CN104992050A - Method for selecting prediction model of time sequence characteristic evaluation based on statistical signal processing - Google Patents
Method for selecting prediction model of time sequence characteristic evaluation based on statistical signal processing Download PDFInfo
- Publication number
- CN104992050A CN104992050A CN201510324354.6A CN201510324354A CN104992050A CN 104992050 A CN104992050 A CN 104992050A CN 201510324354 A CN201510324354 A CN 201510324354A CN 104992050 A CN104992050 A CN 104992050A
- Authority
- CN
- China
- Prior art keywords
- time series
- tendency
- sequence
- trend
- amplitude
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Complex Calculations (AREA)
Abstract
A method for selecting a prediction model of time sequence characteristic evaluation based on statistical signal processing belongs to the field of time sequence prediction and aims to solve the problem that a current switching operation order supporting plate cannot be used conveniently under special cases. The method comprises the steps that amplitude continuity, long memorability, tendency and seasonality of an input time sequence are judged, and a characteristic evaluation type of the input time sequence is determined; at the second step, the predication models of corresponding types are obtained according to mapping relations between the type determined at the first step and the prediction models; and at the third step, the optimal model is selected among the prediction models of the relevant types and used for time sequence prediction. The method for selecting the prediction model of the time sequence characteristic evaluation based on the statistical signal processing provided by the invention is used to predict the time sequence using the time sequence prediction model.
Description
Technical field
The invention belongs to time series forecasting field.
Background technology
Time series predicting model is for the short-term forecasting to future, in actual applications when predicting for certain class time series, needs to select forecast model.When qualitative selection forecast model, characteristic based on seasonal effect in time series characteristic and forecast model realizes mating between time series with forecast model, its object is mainly the scope judging to reduce candidate prediction model by seasonal effect in time series characteristic, is realized the preliminary screening of forecast model by data characteristic.Existing when qualitative selection, exist existing time series kind divide that the index of institute foundation is comparatively single, kind divides not in detail, characteristic covers not comprehensively, with the problem such as forecast model correspondence is unintelligible, that is: there is the problem adopting the prediction effect of the forecast model of existing method choice bad.
Summary of the invention
The object of the invention is, in order to solve the problem problem existing and adopt the prediction effect of the forecast model of existing method choice bad, to the invention provides a kind of forecast model system of selection of time series evaluating characteristics of Corpus--based Method signal transacting.
The forecast model system of selection of the time series evaluating characteristics of Corpus--based Method signal transacting of the present invention,
Described method comprises the steps:
Step one: input time sequence, carry out characteristic judgement:
Amplitude continuity is carried out to sequence input time, Long Memory Properties, tendency and seasonal judgement, determine the evaluating characteristics kind of sequence input time, continuous-random walk time series that described evaluating characteristics kind comprises amplitude, amplitude continuously-short Memorability time series, amplitude continuously-tendency-seasonal time series, amplitude continuously-tendency-exponential trend time series, amplitude continuously-tendency-nonexponential trend time series, amplitude continuously-tendency-complicated trend time series, amplitude continuously-Long Memory Properties-seasonal time series, amplitude continuously-Long Memory Properties-without seasonal time series and amplitude discrete-time series,
Step 2: the mapping relations between the kind determined according to step one and forecast model, obtains the forecast model of corresponding kind;
Step 3: select optimum one in the forecast model of corresponding kind, for time series forecasting.
