JP7235329B2 - Economic indicator estimation system and its program - Google Patents

Description

The present invention relates to an economic indicator estimation system for estimating economic indicators and a program thereof.

Conventionally, there is known a method of predicting trends in the economic environment and financial markets by utilizing various text data such as news articles. For example, Non-Patent Document 1 discloses a method of indexing daily distributed economic news to perform nowcasting of economic trends and predicting asset price volatility using the business sentiment news index. The features of this method are: first, it uses a convolutional neural network (CNN), one of the deep learning models, to estimate business sentiment from economic news; It is a point to measure daily economic trends by indexing. In order to construct the news index, first, using the economic judgment reason collection of the economic watchers survey published by the Cabinet Office as training data, supervised learning is performed by CNN to construct a learner that performs text classification. Next, using the trained learner, the sentences that make up the Japanese articles of economic news are given scores related to business sentiment. Finally, a news index is constructed by aggregating the scored sentences on a monthly and daily basis.

Non-Patent Document 2 describes a method of digitizing and aggregating the text of financial reports at low cost and at high speed by automatically determining the economic sentiment of the text using a recurrent neural network (RNN). is disclosed. After having the RNN learn the task of predicting the economic sentiment (positive/negative) from the text of the economy watcher survey, the economic sentiment of the document is determined. In addition, Non-Patent Document 2 uses such a sentiment estimation model to estimate and index the sentiment of monthly reports issued by the government and the Bank of Japan. , and when calculating the correlation with the Nikkei Stock Average, it showed a higher value than the Bank of Japan Tankan and Economy Watchers Index, which are already widely used as investment indices. ing.

In addition, Non-Patent Document 3 discloses a method of indexing the sentiment of the Bank of Japan's economy for each topic using a topic model and a neural network for texts issued by the Bank of Japan. .

Furthermore, Patent Literature 1 discloses an information processing apparatus that predicts economic indicators regardless of whether news information or the like is published. This information processing device has a model storage unit, an acquisition unit, and a prediction unit. The model storage unit stores a prediction model set based on the location information of the terminal. The acquisition unit acquires location information of one or more terminals. The prediction unit applies the prediction model stored in the model storage unit to the location information acquired by the acquisition unit to predict the specified economic index.

JP 2019-46376 A

Keiichi Goto and two others, ``Construction of business sentiment news index by natural language processing and its application to volatility forecast'', [online], January 2019, IMES DISCUSSION PEPER SERIES, Japan, Institute for Monetary and Economic Studies, Bank of Japan, [2020 Searched on May 28], Internet <URL: http://www.imes.boj.or.jp/research/papers/japanese/19-J-03.pdf> Hiroki Yamamoto et al., "Indexing Financial Reports Using Deep Learning from Economy Watchers Survey", [online], June 6, 2016, 30th National Conference (2016), Japan, Japan Artificial Intelligence Intelligence Society, [Retrieved May 28, 2020], Internet <URL: https://www.ai-gakkai.or.jp/jsai2016/webprogram/2016/pdf/219.pdf> Kyoto Yono and others, "Prediction of real-time BOJ sentiment based on financial reports and macroeconomic indices", [online], May 23, 2017, 31st National Conference (2017), Japan, General Incorporated Association Japan Japanese Society for Artificial Intelligence, [Retrieved May 28, 2020], Internet <URL: https://www.jstage.jst.go.jp/article/pjsai/JSAI2017/0/JSAI2017_2D13/_pdf/-char/ja>

An object of the present invention is to provide a novel technique for estimating economic indicators with arbitrary time resolution.

In order to solve such a problem, a first invention is an economic An economic indicator estimation system for estimating indicators is provided. This system has a digest generator, an economic phenomenon database, a frequency totalizer, a vector totalizer, and a learning processor. The digest generation unit generates an economic event digest by structuring the content of economic events extracted from news groups collected from the outside with a plurality of predetermined items. The economic event database stores economic event digests. The frequency aggregating unit classifies the economic event digests stored in the economic event database into groups of economic event digests stored in the economic event database as aggregation targets, and classifies the economic event digests into economic event patterns having commonalities in content for each predetermined aggregation time unit. Aggregate the frequency of occurrence of economic event patterns. The vector estimator calculates, for each economic event pattern for each aggregated time unit, an economic event frequency vector consisting of non-negative real values mapped onto a common space, which corresponds to the appearance of each economic event pattern for each aggregated time unit. Estimate to reproduce the frequency. The vector aggregation unit aggregates the economic event frequency vectors belonging to the first time unit for each first time unit corresponding to the time resolution of the economic index value, and calculates each economic event pattern within the first time unit. A first economic event aggregation vector representing the degree of occurrence is generated. The learning processing unit learns a regression model for estimating an economic index so that an economic index value temporally corresponding to the input of the first economic event aggregation vector responds.

