High frequency securities trading method and system

Abstract

To provide a high frequency securities trading method and system capable of generating order data based on prediction data for a target item acquired using a machine learning model.SOLUTION: A high frequency securities trading method according to one embodiment of the present disclosure comprises the steps of: calculating a latency for a securities order for each of a plurality of candidate batch sizes; selecting a batch size from the plurality of candidate batch sizes based on the calculated latency; generating input data corresponding to the selected batch size using market data for a target item; generating prediction data for the target item at a future time associated with the selected batch size based on the generated input data using a machine learning model; and generating order data for the target item based on the generated prediction data.SELECTED DRAWING: Figure 12

Description

[Technical field]

[0001] The present disclosure relates to a method and system for high frequency securities trading, in particular, using machine learning models. generating order data based on the prediction data for the target item obtained using the High frequency securities trading methods and systems. [Background technology]

[0002] High frequency trading is the trading of stocks, bonds, and derivatives. By taking advantage of minute price fluctuations of securities such as live commodities, High-frequency securities trading is a method of trading in which transactions are made at high frequencies (hundreds to thousands of times per day). Speed ​​is very important. Generally, the trading algorithm will be generated based on the information entered. The faster you can process and output your data, the more you can take advantage of the transaction. Cut.

[0003] On the other hand, high-frequency securities trading techniques using machine learning models are used to obtain large amounts of Since the data is analyzed, it is not possible to predict the market price of a particular item through existing classical algorithms. Many more factors must be considered that affect the accuracy of the predictions that can be obtained. However, to use machine learning models to analyze large amounts of data, This computation can require a lot of storage space and processing resources. However, existing processors can may not be suited to support such high frequency securities trading techniques.

[0004] In addition, machine learning models require complex calculations on large amounts of data, so When using a learning model, there may be latency for securities orders. This delay time allows the machine learning model to output predictive data for securities. For example, a machine learning model may generate a time lag phenomenon in which the time that the model predicts is already in the past. The predicted data for the target event at T1 future time point (T1 is a positive number) was output. The time when the predicted data is acquired due to the delay time is T1+n (n is a positive number). However, a problem may occur in that the predicted data is data for the past. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Korean Patent Publication No. 10-2001-0016454 Summary of the Invention [Problem to be solved by the invention]

[0006] The present disclosure relates to a method for high frequency securities trading, a computer program, and a method for solving the above problems. The Company provides the program and the system. [Means for solving the problem]

[0007] The present disclosure relates to a method, an apparatus (system) and / or a computer-readable storage medium. The present invention may be embodied in a variety of ways, including as a computer program stored in a computer.

[0008] According to one embodiment of the present disclosure, a high frequency verification method is implemented by at least one processor. The method includes determining a delay time for placing a securities order for each of a plurality of possible batch sizes. A step of calculating a batch size based on the calculated delay time among a plurality of candidate batch sizes. Size selection stage: size was selected using market data for the target item A step of generating input data corresponding to a batch size, using a machine learning model, The future target species associated with the batch size selected based on the input data Generating prediction data for the eye and targeting based on the generated prediction data It may include generating order data for the item.

[0009] A computer for executing a high frequency securities trading method according to an embodiment of the present disclosure. A data program may be provided.

[0010] According to one embodiment of the present disclosure, an information processing system stores one or more instructions. a first memory for executing one or more instructions in the first memory; Calculate the delay time for placing a securities order for each of the multiple candidate batch sizes, and In the complement batch size, select the batch size based on the calculated delay time and The input data corresponding to the selected batch size is used to at least one processor configured to generate data; a second memory for storing an instruction; and executing one or more instructions in the second memory. By using machine learning models, the system can generate a set of selected Generate forecast data for the target item at a future time point related to the batch size generated. and a machine learning module configured to provide the generated prediction data to at least one processor. At least one processor may include a dedicated accelerator for the Generate order data for target items based on forecast data provided by the accelerator The device may be further configured to: Effect of the Invention

[0011] According to some embodiments of the present disclosure, depending on the frequency of currently received or collected stock data, By determining the batch size for input to the machine learning model, the accuracy of the predicted price is improved. Minimize or prevent unnecessary computations in machine learning models while improving or maintaining accuracy. It is possible.

[0012] According to some embodiments of the present disclosure, a machine learning model is The input data corresponding to the maximum batch size that the system can process is generated. The input data can be input to the machine learning model. More accurate predictive data for the target event can be obtained from the machine learning model.

[0013] According to some embodiments of the present disclosure, a delay time for each of a plurality of candidate batch sizes is is predicted, and the batch size at which the predicted delay time ends is ahead of the future time is By selecting this option, unnecessary operations (e.g., operations that cause time divergence) are eliminated from the machine learning model. This may be prevented from being performed in Dell.

[0014] According to some embodiments of the present disclosure, a single machine learning model is used to predict future predictions at multiple points in time. The price of the target item is accurately predicted, and the expected profit is maximized based on the predicted outcome. Thus, securities can be ordered at any future time, thereby maximizing the profitability of trading securities.

[0015] According to some embodiments of the present disclosure, a dedicated accelerator that can be expected to generate the maximum profit among a plurality of dedicated accelerators is A dedicated accelerator is selected, and calculations using the machine learning model are performed through the selected dedicated accelerator. In such cases, the computation speed can be increased, and more rapid order data can be generated and verified. Ticket trading profitability can be maximized.

[0016] The effects of the present disclosure are not limited to the effects mentioned above, and other effects not mentioned are included in the claims. From the description of the scope, a person having ordinary knowledge in the technical field to which the present disclosure pertains (referred to as a "ordinary engineer") This can be clearly understood as [Brief description of the drawings]

[0017] Embodiments of the present disclosure will now be described with reference to the accompanying drawings, in which like references refer to: Numbers refer to similar elements, but are not limited to the above. [Figure 1] FIG. 2 is a schematic diagram illustrating an example of the operation of an information processing system according to an embodiment of the present disclosure. [Diagram 2] 1 is a block diagram showing an internal configuration of an information processing system according to an embodiment of the present disclosure. [Diagram 3] 1 is a diagram illustrating an internal configuration of a processor according to an embodiment of the present disclosure. [Figure 4] This is a visualization of the delay time that occurs between a processor and a dedicated accelerator. [Diagram 5] 1 is a diagram illustrating a visualization of latency for a securities order. [Figure 6] 1 is a diagram illustrating an example of visualization of multiple future time points. [Figure 7] 11 is a diagram visually illustrating delay times calculated for each candidate batch size; [Figure 8] 13 is a diagram visually illustrating delay times calculated for each batch size of each dedicated accelerator. [Figure 9] 1 is a diagram showing an example in which a machine learning model according to an embodiment of the present disclosure outputs output data based on input data. [Figure 10] 1 is a diagram showing an example of a configuration of input data for a machine learning model according to an embodiment of the present disclosure. [Figure 11] FIG. 2 is an exemplary diagram illustrating an artificial neural network model according to an embodiment of the present disclosure. [Figure 12] 1 is a flowchart illustrating a securities trading method according to an embodiment of the present disclosure. [Figure 13] 1 is a flowchart illustrating a method for pre-processing market data according to an embodiment of the present disclosure. [Figure 14] 1 is a diagram illustrating an example of expected profits at various future points in time; [Figure 15] 11 is a flowchart illustrating a securities trading method according to another embodiment of the present disclosure. [Figure 16] 11 is a diagram illustrating an example of a process in which order data is generated based on output data. [Figure 17] FIG. 2 is a block diagram of any computing device associated with generating securities trades in accordance with one embodiment of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] <Summary of the Invention> According to one embodiment of the present disclosure, the step of selecting a batch size comprises: For each future time point, among multiple candidate batch sizes, the calculated end of the delay the largest candidate batch size whose time point precedes each of a plurality of predetermined future time points; The method may include the step of selecting

[0019] Also, the step of selecting a batch size includes selecting a batch size for each of a plurality of predetermined future time points. calculating an expected profit for each of the batch sizes selected by the plurality of forecasts; Within the batch size selected for each future time point determined for The step of selecting the batch size that provides the highest waiting rate can be further included.

[0020] The step of calculating the expected profit includes calculating the expected profit for each of a plurality of predetermined future points in time. , the selected batch size, the return per query and the selected batch size The batch size chosen is based on the computation time of the machine learning model for each of the This may include calculating a respective expected return.

[0021] The high frequency securities trading method also includes transmitting the generated order data to the target securities exchange. The method may further include the step of:

[0022] At least one processor is a Field Programmable Gate Array (FPGA). e Gate Array) and dedicated accelerators for machine learning models. In this case, the step of calculating the delay time is performed for each of a plurality of candidate batch sizes. ,Data rate,The data rate of input / output data between FPGA and dedicated accelerator. Bandwidth, size of input / output data, computation speed of machine learning models using dedicated accelerators, FPGA based on at least one of the following: the processing speed of the The method may include calculating the delay time by using the received signal.

[0023] The latency is the time it takes for the market data to be preprocessed by the FPGA. Time required,The time required to transfer the preprocessed data from the FPGA to the,dedicated accelerator. The time it took for the dedicated accelerator to complete the calculation of the machine learning model, The time it takes for the calculation result to be transferred from the device to the FPGA and the time it takes for the result to be calculated by the FPGA This may include the time it takes for the order data to be generated based on the calculation results.

[0024] In addition, the high-frequency securities trading method requires high data precision. The step of calculating the delay time further includes the step of acquiring a candidate batch size. For each of the input data, calculating a delay time based on the accuracy of the input data. This can be done.

[0025] The delay time is also due to the processing time that increases as the data accuracy of the input data decreases. The computation can be based on a processing element.

[0026] Also, at least one processor includes a first dedicated processor for processing the computation of the machine learning model. and a second dedicated accelerator, and the step of calculating the delay time includes: For each of the above, a first delay time including the calculation time of the first dedicated accelerator and a second dedicated accelerator The step of selecting a batch size includes the step of calculating a second delay time including a calculation time of the device. For each of a plurality of predetermined future times, the calculated first delay time and the Among the two delay times, the end point of the lower delay time is each of several predetermined future times. The step of selecting the largest candidate batch size that is the earliest.

