CN110659767A - Stock trend prediction method based on LSTM-CNN deep learning model - Google Patents

Abstract

A stock trend prediction method based on an LSTM-CNN deep learning model comprises the following steps: step 1: preprocessing data; step 2: predicting the stock price trend by using an LSTM-CNN prediction model; and step 3: and evaluating the standard, wherein the prediction target is to predict the trend position and the bottom turning point, and the accuracy of the two aspects is respectively considered. The invention provides an LSTM-CNN neural network model and a weighting loss function, optimizes a normalization algorithm and a target function trend position, and accurately predicts the stock fluctuation trend.

Description

A stock trend prediction method based on LSTM-CNN deep learning model

technical field

The invention relates to the research of financial time series, in particular to predicting the stock trend by applying the method of combining two neural networks, LSTM and CNN.

Background technique

With the development of the global economy, China's financial market development faces greater opportunities and challenges. The financial market is rich in content, and its products mainly include bonds, futures, foreign exchange, stocks, etc. The price fluctuations are the focus of attention.

As an important part of the financial market, stocks are the weather vane of the financial market. The volatility of the stock market is closely related to the interests of investors, and has become the focus of attention of many investors and financial institutions.

Quantitative analysis of historical stock price fluctuations, transaction data, etc., and the use of data mining, artificial intelligence and other technologies to provide reference is of great significance to investors. In recent years, with the technological breakthroughs in the field of artificial intelligence, artificial intelligence technology has once again attracted the attention of the public.

Hocherative S and Schmidhuber J proposed that the LSTM neural network model is suitable for feature analysis in the data domain with time series characteristics. In the early days of CNN, the prototype of the convolutional neural network was proposed by Fukushima K, which was inclined to identify static features. This feature is also suitable for financial data. The connection is strong in adjacent time, and then local features are extracted.

SUMMARY OF THE INVENTION

Different from the current application of a single neural network to analyze and predict stock data, the stock price fluctuation is analyzed based on the LSTM-CNN network model, and the time series and static features are extracted from the stock data through the neural network. Make predictions.

The technical scheme adopted by the present invention to solve its technical problems is:

A stock trend prediction method based on LSTM-CNN deep learning model, including the following steps:

Step 1: Data Preprocessing

Select the stock data of the stock market for analysis and verification, the overall data set is divided into training set and test set;

Step 1.1: Feature Selection and Normalization

The features are transaction data and technical indicators, and the transaction data needs to be normalized; the technical indicators are calculated from the transaction data, and the normalization process is already included in the calculation process;

First, normalize the original transaction data. In the stock data, when the market continues to rise or fall, new highs or new lows are common phenomena in financial data. Therefore, a normalization method is designed for price and trading volume respectively. ;

Price normalization: Quantify stock data for each time window. Take the first position of the stock sequence as the benchmark price to calculate the relative price increase. The formula is as shown in (1) to separately normalize the data in each time window;

For each training sample, the corresponding closing prices at times t ₀ , t ₁ ,…, t _i , t _m are P ₀ , P ₁ ,…, P _i , P _m . Where μ is the average closing price, and m is the sliding window size;

Normalization of trading volume: Calculated on the overall time series of a stock, the formula is shown in (2):

At time T _i , the trading volume is V _i , the eigenvalue obtained based on the normalization of trading volume is Vx _i , the constant term m is usually equal to the time window size, then the first m time points of a stock are only used for data processing , not included in the sample;

After completing the normalization of closing price and trading volume, the technical indicator values of the calculation sequence are MACD, RSI and KDJ-J respectively, which are all used as input data;

Step 1.2: Calibrate the trend position

Study the medium and long-term price trends of stocks, define the turning points, quantify the trend according to the turning points, get the trend position, and calibrate the data labels accordingly;

Trend turning points include the top turning point at the peak position and the bottom turning point at the trough position. When the price is at the bottom turning point, it means that the price is about to rise, and vice versa. In the actual market, the price has been fluctuating, resulting in fluctuations of different amplitudes, resulting in multiple turning points. The turning points identified by the model are quantified. The specific definitions are as follows:

