Calculating Trends and Predicting Values With FORECAST

In this section:

You can calculate trends in numeric data and predict values beyond the range of those stored in the data source by using the FORECAST feature. FORECAST can be used in a report or graph request.

The calculations you can make to identify trends and forecast values are:

When predicting values in addition to calculating trends, FORECAST continues the same calculations beyond the data points by using the generated trend values as new data points. For the linear regression technique, the calculated regression equation is used to derive trend and predicted values.

FORECAST performs the calculations based on the data provided, but decisions about their use and reliability are the responsibility of the user. Therefore, FORECAST predictions are not always reliable, and many factors determine how accurate a prediction will be.


Top of page

x
FORECAST Processing

How to:

Reference:

You invoke FORECAST processing by including FORECAST in a RECAP command. In this command, you specify the parameters needed for generating estimated values, including the field to use in the calculations, the type of calculation to use, and the number of predictions to generate. The RECAP field that contains the result of FORECAST can be a new field (non-recursive) or the same field used in the FORECAST calculations (recursive):

FORECAST operates on the last ACROSS field in the request. If the request does not contain an ACROSS field, it operates on the last BY field. The FORECAST calculations start over when the highest-level sort field changes its value. In a request with multiple display commands, FORECAST operates on the last ACROSS field (or if there are no ACROSS fields, the last BY field) of the last display command. When using an ACROSS field with FORECAST, the display command must be SUM or COUNT.

Note: Although you pass parameters to FORECAST using an argument list in parentheses, FORECAST is not a function. It can coexist with a function of the same name, as long as the function is not specified in a RECAP command.



x
Syntax: How to Calculate Trends and Predict Values

MOVAVE calculation

ON sortfield RECAP result_field[/fmt] = FORECAST(infield, interval,
 npredict, 'MOVAVE',npoint1)sendstyle

EXPAVE calculation

ON sortfield RECAP result_field[/fmt] = FORECAST(infield, interval,
 npredict, 'EXPAVE',npoint1);

DOUBLEXP calculation

ON sortfield RECAP fld1[/fmt] = FORECAST(infield,
interval, npredict, 'DOUBLEXP',npoint1, npoint2);

SEASONAL calculation

ON sortfield RECAP fld1[/fmt] = FORECAST(infield,
interval, npredict, 'SEASONAL', nperiod, npoint1, npoint2, npoint3);

REGRESS calculation

ON sortfield RECAP result_field[/fmt] = FORECAST(infield, interval,
 npredict, 'REGRESS');

where:

sortfield
Is the last ACROSS field in the request. This field must be in numeric or date format. If the request does not contain an ACROSS field, FORECAST works on the last BY field.
result_field
Is the field containing the result of FORECAST. It can be a new field, or the same as infield. This must be a numeric field; either a real field, a virtual field, or a calculated field.

Note: The word FORECAST and the opening parenthesis must be on the same line as the syntax sortfield=.

fmt
Is the display format for result_field. The default format is D12.2. If result_field was previously reformatted using a DEFINE or COMPUTE command, the format specified in the RECAP command is respected.
infield
Is any numeric field. It can be the same field as result_field, or a different field. It cannot be a date-time field or a numeric field with date display options.
interval
Is the increment to add to each sortfield value (after the last data point) to create the next value. This must be a positive integer. To sort in descending order, use the BY HIGHEST phrase. The result of adding this number to the sortfield values is converted to the same format as sortfield.

For date fields, the minimal component in the format determines how the number is interpreted. For example, if the format is YMD, MDY, or DMY, an interval value of 2 is interpreted as meaning two days; if the format is YM, the 2 is interpreted as meaning two months.

npredict
Is the number of predictions for FORECAST to calculate. It must be an integer greater than or equal to zero. Zero indicates that you do not want predictions, and is only supported with a non-recursive FORECAST. For the SEASONAL method, npredict is the number of periods to calculate. The number of points generated is:
nperiod * npredict
nperiod
For the SEASONAL method, is a positive whole number that specifies the number of data points in a period.
npoint1
Is the number of values to average for the MOVAVE method. For EXPAVE, DOUBLEXP, and SEASONAL, this number is used to calculate the weights for each component in the average. This value must be a positive whole number. The weight, k, is calculated by the following formula:
k=2/(1+npoint1)
npoint2
For DOUBLEXP and SEASONAL, this positive whole number is used to calculate the weights for each term in the trend. The weight, g, is calculated by the following formula:
g=2/(1+npoint2)
npoint3
For SEASONAL, this positive whole number is used to calculate the weights for each term in the seasonal adjustment. The weight, p, is calculated by the following formula:
p=2/(1+npoint3)


x
Reference: Usage Notes for FORECAST


x
Reference: FORECAST Limits

The following are not supported with a RECAP command that uses FORECAST:


