CN110533540A - A kind of whole city multi items finance money guard system based on intelligence dimension Meta-Policy platform - Google Patents

Abstract

The present invention provides a kind of whole city multi items finance based on intelligence dimension Meta-Policy platform to provide guard system, comprising: transaction platform, data platform and Policy Platform.Setting is used to describe the dimension meta-rule in market in Policy Platform, the meta-rule of different dimensions and is combined with each other intersections, generation descriptive power；The tree structure that portrait is formed by tieing up first layout, by the way of layer-by-layer optimizing, the assessment of returning that each node carries out the whole city in tree structure is estimated；Portrait assesses optimization process step by step: as each node that assessment is the multilayer tree structure for being directed to the primary picture for tieing up member layout generation is assessed；One picture of each node on behalf, each as assess one scoring and this corresponding optimized parameter of scoring, the parameter of highest scoring is by the optimized parameter as picture.

Description

A kind of whole city multi items finance money guard system based on intelligence dimension Meta-Policy platform

Technical field

The invention belongs to financial information technology field, in particular to a kind of whole city based on intelligence dimension Meta-Policy platform is more Kind finance provides guard system.

Technical background

Currently, financial industry comes into the big asset management epoch, and while the scale of asset management reaches staggering amount, finance Barriers to competition between the same trade are broken, and are on the one hand that different financial institution is allowed to carry out the asset management business of homogeneity, separately One side is that the extension of original system business is constantly expanded, and intension is enriched constantly, the mutual business cooperation of financial class licence plate Property is remarkably reinforced.Asset management is typical knowledge-intensive, the intensive industry of the talent, and height, which relies on, has professional knowledge and energy The human capital of power.Financial information system business involves a wide range of knowledge, is strongly professional, structure is complicated, outside need get through different field, The barrier of different institutions, different value chain link internodes；Inside needs to get through Product Interface, business interface, between organizational interface Barrier；It in terms of product and service, needs to design competitive product, needs real to the product of different risks and client Apply differentiation price, scientific and effective cutting, recombination carried out to the risk of different type, attribute and feature, with realize risk and Proper Match between income.And the processing of magnanimity finance data, analysis and decision, be no longer simple manpower can and, Financial industry needs the information tool of more intelligent automation, reduces the dependence raising efficiency to manpower.Therefore, it is necessary to it is advanced, Safe and efficient information technology system and asset management overall process closely merge the innovation for leading asset management industry.

Summary of the invention

The purpose of the present invention is to provide a kind of whole city multi items finance based on intelligence dimension Meta-Policy platform to provide piping System, it is therefore an objective to which it can carry out the acquisition, processing and storage of financial market data as intelligent quantization transaction platform, tactful Production returns and surveys, optimizes and store, layout, deployment, operation, air control and the management of policy instance, and then obtains Object of Transaction, hands over Easy signal and outstanding net value curve realize the intelligence money pipe business of full dimension.

Technical solution is as follows:

A kind of whole city multi items finance money guard system based on intelligence dimension Meta-Policy platform characterized by comprising

Transaction platform is responsible for picking declaration form instruction, the execution of algorithm transaction and docking Shi Chang Bao's single function；

Data platform is responsible for the production of service market access and related derivative market；

Policy Platform is responsible for the tactful on-line operation of strategy production and high performance, including off-line strategy production management System and strategy of on-line operational management subsystem, Policy Platform calculate the money of each strategy in combination using machine learning algorithm The weight of gold proportion selects investment combination weight in each period, maximizes its ultimate yield；

Off-line strategy production management subsystem includes 3 submodules: strategy production module, for realizing Row control, report Announcement shows, list query function；Strategy, which returns, surveys optimization module, and the creation and operation, parallel meter of survey task are returned for realizing strategy Calculate task schedule, injecting strategy to time xylometer；Return xylometer module, operation reserve when be used to back survey, transmission market to it is tactful, Strategy calls altogether by fund and the calculating held position and offer strategy API when operation reserve；Strategy of on-line operational management subsystem fortune Between the departure date, factor computing module establishes network linking with xylometer is returned, and receives from the policy signal for returning xylometer, and according to letter Number calculative strategy evaluation points and scoring simple curve, after all policy instance operations finish, calculative strategy is in multiple product Comprehensive score resultant curve in kind；

Setting is used to describe the dimension meta-rule in market in Policy Platform, the meta-rule of different dimensions and is combined with each other intersection, Generate descriptive power；Dimension member is the atom logic that can not be subdivided in a segment description market, is under the jurisdiction of specific dimension meta-rule, has More or short side is to the different members under the same dimension meta-rule are mutual exclusions, all cities in all member aggregation cover time sequences Field state；Portrait is to be used to describe from multiple dimensions more accurately by the permutation and combination of multiple dimension member superposition different cycles Market, formula: as=dimension (member) [period] × dimension (member) [period] × ..., as generation from one or more dimensions (member) The permutation and combination in [period]；Policy Platform forms the tree structure of portrait, by the way of layer-by-layer optimizing, tree by tieing up first layout The assessment of returning that each node carries out the whole city in shape structure is estimated；Portrait assesses optimization process step by step: as assessment is for primary dimension Each node of the multilayer tree structure for the picture that first layout generates is assessed；One picture of each node on behalf, it is each as having assessed There are a scoring and this corresponding optimized parameter of scoring, the parameter of highest scoring is by the optimized parameter as picture.In order to subtract Few calculation amount, the corresponding optimized parameter of superior node can be genetic to next time as lower layer's picture can be on upper layer as the base of optimized parameter On plinth, parameter arrangement-cartesian product is carried out to the dimension member of this layer and realizes that blowing back assessment estimates；

The operation flow of Policy Platform includes: dimension member management: the dimension metacode uploaded being uploaded in system, and is supported Later maintenance；It ties up first layout: multiple dimension members being selected according to dimension, the period carries out fully intermeshing combination, generates picture；As assessment: being directed to Primary a batch for tieing up first layout generation is surveyed as carrying out back, is assessed, is screened, find qualified picture and parameter；Tactful son is single Member management: the code of tactful subelement is uploaded into system, and supports the maintenance in later period；Tactful layout: multiple strategies of selection Subelement carries out fully intermeshing, generates strategy；Policy Filtering: a batch strategy generated according to primary tactful layout is returned one by one It surveys and is screened according to specified screening rule, qualified strategy carries out in-stockroom operation as available strategy；Strategy is suitable Match: for the storage strategy by screening, carrying out parameter adaptation, find the corresponding best trade variety of actual parameter group；If It is set to online: setting qualified policy instance to online；Strategy is online: will shift onto online strategy from policy library On line, into firm offer or mimic panel operating status.

Preferably, different for the evaluation criteria of the picture of different levels, for one layer of picture, according to net profit, the second layer As being then unit time amount of loss.What is generated in order to prevent is effective as there is the problem of over-fitting, be in evaluation process using The resultant curve of full kind carries out platform, while parameter optimization also applies parameter Plain algorithm, assesses effectively as that can participate in To tactful layout.

