CN108241872A - The adaptive Prediction of Stock Index method of Hidden Markov Model based on the multiple features factor - Google Patents
The adaptive Prediction of Stock Index method of Hidden Markov Model based on the multiple features factor Download PDFInfo
- Publication number
- CN108241872A CN108241872A CN201711488697.1A CN201711488697A CN108241872A CN 108241872 A CN108241872 A CN 108241872A CN 201711488697 A CN201711488697 A CN 201711488697A CN 108241872 A CN108241872 A CN 108241872A
- Authority
- CN
- China
- Prior art keywords
- hmm
- model
- prediction
- hidden
- day
- 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
- G06F18/29—Graphical models, e.g. Bayesian networks
- G06F18/295—Markov models or related models, e.g. semi-Markov models; Markov random fields; Networks embedding Markov models
-
- 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
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- 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
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Business, Economics & Management (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- Development Economics (AREA)
- Economics (AREA)
- Marketing (AREA)
- Strategic Management (AREA)
- Technology Law (AREA)
- General Business, Economics & Management (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Financial Or Insurance-Related Operations Such As Payment And Settlement (AREA)
Abstract
The invention discloses the adaptive Prediction of Stock Index method of the Hidden Markov Model based on the multiple features factor, the present invention is mainly made of stock sample data and HMM.The thinking of Prediction of Stock Index based on HMM is the mark by hidden state by the time point of historical data, realize the classification at time point, then by finding the history point consistent with the classification annotation of the previous day day to be predicted, calculating obtains history point and the amount of increase and amount of decrease of their latter days, to estimate the closing price residual error of day to be predicted and proxima luce (prox. luc).Various features attribute has been used, the indexs such as the market value capital, technology, momentum of stock market has been contained as primary election feature, filters out the stronger attribute of predictive ability as feature vector by a variety of methods, predictive ability is compared with better than the method for using feature few.
Description
Technical field
The invention belongs to Prediction of Stock Index technical fields, and the present invention relates to a kind of stock index trend classification prediction sides based on HMM
Method.
Background technology
In the last century 60's, Bao Mu proposes Hidden Markov Model, then by field of speech recognition
Using known to numerous scientific research personnel and fan, Lee opens multiple doctoral thesis within 1988, the speech recognition software quilt based on HMM
It is chosen as important invention.The Renaissance scientific & technical corporation is internationally famous investment institution, because the year that superelevation was maintained at from 1989 is returned
Rate is described as most efficient hedge fund by industry.It is exactly based on the unique mathematical model capture market opportunity and carries out quantization throwing.And
Hidden Markov Model is one of main tool of the said firm.Therefore, HMM is attempted in financial market and carries out prediction with practical
Meaning.
Stock market is divided into generally according to the mode difference of investment:Basic side, poly-talented and quantization type.According to statistics, entirely
The capitalized method of the ball overwhelming majority is all basic side type, this method is based on capital investment.A part of personnel be it is poly-talented based on,
This mode is highly professional, is invested by analyzing various technical indicators, and calculating is cumbersome, based on civil individual.In recent years, in state
The investment way of inside and outside quantization type gradually increases.Using HMM quantitative models in stock market, first have to clear two of understanding and ask
Topic:First, to the problem concerning study of the parameter of HMM, second, the state classification based on observed value.Model parameter is to pass through observed value
What study came, but the model that can reflect shares changing tendency can not be directly established by single shares worth, so as to probe into observation
It is worth the hidden state of behind so that investment research uses.Some important characterization factors, such as shares worth, conclusion of the business can but be extracted
The observations of the secondary informations data as model such as amount, HMM quantitative models are established by multidimensional characteristic factor sequence
Research at present in industry is more to do some statistical analysis class models in financial direction, such as:It is pendulum theoretical model, small
Market value theory, Alpha models, classification, Reasons, alpaca and green turtle strategy etc., these traditional methods have accomplished the day in field
Card can not accomplish breakthrough bottleneck, the method that can utilize machine learning, in the limitation of not excessive financial frame
Under, more the inherent organic growth rule from the essence of data and market is probed into unexpectedly, has better effect, each at present
Kind algorithm in industry did preliminary research, but preferably model still needs do optimize, improve more, to more preferable
And practical problem combine.In the artificial intelligence epoch, artificial intelligence can be more closely applied in Prediction of Stock Index, found a kind of
Efficient Prediction of Stock Index algorithm, has important novelty, creativeness and practical value.