Here, in the first invention, a news filter section may be provided that extracts media names described in a preset news media list from the news group and outputs the extracted media names to the digest generation section. Further, a narrowing processing unit may be provided that extracts economic phenomenon digests that match preset narrowing conditions from among the economic event digests generated by the digest generating unit and stores them in the economic phenomenon database. In this case, the narrowing-down processing unit refers to, as the narrowing-down condition, a performance factor list describing patterns of economic event digests that affect specific economic index values to be estimated, and selects economic event digests. Extraction is preferred.

In the first invention, the vector aggregation unit calculates the sum of economic event frequency vectors normalized so that the sum of the components of each dimension is 1, and normalizes by the appearance frequency in the first time unit Preferably, the first economic event aggregation vector is generated by:

In the first invention, an estimation processing unit that outputs a response of the learned regression model to the input of the second aggregate vector of economic events as an estimated value of an economic indicator temporally corresponding to the second aggregate vector of economic events. may be provided. In this case, the vector aggregating unit aggregates economic event frequency vectors belonging to a second time unit having a time resolution different from that of the economic index value, and expresses the degree of occurrence of each economic event pattern within the second time unit. Preferably, a second economic event aggregation vector is generated. In addition, the vector counting unit calculates the sum of economic event frequency vectors normalized so that the sum of the components of each dimension is 1, and normalizes by the appearance frequency in the second time unit, Preferably, a second economic event aggregation vector is generated.

A second invention is an economic indicator estimation method for estimating an economic indicator with a desired time resolution based on a discrete time series of economic indicator values representing specific economic indicators and economic events affecting the economic indicator values. Offer a program. This program causes the computer to execute processing having the following first to sixth steps. In the first step, an economic event digest is generated by structuring the content of the economic event extracted from the news group collected from the outside with a plurality of predetermined items. In a second step, the economic event digest is stored in the economic event database. In the third step, a group of economic event digests stored in the economic event database is targeted for aggregation, and the economic event digests are classified into economic event patterns having commonalities in content for each predetermined aggregation time unit. aggregate the frequency of occurrence of economic event patterns. In the fourth step, for each aggregated time unit economic event pattern, an economic event frequency vector consisting of non-negative real values mapped onto a common space is obtained as Estimate so as to reproduce the appearance frequency. In the fifth step, for each first time unit corresponding to the time resolution of the economic index value, the economic event frequency vectors belonging to the first time unit are aggregated, and each economic event pattern within the first time unit generates a first economic event aggregation vector representing the degree of occurrence of . In the sixth step, a regression model for estimating economic indices is learned so that economic index values temporally corresponding to the input of the first economic event aggregation vector respond.

Here, in the second invention, the first step may have a step of extracting, from among the news groups, news items with media names described in a preset news media list. Further, the first step may include a step of extracting economic event digests that match preset narrowing conditions and storing them in the economic event database. In this case, the first step refers to, as the narrowing condition, a performance factor list describing patterns of economic event digests that affect specific economic index values to be estimated, and extracts the economic event digests. Extraction is preferred.

In the second invention, the third step calculates the sum of the economic event frequency vectors normalized so that the sum of the components of each dimension is 1, and normalizes the sum of the economic event frequency vectors in the first time unit It is preferable to generate the first economic event aggregation vector by unifying.

In the second invention, in addition to the steps described above, the computer may be caused to execute processing having the following seventh and seventh steps. In the seventh step, the response of the trained regression model to the input of the second aggregate vector of economic events is output as the estimated value of the economic index temporally corresponding to the second aggregate vector of economic events. In the eighth step, economic event frequency vectors belonging to a second time unit with a time resolution different from the economic index value are aggregated to obtain a second time unit representing the degree of occurrence of each economic event pattern within the second time unit. Generate an economic event aggregation vector. In this case, the eighth step calculates the sum of the normalized economic event frequency vectors so that the sum of the components of each dimension is 1, and normalizes by the appearance frequency in the second time unit. It is preferred to generate the second economic event aggregation vector by:

In the first and second inventions, the second time unit may have higher time resolution than the economic indicator value.