Top of page

x
Using a Simple Moving Average

A simple moving average is a series of arithmetic means calculated with a specified number of values from a field. Each new mean in the series is calculated by dropping the first value used in the prior calculation, and adding the next data value to the calculation.

Simple moving averages are sometimes used to analyze trends in stock prices over time. In this scenario, the average is calculated using a specified number of periods of stock prices. A disadvantage to this indicator is that because it drops the oldest values from the calculation as it moves on, it loses its memory over time. Also, mean values are distorted by extreme highs and lows, since this method gives equal weight to each point.

Predicted values beyond the range of the data values are calculated using a moving average that treats the calculated trend values as new data points.

The first complete moving average occurs at the nth data point because the calculation requires n values. This is called the lag. The moving average values for the lag rows are calculated as follows: the first value in the moving average column is equal to the first data value, the second value in the moving average column is the average of the first two data values, and so on until the nth row, at which point there are enough values to calculate the moving average with the number of values specified.



Example: Calculating a New Simple Moving Average Column

This request defines an integer value named PERIOD to use as the independent variable for the moving average. It predicts three periods of values beyond the range of the retrieved data.

DEFINE FILE GGSALES
 SDATE/YYM = DATE;
 SYEAR/Y = SDATE;
 SMONTH/M = SDATE;
 PERIOD/I2 = SMONTH;
END
TABLE FILE GGSALES
  SUM UNITS DOLLARS
  BY  CATEGORY BY PERIOD
  WHERE SYEAR EQ 97 AND CATEGORY NE 'Gifts'
  ON PERIOD RECAP MOVAVE/D10.1= FORECAST(DOLLARS,1,3,'MOVAVE',3);
END

The output is:

output

In the report, the number of values to use in the average is 3 and there are no UNITS or DOLLARS values for the generated PERIOD values.

Each average (MOVAVE value) is computed using DOLLARS values where they exist. The calculation of the moving average begins in the following way:

For predicted values beyond the supplied values, the calculated MOVAVE values are used as new data points to continue the moving average. The predicted MOVAVE values (starting with 694,975.6 for PERIOD 13) are calculated using the previous MOVAVE values as new data points. For example, the first predicted value (694,975.6) is the average of the data points from periods 11 and 12 (620,264 and 762,328) and the moving average for period 12 (702,334.7). The calculation is: 694,975 = (620,264 + 762,328 + 702,334.7)/3.



Example: Using an Existing Field as a Simple Moving Average Column

This request defines an integer value named PERIOD to use as the independent variable for the moving average. It predicts three periods of values beyond the range of the retrieved data. It uses the same name for the RECAP field as the first argument in the FORECAST parameter list. The trend values do not display in the report. The actual data values for DOLLARS are followed by the predicted values in the report column.

DEFINE FILE GGSALES
 SDATE/YYM = DATE;
 SYEAR/Y = SDATE;
 SMONTH/M = SDATE;
 PERIOD/I2 = SMONTH;
END
TABLE FILE GGSALES
  SUM UNITS DOLLARS
  BY  CATEGORY BY PERIOD
  WHERE SYEAR EQ 97 AND CATEGORY NE 'Gifts'
  ON PERIOD RECAP DOLLARS/D10.1 = FORECAST(DOLLARS,1,3,'MOVAVE',3);
END

The output is:

output


Top of page

x
Using Single Exponential Smoothing

The single exponential smoothing method calculates an average that allows you to choose weights to apply to newer and older values.

The following formula determines the weight given to the newest value.

k = 2/(1+n)

where:

k
Is the newest value.
n
Is an integer greater than one. Increasing n increases the weight assigned to the earlier observations (or data instances), as compared to the later ones.

The next calculation of the exponential moving average (EMA) value is derived by the following formula:

EMA = (EMA * (1-k)) + (datavalue * k)

This means that the newest value from the data source is multiplied by the factor k and the current moving average is multiplied by the factor (1-k). These quantities are then summed to generate the new EMA.

