JP2018200513A - Stock price prediction system, stock price prediction method and stock price prediction program - Google Patents

Abstract

To enhance profitability in stock investment.SOLUTION: A system acquires aggregation of a predicted value of a stock price, based on input data a which indicates a change element of the stock price (a rate of change or logarithmic differential of the stock price), by a prediction aggregation acquisition unit 110 (prediction aggregation acquisition means) of a stock price prediction system 100; and acquires collective knowledge of the stock price with a high predictability from the aggregation of the predicted value of the stock price, by a two stages selection unit 140 (collective knowledge acquisition means). Thereby, the system can acquire the collective knowledge of the stock price with the high predictability from the aggregation of the predicted value of the stock price acquired by the prediction aggregation acquisition unit 110, by the two stages selection unit 140.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a stock price prediction system, a stock price prediction method, and a stock price prediction program suitable for stock price prediction.

  With the recent development of ICT (Information and Communication Technology) technology, the movement to apply artificial intelligence also in the financial business is prosperous. In particular, there is something remarkable about technological innovation in the field of financial engineering that combines finance called FinTech and IT (Information Technology).

  Patent Document 1 calculates two types of feature values based on the price scenario of a financial derivative product, and supports the frequency of the degree of variation from each of the two types of feature amounts. Proposes a sales contract support system that calculates distributions, aggregates frequency distributions, and determines the purchase ratio of financial derivatives.

  Further, in Patent Document 2, an asset group is set by generating a uniform random number of an asset and a plurality of other assets for the asset to be predicted, and an observation equation and a state equation are set for the set asset group. We propose a financial asset price prediction system that applies a state space model, estimates the equation structure from past asset prices, and predicts asset prices using the estimated equations. The observation equation is estimated by factor analysis, and the state equation is a state variable having a time series structure estimated from the observation equation and the Kalman filter generated from the state equation.

JP 2003-296577 A Japanese Patent Laid-Open No. 05-108652

  In the above-described sales contract support system of Patent Document 1, the frequency distribution of the degree of variation calculated from two types of feature quantities based on the price scenario is aggregated and the purchase ratio of the financial derivative product is determined. You can make a sales plan.

  On the other hand, in the financial asset price prediction system of Patent Document 2 described above, an equation in which a uniform random number is generated for an asset to be predicted and a plurality of other assets is set, and an equation estimated from past asset prices is established. Since the price of an asset is predicted using an equation based on the structure, the prediction accuracy can be improved.

  By the way, when trying to apply the sales contract support system of Patent Document 1 or the financial asset price prediction system of Patent Document 2 to profitability prediction in stock investment, the following problems may occur.

  That is, although the sales contract support system of Patent Document 1 determines the purchase ratio from the result of counting each frequency distribution, there is a possibility that the accuracy of revenue prediction in the sales plan may decrease because there are many variations in each frequency distribution. .

  In addition, the financial asset price prediction system of Patent Document 2 applies a state space model composed of observation equations and state equations, and estimates the equation structure from past asset prices. For example, depending on the estimation results of state variables, There is a risk that the accuracy of the earnings forecast will be reduced.

  For this reason, it is desired to develop a system that can improve profitability in stock investment by improving the accuracy of earnings prediction.

The present invention has been made in view of such circumstances, and an object of the present invention is to provide a stock price prediction system, a stock price prediction method, and a stock price prediction program that can improve profitability in stock investment.