In the first and second inventions, the second time unit may have higher time resolution than the economic indicator value.

According to the present invention, the regression model is learned so that the economic indicator value temporally corresponding to the input of the first economic event aggregation vector responds. The first economic event aggregation vector represents the degree of occurrence of each economic event pattern within the first time unit. As a learning result of the regression model, an economic event and an economic index value are associated in a form that takes into consideration the frequency of occurrence of patterned economic events, in other words, the degree of influence of an economic event on a certain economic index value. By using the regression model constructed in this way, it becomes possible to estimate economic indicators corresponding to this time resolution as the response of the regression model to the input of arbitrary time resolution.

Block diagram of economic indicator estimation system Diagram showing an example of a time series of economic index values A diagram showing an example of a news group Diagram showing an example of an economic event digest A diagram showing an example of a normalization dictionary Diagram showing an example of narrowing conditions Diagram showing an example of a company list Diagram showing an example of company performance factor data Block diagram of index estimator A diagram showing an example of the results of daily aggregation of the frequency of occurrence of economic event patterns A diagram showing an example of the date latent vector θd Diagram showing an example of the latent vector θi of name (item) A diagram showing an example of the latent vector θj of an element Diagram showing an example of latent vector θk of fluctuation (predicate) A diagram showing an example of a daily unit economic event aggregation vector Diagram showing an example of the relationship between input variables and response variables in a regression model A chart showing an example of smoothed estimates of economic indicators

FIG. 1 is a block diagram of an economic index estimation system according to this embodiment. This economic indicator estimation system 1 estimates economic indicators with a desired time resolution based on known economic indicator values and economic events extracted from a known news group. Here, the "economic index value" represents a specific economic index, and is provided as a discrete time series (economic index data) such as daily or monthly. In addition, "economic event" refers to an event that affects a specific economic index value to be estimated. Economic events and economic index values have some kind of causal relationship, and if an economic event (not limited to one) occurs within a certain period of time, the corresponding economic index value may fluctuate in time.

FIG. 2 shows the time series of the Economy Watchers Manufacturing Index released monthly as an example of the time series of economic index values. The time resolution of this economic index value is monthly, and is obtained from the outside as data to be input to the economic index estimation system 1 . In addition to the economy watchers manufacturing index, such economic indicators include the Bank of Japan's Tankan survey, the industrial production index, the unemployment rate, the number of automobiles sold, and the housing starts statistics, all of which have a causal relationship with economic events. Any economic indicator value, if any, can be used. Moreover, the publication cycle (time resolution) of the economic index values may be quarterly or half-monthly, and may be irregular or random.

FIG. 3 is a diagram showing an example of a news group from which economic events are extracted. This news group is a collection of news articles (may be one article) consisting of an article ID, media name, distribution date and time, and article text, and is collected from various external sources such as the Internet as needed. be.

The economic index estimation system 1 mainly includes a news filter unit 2, a digest generation unit 3, a narrowing processing unit 4, an economic event database 5, and an index estimation unit 6.

The news filter unit 2 extracts only media names that are highly likely to distribute news articles about economic events from the input news group, and excludes other media. This media name is preset as a news media list, and media names not described here are regarded as noise essentially unrelated to economic events. In the example of FIG. 3, news articles such as entertainment newspapers with article ID=“1005” and agricultural newspapers with article ID=“1006”, which are not listed in the news media list, are excluded. By performing such filtering based on the source of the news, the subsequent processing load can be reduced.

