System and methods for forecasting time series with multiple seasonal patterns

• US 7,987,106 B1
• Filed: 06/05/2006
• Issued: 07/26/2011
• Est. Priority Date: 06/05/2006
• Status: Active Grant

• Current Assignee: inContact Incorporated (Nice Ltd)
• Original Assignee: Turgut Aykin
• Inventors: Aykin, Turgut
• Primary Examiner(s): Jeanty; Romain
• Assistant Examiner(s): FEENEY, BRETT A
• Application Number:
  US11/446,711
• Time in Patent Office: 1,877 Days
• Field of Search: None
• US Class Current: 705/7.11
• CPC Class Codes: G06Q 10/063 Operations research, analysis or management

Abstract

A computer implemented system and methods for forecasting the future values of a time series with multiple seasonal cycles are described. The forecasting method of the present invention include 1) two multi-seasonal exponential smoothing methods, 2) the seasonal sub-series method splitting a time series into multiple sub-series, and 3) the aggregate-to-period forecasting method.

Claims

1. A computer implemented method for forecasting the future values of a time series with multiple seasonal patterns comprising:

executing, by one or more computers, the steps of;

(i) providing a plurality of data values for the time series, together with time information relating to said data values;

(ii) inputting history start and end dates for the data values inputted to be used in forecasting;

(iii) inputting a forecast horizon start and end dates to specify over which dates forecasts are to be generated;

(iv) inputting a length of a cycle of the first seasonal pattern and a length of a cycle of the second seasonal pattern wherein the cycle of the second seasonal pattern is different than the cycle of the first seasonal pattern;

(v) determining whether the effect of the first seasonal pattern and the second seasonal pattern, are additive or multiplicative;

(vi) inputting the values of smoothing constants, all between 0 and 1;

(vii) using a hybrid double seasonal exponential smoothing level prediction equation recursively starting from the oldest observation to the newest observation to forecast a level value when the effect of the first seasonal pattern is multiplicative and the second seasonal pattern is additive or using an additive double seasonal exponential smoothing level prediction equation to forecast a level value when the effect of the first seasonal pattern and the second seasonal pattern are both additive;

(viii) using a hybrid double seasonal exponential smoothing trend prediction equation recursively starting from the oldest observation to the newest to forecast a trend value when the effect of the first seasonal pattern is multiplicative and the second seasonal pattern is additive or using an additive double seasonal exponential smoothing trend prediction equation to forecast a trend value when the effect of the first seasonal pattern and the second seasonal pattern are both additive;

(ix) using two hybrid double seasonal exponential smoothing seasonal factor prediction equations recursively starting from the oldest observation to the newest to forecast seasonal factors for each seasonal cycle of the first seasonal pattern and the second seasonal pattern when the effect of the first seasonal pattern is multiplicative and the second seasonal pattern is additive or using an additive double seasonal exponential smoothing seasonal factor prediction equation to forecast seasonal factors for each seasonal cycle of the first seasonal pattern and the second pattern when the effect of the first seasonal pattern and the second seasonal pattern are both additive;

(x) including in steps (iv) to (ix) any additional seasonal patterns that exist in the time series;

(xi) combining the predicted values of level, trend, and seasonal factors for all seasonal patterns and calculating forecasts of the future values of the time series;

(xii) storing the forecasts of the future values of the time series.

2. The method of claim 1 wherein the time series exhibits at least two seasonal patterns.

3. The method of claim 2 wherein the future effects of seasonal patterns may be additive for some seasonal patterns and multiplicative for the others.

4. The method of claim 1 wherein the future values of the level, trend, and seasonal factors are calculated, if some seasonal patterns are additive and some are multiplicative, as follows:

(xii) the level value is calculated as follows;

L(t)=α

((D(t)−

S(t−

m))/C(t−

n))+(1−

α

)(L(t−

1)+T(t−

1))

where

D(t) is an actual recorded value of the time series at time t,

L(t);

is an estimate of the level of the times series at time t,

T(t);

is an estimate of the slope of the time series at time t,

S(t);

is an estimate of the effects of Seasonal Pattern 1 at time t,

C(t);

is an estimate of the effects of Seasonal Pattern 2 at time t,

m;

is a length (in periods) of a cycle of Seasonal Pattern 1,

n;

is a length (in periods) of a cycle of Seasonal Pattern 2,

α

;

is a smoothing constant between zero and one; and

(xiii) the trend value is calculated as follows;

T(t)=β

(L(t)−

L(t−

1))+(1−

β

)T(t−

1)

where

β

;

is a smoothing constant between zero and one; and

(xiv) the seasonal factor values for the seasonal patterns with additive seasonality are calculated as follows;

S(t)=π

(D(t)−

(L(t)C(t−

n)))+(1−

π

)S(t−

m)

where

π

;

is a smoothing constant between zero and one; and

(xv) the seasonal factor values for the seasonal patterns with multiplicative seasonality are calculated as follows;

C(t)=γ

((D(t)−

S(t−

m))/L(t)))+(1−

γ

)C(t−

n)

where

γ

;

is a smoothing constant between zero and one.

