Stock prediction and transaction method based on abnormal jump behavior analysis

Abstract

The invention relates to a stock prediction and transaction method based on abnormal jump behavior analysis, and belongs to the technical field of data processing. The method comprises the following steps: step 1, signal identification: identifying a stock with an abnormal jump behavior through a statistical analysis method, and screening out an abnormal signal for positioning stock data of the abnormal jump behavior; step 2, constructing a data set: extracting related data based on abnormal signals, constructing the data set containing labels and training features, and analyzing a fund relationship before and after an abnormal jump behavior; step 3, selecting a deep learning prediction model: selecting an appropriate deep learning model based on given transaction data, and setting parameters such as batches, network layers, input and output dimensions and the like; step 4, evaluating a prediction model: evaluating the model through accuracy, and checking the prediction effect and effectiveness of the fitting model; and 5, constructing a transaction method: constructing a time-selecting transaction strategy, setting a control bin, designing an interference and loss stopping mechanism, and testing a yield rate based on a classification result of the prediction model. According to the method, the potential relationship between funds and abnormal jump behaviors in the stock market can be better revealed, so that decision support with higher reference value is provided for investors.

Description

Technical Field

[0001] The invention belongs to the technical field of data processing, and in particular relates to a stock prediction and trading method based on abnormal jump behavior analysis. Background Art

[0002] Financial markets are the core of national economic operations. Exploring the laws of change, conducting effective market management, and improving the multi-level financial organizational system have become the main goals of governments and relevant institutions. With the rapid development of the Internet and its related industries, information technology has gradually become dominant, and investors' awareness has also changed significantly. More and more people have begun to pay attention to the financial field. Therefore, how to effectively analyze and predict financial markets in order to provide scientific decision-making support for investors has become a hot topic at present. The trading volume in the financial market is a key indicator: on the one hand, it reveals the market price fluctuation process and helps to study the micro-mechanism of price formation; on the other hand, it is an important basis for trading strategy design and investment risk management. Especially in quantitative trading, the losses caused by uncontrolled risks may be very serious. Therefore, the research on financial transaction prediction has important theoretical value and broad application prospects.

[0003] The stock market is a complex multivariable nonlinear dynamic system, and there is currently a lack of effective modeling methods in academia. At the same time, due to the high uncertainty of the stock market itself, traditional methods that attempt to predict stock prices by building precise models often do not produce ideal prediction results. The stock price itself and the various variables that affect the stock price show significant nonlinear characteristics, which requires the model to have strong nonlinear processing capabilities.

[0004] The correlation between various factors in the stock market is intricate, the primary and secondary relationships are constantly changing, and the quantitative relationship between variables is difficult to accurately extract. Therefore, it is difficult to use traditional forecasting methods to conduct quantitative analysis of the stock market. At the same time, the market is full of random factors, which have a significant impact on stock indexes and stock prices. Prices fluctuate violently, the noise is high, and there is strong nonlinearity and uncertainty. At present, many methods have been proposed for stock market forecasting at home and abroad. These methods have certain guiding significance in practice, but they also have shortcomings. For example, the extrapolation ability of the regression model is poor, the accuracy of the analogy coefficient method is not high, and the neural network has a large amount of calculation and is prone to overfitting. The introduction of deep learning technology effectively solves the shortcomings of traditional forecasting technology in terms of extrapolation, accuracy, calculation and overfitting. Summary of the invention

[0005] In view of the complexity and uncertainty of stock trading data and the defects in the existing technology, the present invention proposes a feature processing technology in combination with a deep learning algorithm, which can construct a financial transaction prediction model to more accurately predict the financial transaction data of any time period or any time in the future based on historical stock trading data.

[0006] The present invention discloses a stock prediction and trading method based on abnormal jump behavior analysis, comprising the following steps:

[0007] Step 1: Signal identification: Identify stocks with abnormal jumping behaviors through statistical analysis methods, filter out abnormal signals, and use them to locate stock data with abnormal jumping behaviors;

[0008] Step 2: Construct a data set: Extract relevant data based on abnormal signals, construct a data set containing labels and training features, and analyze the relationship between funds before and after abnormal jump behavior;

[0009] Step 3: Select a deep learning prediction model: Based on the given transaction data, select a suitable deep learning model and set parameters such as batch, number of network layers, input and output dimensions;

[0010] Step 4: Prediction model evaluation: Evaluate the model through accuracy to test the prediction effect and effectiveness of the fitted model;

[0011] Step 5: Construct a trading method: Based on the classification results of the prediction model, construct a timing trading strategy, set control positions, design a stop-profit and stop-loss mechanism, and test the rate of return.