The method of described step one comprises:
Step is one by one: carry out the judgement of amplitude continuity to sequence input time:
Judge that input time, sequence was continuous time series or discrete-time series, when for continuous time series, proceed to step one two; When being discrete-time series, determine that the evaluating characteristics kind of sequence input time is amplitude discrete-time series;
Step one two: Long Memory Properties judgement is carried out to sequence input time:
Step one 21: the Hurst index being calculated continuous time series by R/S method, when judging whether Hurst index is greater than 0.5 and is less than 1, is if so, proceeded to step one two or four, if not, proceed to step one two or three;
Step one two or three: when Hurst index is for equaling 0.5, continuous time series are random walk sequence, terminate; When Hurst index is greater than 0 and is less than 0.5, continuous time series are short Memorability sequence, terminate;
Step one two or four: KPSS inspection is carried out to continuous time series, judge whether described continuous time series are short Memorability sequence, if, determine the evaluating characteristics kind of sequence input time be amplitude continuously-short Memorability time series, if not, continuous time series are that Long Memory Properties sequence or trend strengthen sequence, proceed to step one three;
Step one three: tendency and seasonal judgement are carried out to sequence input time:
The Long Memory Properties sequence of determining step 1 or trend strengthen sequence and whether have tendency and seasonality, if trendless and have seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-Long Memory Properties-seasonal time series, if have tendency and without seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-exponential trend time series or amplitude continuously-tendency-nonexponential trend time series or amplitude continuously-tendency-complicated trend time series, if trendless and without seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-Long Memory Properties-without seasonal time series, if have tendency and have seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-seasonal time series.
Described step one three, if have tendency and without seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-exponential trend time series or amplitude continuously-tendency-nonexponential trend time series or amplitude continuously-tendency-complicated trend seasonal effect in time series method is:
Steps A: by input time sequence curve and the simple trend fitting curve comparison of setting, judged whether simple trend, if so, then proceeded to step B; If not, then determine the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-complicated trend time series;
Step B: judge sequence input time whether index of coincidence trend, if, then determine the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-exponential trend time series, if not, then determine the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-nonexponential trend time series.
Beneficial effect of the present invention is,
The present invention analyzes the characteristic of sequence input time, determine the classification of sequence input time, what realize between forecast model with data characteristic is corresponding, the mapping relations of multi-to-multi between forecast model and time series kind are set up according to priori, to make every class time series have some feasible candidate families, realize the output that time series kind determines rear corresponding candidate forecast model.The present invention is based on the research that existing time series kind divides, kind for existing single index divides to be expanded, based on multi objective, the classification carrying out time series kind by different level, the characteristic index selected be existing involved, convenience of calculation, result corresponding with characteristic clear and definite, cover most of characteristic and irredundant and reasonable and criterions such as the kind division of mutual exclusion substantially can be carried out to time series based on corresponding statistical property.In conjunction with the separating capacity of different index for forecast model, index is carried out to the division of level, namely determine the order of the index that the first step judges and the index that subsequent serial judges.The division of preliminary time series kind is carried out in combination based on different index, and according to the forecasting problem of reality, forecast model, the time series kind that seasonal effect in time series General Requirements combines for part of properties is accepted or rejected, set up one and can cover more usual time sequence characteristic, can carry out to similar time series the time series kind division system that kind divides, correspondence that is mutually exclusive and forecast model is comparatively clear and definite comparatively accurately.Forecast model prediction effect that the present invention selects is significantly better than existing method choice.
Accompanying drawing explanation
Fig. 1 is the principle schematic of evaluating characteristics kind of sequence input time of the present invention.
Embodiment
Embodiment one: the forecast model system of selection of the time series evaluating characteristics of the Corpus--based Method signal transacting described in present embodiment, described method comprises the steps:
Step one: input time sequence, carry out characteristic judgement:
Amplitude continuity is carried out to sequence input time, Long Memory Properties, tendency and seasonal judgement, determine the evaluating characteristics kind of sequence input time, continuous-random walk time series that described evaluating characteristics kind comprises amplitude, amplitude continuously-short Memorability time series, amplitude continuously-tendency-seasonal time series, amplitude continuously-tendency-exponential trend time series, amplitude continuously-tendency-nonexponential trend time series, amplitude continuously-tendency-complicated trend time series, amplitude continuously-Long Memory Properties-seasonal time series, amplitude continuously-Long Memory Properties-without seasonal time series and amplitude discrete-time series,
Step 2: the mapping relations between the kind determined according to step one and forecast model, obtains the forecast model of corresponding kind;
Step 3: select optimum one in the forecast model of corresponding kind, for time series forecasting.
