JP2001175735A - Device, system and method for predicting time series and storage medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To predict future data by fetching reproducibility and future predictability from time series having strong nonlinearity. SOLUTION: A characteristic extraction processing 12 acquires the characteristic of a time series concerned from past time series data 11. Regression tree preparation processing 13 prepares a regression tree 14 on the basis of acquired characteristic quantity. Meanwhile, characteristic extraction processing 16 acquires characteristic quantity from the current time series data 15 by using the same algorithm as that of the processing 12. Future prediction processing 17 calculates a future predicted value 18 by tracing the regression tree 14 by using the characteristic quantity acquired by the processing 16.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a time series prediction apparatus, system, method, and storage medium for predicting a time series having a strong nonlinearity such as a stock market price movement.

[0002]

2. Description of the Related Art In general, a time series means a time series obtained by arranging variables obtained from physical phenomena in chronological order. And the central method of the analysis was a method called linear regression analysis. This method is based on the theory that when predicting future data from past data, the relationship can be approximated by a linear function. In addition, recently, algorithms that interpret economic phenomena as a kind of time series and predict various economic indices have appeared.

[0003] At present, for example, "Economic time series analysis by adaptive model" (Toshio Sugihara, Engineering Book Co., Ltd. ISBN4-76)
92-0369-1 C3058), there has been proposed a method of predicting an economic time series using an algorithm such as a Kalman filter or a neural network. These methods can cope with a certain degree of nonlinearity in the time series, and have demonstrated very good prediction performance for data such as "time series of power demand in Japan".

[0004]

However, according to the conventional technique, the prediction performance is significantly reduced in a time series having a very strong nonlinearity such as a price movement of each stock in a stock market. There was.

[0005] The time series that appears in economic phenomena is that the economic system itself is complicated and rarely has a simple structure. At present, probabilistic analysis is being performed, treating most of them as completely random time series processes.
However, if the seemingly random time series has some reproducibility and it can be extracted, the method is of great value.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and an object thereof is to extract reproducibility and future predictability from a time series having a strong nonlinearity so that future data can be predicted. It is in.

[0007]

Means for Solving the Problems A time series prediction apparatus according to an aspect of the present invention for achieving the above object has, for example, the following configuration. That is, a feature acquiring unit that acquires a feature of the time series from past time-series data, a creating unit that creates a regression tree based on the feature amount acquired by the feature acquiring unit, and an algorithm that is the same as the feature acquiring unit. The present time series feature acquisition means for acquiring a feature quantity from the current time series data using the feature quantity acquired by the current time series feature acquisition means, and the future prediction using the regression tree created by the creation means Prediction means for obtaining a value.

A time series prediction system according to an aspect of the present invention for achieving the above object distributes the above time series prediction apparatus and a prediction value obtained by the time series prediction apparatus to another information processing apparatus. And distribution means for performing the distribution.

[0009]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A preferred embodiment of the present invention will be described below with reference to the accompanying drawings.

FIG. 1 is a diagram best representing a time-series prediction system according to an embodiment. This system has two phases, a learning phase and a prediction phase. In the learning phase, reference numeral 11 denotes past time-series data, 12 denotes a feature extraction process for extracting a feature amount from the past time-series data 11, and 13 denotes a regression tree using the feature amount obtained from the feature extraction process 12. The regression tree creation process 14 is the “regression tree” obtained in the regression tree creation process 13. The input in the learning phase is the past time-series data 11, and the output is the regression tree 14.

In the prediction phase, 15 is current time series data, 16 is current time series feature extraction processing for calculating feature values from the current time series data 15, and 17 is current feature value extraction processing 16 and regression tree 14. The future prediction processes 17 and 18 for predicting a future time-series value by using these are prediction values obtained from the future prediction process 17. The input in the prediction phase is the current time-series data 15, and the output is the predicted value 1
It becomes 8.

What is important in FIG. 1 is that the time series data of the past week is generally very large. Therefore,
When the feature amount obtained from the feature extraction process 12 is input to the regression tree creation process 13, a very large number of loops are performed.

The feature extraction process 12 and the current time series feature extraction process 16 use exactly the same algorithm. This algorithm will be described in detail with reference to FIG.

