KR20210125773A - A real-time stock price prediction system using LSTM neural network and text miner - Google Patents

Abstract

The present invention relates to a real-time stock price prediction system using a long short-term memory (LSTM) neural network and a text miner which uses an LSTM neural network to use past data of a stock price to train the LSTM neural network, predicts the stock price with the trained neural network, and analyzes real-time stock market news through a text miner to assign a weight to the stock price. The real-time stock price prediction system using an LSTM neural network and a text miner comprises: a neural network module having an LSTM neural network; a neural network training unit using past stock price data to train the LSTM neural network; a text miner extracting mood data representing whether a favorable factor or an unfavorable factor is present in a number by text-mining news data; and a stock price prediction unit inputting stock price data up to the day into the LSTM neural network to acquire outputted prediction stock price data, and giving weight to the acquired prediction stock price data with the mood data to calculate final prediction stock price data. By the system, an accurate stock price can be predicted by reflecting all news environments changing constantly and prediction by time series of the stock price by predicting the stock price through the LSTM neural network and giving a weight with analysis results of real-time news through the text miner.

Description

{ A real-time stock price prediction system using LSTM neural network and text miner }

The present invention uses LSTM (long short-term memory networks) neural network to learn stock prices using historical data, predict stock prices with the learned neural network, and analyze real-time stock news through text miners to give weights to the predicted stock prices. It relates to a real-time stock price prediction system using an LSTM neural network and a text miner.

In general, the current stock price prediction system analyzes past data and infers it through time series [Patent Document 1], or uses artificial intelligence to learn the stock price fluctuation pattern and predict the future [Patent Document 2] ].

However, there are various factors that affect stock prices, and it is virtually impossible to predict the future based on past data. In addition, the stock market does not repeat past trading patterns, but is influenced by new situations at every moment, such as international relations, specific events, and famous statements. Therefore, it is almost impossible to predict the price of stocks that change every moment even if a normal deep learning algorithm is applied.

That is, in the case of a normal artificial neural network, it is assumed that all inputs and outputs are independent and the calculation is performed. However, stock prices are historical data, that is, past stock prices influence future stock prices. Therefore, it is necessary to apply a learning method that reflects these characteristics well.

In addition, it is necessary to predict stock prices by reflecting market conditions based on news as well as time series analysis of stock prices.

Korean Patent Publication No. 10-1508361 (2015.04.08. Announcement) Korean Patent Publication No. 10-148004 (2014.11.04. Announcement)

Sepp Hochreiter, JuRgen Schmidhuber, "Long Short-Term Memory", Neural Computation, Vol. 9, Issue 8, November 1997, pp 1735??1780

An object of the present invention is to solve the above-described problems, by using a long short-term memory networks (LSTM) neural network to learn stock prices using historical data, and predicting stock prices using the learned neural network, It is to provide a real-time stock price prediction system using LSTM neural network and text miner that analyzes real-time stock news and adds weight to the predicted stock price.

That is, an object of the present invention is to generate prediction data by the time series of stock prices using a LSTM neural network, which is a series of recurrent neural networks (RNN), and analyze real-time stock news using text miners (YTextMiner, etc.) This is to provide a real-time stock price prediction system using an LSTM neural network and text miner that predicts the optimal stock price by comparing it with past time series data by adjusting the weight when generating forecast data after judging it as good news or bad news.

In order to achieve the above object, the present invention relates to a real-time stock price prediction system using an LSTM neural network and a text miner, comprising: a neural network module including an LSTM neural network; a neural network learning unit for learning the LSTM neural network using past stock price data; a text miner for extracting mood data that numerically indicates whether news data is good news or bad news by text mining; and a stock price prediction unit that inputs stock price data up to the day to the LSTM neural network to obtain output predicted stock price data, and calculates final predicted stock price data by weighting the obtained predicted stock price data with the mood data. characterized.

In addition, the present invention provides a real-time stock price prediction system using an LSTM neural network and a text miner, the neural network learning unit comprising: (a) collecting time series data of a past stock price; (b) calculating a median value of each stock price; (c) calculating an increment of the median of each stock price; (d) normalizing the calculated increments; (e) generating a sequential model from the normalized incremental time series data; and, (f) training the LSTM neural network with the sequential model.

In addition, the present invention is characterized in that, in a real-time stock price prediction system using an LSTM neural network and a text miner, the increment is normalized by applying a sigmoid function to the increment of the intermediate value in step (c).