[0027] <Detailed Description of the Invention> Hereinafter, specific details for implementing the present disclosure will be described in detail with reference to the accompanying drawings. However, if there is a risk that the gist of the present disclosure may be unnecessarily obscured in the following description, In this specification, a detailed description of the functions and configurations used will be omitted.

[0028] In the accompanying drawings, identical or corresponding components are provided with the same reference numerals. In addition, in the following description of the embodiments, the same or corresponding components may be described in an overlapping manner. However, even if the technology relating to the component is omitted, such component is not intended to be included in certain embodiments.

[0029] Advantages and features of the disclosed embodiments and the manner in which they are accomplished are set forth in the accompanying drawings. This disclosure will become more apparent from the following detailed description of the embodiments, taken in conjunction with the accompanying drawings. The present invention is not limited to the embodiments disclosed herein, but may be embodied in many different forms. However, the present embodiment is intended to be complete and to enable those skilled in the art to fully understand the scope of the invention. It is provided for informational purposes only.

[0030] The following provides a brief explanation of the terms used in this specification and a detailed description of the disclosed embodiments. The terms used in this specification are as simple as possible while taking into consideration their function in this disclosure. The terminology chosen is currently in wide use and common, but this is not to be confused with the intention of engineers working in related fields. Or it may change due to precedents, the emergence of new technologies, etc. In addition, in certain cases, the applicant may In some cases, the meaning of the terms is explained in detail in the description of the invention. Therefore, the terms used in this disclosure are not simply term names. must be defined based on the meaning that it has and the overall content of this disclosure.

[0031] In this specification, the singular expression includes the plural expression unless the context clearly indicates the singular. In addition, a plural expression includes a singular expression unless the context clearly specifies a plural number. Throughout the specification, when a part is said to include a certain element, this is not necessarily the case unless specifically stated to the contrary. Unless otherwise stated, it is not intended to exclude other elements but to include other elements. Taste.

[0032] In addition, the term "module" or "part" as used in the specification means software or A "module" or "part" refers to a hardware component that performs a certain function. However, a "module" or "part" is not limited to software or hardware. A "module" or "part" is something that exists on an addressable storage medium. The method may be configured to regenerate one or more processors. Thus, as an example, a "module" or a "section" may be a software component. , object-oriented software components, such as class components and task components. components, processes, functions, attributes, procedures, subroutines, program code Segments, drivers, firmware, microcode, circuits, data, databases It can contain at least one of the following: Components and "modules" or "parts" are smaller units of functionality provided within them. The components and "modules" or "parts" of the It may be further separated into "modules" or "divisions".

[0033] According to one embodiment of the present disclosure, a "module" or "unit" may be a processor and a memory. A "processor" may be a general-purpose processor, a central processing unit (CPU), a microprocessor, processors, digital signal processors (DSPs), controllers, microcontrollers, state machines, etc. In some environments, the term "processor" may be used interchangeably with "processor" to refer to a specific Application-specific integrated circuits (ASIC), programmable logic devices (PLD), field A "processor" may refer to, for example, a programmable gate array (FPGA), etc. , combination of DSP and microprocessor, combination of multiple microprocessors, A combination of one or more microprocessors coupled with a DSP core, or any other The term "combination of processing devices" may also refer to a combination of such configurations. "Memory" is to be interpreted broadly to include any electronic component capable of storing electronic information. "Memory" can refer to random access memory (RAM), read-only memory ( ROM), Non-Volatile Random Access Memory (NVRAM), Programmable Read-Only Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrical Erasable Programmable Read-Only Memory (EEPROM), flash memory, magnetic or optical data storage devices This may refer to various types of processor-readable media, such as processors, registers, etc. The processor can read and write information to and from memory. If memory is integrated into a processor, the memory is said to be in electronic communication with the processor. The memory is in electronic communication with the processor.

[0034] In this disclosure, a "system" refers to at least one of a server device and a cloud device. For example, a system may include one or more In another example, the system may be configured with one or more cloud devices. In yet another example, the system may be configured with both a server device and a cloud device. As yet another example, the system may operate as a client for high frequency securities orders. This may refer to a device.

[0035] In addition, the terms first, second, A, B, (a), (b), etc. used in the following examples refer to a certain It is merely used to distinguish one component from another, and the term The nature, order or sequence of the components is not limited.

[0036] In the following examples, certain components may be referred to as "connected," "coupled," or "bonded" to other components. When an element is described as being "connected," it means that the element is directly connected to the other element. Each component may be connected to or joined to another component. It should be understood that the terms "connect" and "connected" may also be used interchangeably.

[0037] In addition, the terms "comprises" and / or "includes" used in the following examples "comprising" means that the components, steps, acts and / or elements referred to are does not preclude the presence or addition of one or more other components, steps, operations and / or elements .

[0038] In the present disclosure, "each of a plurality of A's" or "each of a plurality of A's" means that the plurality of A's are included. It refers to all the components included in A, or to some of the components included in multiple A. Each can be referred to.

[0039] Prior to describing the various embodiments of the present disclosure, a brief explanation of terminology will be provided.

[0040] In this disclosure, "item" refers to stocks, bonds, derivatives, etc. that are traded on the securities market. A classification of securities, such as options, futures, etc., based on their content and form. In addition to individual items, there are also index-related items, items related to industry sectors, and specific commercial items. Includes items related to commodities (e.g., crude oil, agricultural products, gold, etc.), and exchange rate related items. can be done.

[0041] For the purposes of this disclosure, a "stock exchange" is an entity that trades securities that are distributed in at least one country. It acts as an intermediary for the listing and trading of securities issued by companies, governments, etc. In one embodiment, a stock exchange may include a stock exchange system. can be done.

[0042] In this disclosure, an "Order Book (OB)" refers to a transaction that exists in the securities market. Buy or sell orders (quote, quantity, buyer, seller, etc.) of existing buyers and sellers The list may include information about the person who is interested in purchasing the product or the person who is interested in selling the product.

[0043] In this disclosure, the "Top of the Book" (ToB ) may contain information for the highest bid and lowest ask prices.

[0044] For purposes of this disclosure, "market data" refers to data for items traded on a securities exchange. For example, market data can include the stocks traded on a stock exchange (among which At least a portion of the order book, announcements, news, etc.

[0045] In this disclosure, a "machine learning model" is a model that generates an answer to a given input. According to one embodiment, the machine learning model may be any model used to infer the The model is an artificial neural network that includes an input layer, multiple hidden layers, and an output layer. A layer can include a layer network model, where each layer contains multiple nodes. In addition, in the present disclosure, the machine learning model is an artificial neural network model. and an artificial neural network model may be referred to as a machine learning model.

[0046] In the embodiments of the present disclosure, "instructions" refer to functions. A set of computer readable instructions bound together by a rule, This refers to a component of a program that is executed by a processor.

[0047] Various embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings.

[0048] FIG. 1 is a schematic diagram illustrating an example of the operation of an information processing system 110 according to an embodiment of the present disclosure. According to one embodiment, the information processing system 110 may generate one or more To predict the market situation at a future time (near future time, for example, after a predetermined time) Based on this, orders for the target items can be generated and traded on the target stock exchange ( In high-frequency securities trading, market data can be transmitted to the It is very important to generate and transmit orders at high speed based on the In securities trading, even microsecond latency must be taken into account. First, to reduce the delay time, the information processing system 110 It can be colocated close to the servers of the stock exchange. .

[0049] According to one embodiment, information processing system 110 receives market data from a first stock exchange. In addition, the information processing system 110 can receive the information from a website other than the first stock exchange. The website can receive market data from one or more It can be a site that aggregates market data generated by exchanges, or a private company. Market data can be generated from orders for multiple products. The market data may include trades, trade orders, trade announcements, news, etc. In one embodiment, the market data is For example, market data can include data for a target item. the top of the order book for the target product, the (active) order list for the target product, The order may include a response from a first stock exchange to a previous order for the stock item, etc.

[0050] Market data can be received dynamically during a unit time period, i.e., depending on the securities market environment. Therefore, the amount of market data received by the information processing system 110 during a unit time is For example, when the stock market is volatile, the number of messages received per unit time may vary. The size of the market data to be collected may be large or the number of data items may be large. In other words, when the volatility of the securities market increases, the magnitude or frequency of fluctuations in the order book also increases. Accordingly, the amount of market data received by the information processing system 110 per unit time is The size or number of may be large.

[0051] In Figure 1, the First Stock Exchange is shown as one stock exchange, but this is for ease of explanation. The primary stock exchange is merely a forum for the purpose of establishing a securities exchange and may include one or more stock exchanges. In Figure 1, the first stock exchange is shown as a separate exchange from the second stock exchange. This is also for the convenience of explanation only, and the First Stock Exchange includes the Second Stock Exchange. and the second stock exchange may include the first stock exchange.

[0052] According to one embodiment, the information processing system 110 analyzes market data to generate orders. For example, the information processing system 110 may receive market data and / or Analyze the data generated based on market data and forecast it to one or more future dates (e.g. , n seconds later, where n is a positive real number) It is possible to predict the future market conditions and generate orders based on this. The process of analyzing the data generated based on the market data and / or machine learning models This can be accomplished by a neural network (e.g., DNN).

[0053] On the other hand, in high-frequency securities trading, market data can be analyzed quickly to generate orders. However, general processors cannot handle the complex and large amount of computation required for machine learning models. It does not have the storage space and computing resources to support the calculation, so it can use a general processor. When using a machine learning algorithm to drive a machine learning model, processing speed and / or efficiency may decrease. In consideration of such points, the information processing system 110 according to an embodiment of the present disclosure uses a machine learning model. Dedicated accelerators for the neural processing units (e.g., Neural Processing Units) The dedicated accelerator can include a neural processing unit (NPU), Integrated circuits for The chip can be implemented using an application specific integrated circuit (ASIC).

