Research On SVM Stock Market Prediction Based On Outlier Feature Extract And Energy Calculation

With the improvement of financial market and investment philosophy in-depth, stock investment is accepted by more and more people. Because the stock market is a complex nonlinear system and affected by many factors, it lead to hard stock market prediction. Domestic and foreign scholars have made some meaningful exploration in the stock market prediction, and high accuracy prediction model can better guide the investors to reduce the risk of investment.This dissertation mainly aimed at the peak point in the stock market prediction is difficult to predict, and the peak point can lead to low accuracy prediction model. From the angle of index outliers to analyze abnormal phenomenon in the stock market, and based on SVM model as forecasting model, it can improve the accuracy of the prediction model. Research carried out in this article is as follows:(1) The stock price fluctuations have stronger mutation, caused it’s difficult to predict stock price movements. A Stock market volatility forecasting model based on characteristics of outliers pattern (SOFSVM) is presented. Firstly, SOFSVM algorithm utilizes Markov Blanket algorithm obtaining local network to shield the effects of other nodes to the target node; Secondly, Analyzing the index of the target node to extract characteristic of outliers pattern from the general behavior; then SOFSVM algorithm capture outlier features using sliding window, and put characteristic of outliers pattern into original SVM model as a prior knowledge. This method can predict peak point and smooth effect of peak point on the predicted results, it also can improve forecasting model robustness.(2) On the case that direction precision is imprecise when outliers happen in SOFSVM algorithm, this dissertation presents a stock market volatility forecasting model based on characteristic energy (EOF-SVM). Firstly, EOF-SVM algorithm based on multiple index outliers and special index type build a Bayesian network, it determine the probability of the stock market abnormal and analyze the risk of stock market. Then, indicator energy is defined according to the degree of support for each indicator and trend indicators, and a weighted combination of related indicators energy together form the total energy used to measure the impact of post-stock movements.Through the plate and the Shanghai composite index data contrast analysis was carried out on the two algorithms respectively, experiments show that prediction model established based on the impact of both stocks and indexes compared to a single prediction models have better effect, in the feature selection comparison with the basic feature selection algorithm has higher prediction accuracy.