The graph depicted at the top of FIG. 2 is a so-called "candlestick" graph of stock prices. In the figure, the history of the stock price fluctuation for 20 days is drawn as a graph. The graph includes information on the opening price, closing price, high price, and low price. The time series prediction that this embodiment is trying to perform is:
For example, given the "candlestick" graph at the top of FIG. 2, it will predict what the stock price will be tomorrow. Hereinafter, the present invention will be described using a system for predicting stock price fluctuations as an embodiment. Of course, the present invention can be applied not only to daily data but also to time-series data having a large time interval between weekly and annual data.

In the present embodiment, for example, the opening price, closing price, high price, and low price of a certain stock in a certain year from the first day to the twentieth day are described in a first time-series data set,
The time series data set is determined such that the opening price, the closing price, the high price, and the low price of the day (from the second day to the 21st day) are referred to as a second time series data set. Then, all the time series data sets of a plurality of brands are used as the same time series model by normalizing each time series data set.

First, the normalization will be described. There are many stocks in the stock market, and their values vary. However, if the regression tree according to the present embodiment is created for each stock price, learning data is chronically insufficient. Therefore, all brands are standardized, and learning proceeds on the premise that they are the same time-series model.

For normalization, each price of the time-series data set is converted into a relative price with the closing price of the last day of the time-series data set to be standardized (the target graph) as a reference (1.0). That is, for each time-series data set, the opening price, the closing price, the high price, and the low price for the past 20 days are divided by the closing price on the last day of the time-series data set, and converted into relative prices. Then, as the feature amount of the time-series pattern, 20 × 4 = 80 real numerical values are obtained.

In the present embodiment, in the learning phase,
Using the standardized time-series data set, a regression tree is created in a procedure described later. Then, in the prediction phase, a standardized time-series data set (the current time-series data 1
Tracing the regression tree using 5), and predicting the relative price of tomorrow's stock price to today's closing price for the desired stock.

As described above, by standardization, all periods of all brands can be used as learning data, and a sufficiently large amount of learning data can be secured. The above example is described in terms of statistics, "Solving a regression problem when a large set of explanatory variables / one-dimensional explained variables described in an 80-dimensional feature space is given as learning data. The problem of "the problem" can be schematized.

In the above example, the open price, the close price, the high price, and the low price
Although the two values are used, the trading volume, season (date), and the like can also be used as values for calculating the feature amount. Further, in the above example, the closing price of today (= the last day of the targeted graph) is used as the reference value, but any value on any day retroactively from the last day may be used as the reference value. Further, the feature value may be calculated using statistical economic data such as the Nikkei Stock Average or the number of housing starts, or subjective economic data such as the economic coincidence index.

Further, in the above example, the relative price was derived by dividing the original price by the price serving as the reference value, but a non-linear process such as taking a natural logarithm thereof is performed to calculate the feature amount. You may. Further, a new feature amount derived by converting data over a plurality of dates, such as “25 day moving average” often used in stock price chart analysis, may be used.

Next, the regression tree 14 in FIG. 1 will be described with reference to FIG. FIG. 3 is a diagram illustrating an example of a regression tree according to the present embodiment. A regression tree is a prediction system with a tree structure. As shown in FIG. 3, the regression tree is composed of "nodes" called "nodes" and "branches" connecting them.
For example, the regression tree of FIG. 3 has 15 nodes and 14 branches. In particular, the origin node (node 0 in FIG. 3) of the regression tree is called a “root node”. In FIG. 3, although a node having a branch always outputs two branches, this is a regression tree called a so-called "binary tree", and in general, the number of branches may be any.

Nodes connected by branches have a parent-child relationship. For example, "child nodes of the root node are the first and sixth nodes, and the parent of the first and sixth nodes is the root node. " Nodes without child nodes are
Call it a leaf node. A special example of a leaf node is a null node. The definition of the null node will be described later in detail. In addition, a null node cannot exist in a regression tree of a binary tree.

Nodes other than leaf nodes are called "internal nodes." The internal nodes have conditional expressions for branching. For example, in the case of stock data prediction in FIG. If the opening price of the previous day is 0.85 or more of the closing price of the current day, the first node, otherwise, the sixth node ". The condition for branching to the child node is called a “node branch condition”. The learning phase is to create node branch conditions based on the time series data of the last week, and the prediction phase is to classify the current time series data according to the node branch conditions.

