CN103985055A - Stock market investment decision-making method based on network analysis and multi-model fusion - Google Patents

Abstract

The invention discloses a stock market investment decision-making method based on network analysis and multi-model fusion. The method includes the steps that fundamental information is grabbed from a network and then network nodes and network connection are constructed; a complex social network model is constructed; an investment portfolio is selected by means of a network analysis method and then data involved in the investment portfolio are input into a multi-model fusion frame, wherein the multi-model fusion frame comprises a plurality of sub-models; all the sub-models conduct market trend prediction with different characteristics according to technical information of different characteristics grabbed in the network generate predicted values of the corresponding sub-models, the predicted values are weighted and summated to obtain a comprehensive market trend predicted value, and then corresponding investment strategies are generated according to the comprehensive market trend predicted value. According to the method, risk factors, which are ignored in general researches, of the investment portfolio are comprehensively considered from multiple angles, real-time requirements of the strategies are guaranteed through the methods such as dimension reduction data preselection of fundamentals and technical feature selection, and consequently the more reliable investment strategies are provided.

Description

A stock market investment decision-making method based on network analysis and multi-model fusion

technical field

The invention belongs to the technical field of stock market investment strategy analysis and relates to a stock market investment decision-making method based on network analysis and multi-model fusion.

Background technique

The financial market is the core of a country's economic operation. Therefore, the issue of securities investment strategy has always been one of the core issues that are most concerned by the investment theory and investment practice circles of various countries. The securities investment strategy reflects our understanding of the laws of the financial market and investor psychology. , is a rule system and action plan to guide investment behavior specified according to investment objectives. Among them, technical analysis and fundamental analysis methods are two main investment analysis methods. Among them, technical analysis is mainly used in the time and space judgment of specific investment operations, and fundamental analysis is mainly used in the selection of investment targets, as an important means to improve the effectiveness and reliability of investment analysis.

With the rapid development of Internet technology, a large amount of information related to the financial market is disseminated on the Internet. These real-time information are huge in scale and in various forms, which imply important information related to investment decisions. How to make comprehensive use of information for market prediction and analysis is an important issue in financial market investment decision-making.

In recent years, many researchers have conducted a series of studies on investment decision-making methods. One of the relatively new research directions is to analyze the relationship between textual information and stock prices. For example, the well-known Arizona State University developed the Arizona Financial Text system, which analyzes and predicts financial news and articles about stock prices; in the most recent 2009, Schumaker's research team analyzed financial news using text-based In 2012, Nizer's research team conducted further research on identifying which news had a visible impact on the stock market. The research results all show that the investment strategy guided by the automatic analysis method can obtain excess profits. However, it is noted that the analysis methods provided by these text analysis-based studies use limited information, and are limited to texts with strong emotional tendencies such as financial news. In fact, in these decision-making methods based on text analysis, the data information used in technical analysis (such as stock prices and stock indexes) is usually ignored. These prediction and analysis methods are obviously methods that use incomplete information.

On the other hand, the research on the traditional analysis method (that is, stock price prediction method) has been done for a long time, and a series of relatively mature research results have been obtained. MIT financial expert Luo Yaozong said: "Technical analysis is an effective method to extract useful information from market prices." There are also some studies in the US Federal Reserve Board and academic circles that support technical analysis. Evidence exists. In China, for example, the research teams of Liu Haiyue (2011), Jiang Long (2012), and Zheng Xiaowei (2013) have modeled and predicted stock price trend changes from neural network, gray RBF network, and LSSVM methods, and have achieved satisfactory results. the results. However, these forecasting methods are based on the laws of historical data for predictive analysis, which belongs to the idea of technical analysis, and its theory is refuted by the efficient market hypothesis, and the information it uses is also limited. And the traditional research only considers the prediction accuracy rate, and considering the frequent fluctuation of the actual stock price, the accuracy rate does not directly correspond to the actual investment income, and even a high prediction accuracy rate may bring negative returns. In addition, such a portfolio May bring higher investment risk.

At the same time, traditional stock price trend analysis and prediction methods do not take into account the scale and timeliness of data, so the amount of information used is very limited, and training is very time-consuming when forecasting under large amounts of data. Considering the actual investment decision-making Real-time requirements have been unable to meet the requirements of stock market analysis and forecasting under the network environment and massive data scale. Moreover, the existing stock forecasts mostly ignore the correlation between stocks, and think that the direct price changes of stocks are independent of each other. On this basis, the research and analysis of the law of stock price trend changes is carried out. This simplified assumption obviously violates our financial knowledge. The general understanding of the market, because the listed companies corresponding to each stock are related to each other and will be influenced and affected by each other.

As mentioned above, technical analysis is mainly used in the time and space judgment of specific investment operations, while fundamental analysis is mainly used in the selection of investment targets, as an important means to improve the effectiveness and reliability of investment analysis.

Contents of the invention

The problem to be solved by the present invention is to provide a stock market investment decision-making method based on network analysis and multi-model fusion, which comprehensively utilizes fundamental analysis and technical analysis to make market investment decisions, which can effectively reduce investment risks and increase investment returns.

The present invention is realized through the following technical solutions:

A stock market investment decision-making method based on network analysis and multi-model fusion, including the following operations:

First, capture fundamental information from the network, build network nodes and network connections on this basis, and construct a complex social network model; use network analysis methods to select investment portfolios, and then input the data involved in the investment portfolios into the multi-model fusion framework middle;

The multi-model fusion framework includes a plurality of sub-models, and each sub-model predicts market trends with different characteristics according to the technical information of different characteristics captured from the network, generates respective prediction values, and then weights the prediction values to obtain and, get the comprehensive market trend prediction value, and generate corresponding investment strategies according to the value;

The feature selector that provides information to the sub-model, the parameters involved in the sub-model, and the weight of the predicted value of the sub-model are all packaged and trained through the univariate distribution estimation algorithm.

The construction of the complex social network model includes the following operations:

1.1) Network nodes

In the vector space model, the fundamental information text captured from the network is represented by a bag of words in binary feature vector mode, as follows:

inf _i ＝(<t ₁ ,w _i1 >,<t ₂ ,w _i2 >,...<t _M ,w _iM >)

Where M is the number of features, w _ik is the weight of the text feature t _k , the weight is calculated by the tf*idf method, and the fixed features are simplified to inf _i =(w _i1 ,w _i2 ,...w _iM );

Perform the following operations on the text in the fundamental information obtained from the network using data mining methods:

1) Filtering: filter out useless parts of the information;

2) Word segmentation: segment the filtered information into multiple words, store the result after word segmentation in the vocabulary database, and identify the part of speech of the words;

3) Carry out further stop word processing to the vocabulary in the vocabulary bank, comprise removing function words and binding negative words;

After obtaining the text features in the fundamental information, calculate its weight, organize the fundamental information into a vector space model, and for time-varying information, the vector space model becomes a time-varying vector:

inf _i (t)=(w _i1 (t), w _i2 (t),...w _iM (t)), where t is a time variable;

1.2) Network connection

For the network G(t)=(V(t),E(t)), use the fundamental information of listed companies to model it as a network node, that is, V(t)={inf _i (t)},E (t)={(i,j,edg _ij (t))|i,j∈V(t)};

V(t) is a collection of network nodes constructed using fundamental information, and E(t) is a collection of two nodes i, j among the network nodes and the connection strength edg _ij (t) between them;

Using cosine similarity $\cos ({\inf }_i\left(t\right),{\inf }_j\left(t\right))=\frac{\underset{t_{\mathrm{no}}\&Element;T_m}{\mathrm{\&Sigma;}}{w}_{\mathrm{in}}\left(t\right)w_{\mathrm{jn}}\left(t\right)}{\sqrt{\underset{t_{\mathrm{no}}\&Element;T_m}{\mathrm{\&Sigma;}}{w}_{\mathrm{in}}{\left(t\right)}^2\underset{t_{\mathrm{no}}\&Element;T_m}{\mathrm{\&Sigma;}}{w}_{\mathrm{jn}}{\left(t\right)}^2}}$ Calculate the network connection strength, where T _M is the complete set of text features of fundamental information, and use the threshold θ to filter, that is, ${\mathrm{edg}}_{\mathrm{ij}}\left(t\right)=\left\{\begin{array}{cc}0,& \cos ({\inf }_i\left(t\right),{\inf }_j\left(t\right))<\mathrm{\&theta;}\\ \cos ({\inf }_i\left(t\right),{\inf }_j\left(t\right)),& \cos ({\inf }_i\left(t\right),{\inf }_j\left(t\right))\&Greater\; Equal;\mathrm{\&theta;}\end{array}\right.,$ θ takes cos45°;

Describe the network nodes with the vector space model shown in inf _i (t)=(w _i1 (t), w _i2 (t),...w _iM (t)), and construct the required network nodes;

Then the network nodes follow the network shown by V(t)={inf _i (t)}, E(t)={(i,j,edg _ij (t))|i,j∈V(t)} Connection means connection, thus complex social network model, which is a dynamic network model.