The digest generator 3 extracts economic events based on the news group filtered by the news filter 2 . Multiple economic events may be extracted from a single news article, or none may be extracted. The contents of the extracted economic event are output in the form of an economic event digest. FIG. 4 is a diagram showing an example of an economic event digest extracted from a news group. An economic event digest is structured by classifying the content of an economic event into a plurality of predetermined items, and expresses the characteristics (feature values) of an economic event in a simple manner without redundancy. . As an example, an economic event digest may consist of a set of "items", "elements" and "predicates". "Name (item)" is an item representing the name of an economic phenomenon, such as "crude oil" or "gasoline". An "element" is an item such as "price" or "demand" that expresses the quantity or tendency of an economic phenomenon. “Predicate” is an item that represents the direction (+/−) of fluctuations of economic phenomena (“elements”), such as “rise” and “increase”. Of these three items, however, the most important ones that characterize economic events are the "element" and the "predicate," i. has “fallen”). Therefore, "element" and "predicate" are indispensable, but "name" can be adopted as necessary, or another item other than these can be added. You may

In addition, "name (item)", "element (element)" and "variation (predicate)" normalize the extracted text to regular text using a normalization dictionary in order to eliminate fluctuations in expression, that is, It is preferable to unify the expressions. FIG. 5 shows an example of a normalization dictionary. For example, with regard to fluctuations (predicate), the extracted texts such as "soaring", "increasing", and "many" are converted to the regular text of "increasing", which has a positive fluctuation direction. is converted to regular text "decrease" with a negative change direction. As a result, even if the expression of the extracted text is different, it becomes possible to uniformly handle it as the same meaning on the system.

The narrowing-down processing unit 4 individually evaluates the economic event digests generated by the digest generating unit 4 according to preset narrowing-down conditions, and identifies economic events related to specific economic index values to be estimated. Extract the digest. FIG. 6 shows an example of narrowing-down conditions. In this embodiment, the narrowing-down condition is defined as a performance factor list describing patterns of economic event digests that affect specific economic index values to be estimated. As a result, among the economic event digests generated by the digest generation unit 4, those that match the pattern described in the performance factor list are extracted, and those that do not match are extracted from the specific economic index value to be estimated. Removed as irrelevant.

The performance factor list can be created, for example, using two types of data: a company list and company performance factor data. As shown in FIG. 7, the company list lists the names of companies that influence specific economic index values to be estimated. The description of the company name may be the name of the company, but in the case of a listed company, the ambiguity can be eliminated by using the brand code or the like. In addition, if classification data such as the 33 industries of the Tokyo Stock Exchange are used, the company list itself can be easily created. On the other hand, as shown in Figure 8, the performance factor data lists sets of "company name," "economic event digest," and "impact." It describes how the event digest affected the company. For example, in the same figure, the economic event digest “white goods/demand/increase” shows the impact of “increase in sales” of “Iroha Denki”, while the economic event digest “dollar yen/market price/decrease” , respectively. For such performance factor data, the method described in Japanese Patent Application Laid-Open No. 2020-24689, which has already been proposed by the present applicant, is assumed, so please refer to it if necessary. 7 are extracted from the company performance factor data shown in FIG. 8, and the performance factor list shown in FIG. 6 is created.

The economic phenomenon digest extracted by the narrowing processing unit 4 is newly added to the economic phenomenon database 5 . The economic event database 5 stores not only the economic event digests to be added this time, but also past economic event digests extracted before that time.

The index estimator 6 estimates an economic index with a desired time resolution based on the economic index value described above and the economic event digest stored in the economic event database 5 . FIG. 9 is a block diagram of the index estimator 6. As shown in FIG. The index estimation unit 6 mainly includes a frequency totalization unit 6a, a vector estimation unit 6b, a vector totalization unit 6c, a learning processing unit 6d, an estimation processing unit 6e, and a regression model 6f. The frequency counting unit 6a and the vector estimating unit 6b perform preprocessing with the group of economic event digests read from the economic event database 5 as input. The vector counting unit 6c and the learning processing unit 6d learn the regression model 6f and associate the economic event digest with the economic index value. The vector counting unit 6c and the estimation processing unit 6e estimate the economic index with a desired time resolution using the learned regression model 6f, and output the estimated value of the economic index as the estimation result.

The frequency aggregating unit 6a classifies economic event digests to be aggregated according to economic event patterns having commonalities in content for each predetermined aggregation time unit, with a group of economic event digests as aggregation targets. Then, the frequency totaling unit 6a totals the frequency of appearance of each of the classified economic event patterns for each total time unit. The aggregate time unit is daily in the present embodiment, but is not limited to this, and can be arbitrarily set such as weekly or monthly, and this aggregate time unit corresponds to the highest time resolution in index estimation. .

