KR20210063767A - Method for learning deep-learning model for stock trading - Google Patents

Abstract

The present invention relates to a deep learning model learning method for stock trading and comprises: a step of collecting transaction data for at least one item; a step of inputting the collected transaction data for at least one item into a first deep learning model based on supervised learning and learning the first deep learning model to predict a stock price change rate for the item for each preset unit time; a step of inputting the collected transaction data for at least one item to a second deep learning model and learning the second deep learning model to derive an order execution strategy that reduces slippage costs and order execution costs for an ultra-short-term market situation in the corresponding item; when a reward for the order execution strategy is determined according to the order execution strategy in a real stock market environment, a step of re-learning the second deep learning model based on the determined reward; and a step of re-learning the first deep learning model using the transaction data of an item for which order execution has been completed according to the order execution strategy in the real stock market environment. Accordingly, it is possible to minimize the loss caused by the execution of a large number of orders.

Description

How to learn a deep learning model for stock trading {METHOD FOR LEARNING DEEP-LEARNING MODEL FOR STOCK TRADING }

The present invention relates to a method for learning a deep learning model for stock trading.

A securities company receives a large amount of securities orders from institutions such as asset management companies, executes securities transactions, and receives transaction fees.

In this case, if a brokerage company sells or buys a large number of entrusted securities orders at once, the impact of a sharp drop or surge in the stock price of the stock market occurs, making it difficult to trade at an appropriate price.

For this reason, securities firms use an algorithmic trading system that divides and executes a large number of securities orders.

However, it was difficult to properly respond to the changing stock market because the existing algorithm and trading system divides and executes orders according to pre-set rules.

For example, in the case of a trading system using the Time Weighted Average Price (TWAP) algorithm, the adaptability and efficiency to the market are inevitably reduced because orders are executed in an evenly distributed quantity over time.

Korean Patent Publication No. 1808259 (Registered on Dec. 6, 2017)

The present invention is to solve the problems of the prior art, collect transaction data for at least one item, and input the collected transaction data for at least one item into a first deep learning model based on supervised learning Thus, the first deep learning model is trained to predict the stock price change rate for the item for each preset unit time, and the collected transaction data for at least one item is input to the second deep learning model to predict the stock price change in the item. We want to train the second deep learning model to derive an order execution strategy that reduces slippage costs and order execution costs for ultra-short market conditions.

In addition, the present invention retrains the second deep learning model based on the determined reward when the reward for the order execution strategy is determined according to the order execution strategy in the real stock market environment, and the real stock market environment to retrain the first deep learning model by using the transaction data of the item for which order execution has been completed according to the order execution strategy.

However, the technical problem to be achieved by the present embodiment is not limited to the technical problems as described above, and other technical problems may exist.

As a technical means for achieving the above technical problem, according to an embodiment of the present invention, in a deep learning model learning method for stock trading in a server, collecting transaction data for at least one stock; inputting the collected transaction data for at least one item into a first deep learning model based on supervised learning and learning the first deep learning model to predict the stock price change rate for the item for each preset unit time; By inputting the collected transaction data for at least one item into a second deep learning model, an order execution strategy that reduces slippage cost and order execution cost for the ultra-short-term market situation in the relevant item is derived. 2 The step of learning the deep learning model, and when the reward for the order execution strategy is determined according to the order execution strategy in the environment of the real stock market, re-learning the second deep learning model based on the determined reward and re-learning the first deep learning model by using the transaction data of the item for which order execution has been completed according to the order execution strategy in an actual securities market environment.

