Exponential Smoothing is a method to smooth real values in time series in order to forecast probable future values.

Exponential Triple Smoothing (ETS) is a set of algorithms in which both trend and periodical (seasonal) influences are processed. Exponential Double Smoothing (EDS) is an algorithm like ETS, but without the periodical influences. EDS produces linear forecasts.

Padoma ikona

See the Wikipedia on Exponential smoothing algorithms for more information.


target (mandatory): A date, time or numeric single value or range. The data point/range for which to calculate a forecast.

values (mandatory): A numeric array or range. values are the historical values, for which you want to forecast the next points.

timeline (mandatory): A numeric array or range. The time line (x-value) range for the historical values.

Piezīmes ikona

The time line doesn't have to to be sorted, the functions will sort it for calculations.
The time line values must have a consistent step between them.
If a constant step can't be identified in the sorted time line, the functions will return the #NUM! error.
If the ranges of the time line and historical values aren't of same size, the functions will return the #N/A error.
If the time line contains less than 2 periods of data, the functions will return the #VALUE! Error.


data_completion (optional): a logical value TRUE or FALSE, a numeric 1 or 0, default is 1 (TRUE). A value of 0 (FALSE) will add missing data points with zero as historical value. A value of 1 (TRUE) will add missing data points by interpolating between the neighboring data points.

Piezīmes ikona

Although the time line requires a constant step between data points, the function support up to 30% missing data points, and will add these data points.


aggregation (optional): A numeric value from 1 to 7, with default 1. The aggregation parameter indicates which method will be used to aggregate identical time values:

Aggregation

Funkcija

1

AVERAGE

2

COUNT

3

COUNTA

4

MAX

5

MEDIAN

6

MIN

7

SUM


Piezīmes ikona

Although the time line requires a constant step between data points, the functions will aggregate multiple points which have the same time stamp.


stat_type (mandatory): A numerical value from 1 to 9. A value indicating which statistic will be returned for the given values and x-range.

The following statistics can be returned:

stat_type

Statistics

1

Alpha smoothing parameter of ETS algorithm (base)

2

Gamma smoothing parameter of ETS algorithm (trend)

3

Beta smoothing parameter of ETS algorithm (periodic deviation)

4

Mean absolute scaled error (MASE) - a measure of the accuracy of forecasts.

5

Symmetric mean absolute percentage error (SMAPE) - an accuracy measure based on percentage errors.

6

Mean absolute error (MAE) – a measure of the accuracy of forecasts.

7

Root mean squared error (RMSE) - a measure of the differences between predicted and observed values.

8

Step size detected time line (x-range). When a stepsize in months/quarters/years is detected, the stepsize is in months, otherwise the stepsize is in days in case of date(time) timeline and numeric in other cases.

9

Number of samples in period – this is the same as argument period_length, or the calculated number in case of argument period_length being 1.


confidence_level (mandatory): A numeric value between 0 and 1 (exclusive), default is 0.95. A value indicating a confidence level for the calculated prediction interval.

Piezīmes ikona

With values <= 0 or >= 1, the functions will return the #NUM! Error.


period_length (optional): A numeric value >= 0, the default is 1. A positive integer indicating the number of samples in a period.

Piezīmes ikona

A value of 1 indicates that Calc is to determine the number of samples in a period automatically.
A value of 0 indicates no periodic effects, a forecast is calculated with EDS algorithms.
For all other positive values, forecasts are calculated with ETS algorithms.
For values that not being a positive whole number, the functions will return the #NUM! Error.


forecast = basevalue + trend * ∆x + periodical_aberration.

forecast = ( basevalue + trend * ∆x ) * periodical_aberration.

Example

The table below contains a timeline and its associated values:

A

B

1

Laika skala

Vērtības

2

01/2013

112

3

02/2013

118

4

03/2013

132

5

04/2013

100

6

05/2013

121

7

06/2013

135

8

07/2013

148

9

08/2013

148

10

09/2013

136

11

10/2013

119

12

11/2013

104

13

12/2013

118


Please support us!