A stock price prediction system according to the present invention includes a prediction set acquisition means for acquiring a set of predicted values of stock prices based on input data indicating change factors of stock prices, and collective knowledge of stock prices with high prediction accuracy from the set of predicted values of stock prices. And collective intelligence acquisition means for acquiring.
Further, the prediction set acquisition unit creates a plurality of data sets from the training data indicated by the input data, replicates a plurality of prediction models, and outputs first output data indicating a set of training data of each prediction model. A random restoration extraction unit that outputs, and a group learning unit that trains a plurality of prediction models using the training data indicated by the first output data and outputs second output data that individually indicates the prediction value. The collective intelligence acquisition means cancels each other's prediction error by taking an average value from the second output data, and outputs the third output data indicating the collective intelligence of the prediction values with high prediction accuracy. It has a value acquisition part.
The collective intelligence acquisition means acquires a set of predicted values with a high consensus ratio from the collective intelligence of predicted values with high prediction accuracy acquired by the average value acquisition unit, and indicates a degree of consensus of the acquired collective intelligence A standard deviation acquisition unit for outputting the output data.
Further, the prediction set acquisition unit creates a plurality of data sets from the training data indicated by the input data, replicates a plurality of prediction models, and outputs first output data indicating a set of training data of each prediction model. A random restoration extraction unit that outputs, and a group learning unit that trains a plurality of prediction models using the training data indicated by the first output data and outputs second output data that individually indicates the prediction value. The collective intelligence acquisition means acquires a set of predicted values having a high consensus ratio from the second output data, and outputs a fourth output data indicating the degree of consensus of the acquired collective intelligence; And an average value acquisition unit that outputs a third output data indicating a collective intelligence of prediction values with high prediction accuracy by offsetting mutual prediction errors by taking an average value from the fourth output data. Features .
In the stock price prediction method of the present invention, a prediction set acquisition unit acquires a set of stock price prediction values based on input data indicating stock price change elements, and a collective intelligence acquisition unit sets the stock price prediction values. And obtaining a collective intelligence of stock prices with high prediction accuracy.
Further, the prediction set acquisition means creates a plurality of data sets from the training data indicated by the input data by a random restoration extraction unit, duplicates a plurality of prediction models, and sets a set of training data of each prediction model. A step of outputting the first output data to be displayed, and a group learning unit to train a plurality of prediction models using the training data indicated by the first output data, and to output second output data individually indicating the prediction values And the collective intelligence acquisition means uses the average value acquisition unit to take the average value from the second output data to cancel each other's prediction error and to generate a set of prediction values with high prediction accuracy. It has the process of outputting the 3rd output data which shows knowledge.
Further, the collective intelligence acquisition means acquires a set of prediction values with a high consensus ratio from the collective intelligence of prediction values with high prediction accuracy acquired by the average value acquisition unit by the standard deviation acquisition unit, and acquires the collective intelligence of the acquired collective intelligence It has the process of outputting the 4th output data which shows an agreement degree.
Further, the prediction set acquisition means creates a plurality of data sets from the training data indicated by the input data by a random restoration extraction unit, duplicates a plurality of prediction models, and sets a set of training data of each prediction model. A step of outputting the first output data to be displayed, and a group learning unit to train a plurality of prediction models using the training data indicated by the first output data, and to output second output data individually indicating the prediction values The collective intelligence acquisition means acquires a set of predicted values having a high consensus ratio from the second output data by the standard deviation acquisition unit, and indicates the degree of consensus of the acquired collective intelligence. The output data of 4 and the average value acquisition unit cancel the mutual prediction errors by taking the average value from the fourth output data, and show the collective intelligence of the prediction values with high prediction accuracy. of Characterized by a step of outputting the force data.
The stock price prediction program of the present invention is a stock price prediction program to be executed by a computer that controls a stock price prediction system, and obtains a set of predicted stock prices based on input data indicating change factors of the stock price by a prediction set acquisition means. And a step of acquiring collective intelligence of stock prices with high prediction accuracy from the set of predicted values of stock prices by collective intelligence acquisition means.
Further, the prediction set acquisition means creates a plurality of data sets from the training data indicated by the input data by a random restoration extraction unit, duplicates a plurality of prediction models, and sets a set of training data of each prediction model. A step of outputting the first output data to be displayed, and a group learning unit to train a plurality of prediction models using the training data indicated by the first output data, and to output second output data individually indicating the prediction values And the collective intelligence acquisition means uses the average value acquisition unit to take the average value from the second output data to cancel each other's prediction error and to generate a set of prediction values with high prediction accuracy. It has the process of outputting the 3rd output data which shows knowledge.
Further, the collective intelligence acquisition means acquires a set of prediction values with a high consensus ratio from the collective intelligence of prediction values with high prediction accuracy acquired by the average value acquisition unit by the standard deviation acquisition unit, and acquires the collective intelligence of the acquired collective intelligence It has the process of outputting the 4th output data which shows an agreement degree.
Further, the prediction set acquisition means creates a plurality of data sets from the training data indicated by the input data by a random restoration extraction unit, duplicates a plurality of prediction models, and sets a set of training data of each prediction model. A step of outputting the first output data to be displayed, and a group learning unit to train a plurality of prediction models using the training data indicated by the first output data, and to output second output data individually indicating the prediction values The collective intelligence acquisition means acquires a set of predicted values having a high consensus ratio from the second output data by the standard deviation acquisition unit, and indicates the degree of consensus of the acquired collective intelligence. The output data of 4 and the average value acquisition unit cancel the mutual prediction errors by taking the average value from the fourth output data, and show the collective intelligence of the prediction values with high prediction accuracy. of Characterized by a step of outputting the force data.
In the stock price prediction system, stock price prediction method, and stock price prediction program of the present invention, the prediction set acquisition unit acquires a set of predicted values of the stock price based on the input data indicating the change factor of the stock price, and the collective intelligence acquisition unit acquires the stock price. Collective intelligence of stock prices with high prediction accuracy is acquired from a set of predicted values.
That is, the collective intelligence acquisition means can acquire collective intelligence of stock prices with high prediction accuracy from the set of prediction values of stock prices acquired by the prediction set acquisition means, and can improve the accuracy of revenue prediction.

  According to the stock price prediction system, the stock price prediction method, and the stock price prediction program of the present invention, the collective intelligence acquisition unit can acquire the collective intelligence of the stock price with high prediction accuracy from the set of prediction values of the stock price acquired by the prediction set acquisition unit, Since the accuracy of the earnings prediction can be improved, the profitability in the stock investment can be improved.