Embodiment two: composition graphs 1 illustrates present embodiment, present embodiment is the further restriction of the forecast model system of selection of time series evaluating characteristics to the Corpus--based Method signal transacting described in embodiment one,
Existing major part research is divided into continuous time series and discrete-time series according to the continuity of seasonal effect in time series amplitude; Whether time series is divided into Long Memory Properties and short Memorability sequence by the existence according to Long Memory Properties; Also research existence that is seasonal according to time series and tendency is had whether time series to be divided into corresponding kind; According to Stability Classification, stationary time series and nonstationary time series can be divided into; Time series can be divided into determinacy time series and randomness time series according to whether comprising uncertain factor in time series, randomness time series can be divided into stationary random sequence, non-stationary random series and white noise sequence again.Existing time series kind divides, and generally carries out simple kind division according to one or two quality versus time sequence.And for forecast model, often requiring seasonal effect in time series will more refinement, the correspondence of the classification that two classification that simple single index carries out time series obtain and time series predicting model is bad, makes the forecast model that a certain class time series is corresponding a large amount of.Therefore, the primary demand of the comprehensive actual forecast demand of present embodiment and model, adopts multi objective to realize classifying to seasonal effect in time series.Seasonal effect in time series variation feature can be divided into four classes: secular trend, seasonal move, cyclical variations and erratic variation.In addition, seasonal effect in time series characteristic is often portrayed by randomness, stationarity and seasonality.
In present embodiment, the characteristic of time employing has: amplitude continuity, Long Memory Properties, tendency, seasonality.
(1) amplitude continuity: for time series, important effect is played in the whether continuous selection to prediction algorithm of amplitude.Generally speaking, the time series of process is continuous, the time-discrete time series of amplitude.And for part-time sequence as some state of a control parameter, its amplitude is also discrete, only obtain finite number value, this just makes the prediction algorithm for continuous time series no longer applicable here, and some should be adopted to realize prediction based on state transition probability as Markov model etc.Therefore, forecast model is divided into two large classes by amplitude continuity, and the forecast model wherein for continuous time series accounts for the overwhelming majority.Like this, if time series is amplitude discrete series, so just can fast by forecast model range shorter.Therefore, select amplitude continuity as first judge index.
(2) Long Memory Properties: whether time series can be divided into Long Memory Properties time series and short Memorability time series according to the existence of Long Memory Properties, and the judge index of the Long Memory Properties mainly fractal degree of seasonal effect in time series.Long Memory Properties seasonal effect in time series coefficient of autocorrelation with postpone the increase of step-length successively decrease very slowly, according to hyperbolicity decay, autocorrelation haracter sex expression fairly obvious, but most of the time sequential forecasting models can not be predicted well for long memory characteristic.Short memory time sequence autocorrelation according to geometry rate attenuation, Long Memory Properties time series considers seasonal effect in time series Long Memory Properties on the basis of sequence short memory time, and therefore short memory time, sequence was commonly called classical time series.In addition, time series can be divided into random walk sequence and nonrandom migration sequence according to whether obeying Markov characteristic.For forecast model, whether most important for Model Selection the existence of Long Memory Properties is, time series predicting model can be carried out clear and definite division according to Long Memory Properties seasonal effect in time series processing power, only have small part time series predicting model can process Long Memory Properties seasonal effect in time series forecasting problem.In addition, for random walk sequence, forecast model is targetedly had partly can to realize good prediction to it.And for this two class feature, all can judge with Hurst index.Therefore, for amplitude continuous print time series, utilize the value of Hurst index just can realize the judgement of Long Memory Properties and random walk characteristic, this two class feature can by Long Memory Properties time series and random walk time series the model fast that is suitable for branch away, large class can be carried out based on the result of calculation of Hurst index to time series and forecast model, the kind of mutual exclusion divides.Therefore second layer characteristic judges to carry out based on Hurst index.
When Hurst index equals 0.5 time, show that time series possesses random walk characteristic, obey Markov characteristic.When Hurst index is between 0 to 0.5 time, illustrate to possess short Memorability, there is long-term negative correlation characteristic between data, if namely go up a moment time series to be in ascendant trend, more may there is downtrending in subsequent time, there is mean reversion phenomenon.When Hurst index is between 0.5 to 1 time, time series is that Long Memory Properties or trend strengthen sequence, if a upper moment time series be in ascendant trend so subsequent time more may keep identical variation tendency, Hurst index is more stronger close to 1 trend, curve is more smooth, tendency is more obvious, and more noise is larger, trend is more unstable close to 0.5 for Hurst index.
