JP2013152656A - Information processor, information processing method and program for determining explanatory variable - Google Patents

Abstract

An object of the present invention is to reduce the amount of calculation performed when finding explanatory variables to be added to a prediction model.
The information processing apparatus calculates, for each of a plurality of objective variables, an error between an actual value of the objective variable and a value calculated by a first prediction model for predicting the value of the objective variable. A first calculation unit to be calculated, a classification unit for classifying a plurality of objective variables into a plurality of groups based on the calculated error, and an error calculated for the objective variable belonging to the group for each of the plurality of groups A second calculation unit for calculating a representative value of the error and a plurality of second prediction models for predicting the representative value for each of the plurality of groups while changing the explanatory variables. A determining unit that determines an explanatory variable to be added to the first prediction model of the objective variable belonging to the group based on a difference between each of the values calculated by the two prediction models and the representative value.
[Selection] Figure 1

Description

  The present technology relates to a prediction model construction technology.

  There is a technique for constructing a prediction model for predicting a value (for example, stock price) of an objective variable that changes with the passage of time by using a value (for example, past stock price) of an explanatory variable.

  It is known that in order to build a prediction model with high accuracy, it is effective to add explanatory variables to the prediction model and predict the value of the objective variable using more explanatory variable values. . For example, when a prediction model is constructed for the stock price of A transportation, the accuracy of the prediction model may be improved by using not only the past stock price of A transportation but also the past price of gasoline.

  However, if a large number of explanatory variables are added to the prediction model to improve accuracy, the prediction model will be specialized for time-series data (hereinafter referred to as learning data) used to construct the prediction model. . Therefore, if the accuracy of the prediction model is verified using other time series data, the accuracy may be lowered. Such a state is called “overfitting” in the field of machine learning.

  It is very difficult to select an explanatory variable that gives the highest accuracy of the prediction model. Simply, it is conceivable that a prediction model is constructed and verified for every combination of explanatory variables, and the prediction model with the highest accuracy is adopted. However, if the number of explanatory variables increases, the amount of calculation becomes enormous due to the combined explosion, which is not realistic.

  Therefore, conventionally, a technique of repeating the process of adding one explanatory variable having the highest usefulness to the prediction model and the process of removing an explanatory variable that has become unnecessary by newly adding an explanatory variable (stepwise variable selection method) ) Is used. Examples of indices representing the usefulness of explanatory variables include F value (F value) for significance as a multiple regression model, AIC (Akaike's Information Criterion), and BIC (Bayesian Information Criterion). In recent years, indexes such as AIC and BIC are often used. Both employ an explanatory variable that minimizes the sum of the squared errors.

  However, there is a problem with the above technique. In this technique, the total calculation amount is approximately (calculation amount required to calculate an index representing the usefulness of one explanatory variable) × (number of objective variables) × (number of explanatory variable candidates). Therefore, when the number of objective variables and the number of explanatory variables increase, there is a problem that the amount of calculation becomes very large.

Paul A. Murtaugh (2009). Performance of several variable-selection methods applied to real ecological data. Ecology Letters, 12: 1061-1068.

  Accordingly, an object of the present technology is, in one aspect, to provide a technique for reducing the amount of calculation performed when finding an explanatory variable to be added to a prediction model.

  An information processing apparatus according to an aspect of the present technology includes: (A) a storage device; and (B) a first value for predicting an actual value of an objective variable and a value of the objective variable for each of a plurality of objective variables. And (C) classifying a plurality of objective variables into a plurality of groups based on the error stored in the storage device. For each of the plurality of groups, a second calculation unit that calculates a representative value of the error using an error calculated for the objective variable belonging to the group and stores the representative value in a storage device; and (D) a plurality of groups For each, a plurality of second prediction models for predicting representative values stored in the storage device are generated while changing the explanatory variables, and each of the values calculated by the generated second prediction models Difference from typical value Based, and a first determining unit for determining the explanatory variable to be added to the first prediction model target variable belonging to the group.

  This will reduce the amount of computation to be performed when finding explanatory variables to be added to the prediction model.

FIG. 1 is a functional block diagram of the information processing apparatus according to the present embodiment. FIG. 2 is a functional block diagram of the grouping processing unit. FIG. 3 is a functional block diagram of the candidate extraction unit. FIG. 4 is a functional block diagram of the determination unit. FIG. 5 is a diagram illustrating an example of learning data stored in the learning data storage unit. FIG. 6 is a diagram showing a main processing flow. FIG. 7 is a diagram for explaining the outline of the grouping process. FIG. 8 is a diagram for explaining an outline of the grouping process. FIG. 9 is a diagram illustrating a processing flow of the grouping process. FIG. 10 is a diagram illustrating an example of data stored in the first model storage unit. FIG. 11 is a diagram illustrating an example of data stored in the first error data storage unit. FIG. 12 is a diagram illustrating an example of data stored in the grouping result storage unit. FIG. 13 is a diagram for explaining an overview of the first candidate extraction process. FIG. 14 is a diagram illustrating a processing flow of the first candidate extraction processing. FIG. 15 is a diagram illustrating an example of data stored in the second model storage unit. FIG. 16 is a diagram illustrating an example of data stored in the second error data storage unit. FIG. 17 is a diagram illustrating an example of data stored in the first evaluation value storage unit. FIG. 18 is a diagram illustrating an example of data stored in the first extraction result storage unit. FIG. 19 is a diagram for explaining the outline of the second candidate extraction process. FIG. 20 is a diagram illustrating a process flow of the second candidate extraction process. FIG. 21 is a diagram illustrating an example of data stored in the third model storage unit. FIG. 22 is a diagram illustrating an example of data stored in the third error data storage unit. FIG. 23 is a diagram illustrating an example of data stored in the second evaluation value storage unit. FIG. 24 is a diagram illustrating an example of data stored in the second extraction result storage unit. FIG. 25 is a diagram illustrating an example of data stored in the candidate storage unit. FIG. 26 is a diagram for explaining the outline of the determination process. FIG. 27 is a diagram illustrating a processing flow of determination processing. FIG. 28 is a diagram illustrating an example of data stored in the fourth model storage unit. FIG. 29 is a diagram illustrating an example of data stored in the fourth error data storage unit. FIG. 30 is a diagram illustrating an example of data stored in the third evaluation value storage unit. FIG. 31 is a diagram illustrating an example of data stored in the determination result storage unit. FIG. 32 is a diagram illustrating an example of screen data to be displayed. FIG. 33 is a diagram illustrating a processing flow of processing for determining an explanatory variable to be added to the prediction model for each brand. FIG. 34 is a diagram illustrating functional blocks of a computer.