According to another embodiment of the present invention, in the method for learning a deep learning model for stock trading in a server, the reference stock price change rate of the first stock for each preset unit time has a first reference stock price change rate at a first time point, and a second time point In the case of having a second reference stock price change rate and an N-th reference stock price change rate at the N-th time point, the first dip at the first time point is output as the transaction data for the first item is input to the first deep learning model. Learning the first deep learning model so that the first stock price change rate of the learning model coincides with the first reference stock price change rate, and the second stock price change rate of the first deep learning model at the second time point outputted thereafter is the second reference stock price Training the first deep learning model to match the rate of change, and training the first deep learning model so that the rate of change of the Nth stock price of the first deep learning model at the output Nth time coincides with the rate of change of the Nth reference stock price. It may further include a step, and may further include the step of determining a reward for the order execution strategy in the reinforcement learning environment, and learning the second deep learning model so that the determined reward is maximized.

The above-described problem solving means are merely exemplary and should not be construed as limiting the present invention. In addition to the above-described exemplary embodiments, there may be additional embodiments described in the drawings and detailed description of the invention.

According to any one of the above-described problem solving means of the present invention, the present invention collects transaction data for at least one item, and uses the collected transaction data for at least one item as a first deep learning model based on supervised learning. The first deep learning model is trained to predict the stock price change rate for the item for each preset unit time by inputting it into the , and the collected transaction data for at least one item is input to the second deep learning model to the corresponding item. The second deep learning model is trained to derive an order execution strategy that reduces slippage costs and order execution costs for ultra-short market conditions in When the reward for the order execution strategy is determined, the second deep learning model is re-trained based on the determined reward, and in the environment of the actual stock market, using the transaction data of the item for which the order execution has been completed according to the order execution strategy By re-learning the first deep learning model, it is possible to derive an appropriate order execution strategy for the current market because a plurality of deep learning models are continuously learned based on market data. In addition, the present invention can minimize the loss due to the execution of a large number of orders by deriving the best order execution strategy according to market changes.

1 is a block diagram of an order execution server, according to an embodiment of the present invention.
2 is a diagram for explaining a method of learning a second deep learning model based on reinforcement learning, according to an embodiment of the present invention.
3 is a view for explaining a method of deriving an order execution strategy for the item in the current cycle, according to an embodiment of the present invention.
4 is a diagram illustrating a method of executing an order execution for a stock for a stock transaction, according to an embodiment of the present invention.
5 is an operation flowchart illustrating a method of performing order execution, according to an embodiment of the present invention.
6 is a diagram illustrating an interface for order execution, according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art can easily implement the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and similar reference numerals are attached to similar parts throughout the specification.

Throughout the specification, when a part is said to be "connected" with another part, this includes not only "directly connected" but also "electrically connected" with another element interposed therebetween. . In addition, when a part "includes" a certain component, it means that other components may be further included rather than excluding other components unless specifically stated to the contrary.

In the present specification, the term "unit" includes a unit realized by hardware, a unit realized by software, and a unit realized using both. Further, one unit may be realized by using two or more hardware, or two or more units may be realized by one piece of hardware.

In this specification, some of the operations or functions described as being performed by the terminal or device may be performed instead in a server connected to the terminal or device. Likewise, some of the operations or functions described as being performed by the server may also be performed by a terminal or device connected to the server.

Hereinafter, with reference to the accompanying configuration diagram or processing flow chart, it will be described in detail for the implementation of the present invention.

1 is a block diagram of an order execution server 10, according to an embodiment of the present invention.

Referring to FIG. 1 , the order execution server 10 may include a data collection unit 100 , an auxiliary prediction value generation unit 110 , a model generation unit 120 , and an order execution unit 130 . Here, the order execution unit 130 may include an order execution strategy derivation unit 132 and an order execution instruction unit 134 . However, the order execution server 10 shown in FIG. 1 is only one implementation example of the present invention, and various modifications are possible based on the components shown in FIG. 1 .

The model generator 120 may generate a first deep learning model based on supervised learning and train the generated first deep learning model. Here, the first deep learning model based on supervised learning may be, for example, a deep neural network model composed of an attention network, a recurrent neural network, and the like.

The model generating unit 120 may input the collected transaction data for at least one item into the first deep learning model and train the first deep learning model to predict the stock price change rate for the item for each preset unit time. have.

