CN107316093B

CN107316093B - Rolling prediction method and device

Info

Publication number: CN107316093B
Application number: CN201610266574.2A
Authority: CN
Inventors: 陈新杰; 赵志洪; 胡楠; 张观侣
Original assignee: Huawei Technologies Co Ltd
Current assignee: Huawei Technologies Co Ltd
Priority date: 2016-04-26
Filing date: 2016-04-26
Publication date: 2021-01-05
Anticipated expiration: 2036-04-26
Also published as: CN107316093A

Abstract

The embodiment of the invention discloses a method and a device for rolling prediction, wherein the method comprises the following steps: determining a lag phase of a hysteresis characteristic of each sequence in the reference sequence; the reference sequence comprises at least two sequences, and the reference sequence comprises an external sequence and/or a self sequence; reordering the hysteresis characteristics of the sequences in the reference sequence according to the hysteresis periods of the hysteresis characteristics of the sequences in the reference sequence, and combining the reordered hysteresis characteristics belonging to the same period to obtain a feature combination to be selected; selecting a feature combination from the feature combinations to be selected to carry out rolling prediction on the verification set with known results, and selecting the feature combination with a rolling prediction result meeting a preset condition as a target feature combination; and performing rolling prediction on the self sequence by using the target feature combination. The influence of the sequence/external sequence on the rolling prediction result is comprehensively considered, and the rolling prediction result is more accurate and stable.

Description

Rolling prediction method and device

Technical Field

The present invention relates to the field of computer technologies, and in particular, to a method and an apparatus for rolling prediction.

Background

Enterprises often have some indexes for measuring self-operation conditions, such as: revenue of the main business, profit, etc. The indexes are accumulated over time to form data with a temporally dependent sequence, the data belong to sequence data, and the temporally dependent sequence of the data is called a time sequence. For the operation condition of an enterprise, people not only care about the current indexes, but also care about possible values of the indexes in the future, and make a strategy preparation in advance to avoid possible risks in the future or guide the decision of production and operation.

In general, in addition to the above-mentioned index in a future period, it is desirable to know the value of the index in a longer time in the future, such as: monthly income value, which the business manager cares about in the future of 1 month, and also wants to know about in the future of 12 months. Therefore, the rolling prediction of the time series to the front for a plurality of periods is widely applied to enterprises.

The methods studied in the academia usually use the regularity of their own sequences for prediction. The self sequence of a certain index A refers to a sequence formed by historical data of the index A. The external sequence is a sequence composed of historical data of other indexes besides A. The self sequence may be generally referred to as a target sequence, and the other sequences may be referred to as external sequences with reference to the target sequence.

With the emphasis of enterprises on internal data, more and more external sequences are collected and used for rolling prediction of target sequences, and the external sequences have certain lag correlation with the data of the target sequences; for example, the shipping data 3 months ago can affect the income of the current month, and the 3 months are called lag phase; therefore, the accuracy of prediction can be improved by selecting a proper lag external sequence.

The following is an implementation scheme for performing rolling prediction by combining an application external sequence and a self sequence, and specifically includes:

firstly, determining a lag period:

calculating the correlation between a reference sequence (comprising a target sequence and an external sequence) and the target sequence, selecting a lag period t with the highest correlation, determining whether the correlation is significant through a correlation test, and determining that the correlation has a hysteresis influence and the lag period is t if the correlation is significant. For the lag phase of the target sequence, an autocorrelation function can be used to obtain autocorrelation values of different periods, whether the correlation is significant is confirmed through correlation check, if the correlation is significant, the corresponding period is determined to be the lag phase, otherwise, no lag phase exists, that is: the lag phase is 0.

Secondly, establishing a model:

a suitable model is built based on the lag phase between the reference sequence and the target sequence, for example: a multiple linear regression model.

And thirdly, rolling prediction:

when the required external features are missing, the feature values of the external sequences can be predicted first, and then the feature values can be input for rolling prediction.

By adopting the scheme to carry out rolling prediction, the accuracy is low and the stability is low.

Disclosure of Invention

The embodiment of the invention provides a method and a device for rolling prediction, which are used for improving the accuracy and the stability of the rolling prediction.

In one aspect, an embodiment of the present invention provides a rolling prediction method, including:

determining a lag phase of a hysteresis characteristic of each sequence in the reference sequence; the reference sequence comprises at least two sequences, and the reference sequence comprises an external sequence and/or a self sequence;

reordering the hysteresis characteristics of the sequences in the reference sequence according to the hysteresis periods of the hysteresis characteristics of the sequences in the reference sequence, and combining the reordered hysteresis characteristics belonging to the same period to obtain a feature combination to be selected;

selecting a feature combination from the feature combinations to be selected to carry out rolling prediction on the verification set with known results, and selecting the feature combination with a rolling prediction result meeting a preset condition as a target feature combination;

and performing rolling prediction on the self sequence by using the target feature combination.