It is a figure which shows an example of a structure of the stock price prediction system of this invention. It is a figure for demonstrating the group learning in the prediction set acquisition part of FIG. It is a figure for demonstrating the back test (Back-test) and the forward test (Forward-test) in the prediction set acquisition part of FIG. It is a figure which shows an example of the input data input into the prediction set acquisition part of FIG. FIG. 5 is a diagram for describing an example of data that is a source of input data in FIG. 4. It is for demonstrating acquisition of collective intelligence based on the output data from the prediction set acquisition part in the 2 step | paragraph selection part of FIG. 1, The same figure (a) demonstrates 1st selection (First selection). FIG. 4B is a diagram for explaining the second selection (Ssecond selection). It is a figure which shows an example of the prediction precision based on the output from the average value acquisition part of the 2 step | paragraph selection part of FIG. FIG. 7 is an output of the two-stage selection unit in FIG. 1, and is a diagram illustrating an example of prediction accuracy for a stock price stock selected in the second selection (Ssecond selection) in FIG. It is a flowchart for demonstrating the process of stock price prediction in the stock price prediction system of FIG. It is a figure which shows an example of the simulation in the 1st period (a)-4th period (d) in the Tokyo stock exchange by the stock price prediction system of FIG. It is a figure which shows an example of the simulation in the 1st period (a)-4th period (d) in the New York stock exchange by the stock price prediction system of FIG. It is a figure for demonstrating the prediction model by a general neural network. It is a figure which shows an example of the prediction precision by the prediction model of FIG.

Hereinafter, an embodiment of the stock price prediction system of the present invention will be described.
Prior to describing an embodiment of the stock price prediction system of the present invention, prediction accuracy by a general neural network will be described with reference to FIGS. The predicted value described below is the rate of return of the stock price. In addition, the prediction accuracy described below means the accuracy of prediction of the return rate of the stock price.

すなわち、未来＝ Ｆ（過去）として記述できる。 ここで、「Ｆ」は過去および将来の動きの関係を意味する。関数Ｆが「過去」のデータを入力として受け取る場合、関数Ｆからの出力として「未来」を予測することができる。 First, the neural network can be described as the following equation 1 regardless of the prediction model.

That is, it can be described as future = F (past). Here, “F” means a relationship between past and future movements. When the function F receives “past” data as an input, “future” can be predicted as an output from the function F.

ここで、｛Ｗl，Ｗ2，…，ＷN｝と数３の変数とは、予測モデルのパラメータである。また、ｄは図示しない第１層のニューロンの数、Ｎは第２層のニューロン１０の数であり、０jは第２の層のｊ番目のニューロン１０からの出力値である。

また、数４で示す式は、［０，１］の値である。 In addition, as shown in FIG. 12, for example, a general neural network prediction model outputs a plurality of neurons 10 having a non-linear filter 11 and a predicted value (a rate of return on stock prices) based on outputs from the plurality of neurons 10. This is shown by a prediction model in which neurons 12 are mixed.
A prediction model in which a plurality of neurons 10 are mixed can be expressed by the following formula 2.

Here, {Wl, W2,..., WN} and the variable of Equation 3 are parameters of the prediction model. Further, d is the number of neurons in the first layer (not shown), N is the number of neurons in the second layer, and 0j is an output value from the j-th neuron 10 in the second layer.

Also, the equation shown in Equation 4 is a value of [0, 1].

ここで、ｊ∈｛１、．．．、Ｎ｝であり、ηは訓練係数である。
また、数７での

は出力データであり、数８の

は、教師データ（Teacher data）である。 In general neural networks, it is necessary to optimize all the parameters w of the prediction model in order to accurately reproduce previous output data from input data.
Here, to explain the outline, the equation shown in Equation 5 is corrected to Equation 6 by the steepest descent method.

Here, j∈ {1,. . . , N}, and η is a training coefficient.
Also, in Equation 7

Is the output data.

Is teacher data.

ここで、ｉ∈｛0、．．．、ｄ｝および j ∈｛1、．．．、N｝である。
そして、数１０の平均二乗誤差Ｅが十分に小さくなるように、上記の数９の修正を繰り返す。

このようなトレーニングプロセスをバックテスト（Back-test）と呼び、数１０におけるαはバックテスト（Back-test）の開始時間であり、βはバックテストのトータルの長さである。 Next, if the number of neurons 10 is as shown in FIG. 12, the prediction model parameter w is corrected as shown in Equation 9 by the back-propagation algorithm.

Here, i∈ {0,. . . , D} and j ∈ {1,. . . , N}.
Then, the correction of Equation 9 is repeated so that the mean square error E of Equation 10 becomes sufficiently small.

Such a training process is called a back test, where α in Equation 10 is the back test start time, and β is the total length of the back test.

ここで、αはバックテスト（Back-test）の開始時間であり、βはテストのトータルの長さである。 Further, the total of training in the back test (Back-test) is represented by the mean square error E of Equation 11.

Here, α is the start time of the back test, and β is the total length of the test.

ここで、数１３における｛ｘ（ｔ）、ｘ（ｔ−１）、．．．、ｘ（ｔ−ｄ）｝は未学習の新しい入力データである。ここでは、この予測をフォワードテスト（Forward-test）という。 And after training all the parameters, new data can be predicted using a neural network. Assuming that the trained neural network is F, the predicted value (the rate of return of the stock price) shown in Expression 12 is expressed by Expression 13.

Here, {x (t), x (t−1),. . . , X (t−d)} is unlearned new input data. Here, this prediction is called a forward test.

  By using the above general neural network, for example, a prediction result in a forward test shown in FIG. 13 can be obtained. That is, FIG. 13 shows a prediction result of the rate of return of the stock price of a specific issue on the Tokyo stock exchange and the New York stock exchange, for example.