Here, the transaction data for at least one item that is input to the first deep learning model may include, for example, transaction price information for the item traded in a securities institution, asking price data, macroeconomic indicator data, and the like.

For example, the model generating unit 120 predicts the rate of change of the stock price in which the output value of the first deep learning model for at least one item per unit time coincides with the reference rate of change of the stock price of the item per unit time which is preset. can learn

To this end, the model generation unit 120 is configured such that the mean squared error (MSE, Mean Squared Error) between the preset standard stock price change rate of the item per unit time and the stock price change rate of the item per unit time, which is the output value of the first deep learning model, is minimized. The first deep learning model may be trained.

For example, the preset reference stock price change rate of the first stock per unit time has the first reference stock price change rate at the first time point (eg, 1 hour later), and the second reference stock price at the second time point (eg, 2 hours later) Assuming that it has a rate of change and has an Nth reference share price change rate at an N th time point (eg, N hours later), the model generator 120 as the transaction data for the first item is input to the first deep learning model The first deep learning model is trained so that the rate of change of the first stock price of the first deep learning model at the output first time point coincides with the first reference rate of change of the stock price, and then the first deep learning model at the second time point is output. 2 The first deep learning model is trained so that the rate of change of stock price coincides with the rate of change of the second reference stock price, and the first deep learning model is then output so that the rate of change of the N th stock price of the first deep learning model at the N th time is output to coincide with the rate of change of the N th standard stock price. Deep learning models can be trained.

The model generator 120 is a reinforcement learning-based second deep including two or more actors, which are neural networks that determine the behavioral rules of the reinforcement learning agent, and Critic, which is a neural network that estimates the behavioral values of the reinforcement learning agents. A learning model may be generated, and the generated second deep learning model may be trained.

The model generator 120 is a reinforcement learning-based second deep learning model based on the collected transaction data for at least one item so that the reinforcement learning-based second deep learning model derives an order execution strategy for at least one item. model can be trained.

The model generation unit 120 inputs the collected transaction data for at least one item into the second deep learning model, so that the slippage cost and the order execution cost for the ultra-short-term market situation in the item are reduced. A second deep learning model may be trained to be derived.

Here, the transaction data for at least one item that is input to the second deep learning model includes, for example, bid window data, market price data (transaction price and transaction volume), long-run time information, order execution amount, etc. for the item. may include

The model generator 120 may determine a reward for the order execution strategy (strategy derived from the second deep learning model) in the reinforcement learning environment, and train the second deep learning model so that the determined reward is maximized. Here, the reinforcement learning environment is a model created by virtualizing the stock market environment. In addition, the reward may include information on whether the order execution was successful for the order execution strategy and the amount of the order execution reduced compared to the volume weighted average transaction (VWAP).

Referring to Figure 2, the second deep learning model 20 is two or more actors (201, 203, 205), which are neural networks that determine the behavioral rules of the reinforcement learning agent, and the neural network that estimates the behavioral values of the reinforcement learning agents, crit ( 209, 211).

Here, the two or more actors 201, 203, and 205 include a first actor 201 that determines an order quantity for at least one item and a second actor 203 that determines an order cancellation amount for at least one item. can do.

The two or more actors 201 , 203 , 205 are configured to determine a final order quantity for at least one stock based on the order quantity determined through the first actor 201 and the order cancellation quantity determined through the second actor 203 . It may include actors 205 .

The model generator 120 may train two or more actors 201 , 203 , and 205 to improve a reward for the order execution strategy from the reinforcement learning environment based on the transaction data.

For example, the model generation unit 120 trains the first actor 201 to determine the order quantity for at least one item, and uses the second actor 203 to determine the order cancellation amount for the order quantity for the item. can learn

Here, if the order quantity determined through the first actor 201 and the order cancellation quantity determined through the second actor 203 are added together, a zero-centered structure appears as a result. This structure helps when learning the determination of the pure order quantity. can give

The model generator 120 may train the third actor 205 to determine the net order amount based on the order amount determined through the first actor 201 and the order cancellation amount determined through the second actor 203 .

