CN110223180A - Portfolio Selection Based method based on depth attention network and intensified learning - Google Patents
Portfolio Selection Based method based on depth attention network and intensified learning Download PDFInfo
- Publication number
- CN110223180A CN110223180A CN201910390018.XA CN201910390018A CN110223180A CN 110223180 A CN110223180 A CN 110223180A CN 201910390018 A CN201910390018 A CN 201910390018A CN 110223180 A CN110223180 A CN 110223180A
- Authority
- CN
- China
- Prior art keywords
- stock
- intensified learning
- victor
- investment
- model
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
- 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
Abstract
The Portfolio Selection Based method based on depth attention network and intensified learning that the invention discloses a kind of, the neural network model for merging attention mechanism is introduced into financial field, using Sharpe Ratio as reward function, income and risk are balanced when generating Portfolio Selection Based using intensified learning frame training pattern.It is also proposed that modeling the correlation between different assets by completely new across assets attention mechanism, and deep exploration is carried out in terms of model interpretation.Superior performance, generalization and strong robustness are achieved in Sino-U.S. history stock market and firm offer simulation.
Description
Technical field
The present invention relates to finance data minings and intensified learning technical field, more particularly to a kind of to be based on depth
The Portfolio Selection Based method of attention network and intensified learning.
Background technique
In recent years, financial innovation and the combination of artificial intelligence technology take in multiple financial fields including quantization transaction
Obtained successful application.In view of being capable of handling large-scale transaction and providing rational decision making support, quantization trading strategies are for a long time
It is used by financial institution and hedge fund always and achieves surprising success.
Traditional quantization trading strategies are typically based on specific financial logic.A kind of viewpoint thinks the money that history is excellent in
Production will continue to keep preferably to show at the next moment, and this phenomenon is known as momentum effect；And the viewpoint of another mean reversion
Think that assets price can be intended to mean value with the time.Although these traditional quantization trading strategies have solid Finance
Support, but the part of properties in financial market is all only focused on, therefore be difficult to obtain good performance on complicated financial market
Performance.
Recently with the rise of deep learning and artificial intelligence technology, it is pre- that share price much is carried out using deep neural network
It surveys, is also suggested using the method that intensified learning generates investment tactics, but at present using deep learning to real financial market
Research still suffers from challenge below: first is that for the balance of income and risk, second is that the correlation between assets how is modeled,
Third is that the interpretation of investment tactics.
Therefore, how to provide that a kind of Generalization Capability is strong, the investment combination method suitable for different complicated market scenarios is this
The problem of field technical staff's urgent need to resolve.
Summary of the invention
In view of this, the present invention provides a kind of Portfolio Selection Based sides based on depth attention network and intensified learning
The neural network model for merging attention mechanism is introduced into financial field by method, using Sharpe Ratio as reward function, is used
Intensified learning frame training pattern balances income and risk when generating Portfolio Selection Based.It is also proposed that by completely new
Across assets attention mechanism model the correlation between different assets, and carried out in terms of model interpretation deep
It explores.Superior performance, generalization and strong robustness are achieved in Sino-U.S. history stock market and firm offer simulation.
To achieve the goals above, the present invention adopts the following technical scheme:
A kind of Portfolio Selection Based method based on depth attention network and intensified learning, comprising:
S1: based on financial field, the classical victor that buys in sells the vanquished's strategy and recently the stock spy of K historical juncture
Vector is levied, preliminary characterization extraction is carried out using the shot and long term memory network with historic state attention mechanism, obtains stock table
Levy r；
S2: r and across assets attention networks are characterized to model the correlativity between stock based on stock, and thus selected
Victor and the vanquished obtain victor's scoring of each branch stock；
S3: it is scored according to the victor of current all stocks and generates investment combination；
S4: it is based on investment combination, using Sharpe Ratio as the reward function of intensified learning, Optimized model parameter.