The closing prices corresponding to the stock price sequence at T ₁ , T ₂ , and T ₃ are P ₁ , P ₂ , and P ₃ respectively, and a parameter is set called the trend threshold δ. When condition (4) is satisfied, and T ₂ If the moment price is the lowest point in the period from T ₁ to T ₃ , the moment T ₂ is considered to be the bottom turning point of a trend;

Similar to the bottom turning point, when the condition ( ₅ ) is satisfied, and the price at time T2 is the highest point in the period from T1 to T3, then the time at T2 is considered as the top turning point _{of a} trend, and its trend is shown in the right picture of Figure ₁ . Show. It is worth noting that the calculation process of the bottom turning point and the top turning point should be based on the low price, otherwise under the same threshold δ, the absolute value of the fluctuation required for the establishment of the bottom turning point and the top turning point will be different, which will cause deviations;

When a bottom turning point is determined, the subsequent sequence is defined as an upward trend until it encounters the next top turning point. Similarly, when a top turning point is determined, the subsequent sequence is defined as a downward trend until it encounters the next top turning point. next bottom turning point;

After determining the bottom and top turning points of the time series, map the current stock price to the relative position in the trend, and use the trend position as the direct prediction target of the LSTM-CNN model to predict the future price trend. The calibration method of the trend position is as follows :

1.2.1, according to the definition of turning point, all bottom turning points and top turning points of a set of sequences can be obtained, the trend position at the bottom turning point is marked as 0, and the trend position at the top turning point is marked as 100;

1.2.2, the trend position value between the bottom turning point and the top turning point is calculated according to formula (6), where _PL and _PH are a set of adjacent bottom and top turning points;

For the last trend of the time series, the top or bottom turning point of the last price cannot be determined, so the undetermined part is not marked and is not included in the training data;

Step 2: Use the LSTM-CNN prediction model to predict the stock price trend, the steps are as follows:

Step 2.1: Network model construction, the steps are as follows:

The LSTM-CNN model consists of three parts, one part is the LSTM time series feature learning layer; the other part is the CNN static feature learning layer, which is synchronized with the LSTM part; the last part is the fully connected output layer, by combining the outputs of the first two models in series, Then build a fully connected neural network.

The first part is composed of three LSTM layers, as shown in part (a) in Figure 3. The input matrix size of LSTM is . where k is the number of training features, and m is the size of the sliding time window. As shown in the figure, the L1 layer contains LSTM units corresponding to M time nodes, and the input of each LSTM unit consists of k features, including transaction data and technical indicators. Each unit in the L1 layer is trained and output as the input of each unit in the L2 layer. Similarly, the output of the L2 layer is used as the input of the L3 layer. The L2 and L3 layers are trained in the same way as the L1 layer. Finally, the output of the LSTM unit corresponding to the last time node of the L3 layer is used as the result of part a;

The second part consists of three convolutional layers and one LSTM layer, as shown in part (b) in Figure 3. In this section, the input matrix is the same as in section (a); the local correlation information between price and technical indicators is first extracted through three convolutional layers. Three convolutional layers followed by an LSTM layer. The temporal relationship in the extracted local dependency information is analyzed through the LSTM layer;

The last part consists of three fully connected layers. LSTM and CNN extract temporal and static information, respectively, and merge into a new feature sequence. This feature sequence is the input to the fully connected layer. Finally, the final prediction result of the model is output, which is the trend position. In addition, applying a random dropout method between fully connected layers can disconnect some nodes between two connected layers, and the final result is not completely dependent on a specific node. The random deactivation method significantly reduces the overfitting problem;

Step 2.2: Activation function

Using ELU as the activation function in the LSTM network, the formula is as (7):

φ(x)=max(0,x) (7)

The network structure used for training is shown in Figure 4. The training process needs to configure the network model. In the LSTM unit, the activation function and input and output dimensions need to be set. In the CNN unit, the activation function and the size of the convolution kernel need to be set. ;