Preferably, Policy Platform includes produce, optimize, factor and strategy-api module, modules Principle of interaction is that up-stream module calls downstream module to service using dobbo, and downstream module notifies up-stream module to use MQ mode； Produce module is responsible for being related to picture during as assessment as assessment mainstream process control, service inquiry, metadata management Return survey when, by dubbo call optimize come realize back survey；Optimize is responsible for picture and returning for strategy is surveyed, and optimizes, returns Task schedule is surveyed, returns and needs to call infrastructure to obtain container, injecting strategy and sending strategy by dubbo when surveying Initial order；Infrastructure is responsible for and tactful container interacts, including strategy of on-line container and off-line strategy container Management；Strategy-api is responsible for tactful online, the start-stop of strategy of on-line；From offline container when factor is responsible for back surveying Correlator accepts an order, and the calculative strategy factor and scoring, calculate complete by send MQ message informing optimize or Person's produce module.

Preferably, the picture evaluation process of Policy Platform: Manager is received as assessment task start request, and is passed through Dubbo is called, and passes to Produce module；Produce starting requires to call as assessment optimizing step by step, every layer of each picture Optimize survey；Optimize, which is received, surveys request as returning, and creation realizes that parameter is fried as returning survey task, by permutation and combination It opens, starts back survey task.

Preferably, it includes: to obtain available tactful container resource, injecting strategy to appearance that survey task is returned in the starting of Policy Platform Device sends enabled instruction to container；Container returns strategy in survey and enters in operation, can send hand position signal to Factor, Factor It receives order and starts calculative strategy performance；Factor calculating finish can send back after tactful performance survey completion notice to Optimize, Optimize can do task according to notice and end processing, including modification task shape body, discharge container resource.

Preferably, the tactful container asynchronous event driven process of Policy Platform: environment is surveyed if it is returning, then by container The monitor in portion directly reads local Sqlite file and sends market；If it is online container, then subscribed to from Data-Source Market；Asynchronous event transmitting: market/signal is all in a manner of asynchronous event in tactful container Internal Transfer；Strategy of on-line is supported Finish Bar, CurrentBar, and the market of Tick complete period and full kind are subscribed to and driving；Tactful internal reading historical quotes number According to using local memory to cache to reduce a large amount of I/O operations, tactful runing time is compressed to Microsecond grade；Strategy generates in container After signal, it is back to survey according to environmental variance judgement or online, if it is survey task is returned, then sends a signal to Factor mould The calculating that block carries out performance can be sent directly to transaction modules, be then forwarded to trade market if it is online container.

Preferably, tactful automated production process: an overall scheduler task is created in produce, receives the page Data are stored in mongo, then start to tie up first layout, after Mission Success, send MQ and notify to give produce module, if some Step executes failure, whole automatic building suspension of task.

Preferably, the function of increasing a row during tactful automated production will tie up first layout, as assessment, strategy Layout and Policy Filtering process are spliced into a process, simplify operating procedure.

Detailed description of the invention

Fig. 1 is the simple logic configuration diagram of muji system.

Fig. 2 is the physical structure deployment schematic diagram of muji system on-line operation part.

Fig. 3 is transaction declaration form flow chart.

Fig. 4 is to move storehouse to change a moon flow chart.

Fig. 5,6,7 are the declaration form of transaction micro services subsystem respectively, remove list and verification certificate flow chart.

Fig. 8 is the automatic monitoring flow chart of storehouse difference.

Fig. 9 is algorithm list process flow diagram.

Figure 10 is data platform hardware deployment architecture diagram.

Figure 11 is the functional flow diagram of data platform.

Figure 12 is the business process map of data platform.

Figure 13 is the building-block of logic of Policy Platform.

Figure 14 is the picture assessment timing diagram of Policy Platform.

Figure 15 is tactful container-asynchronous event driven flow chart.

Figure 16 is Policy Platform parallel computing task schedule schematic diagram.

Figure 17 is Policy Filtering flow chart.

Figure 18 is the automatic flow sheet of strategy.

Figure 19 is strategy production parallel computing architecture diagram.

Figure 20 is strategy operation linear expansion architecture diagram.

Specific embodiment

To make those skilled in the art more fully understand technical solution of the present invention, below with reference to embodiment to the present invention A kind of whole city multi items finance money guard system based on intelligence dimension Meta-Policy platform of offer is described in detail.Implement below Example is merely to illustrate the range of the present invention and is not intended to limit the present invention.

The whole city multi items finance money guard system based on intelligence dimension Meta-Policy platform in the present embodiment is named as the wooden chicken (muji) system.The whole city concept includes domestic market and overseas market, and multi items financial product includes but is not limited to foreign exchange, phase The transaction of goods, stock, bond and digital cash product.

As shown in fig. 1, wooden chicken (muji) system mainly includes three parts: transaction platform, is responsible for picking declaration form instruction, calculate The execution and docking Shi Chang Bao's single function of method transaction；Data platform is responsible for the production of service market access and related derivative market； Policy Platform is responsible for the tactful on-line operation of strategy production and high performance.As shown in Fig. 2, entire wood chicken on-line operation link Using CEP Event processing engine, the circulation that core business is carried out on Apama container is built.Data platform provides market push Service and the inquiry of fundamental data/historical quotes function；Tactful online container reception market data, strategy is based on row Facts part carries out the execution of strategic thinking, generates buying signals for signal and issues transaction container；Transaction container receives transaction letter Number, it will do it a series of logical process, such as mother Dan Shengcheng, algorithm is singulated, and order report is finally entered market；It strikes a bargain in market Afterwards, return event can be pushed to trading channel, transaction container can be entered by then returning event, carry out the maintenance of order data；Most Calculating relevant with fund of holding position is completed afterwards.Table 1 shows that the exploitation environment and running environment of wooden chicken (muji) system.

Table 1

I. transaction platform: transaction platform assumes responsibility for buying signals processing in the wooden chicken system, and algorithm is singulated, docks market, Fund calculates and the multiple player of risk control.The online transaction part of the entire wooden chicken system of support, completes algorithm transaction, wind Danger control, the core function of market route distribution.High safety: the safety of fund, safe, the tactful safety of data, system Safety；High Availabitity: Single Point of Faliure, the isolation for realizing account, the isolation of strategy are avoided；High-performance: performance, the data of data acquisition Performance, the performance of strategy execution, the high concurrent performance of processing；High precision: the accuracy of data processing, buying signals it is accurate Property, the accuracy of capital management；High stable: system long-play fault-free passes through perfect fault restoration machine even if having System, can very short time recovery.Support the whole city: the financial target online transaction of support domestic market, overseas market；Full kind: Support the financial transaction target of the how flat classes such as stock, futures, option, bond, foreign exchange, fund；It is full-automatic: to be based on artificial intelligence skill Art completes All Activity process in such a way that sequencing transaction and algorithm transaction combine；Wholly-owned gold: it can freely be adapted to any The product of capital quantity size exports outstanding net value curve, realizes the target of money pipe value-added service.Table 2 shows that transaction is flat The module of platform.

Table 2

Transaction noun:

Policy instance: policy instance is the specific execution of tactful (strategy combination), is substantially tactful subelement example Diversification combination, it includes to be launched, three kinds of operation, termination states.

Transaction unit: overall process of the simple transaction kind from opening a position hole capital after selling all securities in strategy (strategy combination) example, A referred to as transaction unit, centre include the operation such as adding storehouse, selling shares.

It moves storehouse and changes the moon: after futures main force contract switches over, holding position for the contract firm offer in the wooden chicken system, carry out old conjunction The process to open a position closed a position with new contract about.

Tactful storehouse: the position that policy instance theory is held is generated according to policy signal is accumulative.

Firm offer storehouse: the practical position held of current strategies example, according to sub single transation report-back cumulative calculation.