[0012] Optionally, in step 1, the process of identifying abnormal signals in real-time stock trading price data specifically includes:

[0013] Data acquisition: Download the daily closing prices of all A-share stocks over a period of time from the Wind database.

[0014] Data partitioning: Divide the data into training set and test set.

[0015] Signal recognition condition setting: Set the drastic change condition, that is, within m trading days, there are at least k days with an increase or decrease of more than x.

[0016] Calculate the change: Calculate the change for the day according to the following formula to determine whether it meets the conditions of sharp fluctuations:

[0017] Change on the day = current closing price − previous day’s closing price

[0018] Sliding window method: Set the sliding window size to m, traverse the transaction data, identify the position where the change amplitude exceeds x and record the number of occurrences count. When count is greater than or equal to k, record the point as an abnormal signal.

[0019] Optionally, in step 2, the abnormal signal extraction data and the process of constructing a fund feature data set before and after the abnormal jump behavior specifically include:

[0020] Extract target data: Based on the abnormal signal identified in step 1, obtain the closing price data of t trading days after the signal, recorded as target.

[0021] Price change labeling: Calculate the price change of the target data segment and assign labels to the data according to the following criteria:

[0022] =

[0023] The specific classification criteria for labels include:

[0024] If the change range of the data segment is less than n times, it is identified as an oscillation or downward trend and recorded as 0; if the change range is between n times and 2n times, it is identified as a 1-fold upward trend and recorded as 1; if the change range is more than 2n times, it is identified as a 2-fold upward trend and recorded as 2.

[0025] The 0, 1, and 2 marked here are the targets of the data set, denoted as target_1.

[0026] Extract label data: Based on the abnormal signal identified in step 1, obtain the closing price data of t trading days before the signal, recorded as label.

[0027] Construct a data set. Based on the label, target_1 obtained before and after the abnormal signal identified in step 1, construct the data set {label, target_1} used by the deep learning model.

[0028] Optionally, in step 3, the process of selecting a deep learning model to predict stock market signals specifically includes:

[0029] Write Python programs that can better fit deep learning models for time series;

[0030] Send the data set to the prepared Python program;

[0031] Set the relevant parameters of the model and train the model.

[0032] Optionally, in step 4, the process of evaluating the prediction accuracy of the model by using corresponding evaluation indicators specifically includes:

[0033] In order to verify the effectiveness of the selected model, the present invention uses an out-of-sample test set for testing and uses accuracy as an evaluation indicator. The specific formula is:

[0034] =

[0035] Select the best model, compare the evaluation indicators of each model, use the model corresponding to the best evaluation indicator as the subsequent financial model, and construct a trading method.

[0036] Optionally, in step 5, the classification results of the prediction model corresponding to different timing trading strategies specifically include:

[0037] Prediction classification result 0: indicates that the maximum increase is less than 1 times, which is identified as an oscillation or downward trend, and no operation is performed;

[0038] Prediction classification result 1: indicates that the maximum increase is 1 to 2 times, and it is identified as a 1-fold increase trend. Buy on the first trading day after the abnormal point;

[0039] Prediction classification result 2: Indicates that the maximum increase is more than 2 times, identified as a 2-fold increase trend, and bought on the first trading day after the abnormal point.

[0040] Adding position management and risk control modules to the above trading methods can better control risks and increase profits, including:

[0041] Set position management: after the buy signal is triggered, the single stock purchase amount is calculated according to the preset fixed ratio, where the purchase amount is a fixed percentage cap of the total investment amount to reduce the risk exposure of a single transaction;

[0042] Design a stop-profit and stop-loss mechanism: when the predicted classification result is 1, if the stock price rises to a preset ratio (such as 1 times) compared to the purchase price, a sell operation is triggered; when the predicted classification result is 2, if the stock price rises to another preset ratio (such as 2 times) compared to the purchase price, a sell operation is triggered; after completing the buy operation, if the conditions for triggering a sell signal are not met within the preset time, the sell operation is automatically executed at the preset time point to avoid potential risks caused by long-term holding.