The selection of investment portfolio using network analysis method is to select the most unrelated stocks to form the investment portfolio, including the following multi-diversity division method based on community detection and clustering:

Use the community detection method for division, use the Girvan-Newman clustering method for network clustering, and its evaluation index is modularity $Q\left(t\right)=\underset{i}{\mathrm{\&Sigma;}}(e_i\left(t\right)-{a_i}^2\left(t\right)),$ t is a time variable;

Among them, e _ij (t) represents the proportion of the network connection weights connecting community i and network nodes in community j; a _i (t) represents the proportion of all network connection weights associated with network nodes in community i proportion, including two cases where both network connections are within the community and only one network connection is within the community;

The modularity describes the degree of difference between the divided complex social network model and the random network model, and the community division method with the largest modularity will be the best division;

The basic process of the Girvan-Newman clustering method is as follows:

Step 1: Calculate the betweenness of all network connections;

Step 2: Remove the network connection with the maximum betweenness;

The third step: recalculate the betweenness of the network connections affected in the second step;

Step 4: If there is no remaining network connection, then end the algorithm, otherwise turn to the second step;

Finally, choose a representative from each network community com(t) after optimal division, that is to form the required diversified investment portfolio;

Finally, the strategy model for fundamental analysis to select a portfolio is expressed as:

$\mathrm{<span\; class="notranslate">\; IFA</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; \&ForAll;</span>\mathrm{<span\; class="notranslate">\; com</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; com</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; max</span>}}}^{\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; rep</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; \}</span><span\; class="notranslate">\; .</span>$

The selection of the investment portfolio using the network analysis method is to select the most irrelevant stocks to form the investment portfolio, including the following multi-diversity segmentation method based on the largest fully connected unrelated subnetwork:

The Bron–Kerbosch algorithm is used to extract the largest fully connected subnet in the supplementary network. Its basic form is a recursive backtracking search algorithm. The process is as follows:

Bron-Kerbosch algorithm:

Step1: Given three sets (R, P, X), the initialization sets R and X are empty respectively, and the set P is the set of all network nodes;

Step2: If the sets P and X are empty respectively, output R as the largest group;

Step3: For each network node {v} obtained from the set P, the following processing is performed:

1) Add the network node {v} to the set R, set P, X intersect with the network node {v}'s adjacent network node set N{v}, and then recursively set R, P, X (turn to Step2);

2) Delete the network node {v} from the set P, and add the network node {v} to the set X;

At this time, the strategy model for fundamental analysis to select a portfolio is expressed as:

That is, the largest fully connected subnetwork in the supplementary network generated according to the Bron-Kerbosch algorithm.

The framework of multi-model fusion described is to transform a complex system into several subsystems, one subsystem corresponds to one sub-model, and then combine the sub-models corresponding to these subsystems to jointly describe the same model to improve the model fitting degree ;

The sub-model connection method adopts a weighted summation method, and the output of each sub-model is summed according to a certain weight to obtain the final output;

The sub-models can be added, subtracted or replaced. After basic data training, the overall model will adaptively select the prediction model by adjusting the weight of its own sub-models;

The sub-models described include the following:

1) Trend prediction method based on vector symbol sequence

First, the least squares fitting method is used to vectorize the historical stock price data. If the price of the x _i day is defined as y _i , the error of n days is minimized $S(a,b)=\underset{i=0}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}{({\mathrm{the\; y}}_i-f\left(x_i\right))}^2,f\left(x_i\right)={\mathrm{ax}}_i+b,$ A trend characterized by a slope can be obtained, defined as:

Further discretize the continuous trend vectors to extract macro trend information, use unsupervised clustering methods, and perform k-means-based re-clustering algorithms for the particularity of vectorized data to obtain clustering results and various class center vector;

2) Trend prediction algorithm based on stock time series turning point extraction

On the basis of the definition of time series fluctuation characteristic points, if time series X={x ₁ ,x ₂ ,……,x _N }, then the fluctuation characteristic point x of X in its time interval is {xVD(x) =max(VD( _xi )),i=1,2,...,N}, where

$\mathrm{<span\; class="notranslate">\; VD</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; ;</span>$

3) Investment recommendation algorithm based on vocabulary emotional tendency judgment

For any vocabulary w _i ∈ W, remember:

$\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>\stackrel{<span\; class="notranslate">\; ^</span>}{<span\; class="notranslate">\; =</span>}\left\{\begin{array}{c}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if</span>}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; is\; positive</span>}<span\; class="notranslate">\; ,</span>\\ <span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if</span>}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; is\; negative</span>}<span\; class="notranslate">\; .</span>\end{array}\right.$

T(w _i ) is called the emotional orientation value of w _i , given the emotional orientation relationship network SORN=(W,C,Q), the sequence of lexical numbers passed on a path from w _i to w _j in SORN is denoted as (p ₁ ,p ₂ ,...,p _s ), where 2≤s≤n, p ₁ =i, p _s =j;

If an element (h ₁ ,h ₂ ,…,h _s ) in the set {(p ₁ ,p ₂ ,…,p _s )} satisfies: $q_{h_1{h}_2}+q_{h_2{h}_3}+...+q_{h_{\mathrm{the\; s}-1}{h}_{\mathrm{the\; s}}}=\min \{q_{p_q{p}_2}+q_{p_2{p}_3}+...+q_{p_{\mathrm{the\; s}}-1p_{\mathrm{the\; s}}}\},$ Then (h ₁ ,h ₂ ,…,h _s ) is called the sequence of vocabulary numbers passing through a least noisy path from w _i to w _j in SORN, denoted as D(i,j); called (D(i, 1),..., D(i,i-1), D(i,i+1),...,D(i,n)) is the sequence of D(i,j), denoted as D(i); using Dijkstra's algorithm implemented by adjacency list and Fibonacci heap to calculate D(i);

Select the path with the least noise between two words to identify the emotional tendency relationship between the words. When the corpus is sufficient, the D(i,j) can be obtained by using Q, and the emotional tendency relationship between w _i and w _j can be calculated by using C; there are always more positive words than negative words in the corpus.

In the described clustering algorithm, each trend vector is replaced with the vector corresponding to the cluster center where it is located, and the vector characterizes the type of trend, if it is represented as a symbol, the continuously changing trend vector can finally be Discretize into a series of symbol sequences; using the stock's trading day, opening price, highest price, lowest price, closing price, trading volume and turnover as the main data of technical analysis, the historical information of stock S is expressed as a symbol sequence: {S _i }, $S_i=\{{\mathrm{the\; s}}_i^d,{\mathrm{the\; s}}_i^o,{\mathrm{the\; s}}_i^h,{\mathrm{the\; s}}_i^l,{\mathrm{the\; s}}_i^c,{\mathrm{the\; s}}_i^v,{\mathrm{the\; s}}_i^t\};$

in Corresponding to the trading day, opening price, highest price, lowest price, closing price, trading volume and trading value data of the stock on the i-th day;

And the support vector machine model using the radial basis kernel function is used to learn the historical stock price trend symbol model. The SVM theory pursues the model with the strongest generalization ability after learning. After finding the support vector SV, it is necessary to find The optimal classification hyperplane OHP, the solution process is a quadratic programming problem;

Let SV be the sample point closest to OHP: (l x)+b=0, and the distances between SVs of the same class and OHP are completely equal, and the distances between SVs of different classes and OHP are not necessarily equal;

Find the classification hyperplane for m training samples (x ₁ ,y ₁ ),(x ₂ ,y ₂ ),...,(x _m ,y _m ), the classification hyperplane must meet the conditions of the optimal classification hyperplane :

In order to find the optimal classification hyperplane, the original problem is transformed into a standard quadratic programming problem with the help of Lagrange function:

$\mathrm{<span\; class="notranslate">\; max</span>}\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>$

$\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; .</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; .</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&Greater\; Equal;</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span>$

The key to finding the optimal hyperplane is to find the α _i that satisfies α _i >0 and

And the optimal classification hyperplane is:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; sign</span>}<span\; class="notranslate">\; \{</span>\underset{{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; \}</span>$

Among them, the sample points corresponding to α _i >0 are support vectors, and the quadratic programming problem can be solved by positive method, dual method or interior point algorithm; the trained support vector machine automatically classifies new data and outputs the result of classification The symbol corresponds to the slope of the original cluster center vector, which is the predicted value of the final trend.

The steps of the time series fluctuation feature point extraction algorithm are as follows:

Step 1: Input the start coordinates start and end coordinates end of the sequence to be extracted, and judge whether the distance between start and end satisfies that the subsequence is less than the minimum interval length. If yes, go to the third step; Define to find the point with the largest VD value between the starting point and the end point. If VD is greater than the termination condition of the algorithm range, then add this point as a fluctuation feature point to the result sequence of fluctuation feature points, and record this point as fp;

The second step: use fp to divide the original sequence into two sections, namely start to fp subsection, fp to end subsection, and execute the first step for these two subsections;

Step 3: Save and output the result sequence of fluctuation feature points according to time;

Step 4: Determine the minimum time interval threshold based on the turning point, and use the maximum/minimum principle to extract the turning point seed set from the result sequence of the fluctuation feature points;

Step 5: Based on the turning point seed set, find the turning point between any two consecutive turning points based on the fluctuation feature point sequence and use the principle of maximum reverse volatility, and add it to the turning point seed set, repeat the above operation until the set turning point is extracted parameters can no longer find a new turning point;

Step 6: Save and output the turning point seed set sorted by time;

The turning point seeds extracted above represent the stock price change trend on a large time scale. After training with machine learning methods such as neural networks or support vector machines, the optimal classifier can be obtained, and the classifier can automatically classify new data. , output whether the corresponding date is a turning point, that is, the result of predicting the future trend, and the predicted value of the trend corresponding to the latest turning point is the predicted value of the final trend.

The described investment recommendation algorithm based on vocabulary emotional tendency judgment adopts SWSOA algorithm, and this algorithm starts from any node of SORN maximum connected subgraph, and just carries out emotional tendency classification to all nodes in this subgraph; The input of SWSOA algorithm For a specific SORN, the algorithm is expressed as a function SWSOA(SORN), where SORN represents a SORN variable;

The specific steps of the algorithm are as follows:

Step 1. Utilize the breadth-first traversal algorithm to obtain the maximum connected subgraph Gs of SORN, and the collection of vocabulary nodes contained in Gs is recorded as _WGS ;

Step 2. Designate any node w _i in W _GS , and make W _GS =W _GS -{ _wi }, U={ _wi } and

Step 3. Calculate D(i) to obtain the least noisy path from w _i to any node in _WGS ;

Step 4. For each vocabulary node w _j in WGS, execute a)~c):

a) According to D(i, j) in D(i), calculate the transition relation edge passed on the least noise path from w _i to w _j

quantity e, $e=\underset{l=1}{\overset{l=\mathrm{the\; s}-1}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_{h_l{h}_{l+1}}{c}_{h_l{h}_{l+1}};$

b) Calculate the tendency relationship between w _j and w _i : T(w _j )=(-1) ^e ×T(w _i );

c) If T(w _j )=T(w _i ), then U=U∪{w _j }, otherwise V=V∪{w _j };

Step 5. WSO classification is completed, and the classification results are stored in U and V respectively. If |U|>|V|, then positive words are stored in U and negative words are stored in V, and vice versa;

Extract commentary news reports and comments related to the financial market, use the SWSOA algorithm to identify their emotional tendencies, and determine the listed companies that are positive and their corresponding stocks to be candidate stocks that should be paid attention to;

The predicted value of the final trend of the listed companies and their corresponding stocks judged as positive is 1;

The predicted value of the final trend of the listed companies and their corresponding stocks judged as negative is -1.

