Best Fit Analysis Rules
When Best Fit Analysis is run, it uses rules to determine whether to replace, eliminate, or keep forecast methods in place. The Best Fit Management page contains an additional Status field. This Status field displays a message indicating the results of a rule run on a forecast method; specifically if a forecast method is replaced or eliminated.
• If Best Fit Analysis encounters a hard rule, the application does not generate a forecast or forecast errors, and the Status field displays the appropriate message.
• If Best Fit encounters a soft rule, the application generates a forecast and MAPE, MAD, and RMSE. Composite error is not calculated. The Status field displays the appropriate message.
• If no rules are encountered, the application generates a normal forecast and forecast errors. The Status field does not display a message.
 |
If no message is displayed in the Status field, the normal forecast is generated.
|
The following table provides the hard and soft rules used by Best Fit Analysis.
|
Rule Type
|
Status
|
Description
|
|---|---|---|
|
Soft
|
Trend detected
|
Trend detected. Average eliminated.
|
|
Soft
|
Trend detected
|
Trend detected. Weighted Average eliminated.
|
|
Soft
|
Trend detected
|
Trend detected. Moving Average eliminated.
|
|
Soft
|
LastYearAvg < GS
|
LastYearAverage less than the value of the BEST_FIT_MOVING_AVG_MIN_DEMAND. Moving Average eliminated.
|
|
Soft
|
Trend detected
|
Trend detected. Same As Last Year eliminated.
|
|
Soft
|
Trend detected
|
Trend detected. Single Exponential eliminated.
|
|
Soft
|
#HistSlices < x
|
Number of history slices less than val(minCrostonsHist). Croston's and Intermittence Smoothing eliminated.
|
|
Soft
|
Not intermittent
|
Intermittency test failed. Croston's and Intermittence Smoothing eliminated.
|
|
Soft
|
#HistSlices < z
|
Number of history slices less than val(minDoubleExpHistory). Double Exponential eliminated.
|
|
Soft
|
TotalDemandSlices < 5
|
Number of history slices less than 5. Double Exponential eliminated.
|
|
Soft
|
Horizon forecast < GS
|
The horizon forecast is less than the value of the BESTFIT_MIN_FCST global setting.
|
|
Hard
|
NumHistDemands < 5
|
Last 5 demand slices contains at least one zero demand. Linear Regression eliminated.
|
|
Hard
|
NumHistDemands < 2
|
Total number of historical demands less than 2. Croston's and Intermittence Smoothing eliminated.
|
|
Hard
|
#StreamHistSlices < 12
|
Stream config history slices less than 12. Croston's and Intermittence Smoothing eliminated.
|
|
Hard
|
Insufficient History
|
# of history slices less than 12. Croston's and Intermittence Smoothing eliminated.
|
|
Hard
|
Insufficient History
|
Num History Slices less than 12. Same As Last Year eliminated.
|
|
Hard
|
#StreamHistSlices < 12
|
Stream config history slices less than 12. Same As Last Year eliminated.
|
|
Hard
|
#StreamHistSlices < x
|
Stream config history slices less than the INIT_EXPSMOOTHING_MONTHS global setting + 3. Double Exponential eliminated.
|
|
Hard
|
#HistSlices < y
|
# of history slices less than the INIT_EXPSMOOTHING_MONTHS global setting + 3. Double Exponential eliminated.
|
|
Hard
|
LastYearAverage < 2
|
LastYearAverage less than 2. Winters Multiplicative eliminated.
|
|
Hard
|
#StreamHistSlices < 12
|
Stream config history slices less than 12. Winters Multiplicative eliminated.
|
|
Hard
|
Insufficient History
|
Number of history slices less than 12. Winters Multiplicative eliminated.
|
|
Hard
|
#HistSlices < x
|
Number of history slices less than val(minWintersHist). Winters Multiplicative eliminated.
|
|
Hard
|
Not seasonal
|
Auto correlation test failed. Winters Multiplicative eliminated.
|
|
Hard
|
#Slides < 1
|
Number of slides less than 1. All forecast methods eliminated.
Number of slides refers to the number of passes the best fit forecasting process takes in evaluating the best forecast method.
Example:
For a SKU, assume Minimum # of History Slices = 12, Total # of History Slices = 30, and Hold Out Window = 4. The best fit forecasting process uses 12 months of history at a time and generates forecasts and forecast errors over the holdout window period. Best fit forecast and errors (MAPE, MAD, and RMSE) are generated for each forecast method, 4 slices at a time.
This process is repeated for each slide. This means that the best fit forecasting process slides forward one month (default) and then generates the forecast for the next 4 slices. In this case the total number of slides equals (30 - (12 + 4)) = 15. For each forecast method, the best fit forecasting process generates MAPE, MAD, and RMSE 15 times. The calculated MAD, MAPE, and RMSE for each forecast method (as seen on the Best Fit Management page) is the average error taken across 15 slides.
|
|
Hard
|
LastForwardStartDate = Null
|
LastForwardStartDate = Null. All forecast methods eliminated.
|