The conventional Long Memory Properties method of inspection has that R/S and modified R/S method, KPSS check, LM checks and gph test.Currently adopt R/S method of inspection for the research of Long Memory Properties seasonal effect in time series is main, but this method have a defect be for short memory time sequence Hurst index calculate and there will be larger deviation.Therefore, need in the application in conjunction with different Hurst index calculation method comprehensive descision seasonal effect in time series Long Memory Properties.
In the present embodiment, consider to adopt R/S method and KPSS method two kinds of computing method to carry out the calculating of corresponding statistical indicator.Wherein, the Hurst index that R/S method calculates is between 0 to 1, can judge long memory characteristic by above-mentioned method.KPSS method calculates corresponding statistical value, seasonal effect in time series Long Memory Properties and short Memorability can be judged, short Memorability time series Hurst index can be calculated to R/S method and easily occur that the problem of relatively large deviation makes up, but also there is the problem of oneself, namely KPSS checks for trendless but has the time series of Seasonal Characteristics easily to occur erroneous judgement.The critical value of the statistic that KPSS calculates and significance 10% compares, if statistical value is less than critical value corresponding to 10% significance, show that the significance of assay is greater than 10%, then accept null hypothesis, time series is short Memorability time series, otherwise refusal null hypothesis, time series is Long Memory Properties time series, therefore KPSS can only provide long memory or the judgement of short memory two class seasonal effect in time series, is mainly used to the problem that modified R/S method is larger to short Memorability seasonal effect in time series H value calculation deviation.The corresponding result of calculation of public data collection is as shown in table 1.
Table 1 Long Memory Properties judges statistical indicator result of calculation
In the process that amplitude continuous print 68 time serieses judge, the time series characteristic judged result having 9 data sets to obtain based on distinct methods is inconsistent.Wherein, inconsistent situation can be divided into following two classes:
A () R/S method shows there is long memory characteristic, KPSS verifies as short memory characteristic, and reality is long memory characteristic, and namely KPSS gets the wrong sow by the ear.Show in experimental data, obvious trend characteristic is not all had in amplitude continuous print 68 seasonal effect in time series 4 time serieses, have or be similar to and have Seasonal Characteristics, all in all being easily mistaken as is have mean reversion characteristic, therefore in KPSS deterministic process, is easily mistaken for the stable time series of short memory.Therefore, KPSS verifies as short Memorability and the time series not possessing Seasonal Characteristics is only sequence short memory time.
B () R/S method shows there is long memory characteristic, KPSS verifies as short memory characteristic, and reality is short memory characteristic, namely R/S method for short memory time sequence the calculating of H value there is deviation.Show in experimental data, 5 seasonal effect in time series characteristics in amplitude continuous print 68 time serieses are comparatively close, are the time series of carrying out fluctuating near certain numerical value, obvious short memory characteristic.When time series has obvious short memory characteristic time, easily there is obvious deviation in R/S method Hurst index result of calculation, therefore needs to judge in conjunction with KPSS result.
Can find out, there is certain difference in the possibility of result that the different Long Memory Properties methods of inspection provides, for short Memorability time series, the result of calculation of Hurst index easily occurs deviation, is mistaken for Long Memory Properties; For trendless, seasonal Long Memory Properties time series, KPSS method is easily mistaken for short Memorability, therefore Long Memory Properties, further KPSS are calculated as R/S and verify as the time series of short Memorability, need to carry out seasonal judgement further, if trendless, seasonal time series are then judged as long memory characteristic, otherwise export as short memory characteristic according to KPSS judged result.
