KR20210061598A - Method and apparatus for deep learning-based stock screening and its use for advancing and automating portfolio management - Google Patents

Abstract

A deep learning-based stock screening method, a portfolio automation and advancement method using the same, and a device thereof are provided. According to one embodiment, the deep learning-based stock screening method comprises the steps of: pre-processing to convert financial data into data suitable for deep learning, and removing noise contained in the financial time series using a noise-removing autoencoder; using the noise-removed data to input into a deep learning-based prediction model; predicting a stock price prediction return for each stock item through the trained prediction model; and setting a screening level, and selecting a stock having the stock price prediction return greater than the set screening level.

Description

Deep learning-based stock screening and portfolio automation and advancement method and device using it {METHOD AND APPARATUS FOR DEEP LEARNING-BASED STOCK SCREENING AND ITS USE FOR ADVANCING AND AUTOMATING PORTFOLIO MANAGEMENT}

The following examples relate to a deep learning-based stock screening and a portfolio automation and enhancement method and apparatus using the same.In more detail, the accuracy of the stock price prediction return is improved using deep learning, and based on the predicted return. It relates to how and how to build a portfolio.

Recently, as interest in artificial intelligence technology has been amplified, research on machine learning and deep learning has been actively conducted. In particular, in the fin-tech industry, innovations that can make a big difference to both existing financial services companies are taking place, focusing on payment systems, financial services, bank loan services, investment asset management, insurance and market infrastructure, etc.

When building an optimal portfolio (set of stocks), you need the predicted future returns of each stock. In the case of an average-variance portfolio, one of the representative portfolio methodologies, the future predicted return is used as an input variable, and the weight of each stock is derived according to the average-variance optimization formula, and the weight of the portfolio is greatly influenced by the input variable. Hence, accurate forward-looking returns play a key role in portfolio performance.

However, the conventional portfolio technology is estimated according to the subjective predictions of experts, but since it is extremely difficult to predict the stock price return, the performance of the portfolio is often poor.

Korean Patent Laid-Open Publication No. 10-2019-0064749 discloses a method and apparatus for supporting such intelligent securities investment decision making.

Korean Patent Publication No. 10-2019-0064749

The examples describe a deep learning-based stock screening and a portfolio automation and enhancement method and apparatus using the same, and more specifically, using deep learning to increase the accuracy of the stock price prediction return, and to build a portfolio based on the predicted return. Provide technology.

Examples are deep learning, which can improve the accuracy of prediction in a quantitative way using deep learning and build a portfolio with stocks (screening) with high predictive returns, thereby improving the portfolio automation performance, lowering the risk for investment and increasing the return rate. It is to provide a method and apparatus for automating and upgrading portfolios based on stock screening and using the same.

A deep learning-based stock screening method according to an embodiment includes the steps of: pre-processing to convert financial data into data suitable for deep learning, and removing noise contained in the financial time series using a noise-removing auto-encoder; Inputting the data from which noise has been removed into a prediction model based on deep learning; Predicting a stock price prediction yield for each stock item through the trained prediction model; And setting a screening level to select a stock having the stock price prediction yield greater than the set screening level.

In the selecting of a stock having the stock price prediction yield greater than the screening level, a portfolio may be constructed by selecting a stock having the stock price prediction yield greater than the screening level.

After selecting a stock having the stock price prediction yield greater than the screening level, the method may further include optimizing the constructed portfolio by deriving a portfolio proportion for each stock item through an average-variance optimization process.

It may further include analyzing the performance of the portfolio built using test data.

The financial data may be constructed using trading data and technical indicators of the KOSPI and foreign indices.

The step of removing noise contained in the financial time series using the noise removal autoencoder may be preprocessed by scaling the input financial data to a value between 0 and 1.

The step of removing the noise contained in the financial time series using the noise-removing auto-encoder comprises constructing a stacked noise removal auto-encoder by stacking hidden layers of a plurality of auto-encoders consisting of an input layer, a hidden layer, and an output layer, and the stacking noise removing auto High-level features can be extracted through the encoder.

The step of removing noise contained in the financial time series using the noise-removing autoencoder includes extracting features through the hidden layer using an autoencoder consisting of an input layer, a hidden layer, and an output layer, and converting the hidden layer to the next autoencoder. By using a method of extracting features through a hidden layer using as an input layer, the stacked noise removing auto-encoder may be constructed through a plurality of the auto-encoders, and high-level features may be extracted.