The reference sequence may be only the external sequence or only the self sequence, preferably both the external and self sequences are used. The verification set is a sequence with known results, and the characteristics to be selected are used for predicting the sequence, so that the prediction effect of the characteristics can be known to be better, and the other characteristics are relatively poorer; the predetermined condition is satisfied here in order to select a feature combination with a better prediction effect.

In a possible implementation manner, the selecting a feature combination from the feature combinations to be selected includes:

selecting a feature combination from the feature combinations to be selected in a traversal mode; or selecting a feature combination from the feature combinations to be selected in a random mode; or selecting a feature combination from the feature combinations to be selected in a traversal or random mode, wherein the selected feature combination comprises at least one hysteresis feature of an external sequence and at least one hysteresis feature of a self sequence.

In the above three schemes of selecting feature combinations, the first one of the schemes has a large calculation amount, but has the advantage of being comprehensive, so that the scheme can be used as a preferred implementation scheme under the conditions of strong calculation capacity or more calculation resources; the second calculation amount is relatively small, so that the calculation efficiency can be improved; a third advantage is that the influence of the external sequence and the self sequence on the rolling prediction result are considered at the same time.

In one possible implementation, the target feature combination includes at least two feature combinations meeting a predetermined condition; before using the target feature combination to perform rolling prediction on the self-sequence, the method further comprises:

selecting a characteristic combination from the characteristic combinations meeting the preset conditions to obtain a target combination to be determined;

and performing rolling prediction on the verification set with the known result by using the target combination to be determined, and taking the target combination to be determined as the target feature combination when a preset stopping condition is met.

The present embodiment further combines feature combinations, which helps to further improve the accuracy and stability of the prediction. In addition, a certain screening of the feature combinations is also possible.

In one possible implementation, the predetermined stop condition includes:

adding a new selected feature combination into the target combination to be determined to roll the prediction result and not to be promoted any more;

or selecting the times of the feature combination to reach the preset times;

or, the times of selecting the same feature combination reach the specified times under the condition of improving the prediction result.

In the three stopping conditions, the first one takes the optimal prediction result as a target, the second one is simplest to control, and the smaller preset times can prevent the occurrence of overfitting; the third control is simpler, when the same feature combination is extracted for multiple times, the selection is stopped, and the prediction effect cannot be improved by other feature combinations; the specific number of "multiple times" here may be preset, such as: 2. 5 or other value.

In a possible implementation manner, the target feature combination includes at least two feature combinations, and the performing rolling prediction on the self-sequence by using the target feature combination includes:

and respectively performing rolling prediction on the self sequence by using each feature combination in the target feature combination, and then calculating a weighted average value of rolling prediction results.

In another aspect, an embodiment of the present invention provides a rolling prediction apparatus, including:

the lag phase determining unit is used for determining the lag phase of the lag characteristic of each sequence in the reference sequence; the reference sequence comprises at least two sequences, and the reference sequence comprises an external sequence and/or a self sequence;

a reordering unit, configured to reorder the hysteresis characteristics of each sequence in the reference sequence according to the hysteresis period of the hysteresis characteristics of each sequence in the reference sequence;

the combination unit is used for combining the reordered hysteretic characteristics belonging to the same period to obtain a characteristic combination to be selected;

the characteristic selection unit is used for selecting a characteristic combination from the characteristic combinations to be selected to carry out rolling prediction on the verification set with the known result, and selecting the characteristic combination with the rolling prediction result meeting the preset condition as a target characteristic combination;

and the rolling prediction unit is used for performing rolling prediction on the self sequence by using the target feature combination.

In a possible implementation manner, the feature selection unit is configured to select a feature combination from the feature combinations to be selected in a traversal manner; or selecting a feature combination from the feature combinations to be selected in a random mode; or selecting a feature combination from the feature combinations to be selected in a traversal or random mode, wherein the selected feature combination comprises at least one hysteresis feature of an external sequence and at least one hysteresis feature of a self sequence.

In one possible implementation, the target feature combination includes at least two feature combinations meeting a predetermined condition; the device further comprises:

the combination selection unit is used for selecting a feature combination from the feature combinations meeting the preset conditions to obtain a target combination to be determined before the rolling prediction unit uses the target feature combination to carry out rolling prediction on the self sequence;

the rolling prediction unit is further used for performing rolling prediction on the verification set of the known results by using the target combination to be determined;

and the characteristic selection unit is used for taking the target combination to be determined as the target characteristic combination when a preset stop condition is met in the process of performing rolling prediction by the rolling prediction unit.