Storehouse is poor: the difference in tactful storehouse and firm offer storehouse, poor=tactful storehouse-firm offer storehouse in storehouse；Storehouse difference is timing, and the commission of sending refers to It enables consistent with policy signal command direction；When storehouse difference is negative, the entrusting commands of sending are opposite with policy signal command direction.

Algorithm unijunction beam: refer to female list in completion status, all sub- Dan Jun have return (comprising striking a bargain, removing single success or refuse It is single).

Automatic aligning: referring to the process that buying signals execute, and the position in firm offer storehouse and tactful storehouse are carried out position in storehouse alignment automatically.

Non-automatic alignment: it is corresponding with automatic aligning task, it is different from manual alignment.Refer to that carrying out automatic commission using algorithm hands over A kind of easily mode of doing business without position in storehouse alignment.

Manual alignment: being different from automated transaction (being aligned comprising automatic aligning with non-automatic), refers to manpower intervention using manual The form of intervention makes tactful storehouse and a kind of consistent trading activity in firm offer storehouse.Manual alignment function is under policy instance operating status (including in operation and paused), effectively.

It chases after valence list: after order ticket does not strike a bargain for a period of time, order ticket being carried out to remove list, weight is carried out using new commission valence New declaration form (chasing after list), chases after list to valence according to newly reporting the price into order ticket that can be divided into again, extension valence chases after list, and newest valence chases after list.Have When opponent's valence is also referred to as to valence.

List is chased after to valence: if to buy in, valence being chased after and single sell monovalence with regard to using and buys in.

Extension valence chases after list: if to be bought in, extension valence chases after single to be bought monovalence and is bought in regard to using.

Newest valence chases after list: newest valence chases after list and is bought in or sold using newest valence (LastPrice).

Limit order: being entered with limit order report as defined in exchange, and order ticket needs specified commission price.

Market order: it reports so that market price as defined in exchange is specified into order ticket does not need specified commission price, direct bearing of trading System can automatically participate in business the circuit breaker price that market order is converted to commission direction.Limit-up valence can be converted by buying market order It buys in, sells market order and can be converted into limit down valence and sell.

FOK: otherwise all strike a bargain revocation automatically immediately, and referring to must all strike a bargain in specified price, commission quantity, otherwise certainly It moves by System undoes.As a result it is or all strikes a bargain or all cancel.

FAK: conclusion of the business volume residual cancels automatically immediately, refers to and strikes a bargain in specified price, remaining order is automatically by System undoes.

The day of trade: domestic futures currently follow daily settlement regulation, except it is middle gold institute in addition to be all afternoon 15:00 close, in Golden institute 15:15 closes, and is settled accounts after closing.Refer to the exchange hour period for belonging to same lump-sum settlement the day of trade, usually The firm bargain day at night said and consecutive days are not identical.Such as the night disk in May 30 (Thursday) evening, the day of trade that it belongs to should For May 31.

Modern storehouse: what current trading day opened a position holds position.Yesterday storehouse: what the non-present day of trade opened a position holds position.Flat the present: it opened a position to the same day Hold position and close a position.

Flat yesterday: closing a position to holding position of opening a position of the non-same day, the big quotient institute of current country's futures and quotient Zheng institute, the area Zhong Jinsuobu Point modern storehouse, yesterday storehouse can be used to close a position and flatten all hold position.Last institute and the present Cang Zuocang can be above distinguished, it is necessary to storehouse up to the present It just can be carried out and close a position using flat the present.

Insure type: type of insuring is divided into hedges (referred to as set is protected), arbitrage and speculates, and usually our transaction all belongs to In congenial；It hedges and is mainly used for some enterprise production and managements；Arbitrage is mainly used for the arbitrage deal strategy of some mechanisms, has A little exchanges' (such as middle gold) can relax maximum to arbitrage deal person and remove these singular regulatory requirements.

Contract price minimum change unit: contract price minimum change unit refers to the points of each price minimum change, For example corn minimum change unit is 1, coke minimum change unit is that 0.5, IF minimum change unit is 0.2.

Hop count: for minimum change unit come, such as corn 5 jump corresponding price be 5*1=5；Coke 5 is jumped corresponding Price is 5*0.5=2.5；It is 1 that IF5, which jumps corresponding price,.

Guarantee fund: on forward market, dealer need to only be paid a small amount of fund by the certain ratio of futures contract price and be made It is assured in fulfilment of the financial resources of futures contract, the dealing of futures contract can be participated in, this fund is exactly futures margin.

Freeze guarantee fund: the order ticket that opens a position is quoted when not striking a bargain also, and the guarantee fund collected, which cries, freezes guarantee fund.

As shown in Figure 3, in transaction platform trade container transaction declaration form process are as follows: S1: buying signals reach transaction hold After device, is handled by signal management module logic and realize that signal initialization is associated with tactful storehouse；S2: signal is from signal management module Out after team, the processing of order management modular filtration can be entered and generate female list；S3: female single algoritic module that enters is routed by algorithm, first By the flat modern instruction of preposition processing logical process, then it is routed to corresponding singulated algorithm and carries out singulated, all hand numbers are torn open For the signal of the raw singulated end of female per unit area yield to algorithm postposition module after the completion of single；S4: the sub- Dan Zhi that singulated process generates Sending and receiving are sent to fund correction verification module, carry out wind and raise the check logic with fund, verification can be sent to market routing, otherwise provide after passing through List is refused in gold verification；S5: transaction routing module default transmit a request to trading channel gateway by java card format, by routing It is sent to corresponding trade market；S6: the return sent back to from market is first received by the client built in trading channel, then calls and holds Return is sent back to transaction container by the infrastructure service of device management, and the return quilt list status monitoring device into transaction container receives and touches Circulation of sending out state corresponding, carries out transation report-back and falls library, removes list/refuse single state processing, last triggering fund calculates and mother is single State promote.

It is traded in the transaction declaration form process S2 of container in transaction platform, order management module carries out three layer filtration: processing master The corresponding main force's contract of signal acquisition of power mapping Types Below；Position in storehouse alignment, handles position in storehouse automatic aligning or non-automatic alignment Logic correct declaration form hand number；The secondary-confirmation of strong flat type reinitializes declaration form hand if be not inconsistent with firm offer position in storehouse Number ultimately produces female single.It is traded after the transaction declaration form process S6 of container in transaction platform, transation report-back triggers carry out fund meter It calculates: successively being calculated by way of java plug-in unit, the fund of the dimension of holding position and hold position of policy instance dimension, followed by plan before this The slightly fund of example dimension, is finally the fund of account dimension.In transaction platform after the transaction declaration form process S6 of transaction container, After the female single respective action of triggering, female single state can promote therewith, if female unijunction beam, by result feedback to signal management mould The propulsion of block completion signal condition.

In transaction platform trade container reception to buying signals container signal management module storage after be initially INIT；Each buying signals after storage are in a logic query, and the logical primary key of queue is stratInstId+ symbol+positionDirection；The signal of nontransaction time or Gao You can be advanced directly into part signal after joining the team UNDO state；It is merged if multiple identical signal types are in INIT, the signal being merged is advanced to MERGED；Signal Normally go out team to execute, state circulation waits algorithm to execute to WAITING；Signal enter algorithm execute it is singulated, state becomes RUNNING；Signal in RUNNING state is circulated by after the other signal terminating of high priority as UNDO；Abnormal conditions promote To UNKNOW；Signal, which is finished, under normal circumstances becomes DONE.

