CN106203731A - A kind of stock trend forecasting method based on 2 dimension stream predictions - Google Patents

Abstract

The present invention relates to a kind of stock trend forecasting method based on 2 dimension stream predictions, it is characterised in that extraction historical data, be predicted on horizontal and vertical time point respectively, then the result of prediction merged and obtain final prediction data.The invention have the advantage that and can efficiently solve a step lag issues present in 1 dimension prediction.

Description

A kind of stock trend forecasting method based on 2-dimension stream prediction

Technical field

The present invention relates to a kind of method predicting stock trend.

Background technology

The unstability of securities market and the feature of randomness allow the stock price only predicting tomorrow also be a kind of challenge. The trend of stock market can be preferably estimated by feature set outstanding, favourable structure.Furthermore, when we establish correctly Model obtain when being difficult to the attribute observed of the trend that is continually changing, our predictive ability will obtain raising.In the recent period, Many cities, especially in some big cities, the people carrying out equity investment and purchase is continuously increased, therefore, stock trend prediction skill Art seems increasingly important.

Carrying out the popular method of Prediction of Stock Index at present is to use binary event model.Build on the basis of this model Vertical feature set is better anticipated the future trend of stock market.Bayes and support vector machine is such as used to be predicted, There is higher forecasting accuracy and speed.Another method is theoretical according to Kind of Nonlinear Dynamical System, constructs a personal share Ticket transaction data model, and integrating parallel neutral net carries out e-learning, extracts mode standard, carries out pattern recognition, thus Stock market trend is carried out model prediction.

Said method common problem is: belonging to 1-and tie up prediction algorithm, major defect is that to produce a step delayed Problem, i.e. is predicted producing bigger error after the time point of the raw acute variation of stock certificate data miscarriage.Great majority at present Prediction algorithm all there is such problem, especially complicated at stock certificate data flow or when there is acute variation.

Summary of the invention

Present invention aim to address the step lag issues occurred in 1-dimension stock trend prediction algorithm.

In order to achieve the above object, the technical scheme is that providing a kind of dimension based on 2-flows the stock trend predicted Forecasting Methodology, it is characterised in that comprise the following steps:

Step 1, the stock historical data gathering horizontal dimensions and the stock historical data of vertical latitude, wherein, level is tieed up The stock historical data of degree is by the stock historical data in stock historical data base being entered according to the time period that every day is different Row statistics obtains, the stock historical data of vertical latitude be by by stock historical data in stock historical data base according to The statistics of same time period the most on the same day obtains；

Step 2, stock historical data based on horizontal dimensions, use nonlinear model carry out data modeling, obtain for The nonlinear data model of prediction level dimension stock certificate data；

Stock historical data based on vertical latitude, uses linear model to carry out data modeling, obtains for predicting vertical The linear data model of latitude stock certificate data；

Step 3, obtain the final predictive value P of time t_t, comprise the following steps:

Step 3.1, the stock certificate data collecting horizontal dimensions and the stock certificate data of vertical latitude, unbalanced input respectively Level dimension predictive value x ' is obtained after data model and linear data model_tWith vertical dimension predictive value y '_t, meanwhile, by predictive value p_t ^* It is initialized as 0, weight w is initialized as 0；

Step 3.2, by formula p_t=p_t ^*+WX′_t+(1-w)y′_tIt is calculated intermediate predictor p_t；

Step 3.3, by intermediate predictor p_tValue give predictive value p_t ^*, w is updated to w+A, A ＜ 0.1, returns step 3.2, Until calculating n times intermediate predictor p_tValue, enter step 3.4；

Step 3.4, it is calculated final predictive value P_t, P_t=p_t/N。

Preferably, in described step 2, described nonlinear model uses wavelet neural network.