Preferably, step S1 is specifically included:
LSTM models Temporal dependency: hk=LSTM (hk-1, xk), k ∈ [1, K], wherein hkIt is that LSTM is obtained in kth step coding
Hidden state, the hidden state h that final step obtainsKAs the characterization of stock, it contains Temporal dependency all in X；X is most
The characteristic sequence of nearly K historical juncture composition,
X={ x1..., xk... xK, wherein
The attention mechanism modeling overall situation and long-term dependence for introducing historic state obtain new stock characterization r, meter
Calculate formula are as follows:
Wherein ATT () is that attention calculates function, is defined as:
αk=wT·tanh(W(1)hk+W(2)hK)
wT, W(1), W(2)Learn for corresponding coefficient matrix, random initializtion and being continued to optimize in subsequent training process
To final value.This refers to be continued to optimize using Sharpe Ratio as reward function.
Preferably, step S2 is specifically included:
Give the characterization r of any one stock(i), query vector q is each mapped to by parameter matrix(i), index
Vector k(i)With value vector v(i)
q(i)=W(Q)r(i), k(i)=W(K)r(i), v(i)=W(v)r(i)
The index of stock j is inquired in correlativity modeling between stock j and stock i using stock i, i.e.,
Wherein, DkIt is standardization coefficient；
Using the correlativity after normalization as weight, weight to obtain using the value vector of current all I branch stocks new
Stock characterizes a(i)；
Finally, vector a will be characterized using full articulamentum(i)It is mapped as victor's scoring: s(i)=sigmoid (w(s)T·a(i)+e(s))
Wherein, w(s)For the parameter matrix of full articulamentum, e(s)For the biasing of full articulamentum.
Preferably, step S3 is specifically included:
The victor of all stocks is scored and is arranged according to descending, and obtains sequence number o of any stock i after sequence(i), parameter preset G decision buying long, the stock number sold short respectively；
To any stock i, if o(i)∈ [1, G], then buy long for stock i, and the ratio between investments bought long calculates
It is as follows:
Similarly, if o(i)∈ (I-G, I], then the ratio between investments sold short, and sold short for stock i calculates as follows:
Remaining is not chosen as the stock of victor or the vanquished, then in default of buying long/short sales signal without investment.
Preferably, step S4 is specifically included:
One continuing investment sequence with T holding period of sampling is concentrated from training data every time, it is special to input history respectively
Vector is levied, the investment combination sequence at each moment is obtained according to model
The Sharpe Ratio of entire Investment Sequences and the reward function as intensified learning can be then calculated, is risen using gradient
Method optimizes model parameter, i.e. argmaxHT{B+, B-}；
Wherein,ATIt is the average return of the Investment Sequences each holding period, calculation formula isTCtTake for actual transaction procedure, VTIt is the fluctuation of Investment Sequences, uses the mark of income
Quasi- difference is measured, and calculation formula is
Preferably, after step s4, further includes: model is explained using significance analysis method.
Preferably, described to include: to the specific steps that model explains using significance analysis method
Partial derivative is asked to the various features of input according to victor scoring, according to the size of partial derivative and positive and negative is analyzed
The influence that each feature plays the decision of model.
It can be seen via above technical scheme that compared with prior art, the present disclosure provides one kind to be infused based on depth
The neural network model for merging attention mechanism is introduced into finance by the Portfolio Selection Based method of meaning power network and intensified learning
In field, using Sharpe Ratio as reward function, put down using intensified learning frame training pattern when generating Portfolio Selection Based
The income that weighs and risk.It is also proposed that an investment combination based on across assets attention mechanism generates model, and wherein
Use intensified learning for Optimized model parameter.The it is proposeds of across assets attention mechanism, the different assets of very good solution it
Between the problem that models of correlativity, Generalization Capability is strong.Intensified learning frame enables model to receive with Sharpe Ratio for guiding balance
Benefit and risk.In addition, the method by significance analysis can explore the interpretation of model, deep learning is opened
Internal black box.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
The embodiment of invention for those of ordinary skill in the art without creative efforts, can also basis
The attached drawing of offer obtains other attached drawings.