Invention content
The thinking of Prediction of Stock Index based on HMM is the mark by hidden state by the time point of historical data, realizes the time
The classification of point, then by finding the history point consistent with classification annotation of the previous day day to be predicted, calculating obtain history point with
The amount of increase and amount of decrease of their latter days, to estimate the closing price residual error of day to be predicted and proxima luce (prox. luc).
The present invention is mainly made of stock sample data and HMM.
The adaptive Prediction of Stock Index method of Hidden Markov Model based on the multiple features factor, this method include three steps
Suddenly,
First, HMM model is trained by stock sample data, is established based on HMM model, by the applicable item for limiting model
Part, with the prediction effect being optimal;Limit the applicable elements of model as the length of training sample, the prediction window of observation,
The hidden status number of HMM, the window for avoiding being applicable in Future Data, and optimization processing is carried out to these data parameters.
Second, after trained HMM model, the restriction parameter optimized, totally 53 characterization factors are fitted
With, and price expectation respectively is carried out to observation to be predicted, it is arranged according to the height of the accuracy rate of prediction, chooses high separating capacity
The factor, consider the correlation between characterization factor, extract incoherent character factor, while it is preferable secondary to consider that PCA chooses
Predicted characteristics of the characterization factor as model, as the stock index prediction aspect of model based on HMM finally stablized.
Third obtains convergent HMM parameters by model training, and value sequence mark to be observed is calculated using obtained model
The possibility of each hidden state is denoted as, and labeled as most probable hidden status switch and corresponding likelihood value.Then HMM moulds are utilized
Type calculates the most probable hidden status switch of the sequence of observations to be predicted and likelihood value, and by time and likelihood function value
As the searching foundation of most close history point, time-based weights distribution uses:wm=exp (1/ (i-m+1)), wmFor difference
Weights shared by close history point and latter heave amplitude in the daytime, i are the currently proxima luce (prox. luc) serial number with pre- observation, and m is gone through to be close
The serial number of Shi Tian;So as to obtain the fluctuation limit of training sample day latter day, by this amplitude calculate day to be predicted and
The amount of increase and amount of decrease of proxima luce (prox. luc) further obtains the stock price trend classification of day to be predicted.
Hidden Markov model (HiddenMarkov Model, HMM) is statistical model, is to probe into a markov
The model established in the relational process of process and behind hidden state, i.e., it be used for describing a horse containing implicit unknown parameter
Markov process.It is commonly used to determine the implicit parameter of the sequence from observable sequence, then further utilizes these ginsengs
It counts to research and analyse.For HMM according to background difference is used to be divided into discrete type and continuous type, typical discrete type is hidden state and observation
Worth probability is to correspond, and the probability of the hidden state of continuity HMM and observation is obtained by the probability distribution of hidden state
It arrives.One Hidden Markov Model is represented by a triple:(π, A, B), complete representation are:(N, M, π, A, B),
In:
N:Hidden status number;
M:The corresponding observation number of one hidden state;
Pi:Initial probability distribution π=P { q of hidden state1=Sj, sum (π)=1;
A:Transfering probability distribution between hidden state.Aij=P { qt+1=Sj|qt=Si},1≤i,j≤N.sum(Ai)=1;
B:Certain hidden state corresponds to the probability distribution of observation.Bij=P { ot|qt=Si},1≤i≤M,1≤j≤N;
Discrete type observation probability is distributed:bjk=bj(ot)=P (ot=k | st=j), 1≤k≤U;
Continuous type observation probability is distributed:bj(ot)=∑ wjk·bjk(ot), j=1...N, 1≤k≤M;
The probability distribution of the general observation of continuity is fitted using mixed Gaussian approximation to function, ∑ wjk=1, j=1...N, k
=1...M, if observation sequence is multidimensional,:bjk=bj(ot)=N (O, Ujk,∑jk);
Wherein:Ujk=E [O (t) | Q (t)=j, M (t)=k];∑jk=Cov [O (t) | Q (t)=j, M (t)=k]
When finding history point, it is related to the history section to be found, differentiates the method for similar historical point, the mark of history point
During note, need to solve the computational methods of classification number, when selecting feature, use the stock index of multidimensional.