(3) tendency and seasonality: according to seasonal effect in time series stationarity whether, time series roughly can be divided into stationary time series and nonstationary time series.And this division methods is also common in time series kind divides, many forecast models have clear and definite demand for seasonal effect in time series stationarity.The characteristic of seasonal effect in time series stationarity and data does not change in time, and jiggly time series the most commonly exist tendency seasonal or the two have both.Also whether time series has been carried out the division of corresponding kind according to tendency and seasonal existence in some researchs.Therefore, consider tendency and seasonality to include in determination range here.Because tendency and seasonal time series are random walk, short memory scarcely, this also just divides the time series kind obtained and is connected with Hurst index, whether the Partial Species that can mark off Hurst index carries out the Further Division of time series characteristic based on existence that is seasonal and tendency, therefore selects seasonal and tendency to judge as the characteristic of third layer.But for random walk sequence and short Memorability sequence, there is tendency and seasonality scarcely, the characteristic therefore just without the need to continuing this two kind judges.And Long Memory Properties and trend enhancement (tendency here only considers monotone increasing or downtrending) time series are needed to carry out the judgement of further Seasonal Characteristics, the judgement of trend kind.This is because different forecast models is different for the processing power of Seasonal Characteristics, therefore seasonal judgement is necessary.Part statistical prediction methods is applicable to specific tendency data more by force or only for the predictive ability of particular tendency, and the further judgement therefore for trend kind is also necessary.Therefore seasonality, tendency and trend kind are as third layer characteristic judge index.
Time series kind divides:
According in time series characteristic system, in conjunction with reality, time series is divided into 9 classes, as shown in Figure 1:
First, according to seasonal effect in time series amplitude continuity, time series is divided into continuous time series and discrete-time series two class.Wherein, the applicable models for discrete-time series is less, therefore without the need to carrying out further refinement judgement.And in actual applications, the process problem of discrete-time series is less, be only cover time sequence characteristic comparatively comprehensively, retain discrete-time series classification.
Secondly, for the forecasting problem of the overwhelming majority, faced by be all continuous time series, therefore need to carry out further characteristic division to continuous time series.With reference to part document, time series generally can be divided into steadily and non-stationary two class.If second step divides according to stationarity, there are some problems: the corresponding relation of stationarity and non-stationary time series and forecast model is indefinite, major part forecast model requires not strict for stationarity, in addition, next step division can be caused comparatively complicated if carry out division according to stationarity, the time series kind produced is more, is inconvenient to manage.Therefore, second step divides according to seasonal effect in time series Long Memory Properties, and one is that Long Memory Properties and seasonal effect in time series correspondence are comparatively clear and definite, can rapid drop model scope for part of properties; Two is that alternative is each other obvious, for follow-up judgement advantageously.This step mainly utilizes R/S method to calculate Hurst index in conjunction with KPSS judgement, time series is divided into the class in four classification of the second layer.
Because trend enhancement sequence and Long Memory Properties sequence still may exist Seasonal Characteristics, therefore, in present embodiment, whether three link model judges for the seasonal existence of this two classes seasonal effect in time series.
The method of described step one comprises:
Step is one by one: carry out the judgement of amplitude continuity to sequence input time:
Judge that input time, sequence was continuous time series or discrete-time series, when for continuous time series, proceed to step one two; When being discrete-time series, determine that the evaluating characteristics kind of sequence input time is amplitude discrete-time series;
Step one two: Long Memory Properties judgement is carried out to sequence input time:
Step one 21: the Hurst index being calculated continuous time series by R/S method, when judging whether Hurst index is greater than 0.5 and is less than 1, is if so, proceeded to step one two or four, if not, proceed to step one two or three;
Step one two or three: when Hurst index is for equaling 0.5, continuous time series are random walk sequence, terminate; When Hurst index is greater than 0 and is less than 0.5, continuous time series are short Memorability sequence, terminate;
Step one two or four: KPSS inspection is carried out to continuous time series, judge whether described continuous time series are short Memorability sequence, if, determine the evaluating characteristics kind of sequence input time be amplitude continuously-short Memorability time series, if not, continuous time series are that Long Memory Properties sequence or trend strengthen sequence, proceed to step one three;
Step one three: tendency and seasonal judgement are carried out to sequence input time:
The Long Memory Properties sequence of determining step 1 or trend strengthen sequence and whether have tendency and seasonality, if trendless and have seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-Long Memory Properties-seasonal time series, if have tendency and without seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-exponential trend time series or amplitude continuously-tendency-nonexponential trend time series or amplitude continuously-tendency-complicated trend time series, if trendless and without seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-Long Memory Properties-without seasonal time series, if have tendency and have seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-seasonal time series.
The classification results of public data collection is as shown in table 2.