Note: When the data values are exhausted, the last data value in the sort group is used as the next data value.



Example: Calculating a Single Exponential Smoothing Column

The following defines an integer value named PERIOD to use as the independent variable for the moving average. It predicts three periods of values beyond the range of retrieved data.

DEFINE FILE GGSALES
 SDATE/YYM = DATE;
 SYEAR/Y = SDATE;
 SMONTH/M = SDATE;
 PERIOD/I2 = SMONTH;
END
TABLE FILE GGSALES
  SUM UNITS DOLLARS
  BY  CATEGORY BY PERIOD
  WHERE SYEAR EQ 97 AND CATEGORY NE 'Gifts'
  ON PERIOD RECAP EXPAVE/D10.1= FORECAST(DOLLARS,1,3,'EXPAVE',3);
END

The output is:

output

In the report, three predicted values of EXPAVE are calculated within each value of CATEGORY. For values outside the range of the data, new PERIOD values are generated by adding the interval value (1) to the prior PERIOD value.

Each average (EXPAVE value) is computed using DOLLARS values where they exist. The calculation of the moving average begins in the following way:

For predicted values beyond those supplied, the last EXPAVE value is used as the new data point in the exponential smoothing calculation. The predicted EXPAVE values (starting with 706,741.6) are calculated using the previous average and the new data point. Because the previous average is also used as the new data point, the predicted values are always equal to the last trend value. For example, the previous average for period 13 is 706,741.6, and this is also used as the next data point. Therefore, the average is calculated as follows: (706,741.6 * 0.5) + (706,741.6 * 0.5) = 706,741.6

EXPAVE = (EXPAVE * (1-k)) + (new-DOLLARS * k) = (706741.6*0.5) +
         (706741.6*0.50) =  353370.8 + 353370.8 = 706741.6

Top of page

x
Using Double Exponential Smoothing

Double exponential smoothing produces an exponential moving average that takes into account the tendency of data to either increase or decrease over time without repeating. This is accomplished by using two equations with two constants.

These two equations are solved to derive the smoothed average. The first smoothed average is set to the first data value. The first trend component is set to zero. For choosing the two constants, the best results are usually obtained by minimizing the mean-squared error (MSE) between the data values and the calculated averages. You may need to use nonlinear optimization techniques to find the optimal constants.

The equation used for forecasting beyond the data points with double exponential smoothing is

forecast(t+m) = DOUBLEXP(t) + m * b(t)

where:

m
Is the number of time periods ahead for the forecast.


Example: Calculating a Double Exponential Smoothing Column

The following defines an integer value named PERIOD to use as the independent variable for the moving average. The double exponential smoothing method estimates the trend in the data points better than the single smoothing method:

SET HISTOGRAM = OFF
TABLE FILE CENTSTMT
SUM ACTUAL_YTD
  BY PERIOD
  ON PERIOD RECAP EXP/D15.1 = FORECAST(ACTUAL_YTD,1,0,'EXPAVE',3);
  ON PERIOD RECAP DOUBLEXP/D15.1 = FORECAST(ACTUAL_YTD,1,0,
     'DOUBLEXP',3,3);
WHERE GL_ACCOUNT LIKE '3%%%'
END

The output is:

output

 


Top of page

x
Using Triple Exponential Smoothing

Triple exponential smoothing produces an exponential moving average that takes into account the tendency of data to repeat itself in intervals over time. For example, sales data that is growing and in which 25% of sales always occur during December contains both trend and seasonality. Triple exponential smoothing takes both the trend and seasonality into account by using three equations with three constants.

For triple exponential smoothing you, need to know the number of data points in each time period (designated as L in the following equations). To account for the seasonality, a seasonal index is calculated. The data is divided by the prior season index and then used in calculating the smoothed average.