Fig. 1 attached drawing is the Portfolio Selection Based method provided by the invention based on depth attention network and intensified learning
Schematic diagram；
Fig. 2 attached drawing is that the LSTM stock characterization extraction knot module provided by the invention based on historic state attention mechanism is shown
It is intended to；
Fig. 3 attached drawing is across assets attention schematic network structures provided by the invention；
Fig. 4 attached drawing is the accumulated earnings that model provided by the invention is obtained in US stock market up to 27 years (1990-2016)
Curve and comparison with existing method；
Fig. 5 attached drawing is the interpretation analysis provided by the invention which factor to generate investment combination according to for model.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
The following detailed description of the Portfolio Selection Based side of the deeply study provided by the invention based on attention mechanism
Method.
1, data demand
Since the stock feature vector of input is really the time for including stock exchange data and corresponding company's financial report data
Sequence.Therefore, other than the attributes such as closing price, trading volume, issuance of shares number that stock exchange data itself should include, also
It needs comprising listed company corresponding with stock in the financial report data of synchronization, at least should include p/e ratio, price value ratio, dividend
Equal attributes.
2, technical essential
The present invention relates to the cross discipline application of artificial intelligence field and financial field, by intensified learning frame for realizing
The BWSL strategy of financial field classics.Model mainly includes three parts, and first part uses the LSTM comprising attention mechanism
It is extracted to carry out the characterization of stock, followed by across assets attention networks (Cross Asset Attention
Network, CAAN) modeling stock between correlativity, obtain every stock victor scoring.It is finally scored and is felt according to victor
It obtains and whether the branch stock is bought long or sold short, and distribute ratio between investments, generate final Portfolio Selection Based, refer to attached
Fig. 1.
Entire model is optimized end to end using intensified learning, using Sharpe Ratio as reward function, enables model
Enough pay close attention to the balance of long-term gain and risk.
2.1 financial logics and problem definition
For the ease of introduce, introduce first involved in invention it is basic finance concept, and provide Portfolio Selection Problem and
The mathematical definition of BWSL strategy.
Holding period: holding period is the minimum time unit invested to assets.The present invention divides time shaft in order
For the holding period with regular length, such as day degree or monthly.In the present invention, it is referred to as at the beginning of t-th of holding period
Moment t.
Continuing investment: continuing investment includes one section of continuous holding period, and for t-th of holding period, an investment tactics makes
Assets are invested with capital, and obtain income (may be negative) at the t+1 moment.
Assets price: the price of an assets is defined as a time series:
WhereinI-th of assets is represented in the price of t moment.Stock is used in the present invention to be described as assets,
This can also be expanded into other kinds of assets.
Buying long: buying long transactional operation is to a certain assets in t1Moment is bought in, in t2Moment is sold.In buying long operation, transaction
Person expects that the price of the assets can go up in future.
Short sales: short sale is operated to stock trader in t1Moment borrows a number of assets to sell, in t2Moment buys back identical number
The assets of amount return stock trader.In selling do-nothing operation, dealer expects that the price of the assets can drop in future.
Investment combination: the fixed Assets Pool being made of I assets, an investment combination are defined as a vector b=(b(1)..., b(i)..., b(I))T, wherein b(i)It is the ratio between investments to i-th of assets, and
Zero investment combination: zero investment combination refers to, for a series of investment combination { b1..., bj... bJCombine
When, their gross investment is 0, i.e.,Wherein MjIndicate the investment distributed j-th of investment combination.
BWSL investment tactics: the present invention is by classical the buying in financial field based on victor sells the vanquished (BWSL), BWSL
Core concept buy in the high assets of the following amount of increase (victor, winners), sell the low assets of the following amount of increase (the vanquished,
losers).The present invention executes BWSL strategy: the throwing of a buying long using the zero investment combination mode comprising two investment combinations
Money combination b+For buying in victor, the investment combination b of a short sales-For the vanquished that sells short.
Any moment t gives budgetary restraintsAccording toIn ratio between investments, can calculate separately and sell in do-nothing operation
The borrowing quantity of any stock iThen it is obtained by selling short these the vanquished's stocks
Capital, and then by buy long operation buy in victor's stock, can equally calculate any stock buys in quantity
Ending, sells the stock of buying long, buys back the stock return of short sales to stock trader, can finally calculate t moment strategy
Earning rate are as follows:
By the formula as it can be seen that obtain positive income Rt> 0 then means that the average price amount of increase of victor's stock is higher than and loses
The average price amount of increase of person's stock.That is, even if current time all stock is all dropping, as long as can guarantee victor
The average drop range of stock is less than the vanquished's stock, then BWSL strategy still can obtain positive income.So the core of BWSL strategy is just
It is to consider the relative performance between stock, it is as so correct that select victor and the vanquished as possible according to history feature.