Table 2 public data collection characteristic divides corresponding result
Concentrate 70 public datas, the classification of 3 data sets is had to occur problem, classification results and the intuitive nature of all the other 67 data sets judge consistent, classify accuracy is 95.71%, the time series of makeing mistakes of classifying is as shown in table 3, provides classification directly perceived and eigen system divides the classification obtained in table.
The wrong point Time-series Theory classification of table 3 and classification obtain classification and contrast
Can find out that the mistake of time series kind is divided mainly because the calculating of H value, undulatory property judge to cause.
(1) H value calculates: here for data set tempdub, R/S method and KPSS inspection all provide the judgement of short Memorability, and the actual mistake being long memory trendless seasonal time series and causing KPSS result of calculation, and R/S method is short Memorability to this time series result of calculation is a few cases, obtaining its Hurst index through theory deduction is 0.53, and actual is weak Long Memory Properties.But the introducing due to theoretical Hurst derivation result can cause the mistake of other data sets more to divide, and therefore data set introduces the amendment that characteristic judges framework here, and this framework is not suitable for this data set no longer for this reason.
(2) undulatory property: the basic classification of data set GAGurine and rwalk is correct, but the setting of corresponding classification it is considered that some algorithm is applicable to the time series forecasting of particular tendency, corresponding algorithm requires that seasonal effect in time series undulatory property is less substantially, and overall trend is comparatively obvious, curve is comparatively smooth.And above-mentioned time series does not meet corresponding requirement, even if be classified into corresponding classification, also bad with the matching degree of forecast model.Therefore in tendency deterministic process, be judged as not possessing simple trend, thus entered the judgement of coefficient of dispersion, the coefficient of dispersion of these two data sets is respectively 136.40 and 0.84, then GAGurine directly shows and classifies unsuccessfully, and rwalk can be divided into complicated trend class, still classification error.
Can find out, if rely on corresponding mathematical computations completely, because the applicability of some algorithm and the problem of robustness may cause certain mistake to divide, therefore, export judge to help the modes such as explanation to revise the result of classification in conjunction with multiple method, artificial judgment correction, software.
Except the time series that above-mentioned 3 mistakes are divided, other time serieses belonging to identical category show very high similarity, this taxonomic hierarchies comparatively reasonably can be classified to current data set, each data set can corresponding corresponding kind, the characteristic that the time series of one species shows is comparatively similar, and property sort system possesses certain applicability.
Embodiment three: present embodiment is the further restriction of the forecast model system of selection of time series evaluating characteristics to the Corpus--based Method signal transacting described in embodiment two, described step one three, if have tendency and without seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-exponential trend time series or amplitude continuously-tendency-nonexponential trend time series or amplitude continuously-tendency-complicated trend seasonal effect in time series method is:
Steps A: by input time sequence curve and the simple trend fitting curve comparison of setting, judged whether simple trend, if so, then proceeded to step B; If not, then determine the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-complicated trend time series;
Step B: judge sequence input time whether index of coincidence trend, if, then determine the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-exponential trend time series, if not, then determine the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-nonexponential trend time series.
Because trend enhancement sequence and Long Memory Properties sequence still may exist Seasonal Characteristics, therefore, at three link model, whether the seasonal existence of this two classes seasonal effect in time series is judged in present embodiment.By analyzing a large amount of public data collection, trend is here roughly divided into two classes: simple trend (exponential trend, linear trend, para-curve trend, cubic curve trend etc. can by the trend of the good predict such as exponential smoothing) and complicated trend (can pass through the trend of complicated function matching).Corresponding time series to be divided accordingly, further, for simple trend, because fractional prediction algorithm is fine for the seasonal effect in time series predictive ability of exponential trend, the very applicable seasonal effect in time series prediction carrying out exponential trend, therefore exponential trend is carried out the division of independent kind, and the trend of other characteristics divides as the kind of nonexponential trend.For complicated trend data, if want to provide good predict by the mode of regression fit, then need the higher i.e. trend of seasonal effect in time series signal to noise ratio (S/N ratio) comparatively level and smooth and obvious, therefore the judgement of additional discrete coefficient here, coefficient of dispersion judgement is carried out to the time series removing tendency, with 5 for threshold value, if be less than 5, can incorporate into into complicated trend classification.