As shown in figure 4, a moon process is changed in the shifting storehouse of transaction platform are as follows: S1: transaction container listens to transaction timed task system A moon signal S2 is changed in the shifting storehouse of sending: if there is the execution signal of low priority, being moved storehouse and is changed moon signal and can terminate priority and be lower than Signal in the execution of itself executes into algorithm after then receiving termination return, is otherwise directly entered algorithm.S3: enter algorithm After transaction, start to carry out algorithm calculating, if putting down old hand number is 0, direct more new strategy storehouse is new main force's contract；If put down old Hand number is greater than 0, then according to dealing amount and price, sends declaration form, and establish the monitoring returned current order.The shifting of transaction platform Storehouse is changed after moon process S3, if returned successfully, carries out the update of conclusion of the business record；If return failure, transmission, which is removed, singly asks It asks；If list is refused in return, air control alarm monitoring mechanism is triggered, sends alarm signal.

Fig. 5,6,7 are the declaration form of transaction micro services subsystem respectively, remove list and verification certificate flow chart.Transaction platform includes transaction Declaration form process are as follows: S1: micro services subsystem has the client dynamic routing of signal will in the market routing module of transaction container Request is routed to corresponding trading channel routing gateway；S2: passway will be asked by receiving the secondary business routing of declaration form request progress It asks and is forwarded to the channel CTP or the channel IB；S3: after corresponding channel receives request, first check whether channel is established, if without if It initializes and establishes channel；S4: it checks the logging state of user, is then stepped on again if not logged in；S5: turning for parameter is carried out It changes, be transformed into the parameter that corresponding market needs and declaration form request is changed to redis triggering is asynchronous to fall library；S6: finally request is sent out Past market and the subsequent return for waiting market.After the declaration form process S6 of transaction micro services subsystem, after receiving market reward, first Order information is obtained from redis, then the corresponding event of building transaction container, calls the infrastructure service of Container Management that will return It is sent to transaction container.

The single operation of removing of transaction micro services subsystem includes that the triggering that do not strike a bargain of transaction container removes list or the page cancels manually Periodically remove list after manual list or disk, remove single process are as follows: S1:OTC is received remove single request after, can be according to corresponding brokerID Forward requests to the corresponding channel CTP or the channel IB；S2: it carries out removing forms data encapsulation after carrying out parameter verification；S3: it checks The logging state of user is then stepped on again if not logged in；S4: it after obtaining Account Logon connection, directly transmits to remove and singly ask It asks；S5: synchronous collection carries out succeeding state change to after removing single return.

Transaction micro services subsystem bill time-out removes single time-out, triggers verification certificate business, verification certificate process are as follows: S1: first After OTC receives verification certificate request, the channel CTP or the channel IB are forwarded requests to according to corresponding order number；S2: it is neutralized from caching Verification certificate is assembled in database requests necessary parameter；S3: and then check user logging state, if not logged in, then into Row is stepped on again；S4: query result is returned.

The pause and recovery of the policy instance of transaction micro services subsystem can carry out interface operation by backstage manager.Such as Fruit is pause function: the chicken management platform wooden first initiates suspending event request, via transaction API service to tactful example state into Row verification, verification is by that can be sent to transaction container by basic API service.Whether container of trading can select automatic according to front end It closes a position function, carries out corresponding business execution: if having selected Self Trimming, flat processing by force can be initiated.If do not selected Self Trimming then directly suspends policy instance.If it is recovery function: the chicken management platform wooden first initiates to restore event request, By transaction API service using basic API service, finally it is sent to container starting shifting storehouse and changes the operation such as moon.

The position in storehouse policer operation of transaction micro services subsystem: it is poor that storehouse is obtained by acquisition strategy storehouse data；It can by obtaining Buying signals are sent to be aligned position in storehouse with transaction container concurrent；By not striking a bargain from the acquisition of temporally range or partially being obtained at jiao zi list Storehouse difference reason.

The air control system of transaction platform, which monitors in real time in each policy instance signal implementation procedure, there is supervision indication range Four kinds of outer risk situations: having live signal, and per minute, per hour, daily four seed type: the split warehouse receipt of air control system carries out school It tests, if the abnormal behaviour supervision index for the example belonging to list that opens a position reaches intervention, refuses declaration form；In disk 2 minutes timed tasks and The wind of transation report-back meeting trigger policy example and account raises Risk Calculation interface, calculates wind and raises risk class, if policy instance grade Other wind raises risk class and reaches intervention, then trade-api interface can be called to suspend the policy instance and operation of closing a position；It hands over It whether sufficient calls air control plug-in unit calculative strategy example dimension that can be raised with wind before easy container declaration form, list is directly refused if inadequate； Timed task systematical solutions omit example and account dimension risk indicator monitoring table is scanned, and reach to policy instance and account dimension Risk item to prompting, early warning and intervention levels is alarmed；Air control system to removing, return and transation report-back carries out plan by single commission Slightly the early warning value of example and account dimension is cumulative, can trigger alarm after reaching warning level.

Transaction platform includes timed task subsystem, and 8: 35 can all do a same day knot before timed task disk every night It calculates, mainly count each funds data and ensures next day of trade arm's length dealing.As it is possible that some day, sales counter does not have The day statement of account is given, reconciliation can not be carried out, so the same day can tie failure day, but rear extended meeting is automatically replenished day knot by this task, and just Really execute completion.Process are as follows: S1: in order to guarantee that the day performance of knot does fragment, fragment is carried out for account, carries out day as entrance Knot；S2: the policy instance of all non-INIT states ties condition as day under inquiry account；S3: form function with the policy instance day Hold position record and the conclusion of the business on the same day record of the previous day, does one and brings discs together, obtain hold position data and policy instance fund number According to；S4: and then the fund of discs, update of holding position are calculated every fund of account and updated to database, by all funds It calculates after completing, records corresponding fund snapshot；S5: it is clear finally to carry out day.By service charge, the complete liquidations such as profit and loss of closing a position, Guarantee the normal of next day of trade fund.

Muji system puts the routing forwarding of row world market sales counter by unified network of channels.The password of fund account is adopted Database is stored with the mode of symmetric cryptography.The online transaction process monitoring of transaction platform include trade-server monitoring, Trade-schedule monitoring, trade-risk-server monitoring, trade-risk-server monitoring, futures exchange container, The online container of futures, foreign exchange transaction container, the online container of foreign exchange etc..Transaction platform uses the asynchronous log printing type of log4j2, Log monitoring alarm content includes ctp client control (ctp interface time-out and system exception), order management alarm, algorithm friendship It easily alerts, snapshot data alarm, transaction unit processing exception, database monitoring, user is not landed, all positive normal messages are logical Know, had reached from conclusion of the business number, frequently report removes that single number has reached, wholesale report removes that single number has reached, wind is raised and had reached.

II. data platform: as shown in figure 11, real-time typing world market market data pass through Stream Processing, synthesize K line And index, present quotation is pushed in real time, database service interface is provided, and carries out historical quotes data cleansing and processing.

As shown in Figure 10, data platform realizes the process of multichannel CTP market market access processing are as follows: accesses base from data source Plinth market data, and receive client and use different network lines, guarantee the stabilization of data to greatest extent；Each client is logical Netty is crossed to send received market in market processing service；Monitor whether that each server-side there are market, if there is one Or multiple client does not have market, issues warning information；The basic side information that monitoring service end issues judges whether unanimously, no Warning information is issued if consistent；Corresponding processing, such as processing client application problem are done for the warning information provided, more It is preposition etc. to change futures company.For from the received market data of different clients according to time priority and exchange hand preferential principle Carry out market select it is excellent, the data platform using select it is excellent after market data do subsequent market and process.