In one possible implementation, the predetermined stop condition includes:

or selecting the times of the feature combination to reach the preset times;

or, when the newly selected feature combination is added to the target combination prediction result to be determined and the prediction result is improved, the number of times of selecting the same feature combination reaches the specified number of times.

In a possible implementation manner, the target feature combination includes at least two feature combinations;

and the rolling prediction unit is used for performing rolling prediction on the self sequence by using each feature combination in the target feature combination and then calculating a weighted average value of rolling prediction results.

According to the technical scheme, the embodiment of the invention has the following advantages: at least two sequences are used as a reference and comprise a plurality of hysteresis characteristics, different hysteresis characteristics correspond to hysteresis periods of the hysteresis characteristics, and the influence of self sequences/external sequences on a rolling prediction result is comprehensively considered through rearrangement and determination of characteristic combinations, so that the rolling prediction result is more accurate and stable.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without inventive exercise.

FIG. 1 is a schematic diagram of a system according to an embodiment of the present invention;

FIG. 2 is a block diagram of an integrated prediction module according to an embodiment of the present invention;

FIG. 3 is a schematic flow chart of a method according to an embodiment of the present invention;

FIG. 4 is a schematic structural diagram of an apparatus according to an embodiment of the present invention;

FIG. 5 is a schematic structural diagram of an apparatus according to an embodiment of the present invention;

FIG. 6 is a schematic structural diagram of an apparatus according to an embodiment of the present invention;

fig. 7 is a schematic diagram of a server structure according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention clearer, the present invention will be described in further detail with reference to the accompanying drawings, and it is apparent that the described embodiments are only a part of the embodiments of the present invention, not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

In addition, by adopting the rolling prediction scheme in the background art, when the external sequence is applied to perform forward multi-period rolling prediction on the target sequence, the problem of missing of future multi-period external features is encountered, for example, a relation model between delivery data before 3 months and income of the month is established, when the 4 th month needs to be rolled and predicted, the delivery amount of the month may not be determined and cannot be applied, for the two problems mentioned above, some technologies can be used for solving at present, but certain defects still exist.

The problem to be solved by the embodiment of the invention is how to utilize the common influence of the sequence characteristics of the self sequence and the external hysteresis sequence characteristics on the target sequence prediction; in addition, the influence of future external feature missing can be further avoided, and forward multi-period rolling prediction is realized.

Aiming at the problem of forward m-period rolling prediction, the embodiment of the invention can respectively set the minimum lag period from 1 to m, calculate the hysteresis characteristics of an external sequence and the sequence per se, recombine the characteristics to obtain a prediction model, integrate a plurality of combination models through a forward filtering model integration strategy, consider the influence of the hysteresis characteristics of the sequence per se and the hysteresis characteristics of the external sequence on a prediction target sequence, perform forward rolling prediction integration on prediction results of different minimum lag periods (1 to m), and finally realize forward m-period rolling prediction. The technical scheme of the embodiment of the invention mainly comprises the following aspects:

firstly, generating a hysteresis characteristic set and combining the characteristic set; and generating characteristic combinations of the self sequence and the external sequence.

Secondly, a pre-filtering model integration strategy is proposed, so that the influence of the combination of a plurality of self sequence features and external features is considered;

and thirdly, by rolling prediction of different minimum lag phases, the problem of future external feature loss is avoided, and accordingly forward multi-phase rolling prediction is achieved.

Based on the above description, please refer to fig. 1, which includes: 5 modules, namely: the system comprises a hysteresis characteristic generation module, a characteristic combination generation module, a prediction model module, an integrated prediction module and a rolling prediction integrated module.

1. A hysteresis feature generation module:

the hysteresis characteristic generation module is mainly used for generating a hysteresis characteristic alternative set by the input self sequence and the external sequence according to the minimum hysteresis period k determined in the early stage, and is used for preprocessing the input self sequence and the external sequence aiming at the faced hysteresis characteristic and converting the data of the self sequence and the external sequence into the data required by a subsequent module.

2. A feature combination generation module:

the feature combination generation module mainly combines the candidate sets of the hysteresis features in the hysteresis feature generation module to generate a feature combination set (a set of feature combinations), and performs prediction analysis by taking the feature combinations as a unit, thereby realizing the consideration of the common influence of the external features and the feature combination model on the prediction result.

