CN109242207A - A kind of Financial Time Series prediction technique based on deeply study - Google Patents
A kind of Financial Time Series prediction technique based on deeply study Download PDFInfo
- Publication number
- CN109242207A CN109242207A CN201811179333.XA CN201811179333A CN109242207A CN 109242207 A CN109242207 A CN 109242207A CN 201811179333 A CN201811179333 A CN 201811179333A CN 109242207 A CN109242207 A CN 109242207A
- Authority
- CN
- China
- Prior art keywords
- value
- subsystem
- vector
- network
- time series
- 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
- 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
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- 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
- G06Q40/06—Asset management; Financial planning or analysis
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Strategic Management (AREA)
- Development Economics (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Theoretical Computer Science (AREA)
- Accounting & Taxation (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Finance (AREA)
- Entrepreneurship & Innovation (AREA)
- Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Game Theory and Decision Science (AREA)
- Tourism & Hospitality (AREA)
- Quality & Reliability (AREA)
- Technology Law (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of Financial Time Series prediction techniques based on deeply study.This method includes three main subsystems: data process subsystem, and the function of this subsystem is to carry out data processing to the initial data got from WindAPI；Feature extraction subsystem, the function of this subsystem are one deep neural networks of construction to extract data characteristics；Intensified learning subsystem, the function of this subsystem is based on Actor-Critic algorithm, Constructing Policy network and assessment network, the selection and evaluation of transaction movement are carried out respectively, then continuous iteration, which is more newly arrived, guarantees that whole system gets the newest multidate information in market, and optimal transaction movement is made according to the status information got, finally obtain preferable trading effect.The present invention can constantly go to learn this complicated financial market, capture the transaction movement that may be made a profit in time, realize profit purpose by some essential informations in financial market.
Description
Technical field
The present invention relates to deep learning and intensified learning field, more particularly, to a kind of based on deeply study
Financial Time Series prediction technique.
Background technique
In the today prevailing that globalizes, financial market reaches unprecedented scale, therefore investment field emerges large quantities of excellent
Elegant analyst predicts assets future price by personal experience and the judgement of some subjective analysis, but traditional deal maker grasps
Make obviously that some are inefficient, therefore people also can carry out quantitative research by computer, instead of duplicate some artificial work, at this moment
Also just occur the algorithm model of conventional quantization investment field in succession, many some ground in conjunction with what machine learning was done wherein also including
Study carefully, as well-known Bayes, support vector machines, Recognition with Recurrent Neural Network etc..
But often the index of those of tradition research method all presence considerations is single, manual operation model prediction is extensive
The problems such as ability is inadequate, and traditional machine learning algorithm is applied and similarly there is obvious lack in Financial Time Series
It falls into, many models especially deep neural network will lead to over-fitting in the forecasting problem of Financial Time Series.Financial market
Be it is very changeable, non-stationary property is a generally existing characteristic of Financial Time Series, this allows many model predictions
When also show unstable situation.
In recent years, after defeating utterly major go master-hand with Alpha's go, intensity study enters completely new burning hot exploration
In the stage, same financial field is no exception, but is different algorithm effect naturally also difference, the following Fig. 1 of general frame
It is shown, nitrification enhancement of the invention also can be used Fig. 1 shows.Intensified learning series also has many algorithms, can substantially divide
For the algorithm based on value function (value-based) and the algorithm based on tactful (policy-based).Calculation based on value function
Method typical case is represented as Q-Learning, and the application in financial field is to define state of market to come, then according to ε-
Greedy strategy removes selection transaction movement, the reward that then environment of observation and movement obtain, and then updating maintenance Q value table, with
This is selected to achieve the purpose that optimize transaction movement, but the problem of this algorithm is when state or the dimension of movement are very big
When, it is difficult to it is safeguarded with the form of Q table, and very difficult convergence.Algorithm Typical Representative based on strategy is Policy-
Gradient, the application in financial field is equally to define state of market, then best dynamic according to having policy selection
Make, the reward of the movement is obtained by environmental feedback, then reversely update policing algorithm, is constantly updated in this way so that is selected is dynamic
It more likely makes a profit, although this method can be adapted for high-dimensional motion space, but the algorithm based on strategy is often
What is converged to is local optimal solution and not global optimum, and bout updates that cause the algorithm based on strategy to update relatively low
Effect.
Summary of the invention
The present invention in order to overcome at least one of the drawbacks of the prior art described above, provides a kind of based on deeply
The Financial Time Series prediction technique of habit.