The access of IB market: two IB Gateway are enabled and obtain market data, two ib-client receive two respectively The market data of gateway carry out market screening in market processing service, and the market for selecting main market source carry out subsequent K line The processing such as synthesis.IB market mostly living monitor process: timed task is 1 time per second, considers that the time for avoiding just having opened the set and application are rigid The time of starting is alerted, alarm controls (five points of frequency when main market source or standby market source are more than not have market in 5 seconds Clock only sends out primary alarm)；There is exception if it is main market source, standby market source is normal, needs to switch market source, by the master in caching Market source is set as the coding in standby market source, and main market source code is stored in caching (redis) after switching.

Real-time K line synthesis process: receive to select it is excellent after tick market, the checkout transaction time, in the case where exchange hour just carries out The step of face, sends tick data to mq, obtains the tick of a upper root cache according to whether being market data handled by quotient Zheng, The K line set in caching is obtained, according to low frequency K line periods synthesizer low frequency K line, according to high frequency K line periods synthesizer high frequency K line, The asynchronous storage of Tick data updates redis and caches K line set.Real-time K line dicing process: timed task executes processing per minute, Contract information is obtained, according to whether carrying out the K line in cutting caching to the K line cutting time, traverses the contract of subscription, multithreading executes It is subsequent to cut K sequence of threads, contract time bracket is obtained, the corresponding K line set caching of contract is obtained, calculates accumulative exchange hour Number traverses the K line period, decides whether to cut K line, cut K line if necessary, just the K line in caching is backed up out, be emptied slow It deposits, the asynchronous storage of K line, K line is pushed to tactful container.

Market subscription procedure: relationship is pushed using the subscription that td_subscribe_relation table stores contract and container； Initialization subscribes to contract and container ip, the relationship of port from newly-built middle load；It is monitored after starting by mq and subscribes to message；It receives Subscribing relationship (including subscribe to and unsubscribe) is safeguarded in memory after subscribing to message；Subscribing relationship is persisted to data Library.Market push process: judging whether to push processInBar when synthesis K line；Cutting K line judges whether what push was completed bar；If bar needs to push, bar is pushed to by tcp connection by target container according to subscribing relationship.

Index synthesis process: inquiry contract species data；Tick market are received from mq, judge whether the first stroke tick, It is the timestamp for timing time t1 being set as if the first stroke tick；Compare if not being the first stroke with timing time t1；Difference Value is greater than 500ms, then generates index tick according to the tick data of caching；Tick data are cached, are closed according to index and about tick At K line, K line data loading, K line data-pushing Policy Platform.

The process of data platform world market market working process are as follows: from the different each cities in the data supplier access whole world The basic market data of field；By basic market data-pushing to mq cluster, basic market data landing；Each market working process Using from the market data after mq cluster subscription base market data and processing, market processing, such as the synthesis of kind index are carried out, Across the phase, across kind arbitrage market processing, cross-market derivative data processing etc.；Market data-pushing after processing adds to mq cluster Market data landing after work.

Derivative data embodies the changing rule between two kinds or index or the framework contract quotation of multi items, derivative data Processing purpose is that market support is provided for arbitrage deal, and mainly include three kinds of forms: across phase arbitrage derivative data: reflection is same The variation of difference in selling prices between the same commodity in market different delivery times；Across kind arbitrage derivative data: reflect same city Difference in selling prices between different range of goods variation (generally refer to the kind with correlation, such as corn and cornstarch, The combination such as soybean, soya-bean oil and dregs of beans)；Cross-market arbitrage derivative data: reflect same range of goods in the price difference in different markets Away from variation, the data platform derivative data processing flow are as follows: creation derivative data definition, derivative data definition storage, meter Calculate public time bracket, in mq cluster subscribe to derivative data define in include framework contract, obtain framework contract tick Market, according to derivative data define in calculation formula using regulation engine Aviator calculate price price difference, generate derivative data Tick, derivative data tick close K line, and derivative data tick cuts K line, and push derivative data K line is to tactful container to drive strategy Operation.

Historical data cleaning process: load cleaning target after starting inquires kind, contract data；It is cleaned by the day of trade Tick data；K line number evidence is cleaned by the day of trade, journal file is written into abnormal quotation information.Historical data process: load Kind, contract information；Load with about tick, generates index tick substantially；High frequency index K line and low frequency are generated using index tick Index K line；K line storage.

Data platform provides other platform data query services by dubbo interface.High concurrent is efficiently handled in real time, Gao Ke With guaranteeing that system is stable, reasonable data storage organization and data interaction design guarantee data processing performance and accuracy.

III. Policy Platform: as shown in Figures 12 and 13, muji-manager is the portal of the entire wooden chicken system, before all End access request is all forwarded to the service on backstage via muji-manager.Policy Platform include two major subsystems: 1, from Line strategy production management；2, strategy of on-line operational management.Off-line strategy production includes mainly 3 submodules: 1, strategy production Module (major function be Row control, report show, the functions such as list query)；2, strategy, which returns, surveys optimization module (major function It is creation and operation, parallel computation task schedule, the injecting strategy to time xylometer that strategy returns survey task)；3, xylometer is returned (to return Operation reserve when survey, fund and the calculating held position and to provide strategy API total tactful when sending market to strategy, operation reserve It calls.During strategy of on-line is run, factor computing module establishes network linking with xylometer is returned, and receives from the plan for returning xylometer Slightly signal, and according to signal calculative strategy evaluation points and scoring (simple curve), after all policy instance operations finish, Comprehensive score (resultant curve) of the calculative strategy in multiple kinds.

Setting dimension meta-rule in Policy Platform: giving a definition series of rules in some dimension, for describing market, such as The indexs such as MACD, MA, RSI are exactly set of rule, the meta-rule of different dimensions and are combined with each other intersections, so that generation is preferably retouched State ability.Dimension member: simply dimension member is exactly the atom logic (can not subdivide) in a segment description market, is under the jurisdiction of specific dimension member Rule has more/short side to the different members under the same dimension meta-rule are mutual exclusions, in all member aggregation cover time sequences All state of market.Portrait: pass through the permutation and combination of multiple dimension member superposition different cycles.For more accurately from multiple dimensions Degree description market, formula: as=dimension (member) [period] × dimension (member) [period] × ..., as generation tieed up from one or more The permutation and combination in (member) [period].

Policy Platform includes multiple tactful subelements: portrait subelement, the subelement that opens a position, subelement of closing a position, capital management Subelement, air control subelement etc., every class subelement can only form strategy with inhomogeneous subelement, by multiple tactful subelements Carry out permutation and combination, form multiple strategies, formula: portrait subelement × open a position subelement × is closed a position subelement × capital management Unit × air control subelement.

The operation flow of Policy Platform includes: dimension member management: the dimension metacode uploaded being uploaded in system, and is supported Later maintenance；It ties up first layout: multiple dimension members being selected according to dimension, the period carries out fully intermeshing combination, generates picture；As assessment: being directed to Primary a batch for tieing up first layout generation is surveyed as carrying out back, is assessed, is screened, find qualified picture and parameter；Tactful son is single Member management: the code of tactful subelement is uploaded into system, and supports the maintenance in later period；Tactful layout: multiple strategies of selection Subelement carries out fully intermeshing, generates strategy；Policy Filtering: a batch strategy generated according to primary tactful layout is returned one by one It surveys and is screened according to specified screening rule, qualified strategy carries out in-stockroom operation as available strategy；Strategy is suitable Match: for the storage strategy by screening, carrying out parameter adaptation, find the corresponding best trade variety of actual parameter group；If It is set to online: setting qualified policy instance (strategy+parameter+kind) to online；Strategy is online: will be to online Strategy is shifted on line from policy library, into firm offer or mimic panel operating status.