In addition, the present invention is characterized in that, in a real-time stock price prediction system using an LSTM neural network and a text miner, the normalized increment is calculated by Equation 1 below.

[Formula 1]

,

이고,step,

,

ego,

t is the date, and H(t) and L(t) represent the high and low prices of stock price data on day t, respectively.

In addition, the present invention provides a real-time stock price prediction system using an LSTM neural network and a text miner, in the step (e), time series data S(-N+1), S(-N+2), . . It is characterized in that the sequential model is constructed by configuring the set and the correct answer of the predicted stock price in the case of the set.

As described above, according to the real-time stock price prediction system using the LSTM neural network and the text miner according to the present invention, the stock price is predicted through the LSTM neural network and weighted with the analysis result of the real-time news through the text miner. At the same time as forecasting, the effect of more accurate stock price prediction is obtained by reflecting all the changing news environment from moment to moment.

1 is an exemplary configuration diagram of an entire system for implementing the present invention.
2 is a block diagram of a real-time stock price prediction system using an LSTM neural network and a text miner according to an embodiment of the present invention.
3 is a diagram illustrating the structure of an LSTM neural network according to an embodiment of the present invention.
4 is a flowchart illustrating a learning method of a neural network learning unit according to an embodiment of the present invention.
5 is a table illustrating a stock price data set according to an embodiment of the present invention.
6 is a programming code for a stock price learning method according to an embodiment of the present invention.
7 is programming code for a normalization method according to an embodiment of the present invention;
8 is programming code for a training method according to an embodiment of the present invention;

Hereinafter, specific contents for carrying out the present invention will be described with reference to the drawings.

In addition, in demonstrating this invention, the same part is attached|subjected with the same code|symbol, and the repetition description is abbreviate|omitted.

First, the configuration of the entire system for implementing the present invention will be described with reference to FIG. 1 .

As shown in FIG. 1, the entire system for implementing the present invention includes a user terminal 10, a stock price prediction server 30 for predicting stock prices, a stock price information server 50 for providing stock price information, and news. It is composed of a news server 60 that provides. Additionally, it may be configured to further include a database 40 for storing necessary data.

First, the user terminal 10 is a terminal used by a user, and is a general computer terminal or dedicated terminal having a computing function, such as a smart phone, a tablet PC, a notebook computer, and a personal computer (PC).

The user terminal 10 may access the stock price prediction server 30 to receive a stock price prediction service provided by the stock price prediction server 30 . That is, the user terminal 10 inputs an item from the stock price prediction server 30 , and requests and receives an expected stock price for the input item.

Preferably, the expected stock price is composed of the highest target price and the lowest target price.

Next, the stock price information server 50 is a normal server that provides stock price information, and is a server operated by a securities company, a stock price information provider, and the like. A plurality of stock price information servers 50 may exist, and each stock price information server 50 may provide different types of stock price information.

Preferably, the stock price information consists of a high price, a low price, a closing price, and the like. In addition, each stock price information is composed of information by date (data by date). That is, the stock price information includes stock price information traded on a current or past trading business day.

Next, the news server 60 is a server that provides news, and is a server operated by each newspaper company, broadcasting company, or public institution. The news server 60 provides news online.

News information is news data written in text.

News provided by the news server 60 may be classified according to providers. That is, the news provider is the Financial Supervisory Service, each newspaper or broadcasting company.

Next, the stock price prediction server 30 is a normal server, according to the request of the user or the user terminal 10, predicts a stock price, and provides the predicted stock price information to the user. Preferably, by receiving the name of a specific stock, predicting the stock price for the stock, and providing the predicted stock price data of the stock.

In particular, the stock price prediction server 30 includes a neural network of the LSTM neural network, and learns the LSTM neural network using past stock price information. Then, the current stock price information is input to the trained LSTM neural network, and the predicted stock price data is output.

In addition, the stock price prediction server 30 may provide the predicted stock price data as the highest target price and the lowest target price. In this case, preferably, the highest target price and the lowest target price are extracted by setting an error range in the predicted stock price.

In addition, the stock price prediction server 30 receives and utilizes stock price information from the stock price information server 50 .

In addition, the stock price prediction server 30 includes a text miner, and performs mining analysis on news using the text miner. Through the text miner, whether the current news is positive or negative is extracted as a number (a number for good news or bad news). The figure at this time will be referred to as a news evaluation map.