[0054] On the other hand, when using machine learning models, appropriate pre- and post-processing steps may be required. For example, you can generate input data for a machine learning model from market data, or generate a It may be necessary to generate order data based on the data output from the Pre- and post-processing process may change due to changes in market conditions, regulations, compensation rules for market makers, etc. The processor that handles these pre- and post-processing steps can be used for specific applications. When using custom-made semiconductors (e.g. ASICs) for specific applications, If the pre- / post-treatment process is changed due to design impossibility, the changed pre- / post-treatment process There is a problem that the processor to execute the program must be manufactured again. The process, except for driving the machine learning model, can be reprogrammed and / or redesigned. A programmable processor (e.g., a field programmable gate array (FPGA)) A processor implemented with a FPGA (Field Programmable Gate Array) This can be accomplished.

[0055] As mentioned above, the processor that runs the machine learning model is a dedicated accelerator (e.g., NPU The combination of ASICs enables fast and efficient computation of machine learning models. In addition, the pre- and post-processing steps are reprogrammable or redesignable processes. By using a processor (e.g., FPGA) to process the continuously changing market conditions, In this way, the pre- and post-processing processes can be flexibly changed according to the needs of the transaction. Flexible pre / post processing by using two or more different compatible processors to The implementation of the process and efficient and rapid computation of the machine learning model can be realized at the same time. The internal configuration and data flow of the information processing system 110 will be described in detail with reference to FIGS. This will be described later.

[0056] FIG. 2 is a block diagram showing an internal configuration of an information processing system 110 according to an embodiment of the present disclosure. The information processing system 110 includes a memory 210, a processor 220, a communication module 23, and a 2, the information The processing system 110 utilizes a communications module 230 to communicate information and and / or may be configured to communicate data.

[0057] The memory 210 may include any non-transitory computer-readable storage medium. According to one embodiment, the memory 210 is a read only memory (ROM). , disk drives, SSDs (solid state drives), flash memory Permanent mass storage devices such as flash memory Other examples include: Non-perishable large-capacity storage such as ROM, SSD, flash memory, disk drives, etc. The device is a separate permanent storage device that is separate from the memory and is included in the information processing system 110. The memory 210 may also include an operating system and at least one program. Ramcode (for example, a machine learning model installed and run on the information processing system 110) Codes for calculations, pre / post processing, transmission of securities orders, etc. may be stored. Although memory 210 is shown as a single memory, this is for convenience of illustration only. , the memory 210 may include multiple memories.

[0058] Such software components may be read by a computer separate from the memory 210. The program can be read by a separate computer-readable recording medium. The recording medium may include a recording medium directly connectable to the information processing system 110. For example, flexible disks, disks, tapes, DVD / CD-RO This includes computer-readable recording media such as M drives and memory cards. As another example, a software component may be a computer-readable recording medium. The memory 210 may be loaded through the communication module 230 instead of through the At least one program distributes installation files for the developer or application. The file distribution system provides the file through the communication module 230. computer programs used to analyze market data, predict future market conditions, and place securities orders The program is loaded into the memory 210 based on a program for generating and transmitting the It is possible.

[0059] The processor 220 performs basic arithmetic, logic, and input / output operations. The instructions may be stored in the memory 210 or the processor 212. or communication module 230 to a user terminal (not shown) or other external system. For example, the processor 220 may utilize a machine learning model to generate a prediction based on the input data. and generating forecast data for the target item based on the generated forecast data. The order data thus generated can be used to trade on the target stock exchange. can be transmitted to

[0060] The communication module 230 communicates with a user terminal (not shown) and an information processing system via a network. The information processing system 110 may provide a configuration or function for communicating with each other. The system 110 communicates with an external system (e.g., a separate cloud system). As an example, the information processing system 110 may provide a configuration or function for The control signals, commands, data, etc. provided under the control of the processor 220 are 230 and a communication module of a user terminal and / or an external system via a network For example, the external system ( The stock exchange system may receive order data and the like from the information processing system 110.

[0061] In addition, the input / output interface 240 of the information processing system 110 10 or for input or output that the information processing system 110 may include For example, an input / output interface may be provided. The interface 240 is a PCI express (registered trademark) interface, At least one of the Ethernet interfaces In FIG. 2, the input / output interface 240 is a processor 220. However, the present invention is not limited to this and may include an input / output interface. 240 may be configured to be included in the processor 220. It may contain more than two components.

[0062] The processor 220 of the information processing system 110 may include a plurality of user terminals and / or a plurality of Manage, process and / or store information and / or data received from external systems According to one embodiment, the processor 220 may be configured to It can receive market data from the stock exchange system and the secondary stock exchange system. The processor 220 generates one or more orders for the target item based on the received market data. and generating order data based on the generated forecast data. In FIG. 2, the processor 220 is shown as a single processor. is for convenience of explanation only, and processor 220 may include multiple processors. For example, the processor 220 may be implemented as an FPGA for pre-processing and post-processing. At least one processor, one or more embodied in an ASIC for a machine learning model A dedicated accelerator may be included. In this case, at least one processor implemented in an FPGA may be included. The processor executes one or more instructions stored in the first memory to implement the ASIC. The one or more dedicated accelerators may be implemented using one or more instructions stored in a second memory. can be executed.

[0063] FIG. 3 is a diagram showing an internal configuration of a processor according to an embodiment of the present disclosure. In the following, the processor 300 is embodied in the form of a board. For example, the processor 220 of the information processing system 110 corresponds to the processor 3 of FIG. 00. As another example, the processor 220 of the information processing system 110 may be embodied in the form of may include the processor 300 of FIGS.

[0064] According to one embodiment, the processor 300 includes at least one processor for pre- / post-processing of data. processor 320 and a dedicated accelerator for machine learning models (e.g., implemented in an ASIC) For ease of explanation, in Figs. 3 and 4, At least one processor 320 for pre / post processing is an FPGA 320, The dedicated accelerator 340 for the learning model will be described as being an NPU 340.

[0065] The FPGA 320 includes a data receiving unit 322, a data preprocessing unit 324, and an order generating unit 326. The internal structure of the processor is explained by dividing it into different functions in FIG. It is clear that this does not necessarily mean that the two are physically separated. The internal configuration of the FPGA 320 shown is merely an example and does not include only the essential components.

[0066] According to one embodiment, the data receiver 322 of the FPGA 320 may receive data from one or more stock exchanges (e.g. For example, you can receive market data from a primary stock exchange, a secondary stock exchange, etc. In one embodiment, the one or more stock exchanges may include a target stock exchange. Here, the target stock exchange is a destination to which the order data is transmitted. Based on the data, the purchase or sale of the security may proceed.

[0067] Market data may include data for instruments traded on one or more securities exchanges. For example, market data may be provided for stocks traded on stock exchanges (including at least additional market data may be available for the target For example, market data may include the outperformance of a target item. the top of the order book, a list of (valid) orders for the target item, This may include the target stock exchange's response to a previous order regarding the target stock exchange.

[0068] The data receiver 322 receives market data from the stock exchange whenever it is necessary to update the market data or periodically. On the other hand, the stock market and / or trading If the fluctuation range of the stock market is large, the number of market data received may increase, and the stock market or If the volatility of the target security is small, the frequency of receiving market data will be reduced. For example, if the price fluctuation range of the target item is large, the data receiving unit 322 Receive market data including order books for various stocks more frequently per unit of time. Conversely, if the price fluctuation range of the target item is small, the data reception The 322 exchanges market data including the order book for the target item within a unit time. It can be received in fewer attempts.

[0069] In high-frequency securities trading, it is important to process data at high speeds, so market data The data is transmitted via the User Datagram Protocol (UDP), which has a high data transmission speed. However, in some embodiments, If necessary (e.g., to ensure the reliability of the data), we may use other methods to receive market data. A communication protocol (eg, TCP / IP) may be used.

[0070] The data preprocessing unit 324 performs machine learning processing based on the received one or more market data. Input data for the model may be generated. According to one embodiment, a data preprocessing unit 324 selects one or more input features for one or more items of market data. For example, the data preprocessor 324 may include the input data. The system may include a feature extraction unit for extracting and filtering input features to be processed.

[0071] In one embodiment, the input data may include one or more items that may be a target item. This may include items that may be leading indicators of fluctuations. For example, the target If the stock (spot) of Company A is the stock of Company A, the futures contract related to the stock of Company A, Options related to the formula, options related to Company A included in other exchanges, products related to Company A The input data may include data on futures items (e.g., crude oil) and In one embodiment, the one or more input features included in the input data include a market share of the target event. For example, the input features can include market price (transaction price) and ), the price at the top of the buy order book, the quantity, the price at the top of the sell order book, the quantity Volume, number of buyers, number of sellers, next bid price at top of order book, order - The next level of the sell price at the top of the order book, the variance of the bids and offers included in the order book, etc. Various information extracted from order books for more than one item, processed information and / or The input data may include the reliability of the information. For the configuration of the input data, see FIG. 10. This will be explained in more detail later.

[0072] According to one embodiment, the data preprocessor 324 may be configured to process a plurality of predetermined candidate batch sizes. Select one of them and run the machine learning model to match the selected candidate batch size. It is possible to generate input data for a set of rules, including multiple batch sizes. A list of candidate batch sizes may be generated in advance, and the data preprocessing unit 324 may The data preprocessing unit 324 can select any one of the batch sizes from the list. Calculate the latency for the securities order and, based on the calculated latency, A single batch size can be determined from the batch size list. The delay is calculated. The method will be described in detail below with reference to FIGS.

[0073] The input data generated by the data preprocessing unit 324 is used for the machine learning model. The data is then transmitted to the accelerator, NPU340, and input to a machine learning model (e.g., DNN). According to one embodiment, the NPU 340 is a specialized processor for running machine learning models. The NPU340 is a machine learning model that uses input data. In response to inputting a target event, predictive data for the target event can be obtained. For example, the NPU340 can feed input data to a machine learning model and generate a prediction for one or more future points in time. The output data is a forecast of the price (e.g., market price) of the target item. The machine learning model derives output data related to orders for the target line item. This will be described in detail later with reference to FIGS.