In order to solve the above technical problems, technical scheme is as follows: when a kind of finance based on deeply study
Sequence data predication method, comprising the following steps:
S1: original Financial Time Series are downloaded from the WindAPI of external system；
S2: the original Financial Time Series after downloading are carried out data prediction by building data process subsystem, and defeated
Pretreated data out；
S3: construction feature extracts subsystem, and pretreated data are carried out to the extraction of depth characteristic, exports and extracts
Depth characteristic information；
S4: depth characteristic information is compared with trading environment, carries out intensified learning by building intensified learning subsystem,
Export transaction movement；
S5: the transaction movement generated according to the data being input in system adjusts the target position in storehouse in financial market in real time, reaches
To transaction purpose.
Preferably, the Financial Time Series in the S1 include valence amount information and macroscopic information；
The valence amount information includes: opening price, closing price, highest price, lowest price, exchange hand；
The macroscopic information include: p/e ratio, price value ratio, city's pin rate, city show rate, circulation value, total market capitalisation, rate of gross profit,
Net assets per share, asset-liability ratio, the exchange rate.
Preferably, data prediction includes computing technique index, the inspection of shortage of data value and processing, data in the S2
Standardization, outlier processing, the specific steps are as follows:
S21: the original Financial Time Series vector v of external system WindAPI output is receivedrawCarry out the inspection of shortage of data value
It looks into, and missing values is marked；
S22: original Financial Time Series vector v is utilizedraw, corresponding technology is calculated with macroscopic information according to valence amount information
Indication information；
S23: time slide window is set as window=600, i.e., is calculated within every 600 minutes once, one can be slided every time by calculating
Minute, calculate v in sliding windowrawThe mean μ of each dimensioniAnd standard deviation sigmai, then every dimension is standardized respectivelyWherein viIt is vrawValue of every dimension without transformation, v'iIt is that every dimension is new after standardization
Value, it is 0 that the information of every dimension, which is all standardized as mean value, the data that variance is 1；
S24: exceptional value is determined using box traction substation method, and by these abnormality value removings, with the mean μ being calculated in S23i
To replace exceptional value；
S25: the mean μ that will be calculated in S23iFill up the missing values being labeled in S21；
S26: by S21 to S25 to original vector vrawPretreatment obtain state of market feature vector vfAt data
Manage the output of subsystem.
Preferably, the technical indicator information includes: Exponential Moving Average MACD, averagely tends to Index A DX, is suitable
Gesture index CCI, relative strength index RSI, cloth forest belt BOOL, average true fluctuation ATR, the equal line EMA_5 of index on the 5th, refer within 20th
The equal line EMA_20 of number, the equal line EMA_60 of index on the 60th, the equal line EMA_120 of index on the 120th, price change rate ROC, random momentum refer to
Number SMI, passenger collector-distributor volume AD, momentum line MOM, price concussion percentage PPO, William's variation discrete magnitude WVAD.
Preferably, the processing step of step S4 intensified learning subsystem is:
S41: the output v of feature extraction subsystem is receiveddeepAnd the input as this subsystem, i.e. state vector vstate；
S42: building acts network, with state vector vstateAs input vector, by 128 hidden layer minds of first layer
Full articulamentum through member, then Relu transformation is done, then export one 128 dimension intermediate vector vam1There are 128 hidden layers to the second layer
The full articulamentum of neuron, then carry out a Relu transformation, the intermediate vector v for finally exporting second layer hidden layeram2By one
The full articulamentum of 3 neuron of layer, the movement vector for exporting one 3 dimension indicate target bin position, this three-dimensional respectively indicates position in storehouse and is
It sells shares 1 hand, position in storehouse is no position in storehouse and position in storehouse is plus the Probability p of 1 hand of storehouse, then obtains final transaction movement a by softmax transformation
=A (s), s herein refer to the state in market, and A (s) refers to the mapping value obtained by acting network transformation, that is, trades
Movement；
S43: after executing the transaction movement a in S42, market environment is obtained according to the feedback i.e. incentive message r of current state
And the next NextState s' of market environment；
S44: loss function loss_actor=-log (p) the * td_error of construction movement network, wherein p is step
The probability of the corresponding movement a chosen in S42, td_error is the Timing Difference error obtained from assessment network, acts network
Optimization aim be minimize loss function loss_actor；
S45: building assessment network, is equally with state vector vstateAs input vector, hidden by first layer 128
The full articulamentum of layer neuron, then Relu transformation is done, then export one 128 dimension intermediate vector vcm1To the second layer have 128 it is hidden
The full articulamentum of layer neuron is hidden, then carries out a Relu transformation, the intermediate vector v for finally exporting second layer hidden layercm2Through
The full articulamentum of one layer of 1 neuron is crossed, the value variable v=Q (s) of one 1 dimension is finally exported, s herein refers to market
State, Q (s) refers to characterizing a value of value, also referred to as Q value, table by assessing the obtained mapping value of network transformation
What is levied is the value of the transaction movement；
S46: seeking Timing Difference error td_error=r+ γ * Q (s')-Q (s), and Q (s) here indicates that state s is input to
The Q value that assessment network obtains, and Q (s') then indicates that NextState s' is input to the Q value that assessment network obtains, what γ here referred to
It is the following reward decaying weight, characterization is more following significance level for rewarding consideration；
S47: the loss function loss_critic=td_error of construction assessment network2, and assess the optimization aim of network
It is then minimum loss function loss_critic.