The model generation unit 120 is configured to derive an order execution strategy for at least one item in the current period based on the net order quantity determined by the third actor 205 (that is, a bid window distribution strategy for the net order quantity). 3 Actors 205 can be trained.

Thereafter, the model generator 120 may perform order execution in the reinforcement learning environment 207 according to the order execution strategy for at least one item in the current period and the net order quantity determined by the third actor 205 . .

Meanwhile, the critiques 209 and 211 estimate a state value in order to estimate the final order amount of the plurality of actors 201 , 203 , and 205 in the reinforcement learning environment 207 and a reward for the order execution strategy result. A second crit 211 for updating the corresponding reward may be included by applying a reward averaging method (eg, polyak-ruppert averaging) to the calculated reward in order to reliably estimate it.

Here, the reason for using the reward averaging method is to help the more stable reward estimation by introducing a concept such as a moving average because the reward estimation of the first click 209 is unstable.

The model generation unit 120 may generate an agent model including a plurality of actors 201 , 203 , 205 and a plurality of critiques 209 and 211 .

The model generator 120 may learn the first crit 209 and the second crit 211 .

The model generator 120 may learn the agent model through interaction between the agent model and the environment. Here, the environment is an environment that provides a reward for the action result of the agent model.

The model generator 120 may store result information (eg, the state of the agent model, behavior, reward, etc.) through interaction between the agent model and the environment in the replay buffer.

The model generator 120 may learn the agent model by extracting result information stored in the replay buffer.

In order to estimate a state value for a specific state, the model generator 120 generates a first click 209 and a second click so that when a specific state is input, the difference from the expected state value in the specific state is minimized. (211) can be learned.

The Q-network 213 may estimate the action value of the clicks 209 and 211 based on the reward to which the reward averaging method is applied. The action values of the crit 209 and 211 estimated in this way can be used for learning by a plurality of actors 201 , 203 , 205 .

Critics 209 and 211 can learn in an off-policy form using the Q-network 213 . At this time, the Q-network 213 also performs learning.

For example, the Q-network 213 may learn to generate a value similar to the behavioral value of the bootstrapped clicks 209 and 211 from the reward to which the reward averaging method is applied.

In this case, the Q-network 213 may learn in a way that a mean squared error (MSE) is minimized. Here, the action value of the crit 209 and 211 may be calculated as the sum of the reward and the state value.

The model generator 120 may train two or more actors 201 , 203 , and 205 to improve the action value estimated from the Q-network 213 .

1 and 3 together, the data collection unit 100 may collect transaction data for at least one item. Here, the transaction data for at least one item may include, for example, bid window data for the item, market price data (transaction price and transaction volume), market run time information, order execution amount, and the like.

The auxiliary predictive value generator 110 may generate an auxiliary predicted value for the item by inputting transaction data for at least one item collected in the first deep learning model 30 based on pre-trained supervised learning. Here, the auxiliary prediction value may be, for example, a volume curve. A volume curve is a graph showing the total volume of trading over time during the day. For example, the auxiliary predictive value generator 110 may generate an auxiliary predictive value in which the risk propensity of the customer is reflected through the first deep learning model 30 to which transaction data is input as an input value. Here, the auxiliary prediction value may be a graph (volume curve) showing the total amount of trading to be traded during the day over time.

For example, after the transaction data for at least one item is input to the first deep learning model, when the rate of change of the stock price for the item for each preset unit time is predicted from the first deep learning model, the auxiliary prediction value generator ( 110) may generate an auxiliary predicted value based on the predicted stock price change rate for the corresponding item for each predetermined unit time.

For example, the auxiliary prediction value generating unit 110 responds to the volume-weighted average transaction (VWAP) by reflecting the stock price change rate and the risk tendency information of the customer for each predetermined unit time predicted with respect to the stock price direction of the day. A volume curve, which is an optimal auxiliary prediction value, can be derived by generating a VWAP curve that is

On the other hand, if the order execution is performed according to the volume curve generated by the auxiliary prediction value generating unit 110, there is a high possibility that the order transaction is made at an unfavorable price in a short time.