In the learning phase, past time-series data itself is classified at the same time as node branch conditions are created. That is, division into internal nodes is performed in units of a normalized time-series data set. Although all past data remains in the root node, as the internal node is created, the conditions become more and more strict and the remaining time-series data set decreases. Then, when there is no past time-series data set in the newly created child node, the node is set as a null node. A leaf node is a node that satisfies a certain condition. When the condition is satisfied, no further child nodes are created. This condition is called "leaf node establishment condition". The conditions for establishing a leaf node include “remaining past data becomes smaller than a certain number” or “predicted value of remaining past data (in the present embodiment,
The variance of the actual stock price of the stock on the day after the last day of each time-series data set becomes smaller than a certain value. "

In the learning phase, past data remains at any node other than the null node. A representative value such as an arithmetic mean, a geometric mean, or a median of the predicted values of the past data remaining in a certain node is set as the predicted value of the node. In particular, the predicted value associated with the root node is a representative value (average) of the predicted values of all past time-series data. By setting the predicted value associated with the null node as the predicted value associated with the parent node, the predicted value is allocated to all nodes. For example, in the regression tree of FIG. 3, a number written on the right side of a node indicates a predicted value associated with this node.

In the prediction phase, nodes are traced from the root node based on each node branch condition. At this time, the predicted value associated with the traced node becomes the predicted value of the current time-series data.
For example, when the current time-series data is traced using the regression tree of FIG. 3, if the result is 0, 1, 3, 4, the predicted value is 1.2 / 1.5 / 1.3 / 1. .1. As a result, as the future value of the current time-series data, the predicted value of the traced hierarchy is obtained.

In general, the deeper the hierarchy of the regression tree, the higher the accuracy of this predicted value. Therefore, in the above example, 1.1, which is the predicted value in the deepest hierarchy, may be used as the predicted value for the time-series data. However, beyond a certain depth in the hierarchy, the predicted values may be inaccurate. This phenomenon is a phenomenon called “over-learning” of past time-series data. In order to prevent this, there may be a case where “the predicted value of the nth node from the end is used as the predicted value of the current time-series data”. This processing is a so-called “regression tree cutting”, and is particularly effective when the target time-series pattern includes an essentially random phenomenon.

Next, the regression tree creation processing 13 in FIG.
This will be described with reference to the flowchart of FIG. In the regression tree creation processing 13, nodes of the regression tree are created by a method called a recursive program. For example, in the regression tree of FIG.
Nodes are created in the order of the numbers assigned to the nodes. The “relevant node” (curentnod
There is the concept of e), and the flowchart turns around that node.

First, the root node is set as the node concerned (step S401). Next, the branch condition of the node is determined (step S402). The method for determining the branch condition will be described later in detail. When the branch condition is determined, a loop for creating a child node (step S403)
Divides the past time-series pattern belonging to the node according to the branch condition. For example, if the regression tree is a binary tree, the past time-series pattern remaining at the node is divided into two.

The process from step S404 is a loop process. First, in step S404, the next child node of the node is set as the node. However, at the beginning of the loop, the first child node is the node. The process of step S404 includes the above-described process of dividing the past time-series data. As a result, if no past time-series data remains, that is, if the node is a null node (step S405), a null node is determined (step S406).
And the flow proceeds to loop end determination (step S409).

If the node is not a null node and the condition for establishing a leaf node is satisfied (step S40)
7), a leaf node is determined (step S408), and the process proceeds to loop end determination (step S409). If the node is neither a null node nor a leaf node,
With this node as a parent node, child nodes are created more deeply. That is, the process returns to the determination of the node branch condition of the node (step S402).

In the loop end determination of step S409, the loop ends when the processing of steps S404 to S408 has been executed for all child nodes connected to the parent node of the node. For example, in the case of a binary tree as in the present embodiment, the loop is executed twice. When the loop for creating the child node (steps S404 to S409) ends as described above, it is checked whether or not the parent node exists (step S41).
0). If there is a parent node, the node is set as the node (step S411) and the process returns to step S404 to set the next child node as the node. It should be noted here that since the loop for creating the child node has been completed, the node has moved to the parent node when viewed from the nodes in the “child node creation loop”. That is, in step S410, from the viewpoint of the nodes in the “child node creation loop”, it is determined whether or not there is a parent node of the parent node. Therefore, the node creation proceeds, and when all the child nodes of the root node are created, the parent node of the root node does not exist in the parent node existence determination in step S410, and the regression tree creation in step S412 is completed. The program ends.