Preferably, feature extraction subsystem detailed process described in step S3 is after receiving data process subsystem pretreatment
State of market feature vector vfAs input vector, mentioning for depth characteristic is carried out by the neural network of feature extraction subsystem
It takes, obtained depth characteristic vdeep, output as feature extraction subsystem.
Compared with prior art, the beneficial effect of technical solution of the present invention is: present invention employs more high-dimensional information
Description of the vector as state of market, in addition to traditional valence amount information, the present invention is additionally added macroscopic information and multiple technologies refer to
Mark can more efficiently obtain market trend, improve the capture to profit buying signals；Secondly the feature that the present invention uses mentions
It takes subsystem by deep neural network, the extraction of depth characteristic, the characteristic dimension that can analyze is carried out to state of market information
More much bigger than artificial treatment, analysis ability is also more outstanding than traditional artificial judgement；Based on Actor-Critic algorithm
In intensified learning forecasting system, the continuous iteration of Actor obtains the reasonable probability that each movement is selected under each state,
Critic also continuous iteration constantly improve the rewards and punishments value for selecting each to act under each state, not only learns with this method
Practise efficiency faster and also can be good at processing status space it is very big when the case where.
Detailed description of the invention
Fig. 1 is overall system architecture figure of the invention.
Fig. 2 is the deep neural network structure chart that feature of present invention extracts subsystem.
Fig. 3 is the intensified learning frame of Financial Time Series of the present invention.
Fig. 4 is the algorithm flow of ActorCritic of the present invention.
Specific embodiment
The attached figures are only used for illustrative purposes and cannot be understood as limitating the patent；
The following further describes the technical solution of the present invention with reference to the accompanying drawings and examples.
Embodiment 1
Fig. 1 is the general frame figure based on deeply study Financial Time Series forecasting system of the embodiment of the present invention.
As shown in Figure 1, system of the present invention specifically includes that data process subsystem, feature extraction subsystem and intensified learning subsystem
System.There are also other two parts in figure: external system WindAPI downloads original Financial Time Series and according to intensified learning
The transaction movement that subsystem obtains adjusts target position in storehouse and realizes transaction purpose.
In the specific implementation process, external system WindAPI download original Financial Time Series refer to it is soft by Wind
The api interface that part provides downloads to some about valence amount information and macroscopic information, consults here for facilitating, at data
The computing technique indication information part of reason subsystem is also listed in following tables together, can specifically be seen below:
The raw financial time series data of 1 Wind API of table downloading --- valence amount information description
| Title | Symbol | Description |
| Opening price | open | Each minute first tick price |
| Closing price | close | Each minute the last one tick price |
| Highest price | high | The highest price of target in each minute |
| Lowest price | low | The minimum price of target in each minute |
| Exchange hand | volume | The exchange hand of each minute target |
The raw financial time series data of 2 Wind API of table downloading --- macroscopic information description
The raw financial time series data of 3 Wind API of table downloading --- technical indicator information description
Data process subsystem is then responsible for computing technique index, handles abnormal data and pretreatment work.Technical indicator
Part is just as shown above.
