CN111445339A - Price limiting single prediction analysis method and system for bilinear attention convolution neural network - Google Patents
Price limiting single prediction analysis method and system for bilinear attention convolution neural network Download PDFInfo
- Publication number
- CN111445339A CN111445339A CN202010312349.4A CN202010312349A CN111445339A CN 111445339 A CN111445339 A CN 111445339A CN 202010312349 A CN202010312349 A CN 202010312349A CN 111445339 A CN111445339 A CN 111445339A
- Authority
- CN
- China
- Prior art keywords
- layer
- neural network
- bilinear
- bilinear attention
- network model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000004458 analytical method Methods 0.000 title claims abstract description 26
- 238000013528 artificial neural network Methods 0.000 title claims description 20
- 238000012549 training Methods 0.000 claims abstract description 59
- 238000000034 method Methods 0.000 claims abstract description 31
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 28
- 230000008569 process Effects 0.000 claims abstract description 16
- 230000007246 mechanism Effects 0.000 claims abstract description 14
- 238000011176 pooling Methods 0.000 claims abstract description 14
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims abstract description 11
- 230000006870 function Effects 0.000 claims description 37
- 238000003062 neural network model Methods 0.000 claims description 25
- 238000012360 testing method Methods 0.000 claims description 24
- 238000007781 pre-processing Methods 0.000 claims description 12
- 210000002569 neuron Anatomy 0.000 claims description 11
- 230000004913 activation Effects 0.000 claims description 9
- 239000011159 matrix material Substances 0.000 claims description 8
- 230000008034 disappearance Effects 0.000 claims description 6
- 238000004880 explosion Methods 0.000 claims description 6
- 230000000630 rising effect Effects 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 5
- 230000002441 reversible effect Effects 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 3
- 230000005540 biological transmission Effects 0.000 claims description 2
- 238000011425 standardization method Methods 0.000 claims description 2
- 238000013135 deep learning Methods 0.000 description 5
- 238000011156 evaluation Methods 0.000 description 4
- 238000005457 optimization Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 1
- 238000012733 comparative method Methods 0.000 description 1
- 239000002131 composite material Substances 0.000 description 1
- 230000008451 emotion Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000002759 z-score normalization Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
- G06Q40/06—Asset management; Financial planning or analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention discloses a bilinear attention convolutional neural network price limit single prediction analysis method and a bilinear attention mechanism on the basis of a convolutional neural network, the attention of the network to time information is enhanced, the number of model parameters is reduced through a pooling layer and a Dropout layer to prevent the occurrence of an overfitting phenomenon, the calculated amount is reduced, in the back propagation of a network training process, an Adam algorithm is used for calculating and updating weight, the traditional gradient descent process is optimized, parameters of an updated model are optimized through a loss function L, and therefore the bilinear attention convolutional neural network is optimized.
Description
Technical Field
The invention relates to the field of price limit order analysis, in particular to a bilinear attention convolution neural network price limit order prediction analysis method and system.
Background
The financial market reflects not only the macroscopic economy of a country, but also the future economic trends of various companies or goods, and is a complex and variable large system influenced by the market and market participants, and predicting the future fluctuation of the market is a very challenging task.
The research of deep learning is explosively increased, and the application is also vigorously developed. Deep learning has been widely applied in the fields of voice, image and video, but as many practitioners recognize the field of wind control that the method of deep learning cannot be used, the method is less applied in the financial field because the method is a black box model and has poor interpretability. Convolutional Neural Networks (CNN) have been widely applied to the fields of image motion recognition, text emotion classification, and the like as a representative algorithm for deep learning, but are also rarely applied to the financial field. Judging the future trend of the financial market is one of the most difficult tasks for investors, and finding a neural network model suitable for the stock market is also a difficult problem. The stock market comprises transaction data, financial data, macroscopic data and the like, the transaction data comprises stroke-by-stroke data, time-of-use data, price limit order data and market price order data, and the analysis of the price limit order data is also an important problem.
The limiting Order includes purchase orders and sales orders of multiple investors within a time, can reflect expectations and divergences of investors to the bid, the interaction situation of market and investors, and the trend of short-term population in the future, and does not need to carry out complex text analysis, and does not need a deep model with high complexity, and is suitable for analyzing the financial market.
Therefore, a solution is needed to optimize the convolutional neural network model and enhance the time concern of the model, and on the basis of the optimization, the prediction analysis is performed on the price limit order.
