This appendix covers the following topics:
The forecast method specifies the statistical method used to generate the forecast.
Oracle Demand Planning offers an Automatic forecast method whereby the forecasting engine determines the best statistical forecasting method to use based on the historical performance of each algorithm and the application of decision rules.
You can also choose one of the following statistical forecasting methods:
Linear regression
Polynomial regression
Exponential fit
Logarithmic fit
Asymptotic fit
Exponential Asymptotic fit
Single Exponential Smoothing
Double Exponential Smoothing
Holt-Winters
Croston's Method
When you select a forecasting method, you have the option to specify values for advanced statistical parameters. This enables you to manually fine-tune the forecast.
Setting advanced parameters is discretionary. If you are using the Automatic method, the forecasting engine sets parameters for the selected forecast method when it runs the forecast. If you are using a method other than Automatic, the forecasting engine uses the defaults for parameters that are relevant to the selected method.
The following section lists the advanced statistical parameters in alphabetic order and briefly describes each one. The actual parameters that you see depend on the forecast method that you chose.
For the three methods of the exponential smoothing family (Single Exponential Smoothing, Double Exponential Smoothing, and Holt-Winters), specifies the relative weighting to give to recent changes in mean value.
Alpha Max — The minimum value is 0.0; the maximum value is 1.0; the default is .3.
Alpha Min (available for Double Exponential Smoothing and Holt-Winters) — The minimum value is 0.0; the maximum value is 1.0; the default is 1.0.
Alpha Step (available for Holt-Winters) — The incremental value used to go from Alpha Min to Alpha Max. The minimum value is 0.05; the maximum value is 0.2; the default is 0.1. The value that you enter must evenly divide the difference between Alpha Max and Alpha Min.
For the double exponential smoothing and Holt-Winters forecasting methods, specifies the relative weighting to give to recent changes in trend.
Beta Max — The minimum value is 0.0; the maximum value is 1.0; the default is .3.
Beta Min — The minimum value is 0.0; the maximum value is 1.0; the default is 1.0.
Beta Step — The incremental value used to go from Beta Min to Beta Max. The minimum value is 0.05; the maximum value is 0.2; the default is 0.1. The value that you enter must evenly divide the difference between Beta Max and Beta Min.
For the Automatic, linear and non-linear regression forecast methods, indicates how trends that are calculated based on history will be considered as the forecast time horizon increases. This parameter is useful when the history is large and some cyclical component has been identified.
The parameter value indicates how seriously deviations from baseline activity are considered: a higher value implies slower decay while a lower value implies faster decay for cyclical components. Note that for less history (for example, less than about 1.5 to 2 years) and in the absence of cyclical activity, this parameter might not have any effect on the calculated forecasts.
Cyclical Decay Max — The minimum value is 0.2; the maximum value is 1.0; the default is 1.0.
Cyclical Decay Min — The minimum value is 0.2 the maximum value is 1.0; the default is 0.2.
The difference between the maximum value and the minimum value must be evenly divisible by 4.
You can turn a seasonal or aggregate data filter on or off. You can choose one of the following options:
No seasonal filter — This is the default.
Seasonal filter — Accounts for seasonal patterns in the data.
Moving periodic total filter — An "aggregation" filter that handles sporadic or intermittent time series data. This is available for all methods except Holt-Winters.
For the Holt-Winters forecasting method, specifies the relative weighting to give to recent changes in seasonality.
Gamma Max — The minimum value is 0.0; the maximum value is 1.0; the default is .3.
Gamma Min — The minimum value is 0.0; the maximum value is 1.0; the default is 1.0.
Gamma Step — The incremental value used to go from Gamma Min to Gamma Max. The minimum value is 0.05; the maximum value is 0.2; the default is 0.1. The value that you enter must evenly divide the difference between Gamma Max and Gamma Min.
For all forecasting methods, specifies upper and lower bounds on forecast numbers as a factor or multiple of the historical values.
Minimum Forecast Factor — Sets lower bounds on forecast numbers. The default is 0.
Maximum Forecast Factor — Sets upper bounds on forecast numbers. The default is 100.
For all forecasting methods, specifies whether to prevent over-adjustment to the data by using average, rather than the individual, values of a period. For example, using this parameter for a forecast at the day level would allocate the forecast to all valid days in a weeks rather than forecasting individually at the day level.
For all forecasting methods, historical data is smoothed by averaging data with the time series. You can set the following smoothing parameters:
Median smoothing window — Specifies the length of the median smoothing period. Larger values result in smoother forecasts. Depending on the nature of the data, a window size that is too small might be unable to filter outliers while a window size that is too large might miss data patterns. The minimum value is 1; the maximum value is 27; the default is 3. The time period is based on the level/calendar and history start date.