  FIG. 1 shows a functional block diagram of the information processing apparatus 1 according to the present embodiment. The information processing apparatus 1 includes a grouping processing unit 2, a grouping result storage unit 3, a candidate extraction unit 4, a learning data storage unit 5, a candidate storage unit 6, a determination unit 7, a determination result storage unit 8, And an output unit 9.

  The grouping processing unit 2 performs grouping processing using the data stored in the learning data storage unit 5 and stores the processing result in the grouping result storage unit 3. The candidate extraction unit 4 performs first and second candidate extraction processing using the data stored in the grouping result storage unit 3 and the data stored in the learning data storage unit 5, and the processing result is stored in the candidate storage unit 6. To store. The determination unit 7 performs a determination process using data stored in the grouping result storage unit 3, the learning data storage unit 5, and the candidate storage unit 6, and stores the processing result in the determination result storage unit 8. The output unit 9 generates screen data to be displayed using the data stored in the grouping result storage unit 3, the data stored in the candidate storage unit 6, and the data stored in the determination result storage unit 8. And display it on a display device or the like.

  FIG. 2 shows a functional block diagram of the grouping processing unit 2. The grouping processing unit 2 includes a first model generation unit 21, a first model storage unit 22, a first error calculation unit 23, a first error data storage unit 24, and a group generation unit 25.

  The first model generation unit 21 performs processing using the data stored in the learning data storage unit 5 and stores the processing result in the first model storage unit 22. The first error calculation unit 23 performs processing using the data stored in the learning data storage unit 5 and the data stored in the first model storage unit 22, and the processing result is stored in the first error data storage unit 24. Store. The group generation unit 25 performs processing using the data stored in the first error data storage unit 24 and stores the processing result in the grouping result storage unit 3.

  FIG. 3 shows a functional block diagram of the candidate extraction unit 4. The candidate extraction unit 4 includes a second model generation unit 401, a second model storage unit 402, a third model storage unit 403, a second error calculation unit 404, a second error data storage unit 405, and a third error. Data storage unit 406, first evaluation value calculation unit 407, first evaluation value storage unit 408, second evaluation value storage unit 409, extraction unit 410, first extraction result storage unit 411, and second extraction A result storage unit 412 and a first specifying unit 413 are included.

  The second model generation unit 401 performs processing using the data stored in the grouping result storage unit 3 and the data stored in the learning data storage unit 5, and the processing result is stored in the second model storage unit 402 and the third model. Store in the model storage unit 403. The second error calculation unit 404 performs processing using the data stored in the second model storage unit 402 and stores the processing result in the second error data storage unit 405. In addition, the second error calculation unit 404 performs processing using data stored in the third model storage unit 403 and stores the processing result in the third error data storage unit 406. The first evaluation value calculation unit 407 performs processing using the data stored in the second error data storage unit 405 and stores the processing result in the first evaluation value storage unit 408. The first evaluation value calculation unit 407 performs processing using the data stored in the third error data storage unit 406 and stores the processing result in the second evaluation value storage unit 409. The extraction unit 410 performs processing using the data stored in the first evaluation value storage unit 408 and stores the processing result in the first extraction result storage unit 411. Further, the extraction unit 410 performs processing using the data stored in the second evaluation value storage unit 409 and stores the processing result in the second extraction result storage unit 412. The first specifying unit 413 performs processing using the data stored in the first extraction result storage unit 411 and the data stored in the second extraction result storage unit 412, and stores the processing result in the candidate storage unit 6. To do.

  FIG. 4 shows a functional block diagram of the determination unit 7. The determination unit 7 includes a third model generation unit 71, a fourth model storage unit 72, a third error calculation unit 73, a fourth error data storage unit 74, a second evaluation value calculation unit 75, and a third evaluation value. A value storage unit 76 and a second specifying unit 77 are included.

  The third model generation unit 71 performs processing using the data stored in the learning data storage unit 5 and the data stored in the candidate storage unit 6 and stores the processing result in the fourth model storage unit 72. The third error calculation unit 73 performs processing using the data stored in the learning data storage unit 5 and the data stored in the fourth model storage unit 72, and the processing result is stored in the fourth error data storage unit 74. Store. The second evaluation value calculation unit 75 performs processing using the data stored in the fourth error data storage unit 74 and stores the processing result in the third evaluation value storage unit 76. The second specifying unit 77 performs processing using the data stored in the third evaluation value storage unit 76 and stores the processing result in the determination result storage unit 8.

  FIG. 5 shows an example of learning data stored in the learning data storage unit 5. In the example of FIG. 5, for each date from July 27 to July 31, today's stock price, stock price one day ago, and stock price two days ago are stored. The example of FIG. 5 shows a stock price data set for a specific brand, and the learning data storage unit 5 stores a stock price data set for many brands.

  In the present embodiment, when a prediction model is constructed using the stock price of “today” for a specific stock using the stock price of “1 day ago” and the stock price of “2 days ago” of the stock, It is assumed that explanatory variables to be added are selected in order to improve accuracy.

  Next, the operation of the information processing apparatus 1 shown in FIG. 1 will be described with reference to FIGS. First, the grouping processing unit 2 performs a grouping process (FIG. 6: Step S1). The grouping process will be described with reference to FIGS.