3. A prediction model module:

the prediction model module inputs the feature combination into the prediction model module and outputs a prediction result according to the corresponding model, and the commonly used model can be a linear regression model, a time series model and the like.

4. An integrated prediction module:

the integrated prediction module mainly considers how to select a proper feature combination model for future prediction according to the prediction effect of the feature combination model in the verification set, so that the final prediction effect is more stable.

5. A rolling prediction integration module:

the rolling prediction module mainly summarizes the prediction output at different minimum lag periods k, the minimum lag period k is from 1 to a target m, the prediction results from the forward 1 period to the forward m period can be respectively output, and the final forward multi-period prediction results can be output by integrating the prediction results. The value of k shown in fig. 1 is 1-12, and the specific value of the minimum lag phase may be different in application examples of different self sequences and external sequences, which should not be construed as a limitation on the value range of k.

In the following embodiments, specific implementation schemes of the above five core modules in the embodiments of the present invention will be described in detail, where the above five core modules include: a hysteresis feature generation module, a feature combination generation module, an integrated prediction module, a rolling prediction integration module, and a prediction model module.

Firstly, a hysteresis characteristic generation module:

the lag characteristic generation module converts the input original sequence data into a lag characteristic alternative set. The main steps comprise the following 3 steps:

1. determining a minimum lag period k and a maximum lag period number q;

2. respectively generating hysteresis characteristics with hysteresis stages of k, k +1, … and k + q-1 for each sequence;

3. and reordering the lag characteristics of all the sequences to generate an alternative characteristic set.

The overall process is shown in tables 1-2 and formula (1):

TABLE 1

After the minimum lag phase k and the maximum lag phase number q of the delay characteristic are input, the conversion is shown in table 2:

TABLE 2

Reordering of hysteresis characteristics, as in formula (1):

wherein

In the above example, the single hysteresis characteristic is generated in such a way that the original sequence (which may be an external sequence or a self-sequence) is

The target sequence is

The signature sequence after the first phase corresponds to the target sequence in the following way: then the signature sequence for the lag phase is:

as shown in tables 3 to 5 below:

TABLE 3

Original external characteristic sequence

T1

T2

T3

T4

T5

Extrinsic delay

1

T1

T2

T3

T4

T5

Self time series Y

T1

T2

T3

T4

T5

T6

TABLE 4

X_1,1	Y₁
		X_1,2	Y₂
X_1,3	Y₃
		…	…
X_1,t	Y_t

TABLE 5

NA	Y₁
		X_1,1	Y₂
X_1,2	Y₃
		…	…
X_1,t-1	Y_t

Wherein the hysteresis characteristics of the first column of table 4 are delayed by 1 phase to obtain table 5.

Secondly, a feature combination generation module:

and the feature combination generation module recombines the hysteresis feature sets generated above to obtain a feature combination, namely, the feature combination is selected from each item of the formula (1) and is processed by taking the feature combination as a unit, so that the influence of the external sequence and the self sequence on final prediction is considered at the same time. Generally, 2 to 5 hysteretic features can be selected as a feature combination, and the strategies for selecting the hysteretic features for combination mainly include the following three strategies:

1. and (3) ergodic strategy: that is, for the previous q × p +1 hysteretic features, the traversal takes n hysteretic features as a combination. The strategy has the advantages that the considered combination is comprehensive, and the strategy can be considered when the calculation speed is not required to be considered;

in the traversal strategy, n is uncertain, the value can be 1 at the minimum, and the maximum value can be the total number of the hysteresis characteristics.

2. The randomness strategy is as follows: randomly (with or without putting back) M feature combinations of size n from the feature candidate set. The strategy has the advantages that a small number of feature combinations can be selected, and the calculation efficiency is improved.

The randomness strategy differs from the ergodic strategy in that it defines the number n of selected hysteretic features and defines the number M of feature combinations.

3. The combination strategy of the external sequences and the self sequences comprises the steps of traversing or randomly selecting feature combinations with the size of n from a feature candidate set, wherein n1>0 external sequences (namely x) exist, and n2>0 self sequences (namely y) exist; the advantage of this strategy is that the combination considered must contain both the outer sequence and the self-sequence features.

The outer sequence and self sequence combination strategy differs from the former two strategies in that the hysteresis characteristics of the outer sequence and the inner sequence must be simultaneously contained in the characteristic combination are limited.

Thirdly, a prediction model module:

and inputting the feature combination into a prediction model module, and outputting a prediction result according to the corresponding model.

Fourthly, integrating a prediction module:

the integrated prediction module mainly integrates output results of a plurality of previous combined models (namely feature combinations in the previous text), so that the prediction accuracy is improved, and meanwhile, the prediction is more stable. Two points are mainly considered:

the prediction effect of the combined model in the verification set and the difference of each model in the combined model library.