That is, when orders are executed simply according to the volume curve, it is difficult to trade at the market price or respond to changes in market conditions in a smaller time unit, resulting in an increase in total order cost.

In the present invention, in order to solve this problem, an order execution strategy (eg, a bid window allocation strategy for the total order quantity) is derived through a second deep learning model based on reinforcement learning so that the total order cost is minimized.

That is, according to the present invention, using the second deep learning model based on reinforcement learning, it is possible to optimize the slippage cost and the order execution cost for the ultra-short market situation even for a micro-order execution strategy.

The order execution strategy derivation unit 132 is a reinforcement learning-based second deep learning model 20 trained in advance on transaction data for at least one item collected together with the auxiliary prediction value generated from the first deep learning model 30 . After input to , it is possible to derive an order execution strategy based on the transaction data and auxiliary prediction values from the second deep learning model 20 .

The order execution strategy derivation unit 132 may derive an order execution strategy for at least one item in the current period through the second deep learning model 20 based on the transaction data and the auxiliary prediction value.

For example, the order execution strategy derivation unit 132 determines the order amount for the item based on the auxiliary prediction value through the first actor of the second deep learning model 20, and through the second actor to the item based on the auxiliary prediction value. You can decide the amount of order cancellation for

The order execution strategy derivation unit 132 may determine the final order amount through the third actor based on the order amount for the item determined from the first actor and the order cancellation amount for the item determined from the second actor.

The order execution strategy derivation unit 132 may derive an order execution strategy for the item in the current cycle according to the final order amount determined through the third actor.

Here, the order execution strategy may include at least one of a cumulative order amount, an additional order amount (additional buy/sell order or cancel order) in the current cycle compared to the previous cycle, and ask window allocation information for the additional order amount in the current cycle. Here, the cumulative order amount and the additional order amount may be the final order amount determined by the third actor.

That is, the order execution strategy may include a cumulative order amount or additional order amount and a call window allocation strategy for items that change according to market conditions.

The order execution strategy derivation unit 132 may derive an order execution strategy for distributing the accumulated order amount or the additional order amount to a plurality of quotation windows (including market prices).

Also, the order execution strategy derivation unit 132 may modify the order execution strategy according to a change in market conditions in the current cycle.

For example, the order execution strategy derivation unit 132 may increase or cancel the order for each quotation window based on the accumulated order amount or the additional order amount and the quotation window distribution information.

The present invention can derive an order execution strategy that maximizes future current discounted rewards for at least one item through the reinforcement learning-based second deep learning model 20 .

For example, suppose an order is placed to buy V shares before time T. If the amount of execution to date is v and the elapsed time is time t, the order execution strategy derivation unit 132 is an order to minimize the cost of purchasing the remaining (Vv) shares from the current situation to the T point compared to the V-WAP (VWAP) Implementation strategies can be derived.

In addition, the present invention can provide an effect of minimizing the order execution amount at the present time by optimizing the balance between the short-term compensation according to the order execution strategy and the long-term compensation in the future.

The order execution unit 130 may instruct execution of an order for at least one item in a current cycle by using order information including an order execution strategy.

In addition, the order execution instructing unit 134 may instruct execution of orders for at least one item in the current period by using order information including an order execution strategy modified according to a change in market conditions at the current time.

The order execution instructing unit 134 may instruct the execution of orders for at least one item in the current period by transmitting order information including the order execution strategy to the stock exchange operation server.

Thereafter, the model generator 120 determines a reward for the order execution strategy executed according to the order execution strategy in the environment of the real stock market, and re-learns the second deep learning model 20 based on the determined reward. , it is possible to retrain the first deep learning model using the transaction data of the item for which the order execution has been completed.