  First, an outline of the grouping process will be described. In the grouping process, a plurality of brands to be processed are grouped. The standard for grouping is the error between the predicted value by the prediction model and the actual value. Specifically, as shown in FIG. 7, the error between the predicted value and the actual value is calculated for each date, and grouping is performed based on an error vector having the error for each date as a component.

  The grouping is performed so that brands having similar error vectors (that is, having similar error fluctuation trends) are in the same group. And as shown in FIG. 8, the process which calculates a representative error vector using the error vector of the brand which belongs to the same group is performed for every group.

  Next, the processing flow of the grouping process will be described. The first model generation unit 21 in the grouping processing unit 2 identifies one unprocessed brand among the brands whose learning data is stored in the learning data storage unit 5 (FIG. 9: Step S11).

  The first model generation unit 21 constructs a prediction model using the learning data of the brand specified in step S11, and stores the constructed prediction model data and brand name in the first model storage unit 22 (step). S13). In step S13, a prediction model (for example, an AR (AutoRegressive) model) for predicting the stock price “today” of the stock specified in step S11 using the stock price “1 day ago” and the stock price “2 days ago” is used. To construct. Since the technique for constructing the prediction model is not the main part of the present embodiment, detailed description thereof is omitted.

  FIG. 10 shows an example of data stored in the first model storage unit 22. In the example of FIG. 10, the brand name and the prediction model data are stored.

  The first error calculation unit 23 uses the prediction model constructed in step S13 and the data stored in the learning data storage unit 5 to calculate the value calculated by the prediction model (that is, the prediction value) and the actual value. An error vector based on the error from the value is calculated (step S15). Further, the first error calculation unit 23 stores the brand name, the calculation result, and the like in the first error data storage unit 24.

  FIG. 11 shows an example of data stored in the first error data storage unit 24. In the example of FIG. 11, the brand name and the actual value, predicted value, and error of the stock price for each date are stored.

  Returning to the description of FIG. 9, the first model generation unit 21 determines whether there is an unprocessed brand (step S17). When there is an unprocessed brand (step S17: Yes route), the process returns to step S11.

  On the other hand, when there is no unprocessed brand (step S17: No route), the group generation unit 25 groups the brands using the error vector data stored in the first error data storage unit 24, and groups them. Is temporarily stored in a storage device such as a main memory (step S19). In step S19, grouping is performed using, for example, the K-average method. Note that, for example, a grouping technique such as clustering is well known, and a description thereof will be omitted here.

  The group generation unit 25 calculates a representative error vector for each group, and stores the names of brands belonging to each group and the values of the components of the representative error vector in the grouping result storage unit 3 (step S21). In step S21, for example, a representative error vector is calculated by obtaining an average of error vectors of brands belonging to the group. Then, the process returns to the original process.

  FIG. 12 shows an example of data stored in the grouping result storage unit 3. In the example of FIG. 12, a group identifier, a name of a brand belonging to the group, and a representative error value for each date are stored. The group identifier is a unique number assigned to each group.

  As described above, grouping is performed so that stocks having similar error fluctuation trends are in the same group. This grouping is based on the idea that explanatory variables that are effective to be added to a prediction model for a certain brand can be used for prediction models for other brands belonging to the same group.

  Returning to the description of FIG. 6, the candidate extraction unit 4 performs a first candidate extraction process (step S3). The first candidate extraction process will be described with reference to FIGS.

  First, the outline of the first candidate extraction process will be described with reference to FIG. In order to simplify the explanation, the representative error vector is a one-dimensional vector. In the first candidate extraction process, N (N is a natural number of 2 or more) explanatory variable candidates are extracted for each group based on the accuracy of representative error prediction. That is, an evaluation value is calculated by calculating a sum of squares of differences between a prediction value based on a prediction model for predicting a representative error and a representative error, and N candidates are extracted in order from the candidate having the smallest calculated evaluation value. To do.

  However, in the prediction model for predicting the representative error, not only the explanatory variable candidates to be added, but also the explanatory variables included in the prediction model generated in step S13 for the brands in the group are used. In the example of FIG. 13, for a group including A transportation and B air, a prediction model using learning data of explanatory variables A transportation and B air and learning data of candidate explanatory variables (gasoline or rice) to be added. Is building. This is because when the prediction model includes a plurality of explanatory variables, it is considered that the accuracy of the prediction model may be greatly improved due to a synergistic effect depending on the combination of the explanatory variables. . This makes it difficult to miss candidates that can greatly improve the accuracy of the prediction model when used together with explanatory variables such as A transportation or B air. In addition, in the example of FIG. 13, since the evaluation value calculated about gasoline becomes smaller than the evaluation value calculated about rice, the description which gasoline adds to the group to which A transportation and B air belong belongs. It is preferable as a variable.

  Next, the process flow of the first candidate extraction process will be described. First, the second model generation unit 401 in the candidate extraction unit 4 identifies one unprocessed group among the groups registered in the grouping result storage unit 3 (FIG. 14: step S31).

  The 2nd model production | generation part 401 extracts the explanatory variable used with the prediction model constructed | assembled in step S13 about the brand which belongs to the group specified in step S31 (step S33). For example, when processing is performed for group 1 in FIG. 12, explanatory variables A transportation and B air are extracted.

  The second model generation unit 401 identifies one unprocessed candidate among the explanatory variable candidates to be added (step S35). In addition, the second model generation unit 401 uses the extracted learning data of the explanatory variables and the identified candidate learning data to predict the value of the representative error calculated for the group identified in step S31. Is constructed (step S37). Then, the second model generation unit 401 stores the group identifier, the explanatory variable candidate to be added, and the data of the constructed prediction model in the second model storage unit 402. In step S37, processing is performed using the value of the representative error stored in the grouping result storage unit 2.

  FIG. 15 shows an example of data stored in the second model storage unit 402. In the example of FIG. 15, group identifiers, candidates for explanatory variables to be added, and prediction model data are stored.

  Then, the second error calculation unit 404 calculates an error vector based on the error between the value calculated by the prediction model constructed in step S37 and the value of the representative error, the group identifier, the candidate explanatory variable to be added, and The value of each component of the error vector is stored in the second error data storage unit 405 (step S39).