Claims (3)
1. a forecast model system of selection for the time series evaluating characteristics of Corpus--based Method signal transacting, is characterized in that, described method comprises the steps:
Step one: input time sequence, carry out characteristic judgement:
Amplitude continuity is carried out to sequence input time, Long Memory Properties, tendency and seasonal judgement, determine the evaluating characteristics kind of sequence input time, continuous-random walk time series that described evaluating characteristics kind comprises amplitude, amplitude continuously-short Memorability time series, amplitude continuously-tendency-seasonal time series, amplitude continuously-tendency-exponential trend time series, amplitude continuously-tendency-nonexponential trend time series, amplitude continuously-tendency-complicated trend time series, amplitude continuously-Long Memory Properties-seasonal time series, amplitude continuously-Long Memory Properties-without seasonal time series and amplitude discrete-time series,
Step 2: the mapping relations between the kind determined according to step one and forecast model, obtains the forecast model of corresponding kind;
Step 3: select optimum one in the forecast model of corresponding kind, for time series forecasting.
2. the forecast model system of selection of the time series evaluating characteristics of Corpus--based Method signal transacting according to claim 1, is characterized in that, the method for described step one comprises:
Step is one by one: carry out the judgement of amplitude continuity to sequence input time:
Judge that input time, sequence was continuous time series or discrete-time series, when for continuous time series, proceed to step one two; When being discrete-time series, determine that the evaluating characteristics kind of sequence input time is amplitude discrete-time series;
Step one two: Long Memory Properties judgement is carried out to sequence input time:
Step one 21: the Hurst index being calculated continuous time series by R/S method, when judging whether Hurst index is greater than 0.5 and is less than 1, is if so, proceeded to step one two or four, if not, proceed to step one two or three;
Step one two or three: when Hurst index is for equaling 0.5, continuous time series are random walk sequence, terminate; When Hurst index is greater than 0 and is less than 0.5, continuous time series are short Memorability sequence, terminate;
Step one two or four: KPSS inspection is carried out to continuous time series, judge whether described continuous time series are short Memorability sequence, if, determine the evaluating characteristics kind of sequence input time be amplitude continuously-short Memorability time series, if not, continuous time series are that Long Memory Properties sequence or trend strengthen sequence, proceed to step one three;
Step one three: tendency and seasonal judgement are carried out to sequence input time:
The Long Memory Properties sequence of determining step 1 or trend strengthen sequence and whether have tendency and seasonality, if trendless and have seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-Long Memory Properties-seasonal time series, if have tendency and without seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-exponential trend time series or amplitude continuously-tendency-nonexponential trend time series or amplitude continuously-tendency-complicated trend time series, if trendless and without seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-Long Memory Properties-without seasonal time series, if have tendency and have seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-seasonal time series.
3. the forecast model system of selection of the time series evaluating characteristics of Corpus--based Method signal transacting according to claim 2, it is characterized in that, described step one three, if have tendency and without seasonality, then judge the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-exponential trend time series or amplitude continuously-tendency-nonexponential trend time series or amplitude continuously-tendency-complicated trend seasonal effect in time series method is:
Steps A: by input time sequence curve and the simple trend fitting curve comparison of setting, judged whether simple trend, if so, then proceeded to step B; If not, then determine the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-complicated trend time series;
Step B: judge sequence input time whether index of coincidence trend, if, then determine the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-exponential trend time series, if not, then determine the evaluating characteristics kind of sequence input time be amplitude continuously-tendency-nonexponential trend time series.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510324354.6A CN104992050A (en) | 2015-06-12 | 2015-06-12 | Method for selecting prediction model of time sequence characteristic evaluation based on statistical signal processing |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510324354.6A CN104992050A (en) | 2015-06-12 | 2015-06-12 | Method for selecting prediction model of time sequence characteristic evaluation based on statistical signal processing |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN104992050A true CN104992050A (en) | 2015-10-21 |