Considering the computational complexity and complexity of the entire model due to multi-model fusion, it also includes training, optimization and feature selection methods based on univariate distribution estimation encapsulation methods:

In the univariate distribution estimation algorithm, for data, an n (n=N+Dim) dimensional chromosome is used in the algorithm to encode; the chromosome is divided into two parts, the first part is the encoded part of the encapsulated algorithm parameter, a total of N bits of binary code; and enable the elite retention process, and initialize each stock to store the historical optimal value separately to improve the optimization efficiency and ensure the real-time performance of the technical analysis process: for the population q(t), t is a time variable, let it be An m*n order matrix records the chromosomes arranged from high to low according to the fitness value Q9, the first r rows [q ₁ q ₂ ... q _r ] ^T are reserved as elites, then the estimated distribution probability matrix is P= [P ₁ P ₂ … P _n ], where: R is the number of chromosomes involved in the mutation, and the steps are as follows:

Step 1: Initialize the population q(0), read the historical optimal value and join the initial population, the number of iterations t=0;

The second step: the chromosome population q(t) sorts the chromosomes according to the order of decreasing fitness value, and reserves the first r chromosomes as elites;

Step 3: Calculate the probability P _i of taking 1 in each column of the first R (R is generally greater than r) chromosomes, and form a probability matrix;

Step 4: Randomly generate a number conforming to the 0-1 distribution with the probability P _i as the i-th position of the chromosome, and then use the same algorithm to generate (mr) chromosomes, let t=t+1, combine these chromosomes with the remaining r chromosomes are combined into a new population q(t);

The fifth step: judge t>N, if it is true, then end the algorithm, if not, then turn to the second step;

The parameter optimization part is the application package, which needs to encode and decode the parameters. The binary encoding and decoding method of the weight of each sub-model in the multi-model fusion is as follows:

1) Weight coding

The initialization weight is a set of n×m dimensional long 0 and 1 code sequences, where m is the number of weights to be optimized, n is the number of code bits for each weight, and the code length of each weight is identical;

2) Weight decoding

The value range of weight w _i is w _i ∈ [0,1], assuming that the encoding of the i-th weight is [q ₁ q ₂ … q _n ], where q _i ∈ {0,1}, i∈{ 1,2,…,n}

Then w' _i ＝q ₁ ×2 ⁰ +q ₂ ×2 ¹ +…+q _n ×2 ^n-1

Then map w' _i to the interval [0,1], then

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; \&prime;</span>}}{{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; no</span>}}}$

In the encapsulation method, the data set formed by the technical information is randomly divided into two parts, the training set and the test set, the population is decoded, and the genotype part of the feature is used to select a subset of features for training to obtain a subset whose parameters are set. At the same time, the genotype part of the feature is used to perform feature selection on the test set data, and the test set after feature selection is input to the trained sub-model for testing, and the fitness evaluation is performed according to the classification performance of the test set. If the univariate If the termination condition of the distribution estimation, the optimal sub-model and feature subset are output, otherwise the probability distribution vector is updated according to the univariate distribution estimation algorithm, and a new population is generated by encoding again.

The described use of comprehensive market trend forecasts to design comprehensive investment strategies is based on the following correction factors:

$\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; \{</span>\begin{array}{c}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if</span>}{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&times;</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 1</span>}\\ {\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&times;</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}<span\; class="notranslate">\; ,</span>$

Among them, P and Q correspond to the accuracy rate and evaluation value of the test set in training, respectively, and the correction factor r(t) reflects the investor's confidence in this prediction, that is, the higher the value of r(t), the greater the investor's confidence in this prediction. The more confident the prediction result is, the more likely it is to doubt the prediction result anyway and keep the original strategy; suppose the number of stocks held by investors at time t is N(t), the amount of cash held by investors is c(t), and the stock closing price cp(t) represents VAL(t) as its price, and the operation symbol is Sig(t), then for a stock, it can be expressed mathematically based on the investment strategy as:

$\mathrm{<span\; class="notranslate">\; Sig</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{c}\mathrm{<span\; class="notranslate">\; buy</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if</span>}{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 0</span>}\\ \mathrm{<span\; class="notranslate">\; sell</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if</span>}{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; 0</span>}\\ \mathrm{<span\; class="notranslate">\; Sig</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if</span>}{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}\end{array}\right.<span\; class="notranslate">\; ;</span>$

The closing price is used as the operating standard above, so the final total assets are a(t)=c(t)+N(t)×cp(t);

For the investment portfolio, further consider the stock pre-selection IFA based on network analysis, and define the expectation of investment analysis as E(ITA(t,s))=Y _s (t)×r _s (t), where Y _s is the comprehensive market Trend prediction value, r _s is the degree of confidence in the result, then its transaction amount for stock s can be calculated according to The corresponding general investment strategy is expressed as:

$<span\; class="notranslate">\; \&ForAll;</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; IFA</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>$

$\mathrm{<span\; class="notranslate">\; Sig</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; \{</span>\begin{array}{c}\mathrm{<span\; class="notranslate">\; buy</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; ITA</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if\; ITA</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{<span\; class="notranslate">\; +</span>}\\ \mathrm{<span\; class="notranslate">\; sell</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; ITA</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if\; ITA</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; \&theta;</span>}<span\; class="notranslate">\; \_</span>\\ \mathrm{<span\; class="notranslate">\; Sig</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}<span\; class="notranslate">\; ,</span>$

①: Sig(t,s)＝Buy(s,r)②: Sig(t,s)＝Sell(s,r)

where r=ITA(t,s)

Compared with the prior art, the present invention has the following beneficial technical effects:

The stock market investment decision-making method based on network analysis and multi-model fusion provided by the present invention uses the text information in the fundamental analysis and the numerical information in the technical analysis to complement each other, realizes a more complete comprehensive utilization of information, and utilizes complex social networks The model analyzes the network structure of the market, comprehensively considers the investment portfolio risk factors neglected in general research from the perspective of diversity, and ensures the real-time requirements of decision-making through the pre-selection of dimensionality reduction data on the fundamental side and the feature selection method on the technical side , to provide a more reliable investment strategy.

The stock market investment decision-making method based on network analysis and multi-model fusion provided by the present invention uses a complex social network model to model the financial market to deal with the problem of neglecting the correlation between listed companies in existing methods; due to the use of complex The social network model models the financial market, takes into account the correlation between listed companies, and effectively reduces the risk of the investment portfolio. The calculation shows that the Sharpe index of the selected investment portfolio is significantly improved.

The stock market investment decision-making method based on network analysis and multi-model fusion provided by the present invention complements the text data in the fundamental information and the numerical data in the market, further ensuring the integrity of the information in the research and analysis.

The stock market investment decision-making method based on network analysis and multi-model fusion provided by the present invention considers the real-time problem under the large amount of data, and is more suitable for processing massive data; due to the use of pre-selection process and feature selection, the data is effectively processed Dimensionality reduction, the present invention overcomes the problem that it is difficult to guarantee real-time performance under a large amount of data, and is more suitable for processing massive data.

The stock market investment decision-making method based on network analysis and multi-model fusion provided by the present invention comprehensively uses multiple models to automatically provide corresponding investment strategies for actual investment use or reference; comprehensively uses multi-model fusion methods to flexibly combine various analysis and prediction methods , and through encapsulation method training to achieve adaptive optimization selection, it can be compatible with and utilize existing analysis and prediction methods, and can automatically provide corresponding investment strategies for actual investment use or reference.

Description of drawings

Fig. 1 is the schematic diagram of investment decision-making method of the present invention;

Figure 2 is a re-clustering algorithm based on k-means;

Figure 3 is a distribution estimation algorithm with elite reservation;

Fig. 4 is the basic flowchart of packaging method;

Fig. 5 is the overall schematic diagram of the automatic stock market investment decision-making system of multi-model integration;

Figure 6 shows the simulated investment results in February 2012;

Figure 7-1 and Figure 7-2 are schematic diagrams of simulated investment results and their growth rates in March 2012, respectively.

Detailed ways

The present invention will be further described in detail below in conjunction with specific embodiments, which are explanations of the present invention rather than limitations.

Referring to Figure 1, first capture fundamental information from the network, build network nodes and network connections on this basis, and form a complex social network model; use network analysis methods to select investment portfolios, and then input the data involved in the investment portfolios into In the multi-model fusion framework;

The multi-model fusion framework includes a plurality of sub-models, and each sub-model predicts market trends with different characteristics according to the technical information of different characteristics captured from the network, generates respective prediction values, and then weights the prediction values to obtain And, get the comprehensive market trend prediction value, and generate the corresponding investment strategy according to the value.

The feature selector that provides information to the sub-model, the parameters involved in the sub-model, and the weight of the predicted value of the sub-model are all packaged and trained through the univariate distribution estimation algorithm.

The following will introduce step by step the realization of the method involved in the invention, and finally explain the realization of the complete investment strategy system:

1. Construction of complex social network model

1.1 Network Node

In the vector space model, the fundamental information text captured from the network is represented by a bag of words in binary feature vector mode, as follows:

inf _i ＝(<t ₁ ,w _i1 >,<t ₂ ,w _i2 >,...<t _M ,w _iM >)

Where M is the number of features, w _ik is the weight of the text feature t _k , and the weight is calculated by the tf*idf method. For fixed features, it can be simplified as inf _i =(w _i1 ,w _i2 ,...w _iM ) .