[0043] By obtaining classified signals in real time and combining them with trading methods to make investment decisions, we can ultimately analyze the capital characteristics before and after abnormal jumps in the stock market and achieve profit goals.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] The present invention provides an accurate and universal method for predicting the stock market. Through the method of the present invention, a stock prediction and trading method based on abnormal jump behavior analysis can be efficiently constructed to help investors make more accurate investment decisions in the stock market, improve the accuracy of market predictions, and effectively increase trading returns. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0047] Figure 1 This is a brief flow chart of a stock prediction and trading method based on abnormal jump behavior analysis according to an embodiment of the present invention;

[0048] Figure 2 This is a schematic effect diagram of feature extraction of a stock prediction model based on abnormal jump behavior analysis in an embodiment of the present invention;

[0049] Figure 3 It is a schematic effect diagram of three classifications of predicted classification results in the model data set of an embodiment of the present invention;

[0050] Figure 4 It is a detailed flow chart of a stock prediction model based on abnormal jump behavior analysis according to an embodiment of the present invention;

[0051] Figure 5 This is a detailed flow chart of a transaction analysis method based on abnormal jump behavior according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] A stock prediction and trading method based on abnormal jump behavior analysis, such as Figure 1 As shown, the method includes:

[0054] (1) Identify stocks with abnormal jumping behaviors through statistical analysis methods and screen out relevant abnormal signals. The specific process includes:

[0055] Download the daily closing price data of all A shares from January 2004 to June 2024 from the Wind database.

[0056] The data is divided into a training set (January 2004 to January 2019) and a test set (January 2019 to June 2024).

[0057] Feature engineering can refer to Figure 2 , specifically:

[0058] For example, step 1 S101 mainly includes:

[0059] Set the drastic change condition: within 5 trading days, there are 3 ups and downs exceeding 10%. The change range of signal recognition is calculated as follows:

[0060] =

[0061] Use a sliding window size of 5 to traverse the entire data set, identify the locations where drastic changes occur, and record them.

[0062] There must be a period of one year before and after the abnormal signal to further perform feature analysis on the data in this period. In the trading market, there are 250 trading days a year.

[0063] Therefore, abnormal signals with a length of less than 250 trading days before and after the signal are screened out and deleted.

[0064] (2) Extract relevant data based on abnormal signals, construct a data set containing labels and training features, and analyze the relationship between funds before and after the abnormal jump behavior. The process of step 2 S102 specifically includes:

[0065] Read the abnormal signal recorded in (1) and obtain the closing price data of the 250 trading days after the signal, recorded as target.

[0066] It is necessary to identify the price changes in the target data and then label them. The change is selected by the change range in the target data segment. The calculation process of the change range in the target is as follows:

[0067] =

[0068] refer to Figure 3 , record the target classification target_1, the specific classification is as follows:

[0069] If the change in the data segment is below 1, it is identified as an oscillation or downward trend and classified as 0;

[0070] If the change range of the data segment is between 1 and 2, it is identified as a 1-fold increase trend and classified as 1;

[0071] If the change in the data segment is greater than 2, it is identified as a 2-fold increase trend and classified as 2.

[0072] Read the abnormal signal recorded in (1) and obtain the closing price data of the 250 trading days before the signal, recorded as label.

[0073] Use the above feature engineering to obtain {label, abnormal signal, target_1} data.

[0074] Finally, delete the {abnormal signal} information in each piece of data and construct the data set {label, target_1} to be used by the deep learning model.

[0075] (3) By selecting a deep learning model and adjusting the parameters of the selected model, the prediction of stock trading signals is achieved. In order to better fit the time series data, the selected deep learning models include MLP, LSTM, GRU and Transformer. For details, please refer to Figure 4 , when analyzing the data and building the model, step three includes the following steps:

[0076] Optionally, the MLP network prediction model selected in step 3 S103 is specifically:

[0077] A neural network consists of three layers: input layer, hidden layer, and fully connected layer, where the fully connected layer also serves as the output layer of the network. Neural network training is completed through multiple iterations, each of which includes two stages: forward propagation and back propagation.

[0078] In the forward propagation phase, the input layer only serves as an input channel for data and does not participate in the calculation; each hidden layer performs a linear transformation on the input data, then applies a nonlinear activation function before passing it to the next layer. The output layer is calculated in a similar way to the hidden layer, but uses a different activation function. Subsequently, the loss value is calculated based on the model output and the label.