In the deep learning-based prediction model, the deep learning-based prediction model may be constructed using a feedforward deep neural network.

In the step of inputting the noise-removed data into the deep learning-based prediction model, the stacked noise removal auto-encoder (Stacked) on the input layer of a feedforward deep neural network (FNN), which is a deep learning-based prediction model. A hybrid model (SDAE-FNN) may be constructed by connecting a Denoising Autoencoder (SDAE), and the hybrid model (SDAE-FNN) may be trained using a backpropagation algorithm.

A deep learning-based stock screening apparatus according to another embodiment preprocesses financial data to convert data suitable for deep learning, and uses a noise reduction auto-encoder to remove noise contained in the financial time series. denial; A prediction model inputting into a deep learning-based prediction model using the noise-removed data; A return rate prediction unit that predicts a stock price prediction rate of return for each stock item through the trained prediction model; And a screening unit that sets a screening level and selects stocks having the stock price prediction yield greater than the set screening level.

The screening unit further includes an optimization unit for optimizing the constructed portfolio by selecting a stock having the stock price prediction yield greater than the screening level, and deriving a portfolio proportion for each stock item through an average-distribution optimization process, and the screening unit, A portfolio may be built by selecting stocks having the stock price prediction yield greater than the screening level.

It may further include a performance analysis unit for analyzing the performance of the portfolio built using test data.

The noise removal unit is configured to construct a stacked noise removal autoencoder by stacking hidden layers of a plurality of autoencoders consisting of an input layer, a hidden layer, and an output layer, and extract high-level features through the stacked noise removal autoencoder. I can.

The prediction model is a hybrid model (SDAE-FNN) by connecting the stacked denoising autoencoder (SDAE) to the input layer of a deep learning-based prediction model, a feedforward deep neural network (FNN). And training the hybrid model (SDAE-FNN) using a backpropagation algorithm.

According to embodiments, by using deep learning to improve prediction accuracy in a quantitative method and building a portfolio with stocks (screening) with a high predicted return, the portfolio automation performance can be improved to lower the risk for investment and increase the return. Deep learning-based stock screening and portfolio automation and enhancement methods and devices using the same can be provided.

1 is a block diagram illustrating a deep learning-based stock screening apparatus according to an embodiment.
2 is a flowchart illustrating a deep learning-based stock screening method according to an embodiment.
3 is a diagram for describing a method of constructing an autoencoder for removing stacked noise according to an exemplary embodiment.
4 is a diagram illustrating feature extraction of an autoencoder for removing stacked noise according to an exemplary embodiment.
5 is a diagram illustrating an example of a portfolio in which all stocks are constructed with the same weight according to an embodiment.

Hereinafter, embodiments will be described with reference to the accompanying drawings. However, the described embodiments may be modified in various different forms, and the scope of the present invention is not limited by the embodiments described below. In addition, various embodiments are provided to more completely describe the present invention to those of ordinary skill in the art. In the drawings, the shapes and sizes of elements may be exaggerated for clearer explanation.

The following examples relate to a deep learning-based stock screening and a portfolio automation and enhancement method and apparatus using the same, and a method of increasing the accuracy of the stock price prediction return using deep learning and building a portfolio based on the predicted return. to provide. More specifically, examples use deep learning to improve prediction accuracy in a quantitative way and build a portfolio with stocks (screening) with a high predicted return, thereby improving portfolio automation performance, lowering the risk for investment and increasing the return rate. .

1 is a block diagram illustrating a deep learning-based stock screening apparatus according to an embodiment.

Referring to FIG. 1, a deep learning-based stock screening apparatus 100 according to an embodiment includes a noise removal unit 110, a prediction model 120, a yield prediction unit 130, and a screening unit 140. I can. According to an embodiment, the deep learning-based stock screening apparatus 100 may further include an optimization unit 150 and a performance analysis unit 160. Here, the deep learning-based stock screening apparatus 100 according to an embodiment may be included or included in a deep learning-based stock screening and portfolio automation and enhancement apparatus using the same.

The noise removal unit 110 may pre-process the financial data 101 to convert data suitable for deep learning, and may remove noise contained in the financial time series using a noise removal auto-encoder.

Here, in order to build the stock price prediction model 120, the financial data 101 (i.e., input data) is the trading data (initial price, closing price, low price, high price, trading volume) and technical indicators of the KOSPI and foreign indices (S&P500 index). Can be used.