Perform the following operations on the text in the fundamental information (obtained from the network using data mining methods):

1) Filtering: filtering out useless parts of the information; specific operations include: removing punctuation, separators, time attributes, fixed logos and spam in the information;

2) Word segmentation: segment the filtered information into multiple words, store the result after word segmentation in the vocabulary database, and identify the part of speech of the words;

3) Carry out further stop word processing to the vocabulary in the vocabulary bank, including: remove function words such as auxiliary words, conjunctions and bind negative words, etc.;

After a series of steps such as filtering, word segmentation, and stop words, the text features in the fundamental information are obtained, and their weights are calculated, so that the fundamental information can be organized into the above-mentioned vector space model. For time-varying information (such as monthly reports, annual reports, etc. of listed companies), the vector space model becomes a time-varying vector:

inf _i (t)=(w _i1 (t), w _i2 (t),...w _iM (t)), where t is a time variable;

1.2 Network connection

The structure of financial markets closely resembles complex social networks. Each listed company is similar to members in a society, and the connection between companies is similar to the relationship among members of society. For the network G(t)=(V(t),E(t)), use the fundamental information of listed companies to model it as a network node, that is, V(t)={inf _i (t)},E (t)={(i,j,edg _ij (t))|i,j∈V(t)}.

V(t) is the set of network nodes constructed using fundamental information, and E(t) is the set of two nodes i and j among the network nodes and the connection strength edg _ij (t) between them;

The dynamic network model can clearly and intuitively represent the internal structure of the market, and can realize the automatic processing of real-time changing market information, and provides a series of specific methods for fundamental analysis.

Using cosine similarity $\cos ({\inf }_i\left(t\right),{\inf }_j\left(t\right))=\frac{\underset{t_{\mathrm{no}}\&Element;T_m}{\mathrm{\&Sigma;}}{w}_{\mathrm{in}}\left(t\right)w_{\mathrm{jn}}\left(t\right)}{\sqrt{\underset{t_{\mathrm{no}}\&Element;T_m}{\mathrm{\&Sigma;}}{w}_{\mathrm{in}}{\left(t\right)}^2\underset{t_{\mathrm{no}}\&Element;T_m}{\mathrm{\&Sigma;}}{w}_{\mathrm{jn}}{\left(t\right)}^2}}$ Calculate the network connection strength, where T _M is the complete set of text features of fundamental information, and use the threshold θ to filter, that is, ${\mathrm{edg}}_{\mathrm{ij}}\left(t\right)=\left\{\begin{array}{cc}0,& \cos ({\inf }_i\left(t\right),{\inf }_j\left(t\right))<\mathrm{\&theta;}\\ \cos ({\inf }_i\left(t\right),{\inf }_j\left(t\right)),& \cos ({\inf }_i\left(t\right),{\inf }_j\left(t\right))\&Greater\; Equal;\mathrm{\&theta;}\end{array}\right.,$ θ can take cos45°, which is approximately 0.707.

Like this, describe the network node with the vector space model shown in inf _i (t)=(w _i1 (t), w _i2 (t),...w _iM (t)), construct the required network node; Then the network nodes follow the network shown by V(t)={inf _i (t)}, E(t)={(i,j,edg _ij (t))|i,j∈V(t)} Connection means connection, thus complex social network model, which is a dynamic network model.

2. Portfolio selection through network analysis methods

In order to minimize the risk of investment portfolio while fully reducing the calculation amount of automatic investment analysis. The idea of diversification strategy is introduced in the present invention, that is, in stock investment, if the stocks that are as irrelevant as possible are selected to form a portfolio, the risk of the portfolio will be less than the average risk. Two implementation methods are given below:

2.1 Diversity division method based on community detection clustering

In the present invention, the community detection method is used to divide, because the clustering method will ensure that the similarity distance inside the community is small, and the similarity distance between the communities is large; specifically, the Girvan-Newman clustering method can be used for network clustering, and its evaluation index is Modularity t is a time variable. Among them, e _ij (t) represents the proportion of the weights of network connections connecting community i and network nodes in community j; a _i (t) represents the proportion of weights of all network connections associated with network nodes in community i The proportion of the total includes two cases where both network connections are within the community and only one network connection is within the community. The modularity describes the degree of difference between the divided complex social network model and the random network model, and the community division method with the largest modularity will be regarded as the best division. The basic process of the Girvan-Newman clustering method is as follows:

Step 1: Calculate the betweenness of all network connections;

Step 2: Remove the network connection with the maximum betweenness;

The third step: recalculate the betweenness of the network connections affected in the second step;

Step 4: If there is no remaining network connection, then end the algorithm; otherwise, turn to the second step.

Finally, choose a representative from the optimally divided network communities com(t) to form the required diversified investment portfolio. The selected strategy can be judged according to the importance index rep _i (t) of the network node (such as one of degree, closeness, betweenness number, centrality, etc., and closeness is recommended). Ultimately, the strategy model for selecting a portfolio for fundamental analysis can be expressed as:

$\mathrm{<span\; class="notranslate">\; IFA</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; \&ForAll;</span>\mathrm{<span\; class="notranslate">\; com</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; arg</span>}{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; com</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; max</span>}}}^{\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; rep</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; \}</span>$

The above portfolio selection can also be implemented as follows:

2.2 Multivariate segmentation method based on maximum fully connected unrelated subnetwork:

Note that the clustering method attempts but does not guarantee the maximization of the differences between classifications. The present invention designs another method for diversification, which is to find the largest fully connected subnetwork of the complement network of the complex social network model. Since the complex social network G(t)=(V(t),E(t)) is modeled according to the correlation between listed companies, its complementary network It just represents the uncorrelated relationship between the corresponding stocks, and the fully connected subnetwork selected from it can ensure the strong mutual independence of the stocks in the portfolio, and the largest subnetwork means the maximum coverage of the original market. The present invention adopts the Bron–Kerbosch algorithm to extract the largest fully connected subnet (the largest group) in the supplementary network, and its basic form is a recursive backtracking search algorithm, and the process is as follows:

Bron-Kerbosch algorithm:

Step1: Given three sets (R, P, X), the initialization sets R and X are empty respectively, and the set P is the set of all network nodes (here is the set of network nodes in the supplementary network);

Step2: If the sets P and X are empty respectively, output R as the largest group;

Step3: For each network node {v} obtained from the set P, the following processing is performed:

1) Add the network node {v} to the set R, set P, X intersect with the network node {v}'s adjacent network node set N{v}, and then recursively set R, P, X (turn to Step2);

2) Delete network node {v} from set P, and add network node {v} to set X.

At this time, the strategy model for selecting investment portfolios for fundamental analysis can be expressed as:

That is, the largest fully connected subnetwork in the supplementary network generated according to the Bron-Kerbosch algorithm.

According to the clustering method and the nature of the supplementary network, it can be seen that the above implementation 2.1 highlights the physical meaning of the selection, while the implementation 2.2 emphasizes the irrelevance of the division. During the operation, the operation method can be selected according to the actual needs. Generally, the implementation 2.2 can be selected by default. to operate. At the same time, it is noted that the selection of investment stocks through network analysis can greatly reduce the stocks that need attention, reduce the amount of data and improve the efficiency of subsequent analysis work.

3. A framework for multi-model fusion based on technical information

Considering the mass, complexity and nonlinearity of financial market data, it may be difficult for a single modeling method to accurately describe the model. The usual analysis method compares multiple algorithms, and then adopts the prediction algorithm with relatively better effect, while ignoring the rest of the prediction algorithms. However, this will ignore a lot of useful information. Because of this, the present invention adopts multi-model fusion modeling, and its basic idea is to systematize a complex system into several subsystems, one subsystem corresponds to a sub-model, and then combine the sub-models corresponding to these subsystems to jointly describe the same model to improve model fit.

The neutron model connection method of the present invention adopts a weighted summation method, and the output of each sub-model is summed according to a certain weight value to obtain the final output. Usually, the above weights can be obtained through genetic algorithm and artificial neural network optimization. Different from the "switching" method, the output of each sub-model in the "weighted sum" method contributes to the result. Although this connection method is less sensitive to the data classifier, the calculation amount of the entire model and The complexity has increased a lot. Its connection framework is shown in Figure 2. Let the degree output of each sub-model at time t be y _i (t), the weight after training is u _i , and the final output is: The present invention provides several available sub-models as follows:

3.1 Trend prediction method based on vector symbol sequence

Firstly, the least squares fitting method is used to vectorize the historical stock price data. If the price of the x _i- th day is defined as y _i , minimize the error of n days $S(a,b)=\underset{i=0}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}{({\mathrm{the\; y}}_i-f\left(x_i\right))}^2,f\left(x_i\right)={\mathrm{ax}}_i+b,$ A trend characterized by a slope can be obtained, defined as:

To further discretize the continuous trend vector to extract macro trend information, an unsupervised clustering method can be used to perform k-means-based re-clustering as shown in Figure 3 for the particularity of the vectorized data algorithm to obtain clustering results and various center vectors.

The clustering algorithm shown in Figure 3 can automatically deal with abnormal price changes caused by ex-rights and ex-dividends that sometimes occur in the securities market, so it can effectively eliminate the impact of unexpected events.

Replace each trend vector with the vector corresponding to the cluster center where it is located, and this vector represents the type of trend (such as: big up, up, stable, down, big down, etc.), if it is represented as a symbol, then Finally, the continuously changing trend vector can be discretized into a series of symbol sequences. Generally, the trading day, opening price, highest price, lowest price, closing price, trading volume and turnover of the stock are mainly used as the main data of technical analysis, so the historical information of the stock S can be expressed as a symbol sequence {Si}, $S_i=\{{\mathrm{the\; s}}_i^d,{\mathrm{the\; s}}_i^o,{\mathrm{the\; s}}_i^h,{\mathrm{the\; s}}_i^l,{\mathrm{the\; s}}_i^c,{\mathrm{the\; s}}_i^v,{\mathrm{the\; s}}_i^t\};$ in ${\mathrm{the\; s}}_i^d,{\mathrm{the\; s}}_i^o,{\mathrm{the\; s}}_i^h,{\mathrm{the\; s}}_i^h,{\mathrm{the\; s}}_i^l,{\mathrm{the\; s}}_i^c,{\mathrm{the\; s}}_i^v,{\mathrm{the\; s}}_i^t$ Corresponding to the trading day, opening price, highest price, lowest price, closing price, trading volume and trading value data of the stock on the i-th day respectively. For example, if the closing price is used as the final trading operation price, and when technical analysis is performed, and only m-day data of the previous fixed length is used, then the training data format is

Then, in order to realize automatic technical analysis, the present invention adopts the support vector machine (SVM) model using the radial basis function (RBF) to learn the historical stock price trend symbol model. The SVM theory pursues the model with the strongest generalization ability after learning. Therefore, after finding the support vector SV, it is necessary to find the optimal classification hyperplane OHP. This solving process is a quadratic programming problem.