These equations are solved to derive the triple smoothed average. The first smoothed average is set to the first data value. Initial values for the seasonality factors are calculated based on the maximum number of full periods of data in the data source, while the initial trend is calculated based on two periods of data. These values are calculated with the following steps:

  1. The initial trend factor is calculated by the following formula:
    b(0) = (1/L) ((y(L+1)-y(1))/L + (y(L+2)-y(2))/L + ... + (y(2L) - 
    y(L))/L )
  2. The calculation of the initial seasonality factor is based on the average of the data values within each period, A(j) (1<=j<=N):
    A(j) = ( y((j-1)L+1) + y((j-1)L+2) + ... + y(jL) ) / L
  3. Then, the initial periodicity factor is given by the following formula, where N is the number of full periods available in the data, L is the number of points per period and n is a point within the period (1<= n <= L):
    I(n) = ( y(n)/A(1) + y(L+n)/A(2) + ... + y((N-1)L+n)/A(N) ) / N

The three constants must be chosen carefully. The best results are usually obtained by choosing the constants to minimize the mean-squared error (MSE) between the data values and the calculated averages. Varying the values of npoint1 and npoint2 affect the results, and some values may produce a better approximation. To search for a better approximation, you may want to find values that minimize the MSE.

The equation used to forecast beyond the last data point with triple exponential smoothing is:

forecast(t+m) = (SEASONAL(t) + m * b(t)) / I(t-L+MOD(m/L))

where:

m
Is the number of periods ahead for the forecast.


Example: Calculating a Triple Exponential Smoothing Column

In the following, the data has seasonality but no trend. Therefore, npoint2 is set high (1000) to make the trend factor negligible in the calculation:

SET HISTOGRAM = OFF
TABLE FILE VIDEOTRK
SUM TRANSTOT
BY  TRANSDATE
ON TRANSDATE RECAP SEASONAL/D10.1 = FORECAST(TRANSTOT,1,3,'SEASONAL',
   3,3,1000,1);
WHERE TRANSDATE NE '19910617'
END

In the output, npredict is 3. Therefore, three periods (nine points, nperiod * npredict) are generated.

output


Top of page

x
Using a Linear Regression Equation

The Linear Regression Equation estimates values by assuming that the dependent variable (the new calculated values) and the independent variable (the sort field values) are related by a function that represents a straight line:

y = mx + b

where:

y
Is the dependent variable.
x
Is the independent variable.
m
Is the slope of the line.
b
Is the y-intercept.

REGRESS uses a technique called Ordinary Least Squares to calculate values for m and b that minimize the sum of the squared differences between the data and the resulting line.

The following formulas show how m and b are calculated.

formulas

formulas

where:

n
Is the number of data points.
y
Is the data values (dependent variables).
x
Is the sort field values (independent variables).

Trend values, as well as predicted values, are calculated using the regression line equation.



Example: Calculating a New Linear Regression Field
TABLE FILE CAR
PRINT MPG
BY DEALER_COST
WHERE MPG NE 0.0
  ON DEALER_COST RECAP FORMPG=FORECAST(MPG,1000,3,'REGRESS');
END

The output is:

DEALER_COST      MPG          FORMPG 
      2,886       27           25.51
      4,292       25           23.65
      4,631       21           23.20
      4,915       21           22.82
      5,063       23           22.63
      5,660       21           21.83
                  21           21.83
      5,800       24           21.65
      6,000       24           21.38
      7,427       16           19.49
      8,300       18           18.33
      8,400       18           18.20
     10,000       18           16.08
     11,000       18           14.75
     11,194        9           14.50
     14,940       11            9.53
     15,940        0            8.21
     16,940        0            6.88
     17,940        0            5.55

Note:

DEALER_COST

Calculation

FORMPG

15,940

(-0.001323 * 15,940) + 29.32

8.23

16,940

(-0.001323 * 16,940) + 29.32

6.91

17,940

(-0.001323 * 17,940) + 29.32

5.59


Top of page

x
FORECAST Reporting Techniques

You can use FORECAST multiple times in one request. However, all FORECAST requests must specify the same sort field, interval, and number of predictions. The only things that can change are the RECAP field, method, field used to calculate the FORECAST values, and number of points to average. If you change any of the other parameters, the new parameters are ignored.

If you want to move a FORECAST column in the report output, use an empty COMPUTE command for the FORECAST field as a placeholder. The data type (I, F, P, D) must be the same in the COMPUTE command and the RECAP command.

To make the report output easier to interpret, you can create a field that indicates whether the FORECAST value in each row is a predicted value. To do this, define a virtual field whose value is a constant other than zero. Rows in the report output that represent actual records in the data source will appear with this constant. Rows that represent predicted values will display zero. You can also propagate this field to a HOLD file.