Disclosure of Invention
The invention aims to provide a bilinear attention convolution neural network price limiting single prediction analysis method and system.
In order to solve the technical problem, the invention provides a bilinear attention convolution neural network price limiting single prediction analysis method, which comprises the following steps:
step 1: acquiring a data set of a stock price limit list, and dividing the data set into a training set and a testing set;
step 2: adding a bilinear attention mechanism based on the convolutional neural network, and establishing a bilinear attention convolutional neural network model;
step 3, inputting the training set into a bilinear attention convolution neural network model to train the model, calculating and reversely updating the weight and the offset value of the whole neural network model by adopting a loss function L and an Adam algorithm, and finishing training to obtain the trained bilinear attention convolution neural network model;
and 4, step 4: and inputting the test set into the trained bilinear attention convolution neural network model to obtain corresponding output data, namely a prediction result.
Further, the training set and the test set in step 1 are both preprocessed.
Further, the preprocessing method of the data set is a z-score normalization method for normalizing data, and the specific steps are as follows:
step 1-1: acquiring a data set
And label therefor
Where T is the total number of samples, xtIs a D-dimensional sample corresponding to the time point t, D is a characteristic number, and
each time point t comprises the first n grades of quotations, and the ith grade of quotation comprises the pre-purchase price
And pre-buy volume
And a pre-sale price
And pre-sell volume
ytRepresenting the trend of the stock price corresponding to the time point t at the time point t + m, is a matrix with the size of 3 × 1, each value in the matrix represents rising, falling and invariant respectively, and m is the future time period predicted by the model;
step 1-2: normalizing the original data to ensure that the mean value of the processed data is 0 and the standard deviation is 1, wherein the normalization formula is as follows:
wherein
Is the mean, σ, of the d-th dimension feature in the raw datadStandard deviation of the d-dimension characteristic of the original data;
step 1-3: the normalized data set needs to be processed as an image-like data set in order to use data from a past period of time to predict the price trends of stocks. Class image sample x'tIs composed of normalized samples from time point t-delta t +1 to time point t, the size of class image sample is delta t × D, delta t is time span and is labeled yt；
Step 1-4: after the data set I goes through steps 1-1, 1-2 and 1-3, the data set I is divided into a training set U and a testing set X.
Further, the bilinear attention convolutional neural network model established in the step 2 comprises an 8-layer convolutional network and a bilinear attention mechanism.
Further, the 8-layer convolutional network is:
the layer 1 is a two-dimensional convolutional layer, the convolutional kernel size is 4 × 4n, the number of channels is 16, the layer 2 is a Reshape layer, the layer 3 is a one-dimensional convolutional layer, the convolutional kernel size is 3, the number of channels is 16, the layer 4 is a one-dimensional maximum pooling layer, the layer 5 is a one-dimensional convolutional layer, the convolutional kernel size is 3, the number of channels is 32, the layer 6 is a one-dimensional convolutional layer, the convolutional kernel size is 3, the number of channels is 32, the layer 7 is a one-dimensional maximum pooling layer, the layer 8 is a fully-connected layer, and the number of neurons is 3.
Further, the bilinear attention mechanism includes a bilinear projection layer, a Dropout layer, and a bilinear attention layer 3 layer, which are respectively:
bilinear projection layer, on which the data set U is transformed into U1The related formula is:
U1＝φ(W1UW2+B)
wherein, the d column in the training set contains the variation value of the d characteristic in the time T, and the T row shows the time TBetween points D values of different characteristics, W1∈RΔt'×Δt，W2∈RD×D'，B∈RΔt'×D'The phi function is a Relu activation function, so that the problems of gradient explosion and gradient disappearance can be avoided, the output of the layer is finally obtained, and the information of two dimensions of the input data of the layer is respectively stored in W1And W2In, B is a bias matrix;
the Dropout layer reduces the number of parameters by randomly deleting part of neurons, reduces the calculation amount and prevents an overfitting phenomenon;
bilinear attention layer, on which the data set U1Respectively through W1' obtaining F, by W2' get E, get A by Softmax function, get W3B, phi to obtain U2The related formula is:
F＝U1W1'
E＝W2'F
U2＝φ(W3A+B)
wherein, W1'∈RD'×D”，W2'∈RΔt'×Δt'，W3∈RΔt”×Δt'，B∈RΔt”×D”，aijAnd eijThe elements with the positions (i, j) in the A and the E are respectively represented, and the phi function is a Relu activation function, so that the problems of gradient explosion and gradient disappearance can be avoided, and the output of the layer is finally obtained. In this process, pass W1' obtaining a feature space RD'，W2The diagonal elements of' are all fixed to be 1/delta t, and the weight of each piece of time information is not changed, so that the time information corresponding to the d-th feature is kept unchanged, and the attention to the time information is strengthened.