Let SV be the nearest sample point to OHP: (l·x)+b=0, and the distances between SVs of the same class and OHP are completely equal, and the distances of SVs of different classes to OHP are not necessarily equal.

For m training samples (x ₁ ,y ₁ ),(x ₂ ,y ₂ ),...,(x _m ,y _m ) to find the classification hyperplane, the key is to find the coefficients l and b. Since the support vector machine theory requires that the classification hyperplane has good properties (that is, the hyperplane classification error is small and the generalization ability is strong), the classification hyperplane must meet the conditions of the optimal classification hyperplane:

In order to find the optimal classification hyperplane, according to the optimization theory, the original problem can be transformed into a standard quadratic programming problem with the help of the Lagrange function:

$\mathrm{<span\; class="notranslate">\; max</span>}\mathrm{<span\; class="notranslate">\; W</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>$

$\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; .</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; .</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&Greater\; Equal;</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span>$

The key to finding the optimal hyperplane is to find the α _i that satisfies α _i >0 and

And the optimal classification hyperplane is:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; sign</span>}<span\; class="notranslate">\; \{</span>\underset{{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; \}</span>$

Usually the sample points corresponding to α _i >0 are support vectors. For the above quadratic programming problem, classic solutions such as positive method, dual method, and interior point algorithm can be used to solve it. The trained support vector machine can automatically classify the new data, and output the symbols of the classification results (such as: rising, rising, stable, falling, falling, etc.), which correspond to the slope of the original cluster center vector, which is The predicted value of the final trend.

3.2 Trend prediction algorithm based on stock time series turning point extraction

Time series data usually has the characteristics of frequent short-term fluctuations, high noise interference, and unsteady state. However, in many applications, people usually only care about some key points in the form, and these points are the characteristic points of time series fluctuations. The present invention will introduce an extraction algorithm on the basis of the definition of time series fluctuation feature points: if time series X={x ₁ , x ₂ ,...,x _N }, then the fluctuation characteristics of X in its time interval The point x is {x|VD(x)=max(VD( _xi )), i=1,2,…,N}, where

$\mathrm{<span\; class="notranslate">\; VD</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>$

The steps of the time series fluctuation feature point extraction algorithm are as follows:

Step 1: Input the start coordinates start and end coordinates end of the sequence to be extracted. Determine whether the distance between start and end satisfies that the subsequence is smaller than the minimum interval length. If it is satisfied, go to the third step; if it is not satisfied, find the point with the largest VD value between the starting point and the end point according to the definition of the fluctuation feature point. If VD is greater than the algorithm range The termination condition of , then add this point as a fluctuation feature point to the result sequence of fluctuation feature points, and record this point as fp.

The second step: use fp to divide the original sequence into two sections, namely start to fp subsection, fp to end subsection, and execute the first step for these two subsections.

Step 3: Save and output the result sequence of fluctuation feature points according to time.

Step 4: Determine the minimum time interval threshold based on the turning point, and use the maximum/minimum principle to extract the turning point seed set from the result sequence of fluctuation feature points.

Step 5: Based on the turning point seed set, find the turning point between any two consecutive turning points based on the fluctuation feature point sequence with the principle of maximum reverse amplitude, and add it to the turning point seed set. Repeat the above operations until the parameters are extracted according to the set turning point (if the time interval condition and the fluctuation amplitude condition between the found two points meet the preset threshold), no new turning point can be found;

Step 6: Save and output the turning point seed set sorted by time;

The turning point seeds extracted above represent the trend of stock price changes on a large time scale, and can be trained by machine learning methods such as neural networks or support vector machines to obtain the optimal classifier, which can automatically classify new data. Classification, output whether the corresponding date is a turning point, that is, the result of predicting the future trend. The predicted value of the trend corresponding to the nearest turning point is the predicted value of the final trend.

3.3 Investment Recommendation Algorithm Based on Vocabulary Sentiment Tendency Judgment

For any vocabulary w _i ∈ W, remember:

$\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>\stackrel{<span\; class="notranslate">\; ^</span>}{<span\; class="notranslate">\; =</span>}\left\{\begin{array}{c}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if</span>}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; is\; positive</span>}<span\; class="notranslate">\; ,</span>\\ <span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if</span>}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; is\; negative</span>}<span\; class="notranslate">\; .</span>\end{array}\right.$

Call T(w _i ) the emotional tendency value of w _i . Given the emotional orientation relationship network SORN=(W,C,Q), the sequence of vocabulary numbers passing through a path from w _i to w _j in SORN is recorded as (p ₁ ,p ₂ ,…,p _s ) , where 2≤s≤n, p ₁ =i, p _s =j. If an element (h ₁ ,h ₂ ,…,h _s ) in the set {(p ₁ ,p ₂ ,…,p _s )} satisfies: $q_{h_1{h}_2}+q_{h_2{h}_3}+...+q_{h_{\mathrm{the\; s}-1}{h}_{\mathrm{the\; s}}}=\min \{q_{p_q{p}_2}+q_{p_2{p}_3}+...+q_{p_{\mathrm{the\; s}}-1p_{\mathrm{the\; s}}}\},$ Then (h ₁ , h ₂ , ..., h _s ) is called the sequence of vocabulary numbers passing through a least noisy path from w _i to w _j in SORN, denoted as D(i, j). Call (D(i,1),...,D(i,i-1), D(i,i+1),...,D(i,n)) a sequence of D(i,j), denoted as D(i). D(i) can be calculated using Dijkstra's algorithm implemented by adjacency list and Fibonacci heap.

Obviously, the path with the least noise between two words should be selected to identify the sentiment-oriented relationship between the pair of words. When the corpus is sufficient, the D(i, j) can be obtained by using Q, and then the emotional tendency relationship between w _i and w _j can be calculated by using C. In addition, there are always more positive words than negative words in the corpus. Based on the above analysis, the SWSOA algorithm is proposed. Starting from any node in the maximum connected subgraph of SORN, the algorithm can classify the sentiment orientation of all nodes in the subgraph. The input of SWSOA algorithm is a specific SORN. For ease of use, the algorithm is expressed as a function SWSOA(SORN), where SORN represents the SORN variable. The steps of the algorithm are as follows:

Step 1. Use the breadth-first traversal algorithm to obtain the largest connected subgraph Gs of SORN. The set of vocabulary nodes contained in Gs is denoted as W _GS .

Step 2. Designate any node w _i in W _GS , and make W _GS =W _GS -{ _wi }, U={ _wi } and

Step 3. Compute D(i) to obtain the least noisy path from w _i to any node in _WGS .

Step 4. For each vocabulary node w _j in W _GS , execute a)~c):

d) According to D(i, j) in D(i), calculate the turning point passed on the least noisy path from w _i to w _j

The number of relationship edges e, $e=\underset{l=1}{\overset{l=\mathrm{the\; s}-1}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_{h_l{h}_{l+1}}{c}_{h_l{h}_{l+1}}.$

e) Calculate the propensity relationship between w _j and w _i : T(w _j )=(-1) ^e ×T(w _i ).

f) If T(w _j )=T(w _i ), then U=U∪{w _j }, otherwise V=V∪{w _j }.

Step 5. WSO classification is completed, and the classification results are stored in U and V respectively. If |U|>|V|, then positive words are stored in U and negative words are stored in V, and vice versa.

To extract commentary news reports and comments related to the financial market, you can use the SWSOA algorithm to judge their emotional tendencies. The listed companies and their corresponding stocks that are judged as positive are obviously candidate stocks that should be recommended for attention.

The predicted value of the final trend of the listed companies and their corresponding stocks judged as positive is 1;

The predicted value of the final trend of the listed companies and their corresponding stocks judged as negative is -1.

Note that these sub-models can be added, subtracted or replaced. After basic data training, the overall model will adaptively select the prediction model by adjusting the weights of its own sub-models.

4. Training, optimization and feature selection methods based on univariate distribution estimation encapsulation methods

Considering that the multi-model fusion makes the calculation amount and complexity of the whole model very large, coupled with the massive data information with a lot of redundancy and noise, the time complexity of calculation needs to be considered. For big data projects, the main methods are to improve hardware performance, optimize algorithm performance and reduce data dimensionality. For this reason, the present invention proposes an encapsulation method to realize feature selection reduction of data, and at the same time, the method is also used to adaptively realize model training and parameter optimization in the system.

In the present invention, a univariate distribution estimation algorithm is used for encapsulation training. Univariate distribution estimation algorithm is a kind of genetic algorithm. For the data, an n(n=N+Dim) dimensional chromosome is used in the algorithm to encode. Chromosomes are divided into two parts. The first part is the encoding part of the encapsulated algorithm parameters, with a total of N-bit binary codes. Taking the SVM classifier as an example, there are a total of N=N _c +N _g binary codes, where N _c is the number of coding bits of the SVM parameter C, and N _g is the number of coding bits of the SVM parameter γ. The second part is the feature selection part. The dimension of this part is equal to the dimension of the data to be classified (Dim dimension). 1 means that the feature of the corresponding bit of the data set is retained, and 0 means that the feature of the corresponding bit of the data set is discarded.

