Note: | This topic concerns distribution fitting for forecasts. If you are using distribution fitting to select the best distribution type for an assumption, see Fitting Distributions to Historical Data. |
When analyzing a forecast chart, you can investigate some characteristics of the chart by determining the type of frequency distribution that fits it the best:
You can select Forecast, and then Fit Probability Distribution in the forecast chart menu bar to do a quick fit with the default or currently selected distributions and ranking method. You can also use this command to switch off distribution fitting that is set with either the Forecast menu or Preferences menu.
You can select Preferences, then Forecast, and then Forecast Window in the forecast chart menu bar to specify particular distributions and to select a fit ranking method. Then, you can also change the fit options or use Apply To to set these preferences for other forecasts.
To fit a probability distribution to a forecast chart using the Forecast command on the Preferences menu:
In the forecast chart menu bar, select Preferences, and then Forecast.
In the Forecast Window tab of the Forecast Preferences dialog, select Fit a probability distribution to the forecast and then click Fit Options.
Specify which distributions to fit:
AutoSelect performs a basic analysis of the data to select a distribution fitting option and ranking method. If the data includes only integers, fitting to all discrete distributions (with the exception of Yes-No) is completed using the Chi-square ranking statistic choice.
All Continuous fits the data to all of the built-in continuous distributions (these distributions are displayed as solid shapes on the Distribution Gallery).
All Discrete fits to all discrete distributions except Yes-No.
Choose displays another dialog from which you can select a subset of the distributions to include in the fitting.
Specify how the distributions should be ranked. In ranking the distributions, you can use one of three standard goodness-of-fit tests:
Anderson-Darling — This method closely resembles the Kolmogorov-Smirnov method, except that it weights the differences between the two distributions at their tails greater than at their mid-ranges. This weighting of the tails helps to correct the Kolmogorov-Smirnov method’s tendency to over-emphasize discrepancies in the central region.
Kolmogorov-Smirnov — The result of this test is essentially the largest vertical distance between the two cumulative distributions.
Chi-Square — This test is the oldest and most common of the goodness-of-fit tests. It gauges the general accuracy of the fit. The test breaks down the distribution into areas of equal probability and compares the data points within each area to the number of expected data points.
The first setting, AutoSelect, enables Crystal Ball to select the ranking statistic. If all data values are integers, Chi-square is selected.
Optional: If you know location, shape, or other parameter values that may help create a more accurate fit with certain distributions, select Lock Parameters and enter appropriate values in the Lock Parameters dialog (Locking Parameters When Fitting Distributions).
Optional: By default, values for all appropriate ranking statistics are calculated but only values for the selected ranking statistic are displayed in Goodness Of Fit view. To show values for all three statistics, select Show All Goodness-of-fit Statistics at the bottom of the Distribution Options panel.
During a simulation, Crystal Ball disables distribution fitting on forecast charts and overlay charts after 1,000 trials and until the simulation stops to enhance performance. A final fit is performed at end of the simulation.