The noise removal unit 110 constructs a stacked noise removal autoencoder by stacking the hidden layers of a plurality of autoencoders consisting of an input layer, a hidden layer, and an output layer, and extracts high-level features through the stacked noise removal autoencoder. can do. This will be described in more detail below.

Data from which noise has been removed may be used to input into the deep learning-based prediction model 120. The prediction model 120 may use a feedforward deep neural network (FNN) or a long short-term memory.

In particular, a hybrid model (SDAE-FNN) is constructed by connecting a stacked denoising autoencoder (SDAE) to the input layer of the deep-learning-based predictive model 120, FNN, and backpropagation. A hybrid model (SDAE-FNN) can be trained using the (Backpropagation) algorithm.

The yield prediction unit 130 may predict the stock price prediction yield for each stock item through the trained prediction model 120.

The screening unit 140 may set a screening level to select stocks having a stock price prediction yield greater than the set screening level. Here, the screening unit 140 may build a portfolio by selecting a stock having a stock price prediction yield greater than the screening level.

The optimizer 150 may optimize the constructed portfolio by selecting a stock having a stock price prediction yield greater than the screening level in the screening unit 140, and deriving a portfolio proportion for each stock item through an average-variance optimization process.

The performance analysis unit 160 may analyze the performance of the portfolio built by using the test data.

For example, the deep learning-based stock screening apparatus according to the present embodiments may be implemented through a computer system. The computer system may include a processor, a memory, a persistent storage device, a bus, an input/output interface, and a network interface as components for executing the deep learning-based stock screening method. For example, a processor may include or be part of any device capable of processing any sequence of instructions. The processor may include, for example, a digital processor and/or a processor in a computer processor, mobile device or other electronic device. The processor may be included, for example, in a server computing device, a server computer, a series of server computers, a server farm, a cloud computer, a content platform, a mobile computing device, a smartphone, a tablet, a set-top box, a media player, and the like. The processor can be connected to the memory via a bus.

2 is a flowchart illustrating a deep learning-based stock screening method according to an embodiment.

Referring to FIG. 2, a deep learning-based stock screening method according to an embodiment preprocesses financial data to convert data suitable for deep learning, and removes noise contained in the financial time series using a noise removal autoencoder. Step (S110), inputting into a deep learning-based prediction model using data from which noise has been removed (S120), predicting a stock price prediction yield for each stock item through a trained prediction model (S130), and a screening level. By setting, it may include a step (S140) of selecting a stock having a stock price prediction yield greater than the set screening level.

In addition, after selecting a stock having a stock price prediction yield greater than the screening level, a step (S150) of optimizing the constructed portfolio by deriving a portfolio proportion for each stock item through an average-distribution optimization process may be further included.

In addition, a step (S160) of analyzing the performance of the portfolio built using the test data may be further included.

As described above, according to embodiments, it is possible to increase the accuracy of stock price return prediction by using deep learning, and to reduce the risk of investment and increase the return rate by constructing a threshold-based portfolio using the predicted stock price.

Each step of the stock screening method based on deep learning according to an embodiment will be described in more detail below.

The deep learning-based stock screening method according to an embodiment may be described in more detail with an example of the deep learning-based stock screening apparatus 100 according to an embodiment. The deep learning-based stock screening apparatus 100 according to an embodiment may include a noise removal unit 110, a prediction model 120, a yield prediction unit 130, and a screening unit 140, according to the embodiment. Accordingly, the deep learning-based stock screening apparatus 100 may further include an optimization unit 150 and a performance analysis unit 160.

In step S110, the noise removal unit 110 pre-processes the financial data to convert it into data suitable for deep learning, and may remove noise contained in the financial time series using a noise removal auto-encoder. More specifically, the noise removal unit 110 scales the input data to a value between 0 and 1 in order to convert data suitable for deep learning as a pre-processing operation, and removes noise contained in the financial time series using a noise removal auto-encoder. can do. Here, the financial data may be constructed using the trading data and technical indicators of the KOSPI and foreign indices.