Data set length and prediction window section are adjusted to, while ensure close history point by optimal value by training study
Accurate and characteristic polymorphic, and the self study of adoption status number realizes the method for automatically updating model.
In certain history point section, the method for similar historical point is differentiated, during the mark of history point, the calculating side of classification number
Method realizes the prediction of stock index on the basis of HMM.The present invention also has certain actual application value:First, as engineering
Practise the practical application platform aided education of algorithm;Second is that in the field of current machine study, Prediction of Stock Index is realized, thereby using
HMM handles voice signal;Third, due to the use of hardware cost resource it is less and calculate it is simple and fast, it is easy to accomplish.
There are three typical problems in HMM:
1st, probability problem is calculated:Under the premise of having model parameter, it is λ to calculate the probability of given observation sequence first to have parameter
Hidden Markov Model and an observation sequence collection.If it is desired to calculating in current parameter, this observation is obtained
The probability of sequence can calculate the probability of observation sequence by dynamic computational algorithm forwards algorithms.
2nd, decoding problem:Most probable hidden sequence is found by observation sequence
Decoding problem and probability calculation problem are much like, and different places are, calculating probability problem is asked all hidden
In the case of sequence, the summation of the probability of the observation sequence showed, and decoding problem, it is to solve in which hidden sequence premise
Under, the probability for obtaining this observation sequence is the largest, be not ask it is various and, using Viterbi algorithm, sequence obtains according to the observation
To optimal hidden status switch.
3rd, problem concerning study:Most possible HMM parameters are established by observation sequence collection
In practical applications, HMM can be there are one learning parameter problem in advance, for given observation sequence O, directly
Optimal HMM parameter lambdas are asked for, obtain that P (O | λ) value is maximum, but can be by asking the solution of its local optimum, this algorithm
Referred to as forward-backward algorithm algorithm, also referred to as Baum-Welch are the approximation methods of EM algorithms.
Compared with prior art, the present invention has the advantages that:
1. when applicable Hidden Markov Model does Prediction of Stock Index, it is determined that the hidden status number N of an important parameter sequence
Offering question so that the setting of subparameter is dependent on the property of data in itself, more accurately.
2. when similar historical sample is differentiated, with reference to 2 points:Mark status categories and the likelihood function of classification, phase
More only consider classification, it is too wide in range, it is inaccurate to obtain approximate sample.
3. when determining the weights of amount of increase and amount of decrease of similar sample, dependent on the distance between mark classification likelihood value, simultaneously
It pays close attention to, predicts the time gap of day and similar sample in the daytime.
4. having used various features attribute, the indexs such as the market value capital, technology, momentum of stock market are contained as primary election spy
Sign filters out the stronger attribute of predictive ability as feature vector by a variety of methods, and predictive ability is few compared with better than using feature
Method.
Description of the drawings
Fig. 1 is the process of HMM training.
Fig. 2 is training observation value labeled as each hidden shape probability of state comparison.
Fig. 3 is factor prediction comparison.
Fig. 4 correlations between the factor.
Fig. 5 is the various forms of Comparative results of sample external model.
Fig. 6 is the sample of partial data sample.
Fig. 7 is sample length and accuracy rate.
Fig. 8 is prediction window and accuracy rate.
Fig. 9 is the code for finding history phase near point.
Specific embodiment
Below in conjunction with the accompanying drawings and specific embodiment the invention will be further described.
The present invention is mainly made of stock certificate data and HMM.Stock certificate data is Shanghai and Shenzhen index from 2007.1.4-2017.4.10
Day of trade data, algorithm assumes that the HMM of the state of single order Markov property and current point in time, observation independence.
S1 data preparation:
Step 1:By finance and economics net, Scapy, index calculation formula gather data, as shown in Figure 6.
Step 2:These gather datas are normalized regular wait to pre-process.
Step 3:Training dataset and test data set are divided into the collection stock certificate data of pretreatment.
S2 builds model parameter:
Step 1:By the training dataset of the step 3 in S1 in the factor, with the hmmlearn.hmm of python,
Learn the inner parameter of hmm algorithms, this algorithm is the core algorithm of this model:
(1) it by 55 is that index is ultimately determined to 12 as (S3 feature selectings) that input feature value X, which is,.
