CN108734220A - Adaptive Financial Time Series Forecasting method based on k lines cluster and intensified learning - Google Patents
Adaptive Financial Time Series Forecasting method based on k lines cluster and intensified learning Download PDFInfo
- Publication number
- CN108734220A CN108734220A CN201810501626.9A CN201810501626A CN108734220A CN 108734220 A CN108734220 A CN 108734220A CN 201810501626 A CN201810501626 A CN 201810501626A CN 108734220 A CN108734220 A CN 108734220A
- Authority
- CN
- China
- Prior art keywords
- lines
- length
- cluster
- hachure
- price
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
- G06Q40/04—Trading; Exchange, e.g. stocks, commodities, derivatives or currency exchange
Abstract
The invention discloses a kind of adaptive Financial Time Series Forecasting methods based on k lines cluster and intensified learning, first obtain finance data, the processing of K lineizations is carried out to finance data, and treated that data are calculated for K lineizations, obtains the K line number evidences in the current matching period；The each subdivision of K lines is clustered using Kmeans clustering algorithms, FCM clustering algorithms or on-line talking method based on packing density；Cluster result is input in deeply learning model and carries out parameter training, financial transaction is carried out using trained deeply learning model.Finance data is carried out K lines by the present invention, and is clustered to each subdivision of K lines, and cluster result is input to deeply learning model, obtains the online adaptive prediction that real time financial transaction price is realized based on the deeply learning model for decomposing k lines cluster.
Description
Technical field
The present invention relates to financial transaction fields, and in particular to it is a kind of based on decompose k lines cluster and deeply study from
Adapt to Financial Time Series Forecasting method.
Background technology
It realizes that automated transaction in many developed countries is very universal phenomenon by computer, people is replaced with computer
On the one hand it is for transaction because computer can find the rule that some are difficult or can not capture from huge historical data
Rule and phenomenon, on the other hand can greatly reduce the influence of dealer's mood swing, avoid doing in the case where market is extreme
Go out unreasonable decision.With the rapid development of artificial intelligence, automated transaction is trained by artificial intelligence technology
Model is each thing be engaged in financial transaction and dreamed of with tactful research staff, although the technologies such as machine learning are known in image
Not, the fields such as speech recognition and medical diagnosis achieve many achievements, but the technologies such as machine learning are applied directly to automatically
It is not but an easy thing to change in transaction.Different from common monitoring learning method, carries out automated transaction and may not be used
For the label of study, so just needing by going the feedback constantly explored, and pass through environment in a unknown environment
It constantly updates and Optimal Decision-making, this is also exactly the mode of learning of intensified learning.
The paper that DeepMind was delivered in 2013 " Playing Atari with Deep Reinforcement
It is illustrated in Learning " and how to go to play Atari game with intensified learning training computer, the model obtained by training exists
When playing 7 kinds of different game, wherein 3 kinds of computer games have been even more than the level of the mankind.After 2 years, they again carry out model
It improves, goes to play 49 kinds of different game by this model, the game of half has surmounted the level of the mankind.By 2017
Just, DeepMind is utilized the technological development such as deeply study AlphaGo and defeats the top go hand in the world again.Strengthen
As a branch inside machine learning, it is mainly used to solve the problems, such as Markovian decision type for study.It usually can be with
Intensified learning is divided into two classes, one is based on policy learning, one is based on value study.Based on the normal of value study
There are Q-Learning, TD-Learning and Sarsa with method, they obtain a cost function by study, give cost function
Then input state chooses action by the height of this value, it is dynamic to be frequently utilized for solution state come the value acted
Make the limited discrete data problem in space, and achieves good effect.Although the intensified learning based on value solves greatly
Amount problem, but be difficult that it is applied directly in the transaction in financial market.First, there is a large amount of for the data on financial market
Noise and uncertainty, therefore this can cause such as Q-learning of the intensified learning based on value that can go out in learning process
Existing a series of problems；Secondly as the environment in financial market is too complicated, so that it cannot being approximately one discrete by it
Space；In addition, as Q-learning etc. based on the intensified learning of value when calculating can also use discount to future value, and
The state in market will not generally change because of the action of transaction, and future value is difficult to act；Finally, when
When facing excessive state or motion space, use value function also results in appearance " dimension disaster ".In contrast,
When using intensified learning based on strategy, can from continuous either large-scale state direct output action or action it is general
Rate is distributed；Also, it can select optimal decision without the use of Future Information according to current ambient condition；In addition, it
It, can be more flexible when carrying out parameter optimization to object function.Arbitrarily complicated company has may be implemented in artificial neural network
Continuous function approaches, and deep neural network then can promote the ability of expression by increasing hidden layer, it is possible to pass through
Function approximator of the deep neural network as the intensified learning model based on policy learning.
In recent years, the technologies such as intensified learning are applied also has many theoretical results and practical application with financial transaction field.
Deep learning is applied to the character representation in financial market by Deng et al. for the first time, and builds financial friendship in conjunction with intensified learning
Easy system.Lu carries out characteristic processing and autonomous transaction with LSTM and intensified learning technology.Gabrielsson et al. are by Nison
The Japanese candle line of proposition is merchandised as technical indicator, and by recurrence intensified learning to carry out the high frequency in stock price index futures market.
In short, no matter which type of builds Trading Model in a manner of, the feature in financial market is handled and is indicated all be extremely
A crucial step.Moody et al in his model are used as by price by intensified learning application and transaction system earliest
State is input in model and is trained, although containing all information in price, there is not true in financial market
It is qualitative, for example, the variation of economic policy, company's falseness information, these uncertainties can all influence tendencies of price, so right
It is a vital step that the state of input, which carries out denoising,.Deng et al. by fuzzy learning apply at the denoising to price
Reason can remove noise by the blurring to price, but will increase state space by blurring to a certain extent, compared with
Hardly possible summarizes current general trend of market development feature.Thus how selected characteristic and learnt be when build model extremely key,
In addition the transaction in market is a kind of online dynamic decision behavior, so how Optimal Decision-making makes the maximum revenue of acquisition still
It is technical problem to be solved.
Invention content
In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides one kind based on decomposition k lines cluster and deeply
Finance data is carried out K lines, and is clustered to each subdivision of K lines by the adaptive Financial Time Series Forecasting method of study,
The characterization ability of data can be increased and data are further subjected to denoising, cluster result, which is input to deeply, learns mould
Type can preferably carry out real time data the adaptive optimization of the update and parameter to decision, obtain based on decomposition k lines cluster
Deeply learning model, realize real time financial transaction price online adaptive prediction.