The received input of data process subsystem is the initial data downloaded to from external WindAPI interface, t minutes can
With with a vector(vector dimension is (dimValence amount+dimMacroscopic view) × T, wherein T indicates time span) it indicates, every one-dimensional representation
Be target t minutes valence figureofmerits or macro-indicators, the upper table that the two specific indexs include has been listed, subsequent
dimTechnology.Outlier processing include missing values fill up and the amendment of noise spot, the present invention can check missing values and exception first
Noise figure, the pretreatment being then standardized, obtained mean value and standard deviation.With mean value fill up indicated exceptional value with
And missing values.Final process finishes, it will exports pretreated vector(vector dimension is (dimValence amount+dimMacroscopic view+dimTechnology)
×T)。
Feature extraction subsystem receives from data process subsystem and pre-processes the vector finishedAnd as mind
Input vector through network extracts depth characteristic by neural networkThe specific structure of neural network as shown in Fig. 2,
It is made of three layers of full articulamentum, wherein drawing for convenience, hides and part hidden neuron is gone not draw, every layer of full articulamentum
Hidden neuron number be 16.Input vectorDimension be dimoutput=dimValence amount+dimMacroscopic view+dimTechnology, i.e.,
diminput=31, and output vectorDimension be dimoutput=16.
Algorithm used by intensified learning subsystem is based on movement-assessment algorithm, constantly obtains each minute shape in market
State selects transaction movement by movement network, and makes evaluation to the transaction movement of selection by assessing network, and not dropping out of school, it is excellent to practise
Change and how to choose better transaction movement.The overall intensified learning process of this subsystem is as shown in Fig. 3 intensified learning frame.
System of the invention is the transaction intelligent body in figure, and trading environment refers to market.Transaction intelligent body passes through
Observe the state in market, intelligent body is according to his experience and gets enough multidate informations in the market and judges, and obtains
The transaction movement that should be taken under this state, and transaction movement is gone in trading environment, to obtain trading environment for this
The feedback (reward i.e. in figure) of transaction movement, then intelligent body just according to obtained from system feedback reward value, state of market
And transaction movement carry out intensified learning, on the one hand continually strengthen to obtain should be taken under the state of market of various complexity how
Transaction movement be more conducive to making a profit, on the other hand constantly obtain market latest tendency, guarantee intelligent body can constantly answer
More suitable transaction movement is made to newest fluctuation of price, and according to newest multidate information.
Specifically, movement-assessment algorithm Actor- of intensified learning subsystem of the present invention is introduced below with reference to Fig. 4
Critic, intensified learning subsystem are the output for receiving feature extraction subsystemAnd inputted as the conduct of this subsystem,
That is state vector
As shown in Figure 4, building acts Network Learning Strategies, with state vectorAs input vector, by first layer
There is the full articulamentum of 128 hidden layer neurons, then do Relu transformation, then exports one 128 dimension intermediate vectorTo second
Layer has the full articulamentum of 128 hidden layer neurons, then carries out a Relu transformation, then will be in the full articulamentum generation of the second layer
Between vectorIt is input in the full articulamentum of 3 neurons of the last layer, the movement vector for exporting one 3 dimension indicates target bin
Position, it is 1 hand (- 1) of hole capital after selling all securities that this three-dimensional, which respectively indicates position in storehouse, and position in storehouse is the probability of more 1 hands of storehouse (1) for no position in storehouse (0) and position in storehouse, then
Final transaction movement is obtained by softmax transformation, the whole process for obtaining transaction movement is represented by at=A (st), herein
stRefer to the state in t minutes market；Assessment e-learning is constructed simultaneously to evaluate movement, is equally with state vector
vstateAs input vector, by the full articulamentum of 128 hidden layer neurons of first layer, then Relu transformation is done, then exported
One 128 dimension intermediate vectorThere is the full articulamentum of 128 hidden layer neurons to the second layer, then carry out a Relu transformation,
The intermediate vector that finally second layer hidden layer is exportedThe valence of one 1 dimension is exported by the full articulamentum of one layer of 1 neuron
It is worth vector, the whole process for obtaining value vector is represented by vt=Q (st)。
As seen from Figure 4, as the transaction movement a for receiving system outputtAfterwards, market environment can be according to the transaction made
Act atIntensified learning subsystem two values are returned to, one is the lower minute state s that will jump in markett+1, the other is this
Secondary selection transaction movement atIncentive message
As shown in figure 4, the assessment network portion of intensified learning subsystem can be according to accessed information to movement network
Evaluation is made in the movement of selection, and process will seek Timing Difference error td_error=rat+γ*Q(st+1)-Q(st), here
Q (st) expression state s is input to and assesses the Q value that network obtains, and Q (st+1) then indicate NextState st+1It is input to assessment net
The Q value that network obtains, γ are attenuation coefficients, and the range of γ is 0≤γ < 1.