2.2 stocks characterization is extracted
Original issue stock feature includes two classes, and one kind is the original transaction feature of stock, including price amount of increase, stability bandwidth and friendship
Yi Liang.Another kind of is the financial report data that stock corresponds to listed company, including market value, p/e ratio, price value ratio, dividend.Handle stock
Primitive character composition characteristic vector, the feature vector of any moment tFeature comprising above seven dimensions.
The Portfolio Selection Based of any moment t in order to obtain, the input X to design a model are composition of nearest K historical juncture
Characteristic sequence, i.e. X={ x1..., xk... xK, wherein
First using with historic state attention mechanism shot and long term memory (Long Short-Term Memory,
LSTM it) carries out preliminary characterization to extract, structure is as shown in Figure 2.
LSTM models Temporal dependency: hk=LSTM (hk-1, xk), k ∈ [1, K], wherein hkIt is that LSTM is obtained in kth step coding
Hidden state, the hidden state h that final step obtainsKAs the characterization of stock, it contains Temporal dependency all in X；
Historic state attention mechanism: the attention mechanism modeling overall situation and long-term dependence of historic state are introduced, is obtained
R is characterized to new stock, its calculation formula is:
Wherein ATT () is that attention calculates function, is defined as:
αk=wT·tanh(W(1)hk+W(1)hK)
I.e. first according to hK, hkCalculate the corresponding attention force vector α of any momentk, then calculated using softmax function
Obtain the attention weight ATT (h at corresponding momentK, hk).Wherein wT, W(1), W(2)For corresponding coefficient matrix, random initializtion is simultaneously
Study is continued to optimize in subsequent training process obtains final value.
For i-th stock of t moment, can be expressed as in conjunction with the obtained stock characterization of LSTM and history attention mechanismIt includes the timing and global dependence from t-K+1 to t moment, in model proposed by the present invention, all stock
It all shares the same stock characterization and extracts network, all LSTM-HA (remember, Long Short by the shot and long term of history attention
Term Memory-Historical Attention) extract stock characterization also have stronger generalization.
2.3 victor the vanquisheds selection
In order to fully demonstrate the characteristic that the BWSL strategy of 2.1 section introductions extremely pays close attention to relative performance, sufficiently model stock it
Between correlativity, select victor and the vanquished.Notice that power module (as shown in Figure 3) will be extracted according to 2.2 sections across assets
Stock characterization is that every stock generates corresponding victor's scoring.
Give the characterization r of any one stock(i)(omit time index t and without loss of generality), passes through parameter matrix first
It is each mapped to query vector q(i), index vector k(i)With value vector v(i)；
q(i)=W(Q)r(i), k(i)=W(K)r(i), v(i)=W(v)r(i)
Wherein W(Q), W(K), W(v)It respectively inquires, index, the coefficient matrix that value vector calculates, random initializtion and rear
Study, which is continued to optimize, in continuous training process obtains final value.
The index of stock j is inquired in correlativity modeling between stock j and i using stock i, i.e.,
Wherein DkIt is standardization coefficient, the correlativity after next using normalization uses current all I as weight
The value vector of branch stock weights to obtain new stock characterization a(i), attention coefficient S ATT (q(i), k(j)) equally pass through softmax
Function is calculated.
Finally, use full articulamentum will characterize DUAL PROBLEMS OF VECTOR MAPPING be victor score, victor scoring it is higher represent current time be somebody's turn to do
Stock is more likely to become victor: s(i)=sigmoid (w(s)T·a(i)+e(s)), wherein w(s)For the parameter matrix of full articulamentum.