Policy Platform mainly includes produce/optimize/factor/muji-strategy-api.In order to avoid following Ring relies on, and modules principle of interaction is that up-stream module calls downstream module to service using dobbo, and downstream module notifies upstream mould Block uses MQ mode.Produce module is responsible for as assessment mainstream process control, service inquiry, metadata management, in the mistake as assessment It is related to when returning survey of picture in journey, optimize can be called to survey to realize back by dubbo；Optimize is responsible for picture and strategy Return survey, optimization, return survey task schedule, return survey when need by dubbo call infrastructure obtain container, injection Strategy and sending strategy initial order；Infrastructure is responsible for and tactful container interacts, including strategy of on-line container With the management of off-line strategy container；Muji-strategy-api: it is responsible for online, the start-stop of strategy of on-line of strategy；Factor: negative It accepts an order when blaming back survey from offline container correlator, and the calculative strategy factor and scoring, calculates and complete to pass through transmission MQ message informing optimize or produce module.

It returns xylometer cluster: returning xylometer and concurrently run back survey task, resource consumption is big, using cluster mode, cluster packet Containing 6 machines, one container process of each node deployment, support level extension.

Factor computing cluster: occupancy resource is more, single machine multiple instances deployment, 4 services of one of deployment, another portion 2 services are affixed one's name to, service number and time xylometer quantity are 1:1.

Strategy of on-line container cluster: operation strategy of on-line supports the more containers of single node, support level extension.

As evaluation process:

Manager is received as assessment task start request, and is called by dubbo, and Produce module is passed to；

As assessment optimizing step by step, every layer of each picture requires that optimize is called survey for Produce starting；

Optimize, which is received, surveys request as returning, and creation blows (permutation and combination) as returning survey task, parameter, starts back survey and appoints Business；

Starting back survey task includes: to obtain available tactful container resource, injecting strategy to container, transmission enabled instruction to Container；Container returns survey in strategy enter in operation, hand position signal can be sent receive order to Factor, Factor and start to calculate Tactful performance；Factor calculating can send back survey completion notice to Optimize after finishing tactful performance, and Optimize can root Task is done according to notice to end processing, including modification task shape body, discharge container resource.

As shown in figure 15, tactful container-asynchronous event driven process: different market signals is used for different environment Driving method: surveying environment if it is returning, then directly reads local Sqlite file by the monitor inside container and send row Feelings then subscribe to market from Data-Source if it is online container；Asynchronous event transmitting: market/signal is all with asynchronous thing Part mode is in tactful container Internal Transfer；Strategy of on-line supports the Bar that finishes, CurrentBar, Tick complete period/full kind row Feelings are subscribed to and driving；The tactful internal historical quotes data that read are cached using local memory, reduce a large amount of I/O operations, strategy is transported The row time is compressed to Microsecond grade；In container strategy generate signal after, can according to environmental variance judgement be back survey or online, such as Fruit is back survey task, then the calculating for sending a signal to Factor module progress performance can be directly transmitted if it is online container To transaction modules, it is then forwarded to trade market.

The parallel computation task schedule process of policy module: the primary request that user submits can correspond to father's task, father Task will not be directly entered task waiting list, but first split into the relatively uniform atomic task of granularity, into waiting list, It waits scheduled；Task waiting list is sorted according to task priority set by user, the high close Head of priority, first quilt Operation, equal priority sort according to chronological order；Scheduler module circulation obtains Head corresponding of waiting list Business, then obtains available computing resource, if there is available idling-resource, head task can all be moved to running It is engaged in list, into operating status；Running finishing for task can remove from running task list.

Especially, the operation and resource allocation of the task scheduling modules management atomic task of optimize subsystem, task Tree structure can be constructed during creation, the task that root node on behalf user can see, leaf node is atom Task；Atomic task is according to priority ranking, and into task waiting list, task scheduling modules can be periodically from waiting list head Poll allocates resources to the task if it find that there is available computing resource, and head task is moved on in operation list, is opened Dynamic task, which is returned, surveys operation；Optimize receives the notice from Factor subsystem, starts the process flow that task is completed, such as Fruit is successfully to terminate, then discharges resource, and task schedule is notified to start next task, and in addition there are also some business processing streams Journey；If mission failure, resource can be also discharged, while retry 3 times to task, all be had failed if retrying 3 times, in industry Think that task really fails in business, then can just enter failure handling process, the failure of atomic task will affect the shape of father's task State, the state of the state of the root node including Task Tree and fraternal task.

Policy module forms the tree structure of portrait by tieing up first layout, by the way of layer-by-layer optimizing, in tree structure The assessment of returning that each node carries out the whole city is estimated.Portrait assesses optimization process step by step: being to tie up member layout production for primary as assessing Each node of the multilayer tree structure of raw picture is assessed；One picture of each node on behalf, it is each as having assessed one Scoring and this corresponding optimized parameter that scores, only picture of the scoring greater than 0.6 is considered as effective picture, score in business Highest parameter is by the optimized parameter as picture.In order to reduce calculation amount, the corresponding optimized parameter of superior node (as) can heredity To next time as lower layer's picture can be blown (parameter arrangement-to the dimension member of this layer on basis of the upper layer as optimized parameter Cartesian product) it returns to test and assess and estimate.It is different for the evaluation criteria of the picture of different levels, for one layer of picture, according to net profit, Two layers of picture are then unit time amounts of loss.What is generated in order to prevent is effective as there is the problem of over-fitting, is to use in evaluation process Be that the resultant curve of full kind carries out platform, while parameter optimization also applies parameter Plain algorithm, assesses effectively as meeting Participate in tactful layout.

Clustering performance prioritization scheme technical effect:

By task scheduling engine and resource pool management module, dynamic manages and distribution resource, and one big task is split At small time survey task, multiple small time survey tasks are run simultaneously on different machines.It is carried out using the cluster of 6 server compositions Parallel return is surveyed, and cluster throughput is 185.53 ten thousand bar/ seconds.With six annual datas, member is tieed up using double equal lines, the period is 15 minutes, only There is one layer of picture, each picture blows 492 groups of parameters, and full kind picture assessment is carried out on 6 servers, and 28536 examples are average to consume When be 7.2 minutes, by the resulting estimate, produced with 6 annual datas, can return daily survey 570.72 ten thousand strategies.

Optimizing algorithm solution technique effect step by step:

Tie up first layout by way of permutation and combination by the first layouts of multiple dimensions at multistage tree structure, since the root set It assesses step by step, finds optimized parameter, and optimized parameter is passed into next stage picture.Such benefit is exactly to reduce a large amount of invalid fortune Calculation saves computing resource by taking double lines in 3 periods as an example, and each node is 100 groups of parameters, into after crossing arrangement, first layer It is 2 nodes, the second layer is 4 nodes, and third layer is 8 nodes.