And entire algorithm will be constantly updated, then construct loss function respectively for the movement of subsystem-assessment network.Construction is dynamic
Make loss function loss_actor=-log (p) the * td_error of network, wherein p chooses dynamic for correspondence in step S42
Make atProbability, td_error is from the obtained Timing Difference error of assessment network, this is introduced next, movement network
Optimization aim be minimize loss function loss_actor；The loss function loss_critic=td_ of construction assessment network
error2, and the optimization aim for assessing network is then to minimize loss function loss_critic；
Its each minute specific learning process are as follows:
Movement network and assessment network observe the state of market s of this minute from environmentt, state state here
VectorIt indicates；
Network is acted according to state of market, carries out the prediction a of transaction movementt=A (st), prediction process can be segmented specifically are as follows: defeated
Enter state vectorIt converts, is exported by the full articulamentum of first layer, and by Relu
Wherein ωakAnd bakIt is weight and the biasing for acting the full articulamentum of network kth layer, the ω appeared belowckAnd bckIt then indicates to assess
The weight of the full articulamentum of network kth layer and biasing；It converts, is exported by the full articulamentum of the second layer, and by ReluBy the full articulamentum of third layer, the probability of each transaction movement is obtainedAnd converted by softmax, obtain transaction movement at=softmax (pt)；
Environment gets the output a of movement networkt, two feedbacks can be generated, one is that market will jump to for lower minute
State st+1, equally can also be by state vectorIt indicates, the other is this time selection transaction movement atIncentive message
Network is assessed according to state of market, the prediction v evaluatedt=Q (st), specific prediction process can also be segmented are as follows: input
State vectorIt converts, is exported by the full articulamentum of first layer, and by Relu
It converts, is exported by the full articulamentum of the second layer, and by ReluIt is complete by third layer
Articulamentum obtains required value vector
Seek Timing Difference errorWherein 0≤γ < 1；
Respectively by minimizing loss function come the parameter of iteration update action network and assessment network.Construction acts network
Loss function be loss_actor=-log (p) * td_error, construction assessment network loss function be loss_critic
=td_error2, this minimum process is specifically in the back-propagation completion of neural network.
The overall process of intensified learning subsystem is: one movement network of construction, this network understands basis from transaction ring
The reward that border obtains selects most suitable transaction the probability for taking various transaction movement under different conditions is adjusted
Movement, in addition can also construct an assessment network, this network is a learning network based on value, is updated by single step
Algorithm calculates the rewards and punishments information of each step, and two network integrations get up, and tactful network selects transaction movement, assessment network evaluation choosing
The superiority and inferiority degree of transaction movement out, then continuous iteration updates, and completes the selection to movement (target position in storehouse) and optimizes.
In traditional Q-learning algorithm, the meeting benefit reason when the limited amount of state, indicating to get up is also
Eaily, but when very big in face of state of market space, traditional Q-learning is then helpless.Traditional
In Policy Gradient algorithm, this reward is by covering a complete episode come what is be calculated, this is unavoidably
It is very slow to result in learning rate, takes a long time just acquire thing.The present invention is based on the strong of Actor-Critic algorithm
Chemistry is practised in forecasting system, then can well solve the problem of the two conventional methods encounter, the continuous iteration of Actor obtains
Select the reasonable probability of each movement under each state, Critic also continuous iteration, constantly improve selected under each state it is every
The rewards and punishments value of one movement.So, so that the transaction movement selected every time is become better and better, the excellent of transaction movement is assessed every time
It is bad also more and more quasi-.
The terms describing the positional relationship in the drawings are only for illustration, should not be understood as the limitation to this patent；
Obviously, the above embodiment of the present invention be only to clearly illustrate example of the present invention, and not be pair
The restriction of embodiments of the present invention.For those of ordinary skill in the art, may be used also on the basis of the above description
To make other variations or changes in different ways.There is no necessity and possibility to exhaust all the enbodiments.It is all this
Made any modifications, equivalent replacements, and improvements etc., should be included in the claims in the present invention within the spirit and principle of invention
Protection scope within.
Claims (7)
1. a kind of Financial Time Series prediction technique based on deeply study, it is characterised in that: the following steps are included:
S1: original Financial Time Series are downloaded from the WindAPI of external system；
S2: the original Financial Time Series after downloading are carried out data prediction, and exported pre- by building data process subsystem
Data that treated；
S3: construction feature extracts subsystem, and pretreated data are carried out to the extraction of depth characteristic, export the depth extracted
Characteristic information；
S4: depth characteristic information is compared building intensified learning subsystem with trading environment, carries out intensified learning, output
Transaction movement；
S5: the transaction movement generated according to the data being input in system adjusts the target position in storehouse in financial market in real time, reaches friendship
Easy purpose.