FIG. 10 shows an example of the result of summing up the frequency of occurrence of economic event patterns on a daily basis. As described above, an economic event pattern is defined as a set of "item x element x predicate", and those that have this set in common are counted as the same economic event pattern. As a result, on December 14, 2018 (daily), economic event pattern A (crude oil x price x decline) occurred three times, and economic event pattern B (automobiles x sales x increase) occurred twice. , economic event pattern C (mobile phones x exports x solid) appears once, economic event pattern D (dollar x price x decline) appears once, and economic event pattern E (housing x demand x strong) appears once. The number of occurrences is one, and the number of occurrences of economic phenomenon pattern F (gasoline x demand x increase) is one. The frequency of occurrence of each of the economic event patterns A to F is the economic index value corresponding to this in time (for example, the economic index value on December 14, 2018, or the economic index value for a predetermined period including the same day). There is a correlation, and the higher the frequency of appearance, the greater the degree of impact on this economic indicator value.

The vector estimating unit 6b is configured so that the economic event frequency vector mapped to the common space for each economic event pattern for each aggregated time unit reproduces the appearance frequency of each economic event pattern for each aggregated time unit. estimated to . The purpose of estimating economic event frequency vectors is to convert discrete data such as news into continuous features, such as "automobiles" and "oil", or "sales" and "demand". is not mapped to discrete symbols, but to a common space of r (r≧2) dimensions and expressed on the same scale. This makes it possible to measure the difference and similarity in meaning between objects. In this regard, discrete symbols, such as expressions such as “automobile”=id1 and “oil”=id2, cannot be used for mathematical processing such as four arithmetic operations (the processing of the vector counting unit 6c, which will be described later, cannot be performed). .).

The estimation of the economic event frequency vector is based on the assumptions shown in Equation 1 below. Here, the symbol "~" means that the left side follows the probability distribution of the right side. xdijk is the frequency of occurrence of a particular economic event pattern in a particular aggregated time unit (date). In addition, θd, θi, θj, and θk are r-dimensional latent vectors, and represent continuous numerical values of objects such as date (aggregation time unit), name (item), element (element), and change (predicate). is a parameter with Note that θ satisfies non-negative (not negative).

Specifically, the economic event frequency vector is estimated based on Equation 2 below. Here, xdijk is the appearance frequency (number of observations) of the economic event pattern ijk on date d, and θ _* is the parameter of the economic event frequency vector (estimation target). r is the dimension number of the economic event frequency vector, and the number of dimensions is set as appropriate. In addition, d is the subscript corresponding to the date, m _d is the subscript of the month corresponding to the date d, i is the subscript corresponding to the name of the economic event pattern (item), and j corresponds to the element of the economic event pattern. The subscript, k, is a subscript that corresponds to the predicate of the economic event pattern. In Equation 2, it is assumed that the occurrence frequency of the economic event pattern within the aggregation time unit follows the Poisson distribution. Note that θ can take the form of a function such as e ^θ or a nonlinear function such as a neural network, as long as θ satisfies nonnegativeness.

The parameters .theta.d, .theta.i, .theta.j, and .theta.k in Equation 2 above are the results of mapping discrete symbols onto a common space. If these parameters θd, θi, θj, θk can be estimated (learned) so that the frequency of occurrence xdijk for each economic event pattern can be well reproduced, what kind of economic event (economic In other words, it is possible to express the degree of occurrence of each economic event pattern.

As an estimation result based on Equations 1 and 2 above, the economic event frequency vector is composed of non-negative real values. 11 to 14 are diagrams showing an example of an economic event frequency vector expressed by decomposing into four parameters θ _* (results of θ _* when r=10). Since each dimension of the economic event frequency vector θ _* has a common meaning among objects, for example, different names (items) can be compared on the same scale as non-negative real numbers. In addition, when the real values of estimated parameters between names (items) tend to be similar, economic events including these tend to be observed at the same time.

The vector estimator 6b also normalizes the economic event frequency vector so that the sum of the constituent elements (non-negative real numbers) of each dimension of the economic event frequency vector becomes one. This normalization allows economic event frequency vectors to be evaluated on the same numerical basis. Specifically, normalization is performed by applying Equation 3 below to the economic event frequency vector estimated for each economic event pattern appearing on a specific date. If this normalization is performed, the value of θ _* takes a non-negative real value, so the sum of the economic event frequency vectors Λ _dijkr normalized for each economic event pattern is always 1. Since this result can be interpreted as the potential appearance rate of the economic event pattern, it is possible to add the appearance frequency for each economic event pattern.