Step 2.3: Build a model training strategy, and design a weighted loss function for model training. The steps are as follows:

The value range of the trend position is a limited interval [0, 100], and the application values of different positions corresponding to this interval are different. The present invention performs weighted measurement on it, and proposes a cost-sensitive loss function in the turning area;

进行等比例缩放与平移，使目标值处于区间[-1,1]内，公式(8)为转换方法，c为一较小实数，用来保证对数与分母计算对象不为零，使用arctanh公式(9)，对目标值z与预测值

进行映射，通过映射过程实现权值分配，Y与分别为变换加权后的真实值与预测值；First, compare the target value y with the predicted value

Perform equal scaling and translation to make the target value in the interval [-1,1], formula (8) is the conversion method, c is a small real number, used to ensure that the logarithm and denominator calculation objects are not zero, use arctanh Formula (9), for the target value z and the predicted value

Mapping is carried out, and weight distribution is achieved through the mapping process, and Y and are the true value and the predicted value after the transformation and weighting, respectively;

After the target value and the predicted value are weighted by mapping, MSE is applied to measure the error, and WMSE(10) is obtained, which is the loss function of the prediction model. Through weighting, the training speed can be accelerated, and the model complexity can be reduced to a certain extent;

Step 3: Error Evaluation Criteria

The prediction goal is to predict the trend position and the bottom turning point, and consider the accuracy of these two aspects respectively;

The trend position accuracy is evaluated using the mean absolute error (MAE), and the turning point accuracy is designed with two indicators: bottom precision (LP) and top recall (HR);

Precision and recall are commonly used evaluation indicators for binary classification problems. In the classification algorithm, after comparing the predicted results with the real target, the following four markers indicate the predicted state;

TP: When the true value is a positive class, the predicted value is a positive class;

TN: When the true value is a positive class, the predicted value is a negative class;

FP: When the true value is negative, the predicted value is positive;

FN: The predicted value is the negative class when the true value is negative.

Where T is True, F is False, P is Positive, and N is Negative. And calculate the precision (11) and recall (12) according to their values; from the formula, the precision indicates that when the predicted value is true, the actual value is also true, and the recall indicates that when the true value is true. When the value is true, the predicted value is also true;

For the bottom turning point, every time a turning point is predicted, a strong buy signal is provided, so we hope that each prediction is a turning point, as accurate as possible, otherwise it will cause investment losses, so we need to pay attention to the precision rate;

For the top turning point, every time a turning point is predicted, a sell signal is provided, and the forecast is as accurate as possible when each top appears, so as to avoid missing the selling opportunity, otherwise it will cause losses to investors, so the recall rate needs to be paid attention to.

The beneficial effects of the present invention are mainly manifested in: the present invention proposes the LSTM-CNN neural network model and the weighted loss function, optimizes the normalization algorithm and the trend position of the objective function, and accurately predicts the fluctuation trend of the stock.

Description of drawings

FIG. 1 is a definition diagram of a turning point of the present invention.

FIG. 2 is a diagram showing an example of a trend position of the present invention.

FIG. 3 is a schematic diagram of the structure of the LSTM-CNN network of the present invention.

FIG. 4 is an arctanh mapping graph of the present invention.

FIG. 5 is a target weighting effect diagram of the present invention.

FIG. 6 is the LSTM-CNN prediction diagram of the trend position of Baosteel Co., Ltd. according to the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings.

1 to 6, a stock trend prediction method based on the LSTM-CNN deep learning model, the method includes the following steps:

Step 1: Data Preprocessing

Part of the stock data in the A-share market is used for analysis and verification. The overall data set is divided into training set and test set. The CSI 300 index was selected as the training set, and the SSE 50 index was selected as the test set. It does not include new shares, new shares, long-term suspended stocks, ST shares and stocks with large and abnormal fluctuations. 218 stocks were selected to be included in the training set, and 37 stocks were included in the test set. Select the stock data from 2004 to December 2018 to obtain its weekly opening price, closing price, highest price, trading volume and other transaction data.