The technical solution adopted in the present invention is：
A kind of adaptive Financial Time Series Forecasting method based on decomposition k lines cluster and deeply study, this method
Include the following steps：
Step 1：Financial time series are obtained, the processing of K lineizations are carried out to financial time series, and treated to K lineizations
Data are calculated, and the length of each subdivision of K lines in the current matching period is obtained；
Step 2：Using Kmeans clustering algorithms, FCM clustering algorithms or on-line talking method based on packing density to K lines
The length of each subdivision is clustered, and the cluster centre corresponding to the length of each subdivision of K lines is obtained；
Step 3：It is input to the solid color of cluster result and K lines as learning characteristic in deeply learning model
Parameter training is carried out, financial transaction is carried out using trained deeply learning model.
Further, the K lines include the color of upper hachure, lower hachure, entity and entity.
Further, in the step 1, and the data of K lineizations processing are calculated, is obtained in the current matching period
The step of length of each subdivision of K lines includes：
Highest price, opening price, lowest price and the closing price that the period is corresponded to according to K lines, were calculated in the current matching period
The upper hachures of K lines, lower hachure and entity length；
When the solid color of K lines is red, the upper hachure length of K lines is equal to the difference of highest price and closing price, K lines
Lower hachure length is equal to the difference of opening price and lowest price, and the physical length of K lines is equal to the difference of closing price and opening price；
When the solid color of K lines is green, the upper hachure length of K lines is equal to the difference of highest price and opening price, K lines
Lower hachure length is equal to the difference of closing price and lowest price, and the physical length of K lines is equal to the difference of opening price and closing price.
Further, in the step 2, exist using Kmeans clustering algorithms, FCM clustering algorithms or based on packing density
Line clustering method respectively clusters the upper hachure length of K lines, lower hachure length and physical length, obtains the upper shadow of K lines
Cluster centre corresponding to line length, lower hachure length and physical length.
Further, it is described using Kmeans clustering algorithms to the upper hachure length, lower hachure length or physical length of K lines
The step of being clustered include：
Using the upper hachure length of the K lines in the current matching period, lower hachure length or physical length as input sample, structure
Build training sample set K；
Training sample set K is divided into 5 classes, randomly selects in each class input sample as such initial clustering
Center；
Calculate in each class the square distance of input sample and initial cluster center and, obtain Kmeans object functions, it is right
Kmeans object functions carry out seeking local derviation, calculate the minimum value of the derivative, obtain the cluster centre of each class.
Further, it is described using FCM clustering algorithms to the upper hachure length, lower hachure length or physical length of K lines into
Row cluster the step of include：
Using the color of the upper hachure of the K line number evidences in the current matching period, lower hachure, the length of entity or entity as defeated
Enter sample, structure training sample set K；
Training sample set K is divided into 5 classes, randomly selects in each class input sample as such initial clustering
Center calculates degree of membership of each input sample to all classes, the cluster belonging to input sample is judged by the size of degree of membership
Center；
Calculate in each class input sample and initial cluster center square distance and with degree of membership product, obtain FCM targets
Function carries out FCM object functions to seek local derviation, calculates the minimum value of the derivative, obtain the cluster centre of each class.
Further, the on-line talking method based on packing density is to upper hachure length, lower hachure length or entity
The step of length is clustered include：
Using the color of the upper hachure of the K line number evidences in the current matching period, lower hachure, the length of entity or entity as defeated
Enter sample data；
Calculate the density of each new input sample；
When initialization, chooses current first sample point and to carry out parameter initialization to cluster structure and gather as first
Class center；
Judge that the density of each new input sample is close with the maximum value of all cluster centre density or with all cluster centres
The size of the minimum value of degree, if the density of the input sample is more than the maximum value of all cluster centre density or is less than all clusters
The input sample is then new cluster centre by the minimum value of center density；Otherwise, it is all poly- with other to calculate the input sample
The distance at class center, using the cluster centre of minimum range as the cluster centre belonging to the input sample.
Further, the solid color using cluster result and K lines is input to deeply as learning characteristic
Practising the step of parameter training is carried out in model includes：
According to the sliding time window of setting, the upper hachure length for the K lines that cluster is obtained, lower hachure length, entity are long
The solid color for spending corresponding cluster centre and K lines is input to deep learning network model progress characterology as learning characteristic
It practises；
Transaction movement is obtained by nitrification enhancement, and is applied in environment, using activation primitive by depth nerve
Network converts state, obtains the award of environmental feedback, and store into the data base of deeply learning model, and every
The parameter of deeply learning model is updated every the transaction of certain number；
The weights of deep learning network are initialized using Xavier initial methods；
Object function is constructed, the weights of deep learning network are updated；
Pass through the continuous interaction of deeply learning model and environment, study to the deep learning net with high trade decision
Network model.
Further, the structure object function, the step of being updated to the weights of deep learning network include：
Award based on environmental feedback, constructs object function, and object function is：
maxθL (θ)=∑ log π (at|st, θ) and rt
In formula, atFor the transaction movement of t moment；stFor the state of t moment；θ is the weighting parameter in deep learning network；
rtFor the award of t moment environmental feedback；
Object function is optimized using Adam optimization algorithms, deep learning is updated by the object function after optimization
The weights of network.
A kind of computer installation is used for adaptive Financial Time Series Forecasting, including memory, processor and is stored in
On reservoir and the computer program that can run on a processor, which is characterized in that the processor is realized when executing described program
Following steps, including：
Obtain financial time series, the processing of K lineizations carried out to financial time series, and to K lineizations treated data into
Row calculates, and obtains the length of each subdivision of K lines in the current matching period；
Using Kmeans clustering algorithms, FCM clustering algorithms or on-line talking method based on packing density to each height of K lines
Partial length is clustered, and the cluster centre corresponding to the length of each subdivision of K lines is obtained；
The solid color of cluster result and K lines is input to as learning characteristic in deeply learning model and is joined
Number training carries out financial transaction using trained deeply learning model.
Compared with prior art, the beneficial effects of the invention are as follows：
(1) financial time series are carried out K lines by the present invention, and are clustered to each subdivision of K lines, and data can be increased
Characterization ability and data are further subjected to denoising, cluster result is input to deeply learning model, can be to real-time
Data preferably carry out the adaptive optimization of the update and parameter to decision, obtain based on the deeply for decomposing k lines cluster
Model is practised, the online adaptive prediction of real time financial transaction price is realized；
(2) present invention carries out resolution process to k lines, and for each subdivision of k lines, treated that result is clustered, and passes through
Input state of the corresponding cluster centre that each subdivision of k lines is clustered as model, and it is arranged one for input state
A sliding time window carries out feature learning by deep neural network, and intensified learning carries out decision execution, finally again by not
The disconnected implementation effect that Optimal Decision-making is interacted to environment, carries out the weight of deep neural network by Xavier methods
Initialization, enables deeply learning model to converge to an ideal effect as early as possible, is trained in the gradient to model
When, the weights of neural network are updated using Adam methods, the update to decision can be preferably carried out to real time data
With the adaptive optimization of parameter, better effect is obtained.