Meanwhile, for those skilled in the art, the data collection unit 100 , the auxiliary prediction value generation unit 110 , the model generation unit 120 , the order execution unit 130 , the order execution strategy derivation unit 132 , and the order execution instruction unit 134 are each It will be fully understood that it may be implemented separately, or one or more of these may be implemented integrally.

4 is a diagram illustrating a method of performing order execution for a stock for a stock transaction, according to an embodiment of the present invention.

Referring to FIG. 4 , in step S401 , the order execution server 10 may collect transaction data for at least one item.

In step S403, the order execution server 10 generates a reinforcement learning-based deep learning model including two or more actors, which are neural networks that determine the behavioral rules of the reinforcement learning agent, and Critic, which is a neural network that estimates the behavioral value of the reinforcement learning agent. can do.

In step S405, the order execution server 10 may learn the reinforcement learning-based deep learning model based on the transaction data so that the reinforcement learning-based deep learning model derives an order execution strategy for at least one item.

In step S407, the order execution server 10 may perform order execution for at least one item in the current period by using the order information including the order execution strategy.

In the above description, steps S401 to S407 may be further divided into additional steps or combined into fewer steps, according to an embodiment of the present invention. In addition, some steps may be omitted as necessary, or the order between steps may be changed.

5 is an operation flowchart illustrating a method of performing order execution, according to an embodiment of the present invention.

Referring to FIG. 5 , in step S501 , the user terminal 50 may receive order data for a specific item from the user. Here, the order data may include order type information for a specific item, order deadline information, order quantity information, order code information, and order restriction conditions. For example, the user terminal 50 may receive order data for a specific item through the order setting interface of FIG. 6 . The user terminal 50 provides an account number 601, order deadline information (not shown), order code information 603, order type information 605, order quantity information 607 for a specific item through the order setting interface, Order constraint condition information 609 may be set by the user. Here, the order deadline information (not shown) includes an order deadline for selling or buying a specific item, and the order code information 603 includes code information (eg, code information of a Samsung Electronics item) corresponding to the specific item. can The order type information 605 may include selection information on whether to sell or buy a specific item, and the order quantity information 607 may include an order amount of specific information according to the selected order type information. The order constraint condition 609 may include, for example, a TWAP order constraint range.

In step S503, the user terminal 50 may transmit order data for a specific item to the order execution server 10 .

In step S505, the order execution server 10 may request the stock data providing server 60 for stock data for a specific item at the time of cycle 0. Here, the stock data may include bid window data, transaction tick data, and the like.

In step S507, the order execution server 10 may receive the stock data for a specific item at the time of 0 cycle from the stock data providing server 60.

In step S509, the order execution server 10 may derive the order execution strategy of the 0 cycle time based on the received order data for the specific item and the stock data for the specific item at the 0 cycle time point.

In step S511 , the order execution server 10 may transmit order information including the derived order execution strategy at the time of 0 cycle to the securities transaction management server 70 .

In step S513 , the order execution server 10 may receive execution result information on the order information at the time of cycle 0 from the securities transaction management server 70 .

In step S515 , the order execution server 10 may transmit the execution result information received from the securities transaction management server 70 to the user terminal 50 .

In step S517, the order execution server 10 may request the stock data providing server 60 for stock data for a specific item at the time of one cycle.

In step S519, the order execution server 10 may receive from the stock data providing server 60 securities data for a specific item at the 1st cycle time (the securities data reflecting the execution result of the order information at the 0th cycle time).

In step S521, the order execution server 10 may derive an order execution strategy at the time of one cycle based on the received order data for the specific item and the securities data for the specific item at the time of the first cycle.

In step S523 , the order execution server 10 may transmit order information including the derived order execution strategy at the time of one cycle to the securities transaction operation server 70 .

In step S525, when the stock exchange management server 70 completes all orders for the specific item according to the order quantity information included in the order data for the specific item input by the user terminal 50, the securities transaction management server 70 , it is possible to transmit final execution information for a specific item to the order execution server 10 . If the order of a specific item is not completed, the process after step S513 may be repeated until all orders are completed.