  For example, in the case of the Tokyo Stock Exchange, the first period is 54.8%, the second period is 55.8%, the third period is 51.6%, and the fourth period is 54.8%. This shows that the predicted value (income rate of stock price) was obtained. In the case of the New York Stock Exchange, the first period was 59.1%, the second period was 52.7%, the third period was 52.0%, and the fourth period was 53.2%. It shows that the prediction accuracy (by two choices of ascending or descending) was obtained.

Next, an embodiment of the stock price prediction system of the present invention will be described with reference to FIGS. Note that training data (training data), which is input data a described below, indicates changes in the stock price. As a change factor of the stock price, a stock price change rate represented by the following equation can be used.
Price change rate x (t) = [Stock price (t)-Stock price (t-1)] / Stock price (t-1)

In addition, the change factor of the stock price is not limited to the rate of change of the stock price, and a logarithmic difference represented by the following equation can be used.
Logarithmic difference of stock price x (t) = log Stock price (t)-log Stock price (t-1)

  First, an example of the configuration of the stock price prediction system of the present invention will be described with reference to FIG. The stock price prediction system 100 in FIG. 1 includes a predicted set acquisition unit 110 and a two-stage selection unit 140. The predicted set acquisition unit 110 includes a random restoration extraction unit 120 and a group learning unit 130.

  Although the details will be described later, the random restoration extraction unit 120 creates a plurality of data sets from the training data (training data) indicated by the input data a, replicates a plurality of prediction models, and trains the training data of each prediction model. Output data b indicating the set of. The group learning unit 130 trains a plurality of prediction models using a set of training data indicated by the output data b from the random restoration extraction unit 120, and outputs output data c that individually indicates a set of predicted values.

The two-stage selection unit 140 includes an average value acquisition unit 150 and a standard deviation acquisition unit 160.
The average value acquisition unit 150 will be described in detail later, but the prediction value (stock price of the stock price) having a low prediction accuracy (prediction accuracy) is output from the output data c indicating the set of prediction values (stock price return rate) from the prediction set acquisition unit 110. Output data d indicating the collective intelligence obtained by obtaining the collective intelligence of the prediction value (stock price return rate) with high prediction accuracy (prediction accuracy) by offsetting each other's prediction error by taking the average value of the return rate) Is output. Although the details will be described later, the standard deviation acquisition unit 160 has a high consensus ratio based on collective knowledge of prediction values (stock price return rate) with high prediction accuracy (prediction accuracy) acquired by the average value acquisition unit 150. A set of predicted values (earning rate of stock price) (the most reliable) is acquired, and output data e indicating the degree of consensus of the acquired collective intelligence is output.

  Although details will be described later, the output data d can be used for investment judgment of stock trading, and the output data e can be used for confirmation of confidence level of prediction.

  Next, with reference to FIG. 2, the group learning in the prediction set acquisition part 110 of FIG. 1 is demonstrated. First, the random restoration extraction unit 120 includes a plurality of prediction models 120a to 120n. These prediction models 120a to 120n are duplicated based on the respective data sets a1 to an obtained by re-sampling the training data of the input data a. Details of training data and resampled data sets a1 to an will be described later. These prediction models 120a to 120n execute the above-described back-test and forward-test based on the respective data sets a1 to an, and indicate, for example, Up or Down. Output data b indicating a set of training data is output. Note that the number of data sets (data sets) a1 to an and prediction models 120a to 120n is arbitrary, but is 1000 in the present embodiment, for example.

  Here, the back test and the forward test are processes for preventing overfitting. That is, N and d are optimized so as to optimize the evaluation in the collective learning in the predicted set acquisition unit 110. Next, in order to evaluate the versatility of the prediction models 120a to 120n, a forward test is performed using new untrained teacher data (Teacher data). First, in the back-test, as shown in FIG. 3, for example, training of the model parameter w (Training of w) and evaluation of prediction accuracy (Evaluation of prediction accuracy) are performed at the first stage. In the next stage, training of the model parameter w (Training of w) and evaluation of prediction accuracy (Evaluation of prediction accuracy) are interchanged, and N and d are optimized. Subsequently, in a forward test, a prediction model optimized by a back test is evaluated using new untrained teacher data (Teacher data). In addition, the parameters of {w1, w2,..., WN} in the back test (Back-test) are corrected as shown in Equation 6 above.

In FIG. 2, the collective learning unit 130 individually predicts new data for the forward test by training a plurality of prediction models using the output data d from the random restoration extraction unit 120. To do. In other words, the group learning unit 130 can obtain a predicted value (a rate of return of a stock price) by, for example, the following Equation 14.

In Equation 14, for example, there are 19 independent prediction models, and each prediction accuracy is assumed to be p. At this time, the probability that the x prediction models are correct can be calculated by Equation 14. Next, the probability that the majority decision is correct can be calculated by taking the sum of P (x) where x is 10 or more as shown in Equation 15. If p is 0.53, the probability that the majority decision is correct improves to 0.60.

  That is, the group learning unit 130 can increase the prediction accuracy from 53% to 60%, for example, due to the effect of group learning.

  Next, with reference to FIG. 4, the training data (training data) indicated by the input data a given to the predicted set acquisition unit 110 in FIG. 1 will be described. First, Fig.4 (a) has shown the training data (Training data) for producing | generating the data sets (data set) a1-an. In addition, this training data (Training data) has shown the stock price of a specific brand.

  Training data (Training data) is, for example, as indicated by # 1, data indicated by x (t-1), x (t-2),... X (td-1) and x (t) And teacher data shown.