In particular, by inputting the name of the item in the text minor, the news evaluation for the item is extracted. In addition, the text miner analyzes news data from the news server 60 to extract a news rating.

In addition, the stock price prediction server 30 converts the news rating into a weight, and weights the weight to the previously obtained predicted stock price data to calculate the final predicted stock price data.

Next, the configuration of the real-time stock price prediction system 30 according to an embodiment of the present invention will be described with reference to FIG. 2 .

The real-time stock price prediction system 30 according to the present invention may be implemented in the form of the stock price prediction server 30 above. Alternatively, the real-time stock price prediction system 30 may be implemented in a server-client form, and a client module may be installed in the user terminal 10 .

As shown in FIG. 2 , the real-time stock price prediction system 30 according to an embodiment of the present invention includes a neural network module 31 composed of an LSTM neural network, a neural network learning unit 32 for learning a neural network, and a text for analyzing news. It is composed of a minor 33 and a stock price prediction unit 34 that predicts a stock price.

First, the neural network module 31 includes a long short-term memory networks (LSTM) neural network.

LSTM (Long Short Term Memory networks) neural network is a type of RNN (Recurrent Neural Network), and has a structure in which past data can influence the future. Therefore, the LSTM neural network is the best model for stock price prediction.

In the existing RNN neural network, the values between [-1, 1] are continuously multiplied by the chain rule, which is a long-term dependency problem, so the value gets smaller as it goes forward and eventually disappears. Therefore, the existing RNN method has a problem that parameters are not updated. LSTM is a method proposed to solve this [Non-Patent Document 1].

In the conventional RNN, when calculating the hidden state (St) in the hidden layer, it is simply calculated as St = tanh (Uxt + WSt), but in LSTM, there are a total of four calculation processes.

As shown in FIG. 3 , there are four neural network layers in the hidden layer of the LSTM. That is, it can be confirmed that four neural network layers exist in the central square shape of FIG. 3 . In addition, the core of LSTM is that a very minor linear operation is performed on the horizontal line at the top of the neural (A), and information is transmitted to the next stage without much change through the entire chain.

In addition, the LSTM neural network is trained with the high and low prices of stocks as default values. In particular, for high and low prices, we train with historical data for a specific period (for example, 30 days, less reflection period), predict one day, and compare it with the real data. In this comparison, the training set is learned as it moves into the future in units of a reflection period (30 days as an example). In addition, by analyzing real-time news information, it calculates the ratio of the impact to the stock price at the time by comparing it with similar news in the past. Then, by repeating the adjustment of the weights, the final target price is set to decide whether to buy or sell the stock.

In addition, the LSTM neural network is trained for one event. Therefore, in order to predict the stock price of a new stock (or another stock), the LSTM neural network is initialized and trained again with the historical stock price data of the relevant stock.

Alternatively, as another embodiment, the neural network module 31 includes an LSTM neural network corresponding to each event to be predicted for each event. That is, a plurality of LSTM neural networks corresponding to each stock are provided, and each stock learns from the historical stock price data of the stock.

Next, the neural network learning unit 32 trains the LSTM neural network. That is, by collecting historical data of stock prices, the LSTM neural network is trained using the collected historical data. In particular, preferably, time series data of stock price information up to the day to be predicted is put into the LSTM neural network to train the LSTM neural network.

A specific method for the neural network learning unit 32 to learn the LSTM neural network is illustrated in FIG. 4 .

First, historical data of stock prices are collected (S21). Collect stock price data up to the day before the date you want to predict. As an example, in order to predict the stock price of the next day, historical data up to that day are collected.

Next, as shown in FIG. 5 , the stock price prediction is based on the median value of the past stock price. The median value is obtained as the average of the high and low prices as shown in the following equation (S22).

[Equation 1]

Here, t is the date, M(t) is the median value of day t, and H(t) and L(t) represent the high and low prices of day t, respectively.

The table in Fig. 5 is a data set of Samsung Electronics' stock price from February 2019 to February 2020.

Next, for the reference day, the stock price (or increment) is set to 0 and the increment of the next day is calculated (S23).

In the example of Fig. 5, when No. 1 is the reference, the increment value of the reference day becomes 0. Also, stock prices are shifted less than 1 or greater than -1, indicating a normalized process.