  FIG. 16 shows an example of data stored in the second error data storage unit 405. In the example of FIG. 16, a group identifier, a candidate for an explanatory variable to be added, and a representative error, a predicted value, and an error for each date are stored.

  Then, the first evaluation value calculation unit 407 calculates an evaluation value representing the usefulness of the candidate by squaring each component of the error vector calculated in step S39 to obtain the sum (step S41). Then, the group identifier, the explanatory variable candidate to be added, and the evaluation value are stored in the first evaluation value storage unit 408.

  FIG. 17 shows an example of data stored in the first evaluation value storage unit 408. In the example of FIG. 17, group identifiers, candidate explanatory variables to be added, and evaluation values are stored.

  Then, the second model generation unit 401 determines whether there is an unprocessed candidate (step S43). If there is an unprocessed candidate (step S43: Yes route), the process returns to step S35 to process the next candidate.

  On the other hand, if there is no unprocessed candidate (step S43: No route), the extraction unit 410 sets N explanatory variable candidates as the first evaluation value for the group specified in step S31 in ascending order of evaluation value. Extracted from the storage unit 408 (step S45). Further, the extraction unit 410 stores the group identifier and the extracted explanatory variable candidates in the first extraction result storage unit 411.

  FIG. 18 shows an example of data stored in the first extraction result storage unit 411. In the example of FIG. 18, group identifiers and candidate explanatory variables to be added are stored.

  Then, the second model generation unit 401 determines whether there is an unprocessed group (step S47). If there is an unprocessed group (step S47: Yes route), the process returns to step S31 to process the next group. If there is no unprocessed group (step S47: No route), the process returns to the original process. .

  By carrying out the processing as described above, it is possible to specify an explanatory variable that can improve the accuracy particularly when used together with the explanatory variable used in the prediction model of the addition destination.

  Returning to the description of FIG. 6, the candidate extraction unit 4 performs the second candidate extraction process (step S5). The second candidate extraction process will be described with reference to FIGS.

  First, the outline of the second candidate extraction process will be described with reference to FIG. In order to simplify the explanation, the representative error vector is a one-dimensional vector. In the second candidate extraction process, as in the first candidate extraction process, N (N is a natural number of 2 or more) explanatory variable candidates are extracted for each group based on the accuracy of representative error prediction. That is, an evaluation value is calculated by calculating a sum of squares of differences between a prediction value based on a prediction model for predicting a representative error and a representative error, and N candidates are extracted in order from the candidate having the smallest calculated evaluation value. To do.

  However, in the second candidate extraction process, only the explanatory variable candidates to be added are used in the prediction model for predicting the representative error. In the example of FIG. 19, for a group including A transportation and B air, a prediction model is constructed using only learning data of candidate explanatory variables (gasoline or rice) to be added. This is to identify candidates that are useful for improving the accuracy of the prediction model when it is assumed that there is no synergistic effect due to a plurality of explanatory variables. In the example of FIG. 19, the evaluation value calculated for gasoline is smaller than the evaluation value calculated for rice, so for the group to which A Transport and B Airlines belong, the explanatory variable added by gasoline rather than rice This is preferable as a candidate. Normally, the accuracy of the prediction model is improved by increasing the number of explanatory variables, and therefore a value larger than the evaluation value calculated in the first candidate extraction process is calculated.

  Next, the process flow of the second candidate extraction process will be described. First, the second model generation unit 401 in the candidate extraction unit 4 identifies one unprocessed group among the groups registered in the grouping result storage unit 3 (FIG. 20: Step S51).

  The second model generation unit 401 identifies one unprocessed candidate among the explanatory variable candidates to be added (step S53). Further, the second model generation unit 401 uses the identified candidate learning data to construct a prediction model for predicting the representative error value calculated for the group identified in step S51 (step S55). . Then, the second model generation unit 401 stores the group identifier, the candidate of the explanatory variable to be added, and the data of the constructed prediction model in the third model storage unit 403. In step S55, processing is performed using the value of the representative error stored in the grouping result storage unit 3.

  FIG. 21 shows an example of data stored in the third model storage unit 403. In the example of FIG. 21, group identifiers, candidates for explanatory variables to be added, and prediction model data are stored.

  Then, the second error calculation unit 404 calculates an error vector based on the error between the value calculated by the prediction model constructed in step 55 and the value of the representative error (step S57). Then, the second error calculation unit 404 stores the group identifier, the candidate explanatory variable to be added, the value of each component of the error vector, and the like in the third error data storage unit 406.

  FIG. 22 shows an example of data stored in the third error data storage unit 406. In the example of FIG. 22, a group identifier, a candidate for an explanatory variable to be added, and a representative error, a predicted value, and an error for each date are stored.

  Then, the first evaluation value calculation unit 407 calculates an evaluation value representing the usefulness of the candidate by squaring each component of the error vector calculated in step S57 and obtaining the sum (step S59). The group identifier, the explanatory variable candidate to be added, and the evaluation value are stored in the second evaluation value storage unit 409.

  FIG. 23 shows an example of data stored in the second evaluation value storage unit 409. In the example of FIG. 23, an identifier of a group, a candidate for an explanatory variable to be added, and an evaluation value are stored.

  Then, the second model generation unit 401 determines whether there is an unprocessed candidate (step S61). If there is an unprocessed candidate (step S61: Yes route), the process returns to step S53 to process the next candidate.

  On the other hand, when there is no unprocessed candidate (step S61: No route), the extraction unit 410 selects the second explanatory value for N explanatory variable candidates in ascending order of evaluation value for the group specified in step S51. Extracted from the storage unit 409 (step S63). Further, the extraction unit 410 stores the group identifier and the extracted explanatory variable candidates in the second extraction result storage unit 412.

  FIG. 24 shows an example of data stored in the second extraction result storage unit 412. In the example of FIG. 24, a group identifier and a candidate for an explanatory variable to be added are stored.