Next, the determination of the node branch condition (step S402) in FIG. 4 will be described with reference to FIG. As described above, the node branch condition is, for example, "the left node if the opening price two days before the last is 0.85 or more of the closing price of the last day, otherwise the right node". When the regression tree is a binary tree, in general, this node branch condition may be anything as long as the set of past time-series data remaining at the node is divided into two (in the case of a k-ary tree, , Divided by k). If the past time-series data is described by n-dimensional feature amounts, anything may be used as long as it divides the n-dimensional space. If a node branch condition is to be created using a neural network, the surface that realizes this division is a hypersurface in n dimensions.

However, it is very costly to train each node using a neural network.
Therefore, division is usually performed using a hyperplane. Even if the past thread string data remaining at the node is divided on the hyperplane, the degree of freedom of the normal vector of the hyperplane sharply increases as the dimension of the feature increases. , The learning cost will increase.

Therefore, when the past time-series data is very large and the dimension of the feature is large, it is often the case that the node branch condition (hyperplane) is searched for only the hyperplane perpendicular to the axis of the feature. Done. FIG. 5 shows this state.

Reference numeral 51 denotes the node, and reference numerals 52 and 53 denote respective child nodes. The graph 54 is a graph in which a horizontal axis indicates a certain feature amount and a vertical axis indicates the distribution density of past time-series data remaining at the node. As shown in FIG. 5, good partitioning efficiency when the partition is moved from a smaller feature amount to a larger feature amount is calculated, and a position having the highest partitioning efficiency is defined as a divided surface. The division efficiency described here means that the variance of the predicted value of the parent node is P, the variance of the predicted value of the child node is p1 and p2, and the number of past time-series data remaining in the child node is n1 and n2, respectively. P− {n1 · p1 + n2 · p2} / {n1 + n2}. In the above example, the division efficiency was calculated using the variance of the predicted value of the data remaining in the node. Instead of the variance in ordinary statistics, 1 / nｎ (y-mean) ² , The absolute variance 1 / nｎ | y-median |

As a result of the above processing, the optimum division plane for each feature amount and the division efficiency in the division plane are obtained. Finally, the results of all the feature amounts are totaled, and the division of the feature amount with the highest division efficiency is set as the branch condition of the node.

A method for quickly and stably searching for a branch condition of a node will be described with reference to FIGS. 5, 6, and 7. FIG. In 54 of FIG. 5, when the partition to be divided becomes extremely small or large, the number of time series data remaining in the child node is extremely biased toward one of the child nodes. If there is not much difference in the division efficiency even if the partitioning is performed on the axis of the feature amount, the bias in the number of the child nodes leads to a decrease in the prediction performance. In general, when many random elements are included in past time-series data, a situation in which there is not much difference in the division efficiency at the position of the partition tends to occur. In order to avoid this situation, if the partitioning method is restricted so as to partition around the center of the distribution of the past time-series data remaining at the node as shown at 61 in FIG. become able to. The vicinity of the center can be obtained based on values such as an average value, a median value, and an intermediate value between the maximum value and the minimum value. Furthermore, this method further narrows the search range of the partition and saves the search time for the node branch condition.

The search for the node branch condition described above is based on the distribution density function of past time-series data remaining at the node, such as 54 in FIG. That is, this distribution function is a function unique to the node. Therefore, the center of the distribution as shown in FIG. 6 also differs strictly if the nodes are different. However, even if the partition of each feature is fixed in advance based on the distribution at the root node, the prediction performance is often not significantly affected. As shown in FIG. 7, if “hierarchical fixed partitioning” is performed in advance based on the distribution density function for the feature amount of the past time series data (= all) existing at the root node, the feature amount becomes a continuous value. What is present is the same value as the quantized value.
This not only saves memory, but also significantly reduces the search time for node branch conditions.

The time series prediction system using a binary tree regression tree described above can be applied to a time series prediction system using a quadtree tree or a general n-ary tree regression tree. .

Next, a hardware configuration for realizing the time-series prediction system of the present embodiment will be described.

FIG. 8 is a block diagram showing an example of a hardware configuration of the time-series prediction system according to this embodiment. Reference numeral 81 denotes a line, such as a telephone line, a star line, or a dedicated line. The current time-series data is input from the line 81 to the prediction system. 82 is CP
U, which executes the specific prediction system program described in the embodiment. A memory 83 stores current time-series data and secures a temporary work area used by the program. 8
A hard disk (HDD) 4 stores a control program, a regression tree used in the time-series prediction system, and the like. Finally, reference numeral 85 denotes a display on which various displays including a display of a prediction result and the like are executed.