That is, first, the increment ΔM(t) of day t is obtained by the following equation.

[Equation 2]

That is, the average price (median) of the high (high) and low (low) prices are calculated, and then the increment is calculated.

Next, the increment ΔM(t) is normalized with a sigmoid function (S24). Normalization is as follows.

[Equation 3]

where S(t) is the normalized increment of t days.

Next, a sequential model is generated using the normalized incremental time series data (S25).

Let the size of the entire time series data be N, and set the batch size to n. In this case, N > n, preferably, N and n are set to satisfy N > 2n. That is, to make the number of batches larger than the batch size.

The full normalized incremental time series data is data by a series of dates, S(-N+1), S(-N+2), ...., S(0). S(-k) is the normalized incremental data of the previous k days, and S(0) is the normalized incremental data of the day.

At this time, a batch is created with a batch size n. For a total data size of N, N-n batches are generated as follows.

[ S(-N+1), S(-N+2), S(-N+3), ..., S(-N+n) ], S(-N+n+1)

[ S(-N+2), S(-N+3), S(-N+4), ..., S(-N+n+1) ], S(-N+n+2)

...

[ S(-n-1), S(-n), S(-n+1), ..., S(-2) ], S(-1)

[ S(-n), S(-n+1), S(-n+2), ..., S(-1) ], S(0)

As described above, the sequential model constructs one set of N consecutive time series data (normalized incremental data) with a batch size of n+1, but constructs a set sequentially to form a series of N-n batch sets. At this time, the last data of each batch set is set as the result value (true value) of that set. That is, each batch consists of a data set of n historical stock prices, and the next-day stock price in that set (the real stock price of the predicted stock price) (the correct stock price).

Next, using each batch set as one training data, the LSTM neural network is trained (S26).

6 is a part of stock learning code written in Python language. Seq_len means a training (learning) set, so No. 1 to No. start with 30 We use a training set of 30, which is a 30-day stock trading data set. That is, the batch size is 30 (or 30 days), and the number of batches is 30.

After learning, by predicting the price for the next day, the stock investor can decide whether to sell or buy the stock.

Preferably, the data set (or batch set) is normalized to predict stock prices well, as shown in FIG. 7 . We divide the training data and the test data set that validates the training data in a 9:1 ratio. That is, it is equivalent to the code of "row = int (round (result.shape[0] * 0.9))".

In Figure 7, 0.9 means that this training set is 90% of the data set, and 0.1 means that the test set is 10% of the data set. The ratio may vary depending on the stock to be predicted in the future.

After the sequential model is built, training is started as shown in FIG. 8 to verify whether the model is correct. The batch size (Batch_size) of FIG. 8 is the number of data that can be continuously trained, and epochs are the number of repetitions and data units. For validation, 90% of the identified training set and 10% of the test set are used. It is divided into a training set and a test set for validation, and each set never overlaps.

Next, the text miner collects news data of the corresponding item, performs text mining, and calculates mood data of the corresponding item.

Preferably, the text miner uses a commercially available tool such as YTextMiner. Text Miner (YTextMiner, etc.) is a text analysis tool, and it analyzes real-time stock news to determine whether news corresponding to stock items by keyword is good news or bad news.

Mood data is an indicator that numerically indicates whether a stock is good news or bad news. In case of good news, the value of mood data increases, and in case of bad news, the value of mood data decreases.

Preferably, the mood data (or mood value) has a value between (1-α) and (1+α). In this case, α has a value of 1% to 30%.

Next, the stock price prediction unit 34 inputs the current stock price information to the neural network module 31 or the LSTM neural network to obtain an output value thereof, and weights the mood data obtained in the text minor 33 to the output value as a weight to make a final prediction Calculate stock price information.

First, through the neural network module 31, time series data of stock price information (high price and low price) is input to the LSTM neural network to predict stock prices (or stock prices of the next day).

At this time, the stock price information is converted into a normalized incremental value and input, and on the contrary, the output value is inversely transformed to be converted into a predicted stock price.

In addition, the first predicted stock price information including the highest target price and the lowest target price of the predicted stock price is acquired according to a predetermined error range.

Then, through the text miner 33, mood data is calculated from the news data, and weights are obtained from the calculated mood data.

Then, by weighting the predicted stock price information with a weight, the final predicted stock price information is obtained.