2. a kind of Financial Time Series prediction technique based on deeply study according to claim 1, feature exist
In: the Financial Time Series in the S1 include valence amount information and macroscopic information；
The valence amount information includes: opening price, closing price, highest price, lowest price, exchange hand；
The macroscopic information includes: that p/e ratio, price value ratio, city's pin rate, city show rate, circulation value, total market capitalisation, rate of gross profit, per share
Net assets, asset-liability ratio, the exchange rate.
3. a kind of Financial Time Series prediction technique based on deeply study according to claim 1, feature exist
In: data prediction includes computing technique index, the inspection of shortage of data value and processing, data normalization, exception in the S2
Value processing, the specific steps are as follows:
S21: the original Financial Time Series vector v of external system WindAPI output is receivedrawThe inspection of shortage of data value is carried out, and
Missing values are marked；
S22: original Financial Time Series vector v is utilizedraw, corresponding technical indicator is calculated with macroscopic information according to valence amount information
Information；
S23: setting time slide window as window=600, i.e., calculate within every 600 minutes once, and calculating can slide one minute every time,
Calculate v in sliding windowrawThe mean μ of each dimensioniAnd standard deviation sigmai, then every dimension is standardized respectivelyWherein viIt is vrawValue of every dimension without transformation, vi' it is that every dimension is new after standardization
Value, it is 0 that the information of every dimension, which is all standardized as mean value, the data that variance is 1；
S24: exceptional value is determined using box traction substation method, and by these abnormality value removings, with the mean μ being calculated in S23iCarry out generation
For exceptional value；
S25: the mean μ that will be calculated in S23iFill up the missing values being labeled in S21；
S26: by S21 to S25 to original vector vrawPretreatment obtain state of market feature vector vfAs data processing
The output of system.
4. a kind of Financial Time Series prediction technique based on deeply study according to claim 3, feature exist
In: the technical indicator information includes: Exponential Moving Average MACD, averagely tends to Index A DX, index of taking advantage of a situation CCI, phase
To strong and weak index RSI, cloth forest belt BOOL, average true fluctuation ATR, the equal line EMA_5 of index on the 5th, the equal line EMA_20 of index on the 20th,
The equal line EMA_60 of index on the 60th, the equal line EMA_120 of index on the 120th, price change rate ROC, random momentum index SMI, passenger collector-distributor volume
AD, momentum line MOM, price concussion percentage PPO, William's variation discrete magnitude WVAD.
5. a kind of Financial Time Series prediction technique based on deeply study according to claim 1, feature exist
In: feature extraction subsystem detailed process described in step S3 is to receive the pretreated state of market spy of data process subsystem
Levy vector vfAs input vector, the extraction of depth characteristic, obtained depth are carried out by the neural network of feature extraction subsystem
Spend feature vdeep, output as feature extraction subsystem.
6. a kind of Financial Time Series prediction technique based on deeply study according to claim 1, feature exist
In: the processing step of the intensified learning subsystem of step S4 is:
S41: the output v of feature extraction subsystem is receiveddeepAnd the input as this subsystem, i.e. state vector vstate；
S42: building acts network, with state vector vstateAs input vector, by 128 hidden layer neurons of first layer
Full articulamentum, then do Relu transformation, then export one 128 dimension intermediate vector vam1There are 128 hidden layer nerves to the second layer
The full articulamentum of member, then carry out a Relu transformation, the intermediate vector v for finally exporting second layer hidden layeram2By one layer 3
The full articulamentum of a neuron, the movement vector for exporting one 3 dimension indicate target bin position, and it is to sell shares that this three-dimensional, which respectively indicates position in storehouse,
1 hand, position in storehouse is no position in storehouse and position in storehouse is the Probability p for adding 1 hand of storehouse, then obtains final transaction movement a=A by softmax transformation
(s), s herein refers to the state in market, and A (s) refers to the mapping value obtained by acting network transformation, i.e. transaction is dynamic
Make；
S43: after executing the transaction movement a in S42, obtain market environment according to the feedback i.e. incentive message r of current state and
The next NextState s' of market environment；
S44: loss function loss_actor=-log (p) the * td_error of construction movement network, wherein p is in step S42
The probability of the corresponding movement a chosen, td_error is the Timing Difference error obtained from assessment network, acts the optimization of network
Target is to minimize loss function loss_actor；
S45: building assessment network, is equally with state vector vstateAs input vector, by 128 hidden layer minds of first layer
Full articulamentum through member, then Relu transformation is done, then export one 128 dimension intermediate vector vcm1There are 128 hidden layers to the second layer
The full articulamentum of neuron, then carry out a Relu transformation, the intermediate vector v for finally exporting second layer hidden layercm2By one
The full articulamentum of 1 neuron of layer, finally exports the value variable v=Q (s) of one 1 dimension, and s herein refers to the shape in market
State, Q (s) refer to characterizing a value of value, also referred to as Q value by assessing the obtained mapping value of network transformation, characterization
It is the value of the transaction movement；
S46: Timing Difference error td_error=r+ γ * Q (s')-Q (s), γ here is asked to refer to the following reward decaying power
Weight, characterization is more following significance level for rewarding consideration, and Q (s) here indicates that state s is input to what assessment network obtained
Q value, and Q (s') then indicates that NextState s' is input to the Q value that assessment network obtains；
S47: the loss function loss_critic=td_error of construction assessment network2, and the optimization aim for assessing network is then
Minimize loss function loss_critic.