The configuration of the time-series prediction system shown in FIG. 8 realizes the time-series prediction processing of this embodiment by a so-called “stand-alone” application program. In contrast, FIG. 9 shows a configuration of a time-series prediction system that receives prediction results from a server and displays the results on a client. The contents of the individual clients 92, 93, 94 in FIG. 9 are the same as in FIG. 8, except that the client machines 92, 93, 94 in FIG. 9 do not include the information of the regression tree. That is, the server 91 includes a regression tree.

FIG. 10 shows an example of a screen displayed on the display of FIG. 8 and the display of the client machine of FIG. If the time series to be predicted is a change in the stock price, the time series prediction processing described in the embodiment can display, for example, TOP3 of the expected stock price increase tomorrow and TOP3 of the expected stock price decrease tomorrow.

When the time-series prediction system according to the present embodiment is realized on the client-server system shown in FIG.
It is possible to use a public network such as the Web.
In this case, the data from the server 91 to the clients 92 to 94 becomes very valuable. On the other hand, there is a system in which a user operating a client pays the price. For example, as shown in FIG. 11, by inputting a user ID and a password, money is debited from an account designated in advance by the user. In addition, since the public network is used, SS
It is common practice to securely transfer a user ID and a password by using an encryption technique such as L. By doing so, even if the time-series prediction system of the present embodiment is implemented in a client-server environment using a public line network such as the Web, it is possible to securely charge for services.

As described above, according to the above embodiment, time series data with high randomness such as stock prices
You will be able to predict reproducibility and predict future changes in time series data.

Also, by adding a constraint to the derivation of the branch condition of the regression tree by the method described with reference to FIGS. 6 and 7, there is also an effect that the memory can be saved and the regression tree can be created at high speed. .

An object of the present invention is to supply a storage medium (or a recording medium) on which a program code of software for realizing the functions of the above-described embodiments is recorded to a system or an apparatus, and to provide a computer (a computer) of the system or the apparatus. It is needless to say that the present invention can also be achieved by a CPU or an MPU) reading and executing the program code stored in the storage medium. In this case, the program code itself read from the storage medium implements the functions of the above-described embodiment, and the storage medium storing the program code constitutes the present invention. Also,
When the computer executes the readout program code, not only the functions of the above-described embodiments are realized, but also the operating system (OS) running on the computer based on the instructions of the program code.
It goes without saying that a case where the functions of the above-described embodiments are implemented by performing some or all of the actual processing, and the processing performs the functions of the above-described embodiments.

Further, after the program code read from the storage medium is written into a memory provided in a function expansion card inserted into the computer or a function expansion unit connected to the computer, the program code is read based on the instruction of the program code. Needless to say, the CPU included in the function expansion card or the function expansion unit performs part or all of the actual processing, and the processing realizes the functions of the above-described embodiments.

[0051]

As described above, according to the present invention,
It is possible to extract reproducibility and future predictability from a time series with strong nonlinearity and predict future data.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating a time-series prediction process according to an embodiment.

FIG. 2 is a diagram showing an example of a time series to which the time series prediction system of the present embodiment is applied, and a feature extraction pattern from the time series.

FIG. 3 is a diagram illustrating an example of a regression tree created by the time-series prediction system according to the embodiment.

FIG. 4 is a flowchart showing a regression tree creation process of the time series prediction system of the present invention.

FIG. 5 is a diagram illustrating a method of searching for a node branch condition in a regression tree creation process.

FIG. 6 is a diagram illustrating restrictions imposed for speeding up processing in a node branch condition search in regression tree creation processing.

FIG. 7 is a diagram illustrating quantization of a feature amount in a node branch condition search in a regression tree creation process.

FIG. 8 is a block diagram showing a device configuration for realizing the time-series prediction processing of the present embodiment as a single application program.

FIG. 9 is a block diagram showing a configuration for realizing the time-series prediction processing of the embodiment in a client-server system.

FIG. 10 is a diagram showing an example of a result display screen by the time-series prediction system of the embodiment.

FIG. 11 is a diagram showing an example of a screen used for billing when the present time series prediction system is implemented in a client server system.