The main features are transaction data and technical indicators. The five sets of training features of the data are Cx, Vx, MACD, JDK-J, and RSI, which represent closing price, trading volume, MACD indicator, JDK-J indicator, and RSI indicator respectively. There are about 500 time series points for each stock in the training set, and the sliding window step size is 1. Each training time point corresponds to the trend position. For the target of this sample, the size of the target matrix is the same as the number of samples.

Step 1.1: Feature Selection and Normalization

The main features are transaction data and technical indicators, and the transaction data needs to be normalized. Technical indicators are calculated from transaction data, and the normalization process is already included in the calculation process, so it does not need to be processed again.

First normalize the raw transaction data. In stock data, where there is a continuous rise or a continuous decline, new highs or new lows are a common phenomenon in financial data. Therefore, a normalization method is designed for price and volume respectively.

Price normalization: Quantify stock data for each time window. Taking the first position of the stock sequence as the benchmark price, calculate the relative price increase. The formula is shown in (1) to separately normalize the data in each time window.

For each training sample, the corresponding closing prices at time t ₀ , t ₁ ,..., t _i , t _m are P ₀ , P ₁ ,..., P _i , P _m . Where μ is the average closing price, sliding window m =100;

Normalization of trading volume: Calculated on the overall time series of a stock, the formula is shown in (2):

At time T _i , the trading volume is V _i , the characteristic value obtained based on the normalization of trading volume is Vx _i , the constant term m=100 is usually equal to the time window size, then the first m=100 time points of a stock are only Used for data processing, not included in the sample;

Take the closing price, volume, MACD, RSI and KDJ-J as input data.

Step 1.2: Calibrate the trend position

Studying the mid- to long-term price trends of stocks, turning points are suggested. The trend is quantified according to the turning point, and the trend position is obtained, and the data label is calibrated accordingly.

Trend turning points include the top turning point at the peak position and the bottom turning point at the trough position. When the price is at the bottom turning point, it means that the price is about to rise, and vice versa. In the actual market, the price has been fluctuating, resulting in fluctuations of different amplitudes, resulting in multiple turning points. The turning points identified by the model are quantified. The specific definitions are as follows:

The closing prices corresponding to the stock price sequence at T ₁ , T ₂ , and T ₃ are respectively P ₁ , P ₂ , and P ₃ , and a parameter is set called the trend threshold δ=60%. When the condition (4) is satisfied, And if the price at time T2 is the lowest point in the period from T1 to T3, it is considered that time at T2 is the bottom _turning point _{of a} trend, and its trend is shown in the left figure of Figure ₁ .

Similar to the bottom turning point, when the condition ( ₅ ) is satisfied, and the price at time T2 is the highest point in the period from T1 to T3, then the time at T2 is considered as the top turning point _{of a} trend, and its trend is shown in the right picture of Figure ₁ . Show. It is worth noting that the calculation process of the bottom turning point and the top turning point should be based on the low price. Otherwise, under the same threshold δ, the absolute value of the fluctuation required for the establishment of the bottom turning point and the top turning point will be different, which will cause deviations.

When a bottom turning point is identified, the subsequent series is defined as an uptrend until the next top turning point is encountered. Similarly, when a top turning point is determined, the subsequent sequence is defined as a downtrend until the next bottom turning point is encountered.

After determining the bottom and top turning points of the time series, map the current stock price to the relative position in the trend. The trend position is used as the direct prediction target of the LSTM-CNN model to realize the prediction of future price trends. The calibration method of the trend position is as follows:

1.2.1. According to the definition of turning point, all bottom turning points and top turning points of a set of sequences can be obtained. The trend position at the bottom turning point is marked as 0, and the trend position at the top turning point is marked as 100.

1.2.2. The trend position value between the bottom turning point and the top turning point is calculated according to formula (6), where _PL and _PH are a set of adjacent bottom and top turning points.