  FIG. 4B shows one data set a1 in which training data (training data) # 1 to # 10 in FIG. 4A are arranged at random. Similarly, training data (Training data) # 1 to # 10 in FIG. 4A is also randomly arranged for the other data sets a2 to an. That is, by arranging training data (training data) from # 1 to # 10 in FIG. 4A at random, a plurality of data sets a1 to an are duplicated. In FIG. 4, for convenience of explanation, one set of data is shown as 10 pieces from # 1 to # 10. However, the number of data in one set is not limited to 10 and may be 11 or more.

  Next, with reference to FIG. 5, data serving as the basis of training data will be described. That is, training data corresponds to a stock price of an arbitrary brand. The markets from which stock prices are acquired are, for example, the Tokyo Stock Exchange and the New York Stock Exchange. The stock price acquisition period is divided into a first period (First term), a second period (Second term), a third period (Third term), and a fourth period (Fourth term). Also, every 1st to 4th periods, the data (stock price) used in the back test is 5 years, and the data (stock price) used in the forward test is 2.5 years. It is said.

  In addition, for example, 590 stocks listed on the Tokyo stock exchange. For example, 500 listed stocks are sampled on the New York stock exchange. Since a sample value (stock price) is necessary to determine the sample of a listed stock, for example, a stock with no delisting period is selected. The number of samples is not limited to 590 issues on the Tokyo stock exchange and 500 issues on the New York stock exchange. Optionally, the number of samples may be determined.

  In addition, the sample period of the back test data (stock price) in the first term is 1991/1 to 1995/12. The sample period of the forward test data (stock price) in the first term is 1996/1 to 1998/6. The sample period of the back test data (stock price) in the second term is 1996/1 to 2000/12. In addition, the sample period of the input data (stock price) of the forward test (Forward-test) in the second term is set to 2001/1 to 2003/6. The sample period of back test data (stock price) in the third term is 200/1 to 2005/12. The sample period of forward test data (stock price) in the third term is 2006/1 to 2008/6. The sample period of back test data (stock price) in the fourth term is 2006/1 to 2010/12. In addition, the sample period of the forward test data (stock price) in the fourth term is 2011/1 to 2013/6.

  The sample period of the input data (stock price) used in the back test in the first period (First term) to the fourth period (Fourth term) is 5 years, but it is limited to 5 years. is not. However, if the period is lengthened or shortened, there is a possibility that the prediction accuracy may be lowered, and therefore about 5 years is preferable. Further, the sample period of the input data (stock price) used in the forward test is 2.5 years, but is not limited to 2.5 years. However, if the period is lengthened or shortened, the prediction accuracy may be lowered, and therefore about 2.5 years is preferable.

  Next, the two-stage selection by the average value acquisition unit 150 and the standard deviation acquisition unit 160 of the two-stage selection unit 140 will be described with reference to FIG. First, FIG. 6A is a diagram illustrating the first selection by the average value acquisition unit 150. In FIG. 6A, the vertical axis represents prediction accuracy (Prediction accuracy), and the horizontal axis represents stock number (Stock No.). The average value acquisition unit 150 removes stock issues with low prediction accuracy in order to acquire highly profitable stock issues from the output data c that is a set of prediction values (stock price return rate) of the group learning unit 130 (Removed ) Acquire stocks with high prediction accuracy.

  That is, if a stock issue that is difficult to predict is applied in the back-test in the random restoration extraction unit 120, the set of prediction values (stock price return rate) from the group learning unit 130 is incidental. May contain high consensus. This adversely affects the standard deviation acquisition unit 160, which will be described later, and therefore, stock issues that are difficult to predict are previously removed by the average value acquisition unit 150. For example, 75% of stock issues with low prediction accuracy are removed from the output data c from the group learning unit 130, and 25% stock issues with high prediction accuracy (high profitability) are acquired. Collective intelligence for each acquired stock name is output as output data d.

  Next, FIG. 6B is a diagram illustrating the second selection by the standard deviation acquisition unit 160. In FIG. 6B, the vertical axis represents the consensus ratios (Consensus Ratios), and the horizontal axis represents the daily chart (Daily chart). Here, the standard deviation acquisition unit 160 adaptively detects the most reliable stock name indicating the daily maximum consensus ratios. That is, a consensus ratio C is used in a forward test by the standard deviation acquisition unit 160. The consensus ratio C is not always a high value even if it is a predictable stock brand. For this reason, the standard deviation acquisition unit 160 adaptively detects the most reliable predicted value (the rate of return of the stock price) indicating the maximum consensus ratio every day. That is, the output data e output from the standard deviation acquisition unit 160 indicates the degree of agreement with respect to collective intelligence acquired by the average value acquisition unit 150.

Here, the standard deviation acquisition unit 160 obtains a consensus ratio (Consensus Ratios) C using Equation 16 below in order to improve the prediction accuracy.

  Here, the larger the value of the consensus ratio C, the higher the collective intelligence reliability acquired by the average value acquisition unit 150.