Preferably, stock historical data X of described horizontal dimensions is expressed as:

X={f (t, d) | t=1,2 ..., AT/ Δ t}={x_t| t=1,2 ..., Δ T/ Δ t} (1)

Formula (₁In), (t, is d) the stock historical data of t time period in d days to f, and Δ T is diurnal periodicity, Δ t For time interval；

Then utilize described wavelet neural network to carry out data modeling to comprise the following steps:

The dimension data on flows normalization of step 2A.1, level:

If x_tFor the flow of the upper corresponding moment t of level dimension, then it can be normalized to

$\stackrel{\‾}{X_t}=(V_{\max }-V_{\min })*(X_t-X_{\min })/(X_{\max }-X_{\min })+V_{\min }---\left(2\right)$

In formula (2), V_max=1, Vmin=-1, X_maxMaximum stock certificate data, x in expression level dimension_minExpression level is tieed up Middle minimum stock certificate data；

Step 2A.2, determining the nodes of neutral net every layer, wavelet neural network has input layer, hidden layer and output Layer, if N is input layer number, H is node in hidden layer, and M is output layer nodes, then node in hidden layerIn formula, M=1；

Step 2A.3, training wavelet neural network:

By constantly regulation input layer and the weights of hidden layer, hidden layer and output layer, and the small echo letter for training The coefficient of number, wavelet neural network is finally reached convergence, obtains described nonlinear data model.

Preferably, in described step 2, described linear model uses autoregression integration moving average model ARIMA method.

Preferably, stock historical data Y of described vertical latitude is expressed as:

Y={f (t, d) | d=1,2 ..., Ly}={y_j| j=1,2 ..., Ly} (3)

In formula (1), (t, is d) the stock historical data of t time period in d days to f, and Ly is maximum number of days, then Utilize described Regression-Integral moving average model ARIMA method to carry out data modeling to comprise the following steps:

Step 2B.1, the stock historical data { y of described vertical latitude_j| j=1,2 ..., Ly} is through d rank differential transformation Rear formation stationary time series { Δ y_j| j:1,2 ..., Ly-d}, then have:

In formula (4),It is autocorrelation coefficient, θ_i, i=1,2 ..., q, is partial correlation system Number；

Step 2B.2, determined by the autocorrelogram and partial autocorrelation figure checking formula (4) model ARIMA (p, d, q) in Autoregression item p and rolling average item number q；

Step 2B.3, obtain autocorrelation coefficient by parameter estimationAnd partial correlation coefficient θ_i。

The present invention ties up Forecasting Methodology by proposing a kind of 2-, enterprising from laterally (level dimension) and longitudinal (vertical dimension) time point Row prediction stock stream prediction, then the result merging of bidimensional prediction is obtained final prediction data, 1-dimension can be efficiently solved A step lag issues present in prediction.

Accompanying drawing explanation

Fig. 1 is the algorithm flow chart of the present invention；

Fig. 2 is the stock certificate data schematic diagram of horizontal latitudinal；

Fig. 3 is the stock certificate data schematic diagram of vertical latitude.

Detailed description of the invention

Below in conjunction with specific embodiment, the present invention is expanded on further.Should be understood that these embodiments are merely to illustrate the present invention Rather than restriction the scope of the present invention.In addition, it is to be understood that after having read the content that the present invention lectures, people in the art The present invention can be made various changes or modifications by member, and these equivalent form of values fall within the application appended claims equally and limited Scope.

In conjunction with Fig. 1, the invention provides a kind of stock trend forecasting method based on 2-dimension stream prediction, including following step Rapid:

Step 1, the stock historical data gathering horizontal dimensions and the stock historical data of vertical latitude, wherein, level is tieed up The stock historical data of degree is by the stock historical data in stock historical data base being entered according to the time period that every day is different Row statistics obtains, the stock historical data of vertical latitude be by by stock historical data in stock historical data base according to The statistics of same time period the most on the same day obtains.

If Δ T and Δ t represents the time interval of diurnal periodicity and data sampling respectively.Here Δ T pro data, is 24 little Time, then the stock certificate data of t time period in d days can be expressed as f (t, d), wherein, t ∈ [1, Δ T/ Δ t], d ∈ [1, H], natural law total in being historical data base for h here.Such as f (3,5) represents the stock of the 3rd time period of the 5th day in historical data Data.

The feature of stock historical data can describe in terms of horizontal peacekeeping vertical dimension two.The data of horizontal dimensions are Carrying out statistics according to the change of different time every day to produce, the data of vertical dimension are then to enter according to the data of the most same time Row statistics produces.