In step S527 , the order execution server 10 may transmit the final execution information received from the securities transaction management server 70 to the user terminal 50 .

In the above description, steps S501 to S527 may be further divided into additional steps or combined into fewer steps, according to an embodiment of the present invention. In addition, some steps may be omitted as necessary, or the order between steps may be changed.

An embodiment of the present invention may also be implemented in the form of a recording medium including instructions executable by a computer, such as a program module executed by a computer. Computer-readable media can be any available media that can be accessed by a computer, and includes both volatile and nonvolatile media, removable and non-removable media. Further, the computer-readable medium may include all computer storage media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.

The above description of the present invention is for illustrative purposes only, and those of ordinary skill in the art to which the present invention pertains will be able to understand that other specific forms can be easily modified without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are illustrative and non-limiting in all respects. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as being distributed may also be implemented in a combined form.

The scope of the present invention is indicated by the claims to be described later rather than the detailed description, and all changes or modified forms derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention. .

10: Order Fulfillment Server
100: data collection unit
110: auxiliary prediction value generator
120: model generation unit
130: order execution unit
132: Order execution strategy derivation unit
134: Order Execution Directive

Claims (7)

In a deep learning model learning method for stock trading in a server,
collecting transaction data for at least one item;
inputting the collected transaction data for at least one item into a first deep learning model based on supervised learning to train the first deep learning model to predict the stock price change rate for the item for each preset unit time;
By inputting the collected transaction data for at least one item into a second deep learning model, an order execution strategy that reduces slippage cost and order execution cost for the ultra-short-term market situation in the relevant item is derived. 2 training the deep learning model;
When the reward for the order execution strategy is determined according to the order execution strategy in the real stock market environment, re-learning a second deep learning model based on the determined reward;
Re-learning the first deep learning model by using the transaction data of the item for which order execution has been completed according to the order execution strategy in an actual stock market environment;
How to train a deep learning model for stock trading. The method of claim 1,
The transaction data for at least one item input to the first deep learning model is,
Including one or more of transaction price information, asking price data, and macroeconomic indicator data for stocks traded in a securities institution;
How to train a deep learning model for stock trading. The method of claim 1,
The transaction data for at least one item input to the second deep learning model is,
Including one or more quotation window data, market price data (transaction price and trading volume), market run time information, and order execution volume data for the item,
How to train a deep learning model for stock trading. The method of claim 1,
Further comprising the step of learning the first deep learning model to predict the rate of change of the stock price that the output value of the first deep learning model for at least one item per unit time coincides with the reference price change rate of the item per unit time which is set in advance,
How to train a deep learning model for stock trading. The method of claim 1,
The step of learning the first deep learning model so that the mean squared error (MSE, Mean Squared Error) between the preset standard stock price change rate of the stock per unit time and the stock price change rate of the stock per unit time output value of the first deep learning model is minimized more containing,
How to train a deep learning model for stock trading. The method of claim 1,
When the preset reference price change rate of the first stock per unit time has the first reference price change rate at the first time point, the second reference share price change rate at the second time point, and the Nth reference share price change rate at the Nth time point,
The first deep learning model is formed so that the first stock price change rate of the first deep learning model at the first time point output as the transaction data for the first item is input to the first deep learning model matches the first reference stock price change rate learning;
training the first deep learning model so that the second stock price change rate of the first deep learning model at the second time point outputted thereafter matches the second reference stock price change rate;
Further comprising the step of training the first deep learning model so that the Nth stock price change rate of the first deep learning model at the output Nth time point coincides with the Nth reference stock price change rate,
How to train a deep learning model for stock trading. The method of claim 1,
Further comprising the step of determining a reward for the order execution strategy in the reinforcement learning environment, and training the second deep learning model so that the determined reward is maximized,
How to train a deep learning model for stock trading.

Images