7. according to the method described in claim 1, it is characterized by: trading activity described in step S5, then be according to extensive chemical
The transaction movement a that subsystem obtains, and the practical position in storehouse of the target position in storehouse adjustment firm offer indicated according to transaction movement are practised, it is final complete
The process traded at systematic learning.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811179333.XA CN109242207A (en) | 2018-10-10 | 2018-10-10 | A kind of Financial Time Series prediction technique based on deeply study |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811179333.XA CN109242207A (en) | 2018-10-10 | 2018-10-10 | A kind of Financial Time Series prediction technique based on deeply study |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109242207A true CN109242207A (en) | 2019-01-18 |
Family
ID=65055679
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811179333.XA Pending CN109242207A (en) | 2018-10-10 | 2018-10-10 | A kind of Financial Time Series prediction technique based on deeply study |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109242207A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110059896A (en) * | 2019-05-15 | 2019-07-26 | 浙江科技学院 | A kind of Prediction of Stock Index method and system based on intensified learning |
| CN110322351A (en) * | 2019-07-04 | 2019-10-11 | 天津工业大学 | Multi-source driving quantization investment model under Depth Stratification strategy |
| CN110428333A (en) * | 2019-08-01 | 2019-11-08 | 上海爱富爱克斯网络科技发展有限责任公司 | For the Data Detection and analysis method of financial intermediary MT4 foreign exchange transaction software |
| CN111310915A (en) * | 2020-01-21 | 2020-06-19 | 浙江工业大学 | Data anomaly detection and defense method for reinforcement learning |
| CN111562996A (en) * | 2020-04-11 | 2020-08-21 | 北京交通大学 | Method and system for detecting time sequence abnormality of key performance index data |
| CN111754033A (en) * | 2020-06-15 | 2020-10-09 | 西安工业大学 | Non-stationary time sequence data prediction method based on recurrent neural network |
| CN111798061A (en) * | 2020-07-08 | 2020-10-20 | 洋浦美诺安电子科技有限责任公司 | Financial data prediction system based on block chain and cloud computing |
| CN111798018A (en) * | 2019-04-09 | 2020-10-20 | Oppo广东移动通信有限公司 | Behavior prediction method, behavior prediction device, storage medium and electronic equipment |
| CN112884576A (en) * | 2021-02-02 | 2021-06-01 | 上海卡方信息科技有限公司 | Stock trading method based on reinforcement learning |
| CN114595923A (en) * | 2022-01-11 | 2022-06-07 | 电子科技大学 | Group teaching recommendation system based on deep reinforcement learning |
| CN116934486A (en) * | 2023-09-15 | 2023-10-24 | 深圳格隆汇信息科技有限公司 | Decision evaluation method and system based on deep learning |
| US11809817B2 (en) * | 2022-03-16 | 2023-11-07 | Tata Consultancy Services Limited | Methods and systems for time-series prediction under missing data using joint impute and learn technique |
-
2018
- 2018-10-10 CN CN201811179333.XA patent/CN109242207A/en active Pending
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111798018A (en) * | 2019-04-09 | 2020-10-20 | Oppo广东移动通信有限公司 | Behavior prediction method, behavior prediction device, storage medium and electronic equipment |
| CN110059896A (en) * | 2019-05-15 | 2019-07-26 | 浙江科技学院 | A kind of Prediction of Stock Index method and system based on intensified learning |
| CN110322351A (en) * | 2019-07-04 | 2019-10-11 | 天津工业大学 | Multi-source driving quantization investment model under Depth Stratification strategy |
| CN110428333A (en) * | 2019-08-01 | 2019-11-08 | 上海爱富爱克斯网络科技发展有限责任公司 | For the Data Detection and analysis method of financial intermediary MT4 foreign exchange transaction software |
| CN111310915A (en) * | 2020-01-21 | 2020-06-19 | 浙江工业大学 | Data anomaly detection and defense method for reinforcement learning |
| CN111310915B (en) * | 2020-01-21 | 2023-09-01 | 浙江工业大学 | Data anomaly detection defense method oriented to reinforcement learning |