Description of the drawings
The accompanying drawings which form a part of this application are used for providing further understanding of the present application, and the application's shows
Meaning property embodiment and its explanation do not constitute the improper restriction to the application for explaining the application.
Fig. 1 is based on the adaptive Financial Time Series Forecasting method flow diagram for decomposing k lines cluster and deeply study；
Fig. 2 is K cable architecture schematic diagrames；
Fig. 3 is DRL model framework figures；
Fig. 4 is the price trend figure of three kinds of experimental datas；
Fig. 5 is the deeply learning performance figure of different input feature vectors.
Specific implementation mode
The invention will be further described with embodiment below in conjunction with the accompanying drawings.
It is noted that following detailed description is all illustrative, it is intended to provide further instruction to the application.Unless another
It indicates, all technical and scientific terms used herein has usual with the application person of an ordinary skill in the technical field
The identical meanings of understanding.
It should be noted that term used herein above is merely to describe specific implementation mode, and be not intended to restricted root
According to the illustrative embodiments of the application.As used herein, unless the context clearly indicates otherwise, otherwise singulative
It is also intended to include plural form, additionally, it should be understood that, when in the present specification using term "comprising" and/or " packet
Include " when, indicate existing characteristics, step, operation, device, component and/or combination thereof.
As background technology is introduced, exists in the prior art and how selected characteristic and to be learnt in structure model,
How maximum revenue of the Optimal Decision-making to make acquisition the problem of, in order to solve technical problem as above, present applicant proposes one
Kind is based on the adaptive Financial Time Series Forecasting method for decomposing k lines cluster and deeply study.
A kind of exemplary embodiment of the application, as shown in Figure 1, providing a kind of strong based on decomposition k lines cluster and depth
The adaptive Financial Time Series Forecasting method that chemistry is practised, this approach includes the following steps：
Step 101：Financial time series are obtained, the processing of K lineizations are carried out to financial time series, and to the processing of K lineizations
Data are calculated, and the length of each subdivision of K lines in the current matching period is obtained；
Step 102：Using Kmeans clustering algorithms, FCM clustering algorithms or on-line talking method based on packing density to K
The length of each subdivision of line is clustered, and the cluster centre corresponding to the length of each subdivision of K lines is obtained；
Step 103：It is input to deeply learning model using the solid color of cluster result and K lines as learning characteristic
Middle carry out parameter training carries out financial transaction using trained deeply learning model.
Adaptive financial trade method disclosed by the embodiments of the present invention based on decomposition k lines cluster and deeply study,
Based on the deeply learning model for decomposing k lines cluster, the online adaptive prediction of real time financial transaction price is realized.In order to
The noise for including in market information is reduced, price is carried out the processing of k lineizations by the present invention first, and k lines are using in the specific time period
The price of several feature time points describes this section of time price trend, plays the effect of preliminary denoising, additionally by more k lines
Combination is also used as a kind of signal of price change, then, resolution process is carried out to k lines, is handled for each subdivision of k lines
Result afterwards is clustered.It is used as the input shape of model by the corresponding cluster centre clustered to each subdivision of k lines
State.Due to obtaining the result is that one section of continuous time series, thus also need that a sliding time window is arranged for input state,
Feature learning is carried out by deep neural network, intensified learning carries out decision execution, finally again by constantly being handed over environment
The implementation effect for mutually carrying out Optimal Decision-making also passes through to allow model that can converge to an ideal effect as early as possible in the present invention
Xavier methods initialize the weight of deep neural network, in addition when the gradient to model is trained, use
Famous Adam (adaptive moment estimation) method.
A more detailed embodiment is set forth below in order to make those skilled in the art be better understood by the present invention, this
Inventive embodiments provide a kind of based on the adaptive financial trade method for decomposing k lines cluster and deeply study, this method packet
Include following steps：
Step 201：Financial time series k lines
K line charts originate from the rice market transaction of Japanese feudal times, for calculating the daily ups and downs of the price of rice, by more than 300 years
Development, k line charts have been widely used for the securities markets such as stock, futures and foreign exchange, and the friendship of long campaigns related work
Easy person has summed up certain law by some special areas or form of k line charts, the form gone out by these regularity summarizations
There is great probability to imply the rise or drop of price, some common forms have cross star, tup etc..
The processing of K lineizations is carried out firstly the need of by the financial time series received (Open, High, Low, Close), later
To obtained K lineizations, treated that data calculate, obtain the upper hachure length, lower hachure length and physical length of K lines with
And the color of K line entities (0 represents red, and 1 represents green).The data of the t moment finally obtained can be expressed as：
kt=(ut, lt, bt, ct) (1)
Wherein, ut, lt, bt, ctRespectively the upper hachure length, lower hachure length, physical length of the K lines in corresponding diagram 2 and
The color of entity.
In the present embodiment, the price (highest price, opening price, lowest price, closing price) of period is corresponded to according to K lines to calculate
To the length of each subdivision, the upper hachure of the K lines obtained from, the length of lower hachure and entity；
When solid color is red, upper hachure length=(highest price-closing price), lower hachure length=(opening price-is minimum
Valence), physical length (closing price-opening price)；
When solid color is green, upper hachure length=(highest price-opening price), lower hachure length=(closing price-is minimum
Valence), physical length (opening price-closing price).
Step 202：Decompose k lines cluster
After the upper hachure length, lower hachure length and physical length for obtaining K lines, next just need to the upper of K lines
Hachure length, lower hachure length and physical length carry out clustering processing.It is calculated using static Kmeans clustering algorithms, FCM clusters
Treated that result clusters to each subdivision for method and on-line talking method based on packing density.Kmeans algorithms are
Most classical one of clustering algorithm in clustering algorithm, it is efficient due to the algorithm, so quilt when data are clustered
Extensive use.FCM be it is a kind of with degree of membership come determine each data point belong to some cluster degree algorithm, such clustering algorithm
It is a kind of improvement to the hard clustering algorithm of tradition.Dynamic cluster method based on packing density is one kind by recursive calculation data
Density and the clustering method that data division is carried out by packing density, it can be with adjust automatically cluster centre, and online updating
Parameter need not be iterated training, in addition, it can ignore past data while study to current data input
And only retain important information, new trend can be adapted to rapidly with this.