(2) category label of the output for each sample point, range be in hidden state number N the step 4 of S2 (determine) by
Decoding algorithm Viterbi marks to obtain.
(3) the transition probability matrix A=[a of hmmij] (1≤i, j≤N), define certain day of trade i and rear adjacent day of trade j
Between hidden state transfer probability, 1 rank Markov property principle is utilized.
(4) hmm observes probability distribution B=Bj(k) (1≤j≤N), this matrix define the hidden state quilt of some day of trade
During labeled as j, the feature vector observed is for the general of k-vector.
(5) involved in S2 steps 1 to hmm algorithms in parameter, be by more times of learning algorithm baum-welch
Iterative learning, convergence obtain when obtaining optimal.
Step 2:It determines in total data set first, the length of the every batch of training sample set of division takes under various values
Must be optimal, as shown in Figure 7.
Step 3:After being determined in historical data section, the length of prediction window is adjusted, as shown in Figure 8.
Step 4:The determining of number of the hidden state of the HMM of this operation has program self study to obtain, and is to pass through
OEHS criterion, operation ncomponents.py obtain most stable of hidden state value, the hidden state number N as hmm.
S3 feature selectings:
Step 1:After the model obtained by S2 parts is applicable in optimized parameter, while utilize the characteristic procedure side of filtering type
Method obtains the sequence of each factor predictive ability, as shown in Figure 3.
Step 2:Simultaneously using PCA and pearson coefficients, the preferably incoherent factor of prediction result, such as Fig. 4 are filtered out
It is shown.
Step 3:Obtain input feature vector of 12 dimensions as hmm algorithms in 55 ATTRIBUTE INDEXs by the step 1 of S3,2, i.e., it is special
Sign vector.
S4 models are applicable in:
Step 1:By S2, S3 parts build the optimal of the suitable parameters arrived involved in model and the parameter of hmm algorithms
Value:Single batch training data sample length, forecast interval length, the similar historical points found, the number N of hidden state, 12 dimensions are defeated
Enter feature vector etc..Behind the complete basis of model construction, historical time point and day to be measured are calculated into likelihood value and label minute
Class, as shown in Figure 2.
Step 2:By the distance away from day to be predicted and likelihood function apart from program, most similar history point is filtered out,
Code by weighted average as shown in figure 9, obtain amount of increase and amount of decrease, so as to obtain the tendency classification of day to be measured.
Step 3:By being run on test set, obtaining the models of different situations, the results are shown in Figure 5.
By proposition in financial stock market dynamic modelling method, what is formed using stock market's technical indicator as characterization factor
In data, by different combination of eigenvectors, and by hidden status number mobilism, adjusting training sample length predicts window
Mouthful, most similar historical data point is found, has larger improvement result to the effect of model, specific various different prediction models are to sequence
The prediction effect of row trend is as shown in Figure 5.
By comparative analysis, the prediction effect obtained when 12 denapon are as input feature vector is preferable, highest prediction effect
It is 62.5%, relatively simple feature prediction winning rate improves 3.75%, compares static method as a result, accuracy improves 11.5%,
Effect raising is more notable, this also turns out the superiority of method, and model should be significantly simple with the in due course update of the variation of data
Change time and the accuracy of conventional method, and then solve the problems, such as Speaker Identification from the processing method of neural network.Cause
This can consider that the present invention has very high application value.
Finally it should be noted that:Above example is only to illustrate the present invention and not limits technology described in the invention
Scheme;Therefore, although this specification with reference to above-mentioned each embodiment to present invention has been detailed description, this
Field it is to be appreciated by one skilled in the art that still can modify to the present invention or equivalent replacement;And all do not depart from this
The technical solution of the spirit and scope of invention and its improvement, are intended to be within the scope of the claims of the invention.