Horizontal dimensions is according to the time period that every day is different, stock certificate data in historical data base to be carried out statistics obtain.Figure 2 is the stock certificate data distribution example of horizontal dimensions.Then stock historical data X of horizontal dimensions is expressed as:

X={f (t, d) | t=1,2 ..., Δ T/ Δ t}={xt | t=1,2 ..., Δ T/ Δ t}

The data of vertical dimensions represent the most on the same day at same temporal stock certificate data in historical data base, vertical latitude Stock historical data Y is expressed as:

Y={f (t, d) | d=1,2 ..., Ly}={y_j| j=1,2 ..., Ly}

Step 2, stock historical data based on horizontal dimensions, use nonlinear model carry out data modeling, obtain for The nonlinear data model of prediction level dimension stock certificate data.

The stock historical data of horizontal dimensions illustrate one day in the situation of change of each time period flow, characteristic is to comprise Less random violent data on flows change, is therefore suitable for using nonlinear model to carry out data modeling.The present embodiment uses Wavelet neural network carries out Kodaira dimension and it is predicted.Wavelet neural network combines the advantage of neutral net and wavelet analysis, Being widely used in Forecast of Nonlinear Time Series, concrete steps include:

The dimension data on flows normalization of step 2A.1, level:

If x_tFor the flow of the upper corresponding moment t of level dimension, then it can be normalized to

In formula, v_max=1, v_min=-1, X_maxMaximum stock certificate data, x in expression level dimension_minMinimum stock certificate data in expression level dimension；

Step 2A.2, determining the nodes of neutral net every layer, wavelet neural network has input layer, hidden layer and output Layer, if N is input layer number, H is node in hidden layer, and M is output layer nodes, then node in hidden layerIn formula, M=1；

Step 2A.3, training wavelet neural network:

By constantly regulation input layer and the weights of hidden layer, hidden layer and output layer, and the small echo letter for training The coefficient of number, wavelet neural network is finally reached convergence, obtains described nonlinear data model.

Stock historical data based on vertical latitude, uses linear model to carry out data modeling, obtains for predicting vertical The linear data model of latitude stock certificate data.

The stock historical data of vertical latitude is to flow statistics of variables in the same time the most on the same day, and the sampling period is one day, i.e. 24 hours, because the time cycle is longer, it is thus possible to can comprise change at random and the fluctuation of mass data, applicable employing was linearly built Mould method.The present embodiment uses autoregression integration moving average model ARIMA method to carry out vertical dimension data modeling and prediction. ARIMA is famous Time Series Forecasting Methods, and its main thought is the data sequence being elapsed in time by prediction object and being formed Row are considered as a random sequence, and with mathematical model ARIMA, (p, d, q) carry out this sequence of approximate description, and p is autoregression item here, q For rolling average item number, the difference order done when d becomes steady by time series.

Utilize described Regression-Integral moving average model ARIMA method to carry out data modeling to comprise the following steps:

Step 2B.1, the stock historical data { y of described vertical latitude_j| j=1,2 ..., Ly} is through d rank differential transformation Rear formation stationary time series { Δ y_j| j=1,2 ..., Ly-d}, then have:

In formula (1),It is autocorrelation coefficient, θ_i, i=1,2 ..., q, is partial correlation system Number；

Step 2B.2, determined by the autocorrelogram and partial autocorrelation figure checking formula (1) model ARIMA (p, d, q) in Autoregression item p and rolling average item number q；

Step 2B.3, obtain autocorrelation coefficient by parameter estimationAnd partial correlation coefficient θ_i。

Step 3, obtain the final predictive value P of time t_t, comprise the following steps:

Step 3.1, the stock certificate data collecting horizontal dimensions and the stock certificate data of vertical latitude, unbalanced input respectively Level dimension predictive value x ' is obtained after data model and linear data model_tWith vertical dimension predictive value y '_t, meanwhile, by predictive value p_t ^* It is initialized as 0, weight w is initialized as 0；

Step 3.2, by formula p_t=p_t ^*+w x′_t+(1-w)y′_tIt is calculated intermediate predictor p_t；

Step 3.3, by intermediate predictor p_tValue give predictive value p_t ^*, w is updated to w+A, A ＜ 0.1, returns step 3.2, Until calculating n times intermediate predictor p_tValue, enter step 3.4；

Step 3.4, it is calculated final predictive value P_t, P_t=p_t/N。

Claims (5)

1. a stock trend forecasting method based on 2-dimension stream prediction, it is characterised in that comprise the following steps:

Step 1, the stock historical data gathering horizontal dimensions and the stock historical data of vertical latitude, wherein, horizontal dimensions Stock historical data is by the stock historical data in stock historical data base being united according to the time period that every day is different Meter obtains, the stock historical data of vertical latitude be by by the stock historical data in stock historical data base according to difference The statistics of it same time period obtains；

Step 2, stock historical data based on horizontal dimensions, use nonlinear model to carry out data modeling, obtain for predicting The nonlinear data model of horizontal dimensions stock certificate data；

Stock historical data based on vertical latitude, uses linear model to carry out data modeling, obtains for predicting vertical latitude The linear data model of stock certificate data；

Step 3, obtain the final predictive value P of time t_t, comprise the following steps:

Step 3.1, the stock certificate data collecting horizontal dimensions and the stock certificate data of vertical latitude, unbalanced input data respectively Level dimension predictive value x ' is obtained after model and linear data model_tWith vertical dimension predictive value y '_t, meanwhile, by predictive value p_t ^*Initially Turn to 0, weight w is initialized as 0；

Step 3.2, by formula p_t=p_t ^*+w x′_t+(1-w)y′_tIt is calculated intermediate predictor p_t；

Step 3.3, by intermediate predictor p_tValue give predictive value p_t ^*, w is updated to w+A, A ＜ 0.1, returns step 3.2, until Calculate n times intermediate predictor p_tValue, enter step 3.4；

Step 3.4, it is calculated final predictive value P_t, P_t=p_t/N。

A kind of stock trend forecasting method based on 2-dimension stream prediction, it is characterised in that described In step 2, described nonlinear model uses wavelet neural network.

A kind of stock trend forecasting method based on 2-dimension stream prediction, it is characterised in that described water Stock historical data X of flat dimension is expressed as:

X={f (t, d) | t=1,2 ..., Δ T/ Δ t}={x_t| t=1,2 ..., Δ T/ Δ t} (1)

In formula (1), (t, is d) the stock historical data of t time period in d days to f, and Δ T is diurnal periodicity, when Δ t is Between be spaced；

Then utilize described wavelet neural network to carry out data modeling to comprise the following steps:

The dimension data on flows normalization of step 2A.1, level:

If x_tFor the flow of the upper corresponding moment t of level dimension, then it can be normalized to

${\stackrel{\‾}{x}}_t=(v_{\max }-v_{\min })*(x_t-x_{\min })/(x_{\max }-x_{\min })+v_{\min }---\left(2\right)$

In formula (2), v_max=1, v_min=-1, x_maxMaximum stock certificate data, x in expression level dimension_minIn expression level dimension Little stock certificate data；

Step 2A.2, determining the nodes of neutral net every layer, wavelet neural network has input layer, hidden layer and output Layer, if N is input layer number, H is node in hidden layer, and M is output layer nodes, then node in hidden layerIn formula, M=1；

Step 2A.3, training wavelet neural network:

By constantly regulating input layer and the weights of hidden layer, hidden layer and output layer, and be used for the wavelet function of training Coefficient, wavelet neural network is finally reached convergence, obtains described nonlinear data model.

A kind of stock trend forecasting method based on 2-dimension stream prediction, it is characterised in that described In step 2, described linear model uses autoregression integration moving average model ARIMA method.

A kind of stock trend forecasting method based on 2-dimension stream prediction, it is characterised in that described vertical Stock historical data Y of straight latitude is expressed as:

Y={f (t, d) | d=1,2 ..., Ly}={y_i| j=1,2 ..., Ly} (3)

In formula (1), (t, is d) the stock historical data of t time period in d days to f, and Ly is maximum number of days, then utilize Described Regression-Integral moving average model ARIMA method carries out data modeling and comprises the following steps:

Step 2B.1, the stock historical data { y of described vertical latitude_j| j=1,2 ..., Ly} is formed after the differential transformation of d rank Stationary time series { Δ y_j| j=1,2 ..., Ly-d}, then have:

In formula (4),I=1,2 ..., p, is autocorrelation coefficient, θ_i, i=1,2 ..., q, is partial correlation coefficient；

Step 2B.2, determined by the autocorrelogram and partial autocorrelation figure checking formula (4) model ARIMA (p, d, q) in from Return item p and rolling average item number q；

Step 2B.3, obtain autocorrelation coefficient by parameter estimationAnd partial correlation coefficient θ_i。