Based on the two points, the embodiment of the invention provides a forward filtering model selection integration strategy for integration. There are two main aspects, as shown in fig. 2:

1. and performing rolling prediction on the verification set based on the obtained combined models, such as M combined models obtained by training in a randomness strategy, and selecting the first T combined models with better rolling prediction effect to generate an alternative model library.

2. And selecting a combined model from the alternative model library into the integrated model library, wherein the selected combined model needs to meet the requirement that the prediction effect on the verification set in the integrated model library is improved, and the selection is stopped continuously until the stopping condition is met. Thus, the integrated model library may contain a plurality of combined models.

The stop conditions may be two of the following:

firstly, selecting q times at most; i.e. the selection is stopped after q times of selection. Smaller q values may prevent overfitting from occurring.

And secondly, stopping selection when the same combined model is extracted for multiple times, and determining that other combined models cannot improve the integrated prediction effect. The specific number of "multiple times" here may be preset, such as: 2. 5 or other value.

Fifthly, a rolling prediction integration module:

based on the four modules, the obtained integrated model library can be used for performing rolling prediction on the target sequence, which is as follows:

the rolling prediction integration module mainly collects and integrates prediction results of different minimum lag phases k corresponding to each hysteresis characteristic in the integration model library, so that rolling prediction of a plurality of forward phases (m phases) is realized.

Assume that a time point of t, i.e., y, has been observed at present_t,z_1,t,z_2,t,…z_q*(p+1),tThe former exists, and for each minimum lag phase k, the output is a rolling prediction of the k forward phases, with the different minimum lag phases determining the relationship between the target and what hysteresis characteristics, such as: the minimum lag phase is 1, then the relation established by the combined model is the current-period target y_tAnd the last period z_i,t-j(j ≧ 1), then the next phase target y is predicted_t+1The characteristics used were as follows: z is a radical of_i,t-j+1Because t-j +1 is less than or equal to t, z_i,t-j+1Then it must be present.

Similarly, when the minimum lag phase is m, the target y at the h-th (h.ltoreq.m) phase_t+hThe characteristic used is z_i,t-j+hWherein j is more than or equal to m, so t-j + h is less than or equal to t. As shown in tables 6 and 7 below:

TABLE 6

Table 6 shows the prediction results when the minimum lag phase K is 1

The predicted value of the next period can be output

TABLE 7

Using table 7 as the minimum lag period k ═ m, the predicted values of the next m periods can be output

From the above process, it can be seen that the smaller the minimum lag period k is, the closer the used feature is to the feature of the current period, that is, the smaller k is, the closer the used information is to the latest information. The prediction results of different minimum lag periods are integrated in the module, and the rolling prediction of the previous m periods can be realized by combining the information of different lag periods. The structure is as follows:

i.e. from 1 to m for the minimum lag period k, then for the predicted value at time t + i

There are m-i +1 predicted values, then the final rolling prediction is to weight the multiple prediction results together, i.e. the rolling prediction is to weight the multiple prediction results together

The method comprises the following specific steps:

a lag (k is 1) prediction model, and outputs

The prediction results are obtained as follows: w is a₁；

A lag (k is 2) prediction model, and outputs

The prediction results are obtained as follows: w is a₂；

A lag (k ═ m) prediction model, and outputs

The prediction results are obtained as follows: w is a_m；

Scrolling prediction result output

Wherein

The embodiment of the invention provides a method for predicting by simultaneously considering sequence characteristics of the external sequence and external sequence characteristics, aiming at the problem of forward multi-stage rolling prediction of the external sequence, and the accuracy and stability of prediction are improved while the proper hysteresis characteristics are determined by applying integrated prediction. With the enrichment of external features, more and more external data can be applied to traditional time series prediction, and how to use the external features to improve the prediction accuracy can be a concern for many enterprises in the foreseeable future. We propose here a system mechanism to apply external sequence features for prediction.

In addition, different minimum lag phases k are set, the minimum lag phases k are respectively applied with different information of the hysteresis characteristics from 1 to m, and finally rolling prediction integration is used, so that rolling prediction of forward multiple phases is realized, the characteristics used when predicting future multiple phases are ensured to exist, the use of a predicted value of an external characteristic is avoided, and the prediction uncertainty is increased.

The method provided by the embodiment of the invention can be applied to financial income prediction and is also suitable for other time series problems with external sequences, such as prediction of monthly demand of raw materials and the like.