[0074] The order generation unit 326 can receive the prediction data output from the machine learning model. This allows the generation of order data at the target stock exchange based on the forecast data. For example, the order generator 326 may generate a target type for a future time period inferred from a machine learning model. Targets are set according to pre-determined rules based on price movement forecasts and / or forecast prices. For example, order data for a target item can be generated. If the price of the stock is predicted to rise, the order generator 326 immediately creates a new buy order. A sell order may be created or an existing sell order may be modified. Order data is the type of order for the target item (new order, order cancellation, order modification) , whether to buy or sell, price (quote), quantity, etc.

[0075] The delay in the information processing system until order data is generated based on market data The delay time depends on the data rate, FPGA320 and NP The bandwidth of input / output data between the U340, the size of input / output data, and the functions of the NPU340 The computation speed of the machine learning model, the processing speed of the FPGA320, or / and the visual The time can be calculated based on the busy state.

[0076] A larger batch size can slow down the computation speed of the machine learning model, but it is important to consider the latency. If the batch size is set large without considering the above, the results of the predicted future points in the machine learning model will be A time lag problem may occur where the results are already in the past at the time of output. For example, The machine learning model is then fed with input data corresponding to the target, and the response is calculated as T1 future. If forecast data for the target security for the time point is obtained at time T1+1, the forecast data is At the time of acquisition, T1 is already in the past. In such a case, the output at T1 When it comes to generating order data based on input data, the predictions of machine learning models are It may be difficult to predict the revenues generated from such transactions.

[0077] On the other hand, to reduce latency, we should also consider how to minimize the batch size. However, the smaller the batch size, the lower the prediction accuracy of the machine learning model. Therefore, it is recommended to increase the batch size to increase accuracy and expected profit. Although this can be advantageous, if the batch size is increased too much, the results obtained from the machine learning model may be degraded. This can lead to a time lag problem, where the data collected in the future is already past data. The delay time for securities orders is accurately predicted and expected within an acceptable range without timing discrepancies. Determine the batch size that will give you the highest waiting profit (e.g., the largest possible batch size). It can be important to accurately calculate and maintain the latency for securities orders. Based on the delay time, the expected profit is maximized without causing the timing deviation phenomenon. A large batch size may be selected.

[0078] 4 and 5, a method for calculating a delay time for a securities order according to one embodiment of the present disclosure is shown. This article explains how to do this.

[0079] Figure 4 shows an example of the latency that occurs in a processor and a dedicated accelerator. As shown in Figure 4, when market data is received, it is stored in the The first sub-delay time t1 stored in the FPGA stage (e.g., a receiving buffer, a memory, etc.) is The first sub-delay time t1 may be determined based on the data rate or may be calculated based on at least one of the magnitude of the market data. The higher the data rate, the shorter the first sub-delay time t1 can be. The larger it is, the longer the first sub-delay time t1 can be.

[0080] Additionally, the stored market data must be pre-processed to generate the input data. The required second sub-delay time t2 may occur in the FPGA 320. can be calculated based on the batch size and / or the processing speed of the FPGA 320. That is, the data preprocessing unit 324 preprocesses the input data among multiple candidate batch sizes. A batch size for the data is selected, and input data corresponding to the selected batch size is generated. The larger the batch size, the longer the second sub-delay time t2. The time t2 may be increased. Also, the slower the processing speed of the FPGA 320, the longer the second sub-delay time t 2 may increase.

[0081] The third sub-delay time t3 required between when the input data is received and when it is loaded into the memory. The third sub-delay time t3 depends on the size of the input data and / or and the input / output bandwidth between the FPGA 320 and the NPU 340. For example, , the larger the input / output bandwidth, the shorter the third sub-delay time t3 can be. The larger the third sub-delay time t3, the greater the increase.

[0082] Additionally, it takes time for the machine learning model to be calculated based on the input data loaded. A fourth sub-delay time t4 may occur in the NPU 340. Based on the speed of the U340 machine learning model and / or the size of the input data The larger the input data is, the longer the fourth sub-delay time t4 may be. The slower the computation speed of the learning model, the longer the fourth sub-delay time t4 may be.

[0083] In addition, the result value of the machine learning model (i.e., the predicted data) is transmitted to the FPGA 320. A fifth sub-delay time t5 may occur in the NPU 340 until the fifth sub-delay time t5 is reached. t5 is the size of the predicted data or / and the input / output between the FPGA 320 and the NPU 340. For example, the input / output bandwidth between the FPGA 320 and the NPU 340 can be calculated based on the The larger the bandwidth, the shorter the fifth sub-delay time t5. The 5th sub-delay time t5 may be increased.

[0084] Based on the prediction data (i.e., result value) of the machine learning model received from the NPU 340, The sixth sub-delay time t6 required for generating and transmitting the order data is The sixth sub-delay time t6 may occur due to the processing speed or / and delay time of the FPGA 320. The higher the data rate, the shorter the sixth sub-delay time t6. However, the higher the processing speed of the FPGA 320, the shorter the sixth sub-delay time t6 can be.

[0085] The delay time for a securities order is calculated by adding up each of the sub-delay times t1 to t6 described above. It can be calculated.

[0086] FIG. 5 is a visualization of delay times 500 for a securities order. As shown, Add up the sub-delay times t1 to t6 generated by the FPGA320 and / or the NPU340. Then, the delay time 500 for the securities order can be calculated.

[0087] On the other hand, if the FPGA 320 is busy, the FPGA 320 Sub-delay times due to the state can be added to the delay time 500. If the NPU340 is in a busy state, the sub-delay time due to the busy state of the NPU340 is added to the delay time 500. For example, additional processing may be required due to the busy state of the FPGA 320 and / or the NPU 340. The sub-delay time is determined by the current processing status of the FPGA 320 and / or the NPU 340. The decision can be made taking into account the amount of work that needs to be done in the future. The additional sub-delay time due to the busy state of the NPU320 and / or NPU340 is It may be preset.

[0088] The machine learning model outputs forecast data for securities at a future time based on the input data. In contrast, different future times can be predicted based on input data. In this case, a machine learning model can be designed and trained to The number of future times may be predetermined.

[0089] FIG. 6 is a diagram illustrating an example of a visualization of multiple future points in time. The horizontal axis of FIG. 6 indicates time. In a low data rate environment, as shown in Figure 6(a), the width of the unit time is relatively In a high data rate environment, as shown in FIG. 6(b), the width of the unit time can become relatively long. can be significantly shorter.

[0090] As shown in Figure 6, the machine learning model predicts the price and / or In Figure 6, the time interval k is a multiple of 5. In one embodiment, a processor (e.g., , FPGA) for each predetermined future time point, each candidate batch size securities Calculate the delay time for the order and calculate the order for each future time based on the calculated delay time. A candidate batch size can be selected by the processor. For each future time point, we select the batch size that is the best for the end of the delay time. Select the largest candidate batch size that is prior to each of a plurality of predetermined future time points. That is, the processor can have multiple candidate buffers to choose from at a particular future time. Within the batch size, one or more candidate batches whose delay time ends before a particular future time point Identify the sizes and identify the largest candidate batch size among the identified candidate batch sizes. can be selected as a candidate batch size for the future.

[0091] In Figure 6, the maximum batch size is selected for each future time point. Using Figure 6 as an example, for the fifth (k=5) future time point, the delay The candidate batch sizes whose end time is before the 5th future time are "1" and "2" Accordingly, the maximum batch size, “2”, is the batch size for the fifth (k=5) future point in time. As another example, the 10th (k=10) future batch size For points, the candidate batch size for which the end of the delay time is earlier than the 10th future time point is The size is less than or equal to "4", and therefore the maximum batch size of "4" is the 10th (k=10) can be selected as candidate batch sizes for future time points.

[0092] As explained above, the prediction data of the machine learning model for the kth future time point is effectively used. In order to do this, the order data based on the forecast data must be available at the kth future time point before the target securities However, there may be a long delay in placing a securities order. In this case, the time when the order data is transmitted to the target stock exchange is after the k-th future time. In this case, the predicted prices of the securities may not be valid, especially in high frequency securities. In this case, the candidate batch size is calculated based on the delay time and the future time. A batch size is selected, and the input data corresponding to the selected batch size is generated. This can be important in securities transactions.

[0093] Figure 7 is a visual example of the delay time calculated for each candidate batch size. This shows that the batch size increases from batch 4 to batch 16. From the 1st future time to the 3rd future time, the future time can be farther away than the present time. As shown, we use multiple latency measures based on different candidate batch sizes. However, the first delay time (latency #1 to latency #3) can be calculated. #1) The end time of batch 4 passes the first future time point predicted based on batch 4, and ba When input data corresponding to the size of tch 4 is generated, the machine learning model prediction data is The time when the order data based on the order date is generated may be after the first future time. Similarly, the third delay time (l The end time of the trial #3) is the third future time point predicted based on batch 16. After a while, we generate input data corresponding to the size of batch 16 and run the machine learning model. The time point at which the order data is generated based on the forecast data may be after the third future time point. Due to the time difference, the input data corresponding to batch 4 or batch 16 In some cases, useful results may not be obtained.

[0094] On the other hand, the end time of the second delay time (latency #2) is predicted based on batch 8. Since the second future time point measured has not passed, the size of batch 8 and the corresponding input data When the machine learning model is generated, the output data is validated and the order data is generated at the correct time. This allows the delay time to be reduced as shown in Figure 7. If it is calculated, it corresponds to the maximum candidate batch size that does not cause the time divergence phenomenon. Batch 8 may be selected, and input data corresponding to the selected candidate batch size may be generated. The resulting data can then be input into a machine learning model.