Grade evolutionary operation amount: 100*2+ (100*100*4)+(100*100*100*8)=8040200 is mixed using fully intermeshing；

Using evolutionary operation amount step by step: 100*2+100*4+100*8=1400；

Comparatively, fully intermeshing mixes more than 5700 times that grade evolutionary operation amount is optimizing step by step, using system after optimizing step by step Operand only has the 1/5700 of fully intermeshing, and it is obvious to promote effect.

Poor fitting solves: poor fitting refers to that the sample data volume of general time survey is too small, causes the strategy of fitting can not It meets the requirements, error is larger, and solution is big time range here, and large sample, which returns, to be surveyed, and returns at present for domestic forward market The time range of survey is 2012-01-01 to 2019-3-31 years, nearly 7 years data；Another method is for the whole city Survey, the picture assessment of such as domestic futures and Policy Filtering, preferable 29 kinds of selecting liquidity survey, thus It can effectively avoid the problem that poor fitting.

The solution of over-fitting: over-fitting refers to tactful overfitting historical data, without having generalization ability, solution Certainly scheme be as assessment during restriction strategy picture the maximum number of plies only to 3 grades, while use Plain parameter, limit parameter Isolated island, in addition there are also the interior separated test with outside sample of sample is used, Training strategy in sample, it is to verify sample that sample returns survey outside The validity of the strategy of interior generation.

Parameter Plain algorithm: target elements are determined: from current existing evaluation points, determines a factor as Plain The basic evaluation factor (such as: net profit, year earning rate) of algorithm；Determine neighbour's length: newly-increased, neighbour's length is positive whole Number.For determining the parameter group range in building parameter Plain, default value 2, business personnel can be modified.If some parameter group has M A parameter finds neighbour's parameter group (+N-N step-lengths, when taking neighbour's parameter of the parameter, in parameter group of wherein every 1 parameter Other parameters first immobilize), every 1 parameter have 2N neighbour's parameter group and target elements value (do not include this parameter from Body), there will be M*2N neighbour's parameter groups for M parameter, finally add parameter itself, then share M*2N+1 parameter group and correspondence Target elements value.Such as: in double equal line strategies, corresponding 50 parameters of Fast parameter group (1-50), step-length 1；Slow parameter Group (0-200) corresponds to 21 parameters, step-length 10；Neighbour's length N=2 is set, if selected parameter group Fast=6, Slow=40 When, then parameter group value range be (6,20), (6,30), (6,40), (6,50), (6,60), (4,40), (5,40), (7,40), (8,40)。

Multilevel report form is used in tactful production process: due to producing a large amount of centre during strategy production Ephemeral data occupies a large amount of memory space, in order to save the consumption of storage resource, according to different business needs, using not The report of same level, background program is according to the rank storing data of report.Storage and calculation amount sequence: simple report > summarize report Accuse > scoring report；For the low report of rank, if user needs to see more data, function is unfolded using report, can see To high level data reporting.

Tactful comprehensive score formula: f=n+m/100；

Wherein, n=Nian Huawu lever income/maximum, which is withdrawn, (represents profitability, under risk level, income gets over Gao Ying Sharp ability is stronger) m=(PR-Q)/R (representing validity, P- winning rate, R- profit and loss ratio, Q=1-P), comprehensive score f > 0.6 point, i.e., It is considered available strategy.

The polyfactorial calculating of stock:

Net value performance:

Yearization earning rate: (netValue (end) -1) ^ (250/length (netValue)), netValue (end) is net Value sequence end value, length (netValue) are net value sequence length；

Card Ma ratio: year earning rate/history maximum withdraw；

Sharpe Ratio: (year earning rate -0.03)/year stability bandwidth；Yearization stability bandwidth=std (r_i) * √ 250.

IC (information coefficient): IC (t)=corr (F (t), R (t+1)), current factor value sequence F (t) and next period earning rate The related coefficient (predictive ability of reflection factor pair future profits) of sequence R (t+1), is sequential value.

The average value of IC mean:IC sequence；

The standard deviation of IC std.:IC sequence；

Ann.ICIR:

Ann.ICIR=(mean (IC))/(std (IC)) √ C

Wherein C is that storehouse number is adjusted in 1 year.

Turnover rate turnover: i-th group of the next period recalls the number of share of stock/current i-th group of stock sum；

Yearization turnover rate Ann.Turnover: turnover rate turnover* year tune storehouse number C.

As shown in figure 18, tactful automated production: an overall scheduler task is created in produce, receives the page Data are stored in mongo, then start to tie up first layout, after Mission Success, send MQ and notify to give produce module, if some Step executes failure, whole automatic building suspension of task.Since manual operation generates effective plan from upload dimension member is started to the end Slightly, there are many steps in centre, and the function of a row can be increased in the process by operating comparatively laborious, tactful automated production, will tie up First layout is spliced into a process, simplified operating procedure as assessment, strategy layout and Policy Filtering process.

Policy Platform calculates the weight of the fund proportion of each strategy in combination using machine learning algorithm.Online investment Combined purpose is to select investment combination weight in each period, to maximize its final wealth.

As shown in figure 19, extensive strategy production (parallel computation)-strategy returns when survey overfitting in order to prevent, In All be using long period when survey to strategy, the mode of the whole city carries out back survey, and operand is huge, single machine without Method, which meets back, surveys performance requirement, therefore designs multi-host parallel/single machine multithreading and return time for surveying framework to support magnanimity policy instance It surveys.Strategy returns survey task, and the policy instance of generation is excessive, alreadys exceed a machine and carrys out the service ability upper limit, then automatically should Task splits into subtask, and the fractionation granularity of subtask is subject to the upper limit of single machine while the policy instance of operation.

Technical effect case of comparative examples:

Scheme: using single machine multithreading, and each machine runs 200 threads simultaneously, is without shared number between per thread According to nothing is concurrently locked, and caches frequently-used data and K line market optimization algorithm by using local memory, reduces a large amount of IO behaviour Make, CPU is occupied close to 100% after starting task, and timeslice all concentrates on user time, utilizes multicore to greatest extent Cpu performance.

As a result: preceding six annual data of domestic forward market, one minute data carries out parallel return in 6 servers and surveys, Dan Shi The average time-consuming of example is 0.5 second, and single machine average throughput is 47.83 ten thousand bar/ seconds, compares returning for the mainstream in the industry that big quotient is developed Platform Xquant is surveyed, single machine performance is 12.5 times of Xquant.

As shown in figure 20, extensive strategy operation (linear expansion): the strategy that single strategy container is run simultaneously is on having Limit, it can guarantee process flow operation in general 1000 strategies, it is assumed that a strategy can averagely configure 100,000 money Gold is thus to limit platform to manage the fund (1,000,*10 ten thousand=100,000,000) of 100,000,000 scales, but if to run 100 Ten thousand strategies, the fund of management 100,000,000,000 just necessarily require tactful container to support cluster, and have the ability of level of linearity extension.

Example of the invention is explained in detail above in conjunction with embodiment, but the present invention is not limited to examples detailed above, Within the knowledge of a person skilled in the art, it can also make without departing from the purpose of the present invention Various change also should be regarded as protection scope of the present invention.

