CN108985941A - A kind of stock intelligent Forecasting of combination newsletter archive - Google Patents

A kind of stock intelligent Forecasting of combination newsletter archive Download PDF

Info

Publication number
CN108985941A
CN108985941A CN201810791693.9A CN201810791693A CN108985941A CN 108985941 A CN108985941 A CN 108985941A CN 201810791693 A CN201810791693 A CN 201810791693A CN 108985941 A CN108985941 A CN 108985941A
Authority
CN
China
Prior art keywords
stock
newsletter archive
character representation
moment
vector
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
Application number
CN201810791693.9A
Other languages
Chinese (zh)
Inventor
李晓东
贡诚
冯钧
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hohai University HHU
Original Assignee
Hohai University HHU
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Hohai University HHU filed Critical Hohai University HHU
Priority to CN201810791693.9A priority Critical patent/CN108985941A/en
Publication of CN108985941A publication Critical patent/CN108985941A/en
Pending legal-status Critical Current

Links

Classifications

    • 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
    • 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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/205—Parsing
    • G06F40/216—Parsing using statistical methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities
    • G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Business, Economics & Management (AREA)
  • Data Mining & Analysis (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Finance (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Accounting & Taxation (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Marketing (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Development Economics (AREA)
  • Economics (AREA)
  • Computing Systems (AREA)
  • Strategic Management (AREA)
  • Technology Law (AREA)
  • General Business, Economics & Management (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a kind of stock intelligent Forecastings of combination newsletter archive, pre-process first to newsletter archive, filter Chinese word segmentation and stop words, delete the newsletter archive of not time tag；The prediction duration Δ t for determining stock filters selection newsletter archive according to the time tag of newsletter archive；Character representation is carried out to the newsletter archive for filtering selection, and forms the character representation vector at corresponding moment with the stock certificate data feature vector at corresponding momentSelf-encoding encoder deep learning network is constructed, by character representation vectorInput self-encoding encoder deep learning network carries out compression and feature extraction, obtains low-dimensional character representation vector