[0095] Alternatively, a processor may contain multiple dedicated accelerators. In this case, the processor The delay time is calculated by taking into account the calculation time of each dedicated accelerator. You can choose the candidate batch size and NPU for computing the machine learning model.

[0096] Figure 8 is a visual example of the delay time calculated for each batch size of each dedicated accelerator. As shown in Figure 8, the delay by candidate batch size for dedicated accelerator 1 (NPU1) is The delay time (latency #1, latency #2) is calculated, and the dedicated accelerator 2 (NPU 2) Latency by candidate batch size (latency#3, latency#4) Even with the same batch size, the delay time varies depending on the dedicated accelerator. The reason for the calculation is the speed of the machine learning model and / or the business This may be due to different busy states.

[0097] Based on the delay time for each candidate batch size of each dedicated accelerator, the processor performs machine learning operations. You can select any one of the dedicated accelerators and candidate batch sizes to run. If the delay time is calculated as shown, the processor will be the second dedicated accelerator among multiple dedicated accelerators. We selected the NPU2 as the dedicated accelerator for the computation of the machine learning model, and We can choose batch 8 as the candidate batch size. Once the size and dedicated accelerator are selected, we respond with candidate batch sizes based on market data. The input data is generated, and the generated input data is selected by a dedicated accelerator (NPU2). may be provided to.

[0098] Below, please refer to Figures 9 to 11 to learn the learning method of the machine learning model and the data obtained from the machine learning model. The predicted data will be explained.

[0099] FIG. 9 illustrates a machine learning model according to an embodiment of the present disclosure generating an output data 910 based on input data 910. 9 is a diagram showing an example of outputting data 920. According to one embodiment, the machine learning model 900 Based on the input data 910, output data 920 related to orders for the target item is output. According to one embodiment, the machine learning model may generate unknown results based on the input data 910. Outputs the predicted price (e.g., market price or mid-price) of the target item at a specific time in the future According to another embodiment, the number of future points in time can be calculated based on the input data 910. In this case, the machine learning model 900 is used to output the predicted price of the target item. Based on multi-horizon forecasting For each of a number of future times, the price of the target item may be predicted.

[0100] According to one embodiment, input data 910 input to the machine learning model 900 includes one or more An input feature map (inpu) containing one or more input features for one or more events at a time t feature map). Input data for the machine learning model 900 910 is described in more detail below with reference to FIG.

[0101] According to one embodiment, the machine learning model 900 is generated based on reference market data. The reference input data is used to generate reference output data related to securities orders on the target exchange. For example, the machine learning model 900 may be trained to infer the first stock exchange's Based on the first reference market data and the second reference market data of the second stock exchange The reference input data from time t to time t+M-1, the target species at time t+1, Using the intermediate price data, we apply the time interval containing a total of M consecutive points to the input data. Based on the calculated mean price, the target item can be supervisedly learned to infer the mean price of the target item at the next time point. .

[0102] According to one embodiment, the machine learning model 900 generates market data for a particular item at a particular time. A training set containing the ground truth data and the Based on the data, we estimate forecast data for a specific item at multiple points in time from a specific point in time. Here, the correct answer data is the characteristic of each of multiple points in the future. The price of the fixed item may be the price of the fixed item at a specific future time point output from the machine learning model 900. The inferred price for the item and the price for a specific item at a specific future time included in the ground truth data The rank difference (loss) is calculated, and the calculated difference is reflected in the machine learning model 900 (feed The weights of each node in the artificial neural network can be adjusted by the do.

[0103] The output data 920 output by the machine learning model 900 is the price of the target stock exchange. The information may include information related to a securities order, and may be stored in a processor (e.g., a processor of an information processing system). The processor (processor) determines the target species based on the output data 920 according to a pre-specified rule. Order data for the eye can be generated.

[0104] According to one embodiment, the machine learning model 900 of the present disclosure is an artificial neural network model. For artificial neural network models, This is described in detail below with reference to FIG.

[0105] FIG. 10 illustrates an example of the configuration of input data 1010 for a machine learning model according to an embodiment of the present disclosure. 1 is a diagram showing an information processing system that receives market data from one or more exchanges. The input data 1010 can be generated based on the batch According to one embodiment, the input data 10 may have a size corresponding to the size of the input data 10. 10 is an input feature matrix including one or more input features for one or more events at one or more time points. It can include an input feature map.

[0106] For example, the input feature map is M (where M is a natural number) as shown in FIG. ) input features for K events at time points (where N is a natural number). In the illustrated example, the input feature map included in the input data is The data 1020 at a point (time m in FIG. 10) represents one or more events (time m in FIG. 10) at a particular time. , 1st item, 2nd item, 3rd item, etc.) the top price and quantity of the order book on the buy side, the top price and quantity of the sell side order book, etc. In addition, the input feature map included in the input data can include a specific input feature (see Figure 10) for the nth input feature) is collected at one or more time points (time points t-M+1 to time t) may include specific input features for one or more events. In one embodiment, the input feature map may include one or more input features for different sports. can be generated so as to intersect with each other.

[0107] According to one embodiment, one or more items included in the input data 1010 may be For example, the item that is the subject of the order may be a leading indicator of the market condition of the target item. If the target product is the stock (spot) product of Company A, the futures product related to the stock of Company A, options related to the Company's stock, options related to Company A held on other exchanges, and products related to Company A At least one of the futures contracts for the stock may be a leading benchmark contract. In an example, the one or more events may include a target event. The system will generate target categories based on input data, including data for the target category. In one embodiment, the future market conditions for each input item can be predicted. The relevant information may be contained in a symbol associated with each entry item.

[0108] According to one embodiment, one or more input features included in the input data 1010 are a target species. The input features can contain information that is meaningful to predicting the market situation. For example, the input features can be market prices (trading Buy side order book top price, quantity, sell side order book top price Price, quantity, number of buyers, number of sellers, next bid price at the top of the order book, The next bid price at the top of the order book, the variance of the bid prices included in the order book, etc. Various information that can be extracted from order books for one or more items, processed information, and In one embodiment, such a one may include a degree of reliability of the information. The above input features may be extracted from one or more events respectively.

[0109] As previously mentioned, the configured input data 1010 is fed to a processor (e.g., an FPGA, etc.) and input to the machine learning model. According to one embodiment, a future time and a candidate batch size are determined based on the delay time. Input data 1010 corresponding to the determined candidate batch sizes may be provided to the dedicated accelerator.

[0110] FIG. 11 illustrates an artificial neural network model 1100 according to one embodiment of the present disclosure. FIG. 11 is an example of a machine learning model. Machine learning technology and cognitive science are used to understand biological A statistical learning algorithm or a method thereof based on the structure of a neural network. It is a structure that executes the algorithm.

[0111] According to one embodiment, the artificial neural network model 1100 is a biological neural network. In the case of neural networks, artificial neurons are formed by synaptic connections. A node iteratively adjusts the synaptic weights to determine the correct response to a particular input. The problem is solved by learning to reduce the error between the desired output and the inferred output. It is possible to demonstrate a machine learning model with decision-making capabilities. For example, an artificial neural network The QModel 1100 is an arbitrary model used in artificial intelligence learning methods such as machine learning and deep learning. These may include probability models, neural network models, etc.

[0112] According to one embodiment, the artificial neural network model 1100 is The input data is generated based on the market data of the market, and the target is calculated at a future point in time. Get data related to securities orders at stock exchanges (e.g. price, price fluctuation, etc.) The artificial neural network model may include an artificial neural network model configured to infer According to another embodiment, the artificial neural network model 1100 is a multi-horizon model. Multi-horizon forecasting models, including Data relating to securities orders on the target exchange at a future date (e.g., prices, price fluctuations, etc.) For example, the prediction may be performed using the following method:

[0113] The artificial neural network model 1100 is composed of multiple layers of nodes and connections between them. The multilayer perceptron (MLP) The artificial neural network model 1100 according to the present embodiment includes an MLP. It can be implemented using one of various artificial neural network model structures. As shown in FIG. 1, the artificial neural network model 1100 receives an input signal from the outside. an input layer 1120 that receives data or signals 1110; An output layer 1140 that outputs data 1150, and a , n (herein) input layers 1120 receive signals, extract characteristics, and transmit them to the output layer 1140. The output layer is composed of hidden layers 1130_1 to 1130_n (n is a positive integer). 1140 receives signals from the hidden layers 1130_1 to 1130_n and outputs them to the outside.

[0114] The learning method of the artificial neural network model 1100 involves inputting a teacher signal (correct answer). Supervised learning is a method of learning that optimizes the solution of a problem by Learning method and unsupervised learning method that does not require a teacher signal In one embodiment, an artificial neural network (AI) is used. The network model 1100 infers data related to securities orders at a target exchange. For example, the neural network can be trained in a supervised and / or unsupervised manner. The network model 1100 calculates the target event at one or more future times from the reference input data. It can be supervised learned to infer the reference price.

[0115] The artificial neural network model 1100 thus trained is an information processing system. The communication module and / or memory may store the received signal. Inferring data related to securities orders at a target exchange in response to input of the data provided It is possible.

[0116] According to one embodiment, a method for inferring data related to securities orders at a target securities exchange is provided. The input data for the artificial neural network model is one or more It may include one or more input features for the event. For example, an artificial neural network The input data input to the input layer 1120 of the model 1100 is a single The data containing information on one or more input features for one or more events is stored as a vector data. In response to the input of data, the artificial neural network 1110 may generate a vector 1110 of data elements. The output data output from the output layer 1140 of the neural network model 1100 is the target A vector 1150 representing or characterizing data related to a security order on a stock exchange. That is, the output layer 1140 of the artificial neural network model 1100 may be Represents or characterizes data relating to securities orders on a target exchange at one or more future points in time. In the present disclosure, the artificial neural network may be configured to output a vector that represents the The output data of the work model 1100 is not limited to the types described above, and may include one or more undefined Any information / data that represents data related to securities orders at the target stock exchange at a future date may include.