  Next, an example of the output data d output from the average value acquisition unit 150 will be described with reference to FIG. The output data d is collective knowledge acquired by the average value acquisition unit 150, and is also a predicted value (a rate of return on stock prices) of a specific brand. As shown in FIG. 7, in the case of the Tokyo Stock Exchange, the first term (First term) is 59.9%, the second term (Second term) is 57.9%, and the third term (Third The prediction accuracy was 55.5% for term) and 57.0% for the fourth term (Fourth term). In the case of the New York stock exchange, the first term (First term) is 61.7%, the second term (Second term) is 55.6%, and the third term (Third term) is 55. The prediction accuracy was 3% and the fourth term was 54.6%. Note that the predicted value (the rate of return of the stock price) obtained by the average value acquisition unit 150 can be used for investment judgment of stock trading.

  Next, an example of the output data e output from the standard deviation acquisition unit 160 will be described with reference to FIG. The output data e is obtained by the standard deviation acquisition unit 160 as a consensus ratio for the collective intelligence of the prediction value (stock price return rate) with high prediction accuracy (prediction accuracy) acquired by the average value acquisition unit 150. FIG. 7 is a diagram showing an example of prediction accuracy for a stock price stock selected in the second selection (Ssecond selection) of FIG.

  As shown in FIG. 8, in the case of the Tokyo Stock Exchange, the first term (First term) is 65.9%, the second term (Second term) is 71.6%, the third term (Third The prediction accuracy was 60.4% for term) and 62.1% for the fourth term (Fourth term). In the case of the New York Stock Exchange, the first term (First term) is 84.0%, the second term (Second term) is 69.0%, and the third term (Third term) is 60. The prediction accuracy was 5% and the fourth term (Fourth term) was 57.0%.

  Here, when the prediction accuracy of FIG. 7 is compared with the prediction accuracy of FIG. 8, the collective intelligence having the highest consensus ratio (the most reliable) is obtained from the collective intelligence output from the average value acquisition unit 150. It can be seen that a more accurate predicted value (stock price return) is obtained by acquiring.

  Further, when the prediction accuracy of FIG. 7 and the prediction accuracy of FIG. 8 are compared with the prediction accuracy by the general neural network of FIG. 13, the following can be understood.

  That is, as described above, in the prediction accuracy of FIG. 13, in the case of the Tokyo Stock Exchange, the first term (First term) is 54.8% and the second term (Second term) is 55.8. The prediction accuracy is 51.6% for the third term (Third term) and 54.8% for the fourth term (Fourth term). In the case of the New York stock exchange, the first term (First term) is 59.1%, the second term (Second term) is 52.7%, and the third term (Third term) is 52. The prediction accuracy is 0.0% and the fourth term is 53.2.

  Comparing the prediction accuracy of FIG. 7 with the prediction accuracy of FIG. 13, the prediction accuracy based on collective intelligence that is output data d from the average value acquisition unit 150 is from the first term to the fourth term (Fourth) in FIG. 13. It can be seen that it is higher than any prediction accuracy of term). Further, when comparing the prediction accuracy of FIG. 8 with the prediction accuracy of FIG. 13, the prediction accuracy using the collective intelligence agreement degree that is the output data e from the standard deviation acquisition unit 160 is the first period of FIG. 13. It can be seen that the prediction accuracy is higher than any prediction accuracy from the first term to the fourth term.

  In particular, the prediction accuracy of the first term of the New York stock exchange in FIG. 13 is 59.1%, whereas the New York stock exchange in FIG. Since the prediction accuracy of the first term is 84.0%, it can be seen that the prediction accuracy by the stock price prediction system 100 in this embodiment is further improved.

  Next, with reference to FIG. 9, a stock price prediction process by the stock price prediction system 100 will be described.

(Step S101)
Get a set of training data.
In this case, the random restoration extraction unit 120 of the predicted set acquisition unit 110 duplicates a plurality of data sets from the input data a indicating training data (Training data), and acquires a set of the training data (Training data).

(Step S102)
Output a set of training data.
In this case, output data b indicating a set of training data acquired by the random restoration extraction unit 120 is output.

(Step S103)
A plurality of prediction models are independently trained from a set of training data, and a prediction value (a rate of return of stock price) is acquired from each prediction model.
In this case, the collective learning unit 130 of the prediction set acquisition unit 110 acquires a set of prediction values (stock price return rate) using a plurality of prediction models for the output data b indicating a set of training data.

(Step S104)
Outputs a set of prediction values (stock price return).
In this case, output data c indicating a set of predicted values (stock price profit rate) acquired by the group learning unit 130 is output.

(Step S105)
Collective intelligence with high prediction accuracy is acquired from a set of prediction values (stock price return).
Furthermore, collective intelligence of all stock prices is acquired as a back test, and the prediction accuracy is acquired by comparing with the teacher data. This prediction accuracy is obtained as a result such as the prediction accuracy of FIG. 7 described above.

(Step S106)
Outputs collective intelligence of predicted values (stock price returns) from stock stocks with low prediction accuracy through backtesting.
In this case, the average value acquisition unit 150 removes stock issues with low prediction accuracy, and outputs collective intelligence of prediction values (stock price return rate) of stock issues with high prediction accuracy as output data d.

(Step S107)
In the forward test, a stock issue having a high consensus ratio (Consensus Ratios) is acquired from the collective intelligence of the predicted value (stock price return rate), and the collective intelligence of the predicted value (stock price return rate) is acquired.
In this case, the standard deviation acquisition unit 160 of the two-stage selection unit 140 has a high consensus ratio (the most reliable) from the collective knowledge of stock issues with a high prediction accuracy acquired by the average value acquisition unit 150. Get the collective intelligence prediction value (stock price return).