[Explanation of symbols]

 11 Past time series data 12 Feature extraction processing 13 Regression tree creation processing 14 Regression tree 15 Current time series data 16 Current time series feature extraction processing 17 Future prediction processing 18 Predicted value

Claims (20)

[Claims] A feature acquisition unit configured to acquire a feature of the time series from past time series data; a creation unit configured to create a regression tree based on the feature amount acquired by the feature acquisition unit; A current time series feature acquisition unit that acquires a feature amount from current time series data using the same algorithm as above, a feature amount acquired by the current time series feature acquisition unit, and a regression tree created by the creation unit. A time-series prediction device, comprising: a prediction unit that obtains a future predicted value. 2. The method according to claim 1, wherein the creating unit searches for a division position of a feature amount for determining branching of the regression tree from a plurality of division positions based on each division efficiency. Time series prediction device. 3. The method according to claim 1, wherein the generating unit sets a search range of a feature value division position for determining a branch of the regression tree to a vicinity of an average value or a median value of the feature value distribution, or an intermediate value between a maximum value and a minimum value. 3. The time-series prediction device according to claim 2, wherein the time-series prediction device is configured to restrict the time-series to near. 4. The time-series prediction device according to claim 2, wherein the creation unit searches for the division position from a value obtained by quantizing the feature amount. 5. When tracing a regression tree created by the creation means using the feature quantity acquired by the current time series feature acquisition means, the prediction means stops the prediction value in the middle of the regression tree. The time-series prediction device according to claim 1, wherein the time-series prediction device obtains the time-series prediction data. 6. A time-series prediction device according to claim 1, further comprising: a distribution unit that distributes a prediction value obtained by the time-series prediction device to another information processing device. Time series forecasting system. 7. The distribution means according to claim 5, wherein the distribution means
The time series prediction system according to claim 6, wherein the time series prediction system is formed using an eb system. 8. The method according to claim 6, wherein the distribution unit is formed using a client server system in which the time-series prediction device is a server and the other information processing device is a client. Series prediction system. 9. The time-series prediction system according to claim 6, further comprising a charging unit configured to charge when said distribution value is distributed by said distribution unit based on a contract with a user. . 10. A feature acquiring step of acquiring a feature of the time series from past time series data; a creating step of creating a regression tree based on the feature amount acquired in the feature acquiring step; Using the current time series feature acquisition step of acquiring the feature quantity from the current time series data using the same algorithm as the feature quantity acquired in the current time series feature acquisition step, and the regression tree created in the creation step A prediction step of obtaining a predicted value of the future. 11. The method according to claim 10, wherein, in the creating step, a division position of a feature amount for determining a branch of the regression tree is searched from a plurality of division positions based on each division efficiency. Time series prediction method. 12. The method according to claim 1, wherein a search range of a feature amount division position for determining branching of a regression tree is set near an average value or a median value of the feature value distribution, or an intermediate value between a maximum value and a minimum value. 12. The time-series prediction method according to claim 11, wherein the time-series prediction method is limited to the vicinity of. 13. The time-series prediction method according to claim 11, wherein in the creating step, the division position is searched from a value obtained by quantizing the feature amount. 14. The predicting step, when tracing a regression tree created in the creation step using the feature amount acquired in the current time series feature acquisition step, stops the regression tree in the middle of the regression tree to calculate a prediction value. The time-series prediction method according to claim 10, wherein the time-series prediction method is acquired. 15. A time-series prediction step for executing the time-series prediction method according to claim 10, and a distribution for distributing a prediction value obtained in the time-series prediction step to another information processing apparatus. And a control method for the time-series prediction system. 16. The control method for a time-series prediction system according to claim 15, wherein the distribution step distributes the predicted value using a Web system on the Internet. 17. The method according to claim 15, wherein, in the distribution step, the predicted value is distributed using a client server system in which the time-series prediction method is a server and the other information processing apparatus is a client. The control method of the time-series prediction system described in the above. 18. The time-series prediction system according to claim 15, further comprising a charging step of charging when said predicted value is distributed by said distribution step based on a contract with a user. . 19. A storage medium storing a program for realizing the time-series prediction method according to claim 10. Description: 20. A storage medium for storing a control program for realizing a time series prediction process by a computer, the control program comprising: a feature acquisition step of acquiring a feature of the time series from past time series data. A code of a creation step of creating a regression tree based on the feature amount acquired in the feature acquisition step, and a current time series of acquiring a feature amount from current time series data using the same algorithm as the feature acquisition step A code for a feature acquisition step, a code for a prediction step for determining a future prediction value using the feature amount acquired in the current time series feature acquisition step, and the regression tree created in the creation step. Storage medium.