At the same time, the present invention redesigns the univariate distribution estimation algorithm, enables the elite retention process in the algorithm, and initializes the historical optimal values stored for each stock to improve the optimization efficiency and ensure the real-time performance of the technical analysis process: Population q(t), where t is a time variable, is assumed to be an m*n order matrix, which records the chromosomes arranged from high to low according to the fitness value Q9 (note that Q9 is an evaluation standard that does not depend on data distribution), The first r rows [q ₁ q ₂ ... q _r ] ^T are reserved as elites, then the estimated distribution probability matrix is P=[P ₁ P ₂ ... P _n ], where: R is the number of chromosomes involved in the mutation. The redesigned distribution estimation algorithm is shown in Figure 4. The steps are as follows:

Step 1: Initialize the population q(0), read the historical optimal value and join the initial population, the number of iterations t=0;

The second step: the chromosome population q(t) sorts the chromosomes according to the order of decreasing fitness value, and keeps the first r chromosomes as elites;

Step 3: Calculate the probability P _i of taking 1 in each column of the first R (R is generally greater than r) chromosomes, and form a probability matrix;

Step 4: Randomly generate a number conforming to the 0-1 distribution with the probability P _i as the i-th position of the chromosome, and then use the same algorithm to generate (mr) chromosomes, let t=t+1, combine these chromosomes with the remaining r chromosomes are combined into a new population q(t);

The fifth step: judge t>N, if it is true, then end the algorithm, if not, then go to the second step.

Among them, the parameter optimization part is the application package, which needs to encode and decode the parameters. The following uses the weight of each sub-model in the multi-model fusion as an example to illustrate the binary encoding and decoding method:

1) Weight coding

When optimizing the sub-model weights, the weights need to be coded to participate in the distribution estimation operation. The encoding process can be briefly described as: the initialization weight is a set of n×m dimensional length 0 and 1 encoding sequences, where m is the number of weights to be optimized, n is the number of encoding bits for each weight, and each The encoding lengths of the weights are all the same.

2) Weight decoding

The optimization result returned by the encapsulation method needs to undergo a decoding process corresponding to the encoding process to obtain the actual required weight result. The decoding process can be briefly described as: the value range of the weight w _i is w _i ∈ [0,1], assuming that the encoding of the i-th weight is [q ₁ q ₂ … q _n ], where q _i ∈ {0 ,1},i∈{1,2,…,n}

Then w' _i ＝q ₁ ×2 ⁰ +q ₂ ×2 ¹ +…+q _n ×2 ^n-1

Then map w' _i to the interval [0,1], then

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; \&prime;</span>}}{{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; no</span>}}}$

Finally, the basic flowchart of the encapsulation method is shown in Figure 5. The data set formed by the technical information is randomly divided into two parts, the training set and the test set, the population is decoded, and the feature subset is selected for training using the genotype part of the feature. , get the sub-model whose parameters are set, and use the genotype part of the feature to perform feature selection on the test set data, input the feature-selected test set to the trained sub-model for testing, and perform classification according to the classification performance of the test set For fitness evaluation, if the termination condition of univariate distribution estimation is satisfied, the optimal sub-model and feature subset are output, otherwise, the probability distribution vector is updated according to the univariate distribution estimation algorithm, and a new population is generated by encoding again.

For the training set data, different models to be trained are used to train the univariate distribution estimation algorithm, and the test set data is used for 5-fold cross-validation, and the optimal output result is selected as the optimal sub-model and feature subset, so that it can be adaptively Realize the training and parameter optimization of sub-models in the multi-model fusion framework.

5. Design a comprehensive investment strategy using comprehensive market trend forecasts

In order to formulate an investment strategy for technical analysis, a correction factor is defined in the present invention:

$\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; \{</span>\begin{array}{c}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if</span>}{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&times;</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 1</span>}\\ {\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&times;</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; Q</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}<span\; class="notranslate">\; ,</span>$

Among them, P and Q correspond to the accuracy and evaluation value of the test set in training, respectively. The correction factor r(t) reflects the investor's confidence in the forecast, that is, the higher the value of r(t), the more confident the investor is in the forecast result; otherwise, they will tend to doubt the forecast result and keep the original Strategy. Assuming that the number of stocks held by investors at time t is N(t), the amount of cash held by investors is c(t), and the closing price cp(t) of the stock is used as its price to represent VAL(t), and the operation symbol is Sig(t ), then for a stock, it can be expressed mathematically based on the investment strategy as:

$\mathrm{<span\; class="notranslate">\; Sig</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{c}\mathrm{<span\; class="notranslate">\; buy</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if</span>}{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 0</span>}\\ \mathrm{<span\; class="notranslate">\; sell</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if</span>}{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; 0</span>}\\ \mathrm{<span\; class="notranslate">\; Sig</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if</span>}{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}\end{array}\right.<span\; class="notranslate">\; ;</span>$

The closing price is used as the operating standard above, so the final total assets are a(t)=c(t)+N(t)×cp(t).

For the investment portfolio, further consider the stock pre-selection IFA based on network analysis, and define the expectation of investment analysis as E(ITA(t,s))=Y _s (t)×r _s (t), where Y _s is the comprehensive market Trend prediction value, r _s is the degree of confidence in the result, then its transaction amount for stock s can be calculated according to The corresponding general investment strategy can be expressed as:

$<span\; class="notranslate">\; \&ForAll;</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; IFA</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>$

$\mathrm{<span\; class="notranslate">\; Sig</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; \{</span>\begin{array}{c}\mathrm{<span\; class="notranslate">\; buy</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; ITA</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if\; ITA</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ></span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{<span\; class="notranslate">\; +</span>}\\ \mathrm{<span\; class="notranslate">\; sell</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; ITA</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; if\; ITA</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; \&theta;</span>}<span\; class="notranslate">\; \_</span>\\ \mathrm{<span\; class="notranslate">\; Sig</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; otherwise</span>}\end{array}<span\; class="notranslate">\; ,</span>$

①: Sig(t,s)＝Buy(s,r)②: Sig(t,s)＝Sell(s,r)

where r=ITA(t,s)

The invention models the financial market by using a complex social network model, considers the correlation between listed companies, and effectively reduces the risk of investment portfolios. The calculation shows that the Sharpe index of the selected investment portfolio has increased significantly, as shown in Table 1 below;

Table 1 Portfolio Sharpe Index Comparison

Due to the adoption of the pre-selection process and feature selection, the dimensionality reduction of the data is effectively carried out. The present invention overcomes the problem that the real-time performance is difficult to guarantee under a large amount of data, and is more suitable for processing massive data;

The present invention comprehensively uses the multi-model fusion method, can flexibly combine various analysis and prediction methods, and implements self-adaptive optimization selection through the training of the encapsulation method, can be compatible with and utilize the existing analysis and prediction methods, and can automatically provide corresponding investment strategies For actual investment use or reference.

For the system proposed by the present invention, the actual data of the Shanghai Stock Exchange A-share trading market is simulated, using the trading day, opening price, highest price, lowest price, closing price, trading volume and turnover data of each stock, as well as the business scope of listed companies And quarterly reports, all investment operation strategies within two months are generated within 30 minutes, which is sufficient to meet the real-time requirements. The simulation investment income results are shown in Figure 6.

Figure 6 shows the investment results in February 2012, in which the blue line (IND%) represents the changes in the Shanghai A-share index (that is, the overall trend of the market), and the red line (NP%) represents the investment portfolio that was not pre-selected using network analysis The income situation, the green line (P%) indicates the income situation after using the network analysis to do the pre-selection of the investment portfolio. It can be seen that operating according to the investment strategy provided by this system may obtain excess profits, and the pre-selection by network analysis may contribute to obtaining excess profits.

Figure 7 shows the investment results in March 2012, reflecting the role of network analysis in reducing risk in a downtrend. As can be seen from the figure, the final return of the portfolio after preselection is higher than that without preselection (see left panel), and more importantly, in the period of negative growth rate (that is, the period of decline, see right panel), the growth rate after preselection The rate of loss (green line) is significantly better than the result of preselection without network analysis (red line). It can be seen that through preselection based on diversity, the loss trend during the decline can be contained to a certain extent, that is, the investment risk is reduced.

Claims (10)

1. the Stock Market method that Excavation Cluster Based on Network Analysis and multi-model merge, other are being, comprise following operation:

First from network, capture basic side information, build on this basis network node and be connected with network, build complicated community network model; Utilize the method for network analysis to select investment portfolio, the more related data of investment portfolio are input in multi-model fusion framework;

Described multi-model merges framework and comprises a plurality of submodels, each submodel is for the technological side information of the different characteristic capturing from network, carry out respectively the market trend prediction of different characteristics, generate predicted value separately, again by predicted value weighted sum, obtain comprehensive market trend predicted value, according to this value, generate corresponding investment strategy;

To submodel, provide the weight of the predicted value of the feature selecting device of information, parameter that submodel relates to, submodel all by single argument Estimation of Distribution Algorithm, to encapsulate training.

2. the Stock Market method that Excavation Cluster Based on Network Analysis claimed in claim 1 and multi-model merge, is characterized in that, the structure of described complicated community network model comprises following operation:

1.1) network node

In vector space model, the basic side information text capturing from network represents with the word bag of binary feature vector pattern, as follows:

inf _i＝(<t ₁,w _i1>,<t ₂,w _i2>,...<t _M,w _iM>)

Wherein M is the quantity of feature, w _iktext feature t _kweights, calculate weights by tf*idf method, to fixation features abbreviation, are inf _i=(w _i1, w _i2... w _iM);

To utilizing the text in the basic side information that the method for data mining obtains from network to proceed as follows:

1) filter: filter part useless in information;

2) participle: the information through filtering is divided into a plurality of vocabulary, the result after participle is deposited in lexicon, the part of speech of sign vocabulary;

3) vocabulary in lexicon is further stopped to word and process, comprise and remove function word and bind negative word;

Obtain after the text feature in the middle of basic side information, carry out the calculating of its weights, basic side finish message is become to vector space model, for time the information that becomes, this vector space model becomes a time-varying vector:

Inf _i(t)=(w _i1(t), w _i2(t) ... w _iM(t)), wherein t is time variable;

1.2) network connects

To network G (t)=(V (t), E (t)), the basic side information of use listed company as network node modeling, has V (t)={ inf to it _i(t) }, E (t)={ (i, j, edg _ij(t)) | i, j ∈ V (t) };

V (t) is for utilizing the set of the constructed network node of basic side information, and E (t) is two node i, j and the strength of joint edg between them in the middle of network node _ij(t) set;