(1) the upper hachure length, lower hachure length and physical length of k lines are clustered using Kmeans clustering methods
When being clustered by Kmeans clustering methods, using the upper hachure length of K lines as input sample, training sample is built
Training sample concentration data are divided into 5 classes, then randomly select the initial cluster center of every one kind, the sample each inputted by this collection
This is indicated with the cluster centre belonging to it.It is the object function to Kmeans below, and by asking inclined to object function
The formula of the cluster centre on this corresponding to hachure can be obtained by leading：
In formula (2), ckIndicate k-th of cluster centre, xiFor the i-th input sample；The process of Kmeans clustering algorithms is just
It is first k cluster centre point of initialization, then continuous iteration, makes each sample point to the cluster belonging to it by object function
The distance of central point is minimum, wherein in formula (3), NjIndicate the number of sample point in jth class.
The lower hachure length of k lines and the length of entity are clustered using Kmeans clustering methods, obtain the lower shadow of k lines
The method of cluster centre corresponding to line length and physical length with using Kmeans clustering methods to the upper hachure length of k lines into
Row clustering method is similar, repeats no more.
(2) FCM clustering methods is used to cluster the upper hachure length, lower hachure length and physical length of k lines
When being clustered with FCM, using the upper hachure length of K lines as input sample, training sample set is built, will be trained
Sample intensive data is divided into 5 classes, and degree of membership of each sample to all classes is then calculated, is determined by the size of degree of membership
Determine which cluster centre it belongs to.It is the object function of FCM below, and derivation is carried out to object function and obtains the upper shadow of K lines
Cluster centre corresponding to line：
Wherein in (4) formulaIt is the degree of membership that sample j belongs to the i-th class center, the m in formula is a Weighted Index, one
As application section be [1.5,2.5], m is set as 1 in the present invention；ckIndicate k-th of cluster centre, xjSample is inputted for jth
This；The flow of FCM clustering algorithms is first to initialize subordinated-degree matrix, then so that object function is converged to minimum by iteration
Value, finally determines the cluster result belonging to the upper hachure of K lines according to the result of iteration.
The lower hachure length and physical length of k lines are clustered using FCM clustering methods, the lower hachure for obtaining k lines is long
The method of cluster centre corresponding to degree and physical length clusters the upper hachure length of k lines with using FCM clustering methods
Method is similar, repeats no more.
(3) the on-line talking method based on density clusters the upper hachure length, lower hachure length and physical length of k lines
When being clustered by way of the on-line talking based on density, need first to find out current new input sample and sample
The sum of the distance inputted each of before in this space：
Then the density of k-th of data newly inputted is as follows：
But it will be very when being calculated by above method as the data inputted online are more and more
Difficulty, thus the present invention uses a kind of computational methods of recursive form：
πk(xk)=k (| | xk-uk||2+Xk-||uk||2) (9)
πk(xi)=πk-1(xi)+||xi-uk||2, i=1,2 ..., k-1 (10)
Wherein, ukFor to xkRecursive calculation result；xkFor k-th of sample data；xiFor i-th of sample data；πk(xk)
For xkDensity；πk(xi) it is xiDensity：XkIt is right | | xk||2Recursive calculation result；
Secondly, about πkIt is cumulative and be also to be calculated by this recursive method：
The density of the data of each new input can quickly and be expeditiously calculated by above method.
When its density is calculated in the sample data newly to arrive after, by checking whether that meeting the following conditions determines
Whether it can form a new cluster centre:
THEN(Nk←Nk+1) ⒁
In formula, Dk(xk) it is k-th of data x newly inputtedkDensity；It is ith cluster central point,It is i-th
A cluster centre pointDensity；NkFor the number of cluster centre point.
As soon as so belonging to this new cluster centre if it if a new center can be constituted, it and other are otherwise calculated
Cluster centre distance, the cluster centre of minimum range is as the center belonging to it.
Step 203：Deeply learns
By the solid color for clustering obtained result and K lines one will be generated by the sliding window that a size is 3
The combination of a K lines is input to as input state in deeply study (DRL) model：
st=(cu, cl, cb, cc) (16)
In formula, cuFor the cluster centre corresponding to upper hachure length；clFor the cluster centre corresponding to lower hachure length；cb
For the cluster centre corresponding to physical length；ccFor solid color.
The nitrification enhancement of one standard usually requires to include a process explored and utilized.Exploration can help mould
Type fully understands the state space of its local environment, and optimal action sequence is found using model is then helped.Model is receiving
The process explored and utilized by one after input state acts to choose, this initial in the present invention parameter is set as
0.7, and it can gradually increase with the increase of frequency of training, and its upper limit is at maximum up to 0.95.
When being converted to state by deep neural network, the activation primitive used is relu functions, because
The gradient of sigmoid and tanh is very gentle in zone of saturation, close to 0, it is easy to cause vanishing gradient's
Problem slows down convergence rate.And relu activation primitives are used, the calculation amount of whole process can save very much, in addition, relu can make
The output of a part of neuron is 0, thus causes the sparsity of network, and reduce the relation of interdependence of parameter,
Alleviate the generation of overfitting problem.The probability finally finally each acted with a softmax functions output, select probability is most
Big action.
Finally obtained action a is applied in environment, obtains the award r of environmental feedback, and (s, a, r) is stored in
In data base in DRL models, and current model parameter is just updated to adapt to new ring every 200 transaction
Border.
When being traded, DRL models need to interact with external world's trading environment, by input action to environment,
Obtain the award of environmental feedback.If when DRL is when t moment generates a transaction movement, one t+1 time point of environmental feedback
The true earning (the income less Commission that price difference is brought) obtained after closing a position gives DRL models, and t moment is moved to t+1
Moment.If t moment does not have transaction movement generation, just only t moment is moved to the t+1 moment.It is model shown in Fig. 3
General frame.
As long as cluster number and maximum iteration or serious forgiveness is determined i.e. when either FCM is clustered with Kmeans
It can.Based on the on-line talking of density when being initialized, need to carry out first data input into the initialization to structure.
In training deep neural network because weights initialisation it is reasonable whether when can influence trained to a certain extent
Can network restrain and the speed of training speed, thus parameter initialization is a particularly important step.