Further, the loss function L in step 3 is:
wherein a iscIs an indicator variable, the prediction class is 1 if the prediction class is the same as the actual class, otherwise, the prediction class is 0, pcIs the predicted probability that the sample belongs to class c.
Further, the process of training the bilinear attention convolution neural network model by using the training set in the step 3 is as follows:
step 3-1, inputting the training set U into the network model for forward transmission, and in the reverse process of the training process, calculating and reversely updating the weight and the bias value of the whole neural network model by adopting a loss function L and an Adam algorithm;
step 3-2: and (3) repeating the step (3-1) until a preset training end condition is reached, stopping training, determining the weight value and the bias value of each neuron in the network model at the moment, and obtaining the trained bilinear attention convolution neural network model.
Further, the activation functions used by all convolution layers in the bilinear attention convolution neural network model in step 3 are Re L U functions, and the Softmax function is used by the full-connected layer.
The invention also provides a bilinear attention convolution neural network price limit sheet prediction analysis system, which comprises a data preprocessing module, a network training module and a stock price limit sheet trend prediction module:
in the data preprocessing module, preprocessing a data set to generate a training set and a test set, wherein the training set is input into the network training module, and the test set is input into the stock limit order trend prediction module;
in the network training module, a bilinear attention convolutional neural network model is constructed, the network model is optimized by using a pooling layer, a Dropout layer, a loss function L and an Adam algorithm, a training set is input into the network model to train the network model, and the trained bilinear attention convolutional neural network model is obtained after training is finished;
in the stock price limit single trend prediction module, a test set is input into a trained bilinear attention convolution neural network model to obtain corresponding output data, namely a prediction result.
By the scheme, the bilinear attention convolutional neural network price limiting list prediction analysis method has the advantages that a bilinear attention mechanism is added on the basis of a convolutional neural network, attention of the network to time information is enhanced, the number of model parameters is reduced through a pooling layer and a Dropout layer, the occurrence of an overfitting phenomenon is prevented, the calculated amount is reduced, in the backward propagation of the network training process, the Adam algorithm is used for calculating and updating the weight, the traditional gradient descent process is optimized, parameters of the updated model are optimized through a loss function L, and therefore the bilinear attention convolutional neural network is optimized.
The foregoing description is only an overview of the technical solutions of the present invention, and in order to make the technical solutions of the present invention more clearly understood and to implement them in accordance with the contents of the description, the following detailed description is given with reference to the preferred embodiments of the present invention and the accompanying drawings.
Drawings
Fig. 1 is a schematic structural diagram of a bilinear attention convolution neural network model in the present invention.
FIG. 2 is a flow chart of a method for bilinear attention layer in the present invention.
Fig. 3 is a schematic structural diagram of a bilinear attention convolution neural network price limit order analysis system.
Detailed Description
The following detailed description of embodiments of the present invention is provided in connection with the accompanying drawings and examples. The following examples are intended to illustrate the invention but are not intended to limit the scope of the invention.
This embodiment uses the open data set FI-2010 data set of the stock quotation (data source: Ntakaris A, Magris M, Kanniainen J, et al. benchmark dataset for mid-price for estimating the limit of the book data with the machine learning methods [ J ]. Journal of estimating, 2018,37(8):852 866). The data set comprises five Finnish stocks of the Nordic group of Nardack, a total of 45 ten thousand samples, labels of each limit order are divided into 5 types according to predicted time points, the data set is divided into two parts, one part comprises a bidding period, the other part does not comprise the bidding period, and each sample has 144 characteristics. In this embodiment, 1 of the tags is selected for prediction and does not include a bidding period, and the first 40 characteristics are selected, where the 40 characteristics are the pre-bid price and the pre-bid volume, and the pre-ask price and the pre-ask volume, respectively, of the first ten grades in the limited price list.