| CN111562996A (en) * | 2020-04-11 | 2020-08-21 | 北京交通大学 | Method and system for detecting time sequence abnormality of key performance index data |
| CN111562996B (en) * | 2020-04-11 | 2021-11-23 | 北京交通大学 | Method and system for detecting time sequence abnormality of key performance index data |
| CN111754033A (en) * | 2020-06-15 | 2020-10-09 | 西安工业大学 | Non-stationary time sequence data prediction method based on recurrent neural network |
| CN111798061A (en) * | 2020-07-08 | 2020-10-20 | 洋浦美诺安电子科技有限责任公司 | Financial data prediction system based on block chain and cloud computing |
| CN111798061B (en) * | 2020-07-08 | 2020-12-29 | 深圳市荣亿达科技发展有限公司 | Financial data prediction system based on block chain and cloud computing |
| CN112884576A (en) * | 2021-02-02 | 2021-06-01 | 上海卡方信息科技有限公司 | Stock trading method based on reinforcement learning |
| CN114595923A (en) * | 2022-01-11 | 2022-06-07 | 电子科技大学 | Group teaching recommendation system based on deep reinforcement learning |
| US11809817B2 (en) * | 2022-03-16 | 2023-11-07 | Tata Consultancy Services Limited | Methods and systems for time-series prediction under missing data using joint impute and learn technique |
| CN116934486A (en) * | 2023-09-15 | 2023-10-24 | 深圳格隆汇信息科技有限公司 | Decision evaluation method and system based on deep learning |
| CN116934486B (en) * | 2023-09-15 | 2024-01-12 | 深圳市蓝宇飞扬科技有限公司 | Decision evaluation method and system based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109242207A (en) | A kind of Financial Time Series prediction technique based on deeply study | |
| Tian | Short-term wind speed prediction based on LMD and improved FA optimized combined kernel function LSSVM | |
| Puchalsky et al. | Agribusiness time series forecasting using Wavelet neural networks and metaheuristic optimization: An analysis of the soybean sack price and perishable products demand | |
| Abraham et al. | A neuro-fuzzy approach for modelling electricity demand in Victoria | |
| Cho et al. | Using social network analysis and gradient boosting to develop a soccer win–lose prediction model | |
| Aliahmadi et al. | Comparing artificial neural networks and regression methods for predicting crude oil exports | |
| Omidi et al. | Forecasting stock prices using financial data mining and Neural Network | |
| Almutairi et al. | An intelligent deep learning based prediction model for wind power generation | |
| CN116526450A (en) | Error compensation-based two-stage short-term power load combination prediction method | |
| Tripura et al. | Simultaneous streamflow forecasting based on hybridized neuro-fuzzy method for a river system | |
| de A. Araújo | A quantum‐inspired evolutionary hybrid intelligent approach for stock market prediction | |
| Atashbar et al. | AI and macroeconomic modeling: Deep reinforcement learning in an RBC model | |
| CN110516792A (en) | Non-stable time series forecasting method based on wavelet decomposition and shallow-layer neural network | |
| Lemke et al. | Self-organizing data mining for a portfolio trading system | |
| CN110322351A (en) | Multi-source driving quantization investment model under Depth Stratification strategy | |
| Sprevak | Predictive coding III: The algorithmic level | |
| Asiri et al. | Short-Term Load Forecasting in Smart Grids using Hybrid Deep Learning | |
| Lunga et al. | Online forecasting of stock market movement direction using the improved incremental algorithm | |
| Kokkinos et al. | Efficiency in energy decision support systems using soft computing techniques | |
| Hussain et al. | Optimization of fuzzy neural network using APSO for predicting strength of Malaysian SMEs | |
| Al-Mashhadani et al. | Forecasting Malaysian stock price using artificial neural networks (ANN) | |
| Xue | Prediction of NASDAQ Index Based on Stacking Method | |
| Liu et al. | A Short-Term Load Forecasting Method using Integrated SVR and LSTM Network | |
| Shi | Learning from zero: how to make consumption-saving decisions in a stochastic environment with an AI algorithm | |
| Chaudhari et al. | Anfis based model in decision making to optimize the profit in farm cultivation |
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: 20190118 |