Based on the foregoing description of the embodiment, an embodiment of the present invention provides a method for rolling prediction, as shown in fig. 3, including:

301: determining a lag phase of a hysteresis characteristic of each sequence in the reference sequence; the reference sequence comprises at least two sequences, and the reference sequence comprises an external sequence and/or a self sequence;

in this embodiment, the reference sequence may have only the external sequence or only the self sequence, and preferably both the external sequence and the self sequence are used.

302: reordering the hysteresis characteristics of each sequence in the reference sequence according to the hysteresis period of the hysteresis characteristics of each sequence in the reference sequence, and combining the reordered hysteresis characteristics belonging to the same period to obtain a feature combination to be selected;

different external sequences and self sequences have different lag phases; therefore, different sequences can be calculated, please refer to tables 3-5. The feature combinations can be referred to in terms of formula (1).

303: selecting a feature combination from the feature combinations to be selected to carry out rolling prediction on the verification set with known results, and selecting the feature combination with a rolling prediction result meeting a preset condition as a target feature combination;

the verification set is a sequence with known results, and the characteristics to be selected are used for predicting the sequence, so that the prediction effect of the characteristics can be known to be better, and the other characteristics are relatively poorer; the predetermined condition is met to select a feature combination with a better prediction effect, and how to select, the embodiment of the present invention is not limited uniquely, for example: selecting the optimal feature combination, selecting the top T feature combinations with better preference, and the like, and selecting other algorithm combinations after selection, and the like. The target feature combination is a combination used for finally predicting the self sequence, wherein the self sequence is a target sequence in general, namely: a sequence of rolling predictions is required.

304: and performing rolling prediction on the self sequence by using the target characteristic combination.

After the feature combination including the hysteresis feature is obtained, how to perform the rolling prediction may refer to an existing rolling prediction implementation algorithm, which is not limited in the embodiment of the present invention.

In the embodiment of the invention, at least two sequences are used as references and comprise various hysteresis characteristics, different hysteresis characteristics correspond to hysteresis stages of the sequences, and the influence of self sequences/external sequences on a rolling prediction result is comprehensively considered through rearrangement and characteristic combination determination, so that the rolling prediction result is more accurate and stable.

The embodiment of the invention also provides an implementation scheme for selecting the feature combination, which specifically comprises the following steps: the selecting of the feature combination from the feature combinations to be selected includes:

selecting a feature combination from the feature combinations to be selected in a traversal mode; or selecting a feature combination from the feature combinations to be selected in a random mode; or selecting a feature combination from the feature combinations to be selected in a traversing or random mode, wherein the selected feature combination comprises at least one hysteresis feature of an external sequence and at least one hysteresis feature of a self sequence.

Further, since there may be many feature combinations that meet the conditions, and are not all needed, the embodiment of the present invention provides a solution for simplifying the feature combinations, which is specifically as follows: the target feature combination comprises at least two feature combinations meeting a preset condition; before performing rolling prediction on the self-sequence by using the target feature combination, the method further includes:

The above stopping conditions may be set arbitrarily according to the requirement for improving the prediction accuracy and stability, and in this embodiment, three optional implementation schemes are provided for reference, which are specifically as follows: the predetermined stop condition includes:

adding a new selected feature combination into the rolling prediction result of the target combination to be determined, and not increasing any more;

or selecting the times of the feature combination to reach the preset times;

Because the target feature combination at least includes two hysteresis features, the embodiment of the present invention further provides a calculation scheme of a rolling prediction result, which is specifically as follows: the target feature combination includes at least two feature combinations, and the performing rolling prediction on the self-sequence using the target feature combination includes:

and performing rolling prediction on the self sequence by using each feature combination in the target feature combinations, and calculating a weighted average value of rolling prediction results.

In this embodiment, the target feature combination includes a feature combination, the feature combination includes a hysteresis feature, and an existing rolling prediction algorithm may be referred to for a rolling prediction implementation scheme of each feature combination, which is not uniquely defined in the embodiment of the present invention; the weight value used by different feature combinations can be determined through a training algorithm, the weight value is not limited uniquely in the embodiment of the invention, and in addition, the specific description of the rolling prediction integration module in the previous embodiment can be referred to in the weighted average calculation mode.

An embodiment of the present invention further provides a rolling prediction apparatus, as shown in fig. 4, including:

a lag period determining unit 401, configured to determine lag periods of lag characteristics of each sequence in the reference sequence; the reference sequence comprises at least two sequences, and the reference sequence comprises an external sequence and/or a self sequence;

a reordering unit 402, configured to reorder the hysteresis characteristics of each sequence in the reference sequence according to the hysteresis of each sequence in the reference sequence;

a combining unit 403, configured to combine the reordered hysteresis characteristics belonging to the same period to obtain a candidate characteristic combination;

a feature selection unit 404, configured to select a feature combination from the feature combinations to be selected to perform rolling prediction on the verification set with known results, and select a feature combination with a rolling prediction result meeting a predetermined condition as a target feature combination;

a rolling prediction unit 405 configured to perform rolling prediction on the self sequence using the target feature combination.