The noise removal unit 110 may configure a stacked noise reduction autoencoder by stacking hidden layers of a plurality of autoencoders consisting of an input layer, a hidden layer, and an output layer, and extract high-dimensional features through the stacked noise reduction autoencoder. More specifically, the noise removal unit 110 extracts features through the hidden layer using an auto-encoder consisting of an input layer, a hidden layer, and an output layer, and extracts features through the hidden layer using the hidden layer as an input layer of the next auto-encoder. By using the method, it is possible to construct a multilayer noise reduction autoencoder through a plurality of autoencoders and extract features of a high-dimensional level.

In step S120, the prediction model 120 may input the data from which noise has been removed into the deep learning-based prediction model 120. Here, the deep learning-based prediction model 120 may be configured with a deep learning-based prediction model 120 using a forward deep artificial landscape. A hybrid model (SDAE-FNN) is constructed by connecting a stacked noise canceling auto-encoder (SDAE) to the input layer of the forward deep artificial network (FNN), which is a deep learning-based predictive model (120), and a hybrid model using a backpropagation algorithm. (SDAE-FNN) can be trained.

In step S130, the yield prediction unit 130 may predict the stock price prediction yield for each stock item through the trained prediction model 120. For example, the return rate prediction unit 130 may predict a stock price daily/weekly/monthly rate of return using the trained prediction model 120.

In step S140, the screening unit 140 may set a screening level to select a stock having a stock price prediction yield greater than the set screening level. Here, the screening unit 140 may build a portfolio by selecting a stock having a stock price prediction yield greater than the screening level.

In step S150, the optimizer 150 selects a stock having a stock price prediction rate greater than the screening level, and then derives a portfolio proportion for each stock item through an average-distribution optimization process, thereby optimizing the constructed portfolio.

In step S160, the performance analysis unit 160 may analyze the performance of the portfolio built using the test data. Here, the performance analysis unit 160 may analyze the performance of the portfolio by analyzing the cumulative return and volatility of the portfolio built using the test data.

Hereinafter, a deep learning-based stock screening method according to an embodiment will be described in more detail through an example.

The key goal of these embodiments is to effectively combine stock market data from different time zones. To build a predictive model, you can select 10 stocks belonging to the US stock data S&P500. Here, as an example, Apple (AAPL), Amazon (AMZN), Bank of America Corporation (BAC), Berkshire Hathaway Inc. Class B (BRK-B), General Electric Company (GE), Johnson & Johnson (JNJ), JPMorgan Chase & Co. (JPM), Microsoft Corporation (MSFT), AT & T Inc. (T), and Wells Fargo & Select 10 stocks of Company (WFC). Each data can be downloaded from yahoo finance, for example. In addition, the first 70% of each data can be used for learning and the second 30% for evaluating the predictive model. The data used to build the prediction model can be composed of two indices' transaction data (open price, high price, low price, closing price, trading volume) and technical indicators.

The auto-encoder can basically consist of three layers: an input layer, a hidden layer, and an output layer. In addition, the stacked noise reduction autoencoder can be configured by stacking hidden layers of several autoencoders. The operating principle of the auto-encoder is optimized by minimizing the error between the input data of the input layer and the output data of the output layer. In this process, the hidden layer can extract high-level features. In addition, the features extracted by the hidden layer are used as input data of the next autoencoder, and a high-dimensional level can be extracted in the same manner as the above process. By stacking the hidden layers extracted in this way (left figure, SDAE in the figure below), a stacked noise canceling autoencoder can be built.

In the same way, we can build a stacked noise autoencoder for the S&P500 index.

3 is a diagram for describing a method of constructing an autoencoder for removing stacked noise according to an exemplary embodiment.

Referring to FIG. 3, a method of constructing a stacked noise canceling autoencoder (SDAE) is shown. For example, corrupted data can be created by applying a constant noise to Korea stock data. In addition, features can be extracted using a unit auto-encoder composed of an input layer, a hidden layer, and an output layer. At this time, features can be extracted through the hidden layer of the auto-encoder. The hidden layer can be used as the input layer for the next Ocoencoder. By repeating this process, features are extracted through several hidden layers, and the extracted features are stacked to build a stacked noise reduction auto-encoder (SDAE).

In other words, a hybrid model (SDAE-FNN) is constructed by connecting a stacked noise reduction auto-encoder (SDAE) to the input layer of the front deep artificial network (FNN), and a hybrid model (SDAE-FNN) is trained using a backpropagation algorithm. can do.

4 is a diagram illustrating feature extraction of an autoencoder for removing stacked noise according to an exemplary embodiment.