[0079] In the back-propagation phase, the gradient of the loss to the model output is first calculated, and then the gradient of each parameter of the hidden layer and the output layer is obtained. After obtaining the gradient, each parameter is updated: subtract the gradient and multiply it by the learning rate to complete the back-propagation process.

[0080] The hidden layer adopts the Dropout mechanism, which randomly inactivates some neurons during the training process, and the probability of inactivation of each neuron is 10%.

[0081] Optionally, the LSTM network prediction model selected in step 3 S103 is specifically:

[0082] Specifically, the LSTM network consists of an input gate, a forget gate, and an output gate. The input gate is responsible for storing new information into the memory cell, the forget gate determines which information needs to be removed from the memory cell, and the output gate controls the output of the current time step. Each gate is connected to a multiplication unit, and the input and output of the information flow and the state of the memory cell are controlled by setting the weights connecting the memory cell to the rest of the network. The input of the current time step and the hidden state of the previous time step are input into each gate control unit.

[0083] The mechanism of each gate is implemented through a fully connected layer with a Sigmoid activation function. The Sigmoid function scales the value to the range (0, 1) to indicate the degree of opening or closing of the gate. The calculation formulas for the input gate, forget gate, and output gate at time step t are as follows:

[0084] Input Gate: =

[0085] Forget Gate: =

[0086] Output Gate: =

[0087] in, , , Respectively represent the output of the input, forget and output gates at time step t, σ represents the Sigmoid activation function, represents the input of the current time step, represents the hidden state at the previous time step, , , , is the preset weight matrix, , , is the bias parameter.

[0088] Optionally, the GRU network prediction model selected in step 3 S103 is specifically:

[0089] The GRU (Gated Recurrent Unit) neural network is an improvement on the RNN. It is specifically used to process sequence data and can effectively capture the dependencies between data in the sequence. Compared with the RNN, the GRU alleviates the gradient vanishing and gradient exploding problems by introducing update gates and reset gates, thereby improving the ability to capture long-term dependencies.

[0090] Hidden State The calculation formula is:

[0091] =

[0092] =

[0093] in, and Represent the update gate and reset gate respectively:

[0094] =

[0095] =

[0096] here, , , represents the weight matrix, is the input of the current time step t, , , is the bias term, and σ is the activation function using sigmoid. The calculation of each gate depends on a different parameter matrix, and the dimensions of all gates and hidden states are consistent.

[0097] Optionally, the Transformer network prediction model selected in step 3 S103 is specifically:

[0098] The Encoder structure in Transformer is established as follows:

[0099] Through the Input Embedding operation, the training set data of the capital feature time series obtained in step 2 is converted into a sparse matrix ( ) form, the sparse matrix ( ) After embedding, positional encoding is performed. Sin-cosine position encoding is used, which is generated by sine and cosine functions of different frequencies, and then added to the corresponding position vector. The PE (positional encoding) calculation formula is as follows:

[0100]

[0101]

[0102] Among them, pos represents the absolute position of the element in the sequence vector; d represents the dimension of the vector; 2i and 2i+1 represent parity.

[0103] Then the sparse matrix after position encoding ( ) is sent to the multi-head attention mechanism layer, and the multiple groups of weight matrices in the multi-head attention mechanism are , , Multiply them together and transform them into multiple sets of Query (Q), Keys (K), and Values ​​(V) vectors required to calculate the attenton value.

[0104] Sparse Matrix( ) and the weight matrix , , Multiply to get the Q, K, V vectors:

[0105] Q =

[0106] K =

[0107] V =

[0108] Then display the correlation between the sequences in the form of a matrix:

[0109]

[0110] Through the softmax function, the correlation matrix between sequences is converted into a probability distribution between [0, 1], that is, converted into a probability distribution matrix α with values ​​distributed between [0, 1];

[0111] =

[0112] Multiple z matrices are concatenated through the Concact operation to obtain the matrix Z:

[0113] =

[0114] After obtaining the matrix Z, the sparse matrix Perform residual connection and normalization (Add&Norm) with matrix Z.

[0115] Add a residual block X to the matrix Z (to prevent degradation in deep neural network training and avoid the problem of gradient disappearance), and then normalize the result (Norm):

[0116] Sub_Z = Norm(X + Z)

[0117] The normalized vector of each feature is passed through a fully connected feedforward neural network (Feed Forward). The process is divided into a linear transformation, a ReLU nonlinearity, and another linear transformation to obtain feature extraction.