5. The method of claim 4 wherein forecasts for a period that is τ

(τ

≧

0) periods from time t for which the last recorded value of the time series exists are calculated as follows;

F(t+τ

)=(L(t)+τ

T(t))C(t−

n+τ

)+S(t−

m+τ

).

6. The method of claim 1 wherein the future values of the level, trend, and seasonal factors are calculated, if all seasonal patterns are additive, as follows:

(xvi) the level value is calculated as follows;

L(t)=α

((D(t)−

S(t−

m)−

C(t−

n))+(1−

α

)(L(t−

1)+T(t−

1));

and

(xvii) the trend value is calculated as follows;

T(t)=β

(L(t)−

L(t−

1))+(1−

β

)T(t−

1);

and

(xviii) the seasonal factor values for Seasonal Pattern 1 are calculated as follows;

S(t)=π

(D(t)−

(L(t)+C(t−

n)))+(1−

π

)S(t−

m);

and

(xix) the seasonal factor values for Seasonal Pattern 2 and all other seasonal patterns other than Seasonal Patterns 1 and 2 are calculated as follows;

C(t)=γ

(D(t)−

(L(t)+S(t−

m)))+(1−

γ

)C(t−

n).

7. The method of claim 6 wherein the forecasts are for a period that is τ

(τ

≧

0) periods from time t for which the last recorded value of the time series exists are calculated as follows;

F(t+τ

)=(L(t)+τ

T(t))+C(t−

n+τ

)+S(t−

m+τ

).

8. A computer implemented method for forecasting the future values of a time series with multiple seasonal patterns of claim 1 further comprising the steps of:

(xiii) finding a seasonal cycle with a shortest cycle length;

(xiv) splitting the time series into sub-series, one for each period in a cycle of the seasonal pattern with the shortest cycle length;

(xv) selecting one or more forecasting methods that, in addition to trend, can forecast additive or multiplicative seasonal factors, for one or more seasonal patterns and include them in forecasting the future values of a time series;

(xvi) using the selected forecasting methods to forecast the future values of each sub-series by including the effects of trend and all seasonal patterns that exist in the sub-series;

(xvii) combining the forecasts of all sub-series in a correct time order to obtain the forecasts of the future values of the time series;

(xviii) storing the forecasts of the future values of the time series.

9. The method of claim 8 wherein the time series exhibits at least two seasonal cycles.

10. The method of claim 8 wherein, a set of forecasts for each sub-series is generated independently as:

Sub-series 1;

{F_(1+km+m), F_(1+km+2m), F_(1+km+3m), . . . , F_(1+km+Tm)},

Sub-series 2;

{F_(2+km+m), F_(2+km+2m), F_(2+km+3m), . . . , F_(2+km+Tm)},

Sub-series m;

{F_(m+km+m), F_(m+km+2m), F_(m+km+3m), F_(m+km+Tm)},

where

F_(i+km+jm) is the forecast for the subseries in the j^th future cycle; and

the subseries forecasts are merged to obtain the forecasts for the whole time series as follows

{F_1+km+m), F_(2+km+m), . . . , F_(m+km+m), F_(1+km+2m), F_(2+km+2m), . . . , F_(m+km+2m), F_(1+km+3m), F_(2+km+3m), . . . , F_(m+km+3m), . . . , F_(1+km+Tm), F_(2+km+Tm) . . . , F_(m+km+Tm)}.

11. The computer implemented method for forecasting the future values of a time series with multiple seasonal patterns claim 1 further comprising the steps of:

(xiii) finding the seasonal cycle with the shortest cycle length;

(xiv) for each cycle of a seasonal pattern with a shortest cycle length, aggregating the time series values observed in that cycle to form a time series of aggregated values for cycles aggregated into aggregated periods;

(xv) selecting a first forecasting method that can, in addition to trend, forecast additive or multiplicative seasonal factors, for one or more seasonal patterns and include them in forecasting the future values of the time series;

(xvi) using the first forecasting method to forecast the future values of the aggregated time series with one or more of the seasonal cycles exhibited by the aggregated time series;

(xvii) selecting a second forecasting method that, in addition to trend, can forecast the seasonal factor values for the periods of the seasonal cycle with the shortest cycle length;

(xviii) using the second forecasting method to obtain the period forecasts for the periods of the seasonal cycle with the shortest cycle length using an un-aggregated time series;

(xix) using the period forecasts to forecast the period factor values for the periods of the seasonal cycle with the shortest cycle length;

(xx) breaking the forecasts of the aggregated time series by multiplying the period factors for each period of a cycle with the aggregate period forecast for the cycle to obtain the forecasts for the periods in the future cycle of the seasonal pattern with the shortest length;

(xxi) storing the forecasts of the future values of the time series.