Example (b): a price limiting single prediction analysis method for a bilinear attention convolution neural network comprises the following steps:
step 1: preprocessing a data set of the stock price limit list by adopting a z-score standardization method, and dividing the data set into a training set and a testing set:
step 1-1: acquiring a data set
And labels therefor
Where T is the total number of samples, xtIs a 40-dimensional sample corresponding to the time point t, and 40 is a feature number. Training centralization
Each time point t comprises the first 10 th market, the ith market comprises the pre-purchase price
And pre-buy volume
And a pre-sale price
And pre-sell volume
ytRepresenting the trend of the stock price at time point t + m corresponding to time point t, is a matrix with the size of 3 × 1Each value in (a) represents a rise, a fall and a constant, respectively, and m is the future time period predicted by the model
Step 1-2: normalizing the original data to ensure that the mean value of the processed data is 0 and the standard deviation is 1, wherein the normalization formula is as follows:
wherein
Is the mean, σ, of the d-th dimension feature in the raw datadStandard deviation of the d-dimension characteristic of the original data;
step 1-3: the normalized data set needs to be processed as an image-like data set in order to use data from a past period of time to predict the price trends of stocks. Class image sample x'tIs composed of normalized samples from time point t-delta t +1 to time point t, the size of class image sample is delta t × D, delta t is time span and is labeled yt；
Step 1-4: after the data set I is processed through steps 1-1, 1-2 and 1-3, 70% of the preprocessed data set is selected as a training set U and 30% is selected as a test set X in the example in order to achieve better effect and better generalization ability.
Step 2: and adding a bilinear attention mechanism based on the convolutional neural network, and establishing a bilinear attention convolutional neural network model which comprises an 8-layer convolutional network and the bilinear attention mechanism.
As shown in fig. 1, the 8-layer convolutional network includes:
the layer 1 is a two-dimensional convolution layer, the size of a convolution kernel is 4 × 4n, the number of channels is 16, and the two-dimensional convolution layer is used for learning local features of input data;
the layer 2 is a Reshape layer and is used for changing the dimensionality of data and converting input two-dimensional data into one-dimensional data;
the 3 rd layer is a one-dimensional convolution layer, the size of a convolution kernel is 3, the number of channels is 16, and the 3 rd layer is used for learning the local characteristics of the one-dimensional data output by the upper layer;
the 4 th layer is a one-dimensional maximum pooling layer, the number of parameters output by the upper layer is reduced through a filter of the maximum pooling layer, and an over-fitting phenomenon is prevented;
the 5th layer is a one-dimensional convolution layer, the size of a convolution kernel is 3, the number of channels is 32, and the convolution kernel is used for learning local features output by the pooling layer;
the 6 th layer is a one-dimensional convolution layer, the size of a convolution kernel is 3, the number of channels is 32, the depth of the layer is increased, and the fitting capacity of the model is improved;
the 7 th layer is a one-dimensional maximum pooling layer and is used for reducing the number of parameters output by the previous layer again and optimizing the model;
the 8 th layer is a full connection layer, the number of the neurons is 3, and the neurons are used for realizing classification of prediction results and are divided into 3 types of rising, falling and unchanging.
All the convolution layers use the Re L U function as the activation function, and the full connection layer uses the Softmax function to calculate the output value of the neuron, and the output value is transmitted to the next layer for calculation.
The bilinear attention mechanism comprises a bilinear projection layer, a Dropout layer and a bilinear attention layer 3, which are respectively:
bilinear projection layer, on which the data set U is transformed into U1The formula is as follows:
U1＝φ(W1UW2+B)
wherein the D column in the training set contains the variation value of the D characteristic in time T, the T row represents the values of D different characteristics at T time point, W1∈RΔt'×Δt，W2∈RD×D'，B∈RΔt'×D'The phi function is a Relu activation function, so that the problems of gradient explosion and gradient disappearance can be avoided, the output of the layer is finally obtained, and the information of two dimensions of the input data of the layer is respectively stored in W1And W2In (1), B is a bias matrix.