Referring to FIG. 4, features of a high-dimensional level may be extracted through a stacked noise removal autoencoder (SDAE) 410. The output of each stacked noise encoder (ie, auto-encoder) may be connected to an input layer of a forward deep artificial network (FNN). The entire neural network composed of a stacked noise canceling auto-encoder (SDAE, 410) and a forward deep artificial network (FNN, 420) can tune hyperparameters with a backpropagation algorithm. That is, the stock price prediction yield 402 can be calculated using financial data and technical indicators 401 through the entire neural network composed of a stacked noise reduction auto-encoder (SDAE, 410) and a forward deep artificial network (FNN, 420). .

Here, the optimal number of stacked layers may be determined through tuning, and may be stacked, for example, in 10 layers.

For the hyper-parameter of the prediction model, a parameter that minimizes the mean square root in the verification data can be selected using a Bayesian optimization algorithm.

Here, the Bayesian optimization algorithm can shorten the time required for the calculation process by finding a parameter to be considered preferentially using the calculated parameter value instead of considering the hyper parameter in all cases.

And, the screening level can be selected. For example, in this experiment, two levels of 0.0005 and 0.001 values can be selected as two examples. In other words, by selecting only stocks with a deep learning predicted return of each stock higher than 0.0005 (or 0.001), it is possible to construct a portfolio by selecting only stocks that are predicted to have a good future return.

The weight of each stock can be calculated through the average-variance optimization process for the selected stock. Here, the average-variance optimization process may use the method proposed in Markowitz's Portfolio selection (1952). The above method is the most widely used method in financial investments as a method of obtaining the share of stocks that maximize the expected rate of return to risk. The existing Markowtiz portfolio model uses human-estimated future stock price returns as input information. On the other hand, the deep learning-based stock screening method and apparatus according to the present embodiment can automate the process of building an optimal portfolio by replacing this process with a deep learning method, and increase the return rate of the portfolio with high prediction accuracy of deep learning.

The same calculation can be repeated for the screening level of 0.001.

The evaluation of the deep learning-based stock screening method and apparatus according to the present embodiment can calculate the cumulative return of the portfolio during the test period (a period other than the period used for model construction).

5 is a diagram illustrating an example of a portfolio in which all stocks are constructed with the same weight according to an embodiment.

Referring to FIG. 5, a result of comparing the calculated result with a benchmark (S&P500 index, same weight portfolio) is shown. Here, EW is a portfolio in which all stocks are constructed with the same weight, SP500 is an S&P500 index, and Screening level: 0.0005 and Screening level: 0.001 are portfolios constructed according to this embodiment. It can be seen that the portfolio built according to the present embodiment achieves a higher rate of return when compared with the same weight (EW) and market index (SP500).

That is, the excellent return rate compared to the benchmark index (about twice the S&P 500 index in the case of 0.001, and 1.5 times the EW) shows the excellent performance of the portfolio built according to this embodiment.

The apparatus described above may be implemented as a hardware component, a software component, and/or a combination of a hardware component and a software component. For example, the devices and components described in the embodiments include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable array (FPA), It may be implemented using one or more general purpose or special purpose computers, such as a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications executed on the operating system. Further, the processing device may access, store, manipulate, process, and generate data in response to the execution of software. For the convenience of understanding, although it is sometimes described that one processing device is used, one of ordinary skill in the art, the processing device is a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that it may include. For example, the processing device may include a plurality of processors or one processor and one controller. In addition, other processing configurations are possible, such as a parallel processor.

The software may include a computer program, code, instructions, or a combination of one or more of these, configuring the processing unit to operate as desired or processed independently or collectively. You can command the device. Software and/or data may be interpreted by a processing device or, to provide instructions or data to a processing device, of any type of machine, component, physical device, virtual equipment, computer storage medium or device. Can be embodyed. The software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiment, or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -A hardware device specially configured to store and execute program instructions such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those produced by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

As described above, although the embodiments have been described by the limited embodiments and drawings, various modifications and variations are possible from the above description to those of ordinary skill in the art. For example, the described techniques are performed in a different order from the described method, and/or components such as systems, structures, devices, circuits, etc. described are combined or combined in a form different from the described method, or other components Alternatively, even if substituted or substituted by an equivalent, an appropriate result can be achieved.

Therefore, other implementations, other embodiments, and those equivalent to the claims also fall within the scope of the claims to be described later.