When carrying out weights initialisation to deep neural network, Xavier methods are used, it passes through where neural network
The input dimension n of layer determines the range of parameter distribution with output dimension m：
In addition, when using relu functions as activation primitive, initialized by Xavier methods, it is possible to prevente effectively from
Dead ReLU Problem (certain neurons may never be activated, and cause corresponding parameter that cannot be updated forever).
When building object function, there are the label that can be used for calculating error, thus the present invention different from monitoring learning
It is the how many probability to increase or reduction action generates of award obtained according to some action, therefore in the object function of construction
As shown in (18), wherein being used as an important influence factor by the size rt of award, it can guide gradient to train.
maxθL (θ)=∑ log π (at|st, θ) and rt (18)
In formula, atFor the transaction movement of t moment；stFor the state of t moment；θ is the weighting parameter in deep learning network；
rtFor the award of t moment environmental feedback；
When the parameter to neural network is updated, Adam optimization algorithms have been used.Adam is also that one kind is based on
Gradient declines the algorithm for carrying out optimization object function, but unlike that traditional stochastic gradient descent algorithm, stochastic gradient descent is calculated
Method keeps single learning rate to update all weights, and learning rate can't change in the training process.And Adam passes through
Calculate gradient first moment track and second order moments estimation and be the independent adaptive learning rate of different parameter designings.Thus
Adam is algorithm all the fashion in deep learning field, and the realization that it can be quickly is excellent as a result, some rules of thumb proves
Adam algorithms are had excellent performance in practice, have prodigious advantage relative to other kinds of Stochastic Optimization Algorithms.
In optimization, our purpose is exactly to maximize object function, to train best decision.If in a π
(at|stIn the case of very little, as soon as a prodigious award is obtained from environment, then increasing to its newer degree, with this
To increase the action probability made a profit in a certain state.
Below to it is proposed by the present invention based on decompose k lines cluster and deeply study adaptive financial trade method into
Row experimental verification：
1, Preparatory work of experiment
That is used when being trained and testing to model is derived from the truthful data in financial market, including commodity
Futures and stock price index futures.For commodity future, the present invention has selected screw-thread steel contract and dregs of beans contract；It is of the invention in stock price index futures
Hu-Shen 300 index is selected.Selectively data all allow to open mostly and open do-nothing operation for institute.Opening mostly is meaned when rise in price
It makes a profit, opening sky is made a profit when price drops.
In an experiment, what commodity future and stock price index futures were used is the time between 1 minute data, that is, two prices
Interval is one minute, and when data decimation, since the dregs of beans contract night disk time is short in commodity future and Hu-Shen 300 index contract does not have
Night disk, so the time cycle chosen is longer compared with screw-thread steel contract.
As can be seen from Figure 4 the price trend of these three contracts.A kind of RB and IF300 contracts all becoming with rise
Gesture, for the data of training part all in a lower price, the price of part of detecting data will be apparently higher than trained part.
Then whole fluctuation is bigger for M contracts, and the data price of training and part of detecting is relatively.These types of contract is in transaction
The profit and service charge of generation is all weighed with RMB, and the hand number of transaction is all arranged on the other hand.Here Profit/
Point means the income that one point of every variation can obtain.
2, model training
Firstly the need of training dataset and test data set is divided, preceding 10000 are used for commodity future, stock price index futures
Point is used as training data, and latter 24000 points for testing.When carrying out the initialization to system with training data, it is necessary first to K lines
Change processing training data, is then carried out again come the on-line talking module to Kmeans, FCM and based on density with obtained result
Cluster result, is then continuously inputted into DRL models by construction again, carries out parameter training.
It is tested by the model of last time training as final result.When being tested, for Kmeans,
FCM it is this kind of can not online updating clustering method, we directly by the data generated in environment be successively inputted to training after obtain
Clustering Model in, the cluster result of input data is then obtained by prediction, then result is input to trained DRL models
In.And the on-line talking based on density then need to only input the data from environment, then in demand pairs according to being clustered, and dynamic
Cluster result, is finally also input in DRL models by the parameter for updating dynamic clustering model.
Wherein, when being trained, since model is a non-convex system, so being easily trapped into local minimum
With there is the phenomenon that over-fitting.Other than using above-mentioned and carrying out initialization and Adam optimization algorithms to parameter,
Dropout is also applied in the training process of model, Dropout refer in the training process of deep learning network, for
Neural network unit temporarily abandons it according to certain probability from network, all when can prevent each training in this way
Neuron collective effect, zoom in or out always certain features, easilys lead to over-fitting in this way and reduces generalization ability,
It can effectively avoid occurring these problems in training process by Dropout.In addition, the training for terminating in advance model is also one
Kind is commonly used to the method for avoiding deep neural network over-fitting, thus only has trained 50epoch when being trained to model, so
Model can be allowed to adapt to new test number as early as possible in this way for testing using the model of last time training as training result afterwards
According to.
3, it is compared with the DRL models of other feature learnings
By the DRL moulds proposed by the present invention based on the deeply learning model and other feature learnings that decompose k lines cluster
Type is compared.The first total revenues (Profit) of different models, winning rate (PR), profit and loss ratio (PCR) and (LT) that opens more storehouses
With the number (ST) for opening hole capital after selling all securities.Wherein, the computational methods of winning rate are the total degree of the number divided by transaction got a profit in transaction, profit and loss
Than the ratio for being the average amount of average amount and loss got a profit in transaction, winning rate and profit and loss ratio are assessing a Trading Model
Or it is played an important role on trading strategies.
It is very big from the effect difference caused by different feature learnings that can visually see first in the yield curve of Fig. 5,
By K lineizations cluster DRL modelling effects to be substantially better than other several DRL models, in addition only carry out K lineizations processing without
Cluster can also obtain more good effect, because only carrying out the processing of K lineizations, obtained feature also has certain summary energy
The effect of power and denoising.Using the price of blurring as characteristic effect and bad, although because price is carried out Fuzzy processing
Partial noise is eliminated to a certain extent, but also increases state space simultaneously so that more difficult receipts during training
It holds back.It is worst that price is directly put into the effect obtained in DRL models as feature, because being difficult merely to predict from price
Trend later, in addition, containing many noises in price, it is difficult effectively to extract feature to be put into DRL models.Price is carried out
The result that K lines and cluster obtain not only can be as a kind of summary to price trend, but also can remove the part in data
Noise, the feature obtained in this way be input in DRL models can restrain and achieve the effect that as early as possible in the training process one it is good,
To obtain a preferable performance when test.