Do you want to fit the data on smoothed series? — Specifies whether you want to turn the median smoothing filter on or off. The default is off.
Do you want to interpolate for missing values? — Specifies whether you want smooth the data by interpolating for missing values. This is useful for handling occasional missing values in the time series. The default is no.
For the Automatic, Double Exponential Smoothing and Holt-Winters methods, specifies parameters that determine how large trends detected from recent data affect the forecast.
Trend Min — The minimum value is 0.0; the maximum value is 1.0; the default is 0.4.
Trend Max — The minimum value is 0.0; the maximum value is 0.8; the default is 0.8.
Trend Step — The incremental value used to go from Trend Min to Trend Max. The minimum value is 0.05; the maximum value is 0.2; the default is 0.2. The value that you enter must evenly divide the difference between Min and Max.
Trend Dampening for erratic data — (Active if you chose "Moving Periodic Total Filter" as the data option for Data Filters) Specifies whether to apply trend dampening to erratic data.
For all forecasting methods, provides a ratio that specifies the portion of the data used in the verification phase. This ratio is used to calculate forecast accuracy statistics (MAD, MAPE, and RMSE). For the Automatic method, this ratio is also used to verify the best-fit method. Increasing the size means that the forecasting engine will use a larger portion of the most recent data; decreasing the size means that it will use a smaller portion of the data. The minimum value is 1/26; the maximum value is 1/2; the default is 1/3.
When you enable the Apply Unit of Measure (UOM) when aggregating data property for a measure, Oracle Demand Planning uses the Unit of Measure (UOM) for leaf level product items when aggregating product values to higher levels. For example, suppose that the base Unit of Measure (UOM) for the plan is "Each." If product A has a UOM of DZ (dozen) and a value of 10, and product B has a UOM of Each and a value of 20, then they would be aggregated to a higher level using the UOM conversion of 1 DZ = 12 Each, so 10*12 + 20*1 = 120 + 20 = 140.
You would almost always want to associate UOM's with a new measure, because for volume amounts, it makes the most sense to have them aggregate up this way.
An example of when you might not want to aggregate would be for something like a "Population" stream, for which the UOM's would not make sense.
The Base UOM for the plan is defined in the Demand Planning Server; however, the UOM for specific products may be unavailable or the conversion rates between the leaf products' UOM and the Base UOM for the plan may be unavailable. In these cases, the conversion rates will default to 1, so it will act like a flat aggregate from the leaf level of Product to higher levels. (In the previous example, it would just mean 10 + 20 = 30 if no conversion rates were available.)
When you apply events to a stored measure, Oracle Demand Planning applies the factor specified in the Action phase of the event definition process.
For all event types, events are applied at the day level for Time, values are allocated to the leaf level of the hierarchies, and then aggregated up.
When viewing events and their effects, ensure you are looking at the day level for Time. If you have qualified your event, then ensure that you select the qualified dimension value at the level you selected. When viewing events from other levels you will see an aggregated or allocated value which may not match what you expect to see.
When you associate a promotion event with a measure, Oracle Demand Planning applies the lift value (for each intersection, New Value = Existing Value + Lift) or the lift percent (for each intersection, New Value = Existing Value * [1 + {Lift/100}]) of the absolute lift (for each intersection, New Value = Lift).
When you associate a product introduction event with a measure, Oracle Demand Planning calculates the effect of the event as follows:
For a lifecycle event: Forecast for new product = Sum over all specified base products (specified weights applied to each base product * History for base product, with optional lag). Cannibalizations take away from the specified product a specified fraction of the forecast for the newly introduced product.
For a supercession event: Forecast for new product = Sum over all specified base products (specified weights applied to each product * Forecast for that base product).
When you associate a product phase out event with a measure, Oracle Demand Planning calculates the effect of the event as linear decay rate, with specified start and end dates and the factor of the start day's forecast remaining at the end day. The end day's value is the factor times the start day's value; values between start and end are linearly interpreted.
When a measure is associated with multiple events, Oracle Demand Planning applies the events in the following order:
New product introductions
Product phase outs
Optional promotion events
Mandatory promotion events
Optional events are calculated before mandatory events for the following reason: Assume that you are a store retailer, who offers a storewide 10% discount on Saturdays. The discount is a mandatory event — it happens to all transactions on that day. A customer comes in with a $50 voucher. The voucher is an optional event which may not happen for each transaction. The storewide discount is always taken before the voucher is used. In addition, its unlikely that the vendor would allow a customer to use two optional events for one transaction. For example, you couldn't use a $50 voucher and a voucher for 50% off anything at the same time.