In the last trend of the time series, the top or bottom turning point of the last price cannot be determined by definition, so the undetermined part is not marked and is not included in the training data. The marking effect of a stock sequence is shown in Figure 2.

Step 2: Use the LSTM-CNN prediction model to predict the stock price trend, the steps are as follows;

Step 2.1: Network model construction, the steps are as follows:

The LSTM-CNN model consists of three parts, one part is the LSTM time series feature learning layer; the other part is the CNN static feature learning layer, which is synchronized with the LSTM part; the last part is the fully connected output layer, by combining the outputs of the first two models in series, Then build a fully connected neural network.

The first part is composed of three LSTM layers, as shown in part (a) in Figure 3. The input matrix size of LSTM is . Where k=5 is the number of training features, and m is the size of the sliding time window. As shown in the figure, the L1 layer contains LSTM units corresponding to M=100 time nodes, and the input of each LSTM unit consists of k=5 features, including transaction data and technical indicators. Each unit in the L1 layer is trained and output as the input of each unit in the L2 layer. Similarly, the output of the L2 layer is used as the input of the L3 layer. The L2 and L3 layers are trained in the same way as the L1 layer. Finally, the output of the LSTM unit corresponding to the last time node of the L3 layer is used as the result of part a.

The second part consists of three convolutional layers and one LSTM layer, as shown in part (b) in Figure 3. In this section, the input matrix is the same as in section (a). The local correlation information between price and technical indicators is first extracted through three convolutional layers. Three convolutional layers followed by an LSTM layer. The temporal relationships in the extracted local dependency information are analyzed through LSTM layers.

The last part consists of three fully connected layers. LSTM and CNN extract temporal and static information, respectively, and merge into a new feature sequence. This feature sequence is the input to the fully connected layer. Finally, the final prediction result of the model is output, which is the trend position. In addition, applying a random dropout method between fully connected layers can disconnect some nodes between two connected layers, and the final result is not completely dependent on a specific node. The random deactivation method significantly reduces the overfitting problem.

Step 2.2: Activation function

Using ELU as the activation function in the LSTM network, the formula is as (7):

φ(x)=max(0,x) (7)

The network structure used in the training of the present invention is shown in Figure 4. The training process needs to configure the network model. In the LSTM unit, the activation function and input and output dimensions need to be set. In the CNN unit, the activation function and convolution need to be set. nuclear size. In the experiments of the present invention, the structural parameters of the LSTM-CNN network model are trained using the configurations shown in Table 1.

Table 1

Step 2.3: Build a model training strategy, and design a weighted loss function for model training. The steps are as follows:

The value range of the trend position is a finite interval [0, 100], and different positions corresponding to this interval have different application values. The present invention performs weighted measurement on it, and proposes a cost-sensitive loss function in the turning area.

进行等比例缩放与平移，使目标值处于区间[-1,1]内，公式(8)为转换方法，c为一较小实数，用来保证对数与分母计算对象不为零，使用arctanh公式(9)，对目标值z与预测值

进行映射，通过映射过程实现权值分配，Y与分别为变换加权后的真实值与预测值。First, compare the target value y with the predicted value

Perform equal scaling and translation to make the target value in the interval [-1,1], formula (8) is the conversion method, c is a small real number, used to ensure that the logarithm and denominator calculation objects are not zero, use arctanh Formula (9), for the target value z and the predicted value

Mapping is carried out, and weight distribution is achieved through the mapping process, and Y and are the true value and the predicted value after the transformation and weighting, respectively.

The arctanh mapping curve is shown in Figure 4. It can be seen from the figure that the mapping weighting method gradually increases the changes of the two end values, and the change range in the middle area is relatively small and the whole is smooth. After weighting, the weighted error at the two ends will increase compared to the original error. The closer to the two ends, the larger the error increment, so as to realize the distribution of weights in different regions. The actual effect of weighting is reflected in the actual target. The effect is shown in the figure 5 shown.

After the target value and the predicted value are weighted by mapping, MSE is applied to measure the error, and WMSE(10) is obtained, which is the loss function of the prediction model. By weighting, the training speed can be accelerated while reducing the model complexity to a certain extent.