The economic event frequency vector generated by the vector estimator 6b is stored in a storage device (not shown). The economic event frequency vectors stored in this storage device are read by the vector counting unit 6c whenever the economic event frequency vectors are counted.

The vector totalization unit 6c totalizes the economic event frequency vectors belonging to each predetermined time unit to generate an economic event total vector. The time unit here is a time unit (monthly in this embodiment) corresponding to the time resolution of the economic index value during learning of the regression model 6f. Also, when estimating the index using the regression model 6f, in a time unit different from the time resolution of the economic index value, typically in a time unit with a higher time resolution than the economic index value (daily in this embodiment) be. The economic event aggregate vector represents the degree of occurrence of each economic event pattern within the time unit. That is, during the learning, an economic event aggregation vector (first economic event aggregation vector) representing the degree of occurrence of each economic event pattern in one month is generated as a result of the monthly aggregation. Further, during the index prediction, as a result of the daily aggregation, an economic event aggregation vector (second economic event aggregation vector) representing the degree of appearance of each economic event pattern in one day is generated.

Specifically, the economic event aggregation vector is calculated by Equation 4 below. Using the fact that the sum of Λdijkr is always 1 in the above formula 3, xdijk is the frequency of appearance of the economic event pattern ijk on day d, so xdijkΛdijkr is the name (item) i of date d and the element (element) By taking the sum of all j and normalizing by the frequency of occurrence in the time unit, the economic event aggregate vector for d days can be calculated.

This process is done separately for each predicate k. It should be noted that instead of fluctuation (predicate), it may be performed for each "impact" in the performance factor data of the company shown in FIG. The advantage of using this "influence" is, firstly, that the trend of the value of the regression coefficient can be expressed well, and secondly, the result of the regression coefficient is always positive if the revenue increases, and negative if the revenue decreases. can be constrained. In addition, when taking the sum, it is possible to take the sum using an arbitrary function, or to take the sum assuming that the past economic event aggregate vector affects the current economic event aggregate vector. By performing this process, it becomes possible to interpret which position in the r dimension the event belonging to on the date d is likely to appear.

FIG. 15 is a diagram showing an example of a daily unit economic event aggregation vector. The numerical values shown in the figure represent the characteristics of economic events in each dimension, and represent what kind of economic events are likely to occur on any given day d. In other words, as the degree of appearance of each economic event pattern on the day d, dimension 1 shows a strong characteristic of oil-related events, and dimension 2 shows a tendency of strong characteristics of automobile-related events. This does not necessarily mean that it will happen in terms of calculation, but that it may happen as a result of human interpretation of the calculation results of the computer.

On the other hand, when calculating the economic event aggregate vector for each month, the economic event frequency vector for each economic event pattern belonging to that month should be aggregated for each monthly unit according to Equation 4 above. The reason why the event aggregation vector for learning of the regression model 6f is monthly is to match the time resolution of the economic index values. Therefore, if the time resolution of the economic index value is quarterly, the event aggregation vector for learning is also generated quarterly.

The regression model 6f is a model for estimating the relationship between two variables by a statistical method, and is used for estimating economic indicators with arbitrary time resolution. Any model including ridge regression, Lasso, Gaussian process regression, XGboost, neural network, support vector machine (SVM) and the like can be used as the regression model 6f.

The learning processing unit 6d performs regression so that the economic index value temporally corresponding to the input of the monthly unit economic event aggregate vector (monthly aggregate) generated by the vector aggregate unit 6c responds. The model 6f is trained. FIG. 16 is a diagram showing an example of the relationship between input variables and response variables in the regression model 6f. For example, the learning of the regression model 6f is performed so that the economic index value (57.7) in the same month of January 2014 responds to the input of the economic event aggregation vector in January of the same year.

The estimation processing unit 6e uses the daily economic event aggregation vector (daily aggregation) generated by the vector aggregation unit 6c to estimate an economic indicator value temporally corresponding thereto. Specifically, the economic event aggregate vector is input to the learned regression model 6f, and the response of the regression model 6f to this input is output as the estimated value of the economic index. By performing standardized calculations for the daily and monthly aggregate vectors of economic events in the same way, the sum of the r dimensions can be constrained to 1. Even if the economic index is estimated, it is possible to maintain the estimation accuracy. Estimates of economic indicators can be output regardless of whether the time resolution is higher or lower than the daily unit, and can be output not only for the current time but also for the past. .