  Then, the second model generation unit 401 determines whether there is an unprocessed group (step S65). If there is an unprocessed group (step S65: Yes route), the process returns to step S51 to process the next group. If there is no unprocessed group (step S65: No route), the process returns to the original process. .

  By performing the processing as described above, it is possible to identify candidates that are useful for improving the accuracy of the prediction model when it is assumed that there is no synergistic effect due to a plurality of explanatory variables.

  Returning to the description of FIG. 6, the first specifying unit 413 in the candidate extracting unit 4 determines a final candidate for each group based on the results of the first candidate extracting process and the second candidate extracting process, and determines the group identifier and the final candidate. Store in the candidate storage unit 6 (step S7). Specifically, the candidates for explanatory variables stored in the first extraction result storage unit 411 and stored in the second extraction result storage unit 412 are determined as final candidates.

  FIG. 25 shows an example of data stored in the candidate storage unit 6. In the example of FIG. 25, a group identifier and a candidate for an explanatory variable to be added are stored.

  And the determination part 7 implements a determination process (step S9). The determination process will be described with reference to FIGS.

  First, the outline of the determination process will be described with reference to FIG. In the decision process, for each stock, each final candidate for the group to which the stock belongs is actually added to the prediction model to calculate an evaluation value, and an explanatory variable that is highly useful and represented by the evaluation value is added to the prediction model. Decide which variables to use. For example, when gasoline and diesel oil are the final candidates for Group 1, each of gasoline and diesel oil is actually added to the prediction model to calculate the evaluation value, and the one with the smallest evaluation value is used as the prediction model. Determine the explanatory variable to be added. In the example of FIG. 26, since the evaluation value calculated for gasoline is smaller than the evaluation value calculated for light oil, the explanatory variable added to the prediction model of A transportation is determined to be gasoline.

  Next, the processing flow of the determination process will be described. First, the 3rd model production | generation part 71 in the determination part 7 specifies one unprocessed brand | brand among the brands in which learning data are stored in the learning data storage part 5 (FIG. 27: step S71). In addition, the third model generation unit 71 specifies one unprocessed final candidate from the final candidates for the group to which the brand specified in step S71 belongs (step S73). In step S73, first, the group to which the brand specified in step S71 belongs is specified from the grouping result storage unit 3, and the final candidate corresponding to the specified group is specified from the candidate storage unit 6, and the specified final candidate An unprocessed final candidate is identified from the list.

  Then, the third model generation unit 71 predicts the stock price of “today” of the specified brand using the learning data of the brand specified in step S71 and the learning data of the final candidate specified in step S73. A prediction model is constructed (step S75). Then, the third model generation unit 71 stores the brand name, the candidate explanatory variable to be added, and the data of the constructed prediction model in the fourth model storage unit 72.

  FIG. 28 shows an example of data stored in the fourth model storage unit 72. In the example of FIG. 28, brand names, candidates for explanatory variables to be added, and prediction model data are stored.

  Then, the third error calculation unit 73 calculates an error vector based on an error between the actual value of the stock price of the specified brand and the value calculated by the prediction model constructed in step S75, and adds the brand name. The candidate of explanatory variable, the value of each component of the error vector, and the like are stored in the fourth error data storage unit 74 (step S77).

  FIG. 29 shows an example of data stored in the fourth error data storage unit 74. In the example of FIG. 29, brand names, candidates for explanatory variables to be added, and actual values, predicted values, and errors of stock prices for each date are stored.

  Then, the second evaluation value calculation unit 75 calculates the evaluation value representing the usefulness of the candidate by squaring each component of the error vector calculated in step S77 to obtain the sum, and adds the brand name. The candidate explanatory variables and the evaluation value are stored in the third evaluation value storage unit 76 (step S79).

  FIG. 30 shows an example of data stored in the third evaluation value storage unit 76. In the example of FIG. 30, a brand name, a candidate for an explanatory variable to be added, and an evaluation value are stored.

  Then, the third model generation unit 71 determines whether there is an unprocessed candidate in the candidate storage unit 6 (step S81). If there is an unprocessed candidate (step S81: Yes route), the process returns to step S73 to process the next candidate.

  On the other hand, when there is no unprocessed candidate (step S81: No route), the second specifying unit 77 stores an explanatory variable to be added to the prediction model of the brand specified in step S71 in the third evaluation value storage unit 76. An explanatory variable that is determined on the basis of the evaluation value and is added in association with the brand name is stored in the determination result storage unit 8 (step S83). Specifically, the explanatory variable having the smallest evaluation value is determined.

  FIG. 31 shows an example of data stored in the determination result storage unit 8. In the example of FIG. 31, brand names and explanatory variables added to the prediction model are stored.

  And the 3rd model production | generation part 71 judges whether there exists an unprocessed brand (step S85). If there is an unprocessed brand (step S85: Yes route), the process returns to step S71 to process the next brand. On the other hand, when there is no unprocessed brand (step S85: No route), the process returns to the original process.

  By carrying out the processing as described above, it is possible to specify the optimum explanatory variable for each brand.

  Returning to the description of FIG. 6, the output unit 9 determines using the data stored in the grouping result storage unit 3, the data stored in the candidate storage unit 6, and the data stored in the determination result storage unit 8. Generate screen data to display the results. Then, the generated screen data is displayed on the display device (step S10). Then, the process ends.

  FIG. 32 shows an example of the displayed screen. In the example of FIG. 32, for each group, explanatory variables to be added to the prediction model of the brand belonging to the group, candidates for explanatory variables to be added to the prediction model of the brand belonging to the group, and the representative error vector calculated for the group And are displayed.

  As described above, instead of determining the explanatory variables to be added to the prediction model for each brand, it is possible to reduce the amount of calculation when determining the explanatory variables to be added to the prediction model by determining for each group. become able to.