Step 3: Evaluation Criteria

The main prediction objective of the present invention is to predict the trend position and the bottom turning point, and the accuracy rates of these two aspects are considered respectively.

The trend position accuracy is evaluated using the mean absolute error (MAE), and the turning point accuracy is designed with two indicators: bottom precision (LP) and top recall (HR).

Precision and recall are commonly used evaluation indicators for binary classification problems. In the classification algorithm, after comparing the predicted results with the real target, the following four markers can be used to indicate the predicted state.

TP: When the true value is a positive class, the predicted value is a positive class;

TN: When the true value is a positive class, the predicted value is a negative class;

FP: When the true value is negative, the predicted value is positive;

FN: The predicted value is the negative class when the true value is negative.

Where T is True, F is False, P is Positive, and N is Negative. And according to its values, the precision (11) and recall (12) are calculated. It can be seen from the formula that the precision indicates that when the predicted value is true, the actual value is also true, and the recall indicates that when the actual value is true, the predicted value is also true.

For the turning point at the bottom, when a turning point is predicted each time, it can provide investors with a strong buy signal, so we hope that every time the prediction is a turning point, it will be as accurate as possible, otherwise it will cause investment losses, so you need to pay attention to checking accuracy.

For the top turning point, every time a turning point is predicted, it will provide investors with a sell signal. If investors hold stocks, they should try to predict as accurately as possible when each top appears, so as to avoid missing selling opportunities, otherwise it will cause Investors lose, so the recall rate needs to be concerned.

In the classification problem, there are clear positive and negative classes, and the prediction is correct and wrong. This experiment is a regression algorithm, and the specific trend position value is obtained. There is no clear right or wrong, only the classification. Therefore, in order to use a similar method to analyze the trend The accuracy of the turning point forecast is measured, and the results are measured using the following methods.

The bottom and top are determined according to the trend position, and the turning point is determined based on a certain area. When the trend position value is less than 5, it is the bottom area, and when it is greater than 95, it is the top area. When the predicted value is less than 5, it is considered to be close to the bottom turning point, and the absolute error between it and the true value is measured, which is the bottom precision (Low Precision, LP). Similarly, when the true value is greater than 95, it is considered that it is close to the top turning point, and the absolute error between it and the predicted value is measured, which is the top recall (High Recall, HR).

In classification problems, precision and recall are usually contradictory. In extreme cases, one is very high and the other is very low. Therefore, LP and HR only predict the turning point or turning point of the model. The prediction effect in the region is evaluated. For the overall measurement, the present invention can use the mean absolute error (MAE) described above to reflect the fitting effect of the model to the overall trend because it is a regression calculation. According to the definition of trend position, the trend position is essentially a percentage of the trend. The maximum error between single points is 100, and the minimum error is 0. Therefore, the use of the average absolute error can reflect the prediction effect more intuitively.

The prediction result of this embodiment: when the time window M=100, the trend threshold δ=60%, the error functions LP, HR, and MAE are 8.07%, 11.86%, and 12.72%, respectively. Taking Baosteel Co., Ltd. as an example, the time window M=100, δ=60% prediction effect is shown in Figure 6.

Claims (4)

1. A stock trend prediction method based on an LSTM-CNN deep learning model is characterized by comprising the following steps:

step 1: data pre-processing

Stock data of a stock market are selected for analysis and verification, and an integral data set is divided into a training set and a testing set;

step 1.1: feature selection and normalization

The characteristics are transaction data and technical indexes, and the transaction data needs to be normalized; the technical index is obtained by calculating transaction data, and a normalization process is included in the calculation process;

step 1.2: calibrating trending locations

Researching the price trend of the stocks in the middle and long term, proposing turning points, quantifying the trend according to the turning points to obtain trend positions, and calibrating the data labels according to the trend positions;

step 2: the stock price trend is predicted by using an LSTM-CNN prediction model, and the method comprises the following steps:

step 2.1: constructing a network model, comprising the following steps:

the LSTM-CNN model comprises three parts, wherein one part is an LSTM time sequence characteristic learning layer; the other part is a CNN static characteristic learning layer which is synchronously carried out with the LSTM part; the last part is a fully-connected output layer, and the outputs of the first two models are connected in series and combined to construct a fully-connected neural network;

step 2.2: activating a function

Using the ELU as the activation function, the formula is (7):