In addition, the estimation processing unit 6e performs processing such as smoothing and seasonal adjustment on the estimated values of the economic indicators, which are the responses of the regression model 6f, as necessary. FIG. 17 is a diagram showing an example of estimated economic index values smoothed by a Kalman filter. While the original economic index values shown in Figure 2 are monthly units, the estimated economic index values shown in the figure are daily units, and the time resolution is higher than the original economic index values. , and is changing smoothly.

As described above, according to the present embodiment, the regression model 6f is learned so that the economic index value temporally corresponding to the input of the economic event aggregation vector responds. The economic event aggregate vector represents the degree of occurrence of each economic event pattern within a time unit such as monthly. As a learning result of the regression model 6f, an economic event and an economic index value are associated in a form that takes into consideration the occurrence frequency of the patterned economic event, in other words, the degree of influence of the economic event on a certain economic index value. By using the regression model 7 constructed in this way, it is possible to accurately estimate the economic index corresponding to this time resolution as a response to the input of the economic event aggregation vector of arbitrary time resolution.

Furthermore, the present invention can be regarded as an economic indicator estimation program that equivalently implements the functional blocks shown in FIGS. This economic index estimation program generally causes a computer to execute the following processes. First, an economic event digest is generated for the economic events extracted from the news group. Next, according to the narrowing-down conditions set in advance, economic event digests having relevance to the economic index value are extracted, and the extracted economic event digests are stored in the economic event database 5 . Next, with a group of economic event digests stored in the economic event database 5 as aggregation targets, the economic event digests are classified by economic event pattern for each predetermined aggregation time unit, and the appearance frequency of each economic event pattern is aggregated. do. Next, for each economic event pattern for each aggregated time unit, the economic event frequency vector mapped to the common space is estimated so as to reproduce the appearance frequency of each economic event pattern for each aggregated time unit. . Next, for each monthly unit, economic event frequency vectors belonging to the monthly unit are aggregated to generate an economic event aggregate vector (monthly). Next, the regression model 6f is trained so that the economic index value (monthly) corresponding to the input of the economic event summary vector (monthly) responds. Next, the economic event frequency vectors belonging to the daily unit are aggregated to generate an economic event aggregation vector (daily). Finally, output the response of the trained regression model 6f to the input of the aggregated economic event vector (daily) as the estimated value (daily) of the economic index temporally corresponding to the aggregated economic event vector (daily). .

1 economic index estimation system 2 news filter unit 3 digest generation unit 4 narrowing processing unit 5 economic event database 6 index estimation unit 6a frequency aggregation unit 6b vector estimation unit 6c vector aggregation unit 6d learning processing unit 6e estimation processing unit 6f regression model

Claims (16)