Claims (15)

Pre-processing to convert financial data into data suitable for deep learning, and removing noise contained in the financial time series using a noise-removing auto-encoder;
Inputting the data from which noise has been removed into a prediction model based on deep learning;
Predicting a stock price prediction yield for each stock item through the trained prediction model; And
Setting a screening level, selecting a stock having the stock price prediction yield greater than the set screening level
Containing, deep learning-based stock screening method. The method of claim 1,
Selecting a stock having the stock price prediction yield greater than the screening level,
Building a portfolio by selecting stocks with the stock price prediction yield greater than the screening level
Characterized in, a deep learning-based stock screening method. The method of claim 2,
Selecting a stock having the stock price prediction yield greater than the screening level, optimizing the constructed portfolio by deriving a portfolio proportion for each stock item through an average-variance optimization process
Further comprising a deep learning-based stock screening method. The method according to claim 2 or 3,
Analyzing the performance of the portfolio built using test data
Further comprising a deep learning-based stock screening method. The method of claim 1,
The financial data,
Constructed using technical indicators and trading data of the KOSPI and foreign indices
Characterized in, a deep learning-based stock screening method. The method of claim 1,
The step of removing noise contained in a financial time series using the noise removal autoencoder,
Pre-processing by scaling the input financial data to a value between 0 and 1
Characterized in, a deep learning-based stock screening method. The method of claim 1,
The step of removing noise contained in a financial time series using the noise removal autoencoder,
A stacked-noise-removing auto-encoder is constructed by stacking hidden layers of a plurality of auto-encoders consisting of an input layer, a hidden layer, and an output layer, and high-level features are extracted through the stacked-noise-removing auto-encoder.
Characterized in, a deep learning-based stock screening method. The method of claim 1,
The step of removing noise contained in a financial time series using the noise removal autoencoder,
Using an autoencoder consisting of an input layer, a hidden layer, and an output layer, features are extracted through the hidden layer, and the hidden layer is used as an input layer of the next autoencoder, and features are extracted through the hidden layer. Constructing the stacked noise-removing auto-encoder through an auto-encoder and extracting high-level features
Characterized in, a deep learning-based stock screening method. The method of claim 1,
The deep learning-based prediction model,
The deep learning-based prediction model is constructed using a feedforward deep neural network
Characterized in, a deep learning-based stock screening method. The method of claim 7,
The step of inputting the noise-removed data into a deep learning-based prediction model,
A hybrid model (SDAE-FNN) is constructed by connecting the Stacked Denoising Autoencoder (SDAE) to the input layer of a deep learning-based prediction model, the Feedforward Deep Neural Network (FNN). Training the hybrid model (SDAE-FNN) using a backpropagation algorithm
Characterized in, a deep learning-based stock screening method. A noise removal unit that pre-processes the financial data to convert it into data suitable for deep learning, and removes noise contained in the financial time series using a noise reduction auto-encoder;
A prediction model that inputs the data from which noise has been removed into a prediction model based on deep learning;
A yield prediction unit that predicts a stock price prediction yield for each stock item through the trained prediction model; And
A screening unit that sets a screening level and selects a stock having the stock price prediction yield greater than the set screening level
Containing, deep learning-based stock screening device. The method of claim 11,
The screening unit selects a stock having the stock price prediction rate greater than the screening level, and then derives the proportion of the portfolio for each stock item through an average-variance optimization process, and optimizes the constructed portfolio.
Including more,
The screening unit,
Building a portfolio by selecting stocks with the stock price prediction yield greater than the screening level
Characterized in, a deep learning-based stock screening device. The method of claim 12,
Performance analysis unit that analyzes the performance of the portfolio built using test data
Further comprising a deep learning-based stock screening device. The method of claim 11,
The noise removal unit,
A stacked-noise-removing auto-encoder is constructed by stacking hidden layers of a plurality of auto-encoders consisting of an input layer, a hidden layer, and an output layer, and high-level features are extracted through the stacked-noise-removing auto-encoder.
Characterized in, a deep learning-based stock screening device. The method of claim 14,
The prediction model,
A hybrid model (SDAE-FNN) is constructed by connecting the Stacked Denoising Autoencoder (SDAE) to the input layer of a deep learning-based prediction model, the Feedforward Deep Neural Network (FNN). Training the hybrid model (SDAE-FNN) using a backpropagation algorithm
Characterized in, a deep learning-based stock screening device.

Images