Description

A kind of stock intelligent Forecasting of combination newsletter archive
Technical field
The invention belongs to technical field of data processing, and in particular to a kind of stock intelligent predicting side of combination newsletter archive Method.
Background technique
Financial market, especially stock market, it is both closely bound up with historical quotes, while being also extremely easy by burst gold Melt the influence of media event.At present Common Prediction Method be the forecasting problem of stock market is modeled as recurrence in machine learning or Classification problem.The prior art includes: building forecasting system, and the forecasting problem of stock is reduced to classification problem, by analyzing shadow The some feature distributions for ringing the newsletter archive of stock market are judged；Stock is carried out using the machine learning model of multicore Regression forecasting.However, existing most methods, which often depend upon expert, carrys out selected characteristic.But simple dependence is manually from big It is very difficult for measuring and character representation is found and excavated in complicated stock certificate data and newsletter archive.And for point of machine learning It, can pole if cannot carry out preferable feature vector to data set with vector indicates for generic task and some other task The precision of prediction of big influence classifier.
Deep learning character representation (Deep learned representation, DLR) includes the nerve net of a multilayer Network has the ability for taking out its character representation from sample using the network structure of multilayer, the energy that this extraction feature indicates Power is completely unsupervised.Non-linear but relatively simple activation primitive can be used in each layer of DLR, by the spy of input Sign indicates that vector transforms into more abstract character representation vector.Therefore, the feature by multilayer neural network after abstract Expression can provide a good sample input quantity for the classification task in future.
Application of the deep learning in financial field includes: to construct the model system of investment combination decision, utilizes coding certainly Device (auto-encoder) carries out investment combination, by classifier which determines only on the multitiered network structure of deep learning Stock will do very well in stock market from now on；Analysis and pre- is carried out to stock market using the multilayer neural network of deep learning It surveys.But existing Prediction of Stock Index method, it is usually that the media event that will affect shares changing tendency and historical quotes data are separately examined Consider, and not can be carried out good character representation when handling data, limits Prediction of Stock Index accuracy to a certain extent.
Summary of the invention
To solve the above problems, the present invention proposes a kind of stock intelligent Forecasting of combination newsletter archive, news is realized The combination of event and historical quotes data solves the low technical problem of Prediction of Stock Index accuracy.
The present invention adopts the following technical scheme that a kind of stock intelligent Forecasting of combination newsletter archive, specific steps are such as Under:
1) newsletter archive is pre-processed, filters Chinese word segmentation and stop words, delete the news text of not time tag This；
2) the prediction duration Δ t for determining stock filters selection newsletter archive according to the time tag of newsletter archive；
3) character representation, and the stock certificate data feature vector composition with the corresponding moment are carried out to the newsletter archive for filtering selection The character representation vector at corresponding moment
4) self-encoding encoder deep learning network, the character representation vector that step 3) is obtained are constructedInput self-encoding encoder Deep learning network carries out compression and feature extraction, obtains low-dimensional character representation vector
5) ELM neural network model is constructed, quantificational expression is carried out to the variation degree of share price, determines ELM neural network mould The target output value of type；
6) target output value according to determined by step 5), and the low-dimensional character representation vector that step 4) is obtainedAs the input of ELM neural network model, optimizes ELM neural network model parameter, obtain final prediction model.
Preferably, selection newsletter archive, specific steps are filtered according to the time tag of newsletter archive in the step 2) are as follows:
21) Prediction of Stock Index duration Δ t is determined；
22) data that stock possesses are in [ts, te] in the period, from tsStart, t0∈[ts, teΔ t] filtering selection away from From t0The nearest newsletter archive of+time Δt is as t0The newsletter archive at moment traverses [ts, teΔ t] whole is obtained after the period Newsletter archive.
Preferably, the character representation vector at corresponding moment is formed in the step 3)Specific steps are as follows:
31) reverse document-frequency (the term frequency-inverse document of word frequency-is carried out to newsletter archive Frequency, tf-idf) it calculates, obtain t0The tf-idf value of the newsletter archive word at moment is as t0The newsletter archive at moment Character representation, all newsletter archives include n different words altogether, then t0Moment newsletter archive character representation is [w1, w2..., wn], wherein w1、w2、wnRespectively represent the tf-idf value of first word, second word and n-th of word；
32) according to specific stock certificate data choose needed for index composition stock certificate data feature vector, index include opening price, Closing price, exchange hand and highest price, certain branch stock include m index, then t0The stock certificate data feature vector at moment is [q1, q2..., qm], wherein q1, q2, qm respectively indicates first index, second index and m-th of index in t0Moment it is specific Value；
33) feature that step 31) and step 32) obtain is combined, obtains the newsletter archive and stock at corresponding moment Character representation vector after data combination:
Preferably, low-dimensional character representation vector is obtained in the step 4)Specifically:
Each layer of neural network is all self-encoding encoder, and feature is arranged in the dimension of the character representation vector obtained according to step 3) Indicate vector compression ratio, compression ratio be equal to deep neural network in the number of first layer neuron and the last layer neuron it Than the number of first layer neuron is the dimension for the character representation vector that step 3) obtains, the number of the last layer neuron The dimension of i.e. final low-dimensional character representation vector；Determine the number of plies of deep learning network and the neuron that each layer is included Number is arranged from first layer to the number linear decrease of last one layer of neuron.
Preferably, the target output value of ELM neural network model is determined in the step 5) specifically: ELM neural network The input of model is t0The low-dimensional character representation vector at momentTarget output value is t0The target of+time Δt exports yi:
Wherein, r (x) is comparison function, reflects t0+ time Δt is relative to t0The variation degree at moment, θ are according to stock reality The threshold value of border situation setting, when stock is in t0When the rise degree of+time Δt is greater than threshold θ, then target exports yiIt is+1, instantly When drop degree is greater than threshold θ, target exports yiIt is -1, when variation degree is less than threshold θ, shares changing tendency held stationary, then mesh Mark output yiIt is 0.
Preferably, final prediction model specific steps are obtained in the step 6) are as follows:
61) according to the prediction duration Δ t of stock, the historical quotes data for the stock for needing to predict and newsletter archive are carried out Cutting, and mark that form sample set as follows:
{(X1, Y1), (X2, Y2) ..., (Xl, Yl)}
Sample set includes l sample, wherein XiFor input value, t is indicatediMoment passes through the obtained low-dimensional feature of step 4) Indicate vectorYiFor output valve, indicate to pass through the obtained t of step 5)iThe target output value of+time Δt；
62) sample set is divided into training sample set and test sample collection, using training sample set to common ELM and core letter The neural network of several KELM is trained, and is carried out cross validation to two kinds of ELM using test sample collection, is determined final prediction Model.
It invents achieved the utility model has the advantages that the present invention is a kind of stock intelligent Forecasting of combination newsletter archive, realizes The combination of media event and historical quotes data solves the low technical problem of Prediction of Stock Index accuracy.
For disaggregated model, the learning machine that transfinites (extreme learning machine, ELM) model can reach Higher accuracy rate can be used to analyze a large amount of data sample.Weight between the learning machine network layer that transfinites and partially Initial value can be assigned at random by setting, and the parameter all to model inside is avoided to carry out tuning.The present invention will affect stock price The market of history and the media event of burst, which combine, forms character representation vector, carries out character representation to it using DLR, and by stock City's forecasting problem is converted into a classification problem, and with transfiniting, learning machine carries out classification prediction, significantly improves short-term Prediction of Stock Index Accuracy rate.
Detailed description of the invention
Fig. 1 is the flow chart that newsletter archive is screened in the embodiment of the present invention；
Fig. 2 is newsletter archive screening schematic diagram of the invention；
Fig. 3 is the relational graph of threshold θ and predictablity rate acc of the invention；
Fig. 4 is model training flow chart of the invention.
Specific embodiment
Below according to attached drawing and technical solution of the present invention is further elaborated in conjunction with the embodiments.
A kind of stock intelligent Forecasting of combination newsletter archive, comprising the following steps:
1) as shown in Figure 1, being pre-processed to newsletter archive, Chinese word segmentation and stop words is filtered, newsletter archive is converted For convenience of the data format of computer identification, the newsletter archive of not time tag is deleted；
Common text processing software Jieba or NLTK (natural language toolkit) can be achieved in filtering The pretreatments such as literary participle, stop words；
2) the prediction duration Δ t for determining stock filters selection newsletter archive according to the time tag of newsletter archive；
21) Prediction of Stock Index duration Δ t is determined；
22) data that stock possesses are in [ts, te] in the period, from tsStart, t0∈[ts, teΔ t], filtering selection Distance t0The nearest newsletter archive of+time Δt is as t0The newsletter archive at moment traverses [ts, teΔ t] it is obtained after the period entirely Portion's newsletter archive；
As shown in Fig. 2, wherein t0As current time, t Δ are t0+ time Δt, when only one is new in this period When hearing text, then select the newsletter archive as t0The newsletter archive at moment, when there is a plurality of newsletter archive in this period, then Select distance t0The nearest newsletter archive of+time Δt is as t0The newsletter archive at moment.
With t0The stock certificate data and distance t at moment0The corresponding original number as input of the nearest newsletter archive of+time Δt According to, and t0Reference data of the share price predicted needed for+time Δt as output.
3) character representation, and the stock certificate data feature vector composition with the corresponding moment are carried out to the newsletter archive for filtering selection The character representation vector at corresponding moment；
31) tf-idf calculating is carried out to newsletter archive, obtains t0The tf-idf value of the newsletter archive word at moment is as t0When The character representation of the newsletter archive at quarter, all newsletter archives include n different word { word altogether1, word2... wordn }, Middle word1, word2, wordnFirst word, second word and n-th of word are respectively represented, then corresponds to moment t0News text Eigen is expressed as the vector [w of n dimension1, w2..., wn], wherein w1、w2、wnRespectively represent first word, second word The tf-idf value of language and n-th of word, it may be assumed that
wn=tf-idf (wordn)
32) according to specific stock certificate data choose needed for index composition stock certificate data feature vector, index include opening price, Closing price, exchange hand and highest price, certain branch stock include m index, then t0The stock certificate data feature vector at moment is [q1, q2..., qm], wherein q1, q2, qmFirst index, second index and m-th of index are respectively indicated in t0Moment it is specific Value；
33) feature that step 31) and step 32) obtain is combined, obtains corresponding moment t0Newsletter archive and stock Character representation vector after ticket data combination:
4) self-encoding encoder deep learning network is constructed, the character representation vector that step 3) is obtained inputs self-encoding encoder depth Learning network carries out compression and feature extraction, obtains low-dimensional character representation vector；
Each layer of neural network is all self-encoding encoder, and feature is arranged in the dimension of the character representation vector obtained according to step 3) Indicate vector compression ratio, compression ratio be equal to deep neural network in the number of first layer neuron and the last layer neuron it Than the number of first layer neuron is the dimension for the character representation vector that step 3) obtains, the number of the last layer neuron The dimension of i.e. final low-dimensional character representation vector；Determine the number of plies of deep learning network and the neuron that each layer is included Number is arranged from first layer to the number linear decrease of last one layer of neuron.
The character representation vector exported by multilayer neural networkWith original inputIt compares, in addition to Other than dimension is reduced, the feature after newsletter archive and the combination of historical stock data can be more taken out.
It should be noted that many of prior art Open-Source Tools can build deep learning network, such as sklearn Deep learning tool box (deep-learning-toolbox) in (Scikit learn), keras and matlab, with deep- For learning-toolbox, neural network module therein (neural network, NN) module is chosen, in setup module Parameter, that is, network number of plies and every layer of neuron number, build deep learning network.The layer of network is set in the present embodiment Number is 10, and first layer neuron number is 1011, and the last layer neuron number is 860, from first layer to last one layer of nerve First number obeys linear decrease ordered series of numbers.
5) ELM neural network model is constructed, quantificational expression is carried out to the variation degree of share price, determines ELM neural network mould The target output value of type；
The input of ELM neural network model is the t that step 4) obtains0The low-dimensional character representation vector at moment Target output value is t0The target of+time Δt exports yi:
Wherein, r (x) is comparison function, is t in the present embodiment0+ time Δt stock price subtracts t0Moment stock price, Reflect t0+ time Δt stock is relative to t0The variation degree of moment stock, θ are the threshold values being arranged according to stock actual conditions, when Stock is in t0When the rise degree of+time Δt is greater than threshold θ, stock can rise, then target exports yiIt is+1；When decline degree is big When threshold θ, stock can decline, and target exports yiIt is -1, when variation degree is less than threshold θ, shares changing tendency held stationary, then Target exports yiIt is 0.
In financial field, the situation of change obedience of stock is just distributed very much, and has fertile tail effect (fat tail), and above-mentioned three The probability that kind happens is indicated with following formula:
Wherein pdfGaussian(x) the Gaussian Profile probability density function of stock price, P are indicated-1, P0And P+1It respectively indicates The probability of stock decline, held stationary or rising.By above formula it is found that threshold θ directly determines the probability of three kinds of situations, For the time series of no any priori knowledge, future condition is carried out according to the probability distribution that hypothesis time series is obeyed Prediction, shown in the following formula of the relationship of predictablity rate and probability distribution:
It is symmetrically obtained according to θ and-θ about 0 simultaneously:
P-1=P+1
P0=1-2P+1
In conjunction with above-mentioned formula, can finally obtain:
Acc=6p2-4P+1
P=P-1=P+1, relationship such as Fig. 3 of threshold θ and predictablity rate acc.
From figure 3, it can be seen that accuracy is predictablity rate acc, it is inverted similar to one with the modified-image of threshold θ The curve of SIN function, predictablity rate acc minimum point obtain at p=1/3, and threshold θ is about 0.4 at this time, accurate in prediction Predictablity rate acc is increased with the raising of threshold θ on the right side of rate acc minimum point, but when θ is very big, most of share price Variation is all predicted to be and remains unchanged, then Prediction of Stock Index is nonsensical, on the left of predictablity rate acc minimum point accuracy rate with The raising of θ and reduce, so the smaller predictablity rate of θ is higher, but θ cannot be below every transaction expense of stock market, otherwise cannot Embody the fluctuation of stock market, thus the reasonable value range of θ should be equal to or slightly larger than stock market average every transaction expense, with perfume (or spice) For Hong Kong stock city, θ takes average every transaction expense 0.003 that stock market trades (30bps, bps indicate net assets per share).
6) target output value according to determined by step 5), and the low-dimensional character representation vector that step 4) is obtainedAs input, optimizes ELM neural network model parameter, obtain final prediction model.
61) according to the prediction duration Δ t of stock, the historical quotes data for the stock for needing to predict and newsletter archive are carried out Cutting, and mark that form sample set as follows:
{(X1, Y1), (X2, Y2) ..., (Xl, Yl)}
Sample set includes l sample, wherein XiFor input value, t is indicatediMoment passes through the obtained low-dimensional feature of step 4) Indicate vectorYiFor output valve, indicate to pass through the obtained t of step 5)iThe target output value of+time Δt.
62) as shown in figure 4, sample set is divided into training sample set and test sample collection, using training sample set to basis ELM and the neural network of KELM of kernel function be trained, the input value of training sample set is inputted into neural network, obtains mind Output through network, while error calculation is carried out with the target output value of training sample set, the error of network is obtained, error is anti- To propagation, the weight of every layer of neuron, all training sample sets of loop iteration, until error convergence to a certain range are updated. For ELM, also need to choose two parameters.By taking RBF Radial basis kernel function as an example, Radial basis kernel function needs to be arranged two ginsengs Number γ and C, the range of γ are set as { 2-17, 2-16..., 22, the range of C is set as { 2-5, 2-4..., 214, pass through open source Software chooses the optimal value of the two parameters on training sample set by way of crosscheck.
Cross validation is carried out using neural network of the test sample collection to the ELM on basis and the KELM of kernel function, is determined most Whole prediction model.
It is predicted using the variation tendency of method of the invention to future stock, is made with the stock in 33 Hong Kong in 2001 For test, and only consider newsletter archive or only consider stock certificate data ELM neural network prediction result compare, at 5 points In clock -30 minutes time span of forecasts, accuracy rate rises 0.02, and and consider the principal component analysis of newsletter archive and stock certificate data (principal component analysis, PCA) feature learning method compares, and accuracy rate improves 0.03.

