the historical global exploring ability of the optimal solutions for the particles in the early stage of its evolution,and improve the convergence precision of the particles in the later iterations of algorithm.Also,it solves slow convergence speed in the later iterations of comprehensive learning particle swarm optimizer.Through the comparison of various standard functions,the proposed comprehensive learning particle swarm optimizer is significantly improved in the aspect of function optimization accuracy.On this basis,the comprehensive learning particle swarm optimizer is used to optimize the weights and thresholds of the neural network,and the hyperparameters and the lengths of input features for the neural network are optimized by binary particle swarm optimizer.Finally,the two optimal feature subsets derived from the multi-objective binary particle swarm feature selection algorithm and the binary particle swarm feature selection algorithm are used as inputs in four neural network stock index prediction models respectively.The results show that neural network index prediction model ? composed of the feature selection algorithm based on neural network and multi-objective binary particle swarm optimizer and neural network optimized by the improved comprehensive learning particle swarm optimization hybrid algorithm have the highest prediction accuracy.(3)Stock index prediction model ?: first of all,the feature selection model based on the improved multi-objective comprehensive learning particle swarm optimizer and support vector regression is proposed.Compared to multi-objective comprehensive learning particle swarm optimizer,the research results show that feature selection based on the imporved multi-objective comprehensive learning particle swarm optimizer can effectively reduce the number of features in the optimal feature subset.Secondly,multi-swarm comprehensive learning particle swarm optimizer based on cauchy mutation and gauss mutation is proposed.It introduces the global exploration subpopulation and local search subpopulation to balance global exploring ability and local search ability of the algorithm.Cauchy mutation strategy and gauss mutation strategy are introduced into these two kinds of subpopulation respectively for particles falled into the stagnation,and it helps the particles escape from the local optimal solution and improves the population diversity.Finally,the two optimal feature subsets derived from the two feature selection algorithms are used as inputs for four neural network-based stock index prediction models respectively.The results show that stock index prediction model ? composed of the feature selection model based on the improved multi-objective particle swarm optimizer and the neural network optimized by the improved multi-swarm comprehensive learning particle swarm optimizer has the highest prediction accuracy.(4)It can be seen from the test results of the three stock index prediction models that the stock index prediction model ? is the best model for the Shanghai and Shenzhen 300 index and the FTSE 100 index predictions.The stock index prediction model ? is the best model for the S&P500 index prediction.The stock index prediction model ? is the best model for the N225 index prediction.