Claims (3)
1. the adaptive Prediction of Stock Index method of the Hidden Markov Model based on the multiple features factor, it is characterised in that:This method packet
Three steps are included,
First, HMM model is trained by stock sample data, establishes and is based on HMM model, by limiting the applicable elements of model,
With the prediction effect being optimal;The applicable elements of model are limited as the length of training sample, the prediction window of observation, HMM
Hidden status number, avoid being applicable in the window of Future Data, and optimization processing is carried out to these data parameters;
Second, after trained HMM model, the restriction parameter optimized, totally 53 characterization factors are applicable in, and
Respectively to observation to be predicted carry out price expectation, according to prediction accuracy rate height arrange, choose high separating capacity because
Son considers the correlation between characterization factor, extracts incoherent character factor, while considers that PCA chooses preferable quadratic character
Predicted characteristics of the factor as model, as the stock index prediction aspect of model based on HMM finally stablized;
Third obtains convergent HMM parameters by model training, and calculating value sequence to be observed using obtained model is labeled as
The possibility of each hidden state, and labeled as most probable hidden status switch and corresponding likelihood value;Then HMM model, meter are utilized
The most probable hidden status switch of the sequence of observations to be predicted and likelihood value, and by being used as most in time and likelihood function value
The searching foundation of close history point, time-based weights distribution use:wm=exp (1/ (i-m+1)), wmIt close is gone through to be different
Weights shared by history point and latter heave amplitude in the daytime, i are the currently proxima luce (prox. luc) serial number with pre- observation, and m is close history day
Serial number;L functions are equipped with weights using Euclidean distance formula, so as to obtain the fluctuation limit of training sample day latter day, lead to
The amount of increase and amount of decrease that this amplitude calculates with prediction day and proxima luce (prox. luc) is crossed, further obtains the stock price trend classification of day to be predicted.
2. the adaptive Prediction of Stock Index method of the Hidden Markov Model according to claim 1 based on the multiple features factor,
It is characterized in that:Hidden Markov model is statistical model, is to probe into a Markov process and behind hidden state
The model established in relational process, i.e., it be used for describing a Markov process containing implicit unknown parameter;HMM is according to making
Be divided into discrete type and continuous type with background difference, typical discrete type be hidden state and observation to obtain probability be to correspond, and
The hidden state of continuity HMM and the probability of observation are obtained by the probability distribution of hidden state;One Hidden Markov mould
Type is represented by a triple:(π, A, B), complete representation are:(N, M, π, A, B), wherein:
N:Hidden status number;
M:The corresponding observation number of one hidden state;
Pi:Initial probability distribution π=P { q of hidden state1=Sj, sum (π)=1;
A:Transfering probability distribution between hidden state;Aij=P { qt+1=Sj|qt=Si},1≤i,j≤N.sum(Ai)=1;
B:Certain hidden state corresponds to the probability distribution of observation;Bij=P { ot|qt=Si},1≤i≤M,1≤j≤N;
Discrete type observation probability is distributed:bjk=bj(ot)=P (ot=k | st=j), 1≤k≤U;
Continuous type observation probability is distributed:bj(ot)=∑ wjk·bjk(ot), j=1...N, 1≤k≤M;
The probability distribution of the general observation of continuity is fitted using mixed Gaussian approximation to function, ∑ wjk=1, j=1...N, k=
1...M, if observation sequence is multidimensional,:bjk=bj(ot)=N (O, Ujk,∑jk);
Wherein:Ujk=E [O (t) | Q (t)=j, M (t)=k];∑jk=Cov [O (t) | Q (t)=j, M (t)=k]
When finding history point, it is related to the history section to be found, differentiates the method for similar historical point, the mark of history point
When, it needs to solve the computational methods of classification number, when selecting feature, uses the stock index of multidimensional;
Data set length and prediction window section are adjusted to, while ensure the essence of close history point by optimal value by training study
Accurate and characteristic polymorphic, and the self study of adoption status number realize the method for automatically updating model;
In certain history point section, the method for similar historical point is differentiated, during the mark of history point, the computational methods of classification number,
The prediction of stock index is realized on the basis of HMM;The present invention also has certain actual application value:First, as machine learning algorithm
Practical application platform aided education;Second is that in the field of current machine study, realize Prediction of Stock Index, locate thereby using HMM
Manage voice signal;Third, due to the use of hardware cost resource it is less and calculate it is simple and fast, it is easy to accomplish.