That is, the real-time stock price prediction system 30 of the present invention learns time series data through LSTM and uses it to adjust weights when predicting. Since the stock prediction system through deep learning does not recognize real-time changes in the market, it is used to improve the accuracy by analyzing the data collected through real-time news mining and referring to the final stock price prediction.

As described above, the real-time prediction system of the present invention repeatedly trains past data (for example, 30-day data) using an LSTM model and predicts future daily stock prices, and real-time stock news is final using YTextMiner. Weights are applied when predicting stock prices.

That is, the real-time prediction system of the present invention uses a recurrent neural network (RNN: 　 Recurrent Neural Network) to infer past stock price analysis through time series, and searches for real-time stock news using text miners (YTextMiner, etc.) The short-term stock price is predicted by changing the weight by analyzing the impact on the stock price.

Prediction system development using artificial intelligence can be applied in a variety of ways, such as weather, climate change, earthquakes, typhoon movement routes, virus movement routes, automobile movement, etc. is in progress Among them, stock price prediction is known as the most active technology development field due to its large market size. In order to predict stock prices, it takes a lot of effort because it is necessary to determine the timing of trading through continuous monitoring. Buying and selling can be done through the bot. Through this, selling and buying can proceed at the optimal time without unnecessary monitoring. In addition, by minimizing the case of loss through emotional investment through technical analysis, it is possible to buy and sell stocks more efficiently. Therefore, it is necessary to develop an automated trading bot based on artificial intelligence.

In the case of the United States, most of the system trading is done through an automatic stock trading program. The automatic stock trading program is not simply a machine that automatically buys and sells stocks for no reason, but trades them through an algorithm that has clear standards through technical analysis. Through this, it is possible to proceed with trading by raising the level of professional investors, split selling according to the market price, bid price, and bid price according to the optional trading period according to the approach, and also a machine of artificial intelligence that is not emotional stop loss or profit loss You can conduct trades. In addition, this method throws away the most poisonous emotions in the stock market, and captures the opportunity cost of losing money through surges and sharp declines when time investment is impossible. In the end, both in terms of professionalism and time, it has an effect.

As mentioned above, although the invention made by the present inventors has been described in detail according to the embodiments, the present invention is not limited to the embodiments, and various changes can be made without departing from the gist of the present invention.

10: user terminal 30: stock price prediction server
31: neural network module 32: neural network learning unit
33: text minor 34: stock price prediction unit
40: database 50: stock price information server
60 : news server

Claims (5)

In a real-time stock price prediction system using an LSTM neural network and a text miner,
a neural network module having an LSTM neural network;
a neural network learning unit for learning the LSTM neural network using past stock price data;
a text miner for extracting mood data that numerically indicates whether news data is good news or bad news by text mining; and,
and a stock price prediction unit that inputs stock price data up to the day into the LSTM neural network to obtain output predicted stock price data, and calculates final predicted stock price data by weighting the obtained predicted stock price data with the mood data. Real-time stock price prediction system using LSTM neural network and text miner.
According to claim 1,
The neural network learning unit,
(a) collecting time series data of past stock prices;
(b) calculating a median value of each stock price;
(c) calculating an increment of the median of each stock price;
(d) normalizing the calculated increments;
(e) generating a sequential model from the normalized incremental time series data; and,
(f) a real-time stock price prediction system using an LSTM neural network and a text miner, characterized in that performing a method comprising the step of training the LSTM neural network with the sequential model.
3. The method of claim 2,
In step (c), a real-time stock price prediction system using an LSTM neural network and text miner, characterized in that the increment is normalized by applying a sigmoid function to the increment of the median value.
4. The method of claim 3,
A real-time stock price prediction system using an LSTM neural network and a text miner, characterized in that the normalized increment is calculated by Equation 1 below.
[Formula 1]

step,

,

ego,
t is the date, and H(t) and L(t) represent the high and low prices of stock price data on day t, respectively.
3. The method of claim 2,
In step (e), a plurality of batch sets are sequentially generated from the time series data S(-N+1), S(-N+2), ...., S(0) of the entire normalized increment, A sequential model is constructed by forming a set with a batch size of n+1 to form a series of Nn batch sets, and each batch set consists of n historical stock price data sets and the correct answer of the predicted stock price for the set. Real-time stock price prediction system using LSTM neural network and text miner, characterized in that

Images