  Here, the process in the case where the explanatory variable added to the prediction model is determined for each brand will be briefly described with reference to FIG. First, a processing unit (not shown) in the information processing apparatus 1 identifies one unprocessed brand (FIG. 33: step S101). The processing unit identifies one unprocessed candidate among the explanatory variable candidates to be added (step S103). The processing unit builds a prediction model using the specified brand learning data and the specified candidate learning data (step S105). The processing unit calculates an error vector using the predicted value based on the constructed prediction model and the actual value (step S107). The processing unit calculates an evaluation value representing the usefulness of the candidate by squaring the value of each component of the error vector and obtaining the sum (step S109). If there is an unprocessed candidate (step S111: Yes route), the process returns to step S103. If there is no unprocessed candidate (step S111: No route), the explanatory variable having the highest usefulness represented by the evaluation value is used. Is specified (step S113). If there is an unprocessed brand (step S115: Yes route), the process returns to step S101. If there is no unprocessed brand (step S115: No route), the process is terminated.

  As described in the Background Art section, in this way, the total amount of calculation is approximately (the amount of calculation required to calculate the evaluation value for one explanatory variable) × (the number of objective variables (that is, the number of issues). ) × (number of candidates for explanatory variables). Therefore, when the number of objective variables and the number of explanatory variables increase, there is a problem that the amount of calculation becomes very large.

  On the other hand, according to the processing of the present embodiment as described above, the total calculation amount is approximately (calculation amount required for grouping processing) + (calculation amount required to calculate an evaluation value for one explanatory variable). X (number of groups) x (number of explanatory variable candidates) + (calculation amount required to calculate an evaluation value for one explanatory variable) x (number of groups) x (stocks included in group (objective variable)) Number) × (number of final candidates for explanatory variables). As a result, the amount of calculation can be reduced.

  Although one embodiment of the present technology has been described above, the present technology is not limited to this. For example, the functional block configuration of the information processing apparatus 1 described above does not necessarily correspond to an actual program module configuration.

  Further, the configuration of each table described above is an example, and the configuration as described above is not necessarily required. Further, in the processing flow, the processing order can be changed if the processing result does not change. Further, it may be executed in parallel.

  Although stock price data is used as time series data, the data used is not limited to stock price data, and the present embodiment can be applied to other time series data.

  The data to be used may be data other than time series data. That is, the time does not have to be associated with the objective variable and the explanatory variable.

  Further, in step S7, the explanatory variable candidates stored in the first extraction result storage unit 411 and stored in the second extraction result storage unit 412 are set as final candidates. You may decide. For example, a predetermined number of explanatory variables may be set as final candidates in order from the smallest sum of the evaluation value calculated in the first candidate extraction process and the evaluation value calculated in the second candidate extraction process.

  Further, the processing of the information processing apparatus 1 may be executed by a plurality of computers.

  The information processing apparatus 1 described above is a computer apparatus, and as shown in FIG. 34, a memory 2501, a CPU (Central Processing Unit) 2503, a hard disk drive (HDD: Hard Disk Drive) 2505, and a display device. A display control unit 2507 connected to 2509, a drive device 2513 for the removable disk 2511, an input device 2515, and a communication control unit 2517 for connecting to a network are connected by a bus 2519. An operating system (OS: Operating System) and an application program for executing the processing in this embodiment are stored in the HDD 2505, and are read from the HDD 2505 to the memory 2501 when executed by the CPU 2503. The CPU 2503 controls the display control unit 2507, the communication control unit 2517, and the drive device 2513 according to the processing content of the application program, and performs a predetermined operation. Further, data in the middle of processing is mainly stored in the memory 2501, but may be stored in the HDD 2505. In an embodiment of the present technology, an application program for performing the above-described processing is stored in a computer-readable removable disk 2511 and distributed, and installed from the drive device 2513 to the HDD 2505. In some cases, the HDD 2505 may be installed via a network such as the Internet and the communication control unit 2517. Such a computer apparatus realizes various functions as described above by organically cooperating hardware such as the CPU 2503 and the memory 2501 described above and programs such as the OS and application programs. .

  The embodiments of the present technology described above are summarized as follows.

  The information processing apparatus according to this embodiment is

  For each of (A) a storage device and (B) a plurality of objective variables, an error between the actual value of the objective variable and the value calculated by the first prediction model for predicting the value of the objective variable is calculated. A first calculation unit that calculates and stores in the storage device; and (C) classifies the plurality of objective variables into a plurality of groups based on the error stored in the storage device, and assigns each of the plurality of groups to the group. A second calculation unit that calculates a representative value of the error using the error calculated for the target variable to which the variable belongs, and stores it in the storage device; and (D) a representative value stored in the storage device for each of the plurality of groups. A plurality of second prediction models for predicting the value are generated while changing the explanatory variables, and belong to the group based on the difference between each of the values calculated by the plurality of generated second prediction models and the representative value Purpose change And a first determining unit for determining the explanatory variable to be added to the first predictive model.

  In this way, the calculation performed when determining the explanatory variable to be added to the first prediction model by determining the explanatory variable to be added to the first prediction model by determining for each group instead of determining the explanatory variable to be added to the first prediction model. The amount of can be reduced.

  Further, the first determination unit described above uses (d1) a second for each of the plurality of explanatory variable candidates by using the candidate and the explanatory variable included in the first prediction model of the objective variable belonging to the group. A third calculation unit that generates a first difference between the value calculated by the second prediction model and the representative value and stores the first difference in the storage device; and (d2) the third calculation unit. Based on the calculated first difference, a second determining unit that determines an explanatory variable to be added to the first prediction model of the target variable belonging to the group from among a plurality of explanatory variable candidates. Good. When multiple explanatory variables are included in the same prediction model, the prediction accuracy may be greatly improved due to a synergistic effect. Therefore, as described above, it is difficult to miss an explanatory variable whose prediction accuracy is greatly improved when it is used together with the explanatory variable included in the first prediction model.