3. the adaptive Prediction of Stock Index method of the Hidden Markov Model according to claim 1 based on the multiple features factor,
It is characterized in that:
S1 data preparation:
Step 1:Pass through finance and economics net, Scapy, index calculation formula gather data;
Step 2:These gather datas are normalized regular wait to pre-process;
Step 3:Training dataset and test data set are divided into the collection stock certificate data of pretreatment;
S2 builds model parameter:
Step 1:By the training dataset of the step 3 in S1 in the factor, with the hmmlearn.hmm of python, study
The inner parameter of hmm algorithms, this algorithm are the core algorithms of this model:
(1) it by 55 is that be ultimately determined to 12 be S3 feature selectings to index that input feature value X, which is,;
(2) category label of the output for each sample point, range are that the step 4 of the S2 in hidden state number N determines to be calculated by decoding
Method Viterbi marks to obtain;
(3) the transition probability matrix A=[a of hmmij], 1≤i, j≤N are defined between certain day of trade i and rear adjacent day of trade j
Hidden state transfer probability, 1 rank Markov property principle is utilized;
(4) hmm observes probability distribution B=Bj(k), 1≤j≤N, the hidden state that this matrix defines some day of trade are marked as j
When, the feature vector observed is for the general of k-vector;
(5) involved in S2 steps 1 to hmm algorithms in parameter, be the iteration by more times of learning algorithm baum-welch
Study, convergence obtain when obtaining optimal;
Step 2:It determines in total data set first, the length of the every batch of training sample set of division obtains most under various values
It is excellent;
Step 3:After being determined in historical data section, the length of prediction window is adjusted;
Step 4:The determining of number of the hidden state of the HMM of this operation has program self study to obtain, and is by OEHS standards
Then, operation ncomponents.py obtains most stable of hidden state value, the hidden state number N as hmm;
S3 feature selectings:
Step 1:After the model obtained by S2 parts is applicable in optimized parameter, while using the characteristic procedure method of filtering type, obtain
To the sequence of each factor predictive ability;
Step 2:Simultaneously using PCA and pearson coefficients, the preferably incoherent factor of prediction result is filtered out;
Step 3:Obtain 12 input feature vectors of the dimension as hmm algorithms in 55 ATTRIBUTE INDEXs by the step 1 of S3,2, i.e., feature to
Amount;
S4 models are applicable in:
Step 1:By S2, S3 parts build the optimal value of the suitable parameters arrived involved in model and the parameter of hmm algorithms:It is single
Batch training data sample length, forecast interval length, the similar historical points found, the number N of hidden state, 12 dimension input feature vectors
Vector;Behind the complete basis of model construction, historical time point and day to be measured are calculated into likelihood value and labeled bracketing;
Step 2:By the distance away from day to be predicted and likelihood function apart from program, most similar history point is filtered out, is passed through
Weighted average obtains amount of increase and amount of decrease, so as to obtain the tendency classification of day to be measured;
Step 3:By being run on test set, the result of the model of different situations is obtained.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711488697.1A CN108241872A (en) | 2017-12-30 | 2017-12-30 | The adaptive Prediction of Stock Index method of Hidden Markov Model based on the multiple features factor |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711488697.1A CN108241872A (en) | 2017-12-30 | 2017-12-30 | The adaptive Prediction of Stock Index method of Hidden Markov Model based on the multiple features factor |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108241872A true CN108241872A (en) | 2018-07-03 |
Family
ID=62701274
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711488697.1A Pending CN108241872A (en) | 2017-12-30 | 2017-12-30 | The adaptive Prediction of Stock Index method of Hidden Markov Model based on the multiple features factor |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108241872A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109118368A (en) * | 2018-08-09 | 2019-01-01 | 武汉优品楚鼎科技有限公司 | Financial investment variety analysis method, system and device based on HMM model |
CN109783889A (en) * | 2018-12-26 | 2019-05-21 | 西安工程大学 | Landslide time of origin prediction technique based on mixed Gaussian Hidden Markov Model |