Use cosine similarity $\cos ({\inf }_i\left(t\right),{\inf }_j\left(t\right))=\frac{\underset{t_n\&Element;T_M}{\mathrm{\&Sigma;}}{w}_{\mathrm{in}}\left(t\right)w_{\mathrm{jn}}\left(t\right)}{\sqrt{\underset{t_n\&Element;T_M}{\mathrm{\&Sigma;}}{w}_{\mathrm{in}}{\left(t\right)}^2\underset{t_n\&Element;T_M}{\mathrm{\&Sigma;}}{w}_{\mathrm{jn}}{\left(t\right)}^2}}$ Computational grid strength of joint, wherein T _mfor the complete or collected works of basic side information text feature, and use threshold value θ to filter, have ${\mathrm{edg}}_{\mathrm{ij}}\left(t\right)=\left\{\begin{array}{cc}0,& \cos ({\inf }_i\left(t\right),{\inf }_j\left(t\right))<\mathrm{\&theta;}\\ \cos ({\inf }_i\left(t\right),{\inf }_j\left(t\right)),& \cos ({\inf }_i\left(t\right),{\inf }_j\left(t\right))\&GreaterEqual;\mathrm{\&theta;}\end{array}\right.,$ θ gets cos45 °;

With as inf _i(t)=(w _i1(t), w _i2(t) ... w _iM(t) vector space model) is described network node, is built into needed network node;

Again by network node according to as V (t)={ inf _i(t) }, E (t)={ (i, j, edg _ij(t)) | i, j ∈ V (t) shown in internetwork connection mode connect, thereby complicated community network model, this model is a dynamic network model.

3. the Stock Market method that Excavation Cluster Based on Network Analysis claimed in claim 1 and multi-model merge, it is characterized in that, describedly utilize network analysis method to select investment portfolio to be to select the stock being independent of each other most to form investment portfolio, to comprise the following diversity division methods that detects cluster based on corporations:

Use corporations' detection method to divide, use Girvan-Newman clustering method to carry out network clustering, its evaluation index is modularity $Q\left(t\right)=\underset{i}{\mathrm{\&Sigma;}}(e_{\mathrm{ii}}\left(t\right)-{a_i}^2\left(t\right)),$ T is time variable;

Wherein, e _ij(t) represent to connect the shared ratio of weights that the i of corporations is connected with the network of network node in the j of corporations; a _i(t) the shared ratio of weights that the all-network that expression is associated with network node in the i of corporations is connected, comprises that two networks connect Quan corporations inside and only have a network to be connected to the inner two kinds of situations of corporations;

The complicated community network model that modularity has been described after dividing is big or small with the difference degree of stochastic network model, and the group dividing method of modularity maximum will be as optimum division;

The basic process of Girvan-Newman clustering method is as follows:

The first step: calculate intermediary's degree that all-network connects;

Second step: remove the network connection with maximum intermediary degree;

The 3rd step: intermediary's degree that the network that recalculating affects in second step connects;

The 4th step: if do not have rest network to connect, finish algorithm, otherwise turn to second step;

Finally, in the com of each network corporations (t) from optimal dividing, select a representative, form the investment portfolio of required diversification;

Finally, Fundamental Analysis selects the Policy model of investment portfolio to be expressed as:

$\mathrm{IFA}\left(t\right)=\left\{n\right|\&ForAll;\mathrm{com}\left(t\right),n=\arg {\underset{i\&Element;\mathrm{com}\left(t\right)}{\max }}^N\left({\mathrm{rep}}_i\left(t\right)\right)\}.$

4. the Stock Market method that Excavation Cluster Based on Network Analysis claimed in claim 1 and multi-model merge, it is characterized in that, describedly utilize network analysis method to select investment portfolio to be to select the stock being independent of each other most to form investment portfolio, to comprise the following diversity dividing method based on the irrelevant subnet of maximum full-mesh:

Adopt Bron – Kerbosch algorithm to extract the most complete works of connexon net in net mending, its base form is the searching algorithm of a recursive backtracking, and flow process is as follows:

Bron-Kerbosch algorithm:

Step1: given three set (R, P, X), initialization set R, it is empty that X is respectively, and set P is the set of all-network node;

Step2: if set P, it is empty that X is respectively, and exporting R is Clique;

Step3: for each from set obtain P network node v}, has following processing:

1) by network node v} is added in set R, set P, { it is crossing that v} obtains adjacent_lattice node set N{v}, recursive set R afterwards, P, X (turning Step2) for X and network node;

2) from set, delete network node { v}, and { v} adds in set X by network node P;

Now Fundamental Analysis selects the Policy model of investment portfolio to be expressed as:

the most complete works of connexon net in the net mending generating according to Bron-Kerbosch algorithm.

5. the Stock Market method that Excavation Cluster Based on Network Analysis claimed in claim 1 and multi-model merge, it is characterized in that, the framework that described multi-model merges is that a complicated system is turned to several subsystems, a corresponding submodel of subsystem, the same model of common description that then submodel corresponding to these subsystems combined is to improve model-fitting degree;

Described submodel method of attachment adopts weighted sum mode, and the output of each submodel is sued for peace by certain weights, obtains final output;

Described submodel can be added and subtracted or replace, and through the training of master data, block mold can be by regulating self submodel weights size to carry out adaptive selection to forecast model;

Described submodel comprises following several:

1) trend forecasting method based on vector symbol sequence

First adopt least square fitting method to carry out vector quantization to historical share price data, if definition x _ithe price of day is y _i, minimize the error of n day $S(a,b)=\underset{i=0}{\overset{n}{\mathrm{\&Sigma;}}}{(y_i-f\left(x_i\right))}^2,f\left(x_i\right)={\mathrm{ax}}_i+b,$ Can obtain the trend characterizing with slope, be defined as:

Further the continuous trend vector of discretize, therefrom to extract macroscopical tendency information, adopts unsupervised clustering method, for the singularity of the data of vector quantization, carries out the heavy clustering algorithm based on k-means, obtains cluster result and all kinds of center vector;

2) the trend prediction algorithm extracting based on Stock Index Time Series turning point

On the basis of time series fluctuation characteristic point formal definition, be provided with time series X={x ₁, x ₂..., x _n, the fluctuation characteristic point x of X within its time interval is { xVD (x)=max (VD (x _i)), i=1,2 ..., N}, wherein

$\mathrm{VD}\left(x_i\right)=|x_i-(x_1+(x_N-x_1)-\frac{i-1}{N-1})|;$

3) the investment recommendation algorithm of judging based on vocabulary emotion tendency

To any vocabulary w _i∈ W, note:

$T\left(w_i\right)\widehat{=}\left\{\begin{array}{c}1,\mathrm{if}{w}_i\mathrm{is\; positive},\\ -1,\mathrm{if}{w}_i\mathrm{is\; negative}.\end{array}\right.$

Claim T (w _i) be w _iemotion tendency value, given emotion tendency relational network SORN=(W, C, Q), in SORN one from w _ito w _jpath on the vocabulary numbered sequence of process be designated as (p ₁, p ₂..., p _s), 2≤s≤n wherein, p ₁=i, p _s=j;

If set { (p ₁, p ₂..., p _s) in an element (h ₁, h ₂..., h _s) meet: $q_{h_1{h}_2}+q_{h_2{h}_3}+...+q_{h_{s-1}{h}_s}=\min \{q_{p_q{p}_2}+q_{p_2{p}_3}+...+q_{p_s-1p_s}\},$ Claim (h ₁, h ₂..., h _s) be from w in SORN _ito w _ja minimum noise path on the vocabulary numbered sequence of process, be designated as D (i, j); Claim (D (i, 1) ..., D (i, i-1), D (i, i+1) ..., D (i, n)) and be the sequence of D (i, j), be designated as D (i); The dijkstra's algorithm that adopts adjacency list and Fibonacci heap to realize calculates D (i);

Select the minimum path of noise between two vocabulary to identify this emotion tendency relation to vocabulary, when language material enough abundant, right utilize Q just can obtain D (i, j), recycling C just can calculate w _iand w _jemotion tendency relation; In language material, positive vocabulary is always many than negative vocabulary.

6. the Stock Market method that Excavation Cluster Based on Network Analysis claimed in claim 5 and multi-model merge, it is characterized in that, in described clustering algorithm, each trend vector is replaced with the corresponding vector of cluster centre at its place, and this characterization vector the type of trend, if be characterized by symbol, finally can be by the discrete series of sign sequence that changes into of continually varying trend vector; Use the day of trade, opening price, highest price, lowest price, closing price, trading volume and the transaction value of stock as the general data of technical Analysis, the historical information of stock S is expressed as symbol sebolic addressing: { S _i, $S_i=\{s_i^d,s_i^o,s_i^h,s_i^l,s_i^c,s_i^v,s_i^t\};$

Wherein the respectively corresponding day of trade, opening price, highest price, lowest price, closing price, trading volume and transaction value data at i this strand of day;

And adopted the Support Vector Machine model of use radial basis kernel function to learn historical stock price trend symbolic model, SVM is theoretical to be pursued through study and obtains having the model of strong generalization ability, after obtaining support vector SV, need to obtain optimal classification lineoid OHP, this solution procedure is a quadratic programming problem again;

If SV is from OHP:(lx) the nearest sample point of+b=0, and of a sort SV is completely equal from OHP distance, and inhomogeneous SV not necessarily equates from OHP distance;

To m training sample (x ₁, y ₁), (x ₂, y ₂) ..., (x _m, y _m) ask its classification lineoid, classification lineoid must meet the condition of optimal classification lineoid:

In order to find optimal classification lineoid, by Lagrange function, former problem is changed into and solves standard form quadratic programming problem:

$\max W\left(\mathrm{\&alpha;}\right)=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i-\frac{1}{2}\underset{i,j=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j{y}_i{y}_{i}K(x_i,x_j)$

$s.t.\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{y}_i=0,{\mathrm{\&alpha;}}_i\&GreaterEqual;0,(i=\mathrm{1,2},...,m)$

Ask the key of optimum lineoid to be to obtain and meet α _ithe α of >0 _iand

And optimal classification lineoid is:

$f\left(x\right)=\mathrm{sign}\{\underset{{\mathrm{\&alpha;}}_i>0}{\mathrm{\&Sigma;}}{\mathrm{\&alpha;}}_i{y}_{i}K(x_i,x)-b_0\}$

Wherein, α _ithe sample point that >0 is corresponding is support vector, for quadratic programming problem, can take positive method, Dual Method or interior some Algorithm for Solving; Support Vector Machine after training is automatically classified to new data, the symbol of output category result, and the slope of the corresponding former cluster centre vector of this symbol, is the predicted value of final trend.