φ(x)＝max(0,x) (7)

the network model needs to be configured in the training process, an activation function and input and output dimensions need to be set in an LSTM unit, and the size of the activation function and a convolution kernel need to be set in a CNN unit;

step 2.3: constructing a model training strategy, designing a weighting loss function for the training of the model, and comprising the following steps:

the value range of the trend position is a limited interval [0,100], the application values of different positions corresponding to the interval are different, and the invention performs weighting measurement on the interval to provide a loss function sensitive to the cost of a turning region;

firstly, the target value y and the predicted value are compared

Scaling and translating to make the target value in the interval [ -1,1 [ -1 [ ]]In the method, formula (8) is a conversion method, c is a small real number and is used for ensuring that a logarithm and a denominator calculation object are not zero, and arctaph formula (9) is used for carrying out the conversion on a target value z and a predicted value

Mapping is carried out, and weight assignment, Y and Y are realized through the mapping process

Respectively a real value and a predicted value after the transformation weighting;

after the target value and the predicted value are subjected to mapping weighting, MSE is applied to measure errors of the target value and the predicted value, WMSE (10) which is a loss function of the prediction model is obtained, the training speed can be accelerated through weighting, and meanwhile, the complexity of the model is reduced to a certain extent;

and step 3: evaluation criteria

The prediction target is a prediction trend position and a bottom turning point, and the accuracy rates of the two aspects are respectively considered;

the accuracy rate of the trend position is evaluated by using the average absolute error MAE, the accuracy rate of the turning point is evaluated by designing two indexes of a bottom precision ratio LP and a top recall ratio HR;

the precision ratio and the recall ratio are common evaluation indexes of the two classification problems, and in a classification algorithm, after a prediction result is compared with a real target according to the classification algorithm, the following four marks represent the prediction state of the classification algorithm;

TP: when the true value is the positive class, the predicted value is the positive class;

TN: when the true value is a positive class, the predicted value is a negative class;

FP: when the true value is a negative class, the predicted value is a positive class;

FN: when the true value is a negative class, the predicted value is a negative class;

wherein T is True, F is False, P is Positive, N is Negative, and calculate precision (11) and recall (12) according to each value; according to the formula, the precision ratio represents the condition that the actual value is true when the predicted value is true, and the recall ratio represents the condition that the predicted value is true when the true value is true;

for the bottom turning point, a strong buy signal is provided when the turning point occurs in each prediction, so that it is desirable that each prediction is as accurate as possible, otherwise investment loss is caused, and therefore attention is paid to precision ratio;

for the top turning point, when the turning point appears in each prediction, a selling signal is provided, and the top turning point appears in each prediction, the prediction is accurate as much as possible, so that the selling opportunity is avoided being missed, otherwise, the investment loss is caused, and therefore the recall ratio needs to be paid attention.

2. The method for predicting stock trend based on LSTM-CNN deep learning model as claimed in claim 1, wherein in step 1.1, the original transaction data is first normalized, in the stock data, in the market data, there is a continuously rising or descending market, the innovation height or new height is a common phenomenon in the financial data, therefore, the normalization method is designed for the price and the volume of;

price normalization: quantifying the stock data of each time window, calculating relative fall and rise by taking the stock sequence first value as a reference price, and independently normalizing the data in each time window as shown in a formula (1);

for each training sample, at T₀,T₁,T₂…T_i…T_nThe closing price corresponding to the time is P₀,P₁,P₂…P_i…P_nAt T_iThe characteristic value obtained by time normalization is Cx_iA constant term C;

and (3) normalization of volume of bargaining: the calculation is carried out on the whole time sequence of a stock, and the formula is shown as (2):

at T_iAt the moment, the volume of traffic is V_iThe eigenvalue obtained based on the volume normalization is Vx_iIf the constant term n is generally equal to the size of the time window, the first n time points of one stock are only used for data processing and are not marked into a sample;

and (4) completing the normalization of closing price and volume, and respectively calculating the technical index values of the sequence as MACD, RSI and KDJ-J, wherein the technical index values are all used as input data.

