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.
You can conditionally format forecasted values. For details, see Styling Reports.
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.
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:
Note: The word FORECAST and the opening parenthesis must be on the same line as the syntax 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.
nperiod * npredict
k=2/(1+npoint1)
g=2/(1+npoint2)
p=2/(1+npoint3)
ON TABLE SET STYLE *TYPE=DATA,COLUMN=MYFORECASTSORTFIELD,WHEN=FORECAST,COLOR=RED, $ENDSTYLE
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.
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:
Category PERIOD Unit Sales Dollar Sales EXPAVE -------- ------ ---------- ------------ ------ Coffee 1 61666 801123 801,123.0 2 54870 682340 741,731.5 3 61608 765078 753,404.8 4 57050 691274 722,339.4 5 59229 720444 721,391.7 6 58466 742457 731,924.3 7 60771 747253 739,588.7 8 54633 655896 697,742.3 9 57829 730317 714,029.7 10 57012 724412 719,220.8 11 51110 620264 669,742.4 12 58981 762328 716,035.2 13 0 0 739,181.6 14 0 0 750,754.8 15 0 0 756,541.4 Food 1 54394 672727 672,727.0 2 54894 699073 685,900.0 3 52713 642802 664,351.0 4 58026 718514 691,432.5 5 53289 660740 676,086.3 6 58742 734705 705,395.6 7 60127 760586 732,990.8 8 55622 695235 714,112.9 9 55787 683140 698,626.5 10 57340 713768 706,197.2 11 57459 710138 708,167.6 12 57290 705315 706,741.3 13 0 0 706,028.2 14 0 0 705,671.6 15 0 0 705,493.3
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.
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:
Category PERIOD Unit Sales DOLLARS -------- ------ ---------- ------- Coffee 1 61666 801,123.0 2 54870 682,340.0 3 61608 765,078.0 4 57050 691,274.0 5 59229 720,444.0 6 58466 742,457.0 7 60771 747,253.0 8 54633 655,896.0 9 57829 730,317.0 10 57012 724,412.0 11 51110 620,264.0 12 58981 762,328.0 13 0 694,975.6 14 0 719,879.4 15 0 705,729.9 Food 1 54394 672,727.0 2 54894 699,073.0 3 52713 642,802.0 4 58026 718,514.0 5 53289 660,740.0 6 58742 734,705.0 7 60127 760,586.0 8 55622 695,235.0 9 55787 683,140.0 10 57340 713,768.0 11 57459 710,138.0 12 57290 705,315.0 13 0 708,397.8 14 0 707,817.7 15 0 708,651.9
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:
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.
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:
Category PERIOD Unit Sales Dollar Sales EXPAVE -------- ------ ---------- ------------ ------ Coffee 1 61666 801123 801,123.0 2 54870 682340 741,731.5 3 61608 765078 753,404.8 4 57050 691274 722,339.4 5 59229 720444 721,391.7 6 58466 742457 731,924.3 7 60771 747253 739,588.7 8 54633 655896 697,742.3 9 57829 730317 714,029.7 10 57012 724412 719,220.8 11 51110 620264 669,742.4 12 58981 762328 716,035.2 13 0 0 739,181.6 14 0 0 750,754.8 15 0 0 756,541.4 Food 1 54394 672727 672,727.0 2 54894 699073 685,900.0 3 52713 642802 664,351.0 4 58026 718514 691,432.5 5 53289 660740 676,086.3 6 58742 734705 705,395.6 7 60127 760586 732,990.8 8 55622 695235 714,112.9 9 55787 683140 698,626.5 10 57340 713768 706,197.2 11 57459 710138 708,167.6 12 57290 705315 706,741.3 13 0 0 706,028.2 14 0 0 705,671.6 15 0 0 705,493.3
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:
n=3 (number used to calculate weights)
k = 2/(1+n) = 2/4 = 0.5
EXPAVE = (EXPAVE*(1-k))+(new-DOLLARS*k) = (801123*0.5) + (682340*0.50) = 400561.5 + 341170 = 741731.5
EXPAVE = (EXPAVE*(1-k))+(new-DOLLARS*k) = (741731.5*0.5)+(765078*0.50) = 370865.75 + 382539 = 753404.75
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
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.