7. the Stock Market method that Excavation Cluster Based on Network Analysis claimed in claim 5 and multi-model merge, is characterized in that, the step of described time series fluctuation characteristic point extraction algorithm is as follows:

The first step: input starting point coordinate start and the terminal point coordinate end of sequence to be extracted, judge whether distance between start and end meets subsequence and be less than smallest interval length, turn the 3rd step if meet; If do not meet and find the maximum point of VD value between Origin And Destination by the definition of fluctuation characteristic point, if VD is greater than the end condition of algorithm amplitude, this point is joined in fluctuation characteristic point result sequence as fluctuation characteristic point, and this some note is done to fp;

Second step: with fp, former sequence is divided into two sections, start, to fp subsegment, fp to end subsegment, carries out the first step to these two subsegments;

The 3rd step: fluctuation characteristic is put to result sequence and according to time sequence preserve afterwards and output;

The 4th step: judge minimum interval threshold value based on turning point, extract turning point subset by maximum/minimum value principle in fluctuation characteristic point result sequence;

The 5th step: based on turning point subset, between two continuous turning points, based on fluctuation characteristic point sequence, by the maximum principle of reverse wave amplitude, finding turning point arbitrarily, and join in turning point subset, repeat aforesaid operations, until cannot find again new turning point by the turning point extracting parameter of setting;

The 6th step: turning point subset is according to time sequence preserved and output afterwards;

The turning point seed more than extracting has characterized the change of stock price trend under larger time scale, by machine learning methods such as neural network or support vector machine, train, obtain optimum classifier, this sorter can automatically be classified to new data, whether the corresponding date of output turning point, the i.e. result to future trend prediction, is trend prediction value corresponding to nearest turning point the predicted value of final trend.

8. the Stock Market method that Excavation Cluster Based on Network Analysis claimed in claim 5 and multi-model merge, it is characterized in that, the described investment recommendation algorithm of judging based on vocabulary emotion tendency adopts SWSOA algorithm, this algorithm, from any one node of the largest connected subgraph of SORN, just carries out emotion tendency classification to all nodes in this subgraph; SWSOA algorithm be input as a concrete SORN, this algorithm table is shown as to function SWSOA (SORN), wherein SORN represents SORN variable;

The concrete steps of this algorithm are as follows:

Step 1. utilizes breadth First ergodic algorithm to obtain the largest connected subgraph Gs of SORN, and the set that the lexical node comprising in Gs forms is designated as W _gS;

Step 2. is specified W _gSin any one node w _i, and make W _gS=W _gS-{ w _i, U={w _iand

Step 3. is calculated D (i) and is obtained w _ito W _gSthe minimum noise path of middle arbitrary node;

Step 4. couple W _gSin each lexical node w _j, carry out a)～c):

A), according to the D (i, j) in D (i), calculate from w _ito w _jminimum noise path on the turnover of process

The quantity e that is related to limit, $e=\underset{l=1}{\overset{l=s-1}{\mathrm{\&Sigma;}}}{\mathrm{\&delta;}}_{h_l{h}_{l+1}}{c}_{h_l{h}_{l+1}};$

B) calculate w _jwith w _itendentiousness relation: T (w _j)=(-1) ^e* T (w _i);

C) if T is (w _j)=T (w _i), U=U ∪ { w so _j, otherwise V=V ∪ { w _j;

Step 5.WSO has classified, and classification results leaves in respectively in U and V, if | U|>|V|, in U, depositing so in positive vocabulary and V and depositing negative vocabulary, vice versa;

Extract comment news report and the comment relevant to financial market, utilize SWSOA algorithm to differentiate its emotion tendency, the listed company and the corresponding stock thereof that are judged to be positive are candidate's thighs that recommendation should be paid close attention to;

For the predicted value that is judged to be the final trend of the listed company of positive and the stock of correspondence thereof, be 1;

For the predicted value that is judged to be the final trend of the listed company of negative and the stock of correspondence thereof, be-1.

9. the Stock Market method that Excavation Cluster Based on Network Analysis claimed in claim 5 and multi-model merge, it is characterized in that, consider that multi-model merges calculated amount and the complexity that makes whole model, also comprise based on single argument and distribute and estimate training, optimization and the feature selection approach of method for packing:

Described single argument Estimation of Distribution Algorithm, for data, encodes with a n (n=N+Dim) dimension chromosome in algorithm; Chromosome is divided into two parts, and first is packed algorithm parameter coded portion, altogether N position binary code; And enable elite's retention process, and every stock is stored respectively to historical optimal value and carry out initialization to improve Searching efficiency, guarantee the real-time of technical Analysis process: to population q (t), t is time variable, if it is a m*n rank matrix, recorded the chromosome of arranging from high to low according to fitness value Q9, the capable [q of front r ₁q ₂q _r] ^tbe used as elite and remain, the distribution probability matrix of estimating is P=[P ₁p ₂p _n], wherein: r is for participating in the chromosome quantity of variation, and its step is as follows:

The first step: initialization population q (0), reads historical optimal value and adds initial population, iterations t=0;

Second step: the order that chromosome population q (t) successively decreases according to fitness value sorts to chromosome, and a front r chromosome is retained as elite;

The 3rd step: before calculating, each row of the individual chromosome of R (R is generally greater than r) are got 1 probability P _i, form probability matrix;

The 4th step: with probability P _ione of random generation meets the numeral of 0-1 distribution as chromosome i position, then generates (m-r) individual chromosome with same algorithm, makes t=t+1, by the synthetic new population q (t) of these chromosomes and r the genome remaining;

The 5th step: judgement t>N, if set up, finish algorithm, if be false, turn to second step;

Wherein parameter optimization, partly for application encapsulation, need to be carried out encoding and decoding to parameter, and wherein in multi-model fusion, the scale-of-two decoding method of the weights of each submodel is as follows:

1) weight-codings

Initialization weights are exactly 0 and 1 long coded sequence of one group of n * m dimension, and wherein m is for needing the weights number of optimization, and n is the coding figure place of each weights, and the code length of each weights is identical;

2) weights decoding

Weight w _ispan be w _i∈ [0,1], the coding of supposing i weights is [q ₁q ₂q _n], q wherein _i∈ 0,1}, i ∈ 1,2 ..., n}

W' _i=q ₁* 2 ⁰+ q ₂* 2 ¹+ ... + q _n* 2 ^n-1

Then by w' _ibe mapped to interval [0,1],

$w_i=\frac{w_i^{\&prime;}}{2^n}$

Described method for packing, the data set being formed by technological side information is divided into training set test set two parts at random, population is decoded, utilize the gene phenotype of feature partly to select character subset to train, obtain the submodel that parameter is set, utilize the gene phenotype part of feature to carry out feature selecting to test set data simultaneously, test set after feature selecting is input to the submodel training to be tested, according to the classification performance of test set, carry out fitness assessment, if meet the end condition that single argument distributes and estimates, export optimum submodel and character subset, otherwise upgrade probability distribution vector by single argument Estimation of Distribution Algorithm, again encode and generate new population.

10. the Stock Market method that Excavation Cluster Based on Network Analysis claimed in claim 1 and multi-model merge, is characterized in that, the comprehensive market trend predicted value design synthesis investment strategy of described utilization is based on following modifying factor:

$r\left(t\right)=\{\begin{array}{c}1,\mathrm{if}{P}_s\left(t\right)\&times;(1+Q_s\left(t\right))>1\\ P_s\left(t\right)\&times;(1+Q_s\left(t\right)),\mathrm{otherwise}\end{array},$

Wherein P and Q distinguish accuracy rate and the evaluation of estimate of test set in corresponding training, modifying factor r (t) has reacted the confidence level of investor to this prediction, it is more confident to be that the higher investor of value of r (t) predicts the outcome to this time, and anti regular will tend to suspect and predicts the outcome and keep former strategy; If the t stock quantity that investor holds is constantly N (t), the cash quantity of holding is c (t), and the stock price data cp (t) of usining represents VAL (t) as its price, and functional symbol is Sig (t), to a stock, can be by based on investment strategy, mathematicization is expressed as:

$\mathrm{Sig}\left(t\right)=\left\{\begin{array}{c}\mathrm{Buy},\mathrm{if}{Y}_s\left(t\right)>0\\ \mathrm{Sell},\mathrm{if}{Y}_s\left(t\right)<0\\ \mathrm{Sig}(t-1),\mathrm{if}{Y}_s\left(t\right)=0\end{array}\right.;$

Using closing price above as operation standard, thereby final total assets is a (t)=c (t)+N (t) * cp (t);

To investment portfolio, then further consider the stock preliminary election IFA that gives network analysis, the expectation of definition investment analysis is E (ITA (t, s))=Y _s(t) * r _s(t), Y wherein _sfor comprehensive market trend predicted value, r _sfor to result confidence level, it can be pressed the dealing money of stock s distribute, corresponding general investment strategy is expressed as:

$\&ForAll;s\&Element;\mathrm{IFA}\left(t\right),$

$\mathrm{Sig}(t,s)=\{\begin{array}{c}\mathrm{Buy}(s,\mathrm{ITA}(t,s)),\mathrm{if\; ITA}(t,s)>{\mathrm{\&theta;}}_{+}\\ \mathrm{Sell}(s,|\mathrm{ITA}(t,s)\left|\right),\mathrm{if\; ITA}(t,s)<\mathrm{\&theta;}\_\\ \mathrm{Sig}(t-1,s),\mathrm{otherwise}\end{array},$

①:Sig(t,s)＝Buy(s,r)②:Sig(t,s)＝Sell(s,r)

where?r＝ITA(t,s)