12. The method of claim 11 wherein the time series exhibits at least two seasonal cycles.

13. The method of claim 11 wherein the time series forecasted may consist of values of a statistic calculated over a period of time.

14. The method of claim 13 wherein an aggregate statistic for an aggregate period is obtained by re-calculating it over the aggregate period.

15. The method of claim 13 wherein the period factors for the past periods are obtained by taking the ratio of the value of the statistic for a period to the value of the statistic for the aggregate period including it.

16. The method of claim 13 wherein the period factor forecasts are obtained by applying a third forecasting method to the past period factors.

17. A non-transitory computer readable medium including a computer program stored thereon when executed by a computer performs a method of forecasting the future values of a time series, comprising:

(i) providing a plurality of data values for the time series, together with time information relating to said data values;

(ii) inputting history start and end dates for the data values inputted to be used in forecasting;

(iii) inputting a forecast horizon start and end dates to specify over which dates forecasts are to be generated;

(iv) inputting a length of a cycle of the first seasonal pattern and a length of a cycle of the second seasonal pattern wherein the cycle of the second seasonal pattern is different than the cycle of the first seasonal pattern;

(v) determining whether the effect of the first seasonal pattern and the second seasonal pattern, are additive or multiplicative;

(vi) inputting the values of smoothing constants, all between 0 and 1;

(vii) using a hybrid double seasonal exponential smoothing level prediction equation recursively starting from the oldest observation to the newest observation to forecast a level value when the effect of the first seasonal pattern is multiplicative and the second seasonal pattern is additive or using an additive double seasonal exponential smoothing level prediction equation to forecast a level value when the effect of the first seasonal pattern and the second seasonal pattern are both additive;

(viii) using a hybrid double seasonal exponential smoothing trend prediction equation recursively starting from the oldest observation to the newest to forecast a trend value when the effect of the first seasonal pattern is multiplicative and the second seasonal pattern is additive or using an additive double seasonal exponential smoothing trend prediction equation to forecast a trend value when the effect of the first seasonal pattern and the second seasonal pattern are both additive;

(ix) using two hybrid double seasonal exponential smoothing seasonal factor prediction equations recursively starting from the oldest observation to the newest to forecast seasonal factors for each seasonal cycle of the first seasonal pattern and the second seasonal pattern when the effect of the first seasonal pattern is multiplicative and the second seasonal pattern is additive or using an additive double seasonal exponential smoothing seasonal factor prediction equation to forecast seasonal factors for each seasonal cycle of the first seasonal pattern and the second pattern when the effect of the first seasonal pattern and the second seasonal pattern are both additive;

(x) including in steps (iv) to (ix) any additional seasonal patterns that exist in the time series;

(xi) combining the predicted values of level, trend, and seasonal factors for all seasonal patterns and calculating forecasts of the future values of the time series;

(xii) storing the forecasts of the future values of the time series.

18. The non-transitory computer readable medium of claim 17 further including the program stored thereon when executed by the computer performs the method of forecasting the future values of a time series further comprising:

determining a seasonal pattern with the shortest length;

splitting the time series into sub-series, wherein one said sub-series is selected for each seasonal period in the seasonal cycle with the shortest cycle length;

selecting a third forecasting method that can forecast seasonal factors and include them in forecasting the future values of a time series;

using the third forecasting method to forecast future values of each sub-series with one or more of the seasonal cycles exhibited by each sub-series being predicted and included in forecast by the selected method;

combining the forecast of all sub-series in a correct time order to obtain forecasts of the future values of the time series.

19. The non-transitory computer readable medium of claim 17 further including the program stored thereon when executed by the computer performs the method of forecasting the future values of a time series further comprising:

determining a seasonal pattern with the shortest length;

aggregating the time series values observed in each cycle of the seasonal pattern with the shortest length to form a time series of aggregated values with the periods in each cycle aggregated into an aggregated period;

selecting a fourth forecasting method;

using the fourth selected forecasting method to forecast the future values of the aggregated time series with one or more of the seasonal cycles exhibited by the aggregated time series being forecasted;

selecting a fifth forecasting method that, in addition to trend, can forecast the period factor values for the periods of the seasonal cycle with the shortest cycle length;

using the fifth selected forecasting method and the smoothed time series found after applying the outlier detection algorithm to forecast the period factor values for the periods of the seasonal pattern with the shortest cycle length;

breaking the forecasts of the aggregated time series using the period factor values obtained so that each aggregated forecast value generates the forecasts for the periods in the corresponding seasonal cycle (with the shortest cycle length) of the time series in the future.