Combination Model Of Extremum Guiding Learning For Prediction Of The Financial Time Series Direction

Financial market contains many objective rules,which appear as financial time series.It has played an important role in the development of national economy and investors obtaining benefit.Transactions in any kind of financial markets mainly include spot transactions and future transactions.The futures market has a good effect on the healthy and sustainable development of the enterprise because of its functions of risk aversion,price discovery and hedging.Moreover,in virtue of future market's two-way trading feature,it has greater research value.In this paper,the futures market prediction was chosen as the research problem,and the domestic commodities futures index was taken as the samples.In order to obtain benefit as much as possible and improve the prediction effect,a two-stage combination forecasting model was built to predict and trace the whole time domain direction.In the first stage,this paper compared a series of typical time series forecasting and regression methods to choose one of them as the best direction prediction model,which is used to fit the commodities futures and obtain an effective fitting curve.In the second stage,the support vector machine method was used to enhance the prediction effect and predict the direction of important observation points,which can further strengthen the prediction effect on the basis of the first stage.Based on the sufficient research on the basis of predecessors' achievements,this paper originally completed the following several aspects:(1)According to the imperfections in direction judgment of the typical regression method,a rolling regression comparison method was put forward to improve the prediction effect during the study of the first stage.Because using typical regression method to predict direction was too lag,in this paper,on each point of time series,a regression model was built using reasonable and aptotic length of the time series in order to get the fitting value.Then,this paper compared the adjacent fitting value to predict the direction.There was more than 60 percent probability that using this method to find the turning point was earlier than the typical regression methods.(2)This paper proposed an extremum guiding learning strategy to improve the effect of machine learning during the study of the second stage.According to this strategy,the training set and prediction set were selected differently,which could provide more reasonable and correct training data for machine learning to enhance the prediction effect of the support vector machine.Specifically,this paper studied the partial time series between the extremum and the turning points of the fitting curve,and selected the appropriate points from the partial time series as observation points of the training set.For the prediction set,all the turning points of the fitting curve were part of the observation points of the prediction set directly.On the basis of this,other effective types of turning points were added as the observation points of the prediction set.(3)Some effective characteristic vectors were discovered and designed during the modeling process of the second stage.On the basis of the first stage model,a classification model was built as the second stage model to predict the direction of the observation points of the prediction set using the support vector machine.This paper designed some effective characteristic vectors to describe the observation points of training set,including grid feature,super slope mark and grid difference.Moreover,several prediction plans were established according to the timespan of the training set in order to avoid randomicity of the result and the result of every plan is better than the first stage model.Innovation of this paper lies in:this pater put forward the rolling regression comparison method to improve the imperfections in the direction of judgment of the typical regression method;proposed the extremum guided learning strategy to improve the effect of machine learning,which is used to select observation points of training set and prediction set differently;designed and described effective characteristic vector of the observation points.