Secondly, from three kinds of contract tendencies, the data tendency fluctuation of M contracts is larger, but both may be used in futures exchange
Can open hole capital after selling all securities again to open more storehouses, thus no matter the chance how tendency of price is all in the black.In the survey of RB and IF300 contracts
It tries in data, the tendency of the two on the whole all has a kind of trend of rise, thus to the DRL's of K lineizations cluster from table 1
As can be seen that the number for opening hole capital after selling all securities will be more than by opening the number in more storehouses in transaction count statistics.In addition to this, to winning rate from table 1
Statistics in it can also be seen that winning rate is higher when upward price trend is stronger, and when fluctuation is larger, winning rate is relatively low, illustrates mould
Type is more suitble to trend sexual transaction.
In addition, Kmeans and FCM little from difference in the performance that can be seen that three in the DRL models that these three are clustered
Cluster number be fixed, but shown by experiment (E, Comparing With Other Hyperparameters)
Mostly very few cluster number can all influence the final effect of model, although OnlineMeans can be moved during cluster
The increase cluster centre of state, but finally can also maintain in a fixed cluster number, so from the angle of cluster,
The effect of three does not have much difference.On the other hand, Kmeans and FCM be during training from macroscopically to data into
What row divided, so data division is compared and stresses entirety, and OnlineMeans is then true by coming in line computation packing density
This ownership of random sample still constitutes a new center with current center, so more stressing part when dividing data.Except this
Except, OnlineMeans is calculated by the form of iteration when calculating sample point, so in practical applications
In view of the high frequency of delay and the transaction of network, result can be calculated faster using OnlineMeans.
Finally, in the performance of the DRL models learnt by these types of different characteristic as can be seen that no matter the fluctuation in market compared with
Chance that is big or being relatively steadily all in the black.In addition, the performance of DRL models will be more than that other are several by way of K line clusters
Kind model, while also illustrating a problem, although DRL models can have good performance, preceding summary in marketing
It selects suitable feature to be learnt, is otherwise difficult effectively to learn from market.
The performance of the different Trading Models of table 1
4, it is compared with several frequently seen neural network model
It is tested using 500 indexes of S&P, when realizing contrast model, using the tensorflow frames of Google exploitations
Frame, tensorflow provide much modules for realizing neural network model, can easily be realized by these modules
The neural networks such as RNN, LSTM, FNN.According to the structure of fuzzy inference system, can correspondingly by FNN points for Mamdani types and
TSK types show that the identification precision of TSK type fuzzy neural networks is higher than Mamdani type fuzzy neurals in view of many literature research
Network, thus TSK type fuzzy neural networks have been used when comparison of design is tested, the former piece in TSK structure of fuzzy neural network
Network is used for fuzzy reasoning, and consequent network carries out parameter learning.The structure of LSTM and RNN is substantially similar, only makes in recurrence layer
Unit is respectively LSTM cell and RNN cell.The frequency of training of neural network model is both configured to 50, and time window is big
It is small to be set as 20, the ups and downs of next day price are predicted by the price of input.Since sky cannot be done in stock market,
ClusterDRL models only remain the action done and mostly and held position in training.
As can be seen that all predictablity rates are all just over 50% from the result of table 2, this is the result shows that prediction
The difficulty of the general trend of market development.Nevertheless, still the accuracy rate of ClusterDRL models is still higher than the prediction mould of neural network
Type, neural network model are based on historical data in prediction, but complicated market is seldom to History repeats itself, and price
In contain a large amount of noise, also be difficult to therefrom learn without denoising.And ClusterDRL models then pass through K lineizations and cluster pair
Data carry out denoising, can not only consider in prediction the data of history but will stress with current tendency, thus can be with from table 2
Find out, the number of ClusterDRL to open a position will be less than neural network model, illustrate that ClusterDRL models are carrying out decision
When consider it is more thorough, so effect wants better.
The comparison of table 2 and neural network model
5, the hyper parameter used in model is inquired into
The hyper parameter used in model is inquired into, including the cluster number in Kmeans, FCM, time window and
It closes a position the time.Test the RB contract datas that the data used are 2017 March to 2017 Augusts.Clustering the range that number is inquired into is
From 3 to 6, time window range be from 1 to 5, the time of closing a position be from t+1 to t+3, experimental result respectively Table3,
In Table4, Table5.
From the results shown in Table 3, if when the number of cluster is very few or excessively can all reduce to a certain extent
The effect of model, this is because the number of cluster influences whether the size of state space, if state space is too small to lead to mould
Type is difficult effectively from extracting data feature, and state space is excessive, can influence the convergence of model to a certain extent.Thus,
, when building Model tying, the number acquiescence of cluster uses 5 to behave oneself best in table for we.
When probing into the performance of different time window length, from table 4 it can be found that with time window length increasing
Add, obtained result is gradually deteriorated, and reason is also the increase with state space, and the convergence of model is also more difficult to, but works as the time
If window setting is too small, as state space too small and impact effect.
The Contrast on effect of the different cluster numbers of table 3
The Contrast on effect of 4 different time window length of table
In the financial market of high noisy, high unpredictability, the present invention is proposed from the actual conditions in market
With the deeply learning model that price K lineizations will be handled and clustered, rule can be obtained from history by being handled with K lineizations
Current trend and the preliminary denoising to data are restrained and summarize, then result is carried out cluster to increase the characterization ability of data
And data are further subjected to denoising, input feature vector of the result obtained to this as deeply learning model can be right
Real time data preferably carries out the adaptive optimization of the update and parameter to decision.The present invention is by model in commodity future, stock
Refer to the performance in futures and stock market, demonstrating can with the deeply study that price K lineizations are handled and clustered
To obtain better effect.
Above-mentioned, although the foregoing specific embodiments of the present invention is described with reference to the accompanying drawings, not protects model to the present invention
The limitation enclosed, those skilled in the art should understand that, based on the technical solutions of the present invention, those skilled in the art are not
Need to make the creative labor the various modifications or changes that can be made still within protection scope of the present invention.
Claims (10)
1. a kind of based on the adaptive Financial Time Series Forecasting method for decomposing k lines cluster and deeply study, characterized in that
Include the following steps：
Step 1：Financial time series are obtained, the processing of K lineizations are carried out to financial time series, and to K lineizations treated data
It is calculated, obtains the length of each subdivision of K lines in the current matching period；
Step 2：It is each to K lines using Kmeans clustering algorithms, FCM clustering algorithms or on-line talking method based on packing density
The length of subdivision is clustered, and the cluster centre corresponding to the length of each subdivision of K lines is obtained；
Step 3：The solid color of cluster result and K lines is input to as learning characteristic in deeply learning model and is carried out
Parameter training carries out financial transaction using trained deeply learning model.