[0117] In this way, the input layer 1120 and the output layer 1120 of the artificial neural network model 1100 are In 140, a plurality of input data are matched with a corresponding plurality of output data, Nodes included in layer 1120, hidden layers 1130_1 to 1130_n, and output layer 1140 By adjusting the synaptic values ​​between the Through this learning process, the artificial neural network model can be trained to The characteristics hidden in the input data of the rule 1100 can be known, and the calculation can be performed based on the input data. An artificial neural network model is developed to reduce the error between the output data and the target output. The synaptic values ​​(or weights) between the nodes of 1100 can be adjusted. The artificial neural network model 1100 responds to the input data by generating a target The system can output data related to securities orders on the stock exchange.

[0118] Hereinafter, with reference to FIGS. 12 to 14, a machine learning model according to an embodiment of the present disclosure will be used. Explain how securities are traded.

[0119] FIG. 12 is a flow chart illustrating a method 1200 for trading securities in accordance with one embodiment of the present disclosure. The method illustrated in FIG. 12 is merely one example for achieving the objectives of the present disclosure. It goes without saying that some steps may be added or removed as necessary. The method illustrated in FIG. 12 is executed by at least one processor included in an information processing system. For convenience of explanation, the processor included in the information processing system shown in FIG. The steps shown in FIG. 12 will be described as being performed by a processor. In addition, the processor is divided into two parts: the first processor for pre- and post-processing of data and the second processor for machine learning models. The first processor includes a second processor that includes an accelerator. The first processor may be an FPGA and the second processor may be an NPU.

[0120] The first processor may access at least one of the stock exchange and / or designated websites. The market data may also be received from one or more of the following markets (S1210). Market data includes transaction and valuation information for items traded on a stock exchange. For example, market data can be collected from stock exchanges and / or websites. The aggregated information may include trading, valuation information, etc. for one or more target items.

[0121] Subsequently, the first processor pre-processes the market data to provide input for the machine learning model. Force data may be generated and input data may be provided to the second processor (S1220). The specific method by which the market data is preprocessed to generate input data is shown in Figure 1. This will be explained with reference to 13.

[0122] The second processor then uses the input data to perform computations on the machine learning model, In response to this, prediction data for the target event is obtained, and the obtained prediction The measurement data can be provided to the first processor (S1230). For example, the second processor The processor receives input data from the first processor and inputs the input data into the machine learning model. Then, a series of operations are performed on the input data to generate the output data (prediction data) of the machine learning model. The machine learning model's prediction data includes future time points, target species, Data for price and expected price may be included.

[0123] The first processor then processes the output data of the machine learning model provided by the second processor. After generating order data based on the (predicted data), the order data is sent to the target stock exchange. For example, the first processor may transmit the prediction data to the Order data can be generated to buy or sell the target item. 1 The processor monitors whether the future time included in the prediction data will arrive, The order data generated at a future time or t seconds before the future time is used as the target certificate. In another example, the first processor may transmit the forecast data to a stock exchange. It monitors whether the future time arrives and determines whether the future time arrives or is t seconds from the future time. The order data is generated prior to the transaction and then sent to the target stock exchange. can be transmitted to.

[0124] FIG. 13 illustrates a method 1220 for pre-processing market data, according to one embodiment of the present disclosure. 1 is a flowchart for processing a first processor when market data is received. One of the multiple future time points determined for the purpose of verification is selected as the verification target. In one embodiment, the first processor may select a plurality of predetermined It is possible to select the nearest future time that has not been selected among the future times.

[0125] Then, the first processor calculates the delay time for each candidate batch size for the selected future time point. In one embodiment, the first processor may calculate the data rate (S1320). bandwidth between the first and second processors; the amount of input and output data; The size, the computation speed of the machine learning model by the second processor, the processing speed of the first processor, and / or the busy state of the second processor. The delay time for each can be calculated.

[0126] Subsequently, the first processor sets the end point of the calculated delay time ahead of the selected future time. Identify one or more candidate batch sizes that are suitable for use, and select the largest candidate batch size among the identified candidate batch sizes. The batch size having the required size is used as the candidate batch size for the future time point selected (i.e., during validation). In this way, one future time point can be selected as a future time point (S1330). A candidate batch size may be selected.

[0127] Once the candidate batch size for the selected future time point has been selected, the first processor The expected profit for the current candidate batch size can be calculated (S1340). Here, the expected profit can be the profit expected per unit time until the order data is generated. Revenue is the revenue per query, the candidate batch size, and the machine learning response to the candidate batch size. Based on the computation time of the model, the expected profit can be obtained as follows: This can be obtained through number 1.

number

[0128] Here, "price" represents the revenue per query and is defined as a predetermined constant. In addition, bs represents the batch size, and t_dnn is the computation time for the machine learning model. Therefore, the computation speed of the machine learning model (i.e., the processing speed in the NPU) and / or the input It can be calculated based on the size of the data.

[0129] Once the expected profit for each candidate batch size in the future has been calculated, the first processor Determine whether selection (verification) for all predetermined future times has been completed. (S1350).

[0130] Among the multiple future time points determined in advance, there are still future time points that have not been selected (verified). In response to the determination, the first processor selects a future time point that was not selected as a verification target as a verification target. The first processor can then select the second processor as a verification target again (S1360). The process may continue from step S1320 for other future time points selected as targets. can.

[0131] Meanwhile, as a result of the determination in step S1350, a predetermined future time is selected (verified) In response to determining that no future time points remain that are not yet selected, the first processor Identify the expected profit for the selected candidate batch size, and select the candidate for the future time that has the maximum profit among them. The batch size can then be finally selected (S1370). The processor uses market data to select the inputs corresponding to the candidate batch sizes for the future. and providing the generated input data to the second processor (S 1380).

[0132] According to this embodiment, a batch size that causes a time divergence is not selected, and an inappropriate In addition, according to the present embodiment, it is possible to prevent necessary calculations from being performed in the machine learning model. Select a future time and candidate batch size that can obtain the maximum profit, and select the selected candidate batch size. can be input to the machine learning model.

[0133] Additionally or alternatively, data precision of input data may be n) may be preset, in which case the first processor determines whether the first processor determines the number of input data bits based on the data accuracy of the input data. Here, data accuracy is the degree to which input data is not lost. For example, if the data accuracy is 100%, the input The data is provided to the second processor in its entirety and without any loss to be input into the machine learning model. As another example, if the data accuracy is 50%, the input data is 1 / 2 It is either quantized to the corresponding bits (e.g. int8 to int4) or compressed by a factor of 2 and processed mechanically. It could mean that the data is input to a learning model. In some cases, the market data is compressed by a factor of 2 or quantized by a factor of 2. Note that the input data is generated based on quantized or compressed market data. It could mean.

[0134] According to one embodiment, the second processor (e.g., an NPU, etc.) determines the accuracy of the input data. Processing elements having different sizes For example, processing input data with fixed point 8 The NPU is a processing element for processing fixed point 8. As another example, input data having fixed point 4 can be The NPU for processing is the processing for processing fixed point 4. However, the processor that processes fixed point 8 The size of the processing element is the size of the processing element that processes fixed point 4. Accordingly, even if the NPU is installed in the same space, For example, the processing element that processes fixed point 4 is fixed point The number of processing elements may be about twice as many as the number of processing elements that process t 8. For example, Processing element for processing point 8 (e.g. data accuracy 100%) The number of processing elements for processing fixed point 4 (e.g. data If the probability is twice as high as 50%), the processing time required to process fixed point 4 will be The batch size of input data fed into the NPU containing the filtering element can be increased by a factor of two. In this case, an NPU with processing elements for handling fixed point 8 and an NPU that contains the processing elements for handling fixed point 4. The amount of SRAM required for the calculation and the time required for the calculation may be the same or similar.

[0135] If the data accuracy of the input data is small, the input data is completely input to the machine learning model. This may result in lower forecast accuracy, but the data is accurate. There are more processing elements to process less accurate input data, so the buffer size of the input data is reduced. The size of the results may increase, i.e., the accuracy of the data may decrease, resulting in lower predictions. Accuracy can affect revenue per query, but low data accuracy can also affect revenue per query. The increased batch size due to this can have an impact on the expected profitability. When we look at number 1, the expected revenue falls as the revenue per query falls, but it also increases. For example, the data accuracy is halved and the number of queries per query is increased by the batch size. If we assume that the profit margin decreases by 50 percent and the batch size doubles, the expected profit becomes The benefits may not be changed.

[0136] According to one embodiment, multiple dedicated accelerators may be included in an information processing system. In this case, the first processor performs the first dedicated accelerator's execution for each of the multiple candidate batch sizes. A first delay time including a calculation time of the second dedicated accelerator and a second delay time including a calculation time of the second dedicated accelerator are calculated. The first processor may also determine a time for each of a plurality of predetermined future times. The end point of the shorter delay time among the first delay time and the second delay time is determined based on a plurality of predetermined We can select the largest candidate batch size ahead of each of the specified future times. In addition, the first processor is divided into dedicated accelerators and is assigned to the selected candidate batch size. Calculate the expected profit for each batch and select the candidate batch size, future time, and dedicated processing time with the highest expected profit. It is possible to identify and select the candidate batch size with the highest expected profit. The generated input data is then submitted to the selected dedicated accelerator (the dedicated accelerator with the highest expected return). It can be provided.

[0137] On the other hand, according to another embodiment of the present disclosure, the machine learning model calculates a time series for each of a plurality of future time points. For example, the machine learning model can output prediction data for the target event. The first prediction data for the target event at the time point, the second prediction data for the target event at the time point Outputting the second prediction data and the third prediction data for the target event at the third time point. The forecast data includes target security identification, target price and expected return. It is possible to do so.

[0138] FIG. 14 is a diagram showing an example of expected profits at various future times. The machine learning model infers expected profits for multiple future points in time based on the input data, and You can output forecast data including expected profits and target items for each future time. The target events at each future time point may be the same or different. Expected returns are shown in United States Dollars (USD).