(Step S108)
Outputs the collective intelligence agreement and predicted value.
In this case, the collective intelligence prediction value (stock price return rate) with the highest (highest reliability) consensus ratio (consensus ratio) acquired by the standard deviation acquisition unit 160 and output data e indicating the degree of consensus of the collective intelligence Output.
By using the output data e, for example, a result such as the prediction accuracy of FIG. 8 described above can be obtained. Further, as described above, the output data e can be used to confirm the confidence level of stock prediction.

  Next, a simulation by the stock price prediction system 100 will be described with reference to FIGS.

  First, FIG. 10 shows a simulation result when the sample on the Tokyo stock exchange shown in FIG. 5 is, for example, 590 stocks. FIG. 11 shows a simulation result when the sample on the New York stock exchange shown in FIG. 5 is 500 stocks, for example. However, in the simulations shown in FIGS. 10 and 11, for convenience of explanation, the result of omitting the fee related to stock trading is shown.

  In FIGS. 10 and 11, the symbol h indicates the investment performance (asset amplification factor) of a specific brand predicted by the stock price prediction system 100. The symbol i indicates the investment performance when an investment brand is randomly selected every day (random strategy). Further, the symbol j indicates the average performance of 590 stocks on the Tokyo stock exchange and the average performance of 500 stocks on the New York stock exchange. .

  As can be seen from FIGS. 10 and 11, the investment performance h obtained by the stock price prediction system 100 exceeds the average performance j of the market and the performance i by the random strategy. This means that the prediction value (stock price return rate) by the stock price prediction system 100 is useful for confirming the predictability and profitability in the actual stock market.

  As described above, in the present embodiment, the prediction set acquisition unit 110 (prediction set acquisition unit) of the stock price prediction system 100 determines the stock price based on the input data a indicating the stock price change factor (stock price change rate or logarithmic difference). A set of predicted values is acquired, and collective intelligence of stock prices with high prediction accuracy is acquired from a set of predicted values of stock prices by a two-stage selection unit 140 (collective intelligence acquisition means).

  As a result, the two-stage selection unit 140 (collective intelligence acquisition unit) can acquire collective intelligence of stock prices with high prediction accuracy from the set of predicted values of stock prices acquired by the predicted set acquisition unit 110 (predictive set acquisition unit). Since the accuracy of prediction can be increased, the profitability in investment can be increased.

  That is, the output data d from the average value acquisition unit 150 is obtained from a set of prediction values (stock price return rate) included in the output data c from the group learning unit 130, with a prediction value (stock price return rate) having a low prediction accuracy. By taking the average value, the mutual prediction errors are offset, and the collective knowledge of the prediction value (stock price return rate) with high prediction accuracy is acquired.

  The output data e output from the standard deviation acquisition unit 160 is the degree of consensus for collective intelligence acquired by the average value acquisition unit 150, and is effective for careful selection of stock prices with high prediction accuracy.

  In the present embodiment, the average value acquisition unit 150 uses the output data c indicating the set of prediction values (stock price return rate) from the prediction set acquisition unit 110 to predict a prediction value (stock price profit) with a low prediction accuracy (Prediction accuracy). The average error rate is offset to cancel each other's prediction error, the collective intelligence of the prediction value (stock price return rate) with high prediction accuracy is acquired, and output data d indicating the acquired collective intelligence is obtained. The standard deviation acquisition unit 160 outputs a high consensus ratio (consensus ratio) from the collective knowledge of the prediction value (stock price return rate) with high prediction accuracy (prediction accuracy) acquired by the average value acquisition unit 150. Although a case has been described where a set of predicted values (earnings rate of stock price) is acquired and output data e indicating the degree of consensus of the acquired collective intelligence is output, the present invention is not limited to this example.

  That is, the standard deviation acquisition unit 160 obtains a predicted value (stock price) having a high (highest reliability) consensus ratio (consensus ratio) from the output data c indicating the set of predicted values (stock price return rate) from the predicted set acquisition unit 110. The average value acquisition unit 150 obtains the average value of the prediction values (stock price return rate) with low prediction accuracy (Prediction accuracy) from the collective intelligence acquired by the standard deviation acquisition unit 160. Mutual prediction errors may be offset, and collective intelligence of prediction values (stock price returns) with high prediction accuracy may be acquired.

DESCRIPTION OF SYMBOLS 10, 12 Neuron 11 Nonlinear filter 100 Stock price prediction system 110 Prediction set acquisition part 120a-120n Prediction model 120 Random restoration extraction part 130 Collective learning part 140 Two step selection part 150 Average value acquisition part 160 Standard deviation acquisition part a Input data a1 ~ An Data set b ~ e Output data

Claims (12)