2. according to claim 1 pre- based on the adaptive financial time series for decomposing k lines cluster and deeply study
Survey method, characterized in that the K lines include the color of upper hachure, lower hachure, entity and entity.
3. according to claim 1 pre- based on the adaptive financial time series for decomposing k lines cluster and deeply study
Survey method, characterized in that in the step 1, the price of period is corresponded to according to K lines, the K lines in the current matching period are calculated
The step of length of each subdivision includes：
The K lines in the current matching period are calculated in highest price, opening price, lowest price and the closing price that the period is corresponded to according to K lines
Upper hachure, lower hachure and entity length；
When the solid color of K lines is red, the upper hachure length of K lines is equal to the difference of highest price and closing price, the lower shadow of K lines
Line length is equal to the difference of opening price and lowest price, and the physical length of K lines is equal to the difference of closing price and opening price；
When the solid color of K lines is green, the upper hachure length of K lines is equal to the difference of highest price and opening price, the lower shadow of K lines
Line length is equal to the difference of closing price and lowest price, and the physical length of K lines is equal to the difference of opening price and closing price.
4. according to claim 1 pre- based on the adaptive financial time series for decomposing k lines cluster and deeply study
Survey method, characterized in that in the step 2, exist using Kmeans clustering algorithms, FCM clustering algorithms or based on packing density
Line clustering method respectively clusters the upper hachure length of K lines, lower hachure length and physical length, obtains the upper shadow of K lines
Cluster centre corresponding to line length, lower hachure length and physical length.
5. according to claim 1 pre- based on the adaptive financial time series for decomposing k lines cluster and deeply study
Survey method, characterized in that it is described using Kmeans clustering algorithms to the upper hachure length, lower hachure length or physical length of K lines
The step of being clustered include：
Using the upper hachure length of the K lines in the current matching period, lower hachure length or physical length as input sample, structure instruction
Practice sample set K；
Training sample set K is divided into 5 classes, randomly selects in each class input sample as in such initial clustering
The heart；
Calculate in each class the square distance of input sample and initial cluster center and, obtain Kmeans object functions, it is right
Kmeans object functions carry out seeking local derviation, calculate the minimum value of the derivative, obtain the cluster centre of each class.
6. according to claim 1 pre- based on the adaptive financial time series for decomposing k lines cluster and deeply study
Survey method, characterized in that described that the upper hachure length, lower hachure length or physical length of K lines are carried out using FCM clustering algorithms
The step of cluster includes：
Using the color of the upper hachure of the K line number evidences in the current matching period, lower hachure, the length of entity or entity as input sample
This, structure training sample set K；
Training sample set K is divided into 5 classes, randomly selects in each class input sample as in such initial clustering
The heart calculates degree of membership of each input sample to all classes, is judged in the cluster belonging to input sample by the size of degree of membership
The heart；
Calculate in each class input sample and initial cluster center square distance and with degree of membership product, obtain FCM object functions,
FCM object functions are carried out to seek local derviation, the minimum value of the derivative is calculated, obtains the cluster centre of each class.
7. according to claim 1 pre- based on the adaptive financial time series for decomposing k lines cluster and deeply study
Survey method, characterized in that the on-line talking method based on packing density is long to upper hachure length, lower hachure length or entity
Spending the step of being clustered includes：
Using the color of the upper hachure of the K line number evidences in the current matching period, lower hachure, the length of entity or entity as input sample
Notebook data；
Calculate the density of each new input sample；
When initialization, current first sample point is chosen to carry out parameter initialization to cluster structure and as in first cluster
The heart；
Judge each new density of input sample with the maximum value of all cluster centre density or with all cluster centre density
The size of minimum value, if the density of the input sample is more than the maximum value of all cluster centre density or is less than all cluster centres
The input sample is then new cluster centre by the minimum value of density；Otherwise, it calculates in the input sample and other all clusters
The distance of the heart, using the cluster centre of minimum range as the cluster centre belonging to the input sample.
8. according to claim 1 pre- based on the adaptive financial time series for decomposing k lines cluster and deeply study
Survey method, characterized in that described to be input to deeply study using the solid color of cluster result and K lines as learning characteristic
The step of progress parameter training, includes in model：
According to the sliding time window of setting, upper hachure length, lower hachure length, the physical length pair of the K lines that cluster is obtained
The solid color of the cluster centre and K lines answered is input to deep learning network model as learning characteristic and carries out feature learning；
Transaction movement is obtained by nitrification enhancement, and is applied in environment, passes through deep neural network using activation primitive
State is converted, obtains the award of environmental feedback, and store into the data base of deeply learning model, and every one
The transaction for determining number is updated the parameter of deeply learning model；
The weights of deep learning network are initialized using Xavier initial methods；
Object function is constructed, the weights of deep learning network are updated；
Pass through the continuous interaction of deeply learning model and environment, study to the deep learning network mould with high trade decision
Type.
9. according to claim 7 pre- based on the adaptive financial time series for decomposing k lines cluster and deeply study
Survey method, characterized in that the structure object function, the step of being updated to the weights of deep learning network include：
Award based on environmental feedback, constructs object function, and object function is：
maxθL (θ)=∑ log π (at|st, θ) and rt
In formula, atFor the transaction movement of t moment；stFor the state of t moment；θ is the weighting parameter in deep learning network；rtFor t
The award of moment environmental feedback；
Object function is optimized using Adam optimization algorithms, deep learning network is updated by the object function after optimization
Weights.