2.4 investment combinations generate
Victor's scoring { s of every stock is obtained by the CAAN module of 2.3 sections(1)..., s(i)... s(I), the present invention is to that
A little stocks for obtaining higher victor's scoring carry out buying long operation, and the stock to score for those lower victors carries out short sales behaviour
Make.It is arranged firstly, scoring the victor of all stocks according to descending, and obtains sequence number o of any stock i after sequence(i).Parameter preset G determines the stock number bought long, sold short respectively, and (the stock number that fictitious transaction for convenience, is arranged is identical
All it is G)
To any stock i, if o(i)∈ [1, G], then buy long for stock i, and the ratio between investments bought long calculates
It is as follows:
This is softmax function, and each S (i) is compressed between 0~1 by exponential average, i ' here be for
And i is distinguished, change j into indicate the formula can also with.
Similarly, if o(i)∈ (I-G, I], then the ratio between investments sold short, and sold short for stock i calculates as follows:
Remaining is not chosen as the stock of victor or the vanquished, then in default of buying long/short sales signal without investment.
The optimization of 2.5 intensified learnings
In order to enable model to consider that the income and risk of investment tactics, the present invention use Sharpe Ratio (income wind simultaneously
Dangerous ratio) it is optimized as the reward function of intensified learning.
Sharpe Ratio: Sharpe Ratio has reacted under unit risk, and the risk income rate of investment is beyond risk free return
Degree gives the chain transaction comprising T holding period, and corresponding Sharpe Ratio can calculate with the following method:
Wherein ATIt is the average return of the Investment Sequences each holding period, calculation formula isTCtTake for actual transaction procedure.VTIt is the fluctuation of Investment Sequences, uses the mark of income
Quasi- difference is measured, and calculation formula is
In training process, one continuing investment sequence with T holding period of sampling is concentrated from training data every time, respectively
History feature vector is inputted, the investment combination sequence at each moment is obtained according to model
The Sharpe Ratio of entire Investment Sequences and the reward function as intensified learning can be then calculated, is risen using gradient
Method optimizes model parameter, i.e. argmaxHT{B+, B-}。
Below with reference to specific example to further illustrate the technical scheme of the present invention.
Embodiment is related to the historical simulation transaction performance and explanatory analysis of US stock market.
1, data
US stock market is up to 47 years stock exchange datas and corresponding listed company's financial report numbers from 1970.12-2016.12
Cover many great financial events and different state of market according to, period, for example, 1995-2000 Internet bubble and
The subprime crisis of 2007-2009.Effective stock every year on average is more than 1000, this allows under different state of market
Fully measure actual performance of the invention.
2, evaluation index
The index of the assessment investment combination performance of most standard is exactly to calculate the accumulated earnings obtained in time horizon of vestment, calculation formula
For
3, American market performance
Based on method provided by the invention, model is trained using the data of 1970.1-1989.12, in the U.S.
The result obtained between stock market 1990.1-2016.12 is 27 years total is as follows.
As shown in figure 4, AS, AS-NP, AS-NC, which respectively represent the present invention, proposes that model AlphaStock, AlphaStock are gone
Fall history attention mechanism, AlphaStock removes performance acquired by across assets attention mechanism.Other several this line difference
Represent the leading edge method of deep bid index, financial field and computer field.It can see AlphaStock mould proposed by the present invention
Type and its reduction version outclass the conventional method of other benchmark in the performance of American history mock trading in the market.
4, explanatory analysis
The present invention use sensitivity analysis method, attempt open deep learning model black box, interpretation model how root
The selection of investment combination is carried out according to the history feature of input.
As shown in figure 5, horizontal axis represents the K historical juncture (being set as history in this experiment 12 months) of input, longitudinal axis generation
The correlation that table input feature vector scores with last victor.
From left to right four width figure respectively represents price amount of increase feature, transaction measure feature, stability bandwidth feature and taking-over market
Influence to final result.As can be seen that model proposed by the present invention, which tends to selection, has higher long-term gain, lower fluctuation
Rate, higher intrinsic value and the stock underestimated recently are as victor.