Example: Generating Multiple FORECAST Columns in a Request

This example calculates moving averages and exponential averages for both the DOLLARS and BUDDOLLARS fields in the GGSALES data source. The sort field, interval, and number of predictions are the same for all of the calculations.

DEFINE FILE GGSALES
 SDATE/YYM = DATE;
 SYEAR/Y = SDATE;
 SMONTH/M = SDATE;
 PERIOD/I2 = SMONTH;
END
TABLE FILE GGSALES
  SUM DOLLARS AS 'DOLLARS' BUDDOLLARS AS 'BUDGET'
  BY CATEGORY NOPRINT BY PERIOD AS 'PER'
  WHERE SYEAR EQ 97 AND CATEGORY EQ 'Coffee'
  ON PERIOD RECAP DOLMOVAVE/D10.1= FORECAST(DOLLARS,1,0,'MOVAVE',3);
  ON PERIOD RECAP DOLEXPAVE/D10.1= FORECAST(DOLLARS,1,0,'EXPAVE',4);
  ON PERIOD RECAP BUDMOVAVE/D10.1 = FORECAST(BUDDOLLARS,1,0,'MOVAVE',3);
  ON PERIOD RECAP BUDEXPAVE/D10.1 = FORECAST(BUDDOLLARS,1,0,'EXPAVE',4);
END

The output is shown in the following image.

output



Example: Moving the FORECAST Column

The following example places the DOLLARS field after the MOVAVE field by using an empty COMPUTE command as a placeholder for the MOVAVE field. Both the COMPUTE command and the RECAP command specify formats for MOVAVE (of the same data type), but the format of the RECAP command takes precedence.

DEFINE FILE GGSALES
 SDATE/YYM = DATE;
 SYEAR/Y = SDATE;
 SMONTH/M = SDATE;
 PERIOD/I2 = SMONTH;
END
TABLE FILE GGSALES
SUM   UNITS
COMPUTE MOVAVE/D10.2 = ;
DOLLARS
  BY CATEGORY BY PERIOD
  WHERE SYEAR EQ 97 AND CATEGORY EQ 'Coffee'
  ON PERIOD RECAP MOVAVE/D10.1= FORECAST(DOLLARS,1,3,'MOVAVE',3);
END

The output is shown in the following image.

Category     PERIOD  Unit
Sales        MOVAVE  Dollar
SalesCoffee            1       61666     801,123.0        801123
                  2       54870     741,731.5        682340
                  3       61608     749,513.7        765078
                  4       57050     712,897.3        691274
                  5       59229     725,598.7        720444
                  6       58466     718,058.3        742457
                  7       60771     736,718.0        747253
                  8       54633     715,202.0        655896
                  9       57829     711,155.3        730317
                 10       57012     703,541.7        724412
                 11       51110     691,664.3        620264
                 12       58981     702,334.7        762328
                 13           0     694,975.6             0
                 14           0     719,879.4             0
                 15           0     705,729.9             0


Example: Distinguishing Data Rows From Predicted Rows

In the following example, the DATA_ROW virtual field has the value 1 for each row in the data source. It has the value zero for the predicted rows. The PREDICT field is calculated as YES for predicted rows, and NO for rows containing data.

DEFINE FILE CAR
DATA_ROW/I1 = 1;
END
TABLE FILE CAR
  PRINT DATA_ROW
COMPUTE PREDICT/A3 = IF DATA_ROW EQ 1 THEN 'NO' ELSE 'YES' ;
MPG
BY DEALER_COST
WHERE MPG GE 20
  ON DEALER_COST RECAP FORMPG/D12.2=FORECAST(MPG,1000,3,'REGRESS');
  ON DEALER_COST RECAP MPG         =FORECAST(MPG,1000,3,'REGRESS');
END

The output is:

DEALER_COST  DATA_ROW  PREDICT             MPG          FORMPG 
      2,886         1  NO                27.00           25.65
      4,292         1  NO                25.00           23.91
      4,631         1  NO                21.00           23.49
      4,915         1  NO                21.00           23.14
      5,063         1  NO                23.00           22.95
      5,660         1  NO                21.00           22.21
                    1  NO                21.00           22.21
      5,800         1  NO                24.20           22.04
      6,000         1  NO                24.20           21.79
      7,000         0  YES               20.56           20.56
      8,000         0  YES               19.32           19.32
      9,000         0  YES               18.08           18.08

WebFOCUS