CN110097193A (en) * | 2019-04-28 | 2019-08-06 | 第四范式(北京)技术有限公司 | The method and system of training pattern and the method and system of forecasting sequence data |
WO2020037922A1 (en) * | 2018-08-21 | 2020-02-27 | 平安科技(深圳)有限公司 | Stock index forecasting method, device, and storage medium |
CN113723728A (en) * | 2020-05-26 | 2021-11-30 | 北京九章云极科技有限公司 | Factor checking method and system |
CN115169756A (en) * | 2022-09-07 | 2022-10-11 | 深圳市信润富联数字科技有限公司 | Battery core moisture prediction method, device, equipment and storage medium |
-
2017
- 2017-12-30 CN CN201711488697.1A patent/CN108241872A/en active Pending
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109118368A (en) * | 2018-08-09 | 2019-01-01 | 武汉优品楚鼎科技有限公司 | Financial investment variety analysis method, system and device based on HMM model |
WO2020037922A1 (en) * | 2018-08-21 | 2020-02-27 | 平安科技(深圳)有限公司 | Stock index forecasting method, device, and storage medium |
CN109783889A (en) * | 2018-12-26 | 2019-05-21 | 西安工程大学 | Landslide time of origin prediction technique based on mixed Gaussian Hidden Markov Model |
CN109783889B (en) * | 2018-12-26 | 2023-04-25 | 西安工程大学 | Landslide occurrence time prediction method based on mixed Gaussian hidden Markov model |
CN110097193A (en) * | 2019-04-28 | 2019-08-06 | 第四范式(北京)技术有限公司 | The method and system of training pattern and the method and system of forecasting sequence data |
CN110097193B (en) * | 2019-04-28 | 2021-03-19 | 第四范式(北京)技术有限公司 | Method and system for training model and method and system for predicting sequence data |
CN113723728A (en) * | 2020-05-26 | 2021-11-30 | 北京九章云极科技有限公司 | Factor checking method and system |
CN115169756A (en) * | 2022-09-07 | 2022-10-11 | 深圳市信润富联数字科技有限公司 | Battery core moisture prediction method, device, equipment and storage medium |
CN115169756B (en) * | 2022-09-07 | 2022-12-09 | 深圳市信润富联数字科技有限公司 | Battery core moisture prediction method, device, equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108241872A (en) | The adaptive Prediction of Stock Index method of Hidden Markov Model based on the multiple features factor | |
CN106600059B (en) | Intelligent power grid short-term load prediction method based on improved RBF neural network | |
CN107992976B (en) | Hot topic early development trend prediction system and prediction method | |
CN110751318B (en) | Ultra-short-term power load prediction method based on IPSO-LSTM | |
CN109063911A (en) | A kind of Load aggregation body regrouping prediction method based on gating cycle unit networks | |
CN104751363B (en) | Stock Forecasting of Middle And Long Period Trends method and system based on Bayes classifier | |
CN114841257B (en) | Small sample target detection method based on self-supervision comparison constraint | |
CN112085252B (en) | Anti-fact prediction method for set type decision effect | |
CN110097060A (en) | A kind of opener recognition methods towards trunk image | |
CN106778838A (en) | A kind of method for predicting air quality | |
CN107832259A (en) | A kind of load forecasting method based on time series and Kalman filtering | |
CN108985455A (en) | A kind of computer application neural net prediction method and system | |
CN116187835A (en) | Data-driven-based method and system for estimating theoretical line loss interval of transformer area | |
CN113822419A (en) | Self-supervision graph representation learning operation method based on structural information | |
CN109345381A (en) | A kind of Risk Identification Method and system | |
CN116992779A (en) | Simulation method and system of photovoltaic energy storage system based on digital twin model | |
CN111144462A (en) | Unknown individual identification method and device for radar signals | |
CN117314643A (en) | Data analysis method, device and storage medium based on financial wind tunnel data | |
CN117034060A (en) | AE-RCNN-based flood classification intelligent forecasting method | |
CN107578130A (en) | A kind of method of the copper forecasting of futures prix of the TSK fuzzy systems of structure and parameter combination learning | |
CN115828100A (en) | Mobile phone radiation source spectrogram category increment learning method based on deep neural network | |
CN114372181B (en) | Equipment production intelligent planning method based on multi-mode data | |
Zhao et al. | A Hybrid Time Series Model based on Dilated Conv1D and LSTM with Applications to PM2. 5 Forecasting. | |
CN114463569A (en) | Image matching method and system based on optimization adaptive metric learning | |
CN113296947A (en) | Resource demand prediction method based on improved XGboost model |
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 |
Application publication date: 20180703 |
|
RJ01 | Rejection of invention patent application after publication |