Each embodiment in this specification is described in a progressive manner, the highlights of each of the examples are with other
The difference of embodiment, the same or similar parts in each embodiment may refer to each other.For device disclosed in embodiment
For, since it is corresponded to the methods disclosed in the examples, so being described relatively simple, related place is said referring to method part
It is bright.
The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention.
Various modifications to these embodiments will be readily apparent to those skilled in the art, as defined herein
General Principle can be realized in other embodiments without departing from the spirit or scope of the present invention.Therefore, of the invention
It is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein phase one
The widest scope of cause.
Claims (7)
1. a kind of Portfolio Selection Based method based on depth attention network and intensified learning characterized by comprising
S1: based on financial field it is classical buy in victor sell the vanquished's strategy and recently the stock feature of K historical juncture to
Amount carries out preliminary characterization extraction using the shot and long term memory network with historic state attention mechanism, obtains stock characterization r；
S2: r and across assets attention networks are characterized to model the correlativity between stock based on stock, and thus select victor
And the vanquished, obtain victor's scoring of each branch stock；
S3: it is scored according to the victor of current all stocks and generates investment combination；
S4: it is based on investment combination, using Sharpe Ratio as the reward function of intensified learning, Optimized model parameter.
2. a kind of Portfolio Selection Based method based on depth attention network and intensified learning according to claim 1,
It is characterized in that, step S1 is specifically included:
LSTM models Temporal dependency: hk=LSTM (hk-1, xk), k ∈ [1, K], wherein hkIt is hidden that be LSTM obtain in kth step coding
State, the hidden state h that final step obtainsKAs the characterization of stock, it contains Temporal dependency all in X；X is nearest K
The characteristic sequence of a historical juncture composition, X={ x1..., xk... xK, wherein
The attention mechanism modeling overall situation and long-term dependence for introducing historic state obtain new stock characterization r, calculate public
Formula are as follows:
Wherein ATT () is that attention calculates function, is defined as:
αk=wT·tanh(W(1)h k+W(2)hK)
wT, W(1), W(2)For corresponding coefficient matrix, random initializtion and study is continued to optimize in subsequent training process obtain most
Whole value.
3. a kind of Portfolio Selection Based method based on depth attention network and intensified learning according to claim 2,
It is characterized in that, step S2 is specifically included:
Give the characterization r of any one stock(i), query vector q is each mapped to by parameter matrix(i), index vector k(i)With value vector v(i)
q(i)=W(Q)r(i), k(i)=W(K)r(i), v(i)=W(v)r(i)
The index of stock j is inquired in correlativity modeling between stock j and stock i using stock i, i.e.,
Wherein, DkIt is standardization coefficient；
Using the correlativity after normalization as weight, weight to obtain new stock using the value vector of current all I branch stocks
Characterize a(i)；
Finally, vector a will be characterized using full articulamentum(i)It is mapped as victor's scoring: s(i)=sigmoid (w(s)T·a(i)+e(s))
Wherein, w(s)For the parameter matrix of full articulamentum, e(s)For the biasing of full articulamentum.
4. a kind of Portfolio Selection Based method based on depth attention network and intensified learning according to claim 3,
It is characterized in that, step S3 is specifically included:
The victor of all stocks is scored and is arranged according to descending, and obtains sequence number o of any stock i after sequence(i), in advance
Setting parameter G determines the stock number bought long, sold short respectively；
To any stock i, if o(i)∈ [1, G], then buy long for stock i, and the ratio between investments bought long calculates such as
Under:
Similarly, if o(i)∈ (I-G, I], then the ratio between investments sold short, and sold short for stock i calculates as follows:
Remaining is not chosen as the stock of victor or the vanquished, then in default of buying long/short sales signal without investment.
5. a kind of Portfolio Selection Based method based on depth attention network and intensified learning according to claim 4,
It is characterized in that, step S4 is specifically included:
Concentrate continuing investment sequence of the sampling one with T holding period from training data every time, respectively input history feature to
Amount, obtains the investment combination sequence at each moment according to model
The Sharpe Ratio of entire Investment Sequences and the reward function as intensified learning can be then calculated, the method risen using gradient
Model parameter is optimized, i.e. argmaxHT{B+, B-}；
Wherein,ATIt is the average return of the Investment Sequences each holding period, calculation formula isTCtTake for actual transaction procedure, VTIt is the fluctuation of Investment Sequences, uses the mark of income
Quasi- difference is measured, and calculation formula is
6. a kind of Portfolio Selection Based method based on depth attention network and intensified learning according to claim 5,
It is characterized in that, after step s4, further includes: explained using significance analysis method to model.