The Dropout layer reduces the number of parameters by randomly deleting part of neurons, reduces the calculation amount and prevents an overfitting phenomenon;
as shown in FIG. 2, a bilinear attention layer, at which the data set U is located1Respectively go toOver W1' obtaining F, by W2' get E, get A by Softmax function, get W3B, phi to obtain U2The involved algorithms are:
F＝U1W1'
E＝W2'F
U2＝φ(W3A+B)
wherein, W1'∈RD'×D”，W2'∈RΔt'×Δt'，W3∈RΔt”×Δt'，B∈RΔt”×D”，aijAnd eijThe elements with the positions (i, j) in the A and the E are respectively represented, and the phi function is a Relu activation function, so that the problems of gradient explosion and gradient disappearance can be avoided, and the output of the layer is finally obtained. In this process, pass W1' obtaining a feature space RD'，W2The diagonal elements of' are all fixed to be 1/delta t, and the weight of each piece of time information is not changed, so that the time information corresponding to the d-th feature is kept unchanged, and the attention to the time information is strengthened.
And step 3: use training set
Training is carried out on the bilinear attention-enhancing convolutional neural network shown in fig. 1, and the results of the two parts are combined into a network model output value through the 8-layer convolutional neural network and the bilinear attention mechanism in sequence:
step 3-1, inputting the training set U into the network model for forward transfer, and performing reverse update optimization to update all parameters of the whole neural network model by adopting a loss function L, wherein the loss function L is as follows:
wherein a iscIs an indicator variable, the prediction class is 1 if it is the same as the actual class, otherwiseIs 0, pcIs the predicted probability that the sample belongs to class c; in the reverse updating of the training process, calculating and reversely updating the weight and the bias value of the whole neural network model through an Adam algorithm;
step 3-2: and repeating the step 3-1 until a training end condition is reached, training 200 rounds in the embodiment, stopping training, determining a weight value and a bias value of each neuron in the network model at the moment, and obtaining the trained bilinear attention convolution neural network model.
And 4, step 4: and inputting the test set X into the trained bilinear attention convolution neural network model to obtain a predicted value of the stock at the time point t.
In order to better illustrate the technical effects of the present invention, the examples were conducted to test the present invention. The predicted value of the stock at the time point t represents the trend of the stock in the future t to (t + m), the trend is divided into rising, falling and unchanging, and each kind of results of the predicted value is evaluated, wherein the rising is 1 and the rising is not 0; the dip is 1, not 0; not to 1, but to 0. The predicted values are subjected to mean processing after evaluation and are compared with the true values, and the predicted values are represented by four indexes of Precision (Precision), Recall (Recall), F1 Score (F1-Score) and accuracy (accuracy), and the calculation method is as follows:
wherein, TP, FP, FN and TN are four types of evaluation parameters: TP represents that the predicted value is the same as the actual value, and the predicted value is 1; FP represents that the predicted value is different from the actual value, and the predicted value is 1; TN represents that the predicted value is the same as the actual value, and the predicted value is 0; FN indicates that the predicted value is different from the actual value, and the predicted value is 0. Table 1 is a table of four types of parameters used for evaluation.
| True value (1) | True value (0) | |
| Predicted value (1) | TP | FP |
| Predicted value (0) | FN | TN |
TABLE 1 four-class parameter Table for evaluation
In this embodiment, 3 methods are set for comparison analysis, wherein the SVM, M L P (SVM and M L P are shown in the literature "Tsantekisis A, Passalis N, Tefas A, et al. use missing to detect temporal change indexes in capillary references [ C ]// 201725 th European Signal processing Conference (SIPEUCO). IEEE,2017:2511 2515") and CNN (see the literature "Tsantekisis A, Passalis N, Tefas A, et al. Forxecuting storage test library, the inventor number of using volumetric work [ C ]//2017IEEE 19th coherence business networks (C.19 th/2017 th business networks), the results are shown as accurate learning indexes, and the results are shown in the form of a model of shallow learning index (Cscan) and a deep learning index (Ref) of the SVM 1, and the results are shown as accurate learning index (F-1). The model, the results are shown as accurate learning index, and the results are shown as accurate learning index (Ref. C).
Table 2 is a table comparing the results of the test according to the method of the invention and the 3 comparative methods.
| Method of producing a composite material | Precision ratio (%) | Recall (%) | F1 fraction (%) | Accuracy (%) |
| SVM | 39.62 | 44.92 | 35.88 | - |
| MLP | 47.81 | 60.78 | 48.27 | - |
| CNN | 67.67 | 56.33 | 60.97 | 76.53 |
| The invention | 74.67 | 63.33 | 68.34 | 80.51 |
TABLE 2 comparison of the test results of the method of the invention and SVM, M L P, CNN
As shown in Table 2, the accuracy rate, the recall rate, the F1 score and the accuracy rate of the method are all superior to other comparison methods under the same condition, the method strengthens the attention of the network to time information by optimizing a convolution network layer and increasing a bilinear attention mechanism on the basis of the optimization, prevents the occurrence of an overfitting phenomenon by reducing the number of model parameters through a pooling layer and a Dropout layer, reduces the calculated amount, calculates and updates the weight by using an Adam algorithm in the back propagation of a network training process, optimizes the traditional gradient descent process, optimizes the parameters of an updated model through a loss function L, and optimizes a bilinear attention convolution neural network.