The verification set is a sequence with known results, and the characteristics to be selected are used for predicting the sequence, so that the prediction effect of the characteristics can be known to be better, and the other characteristics are relatively poorer; the predetermined condition is met to select the feature combination with better prediction effect, and how to select, the embodiment of the present invention is not limited uniquely.

The embodiment of the invention also provides an implementation scheme for selecting the feature combination, which specifically comprises the following steps: the feature selecting unit 404 is configured to select a feature combination from the feature combinations to be selected in a traversal manner; or selecting a feature combination from the feature combinations to be selected in a random mode; or selecting a feature combination from the feature combinations to be selected in a traversing or random mode, wherein the selected feature combination comprises at least one hysteresis feature of an external sequence and at least one hysteresis feature of a self sequence.

Further, since there may be many feature combinations that meet the conditions, and are not all needed, the embodiment of the present invention provides a solution for simplifying the feature combinations, which is specifically as follows: as shown in fig. 5, the target feature combination includes at least two feature combinations meeting a predetermined condition; the above-mentioned device still includes:

a combination selecting unit 501, configured to select a feature combination from the feature combinations that meet the predetermined condition to obtain a target combination to be determined before the rolling predicting unit 405 performs rolling prediction on the self sequence using the target feature combination;

the rolling prediction unit 405 is further configured to perform rolling prediction on the verification set of the known result by using the target combination to be determined;

the feature selecting unit 404 is configured to take the target combination to be determined as the target feature combination when a predetermined stop condition is met during the scroll prediction performed by the scroll predicting unit 405.

or selecting the times of the feature combination to reach the preset times;

or, when the newly selected feature combination is added to the condition that the prediction result of the target combination to be determined has the promotion prediction result, the number of times of selecting the same feature combination reaches the specified number of times.

Because the target feature combination at least includes two hysteresis features, the embodiment of the present invention further provides a calculation scheme of a rolling prediction result, which is specifically as follows: the target feature combination comprises at least two feature combinations;

the rolling prediction unit 405 performs rolling prediction on the self-sequence using each of the target feature combinations, and then calculates a weighted average of results of the rolling prediction.

Another electronic device may be provided in an embodiment of the present invention, and as shown in fig. 6, the electronic device may be used for scrolling prediction, and includes: the processor 601, memory 602 may also include an input device 603 and an output device 604; wherein the memory 602 may be used for buffering required for the processor 601 to perform data processing, and also for storing data received by the input device 603 and data to be transmitted by the output device 604;

the processor 601 is configured to determine a lag phase of a hysteresis characteristic of each sequence in the reference sequence; the reference sequence comprises at least two sequences, and the reference sequence comprises an external sequence and/or a self sequence; reordering the hysteresis characteristics of each sequence in the reference sequence according to the hysteresis period of the hysteresis characteristics of each sequence in the reference sequence, and combining the reordered hysteresis characteristics belonging to the same period to obtain a feature combination to be selected; selecting a feature combination from the feature combinations to be selected to carry out rolling prediction on the verification set with known results, and selecting the feature combination with a rolling prediction result meeting a preset condition as a target feature combination; and performing rolling prediction on the self sequence by using the target characteristic combination.

The embodiment of the invention also provides an implementation scheme for selecting the feature combination, which specifically comprises the following steps: the processor 601 is configured to select a feature combination from the feature combinations to be selected, and includes:

the characteristic combination is selected from the characteristic combinations to be selected in a traversal mode; or selecting a feature combination from the feature combinations to be selected in a random mode; or selecting a feature combination from the feature combinations to be selected in a traversing or random mode, wherein the selected feature combination comprises at least one hysteresis feature of an external sequence and at least one hysteresis feature of a self sequence.

Further, since there may be many feature combinations that meet the conditions, and are not all needed, the embodiment of the present invention provides a solution for simplifying the feature combinations, which is specifically as follows: the target feature combination comprises at least two feature combinations meeting a preset condition; the processor 601 is further configured to select a feature combination from the feature combinations meeting the predetermined condition to obtain a target combination to be determined before performing rolling prediction on the self sequence by using the target feature combination;

or selecting the times of the feature combination to reach the preset times;

Because the target feature combination at least includes two hysteresis features, the embodiment of the present invention further provides a calculation scheme of a rolling prediction result, which is specifically as follows: the target feature combination includes at least two feature combinations, and the processor 601 configured to perform rolling prediction on the self-sequence using the target feature combination includes:

and a step of performing rolling prediction on the self-sequence by using each feature combination in the target feature combinations, and then calculating a weighted average of results of the rolling prediction.