  Further, the third calculation unit described above (d11) generates a second prediction model for each of a plurality of explanatory variable candidates using the candidate, and the value calculated by the second prediction model The second difference between the value and the representative value may be calculated and stored in the storage device. (E) Explanation to be added to the first prediction model of the objective variable belonging to the group from among a plurality of candidate explanatory variables based on the first difference and the second difference calculated by the third calculation unit. You may make it further have the 3rd determination part which determines a variable. In this way, when it is assumed that there is no synergistic effect with the explanatory variable included in the first prediction model, it is difficult to miss the explanatory variable effective for improving the prediction accuracy.

  In addition, the first determination unit described above may determine (d3) a plurality of explanatory variables to be added to the first prediction model. Then, the information processing apparatus described above performs (F) for each of the plurality of objective variables, the first determination unit determines the explanatory variable included in the first prediction model of the objective variable and the group to which the objective variable belongs. A plurality of third prediction models for predicting the value of the target variable are generated using each of the determined explanatory variables, and each of the values calculated by the third prediction model and the target A fourth calculation unit that calculates an error from the actual value of the variable and stores the error in the storage device; and (G) a first determination unit based on the error calculated by the fourth calculation unit for each of the plurality of target variables. And a fourth determination unit that determines the most appropriate explanatory variable to be added to the first prediction model from among the plurality of explanatory variables determined by the above. In this way, the most effective explanatory variable for improving the prediction accuracy of each objective variable can be identified.

  The second calculation unit described above may classify (c1) a plurality of objective variables into a plurality of groups by clustering based on the calculated error. For example, a plurality of objective variables can be appropriately classified by using the K average method.

  Further, the representative value of the error described above may be an average value of errors calculated for the objective variable belonging to the group. As a result, an appropriate value can be used as the representative value. In addition, it is not restricted to an average value, For example, it is good also as a median value.

  In the information processing method according to the present embodiment, (H) for each of a plurality of objective variables, the value calculated by the first prediction model for predicting the actual value of the objective variable and the value of the objective variable And (I) classifying a plurality of objective variables into a plurality of groups based on the error stored in the storage device, and (J) for each of the plurality of groups A representative value of the error is calculated using the error calculated for the objective variable belonging to the group, stored in the storage device, and (K) a representative value stored in the storage device is predicted for each of the plurality of groups. A plurality of second prediction models for the target variable belonging to the group are generated based on a difference between each of the values calculated by the plurality of generated second prediction models and the representative value. First Including a process of determining the explanatory variable to be added to the prediction model.

  A program for causing a computer to perform the processing according to the above method can be created. The program can be a computer-readable storage medium such as a flexible disk, a CD-ROM, a magneto-optical disk, a semiconductor memory, a hard disk, or the like. It is stored in a storage device. The intermediate processing result is temporarily stored in a storage device such as a main memory.

  The following supplementary notes are further disclosed with respect to the embodiments including the above examples.

(Appendix 1)
A storage device;
For each of the plurality of objective variables, an error between the actual value of the objective variable and the value calculated by the first prediction model for predicting the value of the objective variable is calculated and stored in the storage device. 1 calculation unit;
The plurality of objective variables are classified into a plurality of groups based on the errors stored in the storage device, and for each of the plurality of groups, the error calculated using the errors calculated for the objective variables belonging to the group. A second calculating unit that calculates a representative value and stores the representative value in the storage device;
For each of the plurality of groups, a plurality of second prediction models for predicting the representative value stored in the storage device are generated while changing the explanatory variables, and the generated plurality of second prediction models are used. A first determination unit that determines an explanatory variable to be added to the first prediction model of an objective variable belonging to the group based on a difference between each of the calculated values and the representative value;
An information processing apparatus.

(Appendix 2)
The first determination unit is
For each of a plurality of explanatory variable candidates, the second prediction model is generated by using the candidate and the explanatory variable included in the first prediction model of the objective variable belonging to the group, and the second prediction Calculating a first difference between the value calculated by the model and the representative value, and storing the first difference in the storage device;
Determining an explanatory variable to be added to the first prediction model of the objective variable belonging to the group from the plurality of explanatory variable candidates based on the first difference calculated by the third calculating unit; 2 decision part;
The information processing apparatus according to claim 1, further comprising:

(Appendix 3)
The third calculation unit is
For each of the plurality of explanatory variable candidates, the second prediction model is generated using the candidate, and a second difference between the value calculated by the second prediction model and the representative value is calculated. , Store in the storage device,
Based on the first difference and the second difference calculated by the third calculation unit, the candidate is added to the first prediction model of the objective variable belonging to the group from the plurality of explanatory variable candidates. The information processing apparatus according to attachment 2, further comprising a third determination unit that determines an explanatory variable.

(Appendix 4)
The first determination unit is
Determining a plurality of explanatory variables to be added to the first prediction model;
For each of the plurality of objective variables, an explanatory variable included in the first prediction model of the objective variable and each of the plurality of explanatory variables determined by the first determination unit for the group to which the objective variable belongs are used. A plurality of third prediction models for predicting the value of the target variable, and calculating an error between each of the values calculated by the third prediction model and the actual value of the target variable, A fourth calculation unit stored in the storage device;
For each of the plurality of objective variables, most of the plurality of explanatory variables determined by the first determination unit based on the error calculated by the fourth calculation unit is added to the first prediction model. A fourth determination unit for determining an appropriate explanatory variable;
The information processing apparatus according to any one of appendices 1 to 3, further comprising:

(Appendix 5)
The second calculation unit is
The information processing apparatus according to any one of attachments 1 to 4, wherein the plurality of objective variables are classified into a plurality of groups by clustering based on the calculated error.

(Appendix 6)
The information processing apparatus according to any one of attachments 1 to 5, wherein the representative value of the error is an average value of the errors calculated for the objective variable belonging to the group.