Claims (8)

1. a kind of whole city multi items finance based on intelligence dimension Meta-Policy platform provides guard system characterized by comprising

Transaction platform is responsible for picking declaration form instruction, the execution of algorithm transaction and docking Shi Chang Bao's single function；

Data platform is responsible for the production of service market access and related derivative market；

Policy Platform is responsible for the tactful on-line operation of strategy production and high performance, including off-line strategy production management subsystem With strategy of on-line operational management subsystem, the fund that Policy Platform calculates each strategy in combination using machine learning algorithm is matched The weight of ratio selects investment combination weight in each period, maximizes its ultimate yield；

Off-line strategy production management subsystem includes 3 submodules: strategy production module, for realizing Row control, report exhibition Existing, list query function；Strategy, which returns, surveys optimization module, returns the creation of survey task for realizing strategy and operation, parallel computation are appointed Business scheduling, injecting strategy to time xylometer；Xylometer module is returned, operation reserve, transmission market to strategy, operation when being used to back survey Strategy calls altogether by fund and the calculating held position and offer strategy API when tactful；The strategy of on-line operational management subsystem runtime Between, factor computing module establishes network linking with xylometer is returned, and receives from the policy signal for returning xylometer, and according to signal meter The tactical comment factor and scoring simple curve are calculated, after all policy instance operations finish, calculative strategy is in multiple kinds Comprehensive score resultant curve；

Setting is used to describe the dimension meta-rule in market in Policy Platform, the meta-rule of different dimensions and is combined with each other intersections, generation Descriptive power；Dimension member be a segment description market the atom logic that can not be subdivided, be under the jurisdiction of specific dimension meta-rule, have mostly or Short side is to the different members under the same dimension meta-rule are mutual exclusions, all market shapes in all member aggregation cover time sequences State；Portrait is the permutation and combination that different cycles are superimposed by multiple dimension members, for describing market from multiple dimensions more accurately, Formula: as=dimension (member) [period] × dimension (member) [period] × ..., as generation from one or more dimensions (member) [period] Permutation and combination；Policy Platform forms the tree structure of portrait, by the way of layer-by-layer optimizing, tree structure by tieing up first layout The assessment of returning that upper each node carries out the whole city is estimated；Portrait assesses optimization process step by step: as assessment is to tie up member layout for primary Each node of the multilayer tree structure of the picture of generation is assessed；One picture of each node on behalf, it is each as having assessed one A scoring and this corresponding optimized parameter of scoring, the parameter of highest scoring is by the optimized parameter as picture；It is calculated to reduce Amount, the corresponding optimized parameter of superior node can be genetic to next time as, lower layer's picture can basis on upper layer as optimized parameter On, parameter arrangement-cartesian product is carried out to the dimension member of this layer and realizes that blowing back assessment estimates；

The operation flow of Policy Platform includes: dimension member management: the dimension metacode uploaded being uploaded in system, and supports the later period Maintenance；It ties up first layout: multiple dimension members being selected according to dimension, the period carries out fully intermeshing combination, generates picture；As assessment: for primary It ties up a batch that first layout generates to survey as carrying out back, assess, qualified picture and parameter are found in screening；Tactful subelement pipe Reason: the code of tactful subelement is uploaded into system, and supports the maintenance in later period；Tactful layout: multiple strategy of selection are single Member carries out fully intermeshing, generates strategy；Policy Filtering: a batch strategy generated according to primary tactful layout is returned survey simultaneously one by one It is screened according to specified screening rule, qualified strategy carries out in-stockroom operation as available strategy；Strategy adaptation: needle To the storage strategy by screening, parameter adaptation is carried out, the corresponding best trade variety of actual parameter group is found；Be set as to It is online: to set qualified policy instance to online；Strategy is online: it will be shifted on line to online strategy from policy library, Into firm offer or mimic panel operating status.

2. the whole city multi items finance according to claim 1 based on intelligence dimension Meta-Policy platform provides guard system, feature It is, different for the evaluation criteria of the picture of different levels, for one layer of picture, according to net profit, second layer picture is then unit Duration amount of loss；What is generated in order to prevent is effective as there is the problem of over-fitting, is using the comprehensive of full kind in evaluation process It closes curve and carries out platform, while parameter optimization also applies parameter Plain algorithm, assess effectively as tactful layout can be participated in.

3. the whole city multi items finance according to claim 2 based on intelligence dimension Meta-Policy platform provides guard system, feature It is, Policy Platform includes produce, optimize, factor and strategy-api module, and modules principle of interaction is Up-stream module calls downstream module to service using dobbo, and downstream module notifies up-stream module to use MQ mode；Produce module It is responsible for picture assessment mainstream process control, service inquiry, metadata management is related to when returning survey of picture during as assessment, Optimize is called to survey to realize back by dubbo；Optimize is responsible for picture and returning for strategy is surveyed, and optimization returns and surveys task schedule, It returns and needs to call infrastructure to obtain container, injecting strategy and sending strategy initial order by dubbo when surveying； Infrastructure is responsible for and tactful container interacts, the management including strategy of on-line container and off-line strategy container； Strategy-api is responsible for tactful online, the start-stop of strategy of on-line；From offline container when factor is responsible for back surveying Correlator accepts an order, and the calculative strategy factor and scoring, calculate complete by send MQ message informing optimize or Person's produce module.

4. the whole city multi items finance according to claim 3 based on intelligence dimension Meta-Policy platform provides guard system, feature Be, the picture evaluation process of Policy Platform: Manager is received as assessment task start request, and is called by dubbo, is passed Pass Produce module；As assessment optimizing step by step, every layer of each picture requires that optimize is called to be returned for Produce starting It surveys；Optimize, which is received, surveys request as returning, and creation realizes that parameter is blown as returning survey task, by permutation and combination, starts back survey and appoints Business.

5. the whole city multi items finance according to claim 4 based on intelligence dimension Meta-Policy platform provides guard system, feature It is, it includes: to obtain available tactful container resource that survey task is returned in the starting of Policy Platform, and injecting strategy to container, transmission is opened Container is arrived in dynamic instruction；Container returns strategy in survey and enters in operation, can send hand position signal to Factor, Factor and receive order Start calculative strategy performance；Factor calculating, which finishes to send back after tactful performance, surveys completion notice to Optimize, Optimize can do task according to notice and end processing, including modification task shape body, discharge container resource.

6. the whole city multi items finance according to claim 5 based on intelligence dimension Meta-Policy platform provides guard system, feature It is, the tactful container asynchronous event driven process of Policy Platform: surveys environment if it is returning, then by inside container Monitor directly reads local Sqlite file and sends market；If it is online container, then market are subscribed to from Data-Source； Asynchronous event transmitting: market/signal is all in a manner of asynchronous event in tactful container Internal Transfer；Strategy of on-line support finishes The market of Bar, CurrentBar, Tick complete period and full kind are subscribed to and driving；The tactful internal historical quotes data that read are adopted It is cached with local memory to reduce a large amount of I/O operations, tactful runing time is compressed to Microsecond grade；Strategy generates signal in container Later, according to environmental variance judgement be back to survey or online, if it is time survey task, then send a signal to Factor module into The calculating of row performance can be sent directly to transaction modules, be then forwarded to trade market if it is online container.

7. the whole city multi items finance according to claim 6 based on intelligence dimension Meta-Policy platform provides guard system, feature It is, tactful automated production process: creates an overall scheduler task in produce, receive page data, deposit Mongo, then starts to tie up first layout, after Mission Success, sends MQ and notifies to give produce module, if some step executes Failure, whole automatic building suspension of task.

8. the whole city multi items finance according to claim 7 based on intelligence dimension Meta-Policy platform provides guard system, feature It is, increases the function of a row during tactful automated production, first layout will be tieed up, as assessment, tactful layout and strategy Screening process is spliced into a process, simplifies operating procedure.