3. The LSTM-CNN deep learning model-based stock trend prediction method of claim 1 or 2, wherein in step 1.2, the trend turning points comprise a top turning point at a peak position and a bottom turning point at a valley position, when the price is at the bottom turning point, it means that the price is about to rise, and vice versa; in an actual market situation, prices are always in fluctuation, fluctuation with different amplitudes is generated, a plurality of turning points are generated, and the turning points of model identification are quantified, and the specific definition is as follows:

stock price is listed at T₁,T₂,T₃The closing prices corresponding to the time are respectively P₁,P₂,P₃Setting a parameter called trend threshold delta, when condition (4) is satisfied, and T₂The time price is T₁To T₃The lowest point in the period, T is considered₂The moment is the bottom turning point of a section of trend;

similar to the bottom turning point, when the condition (5) is satisfied, and T₂The time price is T₁To T₃The highest point in the period, then T is considered₂The moment is a top turning point of a section of trend, and it is worth noting that the calculation processes of the bottom turning point and the top turning point are both based on low price, otherwise, under the same threshold value delta, the bottom turning point and the top turning point are different in the fluctuation absolute value required for establishing, and deviation is generated;

after a certain bottom turning point is determined, the subsequent sequence is defined as an ascending trend until the next top turning point is met, and similarly, after a certain top turning point is determined, the subsequent sequence is defined as a descending trend until the next bottom turning point is met;

after determining the turning points at the bottom and the top of the time sequence, mapping the current stock price to the relative position in the trend, and using the trend position as the direct prediction target of an LSTM-CNN model to realize the prediction of the future price trend, wherein the calibration method of the trend position comprises the following steps:

1.2.1, according to the definition of the turning points, obtaining all bottom turning points and top turning points of a group of sequences, wherein the trend position at the bottom turning point is marked as 0, and the trend position at the top turning point is marked as 100;

1.2.2, the trend position value between the bottom and top turning points is calculated according to equation (6), where P_LAnd P_HA set of adjacent bottom and top inflection points;

the last trend of the time series cannot determine the top or bottom turning point of the last price, so that the part which cannot be determined is not marked and is not included in the training data.

4. The LSTM-CNN deep learning model based stock trend prediction method of claim 1 or 2, wherein in step 2.1, the first part is composed of three LSTM layers, the input matrix size of LSTM is, where k is the number of training features, M is the size of sliding time window, L1 layer contains LSTM units corresponding to M time nodes, the input of each LSTM unit is composed of k features, including transaction data and technical indicators, each unit in L1 layer is trained to output as the input of each unit in L2 layer, similarly, the output of L2 layer is used as the input of L3 layer, the training modes of L2 and L3 layers are the same as that of L1 layer, and finally, the output of LSTM unit corresponding to the last time node in L3 layer is used as the result of part a;

the second part consists of three convolutional layers and one LSTM layer, in this part, the input matrix is the same as in the first part; extracting local related information between price and technical indexes through three convolutional layers, connecting the three convolutional layers with an LSTM layer, and analyzing a time relation in the extracted local dependency information through the LSTM layer;

the last part consists of three completely connected layers, time and static information are respectively extracted by the LSTM and the CNN and are combined into a new characteristic sequence, the characteristic sequence is input by the completely connected layers, and finally, a final prediction result of the model, namely a trend position, is output; in addition, some nodes between two connecting layers can be disconnected by applying a random inactivation method between all connecting layers, the final result is not completely dependent on specific nodes, and the random inactivation method obviously reduces the over-fitting problem.

Images