Claims (6)

1. a kind of stock intelligent Forecasting of combination newsletter archive, which comprises the following steps:
1) newsletter archive is pre-processed, filters Chinese word segmentation and stop words, delete the newsletter archive of not time tag；
2) the prediction duration Δ t for determining stock filters selection newsletter archive according to the time tag of newsletter archive；
3) character representation is carried out to the newsletter archive for filtering selection, and forms and corresponds to the stock certificate data feature vector at corresponding moment The character representation vector at moment
4) self-encoding encoder deep learning network, the character representation vector that step 3) is obtained are constructedInput self-encoding encoder depth Learning network carries out compression and feature extraction, obtains low-dimensional character representation vector
5) ELM neural network model is constructed, quantificational expression is carried out to the variation degree of share price, determines ELM neural network model Target output value；
6) target output value according to determined by step 5), and the low-dimensional character representation vector that step 4) is obtainedMake For the input of ELM neural network model, optimizes ELM neural network model parameter, obtain final prediction model.
2. a kind of stock intelligent Forecasting of combination newsletter archive according to claim 1, which is characterized in that the step It is rapid 2) according to the time tag of newsletter archive filter selection newsletter archive, specific steps are as follows:
21) Prediction of Stock Index duration Δ t is determined；
22) data that stock possesses are in [tS, te] in the period, from tsStart, t0∈[ts, teΔ t] filtering selection distance t0+ The nearest newsletter archive of time Δt is as t0The newsletter archive at moment traverses [ts, teΔ t] whole news are obtained after the period Text.
3. a kind of stock intelligent Forecasting of combination newsletter archive according to claim 1, which is characterized in that the step Rapid 3) the middle character representation vector for forming the corresponding momentSpecific steps are as follows:
31) tf-idf calculating is carried out to newsletter archive, obtains t0The tf-idf value of the newsletter archive word at moment is as t0Moment The character representation of newsletter archive, all newsletter archives include n different words altogether, then t0Moment newsletter archive character representation is [w1, w2..., wn], wherein w1、w2、wnRespectively represent the tf-idf value of first word, second word and n-th of word；
32) feature vector of the composition stock certificate data of index needed for being chosen according to specific stock certificate data, index include opening price, closing quotation Valence, exchange hand and highest price, certain branch stock include m index, then t0The stock certificate data feature vector at moment is [q1, q2..., qm], wherein q1, q2, qmFirst index, second index and m-th of index are respectively indicated in t0The occurrence at moment；
33) feature that step 31) and step 32) obtain is combined, obtains the newsletter archive and stock certificate data at corresponding moment In conjunction with character representation vector later:
4. a kind of stock intelligent Forecasting of combination newsletter archive according to claim 1, which is characterized in that the step It is rapid 4) in obtain low-dimensional character representation vectorSpecifically:
Each layer of neural network is all self-encoding encoder, and character representation is arranged in the dimension of the character representation vector obtained according to step 3) The compression ratio of vector, compression ratio are equal to the ratio between first layer neuron and the number of the last layer neuron in deep neural network, The number of first layer neuron is the dimension for the character representation vector that step 3) obtains, and the number of the last layer neuron is i.e. most The dimension of whole low-dimensional character representation vector；Determine the number of plies of deep learning network and of neuron that each layer is included Number is arranged from first layer to the number linear decrease of last one layer of neuron.
5. a kind of stock intelligent Forecasting of combination newsletter archive according to claim 1, which is characterized in that the step It is rapid 5) in determine ELM neural network model target output value specifically: the input of ELM neural network model be t0Moment it is low Dimensional feature indicates vectorTarget output value is t0The target of+time Δt exports yi:
Wherein, r (x) is comparison function, reflects t0+ time Δt is relative to t0The variation degree at moment, θ are according to the practical feelings of stock The threshold value of condition setting, when stock is in t0When the rise degree of+time Δt is greater than threshold θ, then target exports yiIt is+1, when decline journey When degree is greater than threshold θ, target exports yiIt is -1, when variation degree is less than threshold θ, shares changing tendency held stationary, then target is defeated Y outiIt is 0.
6. a kind of stock intelligent Forecasting of combination newsletter archive according to claim 1, which is characterized in that the step It is rapid 6) in obtain final prediction model specific steps are as follows:
61) according to the prediction duration Δ t of stock, the historical quotes data for needing the stock predicted and newsletter archive are cut Point, and mark that form sample set as follows:
{(X1, Y1), (X2, Y2) ..., (Xl, Yl)}
Sample set includes l sample, wherein XiFor input value, t is indicatediMoment passes through the obtained low-dimensional character representation of step 4) VectorYiFor output valve, indicate to pass through the obtained t of step 5)iThe target output value of+time Δt；
62) sample set is divided into training sample set and test sample collection, using training sample set to common ELM and kernel function The neural network of KELM is trained, and is carried out cross validation to two kinds of ELM using test sample collection, is determined final prediction mould Type.
CN201810791693.9A 2018-07-18 2018-07-18 A kind of stock intelligent Forecasting of combination newsletter archive Pending CN108985941A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810791693.9A CN108985941A (en) 2018-07-18 2018-07-18 A kind of stock intelligent Forecasting of combination newsletter archive