As shown in fig. 3, the bilinear attention convolution neural network price limit order analysis system provided in the embodiment of the present invention includes a data preprocessing module, a network training module, and a stock price limit order trend prediction module:
the data preprocessing module is used for preprocessing the data set to generate a training set and a test set, the training set is input into the network training module, and the test set is input into the stock limit order trend prediction module;
in the network training module, a bilinear attention convolutional neural network model is constructed, the network model is optimized by using a pooling layer, a Dropout layer, a loss function L and an Adam algorithm, a training set is input into the network model to train the network model, and the trained bilinear attention convolutional neural network model is obtained after training is finished;
in the stock price limit single trend prediction module, the test set is input into a trained bilinear attention convolution neural network model to obtain corresponding output data, namely a prediction result.
The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, it should be noted that, for those skilled in the art, many modifications and variations can be made without departing from the technical principle of the present invention, and these modifications and variations should also be regarded as the protection scope of the present invention.
Claims (10)
1. A price limiting single prediction analysis method for a bilinear attention convolution neural network is characterized by comprising the following steps:
step 1: acquiring a data set of a stock price limit list, and dividing the data set into a training set and a testing set;
step 2: adding a bilinear attention mechanism based on the convolutional neural network, and establishing a bilinear attention convolutional neural network model;
step 3, inputting the training set into a bilinear attention convolution neural network model to train the model, calculating and reversely optimizing and updating the weight and the offset value of the whole neural network model by adopting a loss function L and an Adam algorithm, and finishing training to obtain the trained bilinear attention convolution neural network model;
and 4, step 4: and inputting the test set into the trained bilinear attention convolution neural network model to obtain corresponding output data, namely a prediction result.
2. The bilinear attention convolution neural network pricing unit prediction analysis method of claim 1, characterized in that: in the step 1, the training set and the test set are preprocessed, and data is normalized.
3. The bilinear attention convolution neural network pricing unit prediction analysis method of claim 2, characterized in that: the preprocessing method of the data set is a z-score standardization method, and comprises the following specific steps:
step 1-1: obtainFetch datasets
And label therefor
Where T is the total number of samples, xtIs a D-dimensional sample corresponding to the time point t, D is a characteristic number, and
each time point t comprises the first n-th stage market, the ith stage market comprises the pre-purchase price Pi bidAnd pre-buy volume Vi bidAnd a pre-sale price Pi askAnd a pre-sale volume Vi ask，ytRepresenting the trend of the stock price corresponding to the time point t at the time point t + m, is a matrix with the size of 3 × 1, each value in the matrix represents rising, falling and invariant respectively, and m is the future time period predicted by the model;
step 1-2: normalizing the original data to ensure that the mean value of the processed data is 0 and the standard deviation is 1, wherein the normalization formula is as follows:
wherein
Is the mean, σ, of the d-th dimension feature in the raw datadStandard deviation of the d-dimension characteristic of the original data;
step 1-3: the normalized data set needs to be processed as an image-like data set in order to use data from a past period of time to predict the price trends of stocks. Class image sample x'tIs composed of normalized samples from time point t-delta t +1 to time point t, the size of class image sample is delta t × D, delta t is time span and is labeled yt；
Step 1-4: after the data set I goes through steps 1-1, 1-2 and 1-3, the data set I is divided into a training set U and a testing set X.
4. The bilinear attention convolution neural network pricing unit prediction analysis method of claim 1, characterized in that: the bilinear attention convolutional neural network model established in the step 2 comprises an 8-layer convolutional network and a bilinear attention mechanism.