[0139] The first processor generates a forecast including expected profits for each of a plurality of future times as shown in FIG. The system acquires measurement data, generates order data based on the forecast data, and sends it to the target stock exchange. It can be transmitted.

[0140] FIG. 15 is a flowchart illustrating a securities trading method 1500 according to another embodiment of the present disclosure. The method illustrated in FIG. 15 is merely one example for achieving the objectives of the present disclosure. It goes without saying that some steps may be added or deleted as necessary. The method shown in FIG. 15 is executed by at least one processor included in an information processing system. The processor may also include a first processor and a second processor for data pre-processing and post-processing. and a second processor including a dedicated accelerator for the learning model. Here, the first processor may be an FPGA and the second processor may be an NPU. .

[0141] The first processor receives one or more market data from a securities exchange and / or website. Here, the market data can be received from the stock exchange (S1510). It can include transaction and valuation information for the items traded. For example, market data Data is collected from stock exchanges and / or websites for one or more target securities. It may include transactions, evaluation information, etc.

[0142] The first processor then generates input data for the machine learning model based on market data. , and provide input data to the dedicated accelerator (S1520). A processor can generate input data corresponding to a pre-defined batch size. As another example, the first processor may determine the frequency and / or time of received market data. The batch size is determined by the size of the input data, and input data corresponding to the batch size is generated. This can be done.

[0143] The second processor then inputs the input data to the machine learning model and extracts the It is possible to obtain forecast data for the target event for each of multiple future points in time. For example, the second processor may generate a first prediction for the target event at the first time point (S1530). The measured data, the second predicted data for the target event at the second time point, and the target event at the third time point Third prediction data for the target event can be obtained from the machine learning model. The data may include expected returns for each target item at each future time point.

[0144] Then, the first processor receives forecast data for each of a plurality of future times from the second processor. At least one of the multiple future time points can be selected (S1540). In an embodiment, the first processor calculates a delay time for generating the order data, and At least one future time point may be selected from the plurality of future time points based on the Then, the first processor selects the earliest future time after the delay time from among the multiple future times. In another example, the first processor may receive the first signal at one or more future times after the delay time. Identify or calculate the expected revenue for each identified future time point, and then calculate the expected The future time when the profit is the maximum can be selected. For example, the calculation of the expected profit is determined in advance. The above-described algorithm or mathematical formula (e.g., Equation 1) may be used as an example. The expected profit for each can be calculated and output from the machine learning model. In this case, the expected profit is may be included in the forecast data.

[0145] The first processor then generates order data based on the selected future forecast data. After completion of the order, the order data can be transmitted to the target stock exchange (S1550). For example, the first processor may select a target included in the forecast data for a selected future time point. Orders can be generated to buy or sell the item at a target price. The first processor monitors whether the selected future time point arrives and calculates the future time point. The order data generated at a point t seconds before the current or future time is processed by the target stock exchange. In another example, the first processor may transmit the selected future time to the selected location. Monitor whether the event occurs at a future time or t seconds before the future time. After generating the order data, the generated order data can be transmitted to the target stock exchange. can.

[0146] FIG. 16 is a diagram illustrating an example of a process in which order data is generated based on output data. The machine learning model is used to predict the future for multiple time points (k=5, k=10, …, k=25). The prediction data can be obtained by using the latency. The forecast data with the earliest future time point (k=5) is not valid because it is outside the delay time range. It can be treated as valid (i.e., the end of the delay is later than the The second forecast data having the earliest future time point (k=10) among the remaining forecast data (The forecast data shown as a circle in Figure 16) is selected, and the second forecast data is In another example, the order logic may be executed to generate the order data. The future time point with the highest expected profit is selected from the forecast data, and the selected future time point is Order logic is executed based on the forecast data for the time point to generate order data. .

[0147] As yet another example, the first processor may time-order each of the available prediction data. By using these in sequence, multiple order data can be generated and transmitted to the target stock exchange. Using FIG. 16 as an example, the first processor executes the first process at the second future time (k=10). (2) transmitting second order data generated based on the forecast data to the target stock exchange; A third order data is generated based on the third forecast data at a third future time point (k=15). The data can be transmitted to the target stock exchange, and the fourth forecast can be made at the fourth future time point (k=20). and transmitting the fourth order data generated based on the measurement data to the target stock exchange. The fifth order data is generated based on the fifth forecast data at the fifth future time point (k=25). can be transmitted to the target stock exchange.

[0148] FIG. 17 illustrates any computing device related to securities exchange generation in accordance with one embodiment of the present disclosure. 17 is a block diagram of a device 1700. For example, the computing device 1700 may be an information processing system. 10. The system 120 may be associated with a user terminal (not shown). The computing device 1700 includes one or more processors 1720, a bus 1710, and a communications The interface 1730, the computer program executed by the processor 1720 Memory 1740 and computer programs 1 that load RAM 1760 17 may include a storage 1750 for storing the Only components related to the embodiment are shown. A person skilled in the art would recognize that in addition to the components shown in FIG. 17, there are other general-purpose components. It will be appreciated that it may be included.

[0149] The processor 1720 controls the overall operation of each component of the computing device 1700. The processor 1720 is a central processing unit (CPU). ), MPU (Micro Processor Unit), MCU (Micro Co GPU (Graphic Processing Unit), GPU (Graphic Processing Unit) it), NPU (Neural Processing Unit) or the technology of this disclosure The processor may be any type of processor known in the art. The processor 1720 includes at least one application for executing a method according to an embodiment of the present disclosure. The computing device 1700 may perform operations on a program or application. For example, the computing device 1700 may include one or more processors. is a processor implemented in FPGA, and a dedicated machine learning model implemented in ASIC. May include accelerator (NPU ASIC).

[0150] The memory 1740 may store various data, instructions, and / or information. The library 1740 includes storage 175 for performing the methods / operations according to various embodiments of the present disclosure. 0. One or more computer programs 1760 can be loaded from the memory 1 740 may be implemented as a volatile memory such as a RAM, but the scope of the present disclosure is not limited thereto. If the computing device 1700 includes multiple processors, The routing device 1700 includes a first memory for the first processor and a second memory for the second processor. In this case, the first processor may include a second memory for storing the performing one or more of the instructions described above to perform the data pre-processing or The second processor may perform one or more image processing operations stored in the second memory. By executing the execution, the machine learning model is calculated and the prediction data is generated. The first processor may be provided with the first signal.

[0151] The bus 1710 provides communication between the components of the computing device 1700. The bus 1710 is an address bus, a data bus, There are various types of buses such as Data Bus and Control Bus. It can be realized.

[0152] The communication interface 1730 is a wired / wireless interface of the computing device 1700. The communication interface 1730 can also support Internet communication. It is also possible to support various communication methods other than internet communication. The interface 1730 includes a communication module that is well known in the art of the present disclosure. obtain.

[0153] Storage 1750 may non-temporarily store one or more computer programs 1760. The storage 1750 is a ROM (Read Only Memory), EPROM(Erasable Programmable ROM), EEPROM( Electrically Erasable Programmable ROM), Non-volatile memory such as flash memory, hard disks, removable disks, and The present disclosure relates to a computer readable program that can be used in a computer program that is capable of executing the program. The recording medium may include a recording medium.

[0154] When the computer program 1760 is loaded into the memory 1740, the processor 17 20. One or more instruments for performing operations / methods according to various embodiments of the present disclosure. That is, the processor 1 720 executes one or more instructions to implement various embodiments of the present disclosure. The operation / method can be performed.

[0155] The above flow chart and the above description are merely examples, and in some embodiments, different For example, in some embodiments, the order of the steps may be changed or some steps may be omitted. Steps may be repeated, some steps may be omitted, or some steps may be added.

[0156] The above-mentioned method is stored in a computer-readable medium for execution by a computer. The medium may be provided with a computer program stored thereon. It is used to store programs continuously or temporarily for execution or download. The medium may also be a variety of recording media in the form of a single piece of hardware or in the form of a combination of several pieces of hardware. A medium that can be a means or storage means, but is directly connected to any computer system The media is not limited to the above, and may be distributed on a network. , magnetic media such as hard disks, flexible disks and magnetic tapes, CD-R Optical recording media such as OM and DVD, floptical disks magneto optical medium such as a disk, and ROM, RAM, flash memory, etc., in which program instructions are stored. Another example of a medium is a medium for distributing an application. App stores and other sites and servers that supply or distribute various software Also included are recording media or storage media managed by the above.

[0157] The methods, acts, or techniques of the present disclosure may be embodied in various ways. For example, Such techniques may be hardware, firmware, software, or a combination of these. Various exemplary logical blocks described in connection with the disclosure of this application may be embodied in Modules, circuits, and algorithmic stages are electronic hardware, computer software, and It will be understood by those skilled in the art that the present invention may be embodied in a combination of the above-mentioned hardware and software. To clearly account for such substitutions of hardware and software, In the drawings, various exemplary components, blocks, modules, circuits, and steps are illustrated according to their functions. The above has been described generally from the perspective of whether such functionality is embodied in hardware. Whether the technology is implemented in software or not depends on the specific application and the overall system. The design requirements for each system vary. The typical engineer must determine While the described functionality may be implemented in a variety of ways for various applications, such implementations are within the scope of the present disclosure. This document should not be construed as a departure from the scope of the disclosure.

[0158] In a hardware implementation, the processing unit used to perform the techniques is a combination of one or more ASICs, DSPs, digital signal processing devices, l signal processing devices (DSPD), programmable programmable logic devices (PLDs), Field programmable gate array te arrays (FPGA), processor, controller, microcontroller, microprocessor processor, electronic device, or other electronic unit designed to perform the functions described in this disclosure. The present invention may be implemented in a system, a computer, or a combination thereof.

