Machine Learning In Finacial Data Forecasting Applications

Predicting stock prices is a highly complex and very difficult job,with too many factors such as political events,economic conditions,trader expectations and other environmental factors affecting stock prices.In addition,the stock price time series is usually full of noise,dynamic,nonlinear,complex,non-parametric,and confusing.The characteristics of noise-adding are reflected in the incompatibility of complete information,that is,the past behavior of financial markets does not fully capture all the information of future and past prices.Most studies focus on accurate predictions of stock prices.However,different investors tend to have different trading strategies.The prediction model is not necessarily suitable for them based on minimizing the error between actual and predicted values.Accurately predicting the direction of stock price movement is very important for effective market trading strategies.Artificial intelligence is a rapidly evolving technology in information processing applications.AI has been used in different fields,such as business,engineering,management,science,military and finance.Machine learning technology can imitate the volatile stock market,and can produce better forecast results in stock trend forecast than traditional methods.The machine learning algorithms include Support Vector Machine(SVM)?Naive Bayes Classifier(NBC)?Artificial Neutral Network(ANN)?Decision Tree?Random Forest(RF).This paper focuses on improvement of the algorithm of BP neural network and random forest in the prediction of stock price moving direction.The research content of this paper is as follows:In the prediction algorithm,the BP neural network model has become the most widely used model because of its excellent fault tolerance,powerful self-learning ability and nonlinear matching ability.However,the BP algorithm is based on the gradient descent method,especially relying on the initia weights is slower and tends to fall into local minimums.This makes the initial weight and threshold especially important.This paper proposes an adaptive mutation particle swarm optimization(APSO)algorithm to find the optimal BP neural network initial weight and threshold instead of random value,and combine with BP neural network to form a new The APSO-BP prediction model has higher prediction accuracy and does not fall into the local best and is more efficient.When the traditional random forest algorithm faces a large number of feature attributes,it cannot be effectively distinguished because different features may have opposite or repeated effects on the model.The invalid features lead to an increase in the prediction time of the prediction model and a decrease in the prediction accuracy.In order to solve this problem,this paper proposes a binary discrete particle swarm optimization algorithm BPSO combined with RF algorithm to form a new BPSO-RF prediction model algorithm.The BPSO algorithm selects the optimal feature combination,and finally inputs the stock data of the optimal feature combination into the RF prediction model to obtain the trend prediction result.The simulation is compared with the traditional RF algorithm from the perspective of prediction accuracy and pointing efficiency.The simulation experiment proves the validity and feasibility of the proposed BPSO-RF algorithm.