5. The bilinear attention convolution neural network pricing unit prediction analysis method of claim 4, characterized in that: the 8-layer convolutional network is as follows:
the layer 1 is a two-dimensional convolutional layer, the convolutional kernel size is 4 × 4n, the number of channels is 16, the layer 2 is a Reshape layer, the layer 3 is a one-dimensional convolutional layer, the convolutional kernel size is 3, the number of channels is 16, the layer 4 is a one-dimensional maximum pooling layer, the layer 5 is a one-dimensional convolutional layer, the convolutional kernel size is 3, the number of channels is 32, the layer 6 is a one-dimensional convolutional layer, the convolutional kernel size is 3, the number of channels is 32, the layer 7 is a one-dimensional maximum pooling layer, the layer 8 is a fully-connected layer, and the number of neurons is 3.
6. The bilinear attention convolution neural network pricing unit prediction analysis method of claim 4, characterized in that: the bilinear attention mechanism comprises a bilinear projection layer, a Dropout layer and a bilinear attention layer 3, wherein the bilinear attention mechanism comprises the following layers:
bilinear projection layer, on which the data set U is transformed into U1The related formula is:
U1＝φ(W1UW2+B)
wherein the D column in the training set contains the variation value of the D characteristic in time T, the T row represents the values of D different characteristics at T time point, W1∈RΔt'×Δt，W2∈RD×D'，B∈RΔt'×D'The phi function is a Relu activation function, so that the problems of gradient explosion and gradient disappearance can be avoided, the output of the layer is finally obtained, and the information of two dimensions of the input data of the layer is respectively stored in W1And W2In, B is a bias matrix;
the Dropout layer reduces the number of parameters by randomly deleting part of neurons, reduces the calculation amount and prevents an overfitting phenomenon;
bilinear attention layer, on which the data set U1Respectively through W1' obtaining F, by W2' get E, get A by Softmax function, get W3B, phi to obtain U2The related formula is:
F＝U1W1'
E＝W2'F
U2＝φ(W3A+B)
wherein, W1'∈RD'×D”，W2'∈RΔt'×Δt'，W3∈RΔt”×Δt'，B∈RΔt”×D”，aijAnd eijThe elements with the positions (i, j) in the A and the E are respectively represented, and the phi function is a Relu activation function, so that the problems of gradient explosion and gradient disappearance can be avoided, and the output of the layer is finally obtained. In this process, pass W1' obtaining a feature space RD'，W2The diagonal elements of' are all fixed to be 1/delta t, and the weight of each piece of time information is not changed, so that the time information corresponding to the d-th feature is kept unchanged, and the attention to the time information is strengthened.
7. The bilinear attention convolution neural network price limiting univocal prediction analysis method of claim 1, characterized in that the loss function L in the step 3 is:
wherein a iscIs an indicator variable, the prediction class is 1 if the prediction class is the same as the actual class, otherwise, the prediction class is 0, pcIs the predicted probability that a sample belongs to class c。
8. The bilinear attention convolution neural network pricing unit prediction analysis method of claim 1, characterized in that: the process of training the bilinear attention convolution neural network model by using the training set in the step 3 is as follows:
step 3-1, inputting the training set U into the network model for forward transmission, and in the reverse process of the training process, calculating and reversely updating and optimizing the weight and the bias value of the whole neural network model by adopting a loss function L and an Adam algorithm;
step 3-2: and (3) repeating the step (3-1) until a preset training end condition is reached, stopping training, determining the weight value and the bias value of each neuron in the network model at the moment, and obtaining the trained bilinear attention convolution neural network model.
9. The method for finite price single prediction analysis of bilinear attention convolution neural network as claimed in claim 1, wherein the activation functions used by all convolution layers in the bilinear attention convolution neural network model in step 3 are Re L U functions, and the full connection layer uses Softmax function.