(Appendix 7)
For each of the plurality of objective variables, an error between the actual value of the objective variable and the value calculated by the first prediction model for predicting the value of the objective variable is calculated and stored in the storage device.
Classifying the plurality of objective variables into a plurality of groups based on the error stored in the storage device;
For each of the plurality of groups, a representative value of the error is calculated using the error calculated for the objective variable belonging to the group, and is stored in the storage device.
For each of the plurality of groups, a plurality of second prediction models for predicting the representative value stored in the storage device are generated while changing the explanatory variables, and the generated plurality of second prediction models are used. Based on a difference between each calculated value and the representative value, an explanatory variable to be added to the first prediction model of the objective variable belonging to the group is determined.
An information processing method in which processing is executed by a computer.

(Appendix 8)
For each of the plurality of objective variables, an error between the actual value of the objective variable and the value calculated by the first prediction model for predicting the value of the objective variable is calculated and stored in the storage device.
Classifying the plurality of objective variables into a plurality of groups based on the error stored in the storage device;
For each of the plurality of groups, a representative value of the error is calculated using the error calculated for the objective variable belonging to the group, and is stored in the storage device.
For each of the plurality of groups, a plurality of second prediction models for predicting the representative value stored in the storage device are generated while changing the explanatory variables, and the generated plurality of second prediction models are used. Based on a difference between each calculated value and the representative value, an explanatory variable to be added to the first prediction model of the objective variable belonging to the group is determined.
A program that causes a computer to execute processing.

DESCRIPTION OF SYMBOLS 1 Information processing apparatus 2 Grouping process part 3 Grouping result storage part 4 Candidate extraction part 5 Learning data storage part 6 Candidate storage part 7 Determination part 8 Determination result storage part 9 Output part 21 1st model production | generation part 22 1st model storage part 23 First error calculation unit 24 First error data storage unit 25 Group generation unit 401 Second model generation unit 402 Second model storage unit 403 Third model storage unit 404 Second error calculation unit 405 Second error data storage unit 406 Third Error data storage unit 407 First evaluation value calculation unit 408 First evaluation value storage unit 409 Second evaluation value storage unit 410 Extraction unit 411 First extraction result storage unit 412 Second extraction result storage unit 413 First identification unit 71 Third Model generation unit 72 Fourth model storage unit 73 Third error calculation unit 74 Fourth error data storage unit 75 Second evaluation value calculation unit 76 Third evaluation value storage unit 77 Second Part

Claims (6)

A storage device;
For each of the plurality of objective variables, an error between the actual value of the objective variable and the value calculated by the first prediction model for predicting the value of the objective variable is calculated and stored in the storage device. 1 calculation unit;
The plurality of objective variables are classified into a plurality of groups based on the errors stored in the storage device, and for each of the plurality of groups, the error calculated using the errors calculated for the objective variables belonging to the group. A second calculating unit that calculates a representative value and stores the representative value in the storage device;
For each of the plurality of groups, a plurality of second prediction models for predicting the representative value stored in the storage device are generated while changing the explanatory variables, and the generated plurality of second prediction models are used. A first determination unit that determines an explanatory variable to be added to the first prediction model of an objective variable belonging to the group based on a difference between each of the calculated values and the representative value;
An information processing apparatus. The first determination unit is
For each of a plurality of explanatory variable candidates, the second prediction model is generated by using the candidate and the explanatory variable included in the first prediction model of the objective variable belonging to the group, and the second prediction Calculating a first difference between the value calculated by the model and the representative value, and storing the first difference in the storage device;
Determining an explanatory variable to be added to the first prediction model of the objective variable belonging to the group from the plurality of explanatory variable candidates based on the first difference calculated by the third calculating unit; 2 decision part;
The information processing apparatus according to claim 1. The third calculation unit is
For each of the plurality of explanatory variable candidates, the second prediction model is generated using the candidate, and a second difference between the value calculated by the second prediction model and the representative value is calculated. , Store in the storage device,
Based on the first difference and the second difference calculated by the third calculation unit, the candidate is added to the first prediction model of the objective variable belonging to the group from the plurality of explanatory variable candidates. The information processing apparatus according to claim 2, further comprising a third determination unit that determines an explanatory variable. The first determination unit is
Determining a plurality of explanatory variables to be added to the first prediction model;
For each of the plurality of objective variables, an explanatory variable included in the first prediction model of the objective variable and each of the plurality of explanatory variables determined by the first determination unit for the group to which the objective variable belongs are used. A plurality of third prediction models for predicting the value of the target variable, and calculating an error between each of the values calculated by the third prediction model and the actual value of the target variable, A fourth calculation unit stored in the storage device;
For each of the plurality of objective variables, most of the plurality of explanatory variables determined by the first determination unit based on the error calculated by the fourth calculation unit is added to the first prediction model. A fourth determination unit for determining an appropriate explanatory variable;
The information processing apparatus according to any one of claims 1 to 3, further comprising: For each of the plurality of objective variables, an error between the actual value of the objective variable and the value calculated by the first prediction model for predicting the value of the objective variable is calculated and stored in the storage device.
Classifying the plurality of objective variables into a plurality of groups based on the error stored in the storage device;
For each of the plurality of groups, a representative value of the error is calculated using the error calculated for the objective variable belonging to the group, and is stored in the storage device.
For each of the plurality of groups, a plurality of second prediction models for predicting the representative value stored in the storage device are generated while changing the explanatory variables, and the generated plurality of second prediction models are used. Based on a difference between each calculated value and the representative value, an explanatory variable to be added to the first prediction model of the objective variable belonging to the group is determined.
An information processing method in which processing is executed by a computer. For each of the plurality of objective variables, an error between the actual value of the objective variable and the value calculated by the first prediction model for predicting the value of the objective variable is calculated and stored in the storage device.
Classifying the plurality of objective variables into a plurality of groups based on the error stored in the storage device;
For each of the plurality of groups, a representative value of the error is calculated using the error calculated for the objective variable belonging to the group, and is stored in the storage device.
For each of the plurality of groups, a plurality of second prediction models for predicting the representative value stored in the storage device are generated while changing the explanatory variables, and the generated plurality of second prediction models are used. Based on a difference between each calculated value and the representative value, an explanatory variable to be added to the first prediction model of the objective variable belonging to the group is determined.
A program that causes a computer to execute processing.

Images