Family
ID=54303864
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510324354.6A Pending CN104992050A (en) | 2015-06-12 | 2015-06-12 | Method for selecting prediction model of time sequence characteristic evaluation based on statistical signal processing |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104992050A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107563426A (en) * | 2017-08-25 | 2018-01-09 | 清华大学 | A kind of learning method of locomotive operation temporal aspect |
| CN109687875A (en) * | 2018-11-20 | 2019-04-26 | 成都四方伟业软件股份有限公司 | A kind of time series data processing method |
| CN110032670A (en) * | 2019-04-17 | 2019-07-19 | 腾讯科技（深圳）有限公司 | Method for detecting abnormality, device, equipment and the storage medium of time series data |
| CN111639798A (en) * | 2020-05-26 | 2020-09-08 | 华青融天(北京)软件股份有限公司 | Intelligent prediction model selection method and device |
| CN114330908A (en) * | 2021-12-31 | 2022-04-12 | 中国民航信息网络股份有限公司 | Seat booking demand prediction method and device and revenue management system |
-
2015
- 2015-06-12 CN CN201510324354.6A patent/CN104992050A/en active Pending
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107563426A (en) * | 2017-08-25 | 2018-01-09 | 清华大学 | A kind of learning method of locomotive operation temporal aspect |
| CN109687875A (en) * | 2018-11-20 | 2019-04-26 | 成都四方伟业软件股份有限公司 | A kind of time series data processing method |
| CN109687875B (en) * | 2018-11-20 | 2023-03-31 | 成都四方伟业软件股份有限公司 | Time sequence data processing method |
| CN110032670A (en) * | 2019-04-17 | 2019-07-19 | 腾讯科技（深圳）有限公司 | Method for detecting abnormality, device, equipment and the storage medium of time series data |
| CN110032670B (en) * | 2019-04-17 | 2022-11-29 | 腾讯科技（深圳）有限公司 | Method, device and equipment for detecting abnormity of time sequence data and storage medium |
| CN111639798A (en) * | 2020-05-26 | 2020-09-08 | 华青融天(北京)软件股份有限公司 | Intelligent prediction model selection method and device |
| CN114330908A (en) * | 2021-12-31 | 2022-04-12 | 中国民航信息网络股份有限公司 | Seat booking demand prediction method and device and revenue management system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Flach | ROC analysis | |
| WO2020155752A1 (en) | Outlier detection model verification method and apparatus, and computer device and storage medium | |
| CN104992050A (en) | Method for selecting prediction model of time sequence characteristic evaluation based on statistical signal processing | |
| WO2020155756A1 (en) | Method and device for optimizing abnormal point proportion based on clustering and sse | |
| CN108549817A (en) | A kind of software security flaw prediction technique based on text deep learning | |
| CN110826618A (en) | Personal credit risk assessment method based on random forest | |
| Park et al. | Dual features functional support vector machines for fault detection of rechargeable batteries | |
| WO2024036709A1 (en) | Anomalous data detection method and apparatus | |
| WO2018006631A1 (en) | User level automatic segmentation method and system | |
| KR20200010624A (en) | Big Data Integrated Diagnosis Prediction System Using Machine Learning | |
| CN106599924A (en) | Classifier construction method based on three-way decision | |
| CN105488599B (en) | Method and device for predicting article popularity | |
| CN111291824A (en) | Time sequence processing method and device, electronic equipment and computer readable medium | |
| CN111340645B (en) | Improved association analysis method for power load | |
| CN106847306B (en) | Abnormal sound signal detection method and device | |
| CN109543712B (en) | Method for identifying entities on temporal data set | |
| CN110837853A (en) | Rapid classification model construction method | |
| CN115843366A (en) | Predicting compression ratio of data using compressible decisions | |
| Hai-Jun et al. | Fuzzy entropy clustering using possibilistic approach | |
| CN111652384B (en) | Balancing method for data volume distribution and data processing method | |
| CN114020905A (en) | Text classification external distribution sample detection method, device, medium and equipment | |
| CN114186644A (en) | Defect report severity prediction method based on optimized random forest | |
| CN114169542A (en) | Integrated learning tree construction method for incomplete data classification | |
| KR102479531B1 (en) | Method for training fault detection model of injection molding process and method for optimizing recipe of injection molding process | |
| CN115471056B (en) | Data transmission method and data transmission system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20151021 |