10. A bilinear attention convolution neural network price limit single prediction analysis system comprises a data preprocessing module, a network training module and a stock price limit single trend prediction module, and is characterized in that:
in the data preprocessing module, preprocessing a data set to generate a training set and a test set, wherein the training set is input into the network training module, and the test set is input into the stock limit order trend prediction module;
in the network training module, a bilinear attention convolutional neural network model is constructed, the network model is optimized by using a pooling layer, a Dropout layer, a loss function L and an Adam algorithm, a training set is input into the network model to train the network model, and the trained bilinear attention convolutional neural network model is obtained after training is finished;
in the stock price limit single trend prediction module, a test set is input into a trained bilinear attention convolution neural network model to obtain corresponding output data, namely a prediction result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010312349.4A CN111445339A (en) | 2020-04-20 | 2020-04-20 | Price limiting single prediction analysis method and system for bilinear attention convolution neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010312349.4A CN111445339A (en) | 2020-04-20 | 2020-04-20 | Price limiting single prediction analysis method and system for bilinear attention convolution neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111445339A true CN111445339A (en) | 2020-07-24 |
Family
ID=71654241
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010312349.4A Pending CN111445339A (en) | 2020-04-20 | 2020-04-20 | Price limiting single prediction analysis method and system for bilinear attention convolution neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111445339A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113257280A (en) * | 2021-06-07 | 2021-08-13 | 苏州大学 | Speech emotion recognition method based on wav2vec |
| CN114489940A (en) * | 2022-01-18 | 2022-05-13 | 武汉理工大学 | Cold start optimization method based on time domain convolution network in server-free computing environment |
| WO2023040009A1 (en) * | 2021-09-18 | 2023-03-23 | 苏州大学 | Limit order book trend prediction apparatus and method, device, and storage medium |
| CN116205522A (en) * | 2023-01-31 | 2023-06-02 | 中国测绘科学研究院 | Landslide vulnerability evaluation method and system based on multidimensional CNN coupling |
-
2020
- 2020-04-20 CN CN202010312349.4A patent/CN111445339A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113257280A (en) * | 2021-06-07 | 2021-08-13 | 苏州大学 | Speech emotion recognition method based on wav2vec |
| WO2023040009A1 (en) * | 2021-09-18 | 2023-03-23 | 苏州大学 | Limit order book trend prediction apparatus and method, device, and storage medium |
| CN114489940A (en) * | 2022-01-18 | 2022-05-13 | 武汉理工大学 | Cold start optimization method based on time domain convolution network in server-free computing environment |
| CN116205522A (en) * | 2023-01-31 | 2023-06-02 | 中国测绘科学研究院 | Landslide vulnerability evaluation method and system based on multidimensional CNN coupling |
| CN116205522B (en) * | 2023-01-31 | 2023-10-20 | 中国测绘科学研究院 | Landslide vulnerability evaluation method and system based on multidimensional CNN coupling |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Huang et al. | Deep learning in finance and banking: A literature review and classification | |
| Pawar et al. | Stock market price prediction using LSTM RNN | |
| CN111445339A (en) | Price limiting single prediction analysis method and system for bilinear attention convolution neural network | |
| Passalis et al. | Temporal bag-of-features learning for predicting mid price movements using high frequency limit order book data | |
| CN111738532B (en) | Method and system for acquiring influence degree of event on object | |
| Soui et al. | Bankruptcy prediction using stacked auto-encoders | |
| Mandhula et al. | Predicting the customer’s opinion on amazon products using selective memory architecture-based convolutional neural network | |
| Ebiaredoh-Mienye et al. | Artificial neural network technique for improving prediction of credit card default: A stacked sparse autoencoder approach | |
| Ocampo | Decision modeling for manufacturing sustainability with fuzzy analytic hierarchy process | |
| US20210406771A1 (en) | Systems and methods for using artificial intelligence to evaluate lead development | |
| Meena et al. | Identifying emotions from facial expressions using a deep convolutional neural network-based approach | |
| CN114022176A (en) | Method for predicting commodity sales on e-commerce platform and electronic equipment | |
| Etemadi et al. | Earnings per share forecast using extracted rules from trained neural network by genetic algorithm | |
| CN113408582B (en) | Training method and device for feature evaluation model | |
| Jakubik et al. | Incorporating financial news for forecasting Bitcoin prices based on long short-term memory networks | |
| Liu et al. | 1D convolutional neural networks for chart pattern classification in financial time series | |
| Zehtab-Salmasi et al. | Multimodal price prediction | |
| Thesia et al. | A dynamic scenario‐driven technique for stock price prediction and trading | |
| John et al. | Stock market prediction based on deep hybrid RNN model and sentiment analysis | |
| Kuo et al. | Building Graduate Salary Grading Prediction Model Based on Deep Learning. | |
| CN111667307A (en) | Method and device for predicting financial product sales volume | |
| Le et al. | Applying LSTM to predict firm performance based on annual reports: an empirical study from the vietnam stock market | |
| Olajide Olajide et al. | Bank Term Deposit Service Patronage Forecasting using Machine Learning | |
| Mootha et al. | A Stacking Ensemble of Multi Layer Perceptrons to Predict Online Shoppers' Purchasing Intention | |
| Khan | Comparison of machine learning approaches for classification of invoices |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200724 |