Embodiments of the present invention also provide a server, which may be applied to rolling prediction, as shown in fig. 7, which is a schematic structural diagram of a server provided by an embodiment of the present invention, the server 700 may generate relatively large differences due to different configurations or performances, and may include one or more Central Processing Units (CPUs) 722 (e.g., one or more processors) and a memory 732, and one or more storage media 730 (e.g., one or more mass storage devices) storing an application 742 or data 744. Memory 732 and storage medium 730 may be, among other things, transient storage or persistent storage. The program stored in the storage medium 730 may include one or more modules (not shown), each of which may include a series of instruction operations for the server. Further, the central processor 722 may be configured to communicate with the storage medium 730, and execute a series of instruction operations in the storage medium 730 on the server 700.

The server 700 may also include one or more power supplies 726, one or more wired or wireless network interfaces 750, one or more input-output interfaces 758, and/or one or more operating systems 741, such as Windows Server, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, and so forth.

The above-described embodiment may perform the rolling prediction in the foregoing embodiment based on the server structure shown in fig. 7. It should be noted that the implementation scheme of the rolling prediction in the embodiment of the present invention may also be based on a terminal device, and is not limited to a server.

It should be noted that, in the embodiment of the rolling prediction apparatus, the included units are merely divided according to the functional logic, but are not limited to the above division as long as the corresponding functions can be implemented; in addition, specific names of the functional units are only for convenience of distinguishing from each other, and are not used for limiting the protection scope of the present invention.

In addition, it is understood by those skilled in the art that all or part of the steps in the above method embodiments may be implemented by related hardware, and the corresponding program may be stored in a computer readable storage medium, where the above mentioned storage medium may be a read-only memory, a magnetic disk or an optical disk.

The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the embodiment of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method of rolling prediction, comprising:

determining a lag phase of a hysteresis characteristic of each sequence in the reference sequence; the reference sequence comprises at least two sequences, and the reference sequence comprises an external sequence and a self sequence; the lag phase is a lag time when there is a correlation between data of a sequence in the reference sequence and data of a target sequence;

2. The method of claim 1, wherein the selecting a feature combination from the candidate feature combinations comprises:

3. The method according to claim 1 or 2, wherein the target feature combination comprises at least two feature combinations meeting a predetermined condition; before using the target feature combination to perform rolling prediction on the self-sequence, the method further comprises:

4. The method of claim 3, wherein the predetermined stop condition comprises:

or selecting the times of the feature combination to reach the preset times;

5. The method according to claim 1 or 2, wherein the target feature combination comprises at least two feature combinations, and the using the target feature combination to perform the rolling prediction on the self-sequence comprises:

6. The method of claim 3, wherein the target feature combination comprises at least two feature combinations, and wherein the using the target feature combination to perform the rolling prediction on the self-sequence comprises:

7. The method of claim 4, wherein the target feature combination comprises at least two feature combinations, and wherein the using the target feature combination to perform the rolling prediction on the self-sequence comprises:

8. A rolling prediction apparatus, comprising:

the lag phase determining unit is used for determining the lag phase of the lag characteristic of each sequence in the reference sequence; the reference sequence comprises at least two sequences, and the reference sequence comprises an external sequence and a self sequence; the lag phase is a lag time when there is a correlation between data of a sequence in the reference sequence and data of a target sequence;

9. The apparatus of claim 8,

the feature selection unit is used for selecting a feature combination from the feature combinations to be selected in a traversal mode; or selecting a feature combination from the feature combinations to be selected in a random mode; or selecting a feature combination from the feature combinations to be selected in a traversal or random mode, wherein the selected feature combination comprises at least one hysteresis feature of an external sequence and at least one hysteresis feature of a self sequence.

10. The apparatus according to claim 8 or 9, wherein the target feature combination comprises at least two feature combinations meeting a predetermined condition; the device further comprises:

11. The apparatus of claim 10, wherein the predetermined stop condition comprises:

or selecting the times of the feature combination to reach the preset times;

12. The apparatus according to claim 8 or 9, wherein the target feature combination comprises at least two feature combinations;

13. The apparatus of claim 10, wherein the target feature combination comprises at least two feature combinations;

14. The apparatus of claim 11, wherein the target feature combination comprises at least two feature combinations;