An economic indicator estimation system for estimating an economic indicator with a desired time resolution based on a discrete time series of economic indicator values representing a specific economic indicator and economic events that affect the economic indicator values,
a digest generation unit that generates an economic event digest in which the content of the economic event extracted from the news group collected from the outside is structured with a plurality of predetermined items;
an economic event database that stores the economic event digest;
The group of economic event digests stored in the economic event database is subject to aggregation, and the economic event digests are classified into economic event patterns having content commonality for each predetermined aggregation time unit, and each economic event is classified into a frequency counting unit that counts the frequency of occurrence of patterns;
For each of the economic event patterns for each aggregated time unit, an economic event frequency vector composed of non-negative real values mapped to a common space reproduces the appearance frequency of each of the economic event patterns for each aggregated time unit. a vector estimator that estimates such that
For each first time unit corresponding to the time resolution of the economic index value, an economic event frequency vector belonging to the first time unit is aggregated, and the degree of occurrence of each economic event pattern within the first time unit. a vector aggregation unit that generates a first economic event aggregation vector representing
a learning processing unit that learns a regression model for estimating the economic index so that the economic index value temporally corresponding to the input of the first economic event aggregate vector responds to the input An economic index estimation system characterized by: 2. The method according to claim 1, further comprising a news filter unit that extracts media names described in a preset news media list from among the news groups and outputs the extracted media names to the digest generation unit. economic indicator estimation system. The feature further comprises a narrowing processing unit that extracts, from the economic event digests generated by the digest generating unit, those that match preset narrowing conditions and stores them in the economic event database. The economic index estimation system according to claim 1. The narrowing-down processing unit extracts the economic event digest by referring to, as the narrowing-down condition, a performance factor list describing patterns of economic event digests that affect specific economic index values to be estimated. The economic index estimation system according to claim 3, characterized by: The vector aggregation unit calculates the sum of the economic event frequency vectors normalized so that the sum of the components of each dimension is 1, and normalizes by the appearance frequency in the first time unit. , to generate the first economic event aggregation vector. further comprising an estimation processing unit that outputs a response of the learned regression model to the input of the second aggregate vector of economic events as an estimated value of an economic indicator temporally corresponding to the second aggregate vector of economic events;
The vector aggregation unit aggregates economic event frequency vectors belonging to a second time unit having a time resolution different from that of the economic index value, and expresses the degree of occurrence of each economic event pattern within the second time unit. 2. The economic indicator estimation system according to claim 1, wherein the second economic event aggregation vector is generated. The vector aggregation unit calculates the sum of the economic event frequency vectors normalized so that the sum of the components of each dimension is 1, and normalizes by the appearance frequency in the second time unit. , to generate the second economic event aggregation vector. 8. The economic index estimation system according to claim 6, wherein said second time unit has a higher time resolution than said economic index value. In an economic indicator estimation program for estimating an economic indicator with a desired time resolution based on a discrete time series of economic indicator values representing a specific economic indicator and economic events that affect the economic indicator values,
a first step of generating an economic event digest in which the content of the economic event extracted from the news group collected from the outside is structured with a plurality of predetermined items;
a second step of storing the economic event digest in an economic event database;
The group of economic event digests stored in the economic event database is subject to aggregation, and the economic event digests are classified into economic event patterns having content commonality for each predetermined aggregation time unit, and each economic event is classified into a third step of aggregating the frequency of occurrence of patterns;
For each of the economic event patterns for each aggregated time unit, an economic event frequency vector composed of non-negative real values mapped to a common space reproduces the appearance frequency of each of the economic event patterns for each aggregated time unit. a fourth step of estimating that
For each first time unit corresponding to the time resolution of the economic index value, an economic event frequency vector belonging to the first time unit is aggregated, and the degree of occurrence of each economic event pattern within the first time unit. a fifth step of generating a first economic event aggregation vector representing
a sixth step of learning a regression model for estimating the economic index so that the economic index value temporally corresponding to the input of the first economic event aggregation vector responds; An economic indicator estimation program characterized by causing a computer to execute a process comprising: 10. The economic index estimation program according to claim 9, wherein said first step has a step of extracting media names described in a preset news media list from said news group. . 10. The method according to claim 9, wherein said first step has a step of extracting economic event digests that meet preset narrowing conditions and storing them in said economic event database. economic indicator estimation program. The first step extracts the economic event digest by referring to, as the narrowing condition, a performance factor list describing patterns of economic event digests that affect specific economic index values to be estimated. The economic index estimation program according to claim 11, characterized by: In the third step, calculating the sum of the economic event frequency vectors normalized so that the sum of the components of each dimension is 1, and normalizing by the appearance frequency in the first time unit. 10. The economic index estimation program according to claim 9, wherein the first economic event aggregation vector is generated by: a seventh step of outputting a response of the learned regression model to the input of the second aggregate vector of economic events as an estimated value of an economic indicator temporally corresponding to the second aggregate vector of economic events;
aggregating economic event frequency vectors belonging to a second time unit having a time resolution different from that of the economic index value, and a second economic event aggregation representing the degree of occurrence of each economic event pattern within the second time unit; 10. The economic indicator estimation program according to claim 9, further comprising an eighth step of generating vectors. In the eighth step, calculating the sum of the economic event frequency vectors normalized so that the sum of the components of each dimension is 1, and normalizing by the appearance frequency in the second time unit. 15. The economic index estimation program according to claim 14, wherein the second economic event aggregation vector is generated by: 16. The economic index estimation program according to claim 14, wherein said second time unit has a higher time resolution than said economic index value.

Images