[0118] The Feed Forward layer formula is as follows:

[0119] =

[0120] The ReLU function represents the ReLU nonlinear transformation. , are the coefficients and constant terms multiplied by the first linear transformation, It is the constant term added for the second linear change.

[0121] The extracted features are added and normalized again, and then input into the next layer of Encoder.

[0122] The Decoder structure in Transformer is established as follows:

[0123] The initial stage inputs the starting symbol, and then takes the output of the previous moment Transformer (the first type of input) as input. After the embedding operation, a sparse matrix is ​​generated. And apply positional encoding.

[0124] Processing position-encoded sparse matrices And input to the Masked Multi-Head Attention module, which is the first multi-head attention layer and performs a sparse matrix on the input Alignment is performed to prevent the Decoder from accessing future prediction information. Output Q (Query) and pass it to the next layer of multi-head attention mechanism after Add & Norm operations.

[0125] The output of this layer and the K (Keys) and V (Values) (i.e., the second type of input) of the last layer of the encoder enter the second-layer multi-head attention mechanism together to generate an output sequence.

[0126] Output sequence and input Perform residual connection and normalization (Add & Norm) to obtain the normalized connection result.

[0127] The normalized result is further extracted through a fully connected feedforward neural network (Feed Forward Network), and residual connection and normalization (Add & Norm) are performed again to obtain the result vector.

[0128] The result vector is passed to the next layer of Decoder, and the above steps are repeated M times. The last layer of Decoder generates the final vector.

[0129] The final output of the decoder is mapped into an extended vector (logits vector) through the linear layer, and the score is converted into the occurrence probability of each stock classification through the softmax layer. Finally, the category with the highest probability is selected as the prediction result.

[0130] (4) Evaluate the prediction accuracy of the model using corresponding evaluation indicators. Step 4 S104 specifically includes the following steps:

[0131] In order to illustrate that the method proposed in this example can effectively use abnormal jump behavior to classify and analyze the future trend of stock daily closing prices, this example uses an out-of-sample test set for testing. By focusing on the accuracy of the indicators, the effectiveness of this method is evaluated:

[0132]

[0133] Acc is the accuracy, Success indicates the number of successfully predicted categories, and Total indicates the total number of validation sets. The final accuracy results are shown in Appendix 1:

[0134] Table 1

[0135] Model Acc MLP 84.82% LSTM 85.61% GRU 81.52% Transformer 84.92%

[0136] Select the model with the highest accuracy, namely LSTM, as the trading model, predict future stock trends, and build a trading method.

[0137] (5) Using the constructed financial transaction prediction model, a prediction signal is obtained and a transaction method is constructed. Step 5 S105 specifically includes the following steps:

[0138] There are three types of trading signals classified by the deep learning prediction model, corresponding to three timing trading strategies. By designing the corresponding timing trading strategy, the trading method is executed through the signals generated by the model in real-time trading, thereby making a profit. The following is a timing trading strategy designed for these three price changes. The specific trading details are as follows:

[0139] Prediction classification result 0: indicates that the maximum increase is less than 1 times, which is identified as an oscillation or downward trend, and no operation is performed;

[0140] Prediction classification result 1: indicates that the maximum increase is 1 to 2 times, and it is identified as a 1-fold increase trend. Buy on the first trading day after the abnormal point;

[0141] Prediction classification result 2: Indicates that the maximum increase is more than 2 times, identified as a 2-fold increase trend, and bought on the first trading day after the abnormal point.

[0142] In addition, the trading method constructed in this embodiment also includes position management and stop-profit and stop-loss mechanisms, the specific contents of which are as follows:

[0143] Set position management: The single stock purchase amount should take into account the total amount of each investment. In this example, the total amount is set to 100,000,000, and the fixed ratio is set to cap = 1%. When the buy signal is triggered, the amount of the stock purchased is 100,000,000 × 1% = 1,000,000;

[0144] Design a stop-profit and stop-loss mechanism: for predicted classification result 1, sell when the stock price rises to 1 times the ratio of the buying point; for predicted classification result 2, sell when the stock price rises to 2 times the ratio of the buying point; after completing the buy operation, if the sell signal is not triggered within the preset time, the system will automatically execute the sell operation at a fixed time point.