A prediction set acquisition means for acquiring a set of prediction values of the stock price based on input data indicating a change element of the stock price;
Collective intelligence acquisition means for acquiring collective intelligence of stock prices with high prediction accuracy from the set of predicted values of stock prices. The predicted set acquisition means includes
A random restoration extraction unit that creates a plurality of data sets from the training data indicated by the input data, replicates a plurality of prediction models, and outputs first output data indicating a set of training data of each prediction model;
A group learning unit that trains a plurality of prediction models using the training data indicated by the first output data, and outputs second output data that individually indicates predicted values;
The collective intelligence acquisition means includes:
An average value acquisition unit that outputs a third output data indicating a collective intelligence of prediction values with high prediction accuracy by offsetting mutual prediction errors by taking an average value from the second output data, The stock price prediction system according to claim 1.   The collective intelligence acquisition unit acquires a set of prediction values with a high consensus ratio from the collective intelligence of prediction values with high prediction accuracy acquired by the average value acquisition unit, and a fourth output indicating the degree of consensus of the acquired collective intelligence The stock price prediction system according to claim 2, further comprising a standard deviation acquisition unit that outputs data. The predicted set acquisition means includes
A random restoration extraction unit that creates a plurality of data sets from the training data indicated by the input data, replicates a plurality of prediction models, and outputs first output data indicating a set of training data of each prediction model;
A group learning unit that trains a plurality of prediction models using the training data indicated by the first output data, and outputs second output data that individually indicates predicted values;
The collective intelligence acquisition means includes:
A standard deviation acquisition unit that acquires a set of predicted values having a high consensus ratio from the second output data, and outputs fourth output data indicating the degree of consensus of the acquired collective intelligence;
An average value acquisition unit that outputs the third output data indicating the collective intelligence of the prediction values with high prediction accuracy by offsetting each other's prediction error by taking an average value from the fourth output data. The stock price prediction system according to claim 1. A step of acquiring a set of prediction values of the stock price based on input data indicating a change factor of the stock price by the prediction set acquisition means;
And a step of acquiring collective intelligence of stock prices with high prediction accuracy from the set of predicted values of stock prices by collective intelligence acquisition means. The predicted set acquisition means includes
A random restoration extraction unit creates a plurality of data sets from the training data indicated by the input data, replicates a plurality of prediction models, and outputs first output data indicating a set of training data of each prediction model Process,
Training a plurality of prediction models using the training data indicated by the first output data by the group learning unit, and outputting second output data indicating the predicted values individually,
The collective intelligence acquisition means includes:
The average value acquisition unit has a step of canceling each other's prediction error by taking an average value from the second output data and outputting third output data indicating collective intelligence of prediction values with high prediction accuracy. The stock price prediction method according to claim 5, wherein: The collective intelligence acquisition means includes:
4th output data which acquires the set of prediction values with a high consensus ratio from the collective intelligence of the prediction value with the high prediction accuracy which the said average value acquisition part acquired by the standard deviation acquisition part, and shows the agreement degree of the acquired collective intelligence The stock price prediction method according to claim 6, further comprising: The predicted set acquisition means includes
A random restoration extraction unit creates a plurality of data sets from the training data indicated by the input data, replicates a plurality of prediction models, and outputs first output data indicating a set of training data of each prediction model Process,
Training a plurality of prediction models using the training data indicated by the first output data by the group learning unit, and outputting second output data indicating the predicted values individually,
The collective intelligence acquisition means includes:
Obtaining a set of predicted values having a high consensus ratio from the second output data by the standard deviation obtaining unit, and outputting fourth output data indicating the degree of collective intelligence agreement;
A step of offsetting each other's prediction error by taking an average value from the fourth output data and outputting third output data indicating collective intelligence of prediction values with high prediction accuracy by an average value acquisition unit. The stock price prediction method according to claim 5, wherein: A stock price prediction program that is executed by a computer that controls a stock price prediction system,
A step of acquiring a set of prediction values of the stock price based on input data indicating a change factor of the stock price by the prediction set acquisition means;
And a step of acquiring collective intelligence of stock prices with high prediction accuracy from the set of predicted values of stock prices by collective intelligence acquisition means. The predicted set acquisition means includes
A random restoration extraction unit creates a plurality of data sets from the training data indicated by the input data, replicates a plurality of prediction models, and outputs first output data indicating a set of training data of each prediction model Process,
Training a plurality of prediction models using the training data indicated by the first output data by the group learning unit, and outputting second output data indicating the predicted values individually,
The collective intelligence acquisition means includes:
The average value acquisition unit has a step of canceling each other's prediction error by taking an average value from the second output data and outputting third output data indicating collective intelligence of prediction values with high prediction accuracy. The stock price prediction program according to claim 9, wherein: The collective intelligence acquisition means includes:
4th output data which acquires the set of prediction values with a high consensus ratio from the collective intelligence of the prediction value with the high prediction accuracy which the said average value acquisition part acquired by the standard deviation acquisition part, and shows the agreement degree of the acquired collective intelligence The stock price prediction program according to claim 8, further comprising: The predicted set acquisition means includes
A random restoration extraction unit creates a plurality of data sets from the training data indicated by the input data, replicates a plurality of prediction models, and outputs first output data indicating a set of training data of each prediction model Process,
Training a plurality of prediction models using the training data indicated by the first output data by the group learning unit, and outputting second output data indicating the predicted values individually,
The collective intelligence acquisition means includes:
Obtaining a set of predicted values having a high consensus ratio from the second output data by the standard deviation obtaining unit, and outputting fourth output data indicating the degree of collective intelligence agreement;
A step of offsetting each other's prediction error by taking an average value from the fourth output data and outputting third output data indicating collective intelligence of prediction values with high prediction accuracy by an average value acquisition unit. The stock price prediction program according to claim 9, wherein:

Images