7. a kind of Portfolio Selection Based method based on depth attention network and intensified learning according to claim 6,
It is characterized in that, described include: to the specific steps that model explains using significance analysis method
Partial derivative is asked to the various features of input according to victor scoring, according to the size of partial derivative and positive and negative analyzes each
The influence that feature plays the decision of model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910390018.XA CN110223180A (en) | 2019-05-10 | 2019-05-10 | Portfolio Selection Based method based on depth attention network and intensified learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910390018.XA CN110223180A (en) | 2019-05-10 | 2019-05-10 | Portfolio Selection Based method based on depth attention network and intensified learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110223180A true CN110223180A (en) | 2019-09-10 |
Family
ID=67820847
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910390018.XA Pending CN110223180A (en) | 2019-05-10 | 2019-05-10 | Portfolio Selection Based method based on depth attention network and intensified learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110223180A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110909147A (en) * | 2019-12-02 | 2020-03-24 | 支付宝(杭州)信息技术有限公司 | Method and system for training sorting result selection model output standard question method |
| 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 |
| CN112925830A (en) * | 2021-03-15 | 2021-06-08 | 山东大学 | Investment-reinsurance decision method and system |
-
2019
- 2019-05-10 CN CN201910390018.XA patent/CN110223180A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110909147A (en) * | 2019-12-02 | 2020-03-24 | 支付宝(杭州)信息技术有限公司 | Method and system for training sorting result selection model output standard question method |
| 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 |
| CN112925830A (en) * | 2021-03-15 | 2021-06-08 | 山东大学 | Investment-reinsurance decision method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109615531A (en) | Securities market quantifies precisely returning for investment tactics and surveys and assessment system and method | |
| Brázdik et al. | Survey of research on financial sector modeling within DSGE models: What central banks can learn from it | |
| CN110223180A (en) | Portfolio Selection Based method based on depth attention network and intensified learning | |
| Yeyati et al. | Monetary and exchange rate policies | |
| Zheng et al. | Speculative behavior in a housing market: Boom and bust | |
| Smit et al. | Playing at acquisitions: Behavioral option games | |
| Radulescu et al. | Ranking of EU national banking systems using multi-criteria analysis in the light of Brexit | |
| Yeyati | Exchange rate policies and economic development | |
| LiPuma et al. | Currency devaluations and consolidating democracy: the example of the South African rand | |
| Bernstein | The Importance of Strong Labor Demand | |
| Dupor et al. | Do sticky prices need to be replaced with sticky information? | |
| Fisher | A SFC political business cycle: Kalecki’s 1943 model revisited | |
| Free et al. | Bourdieu's logic of practice theory: possibilities for research on management accounting and control | |
| Trivedi et al. | Effective trading in financial markets using technical analysis | |
| Wei | Optimal Combination of Currency Assets and Algorithm Simulation under Exchange Rate Risk | |
| Yang | A DSGE Decision Model for Investigating the LPR Transmission Effect | |
| Roche | Pragmatic Capitalism: What Every Investor Needs to Know about Money and Finance | |
| Tahri | Reinforcement learning approach for algorithmic trading of bitcoin | |
| Han et al. | Bitcoin or Gold? A Financial Investment Model Based on LSTM | |
| Pietrunti et al. | Monetary policy in times of debt | |
| Song et al. | Research of Investment Decision based on the Coupling Model of Metabolic AR and Neural Network | |
| Audzei | Essays on Macroeconomics with Financial Frictions | |
| Ragsdale | Animal Spirits: The Analysis of Fundamental Speculation. | |
| Arnold et al. | Interest rate parity in Excel | |
| Білоцерківський et al. | Methodological recommendations for delivering practical classes of the discipline" Basics of entrepreneurship" |
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: 20190910 |