[0145] Trade according to the constructed trading method and verify the profitability of the method.

[0146] Reference for trading methods based on abnormal jump behavior analysis Figure 5 , the specific process is as follows:

[0147] First, obtain the prediction classification result fitted from the financial transaction model, and determine which category the prediction classification result belongs to. If the prediction classification result is 0, do nothing; otherwise, buy on the first trading day after the abnormal point. Optionally, buy according to the position management method set in this embodiment; after the purchase is completed, make a second judgment. If the prediction classification result is 1, sell when the stock price rises to 1 times the ratio of the buying point; otherwise, sell when the stock price rises to 2 times the ratio of the buying point; finally, calculate the rate of return of the transaction method after the purchase is completed to obtain the profit situation.

[0148] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A stock prediction and trading method based on abnormal jump behavior analysis, characterized in that: Step 1: Signal identification: Identify stocks with abnormal jumping behaviors through statistical analysis methods, filter out abnormal signals, and use them to locate stock data with abnormal jumping behaviors; Step 2: Construct a data set: Extract relevant data based on abnormal signals, construct a data set containing labels and training features, and analyze the relationship between funds before and after abnormal jump behavior; Step 3: Select a deep learning prediction model: Based on the given transaction data, select a suitable deep learning model and set parameters such as batch, number of network layers, input and output dimensions; Step 4: Prediction model evaluation: Evaluate the model through accuracy to test the prediction effect and effectiveness of the fitted model; Step 5: Construct a trading method: Based on the classification results of the prediction model, construct a timing trading strategy, set control positions, design a stop-profit and stop-loss mechanism, and test the rate of return.

2. According to step 1 of claim 1, the signal recognition step is characterized by: Identify abnormal signals through real-time stock trading price data, including: Get historical trading data of stocks and divide them according to the set time range; Set abnormal jump behavior conditions. If the increase or decrease exceeds the set threshold and occurs a certain number of times during a specified trading day, the data point is determined to be an abnormal signal.

3. According to step 2 of claim 1, the specific step of constructing a data set is characterized by: Extract closing price data of several trading days before and after the abnormal signal to construct label and feature data sets; Specifically, according to the increase setting label, all price data of several trading days after the abnormal signal are compared with the initial trading day and divided into three types of trading signals: the maximum increase is less than 1 times, the maximum increase is 1 to 2 times, and the maximum increase is more than 2 times; The price data of several trading days before the abnormal signal are normalized and used as the input features of the model.

4. In step 3 according to claim 1, the key step of establishing a deep learning prediction model is characterized by: Feature screening: Taking the abnormal jump behavior in the stock time series data as the key information, extract the key feature variables from the sequence data of the previous period of the key information and perform normalization processing; Labeling: Taking the abnormal jump behavior in the stock time series data as the key information, analyze the key feature variables of the sequence data for a period of time after the key information and classify them.

5. In step 4 according to claim 1, the model evaluation step is characterized by: Specifically, when evaluating the model, the classification accuracy is set; based on the constructed deep learning model, combined with the test sample set, the model is evaluated through the evaluation indicators to verify the prediction effect and effectiveness of the fitted model.

6. In step 5 of claim 1, the transaction construction method is characterized by: Based on the classification results of the prediction model, different timing trading strategies are corresponding, specifically; Prediction classification result 0: indicates that the maximum increase is less than 1 times, which is identified as an oscillation or downward trend, and no operation is performed; Prediction classification result 1: indicates that the maximum increase is 1 to 2 times, and it is identified as a 1-fold increase trend. Buy on the first trading day after the abnormal point; Prediction classification result 2: Indicates that the maximum increase is more than 2 times, identified as a 2-fold increase trend, and bought on the first trading day after the abnormal point.

7. In step 5 of claim 1, the step of setting the control position is characterized by: After a buy signal is triggered, the single stock purchase amount is calculated based on a preset fixed ratio, where the purchase amount is a fixed percentage of the total investment amount to reduce the risk exposure of a single transaction.

8. In step 5 of claim 1, the stop-profit and stop-loss mechanism is designed, characterized in that: For the prediction classification result 1: sell when the stock price rises to 1 times the buying point; For prediction classification result 2: sell when the stock price rises to 2 times the buying point; After completing the buy operation, if the sell signal is not triggered within the preset time, the system will automatically execute the sell operation at a fixed time point.

Images