DOUBLEXP(t) = k * datavalue(t) + (1-k) * ((DOUBLEXP(t-1) + b(t-1))
b(t) = g * (DOUBLEXP(t)-DOUBLEXP(t-1)) + (1 - g) * (b(t-1))
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:
The following sums the ACTUAL_YTD field of the CENTSTMT data source by period, and calculates a single exponential and double exponential moving average.
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:
YTD PERIOD Actual EXP DOUBLEXP ------ ------ --- -------- 2002/01 12,957,681. 12,957,681.0 12,957,681.0 2002/02 25,441,971. 19,199,826.0 22,439,246.3 2002/03 39,164,321. 29,182,073.5 34,791,885.1 2002/04 52,733,326. 40,957,699.8 48,845,816.0 2002/05 66,765,920. 53,861,809.9 63,860,955.9 2002/06 80,952,492. 67,407,150.9 79,188,052.9
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.
SEASONAL(t) = k * (datavalue(t)/I(t-L)) + (1-k) * (SEASONAL(t-1) + b(t-1))
b(t) = g * (SEASONAL(t)-SEASONAL(t-1)) + (1-g) * (b(t-1))
I(t) = p * (datavalue(t)/SEASONAL(t)) + (1 - p) * I(t-L)
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:
b(0) = (1/L) ((y(L+1)-y(1))/L + (y(L+2)-y(2))/L + ... + (y(2L) - y(L))/L )
A(j) = ( y((j-1)L+1) + y((j-1)L+2) + ... + y(jL) ) / 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:
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.
TRANSDATE TRANSTOT SEASONAL --------- -------- -------- 91/06/18 21.25 21.3 91/06/19 38.17 31.0 91/06/20 14.23 34.6 91/06/21 44.72 53.2 91/06/24 126.28 75.3 91/06/25 47.74 82.7 91/06/26 40.97 73.7 91/06/27 60.24 62.9 91/06/28 31.00 66.3 91/06/29 45.7 91/06/30 94.1 91/07/01 53.4 91/07/02 72.3 91/07/03 140.0 91/07/04 75.8 91/07/05 98.9 91/07/06 185.8 91/07/07 98.2
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:
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.
where:
Trend values, as well as predicted values, are calculated using the regression line equation.
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 is the independent variable (x) and MPG is the dependent variable (y). The equation is used to calculate MPGFORECAST trend and predicted values.
In this case, the equation is approximately as follows:
FORMPG = (-0.001323 * DEALER_COST) + 29.32
The predicted values are (the values are not exactly as calculated by FORECAST because of rounding, but they show the calculation process).
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 |
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.
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:
PER DOLLARS BUDGET DOLMOVAVE DOLEXPAVE BUDMOVAVE BUDEXPAVE --- ------- ------ --------- --------- --------- --------- 1 801123 801375 801,123.0 801,123.0 801,375.0 801,375.0 2 682340 725117 741,731.5 753,609.8 763,246.0 770,871.8 3 765078 810367 749,513.7 758,197.1 778,953.0 786,669.9 4 691274 717688 712,897.3 731,427.8 751,057.3 759,077.1 5 720444 739999 725,598.7 727,034.3 756,018.0 751,445.9 6 742457 742586 718,058.3 733,203.4 733,424.3 747,901.9 7 747253 773136 736,718.0 738,823.2 751,907.0 757,995.6 8 655896 685170 715,202.0 705,652.3 733,630.7 728,865.3 9 730317 753760 711,155.3 715,518.2 737,355.3 738,823.2 10 724412 709397 703,541.7 719,075.7 716,109.0 727,052.7 11 620264 630452 691,664.3 679,551.0 697,869.7 688,412.4 12 762328 718837 702,334.7 712,661.8 686,228.7 700,582.3
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:
Category PERIOD Unit Sales MOVAVE Dollar Sales -------- ------ ---------- ------ ------------ Coffee 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
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
Information Builders |