[0159] Accordingly, the various exemplary logic blocks, modules, and The circuits may be implemented using general purpose processors, DSPs, ASICs, FPGAs or other programmable logic devices. devices, discrete gate or transistor logic, discrete hardware components, or Any combination of components designed to perform the functions described herein may be embodied. A general purpose processor may be a microprocessor, but alternatively may be a Thus, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be a combination of computing devices, e.g. ,DSP and microprocessor, multiple microprocessors, one linked with DSP core It may be embodied in a combination of the above microprocessors, or in any other configuration.

[0160] In firmware and / or software implementations, the techniques include random access Random access memory (RAM), read-only memory ( read-only memory (ROM), non-volatile RAM (non-volatile random access memory: NVRAM), PROM (prog rammable read-only memory), EPROM (erasabl e programmable read-only memory), EEPROM( electrically erasable PROM), flash memory, compact disc (CD), magnetic or optical data storage device The method may be embodied in instructions stored on a computer-readable medium, such as a The instructions may be executable by one or more processors and may be , may be adapted to perform certain aspects of the functionality described in this disclosure.

[0161] When implemented in software, the techniques described above may be implemented by a computer using one or more instructions or codes. The information stored on or transmitted through a computer readable medium A computer readable medium may be transmitted from one place to another. Computer storage media, including any medium that facilitates the transmission of computer programs; Storage media includes both computer-accessible and non-commercially available media. As a non-limiting example, such a computer-readable medium may be The body may store RAM, ROM, EEPROM, CD-ROM or other optical disk storage, Magnetic disk storage or other magnetic storage device, or a desired program It can be used to transport or store code in the form of instructions or data structures, and can be used to Also, any connection may include any other medium that can be accessed by the computer. For use on any suitable computer readable medium.

[0162] For example, the software can be used to connect coaxial cables, fiber optic cables, stranded wires, and digital subscribers. Digital subscriber lines (DSL), or wireless technologies such as infrared, radio, and microwave When transmitted from a remote site, server, or other source, the signal may travel over coaxial cable, optical fiber, cable, twisted wire, digital subscriber line, or other types of communication such as infrared, radio, and microwave. Such wireless technologies are included within the definition of media. Disc is a type of media that includes CD, laser disc (registered trademark), optical disc, DVD (dig ital versatile disc, flexible disc, and Blu-ray Disks are usually used to reproduce data magnetically, whereas disks are used to reproduce data magnetically. Discs use lasers to optically reproduce data. Also included within the scope of computer-readable media.

[0163] The software module can store RAM memory, flash memory, ROM memory, and EPR memory. OM memory, EEPROM memory, register, hard disk, portable disk, CD- It may reside in a ROM, or any other form of storage medium known in the art. enables a processor to read information from and record information on a storage medium, The storage medium may be coupled to the processor. In the alternative, the storage medium may be integral to the processor. The processor and the storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In the alternative, the processor and the storage medium may reside as discrete components in a user terminal. This is also fine.

[0164] The above-described embodiments may be implemented in accordance with the presently disclosed subject matter on one or more stand-alone computer systems. Although described as utilizing the subject aspect, the present disclosure is not limited thereto and may be used in conjunction with a network. and distributed computing environments. Thus, aspects of the subject matter of this disclosure may be embodied in multiple processing chips or devices. Storage may be affected across multiple devices in the same way. Such devices may include PCs, network servers, and handheld devices.

[0165] Although the present disclosure has been described in relation to some embodiments herein, the present disclosure relates to a technology Various modifications and variations within the scope of the present disclosure that are understandable to those of ordinary skill in the art may be made. Such variations and modifications are within the spirit and scope of the appended claims. It should be understood that this falls within the scope. [Explanation of symbols]

[0166] 110 Information Processing Systems 120 First Stock Exchange 130 Second Stock Exchange

Claims

1. 1. A method of high frequency securities trading performed by at least one processor, comprising: the at least one processor includes a first processor for preprocessing and a second processor for a machine learning model, the first processor and the second processor being separate pieces of hardware; receiving market data for a target item for a first unit of time; calculating a delay time for a securities order for each of a plurality of candidate batch sizes; selecting a batch size from the plurality of candidate batch sizes based on the calculated delay time; generating, by the first processor, input data corresponding to the selected batch size using market data for the target item; transmitting, by the first processor, the input data to the second processor; receiving, by the second processor, the input data; generating, by the second processor, prediction data for the target item at a future time point related to the selected batch size and future to the first unit time point based on the generated input data using a machine learning model; generating order data for the target item based on the generated forecast data.

2. The step of selecting a batch size comprises:

2. The method of claim 1, further comprising selecting, for each of a plurality of predetermined future times, the largest candidate batch size from among the plurality of candidate batch sizes whose end point of the calculated delay time precedes each of the plurality of predetermined future times.

3. The step of selecting a batch size comprises: calculating an expected profit for each selected batch size for each of the plurality of predetermined future time points; 3. The method of claim 2, further comprising: selecting, from among the batch sizes selected for each of the plurality of predetermined future times, a batch size that has the highest calculated expected profit.

4. The step of calculating the expected profit comprises:

4. The method of claim 3, further comprising: calculating, for each of the plurality of predetermined future time points, an expected return for each of the selected batch sizes based on the selected batch sizes, a return per query, and a computation time of the machine learning model for each of the selected batch sizes.

5. The step of calculating the delay time comprises:

2. The method of claim 1, further comprising: calculating the delay time for each of the plurality of candidate batch sizes based on at least one of a data rate, a bandwidth of input / output data between the first processor and the second processor, a size of the input / output data, a calculation speed of the machine learning model by the second processor, a processing speed of the first processor, or a busy state of the second processor.

6. The delay time is 6. The high-frequency securities trading method of claim 5, wherein the time taken for the market data to be pre-processed by the first processor, the time taken for the pre-processed data to be transmitted from the first processor to the second processor, the time taken for the second processor to complete the calculation of the machine learning model, the time taken for the second processor to transmit the calculation result from the second processor to the first processor, and the time taken for the first processor to generate the order data based on the calculation result.

7. The method further includes obtaining data precision of the input data; The step of calculating the delay time includes:

2. The method of claim 1, further comprising: calculating a delay time for each of the plurality of candidate batch sizes based on accuracy of the input data.

8. 8. The method of claim 7, wherein the delay time is calculated based on processing factors that increase as data accuracy of the input data decreases.

9. The step of selecting the batch size comprises:

8. The method of claim 7, further comprising selecting a batch size from among the plurality of candidate batch sizes based on data accuracy of the acquired input data.

10. The batch size is:

10. The method of claim 9, wherein the input data is selected based on increasing processing factors as the data accuracy of the input data decreases.

11. the second processor includes a first dedicated accelerator and a second dedicated accelerator for processing operations of a machine learning model; The step of calculating the delay time includes: calculating a first delay time including a calculation time of the first dedicated accelerator and a second delay time including a calculation time of the second dedicated accelerator for each of the plurality of candidate batch sizes; The step of selecting a batch size comprises:

2. The method of claim 1, further comprising the step of selecting, for each of a plurality of predetermined future times, the largest candidate batch size among the calculated first and second delay times, the end point of the shorter delay time preceding each of the plurality of predetermined future times.

12. A computer-readable recording medium storing a program for executing the method according to claim 1 on a computer.

13. A method for programming a programmable logic device, comprising: a first memory for storing one or more instructions; at least one processor including a first processor for pre-processing and a second processor for a machine learning model, the first processor and the second processor being separate pieces of hardware, configured to execute one or more instructions in the first memory to receive market data for a target item during a first unit time, calculate a latency time for placing a securities order for each of a plurality of candidate batch sizes, select a batch size from the plurality of candidate batch sizes based on the calculated latency time, generate input data corresponding to the selected batch size using the market data for the target item by the first processor, and provide the input data to the second processor by the first processor; a second memory storing one or more instructions; a second processor configured to receive the provided input data, utilize a machine learning model to generate forecast data for the target item at a future time associated with the selected batch size based on the generated input data, and provide the generated forecast data to the at least one processor by executing one or more instructions in the second memory; The at least one processor: a high frequency securities trading system further configured to generate order data for the target item based on the forecast data provided by the second processor.

14. Calculating the delay time comprises:

14. The high-frequency securities trading system of claim 13, further comprising: calculating, for each of the plurality of candidate batch sizes, the delay time based on at least one of a data rate, a bandwidth of input / output data between the first processor and the second processor, a size of the input / output data, a calculation speed of the machine learning model by the second processor, a processing speed of the first processor, or a busy state of the second processor.

15. The delay time is 15. The high frequency securities trading system of claim 14, wherein the time taken for the market data to be pre-processed by the first processor, the time taken for the pre-processed data to be transmitted from the first processor to the second processor, the time taken for the second processor to complete the operation of the machine learning model, the time taken for the second processor to transmit the operation result from the second processor to the first processor, and the time taken for the first processor to generate the order data based on the operation result.

16. The at least one processor: further configured to obtain data precision of the input data; Calculating the delay time 14. The high frequency securities trading system of claim 13, further comprising: calculating a delay time for each of the plurality of candidate batch sizes based on accuracy of the input data.

17. 17. The high frequency securities trading system of claim 16, wherein the delay time is calculated based on processing factors that increase as data accuracy of the input data decreases.

18. The step of selecting the batch size comprises:

17. The high frequency securities trading system of claim 16, further comprising selecting a batch size from among the plurality of candidate batch sizes based on data accuracy of the acquired input data.

19. A high frequency securities trading system as described in claim 18, wherein the batch size is selected based on processing elements that increase as data accuracy of the input data decreases.

20. the second processor includes a first dedicated accelerator and a second dedicated accelerator for processing operations of a machine learning model; calculating the delay time includes calculating, for each of the plurality of candidate batch sizes, a first delay time including a calculation time of the first dedicated accelerator and a second delay time including a calculation time of the second dedicated accelerator; selecting the batch size 14. The high frequency securities trading system of claim 13, further comprising: selecting, for each of a plurality of predetermined future times, the largest candidate batch size among the calculated first and second delay times, the end point of the shorter delay time preceding each of the plurality of predetermined future times.

Images