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810791693.9A CN108985941A (en) 2018-07-18 2018-07-18 A kind of stock intelligent Forecasting of combination newsletter archive

Publications (1)

Publication Number Publication Date
CN108985941A true CN108985941A (en) 2018-12-11

Family

ID=64549789

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810791693.9A Pending CN108985941A (en) 2018-07-18 2018-07-18 A kind of stock intelligent Forecasting of combination newsletter archive

Country Status (1)

Country Link
CN (1) CN108985941A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110489631A (en) * 2019-07-10 2019-11-22 平安科技（深圳）有限公司 Stock market development method, apparatus, computer equipment and storage medium
CN112036165A (en) * 2019-05-14 2020-12-04 西交利物浦大学 Method for constructing news characteristic vector and application
CN112116465A (en) * 2020-06-04 2020-12-22 上海金融期货信息技术有限公司 Deep learning model-based transaction position ratio prediction method and system
CN116258579A (en) * 2023-04-28 2023-06-13 成都新希望金融信息有限公司 Training method of user credit scoring model and user credit scoring method

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112036165A (en) * 2019-05-14 2020-12-04 西交利物浦大学 Method for constructing news characteristic vector and application
CN110489631A (en) * 2019-07-10 2019-11-22 平安科技（深圳）有限公司 Stock market development method, apparatus, computer equipment and storage medium
CN112116465A (en) * 2020-06-04 2020-12-22 上海金融期货信息技术有限公司 Deep learning model-based transaction position ratio prediction method and system
CN112116465B (en) * 2020-06-04 2023-09-29 上海金融期货信息技术有限公司 Method and system for predicting trading-holding bin ratio based on deep learning model
CN116258579A (en) * 2023-04-28 2023-06-13 成都新希望金融信息有限公司 Training method of user credit scoring model and user credit scoring method
CN116258579B (en) * 2023-04-28 2023-08-04 成都新希望金融信息有限公司 Training method of user credit scoring model and user credit scoring method

Similar Documents

Publication Publication Date Title
US11086292B2 (en) Methods and apparatus for machine learning predictions of manufacturing processes
US10712727B2 (en) Methods and apparatus for machine learning predictions of manufacture processes
CN108985941A (en) A kind of stock intelligent Forecasting of combination newsletter archive
CN109886020A (en) Software vulnerability automatic classification method based on deep neural network
CN108665093B (en) Deep learning-based expressway traffic accident severity prediction method
Kirange et al. Sentiment Analysis of news headlines for stock price prediction
CN116541911B (en) Packaging design system based on artificial intelligence
Wu et al. Optimized deep learning framework for water distribution data-driven modeling
KR20220090360A (en) Device and method for analyzing and visualizing big data by controlling hyperparameters
CN111177010B (en) Software defect severity identification method
US11410020B2 (en) Machine learning models for predicting time in traffic
Drass et al. Semantic segmentation with deep learning: detection of cracks at the cut edge of glass
CN114584601A (en) User loss identification and intervention method, system, terminal and medium
CN117764227A (en) Customer loss prediction device for gas station
KR20220090358A (en) Device and method for analyzing and visualizing big data
KR20220090359A (en) Device and method for analyzing and visualizing big data by correcting errors
CN116611911A (en) Credit risk prediction method and device based on support vector machine
CN113268929B (en) Short-term load interval prediction method and device
CN113656707A (en) Financing product recommendation method, system, storage medium and equipment
CN113592341A (en) Measurement loss function, sector complexity evaluation method and system
KR20220090361A (en) Device and method for analyzing and visualizing big data by proposing blocks
CN113326882A (en) Model integration method and device based on classification and regression algorithm
CN116341990B (en) Knowledge management evaluation method and system for infrastructure engineering
CN115982646B (en) Management method and system for multisource test data based on cloud platform
Uma et al. EVALUATION OF ENSEMBLE LEARNING APPROACH FOR OPTIMISED STOCK PREDICTION

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: 20181211

RJ01 Rejection of invention patent application after publication