10. a kind of computer installation, be used for adaptive Financial Time Series Forecasting, characterized in that including memory, processor and
Store the computer program that can be run on a memory and on a processor, which is characterized in that the processor executes the journey
Following steps are realized when sequence, including：
Financial time series are obtained, the processing of K lineizations are carried out to financial time series, and treated that data are counted to K lineizations
It calculates, obtains the length of each subdivision of K lines in the current matching period；
Using Kmeans clustering algorithms, FCM clustering algorithms or on-line talking method based on packing density to each subdivision of K lines
Length clustered, obtain the cluster centre corresponding to the length of each subdivision of K lines；
It is input to progress parameter instruction in deeply learning model using the solid color of cluster result and K lines as learning characteristic
Practice, financial transaction is carried out using trained deeply learning model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810501626.9A CN108734220A (en) | 2018-05-23 | 2018-05-23 | Adaptive Financial Time Series Forecasting method based on k lines cluster and intensified learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810501626.9A CN108734220A (en) | 2018-05-23 | 2018-05-23 | Adaptive Financial Time Series Forecasting method based on k lines cluster and intensified learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108734220A true CN108734220A (en) | 2018-11-02 |
Family
ID=63936290
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810501626.9A Pending CN108734220A (en) | 2018-05-23 | 2018-05-23 | Adaptive Financial Time Series Forecasting method based on k lines cluster and intensified learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108734220A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109766745A (en) * | 2018-11-22 | 2019-05-17 | 四川大学 | The combination of intensified learning tri-state long Memory Neural Networks system and training and prediction technique in short-term |
| CN110133643A (en) * | 2019-05-22 | 2019-08-16 | 北京林业大学 | Root system of plant detection method and device |
| CN110263843A (en) * | 2019-06-18 | 2019-09-20 | 苏州梧桐汇智软件科技有限责任公司 | Stock K line recognition methods based on deep neural network |
| CN111369042A (en) * | 2020-02-27 | 2020-07-03 | 山东大学 | Wireless service flow prediction method based on weighted federal learning |
| CN112036754A (en) * | 2020-06-30 | 2020-12-04 | 南昌木本医疗科技有限公司 | Financial big data wind control platform based on block chain |
| CN113376562A (en) * | 2021-06-08 | 2021-09-10 | 国网重庆市电力公司营销服务中心 | CVT (continuously variable transmission) verification method, device and medium based on rolling time window-FCM (fuzzy c-means) clustering |
| CN113411125A (en) * | 2021-06-09 | 2021-09-17 | 中国人民解放军陆军工程大学 | Signal judgment method and communication system for reverse modulation wireless optical communication |
| CN114154582A (en) * | 2021-12-06 | 2022-03-08 | 天津大学 | Deep reinforcement learning method based on environment dynamic decomposition model |
-
2018
- 2018-05-23 CN CN201810501626.9A patent/CN108734220A/en active Pending
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109766745B (en) * | 2018-11-22 | 2022-12-13 | 四川大学 | Reinforced learning tri-state combined long-time and short-time memory neural network system and training and predicting method |
| CN109766745A (en) * | 2018-11-22 | 2019-05-17 | 四川大学 | The combination of intensified learning tri-state long Memory Neural Networks system and training and prediction technique in short-term |
| CN110133643B (en) * | 2019-05-22 | 2021-08-20 | 北京林业大学 | Plant root system detection method and device |
| CN110133643A (en) * | 2019-05-22 | 2019-08-16 | 北京林业大学 | Root system of plant detection method and device |
| CN110263843A (en) * | 2019-06-18 | 2019-09-20 | 苏州梧桐汇智软件科技有限责任公司 | Stock K line recognition methods based on deep neural network |
| WO2021169577A1 (en) * | 2020-02-27 | 2021-09-02 | 山东大学 | Wireless service traffic prediction method based on weighted federated learning |
| CN111369042B (en) * | 2020-02-27 | 2021-09-24 | 山东大学 | Wireless service flow prediction method based on weighted federal learning |
| CN111369042A (en) * | 2020-02-27 | 2020-07-03 | 山东大学 | Wireless service flow prediction method based on weighted federal learning |
| CN112036754A (en) * | 2020-06-30 | 2020-12-04 | 南昌木本医疗科技有限公司 | Financial big data wind control platform based on block chain |
| CN112036754B (en) * | 2020-06-30 | 2022-09-16 | 上海即科智能技术集团有限公司 | Financial big data wind control system based on block chain |
| CN113376562A (en) * | 2021-06-08 | 2021-09-10 | 国网重庆市电力公司营销服务中心 | CVT (continuously variable transmission) verification method, device and medium based on rolling time window-FCM (fuzzy c-means) clustering |
| CN113411125A (en) * | 2021-06-09 | 2021-09-17 | 中国人民解放军陆军工程大学 | Signal judgment method and communication system for reverse modulation wireless optical communication |
| CN114154582A (en) * | 2021-12-06 | 2022-03-08 | 天津大学 | Deep reinforcement learning method based on environment dynamic decomposition model |
| CN114154582B (en) * | 2021-12-06 | 2024-04-19 | 天津大学 | Deep reinforcement learning method based on environment dynamic decomposition model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108734220A (en) | Adaptive Financial Time Series Forecasting method based on k lines cluster and intensified learning | |
| Li et al. | Risk prediction in financial management of listed companies based on optimized BP neural network under digital economy | |
| Wu et al. | Adaptive stock trading strategies with deep reinforcement learning methods | |
| Ha et al. | Response models based on bagging neural networks | |
| CN110751318B (en) | Ultra-short-term power load prediction method based on IPSO-LSTM | |
| CN109118013A (en) | A kind of management data prediction technique, readable storage medium storing program for executing and forecasting system neural network based | |
| CN110084610A (en) | A kind of network trading fraud detection system based on twin neural network | |
| CN109191276B (en) | P2P network lending institution risk assessment method based on reinforcement learning | |
| CN110378382A (en) | Novel quantization transaction system and its implementation based on deeply study | |
| Zhu et al. | Coke price prediction approach based on dense GRU and opposition-based learning salp swarm algorithm | |
| CN109242558A (en) | A kind of management data prediction technique, readable storage medium storing program for executing and forecasting system neural network based | |
| Mu et al. | Transformative computing for products sales forecast based on SCIM | |
| Abudureheman et al. | Performance evaluation of enterprises’ innovation capacity based on fuzzy system model and convolutional neural network | |
| Yinka-Banjo et al. | Stock Market Prediction Using a Hybrid of Deep Learning Models | |
| Iraji et al. | Students classification with adaptive neuro fuzzy | |
| Sari et al. | The effectiveness of hybrid backpropagation Neural Network model and TSK Fuzzy Inference System for inflation forecasting | |
| CN110322351A (en) | Multi-source driving quantization investment model under Depth Stratification strategy | |
| Mohamed et al. | Adaptive trainer for multi-layer perceptron using artificial gorilla troops optimizer algorithm | |
| CN113256425A (en) | Data enhancement deep learning method for stock index prediction | |
| Zhuang et al. | Research on Quantitative Stock Selection Strategy Based on CNN-LSTM | |
| Ushio et al. | The application of deep learning to predict corporate growth | |
| Nishizaki et al. | Simulation analysis using multi-agent systems for social norms | |
| Xue | Prediction of NASDAQ Index Based on Stacking Method | |
| Nur | Comparing the Accuracy of Multiple Discriminant Analyisis, Logistic Regression, and Neural Network to estimate pay and not to pay Dividend | |
| CN112288078B (en) | Self-learning, small sample learning and migration learning method and system based on impulse neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181102